[PATCH 0/24] make atomic_read() behave consistently across all architectures

[PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

As recent discussions[1], and bugs[2] have shown, there is a great deal of
confusion about the expected behavior of atomic_read(), compounded by the
fact that it is not the same on all architectures.  Since users expect calls
to atomic_read() to actually perform a read, it is not desirable to allow
the compiler to optimize this away.  Requiring the use of barrier() in this
case is inefficient, since we only want to re-load the atomic_t variable,
not everything else in scope.

This patchset makes the behavior of atomic_read uniform by removing the
volatile keyword from all atomic_t and atomic64_t definitions that currently
have it, and instead explicitly casts the variable as volatile in
atomic_read().  This leaves little room for creative optimization by the
compiler, and is in keeping with the principles behind "volatile considered
harmful".

Busy-waiters should still use cpu_relax(), but fast paths may be able to
reduce their use of barrier() between some atomic_read() calls.

	-- Chris

1)	http://lkml.org/lkml/2007/7/1/52
2)	http://lkml.org/lkml/2007/8/8/122

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arnd Bergmann]

On Thursday 09 August 2007, Chris Snook wrote:
> This patchset makes the behavior of atomic_read uniform by removing the
> volatile keyword from all atomic_t and atomic64_t definitions that currently
> have it, and instead explicitly casts the variable as volatile in
> atomic_read().  This leaves little room for creative optimization by the
> compiler, and is in keeping with the principles behind "volatile considered
> harmful".
> 

Just an idea: since all architectures already include asm-generic/atomic.h,
why not move the definitions of atomic_t and atomic64_t, as well as anything
that does not involve architecture specific inline assembly into the generic
header?

	Arnd <><

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Arnd Bergmann wrote:
> On Thursday 09 August 2007, Chris Snook wrote:
>> This patchset makes the behavior of atomic_read uniform by removing the
>> volatile keyword from all atomic_t and atomic64_t definitions that currently
>> have it, and instead explicitly casts the variable as volatile in
>> atomic_read().  This leaves little room for creative optimization by the
>> compiler, and is in keeping with the principles behind "volatile considered
>> harmful".
>>
> 
> Just an idea: since all architectures already include asm-generic/atomic.h,
> why not move the definitions of atomic_t and atomic64_t, as well as anything
> that does not involve architecture specific inline assembly into the generic
> header?
> 
> 	Arnd <><

a) chicken and egg: asm-generic/atomic.h depends on definitions in asm/atomic.h

If you can find a way to reshuffle the code and make it simpler, I personally am 
all for it.  I'm skeptical that you'll get much to show for the effort.

b) The definitions aren't precisely identical between all architectures, so it 
would be a mess of special cases, which gets us right back to where we are now.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arnd Bergmann]

On Thursday 09 August 2007, Chris Snook wrote:
> a) chicken and egg: asm-generic/atomic.h depends on definitions in asm/atomic.h

Ok, I see.
 
> If you can find a way to reshuffle the code and make it simpler, I personally am 
> all for it. I'm skeptical that you'll get much to show for the effort. 

I guess it could  be done using more macros or new headers, but I don't see
a way that would actually improve the situation.

> b) The definitions aren't precisely identical between all architectures, so it 
> would be a mess of special cases, which gets us right back to where we are now.

Why are they not identical? Anything beyond the 32/64 bit difference should
be the same afaics, or it might cause more bugs.

	Arnd <><

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 9 Aug 2007, Chris Snook wrote:

> This patchset makes the behavior of atomic_read uniform by removing the
> volatile keyword from all atomic_t and atomic64_t definitions that currently
> have it, and instead explicitly casts the variable as volatile in
> atomic_read().  This leaves little room for creative optimization by the
> compiler, and is in keeping with the principles behind "volatile considered
> harmful".

volatile is generally harmful even in atomic_read(). Barriers control
visibility and AFAICT things are fine.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Christoph Lameter wrote:
> On Thu, 9 Aug 2007, Chris Snook wrote:
> 
>> This patchset makes the behavior of atomic_read uniform by removing the
>> volatile keyword from all atomic_t and atomic64_t definitions that currently
>> have it, and instead explicitly casts the variable as volatile in
>> atomic_read().  This leaves little room for creative optimization by the
>> compiler, and is in keeping with the principles behind "volatile considered
>> harmful".
> 
> volatile is generally harmful even in atomic_read(). Barriers control
> visibility and AFAICT things are fine.

But barriers force a flush of *everything* in scope, which we generally don't 
want.  On the other hand, we pretty much always want to flush atomic_* 
operations.  One way or another, we should be restricting the volatile behavior 
to the thing that needs it.  On most architectures, this patch set just moves 
that from the declaration, where it is considered harmful, to the use, where it 
is considered an occasional necessary evil.

See the resubmitted patchset, which also puts a cast in the atomic[64]_set 
operations.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Tue, 14 Aug 2007, Chris Snook wrote:

> But barriers force a flush of *everything* in scope, which we generally don't
> want.  On the other hand, we pretty much always want to flush atomic_*
> operations.  One way or another, we should be restricting the volatile
> behavior to the thing that needs it.  On most architectures, this patch set
> just moves that from the declaration, where it is considered harmful, to the
> use, where it is considered an occasional necessary evil.

Then we would need

	atomic_read()

and

	atomic_read_volatile()

atomic_read_volatile() would imply an object sized memory barrier before 
and after?

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Tue, 14 Aug 2007, Christoph Lameter wrote:

> On Thu, 9 Aug 2007, Chris Snook wrote:
> 
> > This patchset makes the behavior of atomic_read uniform by removing the
> > volatile keyword from all atomic_t and atomic64_t definitions that currently
> > have it, and instead explicitly casts the variable as volatile in
> > atomic_read().  This leaves little room for creative optimization by the
> > compiler, and is in keeping with the principles behind "volatile considered
> > harmful".
> 
> volatile is generally harmful even in atomic_read(). Barriers control
> visibility and AFAICT things are fine.

Frankly, I don't see the need for this series myself either. Personal
opinion (others may differ), but I consider "volatile" to be a sad /
unfortunate wart in C (numerous threads on this list and on the gcc
lists/bugzilla over the years stand testimony to this) and if we _can_
steer clear of it, then why not -- why use this ill-defined primitive
whose implementation has often differed over compiler versions and
platforms? Granted, barrier() _is_ heavy-handed in that it makes the
optimizer forget _everything_, but then somebody did post a forget()
macro on this thread itself ...

[ BTW, why do we want the compiler to not optimize atomic_read()'s in
  the first place? Atomic ops guarantee atomicity, which has nothing
  to do with "volatility" -- users that expect "volatility" from
  atomic ops are the ones who must be fixed instead, IMHO. ]

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Satyam Sharma wrote:
> 
> On Tue, 14 Aug 2007, Christoph Lameter wrote:
> 
>> On Thu, 9 Aug 2007, Chris Snook wrote:
>>
>>> This patchset makes the behavior of atomic_read uniform by removing the
>>> volatile keyword from all atomic_t and atomic64_t definitions that currently
>>> have it, and instead explicitly casts the variable as volatile in
>>> atomic_read().  This leaves little room for creative optimization by the
>>> compiler, and is in keeping with the principles behind "volatile considered
>>> harmful".
>> volatile is generally harmful even in atomic_read(). Barriers control
>> visibility and AFAICT things are fine.
> 
> Frankly, I don't see the need for this series myself either. Personal
> opinion (others may differ), but I consider "volatile" to be a sad /
> unfortunate wart in C (numerous threads on this list and on the gcc
> lists/bugzilla over the years stand testimony to this) and if we _can_
> steer clear of it, then why not -- why use this ill-defined primitive
> whose implementation has often differed over compiler versions and
> platforms? Granted, barrier() _is_ heavy-handed in that it makes the
> optimizer forget _everything_, but then somebody did post a forget()
> macro on this thread itself ...
> 
> [ BTW, why do we want the compiler to not optimize atomic_read()'s in
>   the first place? Atomic ops guarantee atomicity, which has nothing
>   to do with "volatility" -- users that expect "volatility" from
>   atomic ops are the ones who must be fixed instead, IMHO. ]

Because atomic operations are generally used for synchronization, which requires 
volatile behavior.  Most such codepaths currently use an inefficient barrier(). 
  Some forget to and we get bugs, because people assume that atomic_read() 
actually reads something, and atomic_write() actually writes something.  Worse, 
these are architecture-specific, even compiler version-specific bugs that are 
often difficult to track down.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Tue, 14 Aug 2007, Chris Snook wrote:

> Because atomic operations are generally used for synchronization, which
> requires volatile behavior.  Most such codepaths currently use an inefficient
> barrier().  Some forget to and we get bugs, because people assume that
> atomic_read() actually reads something, and atomic_write() actually writes
> something.  Worse, these are architecture-specific, even compiler
> version-specific bugs that are often difficult to track down.

Looks like we need to have lock and unlock semantics?

atomic_read()

which has no barrier or volatile implications.

atomic_read_for_lock

	Acquire semantics?


atomic_read_for_unlock

	Release semantics?

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

Chris Snook <csnook@redhat.com> wrote:
> 
> Because atomic operations are generally used for synchronization, which requires 
> volatile behavior.  Most such codepaths currently use an inefficient barrier(). 
>  Some forget to and we get bugs, because people assume that atomic_read() 
> actually reads something, and atomic_write() actually writes something.  Worse, 
> these are architecture-specific, even compiler version-specific bugs that are 
> often difficult to track down.

I'm yet to see a single example from the current tree where
this patch series is the correct solution.  So far the only
example has been a buggy piece of code which has since been
fixed with a cpu_relax.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Heiko Carstens]

On Wed, Aug 15, 2007 at 02:49:03PM +0800, Herbert Xu wrote:
> Chris Snook <csnook@redhat.com> wrote:
> > 
> > Because atomic operations are generally used for synchronization, which requires 
> > volatile behavior.  Most such codepaths currently use an inefficient barrier(). 
> >  Some forget to and we get bugs, because people assume that atomic_read() 
> > actually reads something, and atomic_write() actually writes something.  Worse, 
> > these are architecture-specific, even compiler version-specific bugs that are 
> > often difficult to track down.
> 
> I'm yet to see a single example from the current tree where
> this patch series is the correct solution.  So far the only
> example has been a buggy piece of code which has since been
> fixed with a cpu_relax.

Btw.: we still have

include/asm-i386/mach-es7000/mach_wakecpu.h:  while (!atomic_read(deassert));
include/asm-i386/mach-default/mach_wakecpu.h: while (!atomic_read(deassert));

Looks like they need to be fixed as well.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

On 8/15/2007 10:18 AM, Heiko Carstens wrote:
> On Wed, Aug 15, 2007 at 02:49:03PM +0800, Herbert Xu wrote:
>> Chris Snook <csnook@redhat.com> wrote:
>> > 
>> > Because atomic operations are generally used for synchronization, which requires 
>> > volatile behavior.  Most such codepaths currently use an inefficient barrier(). 
>> >  Some forget to and we get bugs, because people assume that atomic_read() 
>> > actually reads something, and atomic_write() actually writes something.  Worse, 
>> > these are architecture-specific, even compiler version-specific bugs that are 
>> > often difficult to track down.
>> 
>> I'm yet to see a single example from the current tree where
>> this patch series is the correct solution.  So far the only
>> example has been a buggy piece of code which has since been
>> fixed with a cpu_relax.
> 
> Btw.: we still have
> 
> include/asm-i386/mach-es7000/mach_wakecpu.h:  while (!atomic_read(deassert));
> include/asm-i386/mach-default/mach_wakecpu.h: while (!atomic_read(deassert));
> 
> Looks like they need to be fixed as well.


I don't know if this here is affected:

/* drivers/ieee1394/ieee1394_core.h */
static inline unsigned int get_hpsb_generation(struct hpsb_host *host)
{
	return atomic_read(&host->generation);
}

/* drivers/ieee1394/nodemgr.c */
static int nodemgr_host_thread(void *__hi)
{
	[...]

	for (;;) {
		[... sleep until bus reset event ...]

		/* Pause for 1/4 second in 1/16 second intervals,
		 * to make sure things settle down. */
		g = get_hpsb_generation(host);
		for (i = 0; i < 4 ; i++) {
			if (msleep_interruptible(63) ||
			    kthread_should_stop())
				goto exit;

	/* Now get the generation in which the node ID's we collect
	 * are valid.  During the bus scan we will use this generation
	 * for the read transactions, so that if another reset occurs
	 * during the scan the transactions will fail instead of
	 * returning bogus data. */

			generation = get_hpsb_generation(host);

	/* If we get a reset before we are done waiting, then
	 * start the waiting over again */

			if (generation != g)
				g = generation, i = 0;
		}

		[... scan bus, using generation ...]

	}
exit:
[...]
}



-- 
Stefan Richter
-=====-=-=== =--- -====
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

Hi Stefan,


On Wed, 15 Aug 2007, Stefan Richter wrote:

> On 8/15/2007 10:18 AM, Heiko Carstens wrote:
> > On Wed, Aug 15, 2007 at 02:49:03PM +0800, Herbert Xu wrote:
> >> Chris Snook <csnook@redhat.com> wrote:
> >> > 
> >> > Because atomic operations are generally used for synchronization, which requires 
> >> > volatile behavior.  Most such codepaths currently use an inefficient barrier(). 
> >> >  Some forget to and we get bugs, because people assume that atomic_read() 
> >> > actually reads something, and atomic_write() actually writes something.  Worse, 
> >> > these are architecture-specific, even compiler version-specific bugs that are 
> >> > often difficult to track down.
> >> 
> >> I'm yet to see a single example from the current tree where
> >> this patch series is the correct solution.  So far the only
> >> example has been a buggy piece of code which has since been
> >> fixed with a cpu_relax.
> > 
> > Btw.: we still have
> > 
> > include/asm-i386/mach-es7000/mach_wakecpu.h:  while (!atomic_read(deassert));
> > include/asm-i386/mach-default/mach_wakecpu.h: while (!atomic_read(deassert));
> > 
> > Looks like they need to be fixed as well.
> 
> 
> I don't know if this here is affected:

Yes, I think it is. You're clearly expecting the read to actually happen
when you call get_hpsb_generation(). It's clearly not a busy-loop, so
cpu_relax() sounds pointless / wrong solution for this case, so I'm now
somewhat beginning to appreciate the motivation behind this series :-)

But as I said, there are ways to achieve the same goals of this series
without using "volatile".

I think I'll submit a RFC/patch or two on this myself (will also fix
the code pieces listed here).


> /* drivers/ieee1394/ieee1394_core.h */
> static inline unsigned int get_hpsb_generation(struct hpsb_host *host)
> {
> 	return atomic_read(&host->generation);
> }
> 
> /* drivers/ieee1394/nodemgr.c */
> static int nodemgr_host_thread(void *__hi)
> {
> 	[...]
> 
> 	for (;;) {
> 		[... sleep until bus reset event ...]
> 
> 		/* Pause for 1/4 second in 1/16 second intervals,
> 		 * to make sure things settle down. */
> 		g = get_hpsb_generation(host);
> 		for (i = 0; i < 4 ; i++) {
> 			if (msleep_interruptible(63) ||
> 			    kthread_should_stop())
> 				goto exit;

Totally unrelated, but this looks weird. IMHO you actually wanted:

	msleep_interruptible(63);
	if (kthread_should_stop())
		goto exit;

here, didn't you? Otherwise the thread will exit even when
kthread_should_stop() != TRUE (just because it received a signal),
and it is not good for a kthread to exit on its own if it uses
kthread_should_stop() or if some other piece of kernel code could
eventually call kthread_stop(tsk) on it.

Ok, probably the thread will never receive a signal in the first
place because it's spawned off kthreadd which ignores all signals
beforehand, but still ...

[PATCH] ieee1394: Fix kthread stopping in nodemgr_host_thread

The nodemgr host thread can exit on its own even when kthread_should_stop
is not true, on receiving a signal (might never happen in practice, as
it ignores signals). But considering kthread_stop() must not be mixed with
kthreads that can exit on their own, I think changing the code like this
is clearer. This change means the thread can cut its sleep short when
receive a signal but looking at the code around, that sounds okay (and
again, it might never actually recieve a signal in practice).

Signed-off-by: Satyam Sharma <satyam@infradead.org>

---

 drivers/ieee1394/nodemgr.c |    3 ++-
 1 files changed, 2 insertions(+), 1 deletions(-)

diff --git a/drivers/ieee1394/nodemgr.c b/drivers/ieee1394/nodemgr.c
index 2ffd534..981a7da 100644
--- a/drivers/ieee1394/nodemgr.c
+++ b/drivers/ieee1394/nodemgr.c
@@ -1721,7 +1721,8 @@ static int nodemgr_host_thread(void *__hi)
 		 * to make sure things settle down. */
 		g = get_hpsb_generation(host);
 		for (i = 0; i < 4 ; i++) {
-			if (msleep_interruptible(63) || kthread_should_stop())
+			msleep_interruptible(63);
+			if (kthread_should_stop())
 				goto exit;
 
 			/* Now get the generation in which the node ID's we collect

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 08:05:38PM +0530, Satyam Sharma wrote:
>
> > I don't know if this here is affected:
> 
> Yes, I think it is. You're clearly expecting the read to actually happen
> when you call get_hpsb_generation(). It's clearly not a busy-loop, so
> cpu_relax() sounds pointless / wrong solution for this case, so I'm now
> somewhat beginning to appreciate the motivation behind this series :-)

Nope, we're calling schedule which is a rather heavy-weight
barrier.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 08:05:38PM +0530, Satyam Sharma wrote:
>>> I don't know if this here is affected:

[...something like]
	b = atomic_read(a);
	for (i = 0; i < 4; i++) {
		msleep_interruptible(63);
		c = atomic_read(a);
		if (c != b) {
			b = c;
			i = 0;
		}
	}

> Nope, we're calling schedule which is a rather heavy-weight
> barrier.

How does the compiler know that msleep() has got barrier()s?
-- 
Stefan Richter
-=====-=-=== =--- -====
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 06:09:35PM +0200, Stefan Richter wrote:
> Herbert Xu wrote:
> > On Wed, Aug 15, 2007 at 08:05:38PM +0530, Satyam Sharma wrote:
> >>> I don't know if this here is affected:
> 
> [...something like]
> 	b = atomic_read(a);
> 	for (i = 0; i < 4; i++) {
> 		msleep_interruptible(63);
> 		c = atomic_read(a);
> 		if (c != b) {
> 			b = c;
> 			i = 0;
> 		}
> 	}
> 
> > Nope, we're calling schedule which is a rather heavy-weight
> > barrier.
> 
> How does the compiler know that msleep() has got barrier()s?

Because msleep_interruptible() is in a separate compilation unit,
the compiler has to assume that it might modify any arbitrary global.
In many cases, the compiler also has to assume that msleep_interruptible()
might call back into a function in the current compilation unit, thus
possibly modifying global static variables.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> On Wed, Aug 15, 2007 at 06:09:35PM +0200, Stefan Richter wrote:
> > Herbert Xu wrote:
> > > On Wed, Aug 15, 2007 at 08:05:38PM +0530, Satyam Sharma wrote:
> > >>> I don't know if this here is affected:
> > 
> > [...something like]
> > 	b = atomic_read(a);
> > 	for (i = 0; i < 4; i++) {
> > 		msleep_interruptible(63);
> > 		c = atomic_read(a);
> > 		if (c != b) {
> > 			b = c;
> > 			i = 0;
> > 		}
> > 	}
> > 
> > > Nope, we're calling schedule which is a rather heavy-weight
> > > barrier.
> > 
> > How does the compiler know that msleep() has got barrier()s?
> 
> Because msleep_interruptible() is in a separate compilation unit,
> the compiler has to assume that it might modify any arbitrary global.
> In many cases, the compiler also has to assume that msleep_interruptible()
> might call back into a function in the current compilation unit, thus
> possibly modifying global static variables.

Yup, I've just verified this with a testcase. So a call to any function
outside of the current compilation unit acts as a compiler barrier. Cool.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> How does the compiler know that msleep() has got barrier()s?
>
> Because msleep_interruptible() is in a separate compilation unit,
> the compiler has to assume that it might modify any arbitrary global.

No; compilation units have nothing to do with it, GCC can optimise
across compilation unit boundaries just fine, if you tell it to
compile more than one compilation unit at once.

What you probably mean is that the compiler has to assume any code
it cannot currently see can do anything (insofar as allowed by the
relevant standards etc.)

> In many cases, the compiler also has to assume that 
> msleep_interruptible()
> might call back into a function in the current compilation unit, thus
> possibly modifying global static variables.

It most often is smart enough to see what compilation-unit-local
variables might be modified that way, though :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 08:31:25PM +0200, Segher Boessenkool wrote:
> >>How does the compiler know that msleep() has got barrier()s?
> >
> >Because msleep_interruptible() is in a separate compilation unit,
> >the compiler has to assume that it might modify any arbitrary global.
> 
> No; compilation units have nothing to do with it, GCC can optimise
> across compilation unit boundaries just fine, if you tell it to
> compile more than one compilation unit at once.

Last I checked, the Linux kernel build system did compile each .c file
as a separate compilation unit.

> What you probably mean is that the compiler has to assume any code
> it cannot currently see can do anything (insofar as allowed by the
> relevant standards etc.)

Indeed.

> >In many cases, the compiler also has to assume that 
> >msleep_interruptible()
> >might call back into a function in the current compilation unit, thus
> >possibly modifying global static variables.
> 
> It most often is smart enough to see what compilation-unit-local
> variables might be modified that way, though :-)

Yep.  For example, if it knows the current value of a given such local
variable, and if all code paths that would change some other variable
cannot be reached given that current value of the first variable.
At least given that gcc doesn't know about multiple threads of execution!

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

[ The Cc: list scares me. Should probably trim it. ]


On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> On Wed, Aug 15, 2007 at 08:31:25PM +0200, Segher Boessenkool wrote:
> > >>How does the compiler know that msleep() has got barrier()s?
> > >
> > >Because msleep_interruptible() is in a separate compilation unit,
> > >the compiler has to assume that it might modify any arbitrary global.
> > 
> > No; compilation units have nothing to do with it, GCC can optimise
> > across compilation unit boundaries just fine, if you tell it to
> > compile more than one compilation unit at once.
> 
> Last I checked, the Linux kernel build system did compile each .c file
> as a separate compilation unit.
> 
> > What you probably mean is that the compiler has to assume any code
> > it cannot currently see can do anything (insofar as allowed by the
> > relevant standards etc.)

I think this was just terminology confusion here again. Isn't "any code
that it cannot currently see" the same as "another compilation unit",
and wouldn't the "compilation unit" itself expand if we ask gcc to
compile more than one unit at once? Or is there some more specific
"definition" for "compilation unit" (in gcc lingo, possibly?)

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 01:24:42AM +0530, Satyam Sharma wrote:
> [ The Cc: list scares me. Should probably trim it. ]

Trim away!  ;-)

> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> 
> > On Wed, Aug 15, 2007 at 08:31:25PM +0200, Segher Boessenkool wrote:
> > > >>How does the compiler know that msleep() has got barrier()s?
> > > >
> > > >Because msleep_interruptible() is in a separate compilation unit,
> > > >the compiler has to assume that it might modify any arbitrary global.
> > > 
> > > No; compilation units have nothing to do with it, GCC can optimise
> > > across compilation unit boundaries just fine, if you tell it to
> > > compile more than one compilation unit at once.
> > 
> > Last I checked, the Linux kernel build system did compile each .c file
> > as a separate compilation unit.
> > 
> > > What you probably mean is that the compiler has to assume any code
> > > it cannot currently see can do anything (insofar as allowed by the
> > > relevant standards etc.)
> 
> I think this was just terminology confusion here again. Isn't "any code
> that it cannot currently see" the same as "another compilation unit",
> and wouldn't the "compilation unit" itself expand if we ask gcc to
> compile more than one unit at once? Or is there some more specific
> "definition" for "compilation unit" (in gcc lingo, possibly?)

This is indeed my understanding -- "compilation unit" is whatever the
compiler looks at in one go.  I have heard the word "module" used for
the minimal compilation unit covering a single .c file and everything
that it #includes, but there might be a better name for this.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> I think this was just terminology confusion here again. Isn't "any 
>> code
>> that it cannot currently see" the same as "another compilation unit",
>> and wouldn't the "compilation unit" itself expand if we ask gcc to
>> compile more than one unit at once? Or is there some more specific
>> "definition" for "compilation unit" (in gcc lingo, possibly?)
>
> This is indeed my understanding -- "compilation unit" is whatever the
> compiler looks at in one go.  I have heard the word "module" used for
> the minimal compilation unit covering a single .c file and everything
> that it #includes, but there might be a better name for this.

Yes, that's what's called "compilation unit" :-)

[/me double checks]

Erm, the C standard actually calls it "translation unit".

To be exact, to avoid any more confusion:

5.1.1.1/1:
A C program need not all be translated at the same time. The
text of the program is kept in units called source files, (or
preprocessing files) in this International Standard. A source
file together with all the headers and source files included
via the preprocessing directive #include is known as a
preprocessing translation unit. After preprocessing, a
preprocessing translation unit is called a translation unit.



Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 10:52:53PM +0200, Segher Boessenkool wrote:
> >>I think this was just terminology confusion here again. Isn't "any 
> >>code
> >>that it cannot currently see" the same as "another compilation unit",
> >>and wouldn't the "compilation unit" itself expand if we ask gcc to
> >>compile more than one unit at once? Or is there some more specific
> >>"definition" for "compilation unit" (in gcc lingo, possibly?)
> >
> >This is indeed my understanding -- "compilation unit" is whatever the
> >compiler looks at in one go.  I have heard the word "module" used for
> >the minimal compilation unit covering a single .c file and everything
> >that it #includes, but there might be a better name for this.
> 
> Yes, that's what's called "compilation unit" :-)
> 
> [/me double checks]
> 
> Erm, the C standard actually calls it "translation unit".
> 
> To be exact, to avoid any more confusion:
> 
> 5.1.1.1/1:
> A C program need not all be translated at the same time. The
> text of the program is kept in units called source files, (or
> preprocessing files) in this International Standard. A source
> file together with all the headers and source files included
> via the preprocessing directive #include is known as a
> preprocessing translation unit. After preprocessing, a
> preprocessing translation unit is called a translation unit.

I am OK with "translation" and "compilation" being near-synonyms.  ;-)

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>> What you probably mean is that the compiler has to assume any code
>>> it cannot currently see can do anything (insofar as allowed by the
>>> relevant standards etc.)
>
> I think this was just terminology confusion here again. Isn't "any code
> that it cannot currently see" the same as "another compilation unit",

It is not; try  gcc -combine  or the upcoming link-time optimisation
stuff, for example.

> and wouldn't the "compilation unit" itself expand if we ask gcc to
> compile more than one unit at once? Or is there some more specific
> "definition" for "compilation unit" (in gcc lingo, possibly?)

"compilation unit" is a C standard term.  It typically boils down
to "single .c file".


Segher

(no subject)

[Satyam Sharma]

On Wed, 15 Aug 2007, Segher Boessenkool wrote:

> > > > What you probably mean is that the compiler has to assume any code
> > > > it cannot currently see can do anything (insofar as allowed by the
> > > > relevant standards etc.)
> > 
> > I think this was just terminology confusion here again. Isn't "any code
> > that it cannot currently see" the same as "another compilation unit",
> 
> It is not; try  gcc -combine  or the upcoming link-time optimisation
> stuff, for example.
> 
> > and wouldn't the "compilation unit" itself expand if we ask gcc to
> > compile more than one unit at once? Or is there some more specific
> > "definition" for "compilation unit" (in gcc lingo, possibly?)
> 
> "compilation unit" is a C standard term.  It typically boils down
> to "single .c file".

As you mentioned later, "single .c file with all the other files (headers
or other .c files) that it pulls in via #include" is actually "translation
unit", both in the C standard as well as gcc docs. "Compilation unit"
doesn't seem to be nearly as standard a term, though in most places it
is indeed meant to be same as "translation unit", but with the new gcc
inter-module-analysis stuff that you referred to above, I suspect one may
reasonably want to call a "compilation unit" as all that the compiler sees
at a given instant.

BTW I did some auditing (only inside include/asm-{i386,x86_64}/ and
arch/{i386,x86_64}/) and found a couple more callsites that don't use
cpu_relax():

arch/i386/kernel/crash.c:101
arch/x86_64/kernel/crash.c:97

that are:

	while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
		mdelay(1);
		msecs--;
	}

where mdelay() becomes __const_udelay() which happens to be in another
translation unit (arch/i386/lib/delay.c) and hence saves this callsite
from being a bug :-)

Curiously, __const_udelay() is still marked as "inline" where it is
implemented in lib/delay.c which is weird, considering it won't ever
be inlined, would it? With the kernel presently being compiled one
translation unit at a time, I don't see how the implementation would
be visible to any callsite out there to be able to inline it.


Satyam

Re: your mail

[Herbert Xu]

On Thu, Aug 16, 2007 at 06:06:00AM +0530, Satyam Sharma wrote:
> 
> that are:
> 
> 	while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
> 		mdelay(1);
> 		msecs--;
> 	}
> 
> where mdelay() becomes __const_udelay() which happens to be in another
> translation unit (arch/i386/lib/delay.c) and hence saves this callsite
> from being a bug :-)

The udelay itself certainly should have some form of cpu_relax in it.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

[ Sorry for empty subject line in previous mail. I intended to make
  a patch so cleared it to change it, but ultimately neither made
  a patch nor restored subject line. Done that now. ]


On Thu, 16 Aug 2007, Herbert Xu wrote:

> On Thu, Aug 16, 2007 at 06:06:00AM +0530, Satyam Sharma wrote:
> > 
> > that are:
> > 
> > 	while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
> > 		mdelay(1);
> > 		msecs--;
> > 	}
> > 
> > where mdelay() becomes __const_udelay() which happens to be in another
> > translation unit (arch/i386/lib/delay.c) and hence saves this callsite
> > from being a bug :-)
> 
> The udelay itself certainly should have some form of cpu_relax in it.

Yes, a form of barrier() must be present in mdelay() or udelay() itself
as you say, having it in __const_udelay() is *not* enough (superflous
actually, considering it is already a separate translation unit and
invisible to the compiler).

However, there are no compiler barriers on the macro-definition-path
between mdelay(1) and __const_udelay(), so the only thing that saves us
from being a bug here is indeed the different-translation-unit concept.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 06:28:42AM +0530, Satyam Sharma wrote:
>
> > The udelay itself certainly should have some form of cpu_relax in it.
> 
> Yes, a form of barrier() must be present in mdelay() or udelay() itself
> as you say, having it in __const_udelay() is *not* enough (superflous
> actually, considering it is already a separate translation unit and
> invisible to the compiler).

As long as __const_udelay does something which has the same
effect as barrier it is enough even if it's in the same unit.
As a matter of fact it does on i386 where __delay either uses
rep_nop or asm/volatile.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

Hi Herbert,


On Thu, 16 Aug 2007, Herbert Xu wrote:

> On Thu, Aug 16, 2007 at 06:28:42AM +0530, Satyam Sharma wrote:
> >
> > > The udelay itself certainly should have some form of cpu_relax in it.
> > 
> > Yes, a form of barrier() must be present in mdelay() or udelay() itself
> > as you say, having it in __const_udelay() is *not* enough (superflous
> > actually, considering it is already a separate translation unit and
> > invisible to the compiler).
> 
> As long as __const_udelay does something which has the same
> effect as barrier it is enough even if it's in the same unit.

Only if __const_udelay() is inlined. But as I said, __const_udelay()
-- although marked "inline" -- will never be inlined anywhere in the
kernel in reality. It's an exported symbol, and never inlined from
modules. Even from built-in targets, the definition of __const_udelay
is invisible when gcc is compiling the compilation units of those
callsites. The compiler has no idea that that function has barriers
or not, so we're saved here _only_ by the lucky fact that
__const_udelay() is in a different compilation unit.


> As a matter of fact it does on i386 where __delay either uses
> rep_nop or asm/volatile.

__delay() can be either delay_tsc() or delay_loop() on i386.

delay_tsc() uses the rep_nop() there for it's own little busy
loop, actually. But for a call site that inlines __const_udelay()
-- if it were ever moved to a .h file and marked inline -- the
call to __delay() will _still_ be across compilation units. So,
again for this case, it does not matter if the callee function
has compiler barriers or not (it would've been a different story
if we were discussing real/CPU barriers, I think), what saves us
here is just the fact that a call is made to a function from a
different compilation unit, which is invisible to the compiler
when compiling the callsite, and hence acting as the compiler
barrier.

Regarding delay_loop(), it uses "volatile" for the "asm" which
has quite different semantics from the C language "volatile"
type-qualifier keyword and does not imply any compiler barrier
at all.


Satyam

Re:

[Segher Boessenkool]

>> "compilation unit" is a C standard term.  It typically boils down
>> to "single .c file".
>
> As you mentioned later, "single .c file with all the other files 
> (headers
> or other .c files) that it pulls in via #include" is actually 
> "translation
> unit", both in the C standard as well as gcc docs.

Yeah.  "single .c file after preprocessing".  Same thing :-)

> "Compilation unit"
> doesn't seem to be nearly as standard a term, though in most places it
> is indeed meant to be same as "translation unit", but with the new gcc
> inter-module-analysis stuff that you referred to above, I suspect one 
> may
> reasonably want to call a "compilation unit" as all that the compiler 
> sees
> at a given instant.

That would be a bit confusing, would it not?  They'd better find
some better name for that if they want to name it at all (remember,
none of these optimisations should have any effect on the semantics
of the program, you just get fewer .o files etc.).


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> No; compilation units have nothing to do with it, GCC can optimise
>> across compilation unit boundaries just fine, if you tell it to
>> compile more than one compilation unit at once.
>
> Last I checked, the Linux kernel build system did compile each .c file
> as a separate compilation unit.

I have some patches to use -combine -fwhole-program for Linux.
Highly experimental, you need a patched bleeding edge toolchain.
If there's interest I'll clean it up and put it online.

David Woodhouse had some similar patches about a year ago.

>>> In many cases, the compiler also has to assume that
>>> msleep_interruptible()
>>> might call back into a function in the current compilation unit, thus
>>> possibly modifying global static variables.
>>
>> It most often is smart enough to see what compilation-unit-local
>> variables might be modified that way, though :-)
>
> Yep.  For example, if it knows the current value of a given such local
> variable, and if all code paths that would change some other variable
> cannot be reached given that current value of the first variable.

Or the most common thing: if neither the address of the translation-
unit local variable nor the address of any function writing to that
variable can "escape" from that translation unit, nothing outside
the translation unit can write to the variable.

> At least given that gcc doesn't know about multiple threads of 
> execution!

Heh, only about the threads it creates itself (not relevant to
the kernel, for sure :-) )


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 11:05:35PM +0200, Segher Boessenkool wrote:
> >>No; compilation units have nothing to do with it, GCC can optimise
> >>across compilation unit boundaries just fine, if you tell it to
> >>compile more than one compilation unit at once.
> >
> >Last I checked, the Linux kernel build system did compile each .c file
> >as a separate compilation unit.
> 
> I have some patches to use -combine -fwhole-program for Linux.
> Highly experimental, you need a patched bleeding edge toolchain.
> If there's interest I'll clean it up and put it online.
> 
> David Woodhouse had some similar patches about a year ago.

Sounds exciting...  ;-)

> >>>In many cases, the compiler also has to assume that
> >>>msleep_interruptible()
> >>>might call back into a function in the current compilation unit, thus
> >>>possibly modifying global static variables.
> >>
> >>It most often is smart enough to see what compilation-unit-local
> >>variables might be modified that way, though :-)
> >
> >Yep.  For example, if it knows the current value of a given such local
> >variable, and if all code paths that would change some other variable
> >cannot be reached given that current value of the first variable.
> 
> Or the most common thing: if neither the address of the translation-
> unit local variable nor the address of any function writing to that
> variable can "escape" from that translation unit, nothing outside
> the translation unit can write to the variable.

But there is usually at least one externally callable function in
a .c file.

> >At least given that gcc doesn't know about multiple threads of 
> >execution!
> 
> Heh, only about the threads it creates itself (not relevant to
> the kernel, for sure :-) )

;-)

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>>> No; compilation units have nothing to do with it, GCC can optimise
>>>> across compilation unit boundaries just fine, if you tell it to
>>>> compile more than one compilation unit at once.
>>>
>>> Last I checked, the Linux kernel build system did compile each .c 
>>> file
>>> as a separate compilation unit.
>>
>> I have some patches to use -combine -fwhole-program for Linux.
>> Highly experimental, you need a patched bleeding edge toolchain.
>> If there's interest I'll clean it up and put it online.
>>
>> David Woodhouse had some similar patches about a year ago.
>
> Sounds exciting...  ;-)

Yeah, the breakage is *quite* spectacular :-)

>>>>> In many cases, the compiler also has to assume that
>>>>> msleep_interruptible()
>>>>> might call back into a function in the current compilation unit, 
>>>>> thus
>>>>> possibly modifying global static variables.
>>>>
>>>> It most often is smart enough to see what compilation-unit-local
>>>> variables might be modified that way, though :-)
>>>
>>> Yep.  For example, if it knows the current value of a given such 
>>> local
>>> variable, and if all code paths that would change some other variable
>>> cannot be reached given that current value of the first variable.
>>
>> Or the most common thing: if neither the address of the translation-
>> unit local variable nor the address of any function writing to that
>> variable can "escape" from that translation unit, nothing outside
>> the translation unit can write to the variable.
>
> But there is usually at least one externally callable function in
> a .c file.

Of course, but often none of those will (indirectly) write a certain
static variable.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 03:23:28AM +0200, Segher Boessenkool wrote:
> >>>>No; compilation units have nothing to do with it, GCC can optimise
> >>>>across compilation unit boundaries just fine, if you tell it to
> >>>>compile more than one compilation unit at once.
> >>>
> >>>Last I checked, the Linux kernel build system did compile each .c 
> >>>file
> >>>as a separate compilation unit.
> >>
> >>I have some patches to use -combine -fwhole-program for Linux.
> >>Highly experimental, you need a patched bleeding edge toolchain.
> >>If there's interest I'll clean it up and put it online.
> >>
> >>David Woodhouse had some similar patches about a year ago.
> >
> >Sounds exciting...  ;-)
> 
> Yeah, the breakage is *quite* spectacular :-)

I bet!!!  ;-)

> >>>>>In many cases, the compiler also has to assume that
> >>>>>msleep_interruptible()
> >>>>>might call back into a function in the current compilation unit, 
> >>>>>thus
> >>>>>possibly modifying global static variables.
> >>>>
> >>>>It most often is smart enough to see what compilation-unit-local
> >>>>variables might be modified that way, though :-)
> >>>
> >>>Yep.  For example, if it knows the current value of a given such 
> >>>local
> >>>variable, and if all code paths that would change some other variable
> >>>cannot be reached given that current value of the first variable.
> >>
> >>Or the most common thing: if neither the address of the translation-
> >>unit local variable nor the address of any function writing to that
> >>variable can "escape" from that translation unit, nothing outside
> >>the translation unit can write to the variable.
> >
> >But there is usually at least one externally callable function in
> >a .c file.
> 
> Of course, but often none of those will (indirectly) write a certain
> static variable.

But there has to be some path to the static functions, assuming that
they are not dead code.  Yes, there can be cases where the compiler
knows enough about the state of the variables to rule out some of code
paths to them, but I have no idea how often this happens in kernel
code.

						Thanx, Paul

[PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Satyam Sharma]

On Wed, 15 Aug 2007, Heiko Carstens wrote:

> [...]
> Btw.: we still have
> 
> include/asm-i386/mach-es7000/mach_wakecpu.h:  while (!atomic_read(deassert));
> include/asm-i386/mach-default/mach_wakecpu.h: while (!atomic_read(deassert));
> 
> Looks like they need to be fixed as well.


[PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

Use cpu_relax() in the busy loops, as atomic_read() doesn't automatically
imply volatility for i386 and x86_64. x86_64 doesn't have this issue because
it open-codes the while loop in smpboot.c:smp_callin() itself that already
uses cpu_relax().

For i386, however, smpboot.c:smp_callin() calls wait_for_init_deassert()
which is buggy for mach-default and mach-es7000 cases.

[ I test-built a kernel -- smp_callin() itself got inlined in its only
  callsite, smpboot.c:start_secondary() -- and the relevant piece of
  code disassembles to the following:

0xc1019704 <start_secondary+12>:        mov    0xc144c4c8,%eax
0xc1019709 <start_secondary+17>:        test   %eax,%eax
0xc101970b <start_secondary+19>:        je     0xc1019709 <start_secondary+17>

  init_deasserted (at 0xc144c4c8) gets fetched into %eax only once and
  then we loop over the test of the stale value in the register only,
  so these look like real bugs to me. With the fix below, this becomes:

0xc1019706 <start_secondary+14>:        pause
0xc1019708 <start_secondary+16>:        cmpl   $0x0,0xc144c4c8
0xc101970f <start_secondary+23>:        je     0xc1019706 <start_secondary+14>

  which looks nice and healthy. ]

Thanks to Heiko Carstens for noticing this.

Signed-off-by: Satyam Sharma <satyam@infradead.org>

---

 include/asm-i386/mach-default/mach_wakecpu.h |    3 ++-
 include/asm-i386/mach-es7000/mach_wakecpu.h  |    3 ++-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/include/asm-i386/mach-default/mach_wakecpu.h b/include/asm-i386/mach-default/mach_wakecpu.h
index 673b85c..3ebb178 100644
--- a/include/asm-i386/mach-default/mach_wakecpu.h
+++ b/include/asm-i386/mach-default/mach_wakecpu.h
@@ -15,7 +15,8 @@
 
 static inline void wait_for_init_deassert(atomic_t *deassert)
 {
-	while (!atomic_read(deassert));
+	while (!atomic_read(deassert))
+		cpu_relax();
 	return;
 }
 
diff --git a/include/asm-i386/mach-es7000/mach_wakecpu.h b/include/asm-i386/mach-es7000/mach_wakecpu.h
index efc903b..84ff583 100644
--- a/include/asm-i386/mach-es7000/mach_wakecpu.h
+++ b/include/asm-i386/mach-es7000/mach_wakecpu.h
@@ -31,7 +31,8 @@ wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
 static inline void wait_for_init_deassert(atomic_t *deassert)
 {
 #ifdef WAKE_SECONDARY_VIA_INIT
-	while (!atomic_read(deassert));
+	while (!atomic_read(deassert))
+		cpu_relax();
 #endif
 	return;
 }

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Denys Vlasenko]

On Thursday 16 August 2007 01:39, Satyam Sharma wrote:
>
>  static inline void wait_for_init_deassert(atomic_t *deassert)
>  {
> -	while (!atomic_read(deassert));
> +	while (!atomic_read(deassert))
> +		cpu_relax();
>  	return;
>  }

For less-than-briliant people like me, it's totally non-obvious that
cpu_relax() is needed for correctness here, not just to make P4 happy.

IOW: "atomic_read" name quite unambiguously means "I will read
this variable from main memory". Which is not true and creates
potential for confusion and bugs.
--
vda

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Andi Kleen]

On Friday 24 August 2007 13:59:32 Denys Vlasenko wrote:
> On Thursday 16 August 2007 01:39, Satyam Sharma wrote:
> >
> >  static inline void wait_for_init_deassert(atomic_t *deassert)
> >  {
> > -	while (!atomic_read(deassert));
> > +	while (!atomic_read(deassert))
> > +		cpu_relax();
> >  	return;
> >  }
> 
> For less-than-briliant people like me, it's totally non-obvious that
> cpu_relax() is needed for correctness here, not just to make P4 happy.

I find it also non obvious. It would be really better to have a barrier
or equivalent (volatile or variable clobber) in the atomic_read()
 
> IOW: "atomic_read" name quite unambiguously means "I will read
> this variable from main memory". Which is not true and creates
> potential for confusion and bugs.

Agreed.

-Andi

RE: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Kenn Humborg]

> On Thursday 16 August 2007 01:39, Satyam Sharma wrote:
> >
> >  static inline void wait_for_init_deassert(atomic_t *deassert)
> >  {
> > -	while (!atomic_read(deassert));
> > +	while (!atomic_read(deassert))
> > +		cpu_relax();
> >  	return;
> >  }
> 
> For less-than-briliant people like me, it's totally non-obvious that
> cpu_relax() is needed for correctness here, not just to make P4 happy.
> 
> IOW: "atomic_read" name quite unambiguously means "I will read
> this variable from main memory". Which is not true and creates
> potential for confusion and bugs.

To me, "atomic_read" means a read which is synchronized with other 
changes to the variable (using the atomic_XXX functions) in such 
a way that I will always only see the "before" or "after"
state of the variable - never an intermediate state while a 
modification is happening.  It doesn't imply that I have to 
see the "after" state immediately after another thread modifies
it.

Perhaps the Linux atomic_XXX functions work like that, or used
to work like that, but it's counter-intuitive to me that "atomic"
should imply a memory read.

Later,
Kenn

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Denys Vlasenko]

On Friday 24 August 2007 13:12, Kenn Humborg wrote:
> > On Thursday 16 August 2007 01:39, Satyam Sharma wrote:
> > >  static inline void wait_for_init_deassert(atomic_t *deassert)
> > >  {
> > > -	while (!atomic_read(deassert));
> > > +	while (!atomic_read(deassert))
> > > +		cpu_relax();
> > >  	return;
> > >  }
> >
> > For less-than-briliant people like me, it's totally non-obvious that
> > cpu_relax() is needed for correctness here, not just to make P4 happy.
> >
> > IOW: "atomic_read" name quite unambiguously means "I will read
> > this variable from main memory". Which is not true and creates
> > potential for confusion and bugs.
>
> To me, "atomic_read" means a read which is synchronized with other
> changes to the variable (using the atomic_XXX functions) in such
> a way that I will always only see the "before" or "after"
> state of the variable - never an intermediate state while a
> modification is happening.  It doesn't imply that I have to
> see the "after" state immediately after another thread modifies
> it.

So you are ok with compiler propagating n1 to n2 here:

n1 += atomic_read(x);
other_variable++;
n2 += atomic_read(x);

without accessing x second time. What's the point? Any sane coder
will say that explicitly anyway:

tmp = atomic_read(x);
n1 += tmp;
other_variable++;
n2 += tmp;

if only for the sake of code readability. Because first code
is definitely hinting that it reads RAM twice, and it's actively *bad*
for code readability when in fact it's not the case!

Locking, compiler and CPU barriers are complicated enough already,
please don't make them even harder to understand.
--
vda

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Linus Torvalds]

On Fri, 24 Aug 2007, Denys Vlasenko wrote:
> 
> So you are ok with compiler propagating n1 to n2 here:
> 
> n1 += atomic_read(x);
> other_variable++;
> n2 += atomic_read(x);
> 
> without accessing x second time. What's the point? Any sane coder
> will say that explicitly anyway:

No.

This is a common mistake, and it's total crap.

Any "sane coder" will often use inline functions, macros, etc helpers to 
do certain abstract things. Those things may contain "atomic_read()" 
calls.

The biggest reason for compilers doing CSE is exactly the fact that many 
opportunities for CSE simple *are*not*visible* on a source code level. 

That is true of things like atomic_read() equally as to things like shared 
offsets inside structure member accesses. No difference what-so-ever.

Yes, we have, traditionally, tried to make it *easy* for the compiler to 
generate good code. So when we can, and when we look at performance for 
some really hot path, we *will* write the source code so that the compiler 
doesn't even have the option to screw it up, and that includes things like 
doing CSE at a source code level so that we don't see the compiler 
re-doing accesses unnecessarily.

And I'm not saying we shouldn't do that. But "performance" is not an 
either-or kind of situation, and we should:

 - spend the time at a source code level: make it reasonably easy for the 
   compiler to generate good code, and use the right algorithms at a 
   higher level (and order structures etc so that they have good cache 
   behaviour).

 - .. *and* expect the compiler to handle the cases we didn't do by hand
   pretty well anyway. In particular, quite often, abstraction levels at a 
   software level means that we give compilers "stupid" code, because some 
   function may have a certain high-level abstraction rule, but then on a 
   particular architecture it's actually a no-op, and the compiler should 
   get to "untangle" our stupid code and generate good end results.

 - .. *and* expect the hardware to be sane and do a good job even when the 
   compiler didn't generate perfect code or there were unlucky cache miss
   patterns etc.

and if we do all of that, we'll get good performance. But you really do 
want all three levels. It's not enough to be good at any one level (or 
even any two).

			Linus

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Satyam Sharma]

Hi Denys,


On Fri, 24 Aug 2007, Denys Vlasenko wrote:

> On Thursday 16 August 2007 01:39, Satyam Sharma wrote:
> >
> >  static inline void wait_for_init_deassert(atomic_t *deassert)
> >  {
> > -	while (!atomic_read(deassert));
> > +	while (!atomic_read(deassert))
> > +		cpu_relax();
> >  	return;
> >  }
> 
> For less-than-briliant people like me, it's totally non-obvious that
> cpu_relax() is needed for correctness here, not just to make P4 happy.

This thread has been round-and-round with exactly the same discussions
:-) I had proposed few such variants to make a compiler barrier implicit
in atomic_{read,set} myself, but frankly, at least personally speaking
(now that I know better), I'm not so much in favour of implicit barriers
(compiler, memory or both) in atomic_{read,set}.

This might sound like an about-turn if you read my own postings to Nick
Piggin from a week back, but I do agree with most his opinions on the
matter now -- separation of barriers from atomic ops is actually good,
beneficial to certain code that knows what it's doing, explicit usage
of barriers stands out more clearly (most people here who deal with it
do know cpu_relax() is an explicit compiler barrier) compared to an
implicit usage in an atomic_read() or such variant ...


> IOW: "atomic_read" name quite unambiguously means "I will read
> this variable from main memory". Which is not true and creates
> potential for confusion and bugs.

I'd have to disagree here -- atomic ops are all about _atomicity_ of
memory accesses, not _making_ them happen (or visible to other CPUs)
_then and there_ itself. The latter are the job of barriers.

The behaviour (and expectations) are quite comprehensively covered in
atomic_ops.txt -- let alone atomic_{read,set}, even atomic_{inc,dec}
are permitted by archs' implementations to _not_ have any memory
barriers, for that matter. [It is unrelated that on x86 making them
SMP-safe requires the use of the LOCK prefix that also happens to be
an implicit memory barrier.]

An argument was also made about consistency of atomic_{read,set} w.r.t.
the other atomic ops -- but clearly, they are all already consistent!
All of them are atomic :-) The fact that atomic_{read,set} do _not_
require any inline asm or LOCK prefix whereas the others do, has to do
with the fact that unlike all others, atomic_{read,set} are not RMW ops
and hence guaranteed to be atomic just as they are in plain & simple C.

But if people do seem to have a mixed / confused notion of atomicity
and barriers, and if there's consensus, then as I'd said earlier, I
have no issues in going with the consensus (eg. having API variants).
Linus would be more difficult to convince, however, I suspect :-)


Satyam

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Christoph Lameter]

On Fri, 24 Aug 2007, Satyam Sharma wrote:

> But if people do seem to have a mixed / confused notion of atomicity
> and barriers, and if there's consensus, then as I'd said earlier, I
> have no issues in going with the consensus (eg. having API variants).
> Linus would be more difficult to convince, however, I suspect :-)

The confusion may be the result of us having barrier semantics in 
atomic_read. If we take that out then we may avoid future confusions.

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Denys Vlasenko]

On Friday 24 August 2007 18:06, Christoph Lameter wrote:
> On Fri, 24 Aug 2007, Satyam Sharma wrote:
> > But if people do seem to have a mixed / confused notion of atomicity
> > and barriers, and if there's consensus, then as I'd said earlier, I
> > have no issues in going with the consensus (eg. having API variants).
> > Linus would be more difficult to convince, however, I suspect :-)
>
> The confusion may be the result of us having barrier semantics in
> atomic_read. If we take that out then we may avoid future confusions.

I think better name may help. Nuke atomic_read() altogether.

n = atomic_value(x);	// doesnt hint as strongly at reading as "atomic_read"
n = atomic_fetch(x);	// yes, we _do_ touch RAM
n = atomic_read_uncached(x); // or this

How does that sound?
--
vda

Re: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Chris Snook]

Denys Vlasenko wrote:
> On Friday 24 August 2007 18:06, Christoph Lameter wrote:
>> On Fri, 24 Aug 2007, Satyam Sharma wrote:
>>> But if people do seem to have a mixed / confused notion of atomicity
>>> and barriers, and if there's consensus, then as I'd said earlier, I
>>> have no issues in going with the consensus (eg. having API variants).
>>> Linus would be more difficult to convince, however, I suspect :-)
>> The confusion may be the result of us having barrier semantics in
>> atomic_read. If we take that out then we may avoid future confusions.
> 
> I think better name may help. Nuke atomic_read() altogether.
> 
> n = atomic_value(x);	// doesnt hint as strongly at reading as "atomic_read"
> n = atomic_fetch(x);	// yes, we _do_ touch RAM
> n = atomic_read_uncached(x); // or this
> 
> How does that sound?

atomic_value() vs. atomic_fetch() should be rather unambiguous. 
atomic_read_uncached() begs the question of precisely which cache we are 
avoiding, and could itself cause confusion.

So, if I were writing atomic.h from scratch, knowing what I know now, I think 
I'd use atomic_value() and atomic_fetch().  The problem is that there are a lot 
of existing users of atomic_read(), and we can't write a script to correctly 
guess their intent.  I'm not sure auditing all uses of atomic_read() is really 
worth the comparatively miniscule benefits.

We could play it safe and convert them all to atomic_fetch(), or we could 
acknowledge that changing the semantics 8 months ago was not at all disastrous, 
and make them all atomic_value(), allowing people to use atomic_fetch() where 
they really care.

	-- Chris

RE: [PATCH] i386: Fix a couple busy loops in mach_wakecpu.h:wait_for_init_deassert()

[Luck, Tony]

>>  static inline void wait_for_init_deassert(atomic_t *deassert)
>>  {
>> -	while (!atomic_read(deassert));
>> +	while (!atomic_read(deassert))
>> +		cpu_relax();
>>  	return;
>>  }
>
> For less-than-briliant people like me, it's totally non-obvious that
> cpu_relax() is needed for correctness here, not just to make P4 happy.

Not just P4 ... there are other threaded cpus where it is useful to
let the core know that this is a busy loop so it would be a good thing
to let other threads have priority.

Even on a non-threaded cpu the cpu_relax() could be useful in the
future to hint to the cpu that it could drop into a lower power
hogging state.

But I agree with your main point that the loop without the cpu_relax()
looks like it ought to work because atomic_read() ought to actually
go out and read memory each time around the loop.

-Tony

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Herbert Xu wrote:
> Chris Snook <csnook@redhat.com> wrote:
>> Because atomic operations are generally used for synchronization, which requires 
>> volatile behavior.  Most such codepaths currently use an inefficient barrier(). 
>>  Some forget to and we get bugs, because people assume that atomic_read() 
>> actually reads something, and atomic_write() actually writes something.  Worse, 
>> these are architecture-specific, even compiler version-specific bugs that are 
>> often difficult to track down.
> 
> I'm yet to see a single example from the current tree where
> this patch series is the correct solution.  So far the only
> example has been a buggy piece of code which has since been
> fixed with a cpu_relax.

Part of the motivation here is to fix heisenbugs.  If I knew where they 
were, I'd be posting patches for them.  Unlike most bugs, where we want 
to expose them as obviously as possible, these can be extremely 
difficult to track down, and are often due to people assuming that the 
atomic_* operations have the same semantics they've historically had. 
Remember that until recently, all SMP architectures except s390 (which 
very few kernel developers outside of IBM, Red Hat, and SuSE do much 
work on) had volatile declarations for atomic_t.  Removing the volatile 
declarations from i386 and x86_64 may have created heisenbugs that won't 
manifest themselves until GCC 6.0 comes out and people start compiling 
kernels with -O5.  We should have consistent semantics for atomic_* 
operations.

The other motivation is to reduce the need for the barriers used to 
prevent/fix such problems which clobber all your registers, and instead 
force atomic_* operations to behave in the way they're actually used. 
After the (resubmitted) patchset is merged, we'll be able to remove a 
whole bunch of barriers, shrinking our source and our binaries, and 
improving performance.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 12:13:12PM -0400, Chris Snook wrote:
>
> Part of the motivation here is to fix heisenbugs.  If I knew where they 

By the same token we should probably disable optimisations
altogether since that too can create heisenbugs.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 07:40:21AM +0800, Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 12:13:12PM -0400, Chris Snook wrote:
> >
> > Part of the motivation here is to fix heisenbugs.  If I knew where they 
> 
> By the same token we should probably disable optimisations
> altogether since that too can create heisenbugs.

Precisely the point -- use of volatile (whether in casts or on asms)
in these cases are intended to disable those optimizations likely to
result in heisenbugs.  But they are also intended to leave other
valuable optimizations in force.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>> Part of the motivation here is to fix heisenbugs.  If I knew where 
>>> they
>>
>> By the same token we should probably disable optimisations
>> altogether since that too can create heisenbugs.
>
> Precisely the point -- use of volatile (whether in casts or on asms)
> in these cases are intended to disable those optimizations likely to
> result in heisenbugs.

The only thing volatile on an asm does is create a side effect
on the asm statement; in effect, it tells the compiler "do not
remove this asm even if you don't need any of its outputs".

It's not disabling optimisation likely to result in bugs,
heisen- or otherwise; _not_ putting the volatile on an asm
that needs it simply _is_ a bug :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 03:30:44AM +0200, Segher Boessenkool wrote:
> >>>Part of the motivation here is to fix heisenbugs.  If I knew where 
> >>>they
> >>
> >>By the same token we should probably disable optimisations
> >>altogether since that too can create heisenbugs.
> >
> >Precisely the point -- use of volatile (whether in casts or on asms)
> >in these cases are intended to disable those optimizations likely to
> >result in heisenbugs.
> 
> The only thing volatile on an asm does is create a side effect
> on the asm statement; in effect, it tells the compiler "do not
> remove this asm even if you don't need any of its outputs".
> 
> It's not disabling optimisation likely to result in bugs,
> heisen- or otherwise; _not_ putting the volatile on an asm
> that needs it simply _is_ a bug :-)

Yep.  And the reason it is a bug is that it fails to disable
the relevant compiler optimizations.  So I suspect that we might
actually be saying the same thing here.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> The only thing volatile on an asm does is create a side effect
>> on the asm statement; in effect, it tells the compiler "do not
>> remove this asm even if you don't need any of its outputs".
>>
>> It's not disabling optimisation likely to result in bugs,
>> heisen- or otherwise; _not_ putting the volatile on an asm
>> that needs it simply _is_ a bug :-)
>
> Yep.  And the reason it is a bug is that it fails to disable
> the relevant compiler optimizations.  So I suspect that we might
> actually be saying the same thing here.

We're not saying the same thing, but we do agree :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> Part of the motivation here is to fix heisenbugs.  If I knew where 
>> they
>
> By the same token we should probably disable optimisations
> altogether since that too can create heisenbugs.

Almost everything is a tradeoff; and so is this.  I don't
believe most people would find disabling all compiler
optimisations an acceptable price to pay for some peace
of mind.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Segher Boessenkool wrote:
>>> Part of the motivation here is to fix heisenbugs.  If I knew where they
>>
>>
>> By the same token we should probably disable optimisations
>> altogether since that too can create heisenbugs.
> 
> 
> Almost everything is a tradeoff; and so is this.  I don't
> believe most people would find disabling all compiler
> optimisations an acceptable price to pay for some peace
> of mind.

So why is this a good tradeoff?

I also think that just adding things to APIs in the hope it might fix
up some bugs isn't really a good road to go down. Where do you stop?

On the actual proposal to make atomic_operators volatile: I think the
better approach in the long term, for both maintainability of the
code and education of coders, is to make the use of barriers _more_
explicit rather than sprinkling these "just in case" ones around.

You may get rid of a few atomic_read heisenbugs (in noise when
compared to all bugs), but if the coder was using a regular atomic
load, or a test_bit (which is also atomic), etc. then they're going
to have problems.

It would be better for Linux if everyone was to have better awareness
of barriers than to hide some of the cases where they're required.
A pretty large number of bugs I see in lock free code in the VM is
due to memory ordering problems. It's hard to find those bugs, or
even be aware when you're writing buggy code if you don't have some
feel for barriers.

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>>> Part of the motivation here is to fix heisenbugs.  If I knew where 
>>>> they
>>>
>>>
>>> By the same token we should probably disable optimisations
>>> altogether since that too can create heisenbugs.
>> Almost everything is a tradeoff; and so is this.  I don't
>> believe most people would find disabling all compiler
>> optimisations an acceptable price to pay for some peace
>> of mind.
>
> So why is this a good tradeoff?

It certainly is better than disabling all compiler optimisations!

> I also think that just adding things to APIs in the hope it might fix
> up some bugs isn't really a good road to go down. Where do you stop?

I look at it the other way: keeping the "volatile" semantics in
atomic_XXX() (or adding them to it, whatever) helps _prevent_ bugs;
certainly most people expect that behaviour, and also that behaviour
is *needed* in some places and no other interface provides that
functionality.


[some confusion about barriers wrt atomics snipped]


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Segher Boessenkool wrote:
>>>>> Part of the motivation here is to fix heisenbugs.  If I knew where 
>>>>> they
>>>>
>>>>
>>>>
>>>> By the same token we should probably disable optimisations
>>>> altogether since that too can create heisenbugs.
>>>
>>> Almost everything is a tradeoff; and so is this.  I don't
>>> believe most people would find disabling all compiler
>>> optimisations an acceptable price to pay for some peace
>>> of mind.
>>
>>
>> So why is this a good tradeoff?
> 
> 
> It certainly is better than disabling all compiler optimisations!

It's easy to be better than something really stupid :)

So i386 and x86-64 don't have volatiles there, and it saves them a
few K of kernel text. What you need to justify is why it is a good
tradeoff to make them volatile (which btw, is much harder to go
the other way after we let people make those assumptions).


>> I also think that just adding things to APIs in the hope it might fix
>> up some bugs isn't really a good road to go down. Where do you stop?
> 
> 
> I look at it the other way: keeping the "volatile" semantics in
> atomic_XXX() (or adding them to it, whatever) helps _prevent_ bugs;

Yeah, but we could add lots of things to help prevent bugs and
would never be included. I would also contend that it helps _hide_
bugs and encourages people to be lazy when thinking about these
things.

Also, you dismiss the fact that we'd actually be *adding* volatile
semantics back to the 2 most widely tested architectures (in terms
of test time, number of testers, variety of configurations, and
coverage of driver code). This is a very important different from
just keeping volatile semantics because it is basically a one-way
API change.


> certainly most people expect that behaviour, and also that behaviour
> is *needed* in some places and no other interface provides that
> functionality.

I don't know that most people would expect that behaviour. Is there
any documentation anywhere that would suggest this?

Also, barrier() most definitely provides the required functionality.
It is overkill in some situations, but volatile is overkill in _most_
situations. If that's what you're worried about, we should add a new
ordering primitive.


> [some confusion about barriers wrt atomics snipped]

What were you confused about?

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Nick Piggin writes:

> So i386 and x86-64 don't have volatiles there, and it saves them a
> few K of kernel text. What you need to justify is why it is a good

I'm really surprised it's as much as a few K.  I tried it on powerpc
and it only saved 40 bytes (10 instructions) for a G5 config.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Paul Mackerras wrote:
> Nick Piggin writes:
> 
> 
>>So i386 and x86-64 don't have volatiles there, and it saves them a
>>few K of kernel text. What you need to justify is why it is a good
> 
> 
> I'm really surprised it's as much as a few K.  I tried it on powerpc
> and it only saved 40 bytes (10 instructions) for a G5 config.
> 
> Paul.
> 

I'm surprised too. Numbers were from the "...use asm() like the other
atomic operations already do" thread. According to them,

   text    data     bss     dec     hex filename
3434150  249176  176128 3859454  3ae3fe atomic_normal/vmlinux
3436203  249176  176128 3861507  3aec03 atomic_volatile/vmlinux

The first one is a stock kenel, the second is with atomic_read/set
cast to volatile. gcc-4.1 -- maybe if you have an earlier gcc it
won't optimise as much?

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Fri, 17 Aug 2007, Nick Piggin wrote:
> 
> I'm surprised too. Numbers were from the "...use asm() like the other
> atomic operations already do" thread. According to them,
> 
>   text    data     bss     dec     hex filename
> 3434150  249176  176128 3859454  3ae3fe atomic_normal/vmlinux
> 3436203  249176  176128 3861507  3aec03 atomic_volatile/vmlinux
> 
> The first one is a stock kenel, the second is with atomic_read/set
> cast to volatile. gcc-4.1 -- maybe if you have an earlier gcc it
> won't optimise as much?

No, see my earlier reply. "volatile" really *is* an incredible piece of 
crap.

Just try it yourself:

	volatile int i;
	int j;

	int testme(void)
	{
	        return i <= 1;
	}

	int testme2(void)
	{
	        return j <= 1;
	}

and compile with all the optimizations you can.

I get:

	testme:
	        movl    i(%rip), %eax
	        subl    $1, %eax
	        setle   %al
	        movzbl  %al, %eax
	        ret

vs

	testme2:
	        xorl    %eax, %eax
	        cmpl    $1, j(%rip)
	        setle   %al
	        ret

(now, whether that "xorl + setle" is better than "setle + movzbl", I don't 
really know - maybe it is. But that's not thepoint. The point is the 
difference between

                movl    i(%rip), %eax
                subl    $1, %eax

and

                cmpl    $1, j(%rip)

and imagine this being done for *every* single volatile access.

Just do a 

	git grep atomic_read

to see how atomics are actually used. A lot of them are exactly the above 
kind of "compare against a value".

			Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 08:50:30PM -0700, Linus Torvalds wrote:
> Just try it yourself:
> 
> 	volatile int i;
> 	int j;
> 
> 	int testme(void)
> 	{
> 	        return i <= 1;
> 	}
> 
> 	int testme2(void)
> 	{
> 	        return j <= 1;
> 	}
> 
> and compile with all the optimizations you can.
> 
> I get:
> 
> 	testme:
> 	        movl    i(%rip), %eax
> 	        subl    $1, %eax
> 	        setle   %al
> 	        movzbl  %al, %eax
> 	        ret
> 
> vs
> 
> 	testme2:
> 	        xorl    %eax, %eax
> 	        cmpl    $1, j(%rip)
> 	        setle   %al
> 	        ret
> 
> (now, whether that "xorl + setle" is better than "setle + movzbl", I don't 
> really know - maybe it is. But that's not thepoint. The point is the 
> difference between
> 
>                 movl    i(%rip), %eax
>                 subl    $1, %eax
> 
> and
> 
>                 cmpl    $1, j(%rip)

gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Fri, Aug 17, 2007 at 04:59:12PM -0700, Paul E. McKenney wrote:
>
> gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)

I had totally forgotten that I'd already filed that bug more
than six years ago until they just closed yours as a duplicate
of mine :)

Good luck in getting it fixed!

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Sat, Aug 18, 2007 at 08:09:13AM +0800, Herbert Xu wrote:
> On Fri, Aug 17, 2007 at 04:59:12PM -0700, Paul E. McKenney wrote:
> >
> > gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)
> 
> I had totally forgotten that I'd already filed that bug more
> than six years ago until they just closed yours as a duplicate
> of mine :)
> 
> Good luck in getting it fixed!

Well, just got done re-opening it for the third time.  And a local
gcc community member advised me not to give up too easily.  But I
must admit that I am impressed with the speed that it was identified
as duplicate.

Should be entertaining!  ;-)

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Fri, 17 Aug 2007, Paul E. McKenney wrote:

> On Sat, Aug 18, 2007 at 08:09:13AM +0800, Herbert Xu wrote:
> > On Fri, Aug 17, 2007 at 04:59:12PM -0700, Paul E. McKenney wrote:
> > >
> > > gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)
> > 
> > I had totally forgotten that I'd already filed that bug more
> > than six years ago until they just closed yours as a duplicate
> > of mine :)
> > 
> > Good luck in getting it fixed!
> 
> Well, just got done re-opening it for the third time.  And a local
> gcc community member advised me not to give up too easily.  But I
> must admit that I am impressed with the speed that it was identified
> as duplicate.
> 
> Should be entertaining!  ;-)

Right. ROTFL... volatile actually breaks atomic_t instead of making it 
safe. x++ becomes a register load, increment and a register store. Without 
volatile we can increment the memory directly. It seems that volatile 
requires that the variable is loaded into a register first and then 
operated upon. Understandable when you think about volatile being used to 
access memory mapped I/O registers where a RMW operation could be 
problematic.

See http://gcc.gnu.org/bugzilla/show_bug.cgi?id=3506

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Christoph Lameter wrote:

> On Fri, 17 Aug 2007, Paul E. McKenney wrote:
> 
> > On Sat, Aug 18, 2007 at 08:09:13AM +0800, Herbert Xu wrote:
> > > On Fri, Aug 17, 2007 at 04:59:12PM -0700, Paul E. McKenney wrote:
> > > >
> > > > gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)
> > > 
> > > I had totally forgotten that I'd already filed that bug more
> > > than six years ago until they just closed yours as a duplicate
> > > of mine :)
> > > 
> > > Good luck in getting it fixed!
> > 
> > Well, just got done re-opening it for the third time.  And a local
> > gcc community member advised me not to give up too easily.  But I
> > must admit that I am impressed with the speed that it was identified
> > as duplicate.
> > 
> > Should be entertaining!  ;-)
> 
> Right. ROTFL... volatile actually breaks atomic_t instead of making it 
> safe. x++ becomes a register load, increment and a register store. Without 
> volatile we can increment the memory directly.

No code does (or would do, or should do):

	x.counter++;

on an "atomic_t x;" anyway.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Sat, 18 Aug 2007, Satyam Sharma wrote:
> 
> No code does (or would do, or should do):
> 
> 	x.counter++;
> 
> on an "atomic_t x;" anyway.

That's just an example of a general problem.

No, you don't use "x.counter++". But you *do* use

	if (atomic_read(&x) <= 1)

and loading into a register is stupid and pointless, when you could just 
do it as a regular memory-operand to the cmp instruction.

And as far as the compiler is concerned, the problem is the 100% same: 
combining operations with the volatile memop.

The fact is, a compiler that thinks that

	movl mem,reg
	cmpl $val,reg

is any better than

	cmpl $val,mem

is just not a very good compiler. But when talking about "volatile", 
that's exactly what ytou always get (and always have gotten - this is 
not a regression, and I doubt gcc is alone in this).

			Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Linus Torvalds wrote:

> On Sat, 18 Aug 2007, Satyam Sharma wrote:
> > 
> > No code does (or would do, or should do):
> > 
> > 	x.counter++;
> > 
> > on an "atomic_t x;" anyway.
> 
> That's just an example of a general problem.
> 
> No, you don't use "x.counter++". But you *do* use
> 
> 	if (atomic_read(&x) <= 1)
> 
> and loading into a register is stupid and pointless, when you could just 
> do it as a regular memory-operand to the cmp instruction.

True, but that makes this a bad/poor code generation issue with the
compiler, not something that affects the _correctness_ of atomic ops if
"volatile" is used for that counter object (as was suggested), because
we'd always use the atomic_inc() etc primitives to do increments, which
are always (should be!) implemented to be atomic.


> And as far as the compiler is concerned, the problem is the 100% same: 
> combining operations with the volatile memop.
> 
> The fact is, a compiler that thinks that
> 
> 	movl mem,reg
> 	cmpl $val,reg
> 
> is any better than
> 
> 	cmpl $val,mem
> 
> is just not a very good compiler.

Absolutely, this is definitely a bug report worth opening with gcc. And
what you've said to explain this previously sounds definitely correct --
seeing "volatile" for any access does appear to just scare the hell out
of gcc and makes it generate such (poor) code.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 09:13:35PM -0700, Linus Torvalds wrote:
> 
> 
> On Sat, 18 Aug 2007, Satyam Sharma wrote:
> > 
> > No code does (or would do, or should do):
> > 
> > 	x.counter++;
> > 
> > on an "atomic_t x;" anyway.
> 
> That's just an example of a general problem.
> 
> No, you don't use "x.counter++". But you *do* use
> 
> 	if (atomic_read(&x) <= 1)
> 
> and loading into a register is stupid and pointless, when you could just 
> do it as a regular memory-operand to the cmp instruction.
> 
> And as far as the compiler is concerned, the problem is the 100% same: 
> combining operations with the volatile memop.
> 
> The fact is, a compiler that thinks that
> 
> 	movl mem,reg
> 	cmpl $val,reg
> 
> is any better than
> 
> 	cmpl $val,mem
> 
> is just not a very good compiler. But when talking about "volatile", 
> that's exactly what ytou always get (and always have gotten - this is 
> not a regression, and I doubt gcc is alone in this).

One of the gcc guys claimed that he thought that the two-instruction
sequence would be faster on some x86 machines.  I pointed out that
there might be a concern about code size.  I chose not to point out
that people might also care about the other x86 machines.  ;-)

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Sat, 18 Aug 2007, Paul E. McKenney wrote:
> 
> One of the gcc guys claimed that he thought that the two-instruction
> sequence would be faster on some x86 machines.  I pointed out that
> there might be a concern about code size.  I chose not to point out
> that people might also care about the other x86 machines.  ;-)

Some (very few) x86 uarchs do tend to prefer "load-store" like code 
generation, and doing a "mov [mem],reg + op reg" instead of "op [mem]" can 
actually be faster on some of them. Not any that are relevant today, 
though.

Also, that has nothing to do with volatile, and should be controlled by 
optimization flags (like -mtune). In fact, I thought there was a separate 
flag to do that (ie something like "-mload-store"), but I can't find it, 
so maybe that's just my fevered brain..

			Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Sat, Aug 18, 2007 at 03:41:13PM -0700, Linus Torvalds wrote:
> 
> 
> On Sat, 18 Aug 2007, Paul E. McKenney wrote:
> > 
> > One of the gcc guys claimed that he thought that the two-instruction
> > sequence would be faster on some x86 machines.  I pointed out that
> > there might be a concern about code size.  I chose not to point out
> > that people might also care about the other x86 machines.  ;-)
> 
> Some (very few) x86 uarchs do tend to prefer "load-store" like code 
> generation, and doing a "mov [mem],reg + op reg" instead of "op [mem]" can 
> actually be faster on some of them. Not any that are relevant today, 
> though.

;-)

> Also, that has nothing to do with volatile, and should be controlled by 
> optimization flags (like -mtune). In fact, I thought there was a separate 
> flag to do that (ie something like "-mload-store"), but I can't find it, 
> so maybe that's just my fevered brain..

Good point, will suggest this if the need arises.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Saturday 18 August 2007 05:13, Linus Torvalds wrote:
> On Sat, 18 Aug 2007, Satyam Sharma wrote:
> > No code does (or would do, or should do):
> >
> > 	x.counter++;
> >
> > on an "atomic_t x;" anyway.
>
> That's just an example of a general problem.
>
> No, you don't use "x.counter++". But you *do* use
>
> 	if (atomic_read(&x) <= 1)
>
> and loading into a register is stupid and pointless, when you could just
> do it as a regular memory-operand to the cmp instruction.

It doesn't mean that (volatile int*) cast is bad, it means that current gcc
is bad (or "not good enough"). IOW: instead of avoiding volatile cast,
it's better to fix the compiler.

> And as far as the compiler is concerned, the problem is the 100% same:
> combining operations with the volatile memop.
>
> The fact is, a compiler that thinks that
>
> 	movl mem,reg
> 	cmpl $val,reg
>
> is any better than
>
> 	cmpl $val,mem
>
> is just not a very good compiler.

Linus, in all honesty gcc has many more cases of suboptimal code,
case of "volatile" is just one of many.

Off the top of my head:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28417

unsigned v;
void f(unsigned A) { v = ((unsigned long long)A) * 365384439 >> (27+32); }

gcc-4.1.1 -S -Os -fomit-frame-pointer t.c

f:
        movl    $365384439, %eax
        mull    4(%esp)
        movl    %edx, %eax <===== ?
        shrl    $27, %eax
        movl    %eax, v
        ret

Why is it moving %edx to %eax?

gcc-4.2.1 -S -Os -fomit-frame-pointer t.c

f:
        movl    $365384439, %eax
        mull    4(%esp)
        movl    %edx, %eax <===== ?
        xorl    %edx, %edx <===== ??!
        shrl    $27, %eax
        movl    %eax, v
        ret

Progress... Now we also zero out %edx afterwards for no apparent reason.
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Fri, 24 Aug 2007, Denys Vlasenko wrote:
>
> > No, you don't use "x.counter++". But you *do* use
> >
> > 	if (atomic_read(&x) <= 1)
> >
> > and loading into a register is stupid and pointless, when you could just
> > do it as a regular memory-operand to the cmp instruction.
> 
> It doesn't mean that (volatile int*) cast is bad, it means that current gcc
> is bad (or "not good enough"). IOW: instead of avoiding volatile cast,
> it's better to fix the compiler.

I would agree that fixing the compiler in this case would be a good thing, 
even quite regardless of any "atomic_read()" discussion.

I just have a strong suspicion that "volatile" performance is so low down 
the list of any C compiler persons interest, that it's never going to 
happen. And quite frankly, I cannot blame the gcc guys for it.

That's especially as "volatile" really isn't a very good feature of the C 
language, and is likely to get *less* interesting rather than more (as 
user space starts to be more and more threaded, "volatile" gets less and 
less useful.

[ Ie, currently, I think you can validly use "volatile" in a "sigatomic_t" 
  kind of way, where there is a single thread, but with asynchronous 
  events. In that kind of situation, I think it's probably useful. But 
  once you get multiple threads, it gets pointless.

  Sure: you could use "volatile" together with something like Dekker's or 
  Peterson's algorithm that doesn't depend on cache coherency (that's 
  basically what the C "volatile" keyword approximates: not atomic 
  accesses, but *uncached* accesses! But let's face it, that's way past 
  insane. ]

So I wouldn't expect "volatile" to ever really generate better code. It 
might happen as a side effect of other improvements (eg, I might hope that 
the SSA work would eventually lead to gcc having a much better defined 
model of valid optimizations, and maybe better code generation for 
volatile accesses fall out cleanly out of that), but in the end, it's such 
an ugly special case in C, and so seldom used, that I wouldn't depend on 
it.

> Linus, in all honesty gcc has many more cases of suboptimal code,
> case of "volatile" is just one of many.

Well, the thing is, quite often, many of those "suboptimal code" 
generations fall into two distinct classes:

 - complex C code. I can't really blame the compiler too much for this. 
   Some things are *hard* to optimize, and for various scalability 
   reasons, you often end up having limits in the compiler where it 
   doesn't even _try_ doing certain optimizations if you have excessive 
   complexity.

 - bad register allocation. Register allocation really is hard, and 
   sometimes gcc just does the "obviously wrong" thing, and you end up 
   having totally unnecessary spills.

> Off the top of my head:

Yes, "unsigned long long" with x86 has always generated atrocious code. In 
fact, I would say that historically it was really *really* bad. These 
days, gcc actually does a pretty good job, but I'm not surprised that it's 
still quite possible to find cases where it did some optimization (in this 
case, apparently noticing that "shift by >= 32 bits" causes the low 
register to be pointless) and then missed *another* optimization (better 
register use) because that optimization had been done *before* the first 
optimization was done.

That's a *classic* example of compiler code generation issues, and quite 
frankly, I think that's very different from the issue of "volatile".

Quite frankly, I'd like there to be more competition in the open source 
compiler game, and that might cause some upheavals, but on the whole, gcc 
actually does a pretty damn good job. 

			Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 06:24:15PM -0700, Christoph Lameter wrote:
> On Fri, 17 Aug 2007, Paul E. McKenney wrote:
> 
> > On Sat, Aug 18, 2007 at 08:09:13AM +0800, Herbert Xu wrote:
> > > On Fri, Aug 17, 2007 at 04:59:12PM -0700, Paul E. McKenney wrote:
> > > >
> > > > gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)
> > > 
> > > I had totally forgotten that I'd already filed that bug more
> > > than six years ago until they just closed yours as a duplicate
> > > of mine :)
> > > 
> > > Good luck in getting it fixed!
> > 
> > Well, just got done re-opening it for the third time.  And a local
> > gcc community member advised me not to give up too easily.  But I
> > must admit that I am impressed with the speed that it was identified
> > as duplicate.
> > 
> > Should be entertaining!  ;-)
> 
> Right. ROTFL... volatile actually breaks atomic_t instead of making it 
> safe. x++ becomes a register load, increment and a register store. Without 
> volatile we can increment the memory directly. It seems that volatile 
> requires that the variable is loaded into a register first and then 
> operated upon. Understandable when you think about volatile being used to 
> access memory mapped I/O registers where a RMW operation could be 
> problematic.
> 
> See http://gcc.gnu.org/bugzilla/show_bug.cgi?id=3506

Yep.  The initial reaction was in fact to close my bug as a duplicate
of 3506.  But I was not asking for atomicity, but rather for smaller
code to be generated, so I reopened it.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Christoph Lameter wrote:
> On Fri, 17 Aug 2007, Paul E. McKenney wrote:
> 
>> On Sat, Aug 18, 2007 at 08:09:13AM +0800, Herbert Xu wrote:
>>> On Fri, Aug 17, 2007 at 04:59:12PM -0700, Paul E. McKenney wrote:
>>>> gcc bugzilla bug #33102, for whatever that ends up being worth.  ;-)
>>> I had totally forgotten that I'd already filed that bug more
>>> than six years ago until they just closed yours as a duplicate
>>> of mine :)
>>>
>>> Good luck in getting it fixed!
>> Well, just got done re-opening it for the third time.  And a local
>> gcc community member advised me not to give up too easily.  But I
>> must admit that I am impressed with the speed that it was identified
>> as duplicate.
>>
>> Should be entertaining!  ;-)
> 
> Right. ROTFL... volatile actually breaks atomic_t instead of making it 
> safe. x++ becomes a register load, increment and a register store. Without 
> volatile we can increment the memory directly. It seems that volatile 
> requires that the variable is loaded into a register first and then 
> operated upon. Understandable when you think about volatile being used to 
> access memory mapped I/O registers where a RMW operation could be 
> problematic.

So, if we want consistent behavior, we're pretty much screwed unless we use 
inline assembler everywhere?

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> Right. ROTFL... volatile actually breaks atomic_t instead of making 
>> it safe. x++ becomes a register load, increment and a register store. 
>> Without volatile we can increment the memory directly. It seems that 
>> volatile requires that the variable is loaded into a register first 
>> and then operated upon. Understandable when you think about volatile 
>> being used to access memory mapped I/O registers where a RMW 
>> operation could be problematic.
>
> So, if we want consistent behavior, we're pretty much screwed unless 
> we use inline assembler everywhere?

Nah, this whole argument is flawed -- "without volatile" we still
*cannot* "increment the memory directly".  On x86, you need a lock
prefix; on other archs, some other mechanism to make the memory
increment an *atomic* memory increment.

And no, RMW on MMIO isn't "problematic" at all, either.

An RMW op is a read op, a modify op, and a write op, all rolled
into one opcode.  But three actual operations.


The advantages of asm code for atomic_{read,set} are:
1) all the other atomic ops are implemented that way already;
2) you have full control over the asm insns selected, in particular,
    you can guarantee you *do* get an atomic op;
3) you don't need to use "volatile <data>" which generates
    not-all-that-good code on archs like x86, and we want to get rid
    of it anyway since it is problematic in many ways;
4) you don't need to use *(volatile <type>*)&<data>, which a) doesn't
    exist in C; b) isn't documented or supported in GCC; c) has a recent
    history of bugginess; d) _still uses volatile objects_; e) _still_
    is problematic in almost all those same ways as in 3);
5) you can mix atomic and non-atomic accesses to the atomic_t, which
    you cannot with the other alternatives.

The only disadvantage I know of is potentially slightly worse
instruction scheduling.  This is a generic asm() problem: GCC
cannot see what actual insns are inside the asm() block.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Russell King]

On Tue, Aug 21, 2007 at 12:04:17AM +0200, Segher Boessenkool wrote:
> And no, RMW on MMIO isn't "problematic" at all, either.
> 
> An RMW op is a read op, a modify op, and a write op, all rolled
> into one opcode.  But three actual operations.

Maybe for some CPUs, but not all.  ARM for instance can't use the
load exclusive and store exclusive instructions to MMIO space.

This means placing atomic_t or bitops into MMIO space is a definite
no-go on ARM.  It breaks.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> And no, RMW on MMIO isn't "problematic" at all, either.
>>
>> An RMW op is a read op, a modify op, and a write op, all rolled
>> into one opcode.  But three actual operations.
>
> Maybe for some CPUs, but not all.  ARM for instance can't use the
> load exclusive and store exclusive instructions to MMIO space.

Sure, your CPU doesn't have RMW instructions -- how to emulate
those if you don't have them is a totally different thing.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Tue, Aug 21, 2007 at 01:02:01AM +0200, Segher Boessenkool wrote:
> >>And no, RMW on MMIO isn't "problematic" at all, either.
> >>
> >>An RMW op is a read op, a modify op, and a write op, all rolled
> >>into one opcode.  But three actual operations.
> >
> >Maybe for some CPUs, but not all.  ARM for instance can't use the
> >load exclusive and store exclusive instructions to MMIO space.
> 
> Sure, your CPU doesn't have RMW instructions -- how to emulate
> those if you don't have them is a totally different thing.

I thought that ARM's load exclusive and store exclusive instructions
were its equivalent of LL and SC, which RISC machines typically use to
build atomic sequences of instructions -- and which normally cannot be
applied to MMIO space.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Russell King]

On Mon, Aug 20, 2007 at 05:05:18PM -0700, Paul E. McKenney wrote:
> On Tue, Aug 21, 2007 at 01:02:01AM +0200, Segher Boessenkool wrote:
> > >>And no, RMW on MMIO isn't "problematic" at all, either.
> > >>
> > >>An RMW op is a read op, a modify op, and a write op, all rolled
> > >>into one opcode.  But three actual operations.
> > >
> > >Maybe for some CPUs, but not all.  ARM for instance can't use the
> > >load exclusive and store exclusive instructions to MMIO space.
> > 
> > Sure, your CPU doesn't have RMW instructions -- how to emulate
> > those if you don't have them is a totally different thing.
> 
> I thought that ARM's load exclusive and store exclusive instructions
> were its equivalent of LL and SC, which RISC machines typically use to
> build atomic sequences of instructions -- and which normally cannot be
> applied to MMIO space.

Absolutely correct.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Russell King]

On Tue, Aug 21, 2007 at 01:02:01AM +0200, Segher Boessenkool wrote:
> >>And no, RMW on MMIO isn't "problematic" at all, either.
> >>
> >>An RMW op is a read op, a modify op, and a write op, all rolled
> >>into one opcode.  But three actual operations.
> >
> >Maybe for some CPUs, but not all.  ARM for instance can't use the
> >load exclusive and store exclusive instructions to MMIO space.
> 
> Sure, your CPU doesn't have RMW instructions -- how to emulate
> those if you don't have them is a totally different thing.

Let me say it more clearly: On ARM, it is impossible to perform atomic
operations on MMIO space.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Russell King writes:

> Let me say it more clearly: On ARM, it is impossible to perform atomic
> operations on MMIO space.

Actually, no one is suggesting that we try to do that at all.

The discussion about RMW ops on MMIO space started with a comment
attributed to the gcc developers that one reason why gcc on x86
doesn't use instructions that do RMW ops on volatile variables is that
volatile is used to mark MMIO addresses, and there was some
uncertainty about whether (non-atomic) RMW ops on x86 could be used on
MMIO.  This is in regard to the question about why gcc on x86 always
moves a volatile variable into a register before doing anything to it.

So the whole discussion is irrelevant to ARM, PowerPC and any other
architecture except x86[-64].

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Andi Kleen]

On Tue, Aug 21, 2007 at 07:33:49PM +1000, Paul Mackerras wrote:
> So the whole discussion is irrelevant to ARM, PowerPC and any other
> architecture except x86[-64].

It's even irrelevant on x86 because all modifying operations on atomic_t 
are coded in inline assembler and will always be RMW no matter
if atomic_t is volatile or not.

[ignoring atomic_set(x, atomic_read(x) + 1) which nobody should do]

The only issue is if atomic_t should have a implicit barrier or not.
My personal opinion is yes -- better safe than sorry. And any code
impact it may have is typically dwarved by the next cache miss anyways,
so it doesn't matter much.

-Andi

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> Let me say it more clearly: On ARM, it is impossible to perform atomic
>> operations on MMIO space.
>
> Actually, no one is suggesting that we try to do that at all.
>
> The discussion about RMW ops on MMIO space started with a comment
> attributed to the gcc developers that one reason why gcc on x86
> doesn't use instructions that do RMW ops on volatile variables is that
> volatile is used to mark MMIO addresses, and there was some
> uncertainty about whether (non-atomic) RMW ops on x86 could be used on
> MMIO.  This is in regard to the question about why gcc on x86 always
> moves a volatile variable into a register before doing anything to it.

This question is GCC PR33102, which was incorrectly closed as a 
duplicate
of PR3506 -- and *that* PR was closed because its reporter seemed to
claim the GCC generated code for an increment on a volatile (namely, 
three
machine instructions: load, modify, store) was incorrect, and it has to
be one machine instruction.

> So the whole discussion is irrelevant to ARM, PowerPC and any other
> architecture except x86[-64].

And even there, it's not something the kernel can take advantage of
before GCC 4.4 is in widespread use, if then.  Let's move on.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Tue, Aug 21, 2007 at 04:48:51PM +0200, Segher Boessenkool wrote:
> >>Let me say it more clearly: On ARM, it is impossible to perform atomic
> >>operations on MMIO space.
> >
> >Actually, no one is suggesting that we try to do that at all.
> >
> >The discussion about RMW ops on MMIO space started with a comment
> >attributed to the gcc developers that one reason why gcc on x86
> >doesn't use instructions that do RMW ops on volatile variables is that
> >volatile is used to mark MMIO addresses, and there was some
> >uncertainty about whether (non-atomic) RMW ops on x86 could be used on
> >MMIO.  This is in regard to the question about why gcc on x86 always
> >moves a volatile variable into a register before doing anything to it.
> 
> This question is GCC PR33102, which was incorrectly closed as a 
> duplicate
> of PR3506 -- and *that* PR was closed because its reporter seemed to
> claim the GCC generated code for an increment on a volatile (namely, 
> three
> machine instructions: load, modify, store) was incorrect, and it has to
> be one machine instruction.
> 
> >So the whole discussion is irrelevant to ARM, PowerPC and any other
> >architecture except x86[-64].
> 
> And even there, it's not something the kernel can take advantage of
> before GCC 4.4 is in widespread use, if then.  Let's move on.

I agree that instant gratification is hard to come by when synching
up compiler and kernel versions.  Nonetheless, it should be possible
to create APIs that are are conditioned on the compiler version.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Valdis.Kletnieks]

[-- Attachment #1: Type: text/plain, Size: 557 bytes --]

On Tue, 21 Aug 2007 09:16:43 PDT, "Paul E. McKenney" said:

> I agree that instant gratification is hard to come by when synching
> up compiler and kernel versions.  Nonetheless, it should be possible
> to create APIs that are are conditioned on the compiler version.

We've tried that, sort of.  See the mess surrounding the whole
extern/static/inline/__whatever boondogle, which seems to have
changed semantics in every single gcc release since 2.95 or so.

And recently mention was made that gcc4.4 will have *new* semantics
in this area. Yee. Hah.






[-- Attachment #2: Type: application/pgp-signature, Size: 226 bytes --]

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Tue, Aug 21, 2007 at 06:51:16PM -0400, Valdis.Kletnieks@vt.edu wrote:
> On Tue, 21 Aug 2007 09:16:43 PDT, "Paul E. McKenney" said:
> 
> > I agree that instant gratification is hard to come by when synching
> > up compiler and kernel versions.  Nonetheless, it should be possible
> > to create APIs that are are conditioned on the compiler version.
> 
> We've tried that, sort of.  See the mess surrounding the whole
> extern/static/inline/__whatever boondogle, which seems to have
> changed semantics in every single gcc release since 2.95 or so.
> 
> And recently mention was made that gcc4.4 will have *new* semantics
> in this area. Yee. Hah.

;-)

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Adrian Bunk]

On Tue, Aug 21, 2007 at 06:51:16PM -0400, Valdis.Kletnieks@vt.edu wrote:
> On Tue, 21 Aug 2007 09:16:43 PDT, "Paul E. McKenney" said:
> 
> > I agree that instant gratification is hard to come by when synching
> > up compiler and kernel versions.  Nonetheless, it should be possible
> > to create APIs that are are conditioned on the compiler version.
> 
> We've tried that, sort of.  See the mess surrounding the whole
> extern/static/inline/__whatever boondogle, which seems to have
> changed semantics in every single gcc release since 2.95 or so.
>...

There is exactly one semantics change in gcc in this area, and that is 
the change of the "extern inline" semantics in gcc 4.3 to the
C99 semantics.

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>>> And no, RMW on MMIO isn't "problematic" at all, either.
>>>>
>>>> An RMW op is a read op, a modify op, and a write op, all rolled
>>>> into one opcode.  But three actual operations.
>>>
>>> Maybe for some CPUs, but not all.  ARM for instance can't use the
>>> load exclusive and store exclusive instructions to MMIO space.
>>
>> Sure, your CPU doesn't have RMW instructions -- how to emulate
>> those if you don't have them is a totally different thing.
>
> Let me say it more clearly: On ARM, it is impossible to perform atomic
> operations on MMIO space.

It's all completely beside the point, see the other subthread, but...

Yeah, you can't do LL/SC to MMIO space; ARM isn't alone in that.
You could still implement atomic operations on MMIO space by taking
a lock elsewhere, in normal cacheable memory space.  Why you would
do this is a separate question, you probably don't want it :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Fri, 17 Aug 2007, Paul Mackerras wrote:
> 
> I'm really surprised it's as much as a few K.  I tried it on powerpc
> and it only saved 40 bytes (10 instructions) for a G5 config.

One of the things that "volatile" generally screws up is a simple

	volatile int i;

	i++;

which a compiler will generally get horribly, horribly wrong.

In a reasonable world, gcc should just make that be (on x86)

	addl $1,i(%rip)

on x86-64, which is indeed what it does without the volatile. But with the 
volatile, the compiler gets really nervous, and doesn't dare do it in one 
instruction, and thus generates crap like

        movl    i(%rip), %eax
        addl    $1, %eax
        movl    %eax, i(%rip)

instead. For no good reason, except that "volatile" just doesn't have any 
good/clear semantics for the compiler, so most compilers will just make it 
be "I will not touch this access in any way, shape, or form". Including 
even trivially correct instruction optimization/combination.

This is one of the reasons why we should never use "volatile". It 
pessimises code generation for no good reason - just because compilers 
don't know what the heck it even means! 

Now, people don't do "i++" on atomics (you'd use "atomic_inc()" for that), 
but people *do* do things like

	if (atomic_read(..) <= 1)
		..

On ppc, things like that probably don't much matter. But on x86, it makes 
a *huge* difference whether you do

	movl i(%rip),%eax
	cmpl $1,%eax

or if you can just use the value directly for the operation, like this:

	cmpl $1,i(%rip)

which is again a totally obvious and totally safe optimization, but is 
(again) something that gcc doesn't dare do, since "i" is volatile.

In other words: "volatile" is a horribly horribly bad way of doing things, 
because it generates *worse*code*, for no good reason. You just don't see 
it on powerpc, because it's already a load-store architecture, so there is 
no "good code" for doing direct-to-memory operations.

		Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 08:42:23PM -0700, Linus Torvalds wrote:
> 
> 
> On Fri, 17 Aug 2007, Paul Mackerras wrote:
> > 
> > I'm really surprised it's as much as a few K.  I tried it on powerpc
> > and it only saved 40 bytes (10 instructions) for a G5 config.
> 
> One of the things that "volatile" generally screws up is a simple
> 
> 	volatile int i;
> 
> 	i++;
> 
> which a compiler will generally get horribly, horribly wrong.
> 
> In a reasonable world, gcc should just make that be (on x86)
> 
> 	addl $1,i(%rip)
> 
> on x86-64, which is indeed what it does without the volatile. But with the 
> volatile, the compiler gets really nervous, and doesn't dare do it in one 
> instruction, and thus generates crap like
> 
>         movl    i(%rip), %eax
>         addl    $1, %eax
>         movl    %eax, i(%rip)

Blech.  Sounds like a chat with some gcc people is in order.  Will
see what I can do.

						Thanx, Paul

> instead. For no good reason, except that "volatile" just doesn't have any 
> good/clear semantics for the compiler, so most compilers will just make it 
> be "I will not touch this access in any way, shape, or form". Including 
> even trivially correct instruction optimization/combination.
> 
> This is one of the reasons why we should never use "volatile". It 
> pessimises code generation for no good reason - just because compilers 
> don't know what the heck it even means! 
> 
> Now, people don't do "i++" on atomics (you'd use "atomic_inc()" for that), 
> but people *do* do things like
> 
> 	if (atomic_read(..) <= 1)
> 		..
> 
> On ppc, things like that probably don't much matter. But on x86, it makes 
> a *huge* difference whether you do
> 
> 	movl i(%rip),%eax
> 	cmpl $1,%eax
> 
> or if you can just use the value directly for the operation, like this:
> 
> 	cmpl $1,i(%rip)
> 
> which is again a totally obvious and totally safe optimization, but is 
> (again) something that gcc doesn't dare do, since "i" is volatile.
> 
> In other words: "volatile" is a horribly horribly bad way of doing things, 
> because it generates *worse*code*, for no good reason. You just don't see 
> it on powerpc, because it's already a load-store architecture, so there is 
> no "good code" for doing direct-to-memory operations.
> 
> 		Linus
> -
> To unsubscribe from this list: send the line "unsubscribe netdev" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

Hi Linus,

[ and others; I think there's a communication gap in a lot of this
  thread, and a little summary would be useful. Hence this posting. ]


On Thu, 16 Aug 2007, Linus Torvalds wrote:

> On Fri, 17 Aug 2007, Paul Mackerras wrote:
> > 
> > I'm really surprised it's as much as a few K.  I tried it on powerpc
> > and it only saved 40 bytes (10 instructions) for a G5 config.
> 
> One of the things that "volatile" generally screws up is a simple
> 
> 	volatile int i;
> 
> 	i++;
> 
> which a compiler will generally get horribly, horribly wrong.
> 
> [...] For no good reason, except that "volatile" just doesn't have any 
> good/clear semantics for the compiler, so most compilers will just make it 
> be "I will not touch this access in any way, shape, or form". Including 
> even trivially correct instruction optimization/combination.
> 
> This is one of the reasons why we should never use "volatile". It 
> pessimises code generation for no good reason - just because compilers 
> don't know what the heck it even means! 
> [...]
> In other words: "volatile" is a horribly horribly bad way of doing things, 
> because it generates *worse*code*, for no good reason. You just don't see 
> it on powerpc, because it's already a load-store architecture, so there is 
> no "good code" for doing direct-to-memory operations.


True, and I bet *everybody* on this list has already agreed for a very
long time that using "volatile" to type-qualify the _declaration_ of an
object itself as being horribly bad (taste-wise, code-generation-wise,
often even buggy for sitations where real CPU barriers should've been
used instead).

However, the discussion on this thread (IIRC) began with only "giving
volatility semantics" to atomic ops. Now that is different, and may not
require the use the "volatile" keyword (at least not in the declaration
of the object) itself.

Sadly, most arch's *still* do type-qualify the declaration of the
"counter" member of atomic_t as "volatile". This is probably a historic
hangover, and I suspect not yet rectified because of lethargy.

Anyway, some of the variants I can think of are:

[1]

#define atomic_read_volatile(v)				\
	({						\
		forget((v)->counter);			\
		((v)->counter);				\
	})

where:

#define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))

[ This is exactly equivalent to using "+m" in the constraints, as recently
  explained on a GCC list somewhere, in response to the patch in my bitops
  series a few weeks back where I thought "+m" was bogus. ]

[2]

#define atomic_read_volatile(v)		(*(volatile int *)&(v)->counter)

This is something that does work. It has reasonably good semantics
guaranteed by the C standard in conjunction with how GCC currently
behaves (and how it has behaved for all supported versions). I haven't
checked if generates much different code than the first variant above,
(it probably will generate similar code to just declaring the object
as volatile, but would still be better in terms of code-clarity and
taste, IMHO), but in any case, we should pick whichever of these variants
works for us and generates good code.

[3]

static inline int atomic_read_volatile(atomic_t *v)
{
	... arch-dependent __asm__ __volatile__ stuff ...
}

I can reasonably bet this variant would often generate worse code than
at least the variant "[1]" above.


Now, why do we even require these "volatility" semantics variants?

Note, "volatility" semantics *know* / assume that it can have a meaning
_only_ as far as the compiler, so atomic_read_volatile() doesn't really
care reading stale values from the cache for certain non-x86 archs, etc.

The first argument is "safety":

Use of atomic_read() (possibly in conjunction with other atomic ops) in
a lot of code out there in the kernel *assumes* the compiler will not
optimize away those ops. (which is possible given current definitions
of atomic_set and atomic_read on archs such as x86 in present code).
An additional argument that builds on this one says that by ensuring
the compiler will not elid or coalesce these ops, we could even avoid
potential heisenbugs in the future.

However, there is a counter-argument:

As Herbert Xu has often been making the point, there is *no* bug out
there involving "atomic_read" in busy-while-loops that should not have
a compiler barrier (or cpu_relax() in fact) anyway. As for non-busy-loops,
they would invariable call schedule() at some point (possibly directly)
and thus have an "implicit" compiler barrier by virtue of calling out
a function that is not in scope of the current compilation unit (although
users in sched.c itself would probably require an explicit compiler
barrier).

The second pro-volatility-in-atomic-ops argument is performance:
(surprise!)

Using a full memory clobber compiler barrier in busy loops will disqualify
optimizations for loop invariants so it probably makes sense to *only*
make the compiler forget *that* particular address of the atomic counter
object, and none other. All 3 variants above would work nicely here.

So the final idea may be to have a cpu_relax_no_barrier() variant as a
rep;nop (pause) *without* an included full memory clobber, and replace
a lot of kernel busy-while-loops out there with:

-	cpu_relax();
+	cpu_relax_no_barrier();
+	forget(x);

or may be just:

-	cpu_relax();
+	cpu_relax_no_barrier();

because the "forget" / "volatility" / specific-variable-compiler-barrier
could be made implicit inside the atomic ops themselves.

This could especially make a difference for register-rich CPUs (probably
not x86) where using a full memory clobber will disqualify a hell lot of
compiler optimizations for loop-invariants.

On x86 itself, cpu_relax_no_barrier() could be:

#define cpu_relax_no_barrier()	__asm__ __volatile__ ("rep;nop":::);

and still continue to do its job as it is doing presently.

However, there is still a counter-argument:

As Herbert Xu and Christoph Lameter have often been saying, giving
"volatility" semantics to the atomic ops will disqualify compiler
optimizations such as eliding / coalescing of atomic ops, etc, and
probably some sections of code in the kernel (Christoph mentioned code
in SLUB, and I saw such code in sched) benefit from such optimizations.

Paul Mackerras has, otoh, mentioned that a lot of such places probably
don't need (or shouldn't use) atomic ops in the first place.
Alternatively, such callsites should probably just cache the atomic_read
in a local variable (which compiler could just as well make a register)
explicitly, and repeating atomic_read() isn't really necessary.

There could still be legitimate uses of atomic ops that don't care about
them being elided / coalesced, but given the loop-invariant-optimization
benefit, personally, I do see some benefit in the use of defining atomic
ops variants with "volatility" semantics (for only the atomic counter
object) but also having a non-volatile atomic ops API side-by-side for
performance critical users (probably sched, slub) that may require that.

Possibly, one of the two APIs above could turn out to be redundant, but
that's still very much the issue of debate presently.


Satyam

[ Sorry if I missed anything important, but this thread has been long
  and noisy, although I've tried to keep up ... ]

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:

> #define atomic_read_volatile(v)				\
> 	({						\
> 		forget((v)->counter);			\
> 		((v)->counter);				\
> 	})
> 
> where:

*vomit* :)

Not only do I hate the keyword volatile, but the barrier is only a
one-sided affair so its probable this is going to have slightly
different allowed reorderings than a real volatile access.

Also, why would you want to make these insane accessors for atomic_t
types? Just make sure everybody knows the basics of barriers, and they
can apply that knowledge to atomic_t and all other lockless memory
accesses as well.


> #define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))

I like order(x) better, but it's not the most perfect name either.

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> 
> > #define atomic_read_volatile(v)				\
> > 	({						\
> > 		forget((v)->counter);			\
> > 		((v)->counter);				\
> > 	})
> > 
> > where:
> 
> *vomit* :)

I wonder if this'll generate smaller and better code than _both_ the
other atomic_read_volatile() variants. Would need to build allyesconfig
on lots of diff arch's etc to test the theory though.


> Not only do I hate the keyword volatile, but the barrier is only a
> one-sided affair so its probable this is going to have slightly
> different allowed reorderings than a real volatile access.

True ...


> Also, why would you want to make these insane accessors for atomic_t
> types? Just make sure everybody knows the basics of barriers, and they
> can apply that knowledge to atomic_t and all other lockless memory
> accesses as well.

Code that looks like:

	while (!atomic_read(&v)) {
		...
		cpu_relax_no_barrier();
		forget(v.counter);
		        ^^^^^^^^
	}

would be uglier. Also think about code such as:

	a = atomic_read();
	if (!a)
		do_something();

	forget();
	a = atomic_read();
	... /* some code that depends on value of a, obviously */

	forget();
	a = atomic_read();
	...

So much explicit sprinkling of "forget()" looks ugly.

	atomic_read_volatile()

on the other hand, looks neater. The "_volatile()" suffix makes it also
no less explicit than an explicit barrier-like macro that this primitive
is something "special", for code clarity purposes.


> > #define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))
> 
> I like order(x) better, but it's not the most perfect name either.

forget(x) is just a stupid-placeholder-for-a-better-name. order(x) sounds
good but we could leave quibbling about function or macro names for later,
this thread is noisy as it is :-)

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:
> 
> On Fri, 17 Aug 2007, Nick Piggin wrote:

>>Also, why would you want to make these insane accessors for atomic_t
>>types? Just make sure everybody knows the basics of barriers, and they
>>can apply that knowledge to atomic_t and all other lockless memory
>>accesses as well.
> 
> 
> Code that looks like:
> 
> 	while (!atomic_read(&v)) {
> 		...
> 		cpu_relax_no_barrier();
> 		forget(v.counter);
> 		        ^^^^^^^^
> 	}
> 
> would be uglier. Also think about code such as:

I think they would both be equally ugly, but the atomic_read_volatile
variant would be more prone to subtle bugs because of the weird
implementation.

And it would be more ugly than introducing an order(x) statement for
all memory operations, and adding an order_atomic() wrapper for it
for atomic types.


> 	a = atomic_read();
> 	if (!a)
> 		do_something();
> 
> 	forget();
> 	a = atomic_read();
> 	... /* some code that depends on value of a, obviously */
> 
> 	forget();
> 	a = atomic_read();
> 	...
> 
> So much explicit sprinkling of "forget()" looks ugly.

Firstly, why is it ugly? It's nice because of those nice explicit
statements there that give us a good heads up and would have some
comments attached to them (also, lack of the word "volatile" is
always a plus).

Secondly, what sort of code would do such a thing? In most cases,
it is probably riddled with bugs anyway (unless it is doing a
really specific sequence of interrupts or something, but in that
case it is very likely to either require locking or busy waits
anyway -> ie. barriers).


> on the other hand, looks neater. The "_volatile()" suffix makes it also
> no less explicit than an explicit barrier-like macro that this primitive
> is something "special", for code clarity purposes.

Just don't use the word volatile, and have barriers both before
and after the memory operation, and I'm OK with it. I don't see
the point though, when you could just have a single barrier(x)
barrier function defined for all memory locations, rather than
this odd thing that only works for atomics (and would have to
be duplicated for atomic_set.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> > 
> > On Fri, 17 Aug 2007, Nick Piggin wrote:
> 
> > > Also, why would you want to make these insane accessors for atomic_t
> > > types? Just make sure everybody knows the basics of barriers, and they
> > > can apply that knowledge to atomic_t and all other lockless memory
> > > accesses as well.
> > 
> > 
> > Code that looks like:
> > 
> > 	while (!atomic_read(&v)) {
> > 		...
> > 		cpu_relax_no_barrier();
> > 		forget(v.counter);
> > 		        ^^^^^^^^
> > 	}
> > 
> > would be uglier. Also think about code such as:
> 
> I think they would both be equally ugly,

You think both these are equivalent in terms of "looks":

					|
while (!atomic_read(&v)) {		|	while (!atomic_read_xxx(&v)) {
	...				|		...
	cpu_relax_no_barrier();		|		cpu_relax_no_barrier();
	order_atomic(&v);		|	}
}					|

(where order_atomic() is an atomic_t
specific wrapper as you mentioned below)

?

Well, taste varies, but ...

> but the atomic_read_volatile
> variant would be more prone to subtle bugs because of the weird
> implementation.

What bugs?

> And it would be more ugly than introducing an order(x) statement for
> all memory operations, and adding an order_atomic() wrapper for it
> for atomic types.

Oh, that order() / forget() macro [forget() was named such by Chuck Ebbert
earlier in this thread where he first mentioned it, btw] could definitely
be generically introduced for any memory operations.

> > 	a = atomic_read();
> > 	if (!a)
> > 		do_something();
> > 
> > 	forget();
> > 	a = atomic_read();
> > 	... /* some code that depends on value of a, obviously */
> > 
> > 	forget();
> > 	a = atomic_read();
> > 	...
> > 
> > So much explicit sprinkling of "forget()" looks ugly.
> 
> Firstly, why is it ugly? It's nice because of those nice explicit
> statements there that give us a good heads up and would have some
> comments attached to them

atomic_read_xxx (where xxx = whatever naming sounds nice to you) would
obviously also give a heads up, and could also have some comments
attached to it.

> (also, lack of the word "volatile" is always a plus).

Ok, xxx != volatile.

> Secondly, what sort of code would do such a thing?

See the nodemgr_host_thread() that does something similar, though not
exactly same.

> > on the other hand, looks neater. The "_volatile()" suffix makes it also
> > no less explicit than an explicit barrier-like macro that this primitive
> > is something "special", for code clarity purposes.
> 
> Just don't use the word volatile,

That sounds amazingly frivolous, but hey, why not. As I said, ok,
xxx != volatile.

> and have barriers both before and after the memory operation,

How could that lead to bugs? (if you can point to existing code,
but just some testcase / sample code would be fine as well).

> [...] I don't see
> the point though, when you could just have a single barrier(x)
> barrier function defined for all memory locations,

As I said, barrier() is too heavy-handed.

> rather than
> this odd thing that only works for atomics

Why would it work only for atomics? You could use that generic macro
for anything you well damn please.

> (and would have to
> be duplicated for atomic_set.

#define atomic_set_xxx for something similar. Big deal ... NOT.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:

>
>On Fri, 17 Aug 2007, Nick Piggin wrote:
>
>>I think they would both be equally ugly,
>>
>
>You think both these are equivalent in terms of "looks":
>
>					|
>while (!atomic_read(&v)) {		|	while (!atomic_read_xxx(&v)) {
>	...				|		...
>	cpu_relax_no_barrier();		|		cpu_relax_no_barrier();
>	order_atomic(&v);		|	}
>}					|
>
>(where order_atomic() is an atomic_t
>specific wrapper as you mentioned below)
>
>?
>

I think the LHS is better if your atomic_read_xxx primitive is using the
crazy one-sided barrier, because the LHS code you immediately know what
barriers are happening, and with the RHS you have to look at the 
atomic_read_xxx
definition.

If your atomic_read_xxx implementation was more intuitive, then both are
pretty well equal. More lines != ugly code.


>>but the atomic_read_volatile
>>variant would be more prone to subtle bugs because of the weird
>>implementation.
>>
>
>What bugs?
>

You can't think for yourself? Your atomic_read_volatile contains a compiler
barrier to the atomic variable before the load. 2 such reads from different
locations look like this:

asm volatile("" : "+m" (v1));
atomic_read(&v1);
asm volatile("" : "+m" (v2));
atomic_read(&v2);

Which implies that the load of v1 can be reordered to occur after the load
of v2. Bet you didn't expect that?

>>Secondly, what sort of code would do such a thing?
>>
>
>See the nodemgr_host_thread() that does something similar, though not
>exactly same.
>

I'm sorry, all this waffling about made up code which might do this and
that is just a waste of time. Seriously, the thread is bloated enough
and never going to get anywhere with all this handwaving. If someone is
saving up all the really real and actually good arguments for why we
must have a volatile here, now is the time to use them.

>>and have barriers both before and after the memory operation,
>>
>
>How could that lead to bugs? (if you can point to existing code,
>but just some testcase / sample code would be fine as well).
>

See above.

>As I said, barrier() is too heavy-handed.
>

Typo. I meant: defined for a single memory location (ie. order(x)).

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> [...]
> > You think both these are equivalent in terms of "looks":
> > 
> > 					|
> > while (!atomic_read(&v)) {		|	while (!atomic_read_xxx(&v)) {
> > 	...				|		...
> > 	cpu_relax_no_barrier();		|
> > cpu_relax_no_barrier();
> > 	order_atomic(&v);		|	}
> > }					|
> > 
> > (where order_atomic() is an atomic_t
> > specific wrapper as you mentioned below)
> > 
> > ?
> 
> I think the LHS is better if your atomic_read_xxx primitive is using the
> crazy one-sided barrier,
  ^^^^^

I'd say it's purposefully one-sided.

> because the LHS code you immediately know what
> barriers are happening, and with the RHS you have to look at the
> atomic_read_xxx
> definition.

No. As I said, the _xxx (whatever the heck you want to name it as) should
give the same heads-up that your "order_atomic" thing is supposed to give.


> If your atomic_read_xxx implementation was more intuitive, then both are
> pretty well equal. More lines != ugly code.
> 
> > [...]
> > What bugs?
> 
> You can't think for yourself? Your atomic_read_volatile contains a compiler
> barrier to the atomic variable before the load. 2 such reads from different
> locations look like this:
> 
> asm volatile("" : "+m" (v1));
> atomic_read(&v1);
> asm volatile("" : "+m" (v2));
> atomic_read(&v2);
> 
> Which implies that the load of v1 can be reordered to occur after the load
> of v2.

And how would that be a bug? (sorry, I really can't think for myself)


> > > Secondly, what sort of code would do such a thing?
> > 
> > See the nodemgr_host_thread() that does something similar, though not
> > exactly same.
> 
> I'm sorry, all this waffling about made up code which might do this and
> that is just a waste of time.

First, you could try looking at the code.

And by the way, as I've already said (why do *require* people to have to
repeat things to you?) this isn't even about only existing code.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Satyam Sharma writes:

> I wonder if this'll generate smaller and better code than _both_ the
> other atomic_read_volatile() variants. Would need to build allyesconfig
> on lots of diff arch's etc to test the theory though.

I'm sure it would be a tiny effect.

This whole thread is arguing about effects that are quite
insignificant.  On the one hand we have the non-volatile proponents,
who want to let the compiler do extra optimizations - which amounts to
letting it elide maybe a dozen loads in the whole kernel, loads which
would almost always be L1 cache hits.

On the other hand we have the volatile proponents, who are concerned
that some code somewhere in the kernel might be buggy without the
volatile behaviour, and who also want to be able to remove some
barriers and thus save a few bytes of code and a few loads here and
there (and possibly some stores too).

Either way the effect on code size and execution time is miniscule.

In the end the strongest argument is actually that gcc generates
unnecessarily verbose code on x86[-64] for volatile accesses.  Even
then we're only talking about ~2000 bytes, or less than 1 byte per
instance of atomic_read on average, about 0.06% of the kernel text
size.

The x86[-64] developers seem to be willing to bear the debugging cost
involved in having the non-volatile behaviour for atomic_read, in
order to save the 2kB.  That's fine with me.  Either way I think
somebody should audit all the uses of atomic_read, not just for
missing barriers, but also to find the places where it's used in a
racy manner.  Then we can work out where the races matter and fix them
if they do.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Paul Mackerras wrote:

> Satyam Sharma writes:
> 
> > I wonder if this'll generate smaller and better code than _both_ the
> > other atomic_read_volatile() variants. Would need to build allyesconfig
> > on lots of diff arch's etc to test the theory though.
> 
> I'm sure it would be a tiny effect.
> 
> This whole thread is arguing about effects that are quite
> insignificant.

Hmm, the fact that this thread became what it did, probably means that
most developers on this list do not mind thinking/arguing about effects
or optimizations that are otherwise "tiny". But yeah, they are tiny
nonetheless.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> #define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))
>
> [ This is exactly equivalent to using "+m" in the constraints, as 
> recently
>   explained on a GCC list somewhere, in response to the patch in my 
> bitops
>   series a few weeks back where I thought "+m" was bogus. ]

[It wasn't explained on a GCC list in response to your patch, as
far as I can see -- if I missed it, please point me to an archived
version of it].

One last time: it isn't equivalent on older (but still supported
by Linux) versions of GCC.  Current versions of GCC allow it, but
it has no documented behaviour at all, so use it at your own risk.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Sat, 18 Aug 2007, Segher Boessenkool wrote:

> > #define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))
> > 
> > [ This is exactly equivalent to using "+m" in the constraints, as recently
> >   explained on a GCC list somewhere, in response to the patch in my bitops
> >   series a few weeks back where I thought "+m" was bogus. ]
> 
> [It wasn't explained on a GCC list in response to your patch, as
> far as I can see -- if I missed it, please point me to an archived
> version of it].

http://gcc.gnu.org/ml/gcc-patches/2007-07/msg01758.html

is a follow-up in the thread on the gcc-patches@gcc.gnu.org mailing list,
which began with:

http://gcc.gnu.org/ml/gcc-patches/2007-07/msg01677.html

that was posted by Jan Kubicka, as he quotes in that initial posting,
after I had submitted:

http://lkml.org/lkml/2007/7/23/252

which was a (wrong) patch to "rectify" what I thought was the "bogus"
"+m" constraint, as per the quoted extract from gcc docs (that was
given in that (wrong) patch's changelog).

That's when _I_ came to know how GCC interprets "+m", but probably
this has been explained on those lists multiple times. Who cares,
anyway?


> One last time: it isn't equivalent on older (but still supported
> by Linux) versions of GCC.  Current versions of GCC allow it, but
> it has no documented behaviour at all, so use it at your own risk.

True.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>> #define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))
>>>
>>> [ This is exactly equivalent to using "+m" in the constraints, as 
>>> recently
>>>   explained on a GCC list somewhere, in response to the patch in my 
>>> bitops
>>>   series a few weeks back where I thought "+m" was bogus. ]
>>
>> [It wasn't explained on a GCC list in response to your patch, as
>> far as I can see -- if I missed it, please point me to an archived
>> version of it].
>
> http://gcc.gnu.org/ml/gcc-patches/2007-07/msg01758.html

Ah yes, that old thread, thank you.

> That's when _I_ came to know how GCC interprets "+m", but probably
> this has been explained on those lists multiple times. Who cares,
> anyway?

I just couldn't find the thread you meant, I thought I missed
have it, that's all :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Geert Uytterhoeven]

On Thu, 16 Aug 2007, Linus Torvalds wrote:
> On Fri, 17 Aug 2007, Paul Mackerras wrote:
> > I'm really surprised it's as much as a few K.  I tried it on powerpc
> > and it only saved 40 bytes (10 instructions) for a G5 config.
> 
> One of the things that "volatile" generally screws up is a simple
> 
> 	volatile int i;
> 
> 	i++;
> 
> which a compiler will generally get horribly, horribly wrong.
> 
> In a reasonable world, gcc should just make that be (on x86)
> 
> 	addl $1,i(%rip)
> 
> on x86-64, which is indeed what it does without the volatile. But with the 
> volatile, the compiler gets really nervous, and doesn't dare do it in one 
> instruction, and thus generates crap like
> 
>         movl    i(%rip), %eax
>         addl    $1, %eax
>         movl    %eax, i(%rip)
> 
> instead. For no good reason, except that "volatile" just doesn't have any 
> good/clear semantics for the compiler, so most compilers will just make it 
> be "I will not touch this access in any way, shape, or form". Including 
> even trivially correct instruction optimization/combination.

Apart from having to fetch more bytes for the instructions (which does
matter), execution time is probably the same on modern processors, as they
convert the single instruction to RISC-style load, modify, store anyway.

Gr{oetje,eeting}s,

						Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
							    -- Linus Torvalds

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Andi Kleen]

On Friday 17 August 2007 05:42, Linus Torvalds wrote:
> On Fri, 17 Aug 2007, Paul Mackerras wrote:
> > I'm really surprised it's as much as a few K.  I tried it on powerpc
> > and it only saved 40 bytes (10 instructions) for a G5 config.
>
> One of the things that "volatile" generally screws up is a simple
>
> 	volatile int i;
>
> 	i++;

But for atomic_t people use atomic_inc() anyways which does this correctly.
It shouldn't really matter for atomic_t.

I'm worrying a bit that the volatile atomic_t change caused subtle code 
breakage like these delay read loops people here pointed out.
Wouldn't it be safer to just re-add the volatile to atomic_read() 
for 2.6.23? Or alternatively make it asm(), but volatile seems more
proven.

-Andi

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Andi Kleen wrote:

> On Friday 17 August 2007 05:42, Linus Torvalds wrote:
> > On Fri, 17 Aug 2007, Paul Mackerras wrote:
> > > I'm really surprised it's as much as a few K.  I tried it on powerpc
> > > and it only saved 40 bytes (10 instructions) for a G5 config.
> >
> > One of the things that "volatile" generally screws up is a simple
> >
> > 	volatile int i;
> >
> > 	i++;
> 
> But for atomic_t people use atomic_inc() anyways which does this correctly.
> It shouldn't really matter for atomic_t.
> 
> I'm worrying a bit that the volatile atomic_t change caused subtle code 
> breakage like these delay read loops people here pointed out.

Umm, I followed most of the thread, but which breakage is this?

> Wouldn't it be safer to just re-add the volatile to atomic_read() 
> for 2.6.23? Or alternatively make it asm(), but volatile seems more
> proven.

The problem with volatile is not just trashy code generation (which also
definitely is a major problem), but definition holes, and implementation
inconsistencies. Making it asm() is not the only other alternative to
volatile either (read another reply to this mail), but considering most
of the thread has been about people not wanting even a
atomic_read_volatile() variant, making atomic_read() itself have volatile
semantics sounds ... strange :-)


PS: http://lkml.org/lkml/2007/8/15/407 was submitted a couple days back,
any word if you saw that?

I have another one for you:


[PATCH] i386, x86_64: __const_udelay() should not be marked inline

Because it can never get inlined in any callsite (each translation unit
is compiled separately for the kernel and so the implementation of
__const_udelay() would be invisible to all other callsites). In fact it
turns out, the correctness of callsites at arch/x86_64/kernel/crash.c:97
and arch/i386/kernel/crash.c:101 explicitly _depends_ upon it not being
inlined, and also it's an exported symbol (modules may want to call
mdelay() and udelay() that often becomes __const_udelay() after some
macro-ing in various headers). So let's not mark it as "inline" either.

Signed-off-by: Satyam Sharma <satyam@infradead.org>

---

 arch/i386/lib/delay.c   |    2 +-
 arch/x86_64/lib/delay.c |    2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/i386/lib/delay.c b/arch/i386/lib/delay.c
index f6edb11..0082c99 100644
--- a/arch/i386/lib/delay.c
+++ b/arch/i386/lib/delay.c
@@ -74,7 +74,7 @@ void __delay(unsigned long loops)
 	delay_fn(loops);
 }
 
-inline void __const_udelay(unsigned long xloops)
+void __const_udelay(unsigned long xloops)
 {
 	int d0;
 
diff --git a/arch/x86_64/lib/delay.c b/arch/x86_64/lib/delay.c
index 2dbebd3..d0cd9cd 100644
--- a/arch/x86_64/lib/delay.c
+++ b/arch/x86_64/lib/delay.c
@@ -38,7 +38,7 @@ void __delay(unsigned long loops)
 }
 EXPORT_SYMBOL(__delay);
 
-inline void __const_udelay(unsigned long xloops)
+void __const_udelay(unsigned long xloops)
 {
 	__delay(((xloops * HZ * cpu_data[raw_smp_processor_id()].loops_per_jiffy) >> 32) + 1);
 }

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> In a reasonable world, gcc should just make that be (on x86)
>
> 	addl $1,i(%rip)
>
> on x86-64, which is indeed what it does without the volatile. But with 
> the
> volatile, the compiler gets really nervous, and doesn't dare do it in 
> one
> instruction, and thus generates crap like
>
>         movl    i(%rip), %eax
>         addl    $1, %eax
>         movl    %eax, i(%rip)
>
> instead. For no good reason, except that "volatile" just doesn't have 
> any
> good/clear semantics for the compiler, so most compilers will just 
> make it
> be "I will not touch this access in any way, shape, or form". Including
> even trivially correct instruction optimization/combination.

It's just a (target-specific, perhaps) missed-optimisation kind
of bug in GCC.  Care to file a bug report?

> but is
> (again) something that gcc doesn't dare do, since "i" is volatile.

Just nobody taught it it can do this; perhaps no one wanted to
add optimisations like that, maybe with a reasoning like "people
who hit the go-slow-in-unspecified-ways button should get what
they deserve" ;-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>>>>> Part of the motivation here is to fix heisenbugs.  If I knew 
>>>>>> where they
>>>>>
>>>>> By the same token we should probably disable optimisations
>>>>> altogether since that too can create heisenbugs.
>>>>
>>>> Almost everything is a tradeoff; and so is this.  I don't
>>>> believe most people would find disabling all compiler
>>>> optimisations an acceptable price to pay for some peace
>>>> of mind.
>>>
>>>
>>> So why is this a good tradeoff?
>> It certainly is better than disabling all compiler optimisations!
>
> It's easy to be better than something really stupid :)

Sure, it wasn't me who made the comparison though.

> So i386 and x86-64 don't have volatiles there, and it saves them a
> few K of kernel text.

Which has to be investigated.  A few kB is a lot more than expected.

> What you need to justify is why it is a good
> tradeoff to make them volatile (which btw, is much harder to go
> the other way after we let people make those assumptions).

My point is that people *already* made those assumptions.  There
are two ways to clean up this mess:

1) Have the "volatile" semantics by default, change the users
    that don't need it;
2) Have "non-volatile" semantics by default, change the users
    that do need it.

Option 2) randomly breaks stuff all over the place, option 1)
doesn't.  Yeah 1) could cause some extremely minor speed or
code size regression, but only temporarily until everything has
been audited.

>>> I also think that just adding things to APIs in the hope it might fix
>>> up some bugs isn't really a good road to go down. Where do you stop?
>> I look at it the other way: keeping the "volatile" semantics in
>> atomic_XXX() (or adding them to it, whatever) helps _prevent_ bugs;
>
> Yeah, but we could add lots of things to help prevent bugs and
> would never be included. I would also contend that it helps _hide_
> bugs and encourages people to be lazy when thinking about these
> things.

Sure.  We aren't _adding_ anything here though, not on the platforms
where it is most likely to show up, anyway.

> Also, you dismiss the fact that we'd actually be *adding* volatile
> semantics back to the 2 most widely tested architectures (in terms
> of test time, number of testers, variety of configurations, and
> coverage of driver code).

I'm not dismissing that.  x86 however is one of the few architectures
where mistakenly leaving out a "volatile" will not easily show up on
user testing, since the compiler will very often produce a memory
reference anyway because it has no registers to play with.

> This is a very important different from
> just keeping volatile semantics because it is basically a one-way
> API change.

That's a good point.  Maybe we should create _two_ new APIs, one
explicitly going each way.

>> certainly most people expect that behaviour, and also that behaviour
>> is *needed* in some places and no other interface provides that
>> functionality.
>
> I don't know that most people would expect that behaviour.

I didn't conduct any formal poll either :-)

> Is there any documentation anywhere that would suggest this?

Not really I think, no.  But not the other way around, either.
Most uses of it seem to expect it though.

>> [some confusion about barriers wrt atomics snipped]
>
> What were you confused about?

Me?  Not much.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 04:38:54AM +0530, Satyam Sharma wrote:
> 
> 
> On Tue, 14 Aug 2007, Christoph Lameter wrote:
> 
> > On Thu, 9 Aug 2007, Chris Snook wrote:
> > 
> > > This patchset makes the behavior of atomic_read uniform by removing the
> > > volatile keyword from all atomic_t and atomic64_t definitions that currently
> > > have it, and instead explicitly casts the variable as volatile in
> > > atomic_read().  This leaves little room for creative optimization by the
> > > compiler, and is in keeping with the principles behind "volatile considered
> > > harmful".
> > 
> > volatile is generally harmful even in atomic_read(). Barriers control
> > visibility and AFAICT things are fine.
> 
> Frankly, I don't see the need for this series myself either. Personal
> opinion (others may differ), but I consider "volatile" to be a sad /
> unfortunate wart in C (numerous threads on this list and on the gcc
> lists/bugzilla over the years stand testimony to this) and if we _can_
> steer clear of it, then why not -- why use this ill-defined primitive
> whose implementation has often differed over compiler versions and
> platforms? Granted, barrier() _is_ heavy-handed in that it makes the
> optimizer forget _everything_, but then somebody did post a forget()
> macro on this thread itself ...
> 
> [ BTW, why do we want the compiler to not optimize atomic_read()'s in
>   the first place? Atomic ops guarantee atomicity, which has nothing
>   to do with "volatility" -- users that expect "volatility" from
>   atomic ops are the ones who must be fixed instead, IMHO. ]

Interactions between mainline code and interrupt/NMI handlers on the same
CPU (for example, when both are using per-CPU variables.  See examples
previously posted in this thread, or look at the rcu_read_lock() and
rcu_read_unlock() implementations in http://lkml.org/lkml/2007/8/7/280.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Satyam Sharma wrote:
> [ BTW, why do we want the compiler to not optimize atomic_read()'s in
>   the first place? Atomic ops guarantee atomicity, which has nothing
>   to do with "volatility" -- users that expect "volatility" from
>   atomic ops are the ones who must be fixed instead, IMHO. ]

LDD3 says on page 125:  "The following operations are defined for the
type [atomic_t] and are guaranteed to be atomic with respect to all
processors of an SMP computer."

Doesn't "atomic WRT all processors" require volatility?
-- 
Stefan Richter
-=====-=-=== =--- -====
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 12:35:31PM +0200, Stefan Richter wrote:
> 
> LDD3 says on page 125:  "The following operations are defined for the
> type [atomic_t] and are guaranteed to be atomic with respect to all
> processors of an SMP computer."
> 
> Doesn't "atomic WRT all processors" require volatility?

Not at all.  We also require this to be atomic without any
hint of volatility.

	extern int foo;
	foo = 4;

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Stefan Richter wrote:

> Satyam Sharma wrote:
> > [ BTW, why do we want the compiler to not optimize atomic_read()'s in
> >   the first place? Atomic ops guarantee atomicity, which has nothing
> >   to do with "volatility" -- users that expect "volatility" from
> >   atomic ops are the ones who must be fixed instead, IMHO. ]
> 
> LDD3 says on page 125:  "The following operations are defined for the
> type [atomic_t] and are guaranteed to be atomic with respect to all
> processors of an SMP computer."
> 
> Doesn't "atomic WRT all processors" require volatility?

No, it definitely doesn't. Why should it?

"Atomic w.r.t. all processors" is just your normal, simple "atomicity"
for SMP systems (ensure that that object is modified / set / replaced
in main memory atomically) and has nothing to do with "volatile"
behaviour.

"Volatile behaviour" itself isn't consistently defined (at least
definitely not consistently implemented in various gcc versions across
platforms), but it is /expected/ to mean something like: "ensure that
every such access actually goes all the way to memory, and is not
re-ordered w.r.t. to other accesses, as far as the compiler can take
care of these". The last "as far as compiler can take care" disclaimer
comes about due to CPUs doing their own re-ordering nowadays.

For example (say on i386):

(A)
$ cat tp1.c
int a;

void func(void)
{
	a = 10;
	a = 20;
}
$ gcc -Os -S tp1.c
$ cat tp1.s
...
movl    $20, a
...

(B)
$ cat tp2.c
volatile int a;

void func(void)
{
	a = 10;
	a = 20;
}
$ gcc -Os -S tp2.c
$ cat tp2.s
...
movl    $10, a
movl    $20, a
...

(C)
$ cat tp3.c
int a;

void func(void)
{
	*(volatile int *)&a = 10;
	*(volatile int *)&a = 20;
}
$ gcc -Os -S tp3.c
$ cat tp3.s
...
movl    $10, a
movl    $20, a
...

In (A) the compiler optimized "a = 10;" away, but the actual store
of the final value "20" to "a" was still "atomic". (B) and (C) also
exhibit "volatile" behaviour apart from the "atomicity".

But as others replied, it seems some callers out there depend upon
atomic ops exhibiting "volatile" behaviour as well, so that answers
my initial question, actually. I haven't looked at the code Paul
pointed me at, but I wonder if that "forget(x)" macro would help
those cases. I'd wish to avoid the "volatile" primitive, personally.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Satyam Sharma wrote:
> On Wed, 15 Aug 2007, Stefan Richter wrote:
>> Doesn't "atomic WRT all processors" require volatility?
> 
> No, it definitely doesn't. Why should it?
> 
> "Atomic w.r.t. all processors" is just your normal, simple "atomicity"
> for SMP systems (ensure that that object is modified / set / replaced
> in main memory atomically) and has nothing to do with "volatile"
> behaviour.
> 
> "Volatile behaviour" itself isn't consistently defined (at least
> definitely not consistently implemented in various gcc versions across
> platforms), but it is /expected/ to mean something like: "ensure that
> every such access actually goes all the way to memory, and is not
> re-ordered w.r.t. to other accesses, as far as the compiler can take
> care of these". The last "as far as compiler can take care" disclaimer
> comes about due to CPUs doing their own re-ordering nowadays.
> 
> For example (say on i386):

[...]

> In (A) the compiler optimized "a = 10;" away, but the actual store
> of the final value "20" to "a" was still "atomic". (B) and (C) also
> exhibit "volatile" behaviour apart from the "atomicity".
> 
> But as others replied, it seems some callers out there depend upon
> atomic ops exhibiting "volatile" behaviour as well, so that answers
> my initial question, actually. I haven't looked at the code Paul
> pointed me at, but I wonder if that "forget(x)" macro would help
> those cases. I'd wish to avoid the "volatile" primitive, personally.

So, looking at load instead of store, understand I correctly that in
your opinion

	int b;

	b = atomic_read(&a);
	if (b)
		do_something_time_consuming();

	b = atomic_read(&a);
	if (b)
		do_something_more();

should be changed to explicitly forget(&a) after
do_something_time_consuming?

If so, how about the following:

static inline void A(atomic_t *a)
{
	int b = atomic_read(a);
	if (b)
		do_something_time_consuming();
}

static inline void B(atomic_t *a)
{
	int b = atomic_read(a);
	if (b)
		do_something_more();
}

static void C(atomic_t *a)
{
	A(a);
	B(b);
}

Would this need forget(a) after A(a)?

(Is the latter actually answered in C99 or is it compiler-dependent?)
-- 
Stefan Richter
-=====-=-=== =--- -====
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

I wrote:
> static inline void A(atomic_t *a)
> {
> 	int b = atomic_read(a);
> 	if (b)
> 		do_something_time_consuming();
> }
> 
> static inline void B(atomic_t *a)
> {
> 	int b = atomic_read(a);
> 	if (b)
> 		do_something_more();
> }
> 
> static void C(atomic_t *a)
> {
> 	A(a);
> 	B(b);
	/* ^ typo */
	B(a);
> }
> 
> Would this need forget(a) after A(a)?
> 
> (Is the latter actually answered in C99 or is it compiler-dependent?)


-- 
Stefan Richter
-=====-=-=== =--- -====
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Stefan Richter wrote:

> Satyam Sharma wrote:
> > On Wed, 15 Aug 2007, Stefan Richter wrote:
> >> Doesn't "atomic WRT all processors" require volatility?
> > 
> > No, it definitely doesn't. Why should it?
> > 
> > "Atomic w.r.t. all processors" is just your normal, simple "atomicity"
> > for SMP systems (ensure that that object is modified / set / replaced
> > in main memory atomically) and has nothing to do with "volatile"
> > behaviour.
> > 
> > "Volatile behaviour" itself isn't consistently defined (at least
> > definitely not consistently implemented in various gcc versions across
> > platforms), but it is /expected/ to mean something like: "ensure that
> > every such access actually goes all the way to memory, and is not
> > re-ordered w.r.t. to other accesses, as far as the compiler can take
> > care of these". The last "as far as compiler can take care" disclaimer
> > comes about due to CPUs doing their own re-ordering nowadays.
> > 
> > For example (say on i386):
> 
> [...]
> 
> > In (A) the compiler optimized "a = 10;" away, but the actual store
> > of the final value "20" to "a" was still "atomic". (B) and (C) also
> > exhibit "volatile" behaviour apart from the "atomicity".
> > 
> > But as others replied, it seems some callers out there depend upon
> > atomic ops exhibiting "volatile" behaviour as well, so that answers
> > my initial question, actually. I haven't looked at the code Paul
> > pointed me at, but I wonder if that "forget(x)" macro would help
> > those cases. I'd wish to avoid the "volatile" primitive, personally.
> 
> So, looking at load instead of store, understand I correctly that in
> your opinion
> 
> 	int b;
> 
> 	b = atomic_read(&a);
> 	if (b)
> 		do_something_time_consuming();
> 
> 	b = atomic_read(&a);
> 	if (b)
> 		do_something_more();
> 
> should be changed to explicitly forget(&a) after
> do_something_time_consuming?

No, I'd actually prefer something like what Christoph Lameter suggested,
i.e. users (such as above) who want "volatile"-like behaviour from atomic
ops can use alternative functions. How about something like:

#define atomic_read_volatile(v)			\
	({					\
		forget(&(v)->counter);		\
		((v)->counter);			\
	})

Or possibly, implement these "volatile" atomic ops variants in inline asm
like the patch that Sebastian Siewior has submitted on another thread just
a while back.

Of course, if we find there are more callers in the kernel who want the
volatility behaviour than those who don't care, we can re-define the
existing ops to such variants, and re-name the existing definitions to
somethine else, say "atomic_read_nonvolatile" for all I care.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 07:17:29PM +0530, Satyam Sharma wrote:
> On Wed, 15 Aug 2007, Stefan Richter wrote:
> > Satyam Sharma wrote:
> > > On Wed, 15 Aug 2007, Stefan Richter wrote:
> > >> Doesn't "atomic WRT all processors" require volatility?
> > > 
> > > No, it definitely doesn't. Why should it?
> > > 
> > > "Atomic w.r.t. all processors" is just your normal, simple "atomicity"
> > > for SMP systems (ensure that that object is modified / set / replaced
> > > in main memory atomically) and has nothing to do with "volatile"
> > > behaviour.
> > > 
> > > "Volatile behaviour" itself isn't consistently defined (at least
> > > definitely not consistently implemented in various gcc versions across
> > > platforms), but it is /expected/ to mean something like: "ensure that
> > > every such access actually goes all the way to memory, and is not
> > > re-ordered w.r.t. to other accesses, as far as the compiler can take
> > > care of these". The last "as far as compiler can take care" disclaimer
> > > comes about due to CPUs doing their own re-ordering nowadays.
> > > 
> > > For example (say on i386):
> > 
> > [...]
> > 
> > > In (A) the compiler optimized "a = 10;" away, but the actual store
> > > of the final value "20" to "a" was still "atomic". (B) and (C) also
> > > exhibit "volatile" behaviour apart from the "atomicity".
> > > 
> > > But as others replied, it seems some callers out there depend upon
> > > atomic ops exhibiting "volatile" behaviour as well, so that answers
> > > my initial question, actually. I haven't looked at the code Paul
> > > pointed me at, but I wonder if that "forget(x)" macro would help
> > > those cases. I'd wish to avoid the "volatile" primitive, personally.
> > 
> > So, looking at load instead of store, understand I correctly that in
> > your opinion
> > 
> > 	int b;
> > 
> > 	b = atomic_read(&a);
> > 	if (b)
> > 		do_something_time_consuming();
> > 
> > 	b = atomic_read(&a);
> > 	if (b)
> > 		do_something_more();
> > 
> > should be changed to explicitly forget(&a) after
> > do_something_time_consuming?
> 
> No, I'd actually prefer something like what Christoph Lameter suggested,
> i.e. users (such as above) who want "volatile"-like behaviour from atomic
> ops can use alternative functions. How about something like:
> 
> #define atomic_read_volatile(v)			\
> 	({					\
> 		forget(&(v)->counter);		\
> 		((v)->counter);			\
> 	})

Wouldn't the above "forget" the value, throw it away, then forget
that it forgot it, giving non-volatile semantics?

> Or possibly, implement these "volatile" atomic ops variants in inline asm
> like the patch that Sebastian Siewior has submitted on another thread just
> a while back.

Given that you are advocating a change (please keep in mind that
atomic_read() and atomic_set() had volatile semantics on almost all
platforms), care to give some example where these historical volatile
semantics are causing a problem?

> Of course, if we find there are more callers in the kernel who want the
> volatility behaviour than those who don't care, we can re-define the
> existing ops to such variants, and re-name the existing definitions to
> somethine else, say "atomic_read_nonvolatile" for all I care.

Do we really need another set of APIs?  Can you give even one example
where the pre-existing volatile semantics are causing enough of a problem
to justify adding yet more atomic_*() APIs?

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 07:25:16AM -0700, Paul E. McKenney wrote:
> 
> Do we really need another set of APIs?  Can you give even one example
> where the pre-existing volatile semantics are causing enough of a problem
> to justify adding yet more atomic_*() APIs?

Let's turn this around.  Can you give a single example where
the volatile semantics is needed in a legitimate way?

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 11:33:36PM +0800, Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 07:25:16AM -0700, Paul E. McKenney wrote:
> > 
> > Do we really need another set of APIs?  Can you give even one example
> > where the pre-existing volatile semantics are causing enough of a problem
> > to justify adding yet more atomic_*() APIs?
> 
> Let's turn this around.  Can you give a single example where
> the volatile semantics is needed in a legitimate way?

Sorry, but you are the one advocating for the change.

Nice try, though!  ;-)

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> On Wed, Aug 15, 2007 at 11:33:36PM +0800, Herbert Xu wrote:
> > On Wed, Aug 15, 2007 at 07:25:16AM -0700, Paul E. McKenney wrote:
> > > 
> > > Do we really need another set of APIs?  Can you give even one example
> > > where the pre-existing volatile semantics are causing enough of a problem
> > > to justify adding yet more atomic_*() APIs?
> > 
> > Let's turn this around.  Can you give a single example where
> > the volatile semantics is needed in a legitimate way?
> 
> Sorry, but you are the one advocating for the change.

Not for i386 and x86_64 -- those have atomic ops without any "volatile"
semantics (currently as per existing definitions).

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 10:48:28PM +0530, Satyam Sharma wrote:
> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> > On Wed, Aug 15, 2007 at 11:33:36PM +0800, Herbert Xu wrote:
> > > On Wed, Aug 15, 2007 at 07:25:16AM -0700, Paul E. McKenney wrote:
> > > > 
> > > > Do we really need another set of APIs?  Can you give even one example
> > > > where the pre-existing volatile semantics are causing enough of a problem
> > > > to justify adding yet more atomic_*() APIs?
> > > 
> > > Let's turn this around.  Can you give a single example where
> > > the volatile semantics is needed in a legitimate way?
> > 
> > Sorry, but you are the one advocating for the change.
> 
> Not for i386 and x86_64 -- those have atomic ops without any "volatile"
> semantics (currently as per existing definitions).

I claim unit volumes with arm, and the majority of the architectures, but
I cannot deny the popularity of i386 and x86_64 with many developers.  ;-)

However, I am not aware of code in the kernel that would benefit
from the compiler coalescing multiple atomic_set() and atomic_read()
invocations, thus I don't see the downside to volatility in this case.
Are there some performance-critical code fragments that I am missing?

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> On Wed, Aug 15, 2007 at 10:48:28PM +0530, Satyam Sharma wrote:
> > [...]
> > Not for i386 and x86_64 -- those have atomic ops without any "volatile"
> > semantics (currently as per existing definitions).
> 
> I claim unit volumes with arm, and the majority of the architectures, but
> I cannot deny the popularity of i386 and x86_64 with many developers.  ;-)

Hmm, does arm really need that "volatile int counter;"? Hopefully RMK will
take a patch removing that "volatile" ... ;-)


> However, I am not aware of code in the kernel that would benefit
> from the compiler coalescing multiple atomic_set() and atomic_read()
> invocations, thus I don't see the downside to volatility in this case.
> Are there some performance-critical code fragments that I am missing?

I don't know, and yes, code with multiple atomic_set's and atomic_read's
getting optimized or coalesced does sound strange to start with. Anyway,
I'm not against "volatile semantics" per se. As replied elsewhere, I do
appreciate the motivation behind this series (to _avoid_ gotchas, not to
fix existing ones). Just that I'd like to avoid using "volatile", for
aforementioned reasons, especially given that there are perfectly
reasonable alternatives to achieve the same desired behaviour.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

Hi Paul,


On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> On Wed, Aug 15, 2007 at 07:17:29PM +0530, Satyam Sharma wrote:
> > [...]
> > No, I'd actually prefer something like what Christoph Lameter suggested,
> > i.e. users (such as above) who want "volatile"-like behaviour from atomic
> > ops can use alternative functions. How about something like:
> > 
> > #define atomic_read_volatile(v)			\
> > 	({					\
> > 		forget(&(v)->counter);		\
> > 		((v)->counter);			\
> > 	})
> 
> Wouldn't the above "forget" the value, throw it away, then forget
> that it forgot it, giving non-volatile semantics?

Nope, I don't think so. I wrote the following trivial testcases:
[ See especially tp4.c and tp4.s (last example). ]

==============================================================================
$ cat tp1.c # Using volatile access casts

#define atomic_read(a)	(*(volatile int *)&a)

int a;

void func(void)
{
	a = 0;
	while (atomic_read(a))
		;
}
==============================================================================
$ gcc -Os -S tp1.c; cat tp1.s

func:
	pushl	%ebp
	movl	%esp, %ebp
	movl	$0, a
.L2:
	movl	a, %eax
	testl	%eax, %eax
	jne	.L2
	popl	%ebp
	ret
==============================================================================
$ cat tp2.c # Using nothing; gcc will optimize the whole loop away

#define forget(x)

#define atomic_read(a)		\
	({			\
		forget(&(a));	\
		(a);		\
	})

int a;

void func(void)
{
	a = 0;
	while (atomic_read(a))
		;
}
==============================================================================
$ gcc -Os -S tp2.c; cat tp2.s

func:
	pushl	%ebp
	movl	%esp, %ebp
	popl	%ebp
	movl	$0, a
	ret
==============================================================================
$ cat tp3.c # Using a full memory clobber barrier

#define forget(x)	asm volatile ("":::"memory")

#define atomic_read(a)		\
	({			\
		forget(&(a));	\
		(a);		\
	})

int a;

void func(void)
{
	a = 0;
	while (atomic_read(a))
		;
}
==============================================================================
$ gcc -Os -S tp3.c; cat tp3.s

func:
	pushl	%ebp
	movl	%esp, %ebp
	movl	$0, a
.L2:
	cmpl	$0, a
	jne	.L2
	popl	%ebp
	ret
==============================================================================
$ cat tp4.c # Using a forget(var) macro

#define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))

#define atomic_read(a)		\
	({			\
		forget(a);	\
		(a);		\
	})

int a;

void func(void)
{
	a = 0;
	while (atomic_read(a))
		;
}
==============================================================================
$ gcc -Os -S tp4.c; cat tp4.s

func:
	pushl	%ebp
	movl	%esp, %ebp
	movl	$0, a
.L2:
	cmpl	$0, a
	jne	.L2
	popl	%ebp
	ret
==============================================================================


Possibly these were too trivial to expose any potential problems that you
may have been referring to, so would be helpful if you could write a more
concrete example / sample code.


> > Or possibly, implement these "volatile" atomic ops variants in inline asm
> > like the patch that Sebastian Siewior has submitted on another thread just
> > a while back.
> 
> Given that you are advocating a change (please keep in mind that
> atomic_read() and atomic_set() had volatile semantics on almost all
> platforms), care to give some example where these historical volatile
> semantics are causing a problem?
> [...]
> Can you give even one example
> where the pre-existing volatile semantics are causing enough of a problem
> to justify adding yet more atomic_*() APIs?

Will take this to the other sub-thread ...


> > Of course, if we find there are more callers in the kernel who want the
> > volatility behaviour than those who don't care, we can re-define the
> > existing ops to such variants, and re-name the existing definitions to
> > somethine else, say "atomic_read_nonvolatile" for all I care.
> 
> Do we really need another set of APIs?

Well, if there's one set of users who do care about volatile behaviour,
and another set that doesn't, it only sounds correct to provide both
those APIs, instead of forcing one behaviour on all users.


Thanks,
Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 11:25:05PM +0530, Satyam Sharma wrote:
> Hi Paul,
> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> 
> > On Wed, Aug 15, 2007 at 07:17:29PM +0530, Satyam Sharma wrote:
> > > [...]
> > > No, I'd actually prefer something like what Christoph Lameter suggested,
> > > i.e. users (such as above) who want "volatile"-like behaviour from atomic
> > > ops can use alternative functions. How about something like:
> > > 
> > > #define atomic_read_volatile(v)			\
> > > 	({					\
> > > 		forget(&(v)->counter);		\
> > > 		((v)->counter);			\
> > > 	})
> > 
> > Wouldn't the above "forget" the value, throw it away, then forget
> > that it forgot it, giving non-volatile semantics?
> 
> Nope, I don't think so. I wrote the following trivial testcases:
> [ See especially tp4.c and tp4.s (last example). ]

Right.  I should have said "wouldn't the compiler be within its rights
to forget the value, throw it away, then forget that it forgot it".
The value coming out of the #define above is an unadorned ((v)->counter),
which has no volatile semantics.

> ==============================================================================
> $ cat tp1.c # Using volatile access casts
> 
> #define atomic_read(a)	(*(volatile int *)&a)
> 
> int a;
> 
> void func(void)
> {
> 	a = 0;
> 	while (atomic_read(a))
> 		;
> }
> ==============================================================================
> $ gcc -Os -S tp1.c; cat tp1.s
> 
> func:
> 	pushl	%ebp
> 	movl	%esp, %ebp
> 	movl	$0, a
> .L2:
> 	movl	a, %eax
> 	testl	%eax, %eax
> 	jne	.L2
> 	popl	%ebp
> 	ret
> ==============================================================================
> $ cat tp2.c # Using nothing; gcc will optimize the whole loop away
> 
> #define forget(x)
> 
> #define atomic_read(a)		\
> 	({			\
> 		forget(&(a));	\
> 		(a);		\
> 	})
> 
> int a;
> 
> void func(void)
> {
> 	a = 0;
> 	while (atomic_read(a))
> 		;
> }
> ==============================================================================
> $ gcc -Os -S tp2.c; cat tp2.s
> 
> func:
> 	pushl	%ebp
> 	movl	%esp, %ebp
> 	popl	%ebp
> 	movl	$0, a
> 	ret
> ==============================================================================
> $ cat tp3.c # Using a full memory clobber barrier
> 
> #define forget(x)	asm volatile ("":::"memory")
> 
> #define atomic_read(a)		\
> 	({			\
> 		forget(&(a));	\
> 		(a);		\
> 	})
> 
> int a;
> 
> void func(void)
> {
> 	a = 0;
> 	while (atomic_read(a))
> 		;
> }
> ==============================================================================
> $ gcc -Os -S tp3.c; cat tp3.s
> 
> func:
> 	pushl	%ebp
> 	movl	%esp, %ebp
> 	movl	$0, a
> .L2:
> 	cmpl	$0, a
> 	jne	.L2
> 	popl	%ebp
> 	ret
> ==============================================================================
> $ cat tp4.c # Using a forget(var) macro
> 
> #define forget(a)	__asm__ __volatile__ ("" :"=m" (a) :"m" (a))
> 
> #define atomic_read(a)		\
> 	({			\
> 		forget(a);	\
> 		(a);		\
> 	})
> 
> int a;
> 
> void func(void)
> {
> 	a = 0;
> 	while (atomic_read(a))
> 		;
> }
> ==============================================================================
> $ gcc -Os -S tp4.c; cat tp4.s
> 
> func:
> 	pushl	%ebp
> 	movl	%esp, %ebp
> 	movl	$0, a
> .L2:
> 	cmpl	$0, a
> 	jne	.L2
> 	popl	%ebp
> 	ret
> ==============================================================================
> 
> Possibly these were too trivial to expose any potential problems that you
> may have been referring to, so would be helpful if you could write a more
> concrete example / sample code.

The trick is to have a sufficiently complicated expression to force
the compiler to run out of registers.  If the value is non-volatile,
it will refetch it (and expect it not to have changed, possibly being
disappointed by an interrupt handler running on that same CPU).

> > > Or possibly, implement these "volatile" atomic ops variants in inline asm
> > > like the patch that Sebastian Siewior has submitted on another thread just
> > > a while back.
> > 
> > Given that you are advocating a change (please keep in mind that
> > atomic_read() and atomic_set() had volatile semantics on almost all
> > platforms), care to give some example where these historical volatile
> > semantics are causing a problem?
> > [...]
> > Can you give even one example
> > where the pre-existing volatile semantics are causing enough of a problem
> > to justify adding yet more atomic_*() APIs?
> 
> Will take this to the other sub-thread ...

OK.

> > > Of course, if we find there are more callers in the kernel who want the
> > > volatility behaviour than those who don't care, we can re-define the
> > > existing ops to such variants, and re-name the existing definitions to
> > > somethine else, say "atomic_read_nonvolatile" for all I care.
> > 
> > Do we really need another set of APIs?
> 
> Well, if there's one set of users who do care about volatile behaviour,
> and another set that doesn't, it only sounds correct to provide both
> those APIs, instead of forcing one behaviour on all users.

Well, if the second set doesn't care, they should be OK with the volatile
behavior in this case.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> Possibly these were too trivial to expose any potential problems that 
>> you
>> may have been referring to, so would be helpful if you could write a 
>> more
>> concrete example / sample code.
>
> The trick is to have a sufficiently complicated expression to force
> the compiler to run out of registers.

You can use -ffixed-XXX to keep the testcase simple.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>> Of course, if we find there are more callers in the kernel who want 
>>> the
>>> volatility behaviour than those who don't care, we can re-define the
>>> existing ops to such variants, and re-name the existing definitions 
>>> to
>>> somethine else, say "atomic_read_nonvolatile" for all I care.
>>
>> Do we really need another set of APIs?
>
> Well, if there's one set of users who do care about volatile behaviour,
> and another set that doesn't, it only sounds correct to provide both
> those APIs, instead of forcing one behaviour on all users.

But since there currently is only one such API, and there are
users expecting the stronger behaviour, the only sane thing to
do is let the API provide that behaviour.  You can always add
a new API with weaker behaviour later, and move users that are
okay with it over to that new API.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[David Howells]

Herbert Xu <herbert@gondor.apana.org.au> wrote:

> Let's turn this around.  Can you give a single example where
> the volatile semantics is needed in a legitimate way?

Accessing H/W registers?  But apart from that...

David

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 07:19:57PM +0100, David Howells wrote:
> Herbert Xu <herbert@gondor.apana.org.au> wrote:
> 
> > Let's turn this around.  Can you give a single example where
> > the volatile semantics is needed in a legitimate way?
> 
> Accessing H/W registers?  But apart from that...

Communicating between process context and interrupt/NMI handlers using
per-CPU variables.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 11:45:20AM -0700, Paul E. McKenney wrote:
> On Wed, Aug 15, 2007 at 07:19:57PM +0100, David Howells wrote:
> > Herbert Xu <herbert@gondor.apana.org.au> wrote:
> > 
> > > Let's turn this around.  Can you give a single example where
> > > the volatile semantics is needed in a legitimate way?
> > 
> > Accessing H/W registers?  But apart from that...
> 
> Communicating between process context and interrupt/NMI handlers using
> per-CPU variables.

Remeber we're talking about atomic_read/atomic_set.  Please
cite the actual file/function name you have in mind.

Thanks,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 07:41:46AM +0800, Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 11:45:20AM -0700, Paul E. McKenney wrote:
> > On Wed, Aug 15, 2007 at 07:19:57PM +0100, David Howells wrote:
> > > Herbert Xu <herbert@gondor.apana.org.au> wrote:
> > > 
> > > > Let's turn this around.  Can you give a single example where
> > > > the volatile semantics is needed in a legitimate way?
> > > 
> > > Accessing H/W registers?  But apart from that...
> > 
> > Communicating between process context and interrupt/NMI handlers using
> > per-CPU variables.
> 
> Remeber we're talking about atomic_read/atomic_set.  Please
> cite the actual file/function name you have in mind.

Yep, we are indeed talking about atomic_read()/atomic_set().

We have been through this issue already in this thread.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 04:53:35PM -0700, Paul E. McKenney wrote:
>
> > > Communicating between process context and interrupt/NMI handlers using
> > > per-CPU variables.
> > 
> > Remeber we're talking about atomic_read/atomic_set.  Please
> > cite the actual file/function name you have in mind.
> 
> Yep, we are indeed talking about atomic_read()/atomic_set().
> 
> We have been through this issue already in this thread.

Sorry, but I must've missed it.  Could you cite the file or
function for my benefit?

Thanks,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 08:12:48AM +0800, Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 04:53:35PM -0700, Paul E. McKenney wrote:
> >
> > > > Communicating between process context and interrupt/NMI handlers using
> > > > per-CPU variables.
> > > 
> > > Remeber we're talking about atomic_read/atomic_set.  Please
> > > cite the actual file/function name you have in mind.
> > 
> > Yep, we are indeed talking about atomic_read()/atomic_set().
> > 
> > We have been through this issue already in this thread.
> 
> Sorry, but I must've missed it.  Could you cite the file or
> function for my benefit?

I might summarize the thread if there is interest, but I am not able to
do so right this minute.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 05:23:10PM -0700, Paul E. McKenney wrote:
> On Thu, Aug 16, 2007 at 08:12:48AM +0800, Herbert Xu wrote:
> > On Wed, Aug 15, 2007 at 04:53:35PM -0700, Paul E. McKenney wrote:
> > >
> > > > > Communicating between process context and interrupt/NMI handlers using
> > > > > per-CPU variables.
> > > > 
> > > > Remeber we're talking about atomic_read/atomic_set.  Please
> > > > cite the actual file/function name you have in mind.
> > > 
> > > Yep, we are indeed talking about atomic_read()/atomic_set().
> > > 
> > > We have been through this issue already in this thread.
> > 
> > Sorry, but I must've missed it.  Could you cite the file or
> > function for my benefit?
> 
> I might summarize the thread if there is interest, but I am not able to
> do so right this minute.

Thanks.  But I don't need a summary of the thread, I'm asking
for an extant code snippet in our kernel that benefits from
the volatile change and is not part of a busy-wait.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 08:30:23AM +0800, Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 05:23:10PM -0700, Paul E. McKenney wrote:
> > On Thu, Aug 16, 2007 at 08:12:48AM +0800, Herbert Xu wrote:
> > > On Wed, Aug 15, 2007 at 04:53:35PM -0700, Paul E. McKenney wrote:
> > > >
> > > > > > Communicating between process context and interrupt/NMI handlers using
> > > > > > per-CPU variables.
> > > > > 
> > > > > Remeber we're talking about atomic_read/atomic_set.  Please
> > > > > cite the actual file/function name you have in mind.
> > > > 
> > > > Yep, we are indeed talking about atomic_read()/atomic_set().
> > > > 
> > > > We have been through this issue already in this thread.
> > > 
> > > Sorry, but I must've missed it.  Could you cite the file or
> > > function for my benefit?
> > 
> > I might summarize the thread if there is interest, but I am not able to
> > do so right this minute.
> 
> Thanks.  But I don't need a summary of the thread, I'm asking
> for an extant code snippet in our kernel that benefits from
> the volatile change and is not part of a busy-wait.

Sorry, can't help you there.  I really do believe that the information
you need (as opposed to the specific item you are asking for) really
has been put forth in this thread.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Wed, Aug 15, 2007 at 05:49:50PM -0700, Paul E. McKenney wrote:
> On Thu, Aug 16, 2007 at 08:30:23AM +0800, Herbert Xu wrote:
>
> > Thanks.  But I don't need a summary of the thread, I'm asking
> > for an extant code snippet in our kernel that benefits from
> > the volatile change and is not part of a busy-wait.
> 
> Sorry, can't help you there.  I really do believe that the information
> you need (as opposed to the specific item you are asking for) really
> has been put forth in this thread.

That only leads me to believe that such a code snippet simply
does not exist.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 08:53:16AM +0800, Herbert Xu wrote:
> On Wed, Aug 15, 2007 at 05:49:50PM -0700, Paul E. McKenney wrote:
> > On Thu, Aug 16, 2007 at 08:30:23AM +0800, Herbert Xu wrote:
> >
> > > Thanks.  But I don't need a summary of the thread, I'm asking
> > > for an extant code snippet in our kernel that benefits from
> > > the volatile change and is not part of a busy-wait.
> > 
> > Sorry, can't help you there.  I really do believe that the information
> > you need (as opposed to the specific item you are asking for) really
> > has been put forth in this thread.
> 
> That only leads me to believe that such a code snippet simply
> does not exist.

Whatever...

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> "Volatile behaviour" itself isn't consistently defined (at least
> definitely not consistently implemented in various gcc versions across
> platforms),

It should be consistent across platforms; if not, file a bug please.

> but it is /expected/ to mean something like: "ensure that
> every such access actually goes all the way to memory, and is not
> re-ordered w.r.t. to other accesses, as far as the compiler can take
> care of these". The last "as far as compiler can take care" disclaimer
> comes about due to CPUs doing their own re-ordering nowadays.

You can *expect* whatever you want, but this isn't in line with
reality at all.

volatile _does not_ make accesses go all the way to memory.
volatile _does not_ prevent reordering wrt other accesses.

What volatile does are a) never optimise away a read (or write)
to the object, since the data can change in ways the compiler
cannot see; and b) never move stores to the object across a
sequence point.  This does not mean other accesses cannot be
reordered wrt the volatile access.

If the abstract machine would do an access to a volatile-
qualified object, the generated machine code will do that
access too.  But, for example, it can still be optimised
away by the compiler, if it can prove it is allowed to.

If you want stuff to go all the way to memory, you need some
architecture-specific flush sequence; to make a store globally
visible before another store, you need mb(); before some following
read, you need mb(); to prevent reordering you need a barrier.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Segher Boessenkool wrote:

> > "Volatile behaviour" itself isn't consistently defined (at least
> > definitely not consistently implemented in various gcc versions across
> > platforms),
> 
> It should be consistent across platforms; if not, file a bug please.
> 
> > but it is /expected/ to mean something like: "ensure that
> > every such access actually goes all the way to memory, and is not
> > re-ordered w.r.t. to other accesses, as far as the compiler can take
                              ^
                              (volatile)

(or, alternatively, "other accesses to the same volatile object" ...)

> > care of these". The last "as far as compiler can take care" disclaimer
> > comes about due to CPUs doing their own re-ordering nowadays.
> 
> You can *expect* whatever you want, but this isn't in line with
> reality at all.
> 
> volatile _does not_ prevent reordering wrt other accesses.
> [...]
> What volatile does are a) never optimise away a read (or write)
> to the object, since the data can change in ways the compiler
> cannot see; and b) never move stores to the object across a
> sequence point.  This does not mean other accesses cannot be
> reordered wrt the volatile access.
> 
> If the abstract machine would do an access to a volatile-
> qualified object, the generated machine code will do that
> access too.  But, for example, it can still be optimised
> away by the compiler, if it can prove it is allowed to.

As (now) indicated above, I had meant multiple volatile accesses to
the same object, obviously.

BTW:

#define atomic_read(a)	(*(volatile int *)&(a))
#define atomic_set(a,i)	(*(volatile int *)&(a) = (i))

int a;

void func(void)
{
	int b;

	b = atomic_read(a);
	atomic_set(a, 20);
	b = atomic_read(a);
}

gives:

func:
	pushl	%ebp
	movl	a, %eax
	movl	%esp, %ebp
	movl	$20, a
	movl	a, %eax
	popl	%ebp
	ret

so the first atomic_read() wasn't optimized away.


> volatile _does not_ make accesses go all the way to memory.
> [...]
> If you want stuff to go all the way to memory, you need some
> architecture-specific flush sequence; to make a store globally
> visible before another store, you need mb(); before some following
> read, you need mb(); to prevent reordering you need a barrier.

Sure, which explains the "as far as the compiler can take care" bit.
Poor phrase / choice of words, probably.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> What volatile does are a) never optimise away a read (or write)
>> to the object, since the data can change in ways the compiler
>> cannot see; and b) never move stores to the object across a
>> sequence point.  This does not mean other accesses cannot be
>> reordered wrt the volatile access.
>>
>> If the abstract machine would do an access to a volatile-
>> qualified object, the generated machine code will do that
>> access too.  But, for example, it can still be optimised
>> away by the compiler, if it can prove it is allowed to.
>
> As (now) indicated above, I had meant multiple volatile accesses to
> the same object, obviously.

Yes, accesses to volatile objects are never reordered with
respect to each other.

> BTW:
>
> #define atomic_read(a)	(*(volatile int *)&(a))
> #define atomic_set(a,i)	(*(volatile int *)&(a) = (i))
>
> int a;
>
> void func(void)
> {
> 	int b;
>
> 	b = atomic_read(a);
> 	atomic_set(a, 20);
> 	b = atomic_read(a);
> }
>
> gives:
>
> func:
> 	pushl	%ebp
> 	movl	a, %eax
> 	movl	%esp, %ebp
> 	movl	$20, a
> 	movl	a, %eax
> 	popl	%ebp
> 	ret
>
> so the first atomic_read() wasn't optimized away.

Of course.  It is executed by the abstract machine, so
it will be executed by the actual machine.  On the other
hand, try

	b = 0;
	if (b)
		b = atomic_read(a);

or similar.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Segher Boessenkool wrote:

> [...]
> > BTW:
> > 
> > #define atomic_read(a)	(*(volatile int *)&(a))
> > #define atomic_set(a,i)	(*(volatile int *)&(a) = (i))
> > 
> > int a;
> > 
> > void func(void)
> > {
> > 	int b;
> > 
> > 	b = atomic_read(a);
> > 	atomic_set(a, 20);
> > 	b = atomic_read(a);
> > }
> > 
> > gives:
> > 
> > func:
> > 	pushl	%ebp
> > 	movl	a, %eax
> > 	movl	%esp, %ebp
> > 	movl	$20, a
> > 	movl	a, %eax
> > 	popl	%ebp
> > 	ret
> > 
> > so the first atomic_read() wasn't optimized away.
> 
> Of course.  It is executed by the abstract machine, so
> it will be executed by the actual machine.  On the other
> hand, try
> 
> 	b = 0;
> 	if (b)
> 		b = atomic_read(a);
> 
> or similar.

Yup, obviously. Volatile accesses (or any access to volatile objects),
or even "__volatile__ asms" (which gcc normally promises never to elid)
can always be optimized for cases such as these where the compiler can
trivially determine that the code in question is not reachable.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Satyam Sharma writes:

> > Doesn't "atomic WRT all processors" require volatility?
> 
> No, it definitely doesn't. Why should it?
> 
> "Atomic w.r.t. all processors" is just your normal, simple "atomicity"
> for SMP systems (ensure that that object is modified / set / replaced
> in main memory atomically) and has nothing to do with "volatile"
> behaviour.

Atomic variables are "volatile" in the sense that they are liable to
be changed at any time by mechanisms that are outside the knowledge of
the C compiler, namely, other CPUs, or this CPU executing an interrupt
routine.

In the kernel we use atomic variables in precisely those situations
where a variable is potentially accessed concurrently by multiple
CPUs, and where each CPU needs to see updates done by other CPUs in a
timely fashion.  That is what they are for.  Therefore the compiler
must not cache values of atomic variables in registers; each
atomic_read must result in a load and each atomic_set must result in a
store.  Anything else will just lead to subtle bugs.

I have no strong opinion about whether or not the best way to achieve
this is through the use of the "volatile" C keyword.  Segher's idea of
using asm instead seems like a good one to me.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> In the kernel we use atomic variables in precisely those situations
> where a variable is potentially accessed concurrently by multiple
> CPUs, and where each CPU needs to see updates done by other CPUs in a
> timely fashion.  That is what they are for.  Therefore the compiler
> must not cache values of atomic variables in registers; each
> atomic_read must result in a load and each atomic_set must result in a
> store.  Anything else will just lead to subtle bugs.

This may have been the intend. However, today the visibility is controlled 
using barriers. And we have barriers that we use with atomic operations. 
Having volatile be the default just lead to confusion. Atomic read should 
just read with no extras. Extras can be added by using variants like 
atomic_read_volatile or so.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Christoph Lameter writes:

> On Thu, 16 Aug 2007, Paul Mackerras wrote:
> 
> > In the kernel we use atomic variables in precisely those situations
> > where a variable is potentially accessed concurrently by multiple
> > CPUs, and where each CPU needs to see updates done by other CPUs in a
> > timely fashion.  That is what they are for.  Therefore the compiler
> > must not cache values of atomic variables in registers; each
> > atomic_read must result in a load and each atomic_set must result in a
> > store.  Anything else will just lead to subtle bugs.
> 
> This may have been the intend. However, today the visibility is controlled 
> using barriers. And we have barriers that we use with atomic operations. 

Those barriers are for when we need ordering between atomic variables
and other memory locations.  An atomic variable by itself doesn't and
shouldn't need any barriers for other CPUs to be able to see what's
happening to it.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> Those barriers are for when we need ordering between atomic variables
> and other memory locations.  An atomic variable by itself doesn't and
> shouldn't need any barriers for other CPUs to be able to see what's
> happening to it.

It does not need any barriers. As soon as one cpu acquires the 
cacheline for write it will be invalidated in the caches of the others. So 
the other cpu will have to refetch. No need for volatile.

The issue here may be that the compiler has fetched the atomic variable 
earlier and put it into a register. However, that prefetching is limited 
because it cannot cross functions calls etc. The only problem could be 
loops where the compiler does not refetch the variable since it assumes 
that it does not change and there are no function calls in the body of the 
loop. But AFAIK these loops need cpu_relax and other measures anyways to 
avoid bad effects from busy waiting.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 05:26:34PM -0700, Christoph Lameter wrote:
> On Thu, 16 Aug 2007, Paul Mackerras wrote:
> 
> > In the kernel we use atomic variables in precisely those situations
> > where a variable is potentially accessed concurrently by multiple
> > CPUs, and where each CPU needs to see updates done by other CPUs in a
> > timely fashion.  That is what they are for.  Therefore the compiler
> > must not cache values of atomic variables in registers; each
> > atomic_read must result in a load and each atomic_set must result in a
> > store.  Anything else will just lead to subtle bugs.
> 
> This may have been the intend. However, today the visibility is controlled 
> using barriers. And we have barriers that we use with atomic operations. 
> Having volatile be the default just lead to confusion. Atomic read should 
> just read with no extras. Extras can be added by using variants like 
> atomic_read_volatile or so.

Seems to me that we face greater chance of confusion without the
volatile than with, particularly as compiler optimizations become
more aggressive.  Yes, we could simply disable optimization, but
optimization can be quite helpful.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> Seems to me that we face greater chance of confusion without the
> volatile than with, particularly as compiler optimizations become
> more aggressive.  Yes, we could simply disable optimization, but
> optimization can be quite helpful.

A volatile default would disable optimizations for atomic_read. 
atomic_read without volatile would allow for full optimization by the 
compiler. Seems that this is what one wants in many cases.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 05:42:07PM -0700, Christoph Lameter wrote:
> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> 
> > Seems to me that we face greater chance of confusion without the
> > volatile than with, particularly as compiler optimizations become
> > more aggressive.  Yes, we could simply disable optimization, but
> > optimization can be quite helpful.
> 
> A volatile default would disable optimizations for atomic_read. 
> atomic_read without volatile would allow for full optimization by the 
> compiler. Seems that this is what one wants in many cases.

The volatile cast should not disable all that many optimizations,
for example, it is much less hurtful than barrier().  Furthermore,
the main optimizations disabled (pulling atomic_read() and atomic_set()
out of loops) really do need to be disabled.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> The volatile cast should not disable all that many optimizations,
> for example, it is much less hurtful than barrier().  Furthermore,
> the main optimizations disabled (pulling atomic_read() and atomic_set()
> out of loops) really do need to be disabled.

In many cases you do not need a barrier. Having volatile there *will* 
impact optimization because the compiler cannot use a register that may 
contain the value that was fetched earlier. And the compiler cannot choose 
freely when to fetch the value. The order of memory accesses are fixed if 
you use volatile. If the variable is not volatile then the compiler can 
arrange memory accesses any way they fit and thus generate better code.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 05:59:41PM -0700, Christoph Lameter wrote:
> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> 
> > The volatile cast should not disable all that many optimizations,
> > for example, it is much less hurtful than barrier().  Furthermore,
> > the main optimizations disabled (pulling atomic_read() and atomic_set()
> > out of loops) really do need to be disabled.
> 
> In many cases you do not need a barrier. Having volatile there *will* 
> impact optimization because the compiler cannot use a register that may 
> contain the value that was fetched earlier. And the compiler cannot choose 
> freely when to fetch the value. The order of memory accesses are fixed if 
> you use volatile. If the variable is not volatile then the compiler can 
> arrange memory accesses any way they fit and thus generate better code.

Understood.  My point is not that the impact is precisely zero, but
rather that the impact on optimization is much less hurtful than the
problems that could arise otherwise, particularly as compilers become
more aggressive in their optimizations.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Wed, 15 Aug 2007, Paul E. McKenney wrote:

> Understood.  My point is not that the impact is precisely zero, but
> rather that the impact on optimization is much less hurtful than the
> problems that could arise otherwise, particularly as compilers become
> more aggressive in their optimizations.

The problems arise because barriers are not used as required. Volatile 
has wishy washy semantics and somehow marries memory barriers with data 
access. It is clearer to separate the two. Conceptual cleanness usually 
translates into better code. If one really wants the volatile then lets 
make it explicit and use

	atomic_read_volatile()

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Wed, 15 Aug 2007, Christoph Lameter wrote:

> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> 
> > Understood.  My point is not that the impact is precisely zero, but
> > rather that the impact on optimization is much less hurtful than the
> > problems that could arise otherwise, particularly as compilers become
> > more aggressive in their optimizations.
> 
> The problems arise because barriers are not used as required. Volatile 
> has wishy washy semantics and somehow marries memory barriers with data 
> access. It is clearer to separate the two. Conceptual cleanness usually 
> translates into better code. If one really wants the volatile then lets 
> make it explicit and use
> 
> 	atomic_read_volatile()

Completely agreed, again. To summarize again (had done so about ~100 mails
earlier in this thread too :-) ...

atomic_{read,set}_volatile() -- guarantees volatility also along with
atomicity (the two _are_ different concepts after all, irrespective of
whether callsites normally want one with the other or not)

atomic_{read,set}_nonvolatile() -- only guarantees atomicity, compiler
free to elid / coalesce / optimize such accesses, can keep the object
in question cached in a local register, leads to smaller text, etc.

As to which one should be the default atomic_read() is a question of
whether majority of callsites (more weightage to important / hot
codepaths, lesser to obscure callsites) want a particular behaviour.

Do we have a consensus here? (hoping against hope, probably :-)


[ This thread has gotten completely out of hand ... for my mail client
  alpine as well, it now seems. Reminds of that 1000+ GPLv3 fest :-) ]

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 07:45:44AM +0530, Satyam Sharma wrote:
>
> Completely agreed, again. To summarize again (had done so about ~100 mails
> earlier in this thread too :-) ...
> 
> atomic_{read,set}_volatile() -- guarantees volatility also along with
> atomicity (the two _are_ different concepts after all, irrespective of
> whether callsites normally want one with the other or not)
> 
> atomic_{read,set}_nonvolatile() -- only guarantees atomicity, compiler
> free to elid / coalesce / optimize such accesses, can keep the object
> in question cached in a local register, leads to smaller text, etc.
> 
> As to which one should be the default atomic_read() is a question of
> whether majority of callsites (more weightage to important / hot
> codepaths, lesser to obscure callsites) want a particular behaviour.
> 
> Do we have a consensus here? (hoping against hope, probably :-)

I can certainly agree with this.

But I have to say that I still don't know of a single place
where one would actually use the volatile variant.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Herbert Xu wrote:

> > Do we have a consensus here? (hoping against hope, probably :-)
> 
> I can certainly agree with this.

I agree too.

> But I have to say that I still don't know of a single place
> where one would actually use the volatile variant.

I suspect that what you say is true after we have looked at all callers.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Christoph Lameter writes:

> > But I have to say that I still don't know of a single place
> > where one would actually use the volatile variant.
> 
> I suspect that what you say is true after we have looked at all callers.

It seems that there could be a lot of places where atomic_t is used in
a non-atomic fashion, and that those uses are either buggy, or there
is some lock held at the time which guarantees that other CPUs aren't
changing the value.  In both cases there is no point in using
atomic_t; we might as well just use an ordinary int.

In particular, atomic_read seems to lend itself to buggy uses.  People
seem to do things like:

	atomic_add(&v, something);
	if (atomic_read(&v) > something_else) ...

and expect that there is some relationship between the value that the
atomic_add stored and the value that the atomic_read will return,
which there isn't.  People seem to think that using atomic_t magically
gets rid of races.  It doesn't.

I'd go so far as to say that anywhere where you want a non-"volatile"
atomic_read, either your code is buggy, or else an int would work just
as well.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 01:23:06PM +1000, Paul Mackerras wrote:
>
> In particular, atomic_read seems to lend itself to buggy uses.  People
> seem to do things like:
> 
> 	atomic_add(&v, something);
> 	if (atomic_read(&v) > something_else) ...

If you're referring to the code in sk_stream_mem_schedule
then it's working as intended.  The atomicity guarantees
that the atomic_add/atomic_sub won't be seen in parts by
other readers.

We certainly do not need to see other atomic_add/atomic_sub
operations immediately.

If you're referring to another code snippet please cite.

> I'd go so far as to say that anywhere where you want a non-"volatile"
> atomic_read, either your code is buggy, or else an int would work just
> as well.

An int won't work here because += and -= do not have the
atomicity guarantees that atomic_add/atomic_sub do.  In
particular, this may cause an atomic_read on another CPU
to give a bogus reading.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> If you're referring to the code in sk_stream_mem_schedule
> then it's working as intended.  The atomicity guarantees

You mean it's intended that *sk->sk_prot->memory_pressure can end up
as 1 when sk->sk_prot->memory_allocated is small (less than
->sysctl_mem[0]), or as 0 when ->memory_allocated is large (greater
than ->sysctl_mem[2])?  Because that's the effect of the current code.
If so I wonder why you bother computing it.

> that the atomic_add/atomic_sub won't be seen in parts by
> other readers.
> 
> We certainly do not need to see other atomic_add/atomic_sub
> operations immediately.
> 
> If you're referring to another code snippet please cite.
> 
> > I'd go so far as to say that anywhere where you want a non-"volatile"
> > atomic_read, either your code is buggy, or else an int would work just
> > as well.
> 
> An int won't work here because += and -= do not have the
> atomicity guarantees that atomic_add/atomic_sub do.  In
> particular, this may cause an atomic_read on another CPU
> to give a bogus reading.

The point is that guaranteeing the atomicity of the increment or
decrement does not suffice to make the code race-free.  In this case
the race arises from the fact that reading ->memory_allocated and
setting *->memory_pressure are separate operations.  To make that code
work properly you need a lock.  And once you have the lock an ordinary
int would suffice for ->memory_allocated.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 01:48:32PM +1000, Paul Mackerras wrote:
> Herbert Xu writes:
> 
> > If you're referring to the code in sk_stream_mem_schedule
> > then it's working as intended.  The atomicity guarantees
> 
> You mean it's intended that *sk->sk_prot->memory_pressure can end up
> as 1 when sk->sk_prot->memory_allocated is small (less than
> ->sysctl_mem[0]), or as 0 when ->memory_allocated is large (greater
> than ->sysctl_mem[2])?  Because that's the effect of the current code.
> If so I wonder why you bother computing it.

You need to remember that there are three different limits:
minimum, pressure, and maximum.  By default we should never
be in a situation where what you say can occur.

If you set all three limits to the same thing, then yes it
won't work as intended but it's still well-behaved.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> > You mean it's intended that *sk->sk_prot->memory_pressure can end up
> > as 1 when sk->sk_prot->memory_allocated is small (less than
> > ->sysctl_mem[0]), or as 0 when ->memory_allocated is large (greater
> > than ->sysctl_mem[2])?  Because that's the effect of the current code.
> > If so I wonder why you bother computing it.
> 
> You need to remember that there are three different limits:
> minimum, pressure, and maximum.  By default we should never
> be in a situation where what you say can occur.
> 
> If you set all three limits to the same thing, then yes it
> won't work as intended but it's still well-behaved.

I'm not talking about setting all three limits to the same thing.

I'm talking about this situation:

CPU 0 comes into __sk_stream_mem_reclaim, reads memory_allocated, but
then before it can do the store to *memory_pressure, CPUs 1-1023 all
go through sk_stream_mem_schedule, collectively increase
memory_allocated to more than sysctl_mem[2] and set *memory_pressure.
Finally CPU 0 gets to do its store and it sets *memory_pressure back
to 0, but by this stage memory_allocated is way larger than
sysctl_mem[2].

Yes, it's unlikely, but that is the nature of race conditions - they
are unlikely, and only show up at inconvenient times, never when
someone who could fix the bug is watching. :)

Similarly it would be possible for other CPUs to decrease
memory_allocated from greater than sysctl_mem[2] to less than
sysctl_mem[0] in the interval between when we read memory_allocated
and set *memory_pressure to 1.  And it's quite possible for their
setting of *memory_pressure to 0 to happen before our setting of it to
1, so that it ends up at 1 when it should be 0.

Now, maybe it's the case that it doesn't really matter whether
*->memory_pressure is 0 or 1.  But if so, why bother computing it at
all?

People seem to think that using atomic_t means they don't need to use
a spinlock.  That's fine if there is only one variable involved, but
as soon as there's more than one, there's the possibility of a race,
whether or not you use atomic_t, and whether or not atomic_read has
"volatile" behaviour.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 02:34:25PM +1000, Paul Mackerras wrote:
>
> I'm talking about this situation:
> 
> CPU 0 comes into __sk_stream_mem_reclaim, reads memory_allocated, but
> then before it can do the store to *memory_pressure, CPUs 1-1023 all
> go through sk_stream_mem_schedule, collectively increase
> memory_allocated to more than sysctl_mem[2] and set *memory_pressure.
> Finally CPU 0 gets to do its store and it sets *memory_pressure back
> to 0, but by this stage memory_allocated is way larger than
> sysctl_mem[2].

It doesn't matter.  The memory pressure flag is an *advisory*
flag.  If we get it wrong the worst that'll happen is that we'd
waste some time doing work that'll be thrown away.

Please look at the places where it's used before jumping to
conclusions.

> Now, maybe it's the case that it doesn't really matter whether
> *->memory_pressure is 0 or 1.  But if so, why bother computing it at
> all?

As long as we get it right most of the time (and I think you
would agree that we do get it right most of the time), then
this flag has achieved its purpose.

> People seem to think that using atomic_t means they don't need to use
> a spinlock.  That's fine if there is only one variable involved, but
> as soon as there's more than one, there's the possibility of a race,
> whether or not you use atomic_t, and whether or not atomic_read has
> "volatile" behaviour.

In any case, this actually illustrates why the addition of
volatile is completely pointless.  Even if this code was
broken, which it definitely is not, having the volatile
there wouldn't have helped at all.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> It doesn't matter.  The memory pressure flag is an *advisory*
> flag.  If we get it wrong the worst that'll happen is that we'd
> waste some time doing work that'll be thrown away.

Ah, so it's the "racy but I don't care because it's only an
optimization" case.  That's fine.  Somehow I find it hard to believe
that all the racy uses of atomic_read in the kernel are like that,
though. :)

> In any case, this actually illustrates why the addition of
> volatile is completely pointless.  Even if this code was
> broken, which it definitely is not, having the volatile
> there wouldn't have helped at all.

Yes, adding volatile to racy code doesn't somehow make it race-free.
Neither does using atomic_t, despite what some seem to believe.

I have actually started going through all the uses of atomic_read in
the kernel.  So far out of the first 100 I have found none where we
have two atomic_reads of the same variable and the compiler could
usefully use the value from the first as the result of the second.
But there's still > 2500 to go...

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> Herbert Xu writes:
> 
> > It doesn't matter.  The memory pressure flag is an *advisory*
> > flag.  If we get it wrong the worst that'll happen is that we'd
> > waste some time doing work that'll be thrown away.
> 
> Ah, so it's the "racy but I don't care because it's only an
> optimization" case.  That's fine.  Somehow I find it hard to believe
> that all the racy uses of atomic_read in the kernel are like that,
> though. :)

My use of atomic_read in SLUB is like that. Volatile does not magically 
sync up reads somehow.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> 
> It seems that there could be a lot of places where atomic_t is used in
> a non-atomic fashion, and that those uses are either buggy, or there
> is some lock held at the time which guarantees that other CPUs aren't
> changing the value.  In both cases there is no point in using
> atomic_t; we might as well just use an ordinary int.

The point of atomic_t is to do atomic *changes* to the variable.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> I'd go so far as to say that anywhere where you want a non-"volatile"
> atomic_read, either your code is buggy, or else an int would work just
> as well.

Even, the only way to implement a "non-volatile" atomic_read() is
essentially as a plain int (you can do some tricks so you cannot
assign to the result and stuff like that, but that's not the issue
here).

So if that would be the behaviour we wanted, just get rid of that
whole atomic_read() thing, so no one can misuse it anymore.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Friesen]

Herbert Xu wrote:

> But I have to say that I still don't know of a single place
> where one would actually use the volatile variant.

Given that many of the existing users do currently have "volatile", are 
you comfortable simply removing that behaviour from them?  Are you sure 
that you will not introduce any issues?

Forcing a re-read is only a performance penalty.  Removing it can cause 
behavioural changes.

I would be more comfortable making the default match the majority of the 
current implementations (ie: volatile semantics).  Then, if someone 
cares about performance they can explicitly validate the call path and 
convert it over to the non-volatile version.

Correctness before speed...

Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Wed, Aug 15, 2007 at 06:41:40PM -0700, Christoph Lameter wrote:
> On Wed, 15 Aug 2007, Paul E. McKenney wrote:
> 
> > Understood.  My point is not that the impact is precisely zero, but
> > rather that the impact on optimization is much less hurtful than the
> > problems that could arise otherwise, particularly as compilers become
> > more aggressive in their optimizations.
> 
> The problems arise because barriers are not used as required. Volatile 
> has wishy washy semantics and somehow marries memory barriers with data 
> access. It is clearer to separate the two. Conceptual cleanness usually 
> translates into better code. If one really wants the volatile then lets 
> make it explicit and use
> 
> 	atomic_read_volatile()

There are indeed architectures where you can cause gcc to emit memory
barriers in response to volatile.  I am assuming that we are -not-
making gcc do this.  Given this, then volatiles and memory barrier
instructions are orthogonal -- one controls the compiler, the other
controls the CPU.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Christoph Lameter writes:

> A volatile default would disable optimizations for atomic_read. 
> atomic_read without volatile would allow for full optimization by the 
> compiler. Seems that this is what one wants in many cases.

Name one such case.

An atomic_read should do a load from memory.  If the programmer puts
an atomic_read() in the code then the compiler should emit a load for
it, not re-use a value returned by a previous atomic_read.  I do not
believe it would ever be useful for the compiler to collapse two
atomic_read statements into a single load.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 11:51:42AM +1000, Paul Mackerras wrote:
> 
> Name one such case.

See sk_stream_mem_schedule in net/core/stream.c:

        /* Under limit. */
        if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
                if (*sk->sk_prot->memory_pressure)
                        *sk->sk_prot->memory_pressure = 0;
                return 1;
        }

        /* Over hard limit. */
        if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
                sk->sk_prot->enter_memory_pressure();
                goto suppress_allocation;
        }

We don't need to reload sk->sk_prot->memory_allocated here.

Now where is your example again?

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> See sk_stream_mem_schedule in net/core/stream.c:
> 
>         /* Under limit. */
>         if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
>                 if (*sk->sk_prot->memory_pressure)
>                         *sk->sk_prot->memory_pressure = 0;
>                 return 1;
>         }
> 
>         /* Over hard limit. */
>         if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
>                 sk->sk_prot->enter_memory_pressure();
>                 goto suppress_allocation;
>         }
> 
> We don't need to reload sk->sk_prot->memory_allocated here.

Are you sure?  How do you know some other CPU hasn't changed the value
in between?

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 12:05:56PM +1000, Paul Mackerras wrote:
> Herbert Xu writes:
> 
> > See sk_stream_mem_schedule in net/core/stream.c:
> > 
> >         /* Under limit. */
> >         if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
> >                 if (*sk->sk_prot->memory_pressure)
> >                         *sk->sk_prot->memory_pressure = 0;
> >                 return 1;
> >         }
> > 
> >         /* Over hard limit. */
> >         if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
> >                 sk->sk_prot->enter_memory_pressure();
> >                 goto suppress_allocation;
> >         }
> > 
> > We don't need to reload sk->sk_prot->memory_allocated here.
> 
> Are you sure?  How do you know some other CPU hasn't changed the value
> in between?

Yes I'm sure, because we don't care if others have increased
the reservation.

Note that even if we did we'd be using barriers so volatile
won't do us any good here.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 10:11:05AM +0800, Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 12:05:56PM +1000, Paul Mackerras wrote:
> > Herbert Xu writes:
> > 
> > > See sk_stream_mem_schedule in net/core/stream.c:
> > > 
> > >         /* Under limit. */
> > >         if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
> > >                 if (*sk->sk_prot->memory_pressure)
> > >                         *sk->sk_prot->memory_pressure = 0;
> > >                 return 1;
> > >         }
> > > 
> > >         /* Over hard limit. */
> > >         if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
> > >                 sk->sk_prot->enter_memory_pressure();
> > >                 goto suppress_allocation;
> > >         }
> > > 
> > > We don't need to reload sk->sk_prot->memory_allocated here.
> > 
> > Are you sure?  How do you know some other CPU hasn't changed the value
> > in between?
> 
> Yes I'm sure, because we don't care if others have increased
> the reservation.
> 
> Note that even if we did we'd be using barriers so volatile
> won't do us any good here.

If the load-coalescing is important to performance, why not load into
a local variable?

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> > Are you sure?  How do you know some other CPU hasn't changed the value
> > in between?
> 
> Yes I'm sure, because we don't care if others have increased
> the reservation.

But others can also reduce the reservation.  Also, the code sets and
clears *sk->sk_prot->memory_pressure nonatomically with respect to the
reads of sk->sk_prot->memory_allocated, so in fact the code doesn't
guarantee any particular relationship between the two.

That code looks like a beautiful example of buggy, racy code where
someone has sprinkled magic fix-the-races dust (otherwise known as
atomic_t) around in a vain attempt to fix the races.

That's assuming that all that stuff actually performs any useful
purpose, of course, and that there isn't some lock held by the
callers.  In the latter case it is pointless using atomic_t.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 01:15:05PM +1000, Paul Mackerras wrote:
> 
> But others can also reduce the reservation.  Also, the code sets and
> clears *sk->sk_prot->memory_pressure nonatomically with respect to the
> reads of sk->sk_prot->memory_allocated, so in fact the code doesn't
> guarantee any particular relationship between the two.

Yes others can reduce the reservation, but the point of this
is that the code doesn't care.  We'll either see the value
before or after the reduction and in either case we'll do
something sensible.

The worst that can happen is when we're just below the hard
limit and multiple CPUs fail to allocate but that's not really
a problem because if the machine is making progress at all
then we will eventually scale back and allow these allocations
to succeed.

As to the non-atomic operation on memory_pressue, that's OK
because we only ever assign values to it and never do other
operations such as += or -=.  Remember that int/long assignments
must be atomic or Linux won't run on your architecture.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> > We don't need to reload sk->sk_prot->memory_allocated here.
> 
> Are you sure?  How do you know some other CPU hasn't changed the value
> in between?

The cpu knows because the cacheline was not invalidated.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Wed, 15 Aug 2007, Christoph Lameter wrote:

> On Thu, 16 Aug 2007, Paul Mackerras wrote:
> 
> > > We don't need to reload sk->sk_prot->memory_allocated here.
> > 
> > Are you sure?  How do you know some other CPU hasn't changed the value
> > in between?
> 
> The cpu knows because the cacheline was not invalidated.

Crap my statement above is wrong..... We do not care that the 
value was changed otherwise we would have put a barrier in there.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> Herbert Xu writes:
> 
> > See sk_stream_mem_schedule in net/core/stream.c:
> > 
> >         /* Under limit. */
> >         if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
> >                 if (*sk->sk_prot->memory_pressure)
> >                         *sk->sk_prot->memory_pressure = 0;
> >                 return 1;
> >         }
> > 
> >         /* Over hard limit. */
> >         if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
> >                 sk->sk_prot->enter_memory_pressure();
> >                 goto suppress_allocation;
> >         }
> > 
> > We don't need to reload sk->sk_prot->memory_allocated here.
> 
> Are you sure?  How do you know some other CPU hasn't changed the value
> in between?

I can't speak for this particular case, but there could be similar code
examples elsewhere, where we do the atomic ops on an atomic_t object
inside a higher-level locking scheme that would take care of the kind of
problem you're referring to here. It would be useful for such or similar
code if the compiler kept the value of that atomic object in a register.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Thu, 16 Aug 2007, Satyam Sharma wrote:

> On Thu, 16 Aug 2007, Paul Mackerras wrote:
> > Herbert Xu writes:
> > 
> > > See sk_stream_mem_schedule in net/core/stream.c:
> > > 
> > >         /* Under limit. */
> > >         if (atomic_read(sk->sk_prot->memory_allocated) < sk->sk_prot->sysctl_mem[0]) {
> > >                 if (*sk->sk_prot->memory_pressure)
> > >                         *sk->sk_prot->memory_pressure = 0;
> > >                 return 1;
> > >         }
> > > 
> > >         /* Over hard limit. */
> > >         if (atomic_read(sk->sk_prot->memory_allocated) > sk->sk_prot->sysctl_mem[2]) {
> > >                 sk->sk_prot->enter_memory_pressure();
> > >                 goto suppress_allocation;
> > >         }
> > > 
> > > We don't need to reload sk->sk_prot->memory_allocated here.
> > 
> > Are you sure?  How do you know some other CPU hasn't changed the value
> > in between?
> 
> I can't speak for this particular case, but there could be similar code
> examples elsewhere, where we do the atomic ops on an atomic_t object
> inside a higher-level locking scheme that would take care of the kind of
> problem you're referring to here. It would be useful for such or similar
> code if the compiler kept the value of that atomic object in a register.

We might not be using atomic_t (and ops) if we already have a higher-level
locking scheme, actually. So as Herbert mentioned, such cases might just
not care. [ Too much of this thread, too little sleep, sorry! ]

Anyway, the problem, of course, is that this conversion to a stronger /
safer-by-default behaviour doesn't happen with zero cost to performance.
Converting atomic ops to "volatile" behaviour did add ~2K to kernel text
for archs such as i386 (possibly to important codepaths) that didn't have
those semantics already so it would be constructive to actually look at
those differences and see if there were really any heisenbugs that got
rectified. Or if there were legitimate optimizations that got wrongly
disabled. Onus lies on those proposing the modifications, I'd say ;-)

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Satyam Sharma writes:

> Anyway, the problem, of course, is that this conversion to a stronger /
> safer-by-default behaviour doesn't happen with zero cost to performance.
> Converting atomic ops to "volatile" behaviour did add ~2K to kernel text
> for archs such as i386 (possibly to important codepaths) that didn't have
> those semantics already so it would be constructive to actually look at
> those differences and see if there were really any heisenbugs that got
> rectified. Or if there were legitimate optimizations that got wrongly
> disabled. Onus lies on those proposing the modifications, I'd say ;-)

The uses of atomic_read where one might want it to allow caching of
the result seem to me to fall into 3 categories:

1. Places that are buggy because of a race arising from the way it's
   used.

2. Places where there is a race but it doesn't matter because we're
   doing some clever trick.

3. Places where there is some locking in place that eliminates any
   potential race.

In case 1, adding volatile won't solve the race, of course, but it's
hard to argue that we shouldn't do something because it will slow down
buggy code.  Case 2 is hopefully pretty rare and accompanied by large
comment blocks, and in those cases caching the result of atomic_read
explicitly in a local variable would probably make the code clearer.
And in case 3 there is no reason to use atomic_t at all; we might as
well just use an int.

So I don't see any good reason to make the atomic API more complex by
having "volatile" and "non-volatile" versions of atomic_read.  It
should just have the "volatile" behaviour.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 02:11:43PM +1000, Paul Mackerras wrote:
>
> The uses of atomic_read where one might want it to allow caching of
> the result seem to me to fall into 3 categories:
> 
> 1. Places that are buggy because of a race arising from the way it's
>    used.
> 
> 2. Places where there is a race but it doesn't matter because we're
>    doing some clever trick.
> 
> 3. Places where there is some locking in place that eliminates any
>    potential race.

Agreed.

> In case 1, adding volatile won't solve the race, of course, but it's
> hard to argue that we shouldn't do something because it will slow down
> buggy code.  Case 2 is hopefully pretty rare and accompanied by large
> comment blocks, and in those cases caching the result of atomic_read
> explicitly in a local variable would probably make the code clearer.
> And in case 3 there is no reason to use atomic_t at all; we might as
> well just use an int.

Since adding volatile doesn't help any of the 3 cases, and
takes away optimisations from both 2 and 3, I wonder what
is the point of the addition after all?

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> On Thu, Aug 16, 2007 at 02:11:43PM +1000, Paul Mackerras wrote:
> >
> > The uses of atomic_read where one might want it to allow caching of
> > the result seem to me to fall into 3 categories:
> > 
> > 1. Places that are buggy because of a race arising from the way it's
> >    used.
> > 
> > 2. Places where there is a race but it doesn't matter because we're
> >    doing some clever trick.
> > 
> > 3. Places where there is some locking in place that eliminates any
> >    potential race.
> 
> Agreed.
> 
> > In case 1, adding volatile won't solve the race, of course, but it's
> > hard to argue that we shouldn't do something because it will slow down
> > buggy code.  Case 2 is hopefully pretty rare and accompanied by large
> > comment blocks, and in those cases caching the result of atomic_read
> > explicitly in a local variable would probably make the code clearer.
> > And in case 3 there is no reason to use atomic_t at all; we might as
> > well just use an int.
> 
> Since adding volatile doesn't help any of the 3 cases, and
> takes away optimisations from both 2 and 3, I wonder what
> is the point of the addition after all?

Note that I said these are the cases _where one might want to allow
caching_, so of course adding volatile doesn't help _these_ cases.
There are of course other cases where one definitely doesn't want to
allow the compiler to cache the value, such as when polling an atomic
variable waiting for another CPU to change it, and from my inspection
so far these cases seem to be the majority.

The reasons for having "volatile" behaviour of atomic_read (whether or
not that is achieved by use of the "volatile" C keyword) are

- It matches the normal expectation based on the name "atomic_read"
- It matches the behaviour of the other atomic_* primitives
- It avoids bugs in the cases where "volatile" behaviour is required

To my mind these outweigh the small benefit for some code of the
non-volatile (caching-allowed) behaviour.  In fact it's pretty minor
either way, and since x86[-64] has this behaviour, one can expect the
potential bugs in generic code to have mostly been found, although
perhaps not all of them since x86[-64] has less aggressive reordering
of memory accesses and fewer registers in which to cache things than
some other architectures.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 04:56:21PM +1000, Paul Mackerras wrote:
>
> Note that I said these are the cases _where one might want to allow
> caching_, so of course adding volatile doesn't help _these_ cases.
> There are of course other cases where one definitely doesn't want to
> allow the compiler to cache the value, such as when polling an atomic
> variable waiting for another CPU to change it, and from my inspection
> so far these cases seem to be the majority.

We've been through that already.  If it's a busy-wait it
should use cpu_relax.  If it's scheduling away that already
forces the compiler to reread anyway.

Do you have an actual example where volatile is needed?

> - It matches the normal expectation based on the name "atomic_read"
> - It matches the behaviour of the other atomic_* primitives

Can't argue since you left out what those expectations
or properties are.

> - It avoids bugs in the cases where "volatile" behaviour is required

Do you (or anyone else for that matter) have an example of this?

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 04:56:21PM +1000, Paul Mackerras wrote:
>>
>> Note that I said these are the cases _where one might want to allow
>> caching_, so of course adding volatile doesn't help _these_ cases.
>> There are of course other cases where one definitely doesn't want to
>> allow the compiler to cache the value, such as when polling an atomic
>> variable waiting for another CPU to change it, and from my inspection
>> so far these cases seem to be the majority.
> 
> We've been through that already.  If it's a busy-wait it
> should use cpu_relax.  If it's scheduling away that already
> forces the compiler to reread anyway.
> 
> Do you have an actual example where volatile is needed?
> 
>> - It matches the normal expectation based on the name "atomic_read"
>> - It matches the behaviour of the other atomic_* primitives
> 
> Can't argue since you left out what those expectations
> or properties are.

We use atomic_t for data that is concurrently locklessly written and
read at arbitrary times.  My naive expectation as driver author (driver
maintainer) is that all atomic_t accessors, including atomic_read, (and
atomic bitops) work with the then current value of the atomic data.

>> - It avoids bugs in the cases where "volatile" behaviour is required
> 
> Do you (or anyone else for that matter) have an example of this?

The only code I somewhat know, the ieee1394 subsystem, was perhaps
authored and is currently maintained with the expectation that each
occurrence of atomic_read actually results in a load operation, i.e. is
not optimized away.  This means all atomic_t (bus generation, packet and
buffer refcounts, and some other state variables)* and likewise all
atomic bitops in that subsystem.

If that assumption is wrong, then what is the API or language primitive
to force a load operation to occur?


*)  Interesting what a quick LXR session in search for all atomic_t
usages in 'my' subsystem brings to light.  I now noticed an apparently
unused struct member in the bitrotting pcilynx driver, and more
importantly, a pairing of two atomic_t variables in raw1394 that should
be audited for race conditions and for possible replacement by plain int.
-- 
Stefan Richter
-=====-=-=== =--- =----
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 10:06:31AM +0200, Stefan Richter wrote:
> > 
> > Do you (or anyone else for that matter) have an example of this?
> 
> The only code I somewhat know, the ieee1394 subsystem, was perhaps
> authored and is currently maintained with the expectation that each
> occurrence of atomic_read actually results in a load operation, i.e. is
> not optimized away.  This means all atomic_t (bus generation, packet and
> buffer refcounts, and some other state variables)* and likewise all
> atomic bitops in that subsystem.

Can you find an actual atomic_read code snippet there that is
broken without the volatile modifier?

Thanks,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 10:06:31AM +0200, Stefan Richter wrote:
>> > 
>> > Do you (or anyone else for that matter) have an example of this?
>> 
>> The only code I somewhat know, the ieee1394 subsystem, was perhaps
>> authored and is currently maintained with the expectation that each
>> occurrence of atomic_read actually results in a load operation, i.e. is
>> not optimized away.  This means all atomic_t (bus generation, packet and
>> buffer refcounts, and some other state variables)* and likewise all
>> atomic bitops in that subsystem.
> 
> Can you find an actual atomic_read code snippet there that is
> broken without the volatile modifier?

What do I have to look for?  atomic_read after another read or write
access to the same variable, in the same function scope?  Or in the sum
of scopes of functions that could be merged by function inlining?

One example was discussed here earlier:  The for (;;) loop in
nodemgr_host_thread.  There an msleep_interruptible implicitly acted as
barrier (at the moment because it's in a different translation unit; if
it were the same, then because it hopefully has own barriers).  So that
happens to work, although such an implicit barrier is bad style:  Better
enforce the desired behaviour (== guaranteed load operation) *explicitly*.
-- 
Stefan Richter
-=====-=-=== =--- =----
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

I wrote:
> Herbert Xu wrote:
>> On Thu, Aug 16, 2007 at 10:06:31AM +0200, Stefan Richter wrote:
[...]
>>> expectation that each
>>> occurrence of atomic_read actually results in a load operation, i.e. is
>>> not optimized away.
[...]
>> Can you find an actual atomic_read code snippet there that is
>> broken without the volatile modifier?

PS:  Just to clarify, I'm not speaking for the volatile modifier.  I'm
not speaking for any particular implementation of atomic_t and its
accessors at all.  All I am saying is that
  - we use atomically accessed data types because we concurrently but
    locklessly access this data,
  - hence a read access to this data that could be optimized away
    makes *no sense at all*.

The only sensible read accessor to an atomic datatype is a read accessor
that will not be optimized away.

So, the architecture guys can implement atomic_read however they want
--- as long as it cannot be optimized away.*

PPS:  If somebody has code where he can afford to let the compiler
coalesce atomic_read with a previous access to the same data, i.e.
doesn't need and doesn't want all guarantees that the atomic_read API
makes (or IMO should make), then he can replace the atomic_read by a
local temporary variable.


*) Exceptions:
	if (known_to_be_false)
		read_access(a);
and the like.
-- 
Stefan Richter
-=====-=-=== =--- =----
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 12:31:03PM +0200, Stefan Richter wrote:
> 
> PS:  Just to clarify, I'm not speaking for the volatile modifier.  I'm
> not speaking for any particular implementation of atomic_t and its
> accessors at all.  All I am saying is that
>   - we use atomically accessed data types because we concurrently but
>     locklessly access this data,
>   - hence a read access to this data that could be optimized away
>     makes *no sense at all*.

No sane compiler can optimise away an atomic_read per se.
That's only possible if there's a preceding atomic_set or
atomic_read, with no barriers in the middle.

If that's the case, then one has to conclude that doing
away with the second read is acceptable, as otherwise
a memory (or at least a compiler) barrier should have been
used.

In fact, volatile doesn't guarantee that the memory gets
read anyway.  You might be reading some stale value out
of the cache.  Granted this doesn't happen on x86 but
when you're coding for the kernel you can't make such
assumptions.

So the point here is that if you don't mind getting a stale
value from the CPU cache when doing an atomic_read, then
surely you won't mind getting a stale value from the compiler
"cache".

> So, the architecture guys can implement atomic_read however they want
> --- as long as it cannot be optimized away.*

They can implement it however they want as long as it stays
atomic.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 06:42:50PM +0800, Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 12:31:03PM +0200, Stefan Richter wrote:
> > 
> > PS:  Just to clarify, I'm not speaking for the volatile modifier.  I'm
> > not speaking for any particular implementation of atomic_t and its
> > accessors at all.  All I am saying is that
> >   - we use atomically accessed data types because we concurrently but
> >     locklessly access this data,
> >   - hence a read access to this data that could be optimized away
> >     makes *no sense at all*.
> 
> No sane compiler can optimise away an atomic_read per se.
> That's only possible if there's a preceding atomic_set or
> atomic_read, with no barriers in the middle.
> 
> If that's the case, then one has to conclude that doing
> away with the second read is acceptable, as otherwise
> a memory (or at least a compiler) barrier should have been
> used.

The compiler can also reorder non-volatile accesses.  For an example
patch that cares about this, please see:

	http://lkml.org/lkml/2007/8/7/280

This patch uses an ORDERED_WRT_IRQ() in rcu_read_lock() and
rcu_read_unlock() to ensure that accesses aren't reordered with respect
to interrupt handlers and NMIs/SMIs running on that same CPU.

> In fact, volatile doesn't guarantee that the memory gets
> read anyway.  You might be reading some stale value out
> of the cache.  Granted this doesn't happen on x86 but
> when you're coding for the kernel you can't make such
> assumptions.
> 
> So the point here is that if you don't mind getting a stale
> value from the CPU cache when doing an atomic_read, then
> surely you won't mind getting a stale value from the compiler
> "cache".

Absolutely disagree.  An interrupt/NMI/SMI handler running on the CPU
will see the same value (whether in cache or in store buffer) that
the mainline code will see.  In this case, we don't care about CPU
misordering, only about compiler misordering.  It is easy to see
other uses that combine communication with handlers on the current
CPU with communication among CPUs -- again, see prior messages in
this thread.

> > So, the architecture guys can implement atomic_read however they want
> > --- as long as it cannot be optimized away.*
> 
> They can implement it however they want as long as it stays
> atomic.

Precisely.  And volatility is a key property of "atomic".  Let's please
not throw it away.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 09:34:41AM -0700, Paul E. McKenney wrote:
>
> The compiler can also reorder non-volatile accesses.  For an example
> patch that cares about this, please see:
> 
> 	http://lkml.org/lkml/2007/8/7/280
> 
> This patch uses an ORDERED_WRT_IRQ() in rcu_read_lock() and
> rcu_read_unlock() to ensure that accesses aren't reordered with respect
> to interrupt handlers and NMIs/SMIs running on that same CPU.

Good, finally we have some code to discuss (even though it's
not actually in the kernel yet).

First of all, I think this illustrates that what you want
here has nothing to do with atomic ops.  The ORDERED_WRT_IRQ
macro occurs a lot more times in your patch than atomic
reads/sets.  So *assuming* that it was necessary at all,
then having an ordered variant of the atomic_read/atomic_set
ops could do just as well.

However, I still don't know which atomic_read/atomic_set in
your patch would be broken if there were no volatile.  Could
you please point them out?

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 07:59:02AM +0800, Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 09:34:41AM -0700, Paul E. McKenney wrote:
> >
> > The compiler can also reorder non-volatile accesses.  For an example
> > patch that cares about this, please see:
> > 
> > 	http://lkml.org/lkml/2007/8/7/280
> > 
> > This patch uses an ORDERED_WRT_IRQ() in rcu_read_lock() and
> > rcu_read_unlock() to ensure that accesses aren't reordered with respect
> > to interrupt handlers and NMIs/SMIs running on that same CPU.
> 
> Good, finally we have some code to discuss (even though it's
> not actually in the kernel yet).

There was some earlier in this thread as well.

> First of all, I think this illustrates that what you want
> here has nothing to do with atomic ops.  The ORDERED_WRT_IRQ
> macro occurs a lot more times in your patch than atomic
> reads/sets.  So *assuming* that it was necessary at all,
> then having an ordered variant of the atomic_read/atomic_set
> ops could do just as well.

Indeed.  If I could trust atomic_read()/atomic_set() to cause the compiler
to maintain ordering, then I could just use them instead of having to
create an  ORDERED_WRT_IRQ().  (Or ACCESS_ONCE(), as it is called in a
different patch.)

> However, I still don't know which atomic_read/atomic_set in
> your patch would be broken if there were no volatile.  Could
> you please point them out?

Suppose I tried replacing the ORDERED_WRT_IRQ() calls with
atomic_read() and atomic_set().  Starting with __rcu_read_lock():

o	If "ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])++"
	was ordered by the compiler after
	"ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1", then
	suppose an NMI/SMI happened after the rcu_read_lock_nesting but
	before the rcu_flipctr.

	Then if there was an rcu_read_lock() in the SMI/NMI
	handler (which is perfectly legal), the nested rcu_read_lock()
	would believe that it could take the then-clause of the
	enclosing "if" statement.  But because the rcu_flipctr per-CPU
	variable had not yet been incremented, an RCU updater would
	be within its rights to assume that there were no RCU reads
	in progress, thus possibly yanking a data structure out from
	under the reader in the SMI/NMI function.

	Fatal outcome.  Note that only one CPU is involved here
	because these are all either per-CPU or per-task variables.

o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1"
	was ordered by the compiler to follow the
	"ORDERED_WRT_IRQ(me->rcu_flipctr_idx) = idx", and an NMI/SMI
	happened between the two, then an __rcu_read_lock() in the NMI/SMI
	would incorrectly take the "else" clause of the enclosing "if"
	statement.  If some other CPU flipped the rcu_ctrlblk.completed
	in the meantime, then the __rcu_read_lock() would (correctly)
	write the new value into rcu_flipctr_idx.

	Well and good so far.  But the problem arises in
	__rcu_read_unlock(), which then decrements the wrong counter.
	Depending on exactly how subsequent events played out, this could
	result in either prematurely ending grace periods or never-ending
	grace periods, both of which are fatal outcomes.

And the following are not needed in the current version of the
patch, but will be in a future version that either avoids disabling
irqs or that dispenses with the smp_read_barrier_depends() that I
have 99% convinced myself is unneeded:

o	nesting = ORDERED_WRT_IRQ(me->rcu_read_lock_nesting);

o	idx = ORDERED_WRT_IRQ(rcu_ctrlblk.completed) & 0x1;

Furthermore, in that future version, irq handlers can cause the same
mischief that SMI/NMI handlers can in this version.

Next, looking at __rcu_read_unlock():

o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting - 1"
	was reordered by the compiler to follow the
	"ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])--",
	then if an NMI/SMI containing an rcu_read_lock() occurs between
	the two, this nested rcu_read_lock() would incorrectly believe
	that it was protected by an enclosing RCU read-side critical
	section as described in the first reversal discussed for
	__rcu_read_lock() above.  Again, fatal outcome.

This is what we have now.  It is not hard to imagine situations that
interact with -both- interrupt handlers -and- other CPUs, as described
earlier.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Thu, 16 Aug 2007, Paul E. McKenney wrote:

> On Fri, Aug 17, 2007 at 07:59:02AM +0800, Herbert Xu wrote:
> > On Thu, Aug 16, 2007 at 09:34:41AM -0700, Paul E. McKenney wrote:
> > >
> > > The compiler can also reorder non-volatile accesses.  For an example
> > > patch that cares about this, please see:
> > > 
> > > 	http://lkml.org/lkml/2007/8/7/280
> > > 
> > > This patch uses an ORDERED_WRT_IRQ() in rcu_read_lock() and
> > > rcu_read_unlock() to ensure that accesses aren't reordered with respect
> > > to interrupt handlers and NMIs/SMIs running on that same CPU.
> > 
> > Good, finally we have some code to discuss (even though it's
> > not actually in the kernel yet).
> 
> There was some earlier in this thread as well.

Hmm, I never quite got what all this interrupt/NMI/SMI handling and
RCU business you mentioned earlier was all about, but now that you've
pointed to the actual code and issues with it ...


> > First of all, I think this illustrates that what you want
> > here has nothing to do with atomic ops.  The ORDERED_WRT_IRQ
> > macro occurs a lot more times in your patch than atomic
> > reads/sets.  So *assuming* that it was necessary at all,
> > then having an ordered variant of the atomic_read/atomic_set
> > ops could do just as well.
> 
> Indeed.  If I could trust atomic_read()/atomic_set() to cause the compiler
> to maintain ordering, then I could just use them instead of having to
> create an  ORDERED_WRT_IRQ().  (Or ACCESS_ONCE(), as it is called in a
> different patch.)

+#define WHATEVER(x)	(*(volatile typeof(x) *)&(x))

I suppose one could want volatile access semantics for stuff that's
a bit-field too, no?

Also, this gives *zero* "re-ordering" guarantees that your code wants
as you've explained it below) -- neither w.r.t. CPU re-ordering (which
probably you don't care about) *nor* w.r.t. compiler re-ordering
(which you definitely _do_ care about).


> > However, I still don't know which atomic_read/atomic_set in
> > your patch would be broken if there were no volatile.  Could
> > you please point them out?
> 
> Suppose I tried replacing the ORDERED_WRT_IRQ() calls with
> atomic_read() and atomic_set().  Starting with __rcu_read_lock():
> 
> o	If "ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])++"
> 	was ordered by the compiler after
> 	"ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1", then
> 	suppose an NMI/SMI happened after the rcu_read_lock_nesting but
> 	before the rcu_flipctr.
> 
> 	Then if there was an rcu_read_lock() in the SMI/NMI
> 	handler (which is perfectly legal), the nested rcu_read_lock()
> 	would believe that it could take the then-clause of the
> 	enclosing "if" statement.  But because the rcu_flipctr per-CPU
> 	variable had not yet been incremented, an RCU updater would
> 	be within its rights to assume that there were no RCU reads
> 	in progress, thus possibly yanking a data structure out from
> 	under the reader in the SMI/NMI function.
> 
> 	Fatal outcome.  Note that only one CPU is involved here
> 	because these are all either per-CPU or per-task variables.

Ok, so you don't care about CPU re-ordering. Still, I should let you know
that your ORDERED_WRT_IRQ() -- bad name, btw -- is still buggy. What you
want is a full compiler optimization barrier().

[ Your code probably works now, and emits correct code, but that's
  just because of gcc did what it did. Nothing in any standard,
  or in any documented behaviour of gcc, or anything about the real
  (or expected) semantics of "volatile" is protecting the code here. ]


> o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1"
> 	was ordered by the compiler to follow the
> 	"ORDERED_WRT_IRQ(me->rcu_flipctr_idx) = idx", and an NMI/SMI
> 	happened between the two, then an __rcu_read_lock() in the NMI/SMI
> 	would incorrectly take the "else" clause of the enclosing "if"
> 	statement.  If some other CPU flipped the rcu_ctrlblk.completed
> 	in the meantime, then the __rcu_read_lock() would (correctly)
> 	write the new value into rcu_flipctr_idx.
> 
> 	Well and good so far.  But the problem arises in
> 	__rcu_read_unlock(), which then decrements the wrong counter.
> 	Depending on exactly how subsequent events played out, this could
> 	result in either prematurely ending grace periods or never-ending
> 	grace periods, both of which are fatal outcomes.
> 
> And the following are not needed in the current version of the
> patch, but will be in a future version that either avoids disabling
> irqs or that dispenses with the smp_read_barrier_depends() that I
> have 99% convinced myself is unneeded:
> 
> o	nesting = ORDERED_WRT_IRQ(me->rcu_read_lock_nesting);
> 
> o	idx = ORDERED_WRT_IRQ(rcu_ctrlblk.completed) & 0x1;
> 
> Furthermore, in that future version, irq handlers can cause the same
> mischief that SMI/NMI handlers can in this version.
> 
> Next, looking at __rcu_read_unlock():
> 
> o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting - 1"
> 	was reordered by the compiler to follow the
> 	"ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])--",
> 	then if an NMI/SMI containing an rcu_read_lock() occurs between
> 	the two, this nested rcu_read_lock() would incorrectly believe
> 	that it was protected by an enclosing RCU read-side critical
> 	section as described in the first reversal discussed for
> 	__rcu_read_lock() above.  Again, fatal outcome.
> 
> This is what we have now.  It is not hard to imagine situations that
> interact with -both- interrupt handlers -and- other CPUs, as described
> earlier.

It's not about interrupt/SMI/NMI handlers at all! What you clearly want,
simply put, is that a certain stream of C statements must be emitted
by the compiler _as they are_ with no re-ordering optimizations! You must
*definitely* use barrier(), IMHO.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 01:09:08PM +0530, Satyam Sharma wrote:
> 
> 
> On Thu, 16 Aug 2007, Paul E. McKenney wrote:
> 
> > On Fri, Aug 17, 2007 at 07:59:02AM +0800, Herbert Xu wrote:
> > > On Thu, Aug 16, 2007 at 09:34:41AM -0700, Paul E. McKenney wrote:
> > > >
> > > > The compiler can also reorder non-volatile accesses.  For an example
> > > > patch that cares about this, please see:
> > > > 
> > > > 	http://lkml.org/lkml/2007/8/7/280
> > > > 
> > > > This patch uses an ORDERED_WRT_IRQ() in rcu_read_lock() and
> > > > rcu_read_unlock() to ensure that accesses aren't reordered with respect
> > > > to interrupt handlers and NMIs/SMIs running on that same CPU.
> > > 
> > > Good, finally we have some code to discuss (even though it's
> > > not actually in the kernel yet).
> > 
> > There was some earlier in this thread as well.
> 
> Hmm, I never quite got what all this interrupt/NMI/SMI handling and
> RCU business you mentioned earlier was all about, but now that you've
> pointed to the actual code and issues with it ...

Glad to help...

> > > First of all, I think this illustrates that what you want
> > > here has nothing to do with atomic ops.  The ORDERED_WRT_IRQ
> > > macro occurs a lot more times in your patch than atomic
> > > reads/sets.  So *assuming* that it was necessary at all,
> > > then having an ordered variant of the atomic_read/atomic_set
> > > ops could do just as well.
> > 
> > Indeed.  If I could trust atomic_read()/atomic_set() to cause the compiler
> > to maintain ordering, then I could just use them instead of having to
> > create an  ORDERED_WRT_IRQ().  (Or ACCESS_ONCE(), as it is called in a
> > different patch.)
> 
> +#define WHATEVER(x)	(*(volatile typeof(x) *)&(x))
> 
> I suppose one could want volatile access semantics for stuff that's
> a bit-field too, no?

One could, but this is not supported in general.  So if you want that,
you need to use the usual bit-mask tricks and (for setting) atomic
operations.

> Also, this gives *zero* "re-ordering" guarantees that your code wants
> as you've explained it below) -- neither w.r.t. CPU re-ordering (which
> probably you don't care about) *nor* w.r.t. compiler re-ordering
> (which you definitely _do_ care about).

You are correct about CPU re-ordering (and about the fact that this
example doesn't care about it), but not about compiler re-ordering.

The compiler is prohibited from moving a volatile access across a sequence
point.  One example of a sequence point is a statement boundary.  Because
all of the volatile accesses in this code are separated by statement
boundaries, a conforming compiler is prohibited from reordering them.

> > > However, I still don't know which atomic_read/atomic_set in
> > > your patch would be broken if there were no volatile.  Could
> > > you please point them out?
> > 
> > Suppose I tried replacing the ORDERED_WRT_IRQ() calls with
> > atomic_read() and atomic_set().  Starting with __rcu_read_lock():
> > 
> > o	If "ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])++"
> > 	was ordered by the compiler after
> > 	"ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1", then
> > 	suppose an NMI/SMI happened after the rcu_read_lock_nesting but
> > 	before the rcu_flipctr.
> > 
> > 	Then if there was an rcu_read_lock() in the SMI/NMI
> > 	handler (which is perfectly legal), the nested rcu_read_lock()
> > 	would believe that it could take the then-clause of the
> > 	enclosing "if" statement.  But because the rcu_flipctr per-CPU
> > 	variable had not yet been incremented, an RCU updater would
> > 	be within its rights to assume that there were no RCU reads
> > 	in progress, thus possibly yanking a data structure out from
> > 	under the reader in the SMI/NMI function.
> > 
> > 	Fatal outcome.  Note that only one CPU is involved here
> > 	because these are all either per-CPU or per-task variables.
> 
> Ok, so you don't care about CPU re-ordering. Still, I should let you know
> that your ORDERED_WRT_IRQ() -- bad name, btw -- is still buggy. What you
> want is a full compiler optimization barrier().

No.  See above.

> [ Your code probably works now, and emits correct code, but that's
>   just because of gcc did what it did. Nothing in any standard,
>   or in any documented behaviour of gcc, or anything about the real
>   (or expected) semantics of "volatile" is protecting the code here. ]

Really?  Why doesn't the prohibition against moving volatile accesses
across sequence points take care of this?

> > o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1"
> > 	was ordered by the compiler to follow the
> > 	"ORDERED_WRT_IRQ(me->rcu_flipctr_idx) = idx", and an NMI/SMI
> > 	happened between the two, then an __rcu_read_lock() in the NMI/SMI
> > 	would incorrectly take the "else" clause of the enclosing "if"
> > 	statement.  If some other CPU flipped the rcu_ctrlblk.completed
> > 	in the meantime, then the __rcu_read_lock() would (correctly)
> > 	write the new value into rcu_flipctr_idx.
> > 
> > 	Well and good so far.  But the problem arises in
> > 	__rcu_read_unlock(), which then decrements the wrong counter.
> > 	Depending on exactly how subsequent events played out, this could
> > 	result in either prematurely ending grace periods or never-ending
> > 	grace periods, both of which are fatal outcomes.
> > 
> > And the following are not needed in the current version of the
> > patch, but will be in a future version that either avoids disabling
> > irqs or that dispenses with the smp_read_barrier_depends() that I
> > have 99% convinced myself is unneeded:
> > 
> > o	nesting = ORDERED_WRT_IRQ(me->rcu_read_lock_nesting);
> > 
> > o	idx = ORDERED_WRT_IRQ(rcu_ctrlblk.completed) & 0x1;
> > 
> > Furthermore, in that future version, irq handlers can cause the same
> > mischief that SMI/NMI handlers can in this version.
> > 
> > Next, looking at __rcu_read_unlock():
> > 
> > o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting - 1"
> > 	was reordered by the compiler to follow the
> > 	"ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])--",
> > 	then if an NMI/SMI containing an rcu_read_lock() occurs between
> > 	the two, this nested rcu_read_lock() would incorrectly believe
> > 	that it was protected by an enclosing RCU read-side critical
> > 	section as described in the first reversal discussed for
> > 	__rcu_read_lock() above.  Again, fatal outcome.
> > 
> > This is what we have now.  It is not hard to imagine situations that
> > interact with -both- interrupt handlers -and- other CPUs, as described
> > earlier.
> 
> It's not about interrupt/SMI/NMI handlers at all! What you clearly want,
> simply put, is that a certain stream of C statements must be emitted
> by the compiler _as they are_ with no re-ordering optimizations! You must
> *definitely* use barrier(), IMHO.

Almost.  I don't care about most of the operations, only about the loads
and stores marked volatile.  Again, although the compiler is free to
reorder volatile accesses that occur -within- a single statement, it
is prohibited by the standard from moving volatile accesses from one
statement to another.  Therefore, this code can legitimately use volatile.

Or am I missing something subtle?

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Paul E. McKenney wrote:

> On Fri, Aug 17, 2007 at 01:09:08PM +0530, Satyam Sharma wrote:
> > 
> > On Thu, 16 Aug 2007, Paul E. McKenney wrote:
> > 
> > > On Fri, Aug 17, 2007 at 07:59:02AM +0800, Herbert Xu wrote:
> > > > 
> > > > First of all, I think this illustrates that what you want
> > > > here has nothing to do with atomic ops.  The ORDERED_WRT_IRQ
> > > > macro occurs a lot more times in your patch than atomic
> > > > reads/sets.  So *assuming* that it was necessary at all,
> > > > then having an ordered variant of the atomic_read/atomic_set
> > > > ops could do just as well.
> > > 
> > > Indeed.  If I could trust atomic_read()/atomic_set() to cause the compiler
> > > to maintain ordering, then I could just use them instead of having to
> > > create an  ORDERED_WRT_IRQ().  (Or ACCESS_ONCE(), as it is called in a
> > > different patch.)
> > 
> > +#define WHATEVER(x)	(*(volatile typeof(x) *)&(x))
> > [...]
> > Also, this gives *zero* "re-ordering" guarantees that your code wants
> > as you've explained it below) -- neither w.r.t. CPU re-ordering (which
> > probably you don't care about) *nor* w.r.t. compiler re-ordering
> > (which you definitely _do_ care about).
> 
> You are correct about CPU re-ordering (and about the fact that this
> example doesn't care about it), but not about compiler re-ordering.
> 
> The compiler is prohibited from moving a volatile access across a sequence
> point.  One example of a sequence point is a statement boundary.  Because
> all of the volatile accesses in this code are separated by statement
> boundaries, a conforming compiler is prohibited from reordering them.

Yes, you're right, and I believe precisely this was discussed elsewhere
as well today.

But I'd call attention to what Herbert mentioned there. You're using
ORDERED_WRT_IRQ() on stuff that is _not_ defined to be an atomic_t at all:

* Member "completed" of struct rcu_ctrlblk is a long.
* Per-cpu variable rcu_flipctr is an array of ints.
* Members "rcu_read_lock_nesting" and "rcu_flipctr_idx" of
  struct task_struct are ints.

So are you saying you're "having to use" this volatile-access macro
because you *couldn't* declare all the above as atomic_t and thus just
expect the right thing to happen by using the atomic ops API by default,
because it lacks volatile access semantics (on x86)?

If so, then I wonder if using the volatile access cast is really the
best way to achieve (at least in terms of code clarity) the kind of
re-ordering guarantees it wants there. (there could be alternative
solutions, such as using barrier(), or that at bottom of this mail)

What I mean is this: If you switch to atomic_t, and x86 switched to
make atomic_t have "volatile" semantics by default, the statements
would be simply a string of: atomic_inc(), atomic_add(), atomic_set(),
and atomic_read() statements, and nothing in there that clearly makes
it *explicit* that the code is correct (and not buggy) simply because
of the re-ordering guarantees that the C "volatile" type-qualifier
keyword gives us as per the standard. But now we're firmly in
"subjective" territory, so you or anybody could legitimately disagree.


> > > Suppose I tried replacing the ORDERED_WRT_IRQ() calls with
> > > atomic_read() and atomic_set().  Starting with __rcu_read_lock():
> > > 
> > > o	If "ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])++"
> > > 	was ordered by the compiler after
> > > 	"ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1", then
> > > 	suppose an NMI/SMI happened after the rcu_read_lock_nesting but
> > > 	before the rcu_flipctr.
> > > 
> > > 	Then if there was an rcu_read_lock() in the SMI/NMI
> > > 	handler (which is perfectly legal), the nested rcu_read_lock()
> > > 	would believe that it could take the then-clause of the
> > > 	enclosing "if" statement.  But because the rcu_flipctr per-CPU
> > > 	variable had not yet been incremented, an RCU updater would
> > > 	be within its rights to assume that there were no RCU reads
> > > 	in progress, thus possibly yanking a data structure out from
> > > 	under the reader in the SMI/NMI function.
> > > 
> > > 	Fatal outcome.  Note that only one CPU is involved here
> > > 	because these are all either per-CPU or per-task variables.
> > 
> > Ok, so you don't care about CPU re-ordering. Still, I should let you know
> > that your ORDERED_WRT_IRQ() -- bad name, btw -- is still buggy. What you
> > want is a full compiler optimization barrier().
> 
> No.  See above.

True, *(volatile foo *)& _will_ work for this case.

But multiple calls to barrier() (granted, would invalidate all other
optimizations also) would work as well, would it not?

[ Interestingly, if you declared all those objects mentioned earlier as
  atomic_t, and x86(-64) switched to an __asm__ __volatile__ based variant
  for atomic_{read,set}_volatile(), the bugs you want to avoid would still
  be there. "volatile" the C language type-qualifier does have compiler
  re-ordering semantics you mentioned earlier, but the "volatile" that
  applies to inline asm()s gives no re-ordering guarantees. ]


> > > o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1"
> > > 	was ordered by the compiler to follow the
> > > 	"ORDERED_WRT_IRQ(me->rcu_flipctr_idx) = idx", and an NMI/SMI
> > > 	happened between the two, then an __rcu_read_lock() in the NMI/SMI
> > > 	would incorrectly take the "else" clause of the enclosing "if"
> > > 	statement.  If some other CPU flipped the rcu_ctrlblk.completed
> > > 	in the meantime, then the __rcu_read_lock() would (correctly)
> > > 	write the new value into rcu_flipctr_idx.
> > > 
> > > 	Well and good so far.  But the problem arises in
> > > 	__rcu_read_unlock(), which then decrements the wrong counter.
> > > 	Depending on exactly how subsequent events played out, this could
> > > 	result in either prematurely ending grace periods or never-ending
> > > 	grace periods, both of which are fatal outcomes.
> > > 
> > > And the following are not needed in the current version of the
> > > patch, but will be in a future version that either avoids disabling
> > > irqs or that dispenses with the smp_read_barrier_depends() that I
> > > have 99% convinced myself is unneeded:
> > > 
> > > o	nesting = ORDERED_WRT_IRQ(me->rcu_read_lock_nesting);
> > > 
> > > o	idx = ORDERED_WRT_IRQ(rcu_ctrlblk.completed) & 0x1;
> > > 
> > > Furthermore, in that future version, irq handlers can cause the same
> > > mischief that SMI/NMI handlers can in this version.

So don't remove the local_irq_save/restore, which is well-established and
well-understood for such cases (it doesn't help you with SMI/NMI,
admittedly). This isn't really about RCU or per-cpu vars as such, it's
just about racy code where you don't want to get hit by a concurrent
interrupt (it does turn out that doing things in a _particular order_ will
not cause fatal/buggy behaviour, but it's still a race issue, after all).


> > > Next, looking at __rcu_read_unlock():
> > > 
> > > o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting - 1"
> > > 	was reordered by the compiler to follow the
> > > 	"ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])--",
> > > 	then if an NMI/SMI containing an rcu_read_lock() occurs between
> > > 	the two, this nested rcu_read_lock() would incorrectly believe
> > > 	that it was protected by an enclosing RCU read-side critical
> > > 	section as described in the first reversal discussed for
> > > 	__rcu_read_lock() above.  Again, fatal outcome.
> > > 
> > > This is what we have now.  It is not hard to imagine situations that
> > > interact with -both- interrupt handlers -and- other CPUs, as described
> > > earlier.

Unless somebody's going for a lockless implementation, such situations
normally use spin_lock_irqsave() based locking (or local_irq_save for
those who care only for current CPU) -- problem with the patch in question,
is that you want to prevent races with concurrent SMI/NMIs as well, which
is not something that a lot of code needs to consider.

[ Curiously, another thread is discussing something similar also:
  http://lkml.org/lkml/2007/8/15/393 "RFC: do get_rtc_time() correctly" ]

Anyway, I didn't look at the code in that patch very much in detail, but
why couldn't you implement some kind of synchronization variable that lets
rcu_read_lock() or rcu_read_unlock() -- when being called from inside an
NMI or SMI handler -- know that it has concurrently interrupted an ongoing
rcu_read_{un}lock() and so must do things differently ... (?)

I'm also wondering if there's other code that's not using locking in the
kernel that faces similar issues, and what they've done to deal with it
(if anything). Such bugs would be subtle, and difficult to diagnose.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Sat, Aug 18, 2007 at 12:01:38AM +0530, Satyam Sharma wrote:
> 
> 
> On Fri, 17 Aug 2007, Paul E. McKenney wrote:
> 
> > On Fri, Aug 17, 2007 at 01:09:08PM +0530, Satyam Sharma wrote:
> > > 
> > > On Thu, 16 Aug 2007, Paul E. McKenney wrote:
> > > 
> > > > On Fri, Aug 17, 2007 at 07:59:02AM +0800, Herbert Xu wrote:
> > > > > 
> > > > > First of all, I think this illustrates that what you want
> > > > > here has nothing to do with atomic ops.  The ORDERED_WRT_IRQ
> > > > > macro occurs a lot more times in your patch than atomic
> > > > > reads/sets.  So *assuming* that it was necessary at all,
> > > > > then having an ordered variant of the atomic_read/atomic_set
> > > > > ops could do just as well.
> > > > 
> > > > Indeed.  If I could trust atomic_read()/atomic_set() to cause the compiler
> > > > to maintain ordering, then I could just use them instead of having to
> > > > create an  ORDERED_WRT_IRQ().  (Or ACCESS_ONCE(), as it is called in a
> > > > different patch.)
> > > 
> > > +#define WHATEVER(x)	(*(volatile typeof(x) *)&(x))
> > > [...]
> > > Also, this gives *zero* "re-ordering" guarantees that your code wants
> > > as you've explained it below) -- neither w.r.t. CPU re-ordering (which
> > > probably you don't care about) *nor* w.r.t. compiler re-ordering
> > > (which you definitely _do_ care about).
> > 
> > You are correct about CPU re-ordering (and about the fact that this
> > example doesn't care about it), but not about compiler re-ordering.
> > 
> > The compiler is prohibited from moving a volatile access across a sequence
> > point.  One example of a sequence point is a statement boundary.  Because
> > all of the volatile accesses in this code are separated by statement
> > boundaries, a conforming compiler is prohibited from reordering them.
> 
> Yes, you're right, and I believe precisely this was discussed elsewhere
> as well today.
> 
> But I'd call attention to what Herbert mentioned there. You're using
> ORDERED_WRT_IRQ() on stuff that is _not_ defined to be an atomic_t at all:
> 
> * Member "completed" of struct rcu_ctrlblk is a long.
> * Per-cpu variable rcu_flipctr is an array of ints.
> * Members "rcu_read_lock_nesting" and "rcu_flipctr_idx" of
>   struct task_struct are ints.
> 
> So are you saying you're "having to use" this volatile-access macro
> because you *couldn't* declare all the above as atomic_t and thus just
> expect the right thing to happen by using the atomic ops API by default,
> because it lacks volatile access semantics (on x86)?
> 
> If so, then I wonder if using the volatile access cast is really the
> best way to achieve (at least in terms of code clarity) the kind of
> re-ordering guarantees it wants there. (there could be alternative
> solutions, such as using barrier(), or that at bottom of this mail)
> 
> What I mean is this: If you switch to atomic_t, and x86 switched to
> make atomic_t have "volatile" semantics by default, the statements
> would be simply a string of: atomic_inc(), atomic_add(), atomic_set(),
> and atomic_read() statements, and nothing in there that clearly makes
> it *explicit* that the code is correct (and not buggy) simply because
> of the re-ordering guarantees that the C "volatile" type-qualifier
> keyword gives us as per the standard. But now we're firmly in
> "subjective" territory, so you or anybody could legitimately disagree.

In any case, given Linus's note, it appears that atomic_read() and
atomic_set() won't consistently have volatile semantics, at least
not while the compiler generates such ugly code for volatile accesses.
So I will continue with my current approach.

In any case, I will not be using atomic_inc() or atomic_add() in this
code, as doing so would more than double the overhead, even on machines
that are the most efficient at implementing atomic operations.

> > > > Suppose I tried replacing the ORDERED_WRT_IRQ() calls with
> > > > atomic_read() and atomic_set().  Starting with __rcu_read_lock():
> > > > 
> > > > o	If "ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])++"
> > > > 	was ordered by the compiler after
> > > > 	"ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1", then
> > > > 	suppose an NMI/SMI happened after the rcu_read_lock_nesting but
> > > > 	before the rcu_flipctr.
> > > > 
> > > > 	Then if there was an rcu_read_lock() in the SMI/NMI
> > > > 	handler (which is perfectly legal), the nested rcu_read_lock()
> > > > 	would believe that it could take the then-clause of the
> > > > 	enclosing "if" statement.  But because the rcu_flipctr per-CPU
> > > > 	variable had not yet been incremented, an RCU updater would
> > > > 	be within its rights to assume that there were no RCU reads
> > > > 	in progress, thus possibly yanking a data structure out from
> > > > 	under the reader in the SMI/NMI function.
> > > > 
> > > > 	Fatal outcome.  Note that only one CPU is involved here
> > > > 	because these are all either per-CPU or per-task variables.
> > > 
> > > Ok, so you don't care about CPU re-ordering. Still, I should let you know
> > > that your ORDERED_WRT_IRQ() -- bad name, btw -- is still buggy. What you
> > > want is a full compiler optimization barrier().
> > 
> > No.  See above.
> 
> True, *(volatile foo *)& _will_ work for this case.
> 
> But multiple calls to barrier() (granted, would invalidate all other
> optimizations also) would work as well, would it not?

They work, but are a bit slower.  So they do work, but not as well.

> [ Interestingly, if you declared all those objects mentioned earlier as
>   atomic_t, and x86(-64) switched to an __asm__ __volatile__ based variant
>   for atomic_{read,set}_volatile(), the bugs you want to avoid would still
>   be there. "volatile" the C language type-qualifier does have compiler
>   re-ordering semantics you mentioned earlier, but the "volatile" that
>   applies to inline asm()s gives no re-ordering guarantees. ]

Well, that certainly would be a point in favor of "volatile" over inline
asms.  ;-)

> > > > o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting + 1"
> > > > 	was ordered by the compiler to follow the
> > > > 	"ORDERED_WRT_IRQ(me->rcu_flipctr_idx) = idx", and an NMI/SMI
> > > > 	happened between the two, then an __rcu_read_lock() in the NMI/SMI
> > > > 	would incorrectly take the "else" clause of the enclosing "if"
> > > > 	statement.  If some other CPU flipped the rcu_ctrlblk.completed
> > > > 	in the meantime, then the __rcu_read_lock() would (correctly)
> > > > 	write the new value into rcu_flipctr_idx.
> > > > 
> > > > 	Well and good so far.  But the problem arises in
> > > > 	__rcu_read_unlock(), which then decrements the wrong counter.
> > > > 	Depending on exactly how subsequent events played out, this could
> > > > 	result in either prematurely ending grace periods or never-ending
> > > > 	grace periods, both of which are fatal outcomes.
> > > > 
> > > > And the following are not needed in the current version of the
> > > > patch, but will be in a future version that either avoids disabling
> > > > irqs or that dispenses with the smp_read_barrier_depends() that I
> > > > have 99% convinced myself is unneeded:
> > > > 
> > > > o	nesting = ORDERED_WRT_IRQ(me->rcu_read_lock_nesting);
> > > > 
> > > > o	idx = ORDERED_WRT_IRQ(rcu_ctrlblk.completed) & 0x1;
> > > > 
> > > > Furthermore, in that future version, irq handlers can cause the same
> > > > mischief that SMI/NMI handlers can in this version.
> 
> So don't remove the local_irq_save/restore, which is well-established and
> well-understood for such cases (it doesn't help you with SMI/NMI,
> admittedly). This isn't really about RCU or per-cpu vars as such, it's
> just about racy code where you don't want to get hit by a concurrent
> interrupt (it does turn out that doing things in a _particular order_ will
> not cause fatal/buggy behaviour, but it's still a race issue, after all).

The local_irq_save/restore are something like 30% of the overhead of
these two functions, so will be looking hard at getting rid of them.
Doing so allows the scheduling-clock interrupt to get into the mix,
and also allows preemption.  The goal would be to find some trick that
suppresses preemption, fends off the grace-period-computation code
invoked from the the scheduling-clock interrupt, and otherwise keeps
things on an even keel.

> > > > Next, looking at __rcu_read_unlock():
> > > > 
> > > > o	If "ORDERED_WRT_IRQ(me->rcu_read_lock_nesting) = nesting - 1"
> > > > 	was reordered by the compiler to follow the
> > > > 	"ORDERED_WRT_IRQ(__get_cpu_var(rcu_flipctr)[idx])--",
> > > > 	then if an NMI/SMI containing an rcu_read_lock() occurs between
> > > > 	the two, this nested rcu_read_lock() would incorrectly believe
> > > > 	that it was protected by an enclosing RCU read-side critical
> > > > 	section as described in the first reversal discussed for
> > > > 	__rcu_read_lock() above.  Again, fatal outcome.
> > > > 
> > > > This is what we have now.  It is not hard to imagine situations that
> > > > interact with -both- interrupt handlers -and- other CPUs, as described
> > > > earlier.
> 
> Unless somebody's going for a lockless implementation, such situations
> normally use spin_lock_irqsave() based locking (or local_irq_save for
> those who care only for current CPU) -- problem with the patch in question,
> is that you want to prevent races with concurrent SMI/NMIs as well, which
> is not something that a lot of code needs to consider.

Or that needs to resolve similar races with IRQs without disabling them.
One reason to avoid disabling IRQs is to avoid degrading scheduling
latency.  In any case, I do agree that the amount of code that must
worry about this is quite small at the moment.  I believe that it
will become more common, but would imagine that this belief might not
be universal.  Yet, anyway.  ;-)

> [ Curiously, another thread is discussing something similar also:
>   http://lkml.org/lkml/2007/8/15/393 "RFC: do get_rtc_time() correctly" ]
> 
> Anyway, I didn't look at the code in that patch very much in detail, but
> why couldn't you implement some kind of synchronization variable that lets
> rcu_read_lock() or rcu_read_unlock() -- when being called from inside an
> NMI or SMI handler -- know that it has concurrently interrupted an ongoing
> rcu_read_{un}lock() and so must do things differently ... (?)

Given some low-level details of the current implementation, I could
imagine manipulating rcu_read_lock_nesting on entry to and exit from
all NMI/SMI handlers, but would like to avoid that kind of architecture
dependency.  I am not confident of locating all of them, for one thing...

> I'm also wondering if there's other code that's not using locking in the
> kernel that faces similar issues, and what they've done to deal with it
> (if anything). Such bugs would be subtle, and difficult to diagnose.

Agreed!

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Paul E. McKenney wrote:
> On Thu, Aug 16, 2007 at 06:42:50PM +0800, Herbert Xu wrote:

>>In fact, volatile doesn't guarantee that the memory gets
>>read anyway.  You might be reading some stale value out
>>of the cache.  Granted this doesn't happen on x86 but
>>when you're coding for the kernel you can't make such
>>assumptions.
>>
>>So the point here is that if you don't mind getting a stale
>>value from the CPU cache when doing an atomic_read, then
>>surely you won't mind getting a stale value from the compiler
>>"cache".
> 
> 
> Absolutely disagree.  An interrupt/NMI/SMI handler running on the CPU
> will see the same value (whether in cache or in store buffer) that
> the mainline code will see.  In this case, we don't care about CPU
> misordering, only about compiler misordering.  It is easy to see
> other uses that combine communication with handlers on the current
> CPU with communication among CPUs -- again, see prior messages in
> this thread.

I still don't agree with the underlying assumption that everybody
(or lots of kernel code) treats atomic accesses as volatile.

Nobody that does has managed to explain my logic problem either:
loads and stores to long and ptr have always been considered to be
atomic, test_bit is atomic; so why are another special subclass of
atomic loads and stores? (and yes, it is perfectly legitimate to
want a non-volatile read for a data type that you also want to do
atomic RMW operations on)

Why are people making these undocumented and just plain false
assumptions about atomic_t? If they're using lockless code (ie.
which they must be if using atomics), then they actually need to be
thinking much harder about memory ordering issues. If that is too
much for them, then they can just use locks.


>>>So, the architecture guys can implement atomic_read however they want
>>>--- as long as it cannot be optimized away.*
>>
>>They can implement it however they want as long as it stays
>>atomic.
> 
> 
> Precisely.  And volatility is a key property of "atomic".  Let's please
> not throw it away.

It isn't, though (at least not since i386 and x86-64 don't have it).
_Adding_ it is trivial, and can be done any time. Throwing it away
(ie. making the API weaker) is _hard_. So let's not add it without
really good reasons. It most definitely results in worse code
generation in practice.

I don't know why people would assume volatile of atomics. AFAIK, most
of the documentation is pretty clear that all the atomic stuff can be
reordered etc. except for those that modify and return a value.

It isn't even intuitive: `*lp = value` is like the most fundamental
atomic operation in Linux.

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Nick Piggin writes:

> Why are people making these undocumented and just plain false
> assumptions about atomic_t?

Well, it has only been false since December 2006.  Prior to that
atomics *were* volatile on all platforms.

> If they're using lockless code (ie.
> which they must be if using atomics), then they actually need to be
> thinking much harder about memory ordering issues.

Indeed.  I believe that most uses of atomic_read other than in polling
loops or debug printk statements are actually racy.  In some cases the
race doesn't seem to matter, but I'm sure there are cases where it
does.

> If that is too
> much for them, then they can just use locks.

Why use locks when you can just sprinkle magic fix-the-races dust (aka
atomic_t) over your code? :) :)

> > Precisely.  And volatility is a key property of "atomic".  Let's please
> > not throw it away.
> 
> It isn't, though (at least not since i386 and x86-64 don't have it).

Conceptually it is, because atomic_t is specifically for variables
which are liable to be modified by other CPUs, and volatile _means_
"liable to be changed by mechanisms outside the knowledge of the
compiler".

> _Adding_ it is trivial, and can be done any time. Throwing it away
> (ie. making the API weaker) is _hard_. So let's not add it without

Well, in one sense it's not that hard - Linus did it just 8 months ago
in commit f9e9dcb3. :)

> really good reasons. It most definitely results in worse code
> generation in practice.

0.0008% increase in kernel text size on powerpc according to my
measurement. :)

> I don't know why people would assume volatile of atomics. AFAIK, most

By making something an atomic_t you're saying "other CPUs are going to
be modifying this, so treat it specially".  It's reasonable to assume
that special treatment extends to reading and setting it.

> of the documentation is pretty clear that all the atomic stuff can be
> reordered etc. except for those that modify and return a value.

Volatility isn't primarily about reordering (though as Linus says it
does restrict reordering to some extent).

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Paul Mackerras wrote:
> Nick Piggin writes:
> 
> 
>>Why are people making these undocumented and just plain false
>>assumptions about atomic_t?
> 
> 
> Well, it has only been false since December 2006.  Prior to that
> atomics *were* volatile on all platforms.

Hmm, although I don't think it has ever been guaranteed by the
API documentation (concede documentation is often not treated
as the authoritative source here, but for atomic it is actually
very good and obviously indispensable as the memory ordering
reference).


>>If they're using lockless code (ie.
>>which they must be if using atomics), then they actually need to be
>>thinking much harder about memory ordering issues.
> 
> 
> Indeed.  I believe that most uses of atomic_read other than in polling
> loops or debug printk statements are actually racy.  In some cases the
> race doesn't seem to matter, but I'm sure there are cases where it
> does.
> 
> 
>>If that is too
>>much for them, then they can just use locks.
> 
> 
> Why use locks when you can just sprinkle magic fix-the-races dust (aka
> atomic_t) over your code? :) :)

I agree with your skepticism of a lot of lockless code. But I think
a lot of the more subtle race problems will not be fixed with volatile.
The big, dumb infinite loop bugs would be fixed, but they're pretty
trivial to debug and even audit for.


>>>Precisely.  And volatility is a key property of "atomic".  Let's please
>>>not throw it away.
>>
>>It isn't, though (at least not since i386 and x86-64 don't have it).
> 
> 
> Conceptually it is, because atomic_t is specifically for variables
> which are liable to be modified by other CPUs, and volatile _means_
> "liable to be changed by mechanisms outside the knowledge of the
> compiler".

Usually that is the case, yes. But also most of the time we don't
care that it has been changed and don't mind it being reordered or
eliminated.

One of the only places we really care about that at all is for
variables that are modified by the *same* CPU.


>>_Adding_ it is trivial, and can be done any time. Throwing it away
>>(ie. making the API weaker) is _hard_. So let's not add it without
> 
> 
> Well, in one sense it's not that hard - Linus did it just 8 months ago
> in commit f9e9dcb3. :)

Well it would have been harder if the documentation also guaranteed
that atomic_read/atomic_set was ordered. Or it would have been harder
for _me_ to make such a change, anyway ;)


>>really good reasons. It most definitely results in worse code
>>generation in practice.
> 
> 
> 0.0008% increase in kernel text size on powerpc according to my
> measurement. :)

I don't think you're making a bad choice by keeping it volatile on
powerpc and waiting for others to shake out more of the bugs. You
get to fix everybody else's memory ordering bugs :)


>>I don't know why people would assume volatile of atomics. AFAIK, most
> 
> 
> By making something an atomic_t you're saying "other CPUs are going to
> be modifying this, so treat it specially".  It's reasonable to assume
> that special treatment extends to reading and setting it.

But I don't actually know what that "special treatment" is. Well
actually, I do know that operations will never result in a partial
modification being exposed. I also know that the operators that
do not modify and return are not guaranteed to have any sort of
ordering constraints.

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Nick Piggin wrote:
> I don't know why people would assume volatile of atomics. AFAIK, most
> of the documentation is pretty clear that all the atomic stuff can be
> reordered etc. except for those that modify and return a value.

Which documentation is there?

For driver authors, there is LDD3.  It doesn't specifically cover
effects of optimization on accesses to atomic_t.

For architecture port authors, there is Documentation/atomic_ops.txt.
Driver authors also can learn something from that document, as it
indirectly documents the atomic_t and bitops APIs.

Prompted by this thread, I reread this document, and indeed, the
sentence "Unlike the above routines, it is required that explicit memory
barriers are performed before and after [atomic_{inc,dec}_return]"
indicates that atomic_read (one of the "above routines") is very
different from all other atomic_t accessors that return values.

This is strange.  Why is it that atomic_read stands out that way?  IMO
this API imbalance is quite unexpected by many people.  Wouldn't it be
beneficial to change the atomic_read API to behave the same like all
other atomic_t accessors that return values?

OK, it is also different from the other accessors that return data in so
far as it doesn't modify the data.  But as driver "author", i.e. user of
the API, I can't see much use of an atomic_read that can be reordered
and, more importantly, can be optimized away by the compiler.  Sure, now
that I learned of these properties I can start to audit code and insert
barriers where I believe they are needed, but this simply means that
almost all occurrences of atomic_read will get barriers (unless there
already are implicit but more or less obvious barriers like msleep).
-- 
Stefan Richter
-=====-=-=== =--- =---=
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Stefan Richter wrote:
> Nick Piggin wrote:
> 
>>I don't know why people would assume volatile of atomics. AFAIK, most
>>of the documentation is pretty clear that all the atomic stuff can be
>>reordered etc. except for those that modify and return a value.
> 
> 
> Which documentation is there?

Documentation/atomic_ops.txt


> For driver authors, there is LDD3.  It doesn't specifically cover
> effects of optimization on accesses to atomic_t.
> 
> For architecture port authors, there is Documentation/atomic_ops.txt.
> Driver authors also can learn something from that document, as it
> indirectly documents the atomic_t and bitops APIs.
>

"Semantics and Behavior of Atomic and Bitmask Operations" is
pretty direct :)

Sure, it says that it's for arch maintainers, but there is no
reason why users can't make use of it.


> Prompted by this thread, I reread this document, and indeed, the
> sentence "Unlike the above routines, it is required that explicit memory
> barriers are performed before and after [atomic_{inc,dec}_return]"
> indicates that atomic_read (one of the "above routines") is very
> different from all other atomic_t accessors that return values.
> 
> This is strange.  Why is it that atomic_read stands out that way?  IMO

It is not just atomic_read of course. It is atomic_add,sub,inc,dec,set.


> this API imbalance is quite unexpected by many people.  Wouldn't it be
> beneficial to change the atomic_read API to behave the same like all
> other atomic_t accessors that return values?

It is very consistent and well defined. Operations which both modify
the data _and_ return something are defined to have full barriers
before and after.

What do you want to add to the other atomic accessors? Full memory
barriers? Only compiler barriers? It's quite likely that if you think
some barriers will fix bugs, then there are other bugs lurking there
anyway.

Just use spinlocks if you're not absolutely clear about potential
races and memory ordering issues -- they're pretty cheap and simple.


> OK, it is also different from the other accessors that return data in so
> far as it doesn't modify the data.  But as driver "author", i.e. user of
> the API, I can't see much use of an atomic_read that can be reordered
> and, more importantly, can be optimized away by the compiler.

It will return to you an atomic snapshot of the data (loaded from
memory at some point since the last compiler barrier). All you have
to be aware of compiler barriers and the Linux SMP memory ordering
model, which should be a given if you are writing lockless code.


> Sure, now
> that I learned of these properties I can start to audit code and insert
> barriers where I believe they are needed, but this simply means that
> almost all occurrences of atomic_read will get barriers (unless there
> already are implicit but more or less obvious barriers like msleep).

You might find that these places that appear to need barriers are
buggy for other reasons anyway. Can you point to some in-tree code
we can have a look at?

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Stefan Richter wrote:
> [...]
> Just use spinlocks if you're not absolutely clear about potential
> races and memory ordering issues -- they're pretty cheap and simple.

I fully agree with this. As Paul Mackerras mentioned elsewhere,
a lot of authors sprinkle atomic_t in code thinking they're somehow
done with *locking*. This is sad, and I wonder if it's time for a
Documentation/atomic-considered-dodgy.txt kind of document :-)


> > Sure, now
> > that I learned of these properties I can start to audit code and insert
> > barriers where I believe they are needed, but this simply means that
> > almost all occurrences of atomic_read will get barriers (unless there
> > already are implicit but more or less obvious barriers like msleep).
> 
> You might find that these places that appear to need barriers are
> buggy for other reasons anyway. Can you point to some in-tree code
> we can have a look at?

Such code was mentioned elsewhere (query nodemgr_host_thread in cscope)
that managed to escape the requirement for a barrier only because of
some completely un-obvious compilation-unit-scope thing. But I find such
an non-explicit barrier quite bad taste. Stefan, do consider plunking an
explicit call to barrier() there.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:
> 
> On Fri, 17 Aug 2007, Nick Piggin wrote:

>>>Sure, now
>>>that I learned of these properties I can start to audit code and insert
>>>barriers where I believe they are needed, but this simply means that
>>>almost all occurrences of atomic_read will get barriers (unless there
>>>already are implicit but more or less obvious barriers like msleep).
>>
>>You might find that these places that appear to need barriers are
>>buggy for other reasons anyway. Can you point to some in-tree code
>>we can have a look at?
> 
> 
> Such code was mentioned elsewhere (query nodemgr_host_thread in cscope)
> that managed to escape the requirement for a barrier only because of
> some completely un-obvious compilation-unit-scope thing. But I find such
> an non-explicit barrier quite bad taste. Stefan, do consider plunking an
> explicit call to barrier() there.

It is very obvious. msleep calls schedule() (ie. sleeps), which is
always a barrier.

The "unobvious" thing is that you wanted to know how the compiler knows
a function is a barrier -- answer is that if it does not *know* it is not
a barrier, it must assume it is a barrier. If the whole msleep call chain
including the scheduler were defined static in the current compilation
unit, then it would still be a barrier because it would actually be able
to see the barriers in schedule(void), if nothing else.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> > 
> > On Fri, 17 Aug 2007, Nick Piggin wrote:
> 
> > > > Sure, now
> > > > that I learned of these properties I can start to audit code and insert
> > > > barriers where I believe they are needed, but this simply means that
> > > > almost all occurrences of atomic_read will get barriers (unless there
> > > > already are implicit but more or less obvious barriers like msleep).
> > > 
> > > You might find that these places that appear to need barriers are
> > > buggy for other reasons anyway. Can you point to some in-tree code
> > > we can have a look at?
> > 
> > 
> > Such code was mentioned elsewhere (query nodemgr_host_thread in cscope)
> > that managed to escape the requirement for a barrier only because of
> > some completely un-obvious compilation-unit-scope thing. But I find such
> > an non-explicit barrier quite bad taste. Stefan, do consider plunking an
> > explicit call to barrier() there.
> 
> It is very obvious. msleep calls schedule() (ie. sleeps), which is
> always a barrier.

Probably you didn't mean that, but no, schedule() is not barrier because
it sleeps. It's a barrier because it's invisible.

> The "unobvious" thing is that you wanted to know how the compiler knows
> a function is a barrier -- answer is that if it does not *know* it is not
> a barrier, it must assume it is a barrier.

True, that's clearly what happens here. But are you're definitely joking
that this is "obvious" in terms of code-clarity, right?

Just 5 minutes back you mentioned elsewhere you like seeing lots of
explicit calls to barrier() (with comments, no less, hmm? :-)

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:

>
>On Fri, 17 Aug 2007, Nick Piggin wrote:
>
>
>>Satyam Sharma wrote:
>>
>>It is very obvious. msleep calls schedule() (ie. sleeps), which is
>>always a barrier.
>>
>
>Probably you didn't mean that, but no, schedule() is not barrier because
>it sleeps. It's a barrier because it's invisible.
>

Where did I say it is a barrier because it sleeps?

It is always a barrier because, at the lowest level, schedule() (and thus
anything that sleeps) is defined to always be a barrier. Regardless of
whatever obscure means the compiler might need to infer the barrier.

In other words, you can ignore those obscure details because schedule() is
always going to have an explicit barrier in it.


>>The "unobvious" thing is that you wanted to know how the compiler knows
>>a function is a barrier -- answer is that if it does not *know* it is not
>>a barrier, it must assume it is a barrier.
>>
>
>True, that's clearly what happens here. But are you're definitely joking
>that this is "obvious" in terms of code-clarity, right?
>

No. If you accept that barrier() is implemented correctly, and you know
that sleeping is defined to be a barrier, then its perfectly clear. You
don't have to know how the compiler "knows" that some function contains
a barrier.


>Just 5 minutes back you mentioned elsewhere you like seeing lots of
>explicit calls to barrier() (with comments, no less, hmm? :-)
>

Sure, but there are well known primitives which contain barriers, and
trivial recognisable code sequences for which you don't need comments.
waiting-loops using sleeps or cpu_relax() are prime examples.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> > On Fri, 17 Aug 2007, Nick Piggin wrote:
> > > Satyam Sharma wrote:
> > > 
> > > It is very obvious. msleep calls schedule() (ie. sleeps), which is
> > > always a barrier.
> > 
> > Probably you didn't mean that, but no, schedule() is not barrier because
> > it sleeps. It's a barrier because it's invisible.
> 
> Where did I say it is a barrier because it sleeps?

Just below. What you wrote:

> It is always a barrier because, at the lowest level, schedule() (and thus
> anything that sleeps) is defined to always be a barrier.

"It is always a barrier because, at the lowest level, anything that sleeps
is defined to always be a barrier".


> Regardless of
> whatever obscure means the compiler might need to infer the barrier.
> 
> In other words, you can ignore those obscure details because schedule() is
> always going to have an explicit barrier in it.

I didn't quite understand what you said here, so I'll tell what I think:

* foo() is a compiler barrier if the definition of foo() is invisible to
  the compiler at a callsite.

* foo() is also a compiler barrier if the definition of foo() includes
  a barrier, and it is inlined at the callsite.

If the above is wrong, or if there's something else at play as well,
do let me know.

> > > The "unobvious" thing is that you wanted to know how the compiler knows
> > > a function is a barrier -- answer is that if it does not *know* it is not
> > > a barrier, it must assume it is a barrier.
> > 
> > True, that's clearly what happens here. But are you're definitely joking
> > that this is "obvious" in terms of code-clarity, right?
> 
> No. If you accept that barrier() is implemented correctly, and you know
> that sleeping is defined to be a barrier,

Curiously, that's the second time you've said "sleeping is defined to
be a (compiler) barrier". How does the compiler even know if foo() is
a function that "sleeps"? Do compilers have some notion of "sleeping"
to ensure they automatically assume a compiler barrier whenever such
a function is called? Or are you saying that the compiler can see the
barrier() inside said function ... nopes, you're saying quite the
opposite below.


> then its perfectly clear. You
> don't have to know how the compiler "knows" that some function contains
> a barrier.

I think I do, why not? Would appreciate if you could elaborate on this.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:

>
>On Fri, 17 Aug 2007, Nick Piggin wrote:
>
>
>>Satyam Sharma wrote:
>>
>>>On Fri, 17 Aug 2007, Nick Piggin wrote:
>>>
>>>>Satyam Sharma wrote:
>>>>
>>>>It is very obvious. msleep calls schedule() (ie. sleeps), which is
>>>>always a barrier.
>>>>
>>>Probably you didn't mean that, but no, schedule() is not barrier because
>>>it sleeps. It's a barrier because it's invisible.
>>>
>>Where did I say it is a barrier because it sleeps?
>>
>
>Just below. What you wrote:
>
>
>>It is always a barrier because, at the lowest level, schedule() (and thus
>>anything that sleeps) is defined to always be a barrier.
>>
>
>"It is always a barrier because, at the lowest level, anything that sleeps
>is defined to always be a barrier".
>

... because it must call schedule and schedule is a barrier.


>>Regardless of
>>whatever obscure means the compiler might need to infer the barrier.
>>
>>In other words, you can ignore those obscure details because schedule() is
>>always going to have an explicit barrier in it.
>>
>
>I didn't quite understand what you said here, so I'll tell what I think:
>
>* foo() is a compiler barrier if the definition of foo() is invisible to
>  the compiler at a callsite.
>
>* foo() is also a compiler barrier if the definition of foo() includes
>  a barrier, and it is inlined at the callsite.
>
>If the above is wrong, or if there's something else at play as well,
>do let me know.
>

Right.


>>>>The "unobvious" thing is that you wanted to know how the compiler knows
>>>>a function is a barrier -- answer is that if it does not *know* it is not
>>>>a barrier, it must assume it is a barrier.
>>>>
>>>True, that's clearly what happens here. But are you're definitely joking
>>>that this is "obvious" in terms of code-clarity, right?
>>>
>>No. If you accept that barrier() is implemented correctly, and you know
>>that sleeping is defined to be a barrier,
>>
>
>Curiously, that's the second time you've said "sleeping is defined to
>be a (compiler) barrier".
>

_In Linux,_ sleeping is defined to be a compiler barrier.

>How does the compiler even know if foo() is
>a function that "sleeps"? Do compilers have some notion of "sleeping"
>to ensure they automatically assume a compiler barrier whenever such
>a function is called? Or are you saying that the compiler can see the
>barrier() inside said function ... nopes, you're saying quite the
>opposite below.
>

You're getting too worried about the compiler implementation. Start
by assuming that it does work ;)


>>then its perfectly clear. You
>>don't have to know how the compiler "knows" that some function contains
>>a barrier.
>>
>
>I think I do, why not? Would appreciate if you could elaborate on this.
>

If a function is not completely visible to the compiler (so it can't
determine whether a barrier could be in it or not), then it must always
assume it will contain a barrier so it always does the right thing.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> 
> > I didn't quite understand what you said here, so I'll tell what I think:
> > 
> > * foo() is a compiler barrier if the definition of foo() is invisible to
> >  the compiler at a callsite.
> > 
> > * foo() is also a compiler barrier if the definition of foo() includes
> >  a barrier, and it is inlined at the callsite.
> > 
> > If the above is wrong, or if there's something else at play as well,
> > do let me know.
> 
> [...]
> If a function is not completely visible to the compiler (so it can't
> determine whether a barrier could be in it or not), then it must always
> assume it will contain a barrier so it always does the right thing.

Yup, that's what I'd said just a few sentences above, as you can see. I
was actually asking for "elaboration" on "how a compiler determines that
function foo() (say foo == schedule), even when it cannot see that it has
a barrier(), as you'd mentioned, is a 'sleeping' function" actually, and
whether compilers have a "notion of sleep to automatically assume a
compiler barrier whenever such a sleeping function foo() is called". But
I think you've already qualified the discussion to this kernel, so okay,
I shouldn't nit-pick anymore.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Nick Piggin wrote:
> Stefan Richter wrote:
>> For architecture port authors, there is Documentation/atomic_ops.txt.
>> Driver authors also can learn something from that document, as it
>> indirectly documents the atomic_t and bitops APIs.
> 
> "Semantics and Behavior of Atomic and Bitmask Operations" is
> pretty direct :)

"Indirect", "pretty direct"... It's subjective.

(It is not an API documentation; it is an implementation specification.)

> Sure, it says that it's for arch maintainers, but there is no
> reason why users can't make use of it.
> 
> 
>> Prompted by this thread, I reread this document, and indeed, the
>> sentence "Unlike the above routines, it is required that explicit memory
>> barriers are performed before and after [atomic_{inc,dec}_return]"
>> indicates that atomic_read (one of the "above routines") is very
>> different from all other atomic_t accessors that return values.
>> 
>> This is strange.  Why is it that atomic_read stands out that way?  IMO
> 
> It is not just atomic_read of course. It is atomic_add,sub,inc,dec,set.

Yes, but unlike these, atomic_read returns a value.

Without me (the API user) providing extra barriers, that value may
become something else whenever someone touches code in the vicinity of
the atomic_read.

>> this API imbalance is quite unexpected by many people.  Wouldn't it be
>> beneficial to change the atomic_read API to behave the same like all
>> other atomic_t accessors that return values?
> 
> It is very consistent and well defined. Operations which both modify
> the data _and_ return something are defined to have full barriers
> before and after.

You are right, atomic_read is not only different from accessors that
don't retunr values, it is also different from all other accessors that
return values (because they all also modify the value).  There is just
no actual API documentation, which contributes to the issue that some
people (or at least one: me) learn a little bit late how special
atomic_read is.

> What do you want to add to the other atomic accessors? Full memory
> barriers? Only compiler barriers? It's quite likely that if you think
> some barriers will fix bugs, then there are other bugs lurking there
> anyway.

A lot of different though related issues are discussed in this thread,
but I personally am only occupied by one particular thing:  What kind of
return values do I get from atomic_read.

> Just use spinlocks if you're not absolutely clear about potential
> races and memory ordering issues -- they're pretty cheap and simple.

Probably good advice, like generally if driver guys consider lockless
algorithms.

>> OK, it is also different from the other accessors that return data in so
>> far as it doesn't modify the data.  But as driver "author", i.e. user of
>> the API, I can't see much use of an atomic_read that can be reordered
>> and, more importantly, can be optimized away by the compiler.
> 
> It will return to you an atomic snapshot of the data (loaded from
> memory at some point since the last compiler barrier). All you have
> to be aware of compiler barriers and the Linux SMP memory ordering
> model, which should be a given if you are writing lockless code.

OK, that's what I slowly realized during this discussion, and I
appreciate the explanations that were given here.

>> Sure, now
>> that I learned of these properties I can start to audit code and insert
>> barriers where I believe they are needed, but this simply means that
>> almost all occurrences of atomic_read will get barriers (unless there
>> already are implicit but more or less obvious barriers like msleep).
> 
> You might find that these places that appear to need barriers are
> buggy for other reasons anyway. Can you point to some in-tree code
> we can have a look at?

I could, or could not, if I were through with auditing the code.  I
remembered one case and posted it (nodemgr_host_thread) which was safe
because msleep_interruptible provided the necessary barrier there, and
this implicit barrier is not in danger to be removed by future patches.
-- 
Stefan Richter
-=====-=-=== =--- =---=
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

I wrote:
> Nick Piggin wrote:
>> You might find that these places that appear to need barriers are
>> buggy for other reasons anyway. Can you point to some in-tree code
>> we can have a look at?
> 
> I could, or could not, if I were through with auditing the code.  I
> remembered one case and posted it (nodemgr_host_thread) which was safe
> because msleep_interruptible provided the necessary barrier there, and
> this implicit barrier is not in danger to be removed by future patches.

PS, just in case anybody holds his breath for more example code from me,
I don't plan to continue with an actual audit of the drivers I maintain.
It's an important issue, but my current time budget will restrict me to
look at it ad hoc, per case.  (Open bugs have higher priority than
potential bugs.)
-- 
Stefan Richter
-=====-=-=== =--- =---=
http://arcgraph.de/sr/

LDD3 pitfalls (was Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures)

[Stefan Richter]

Nick Piggin wrote:
> Stefan Richter wrote:
>> Nick Piggin wrote:
>>
>>> I don't know why people would assume volatile of atomics. AFAIK, most
>>> of the documentation is pretty clear that all the atomic stuff can be
>>> reordered etc. except for those that modify and return a value.
>>
>>
>> Which documentation is there?
> 
> Documentation/atomic_ops.txt
> 
> 
>> For driver authors, there is LDD3.  It doesn't specifically cover
>> effects of optimization on accesses to atomic_t.
>>
>> For architecture port authors, there is Documentation/atomic_ops.txt.
>> Driver authors also can learn something from that document, as it
>> indirectly documents the atomic_t and bitops APIs.
>>
> 
> "Semantics and Behavior of Atomic and Bitmask Operations" is
> pretty direct :)
> 
> Sure, it says that it's for arch maintainers, but there is no
> reason why users can't make use of it.


Note, LDD3 page 238 says:  "It is worth noting that most of the other
kernel primitives dealing with synchronization, such as spinlock and
atomic_t operations, also function as memory barriers."

I don't know about Linux 2.6.10 against which LDD3 was written, but
currently only _some_ atomic_t operations function as memory barriers.

Besides, judging from some posts in this thread, saying that atomic_t
operations dealt with synchronization may not be entirely precise.
-- 
Stefan Richter
-=====-=-=== =--- =--=-
http://arcgraph.de/sr/

Re: LDD3 pitfalls (was Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures)

[Chris Snook]

Stefan Richter wrote:
> Nick Piggin wrote:
>> Stefan Richter wrote:
>>> Nick Piggin wrote:
>>>
>>>> I don't know why people would assume volatile of atomics. AFAIK, most
>>>> of the documentation is pretty clear that all the atomic stuff can be
>>>> reordered etc. except for those that modify and return a value.
>>>
>>> Which documentation is there?
>> Documentation/atomic_ops.txt
>>
>>
>>> For driver authors, there is LDD3.  It doesn't specifically cover
>>> effects of optimization on accesses to atomic_t.
>>>
>>> For architecture port authors, there is Documentation/atomic_ops.txt.
>>> Driver authors also can learn something from that document, as it
>>> indirectly documents the atomic_t and bitops APIs.
>>>
>> "Semantics and Behavior of Atomic and Bitmask Operations" is
>> pretty direct :)
>>
>> Sure, it says that it's for arch maintainers, but there is no
>> reason why users can't make use of it.
> 
> 
> Note, LDD3 page 238 says:  "It is worth noting that most of the other
> kernel primitives dealing with synchronization, such as spinlock and
> atomic_t operations, also function as memory barriers."
> 
> I don't know about Linux 2.6.10 against which LDD3 was written, but
> currently only _some_ atomic_t operations function as memory barriers.
> 
> Besides, judging from some posts in this thread, saying that atomic_t
> operations dealt with synchronization may not be entirely precise.

atomic_t is often used as the basis for implementing more sophisticated 
synchronization mechanisms, such as rwlocks.  Whether or not they are designed 
for that purpose, the atomic_* operations are de facto synchronization primitives.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> (and yes, it is perfectly legitimate to
> want a non-volatile read for a data type that you also want to do
> atomic RMW operations on)

...which is undefined behaviour in C (and GCC) when that data is
declared volatile, which is a good argument against implementing
atomics that way in itself.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> So the point here is that if you don't mind getting a stale
> value from the CPU cache when doing an atomic_read, then
> surely you won't mind getting a stale value from the compiler
> "cache".

No, that particular argument is bogus, because there is a cache
coherency protocol operating to keep the CPU cache coherent with
stores from other CPUs, but there isn't any such protocol (nor should
there be) for a register used as a "cache".

(Linux requires SMP systems to keep any CPU caches coherent as far as
accesses by other CPUs are concerned.  It doesn't support any SMP
systems that are not cache-coherent as far as CPU accesses are
concerned.  It does support systems with non-cache-coherent DMA.)

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 11:54:44AM +0200, Stefan Richter wrote:
> 
> One example was discussed here earlier:  The for (;;) loop in
> nodemgr_host_thread.  There an msleep_interruptible implicitly acted as
> barrier (at the moment because it's in a different translation unit; if
> it were the same, then because it hopefully has own barriers).  So that
> happens to work, although such an implicit barrier is bad style:  Better
> enforce the desired behaviour (== guaranteed load operation) *explicitly*.

Hmm, it's not bad style at all.  Let's assume that everything
is in the same scope.  Such a loop must either call a function
that busy-waits, which should always have a cpu_relax or
something equivalent, or it'll call a function that schedules
away which immediately invalidates any values the compiler might
have cached for the atomic_read.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 10:06:31AM +0200, Stefan Richter wrote:
>>> Do you (or anyone else for that matter) have an example of this?
>> The only code I somewhat know, the ieee1394 subsystem, was perhaps
>> authored and is currently maintained with the expectation that each
>> occurrence of atomic_read actually results in a load operation, i.e. is
>> not optimized away.  This means all atomic_t (bus generation, packet and
>> buffer refcounts, and some other state variables)* and likewise all
>> atomic bitops in that subsystem.
> 
> Can you find an actual atomic_read code snippet there that is
> broken without the volatile modifier?

A whole bunch of atomic_read uses will be broken without the volatile 
modifier once we start removing barriers that aren't needed if volatile 
behavior is guaranteed.

barrier() clobbers all your registers.  volatile atomic_read() only 
clobbers one register, and more often than not it's a register you 
wanted to clobber anyway.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 03:48:54PM -0400, Chris Snook wrote:
>
> >Can you find an actual atomic_read code snippet there that is
> >broken without the volatile modifier?
> 
> A whole bunch of atomic_read uses will be broken without the volatile 
> modifier once we start removing barriers that aren't needed if volatile 
> behavior is guaranteed.

Could you please cite the file/function names so we can
see whether removing the barrier makes sense?

Thanks,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 03:48:54PM -0400, Chris Snook wrote:
>>> Can you find an actual atomic_read code snippet there that is
>>> broken without the volatile modifier?
>> A whole bunch of atomic_read uses will be broken without the volatile 
>> modifier once we start removing barriers that aren't needed if volatile 
>> behavior is guaranteed.
> 
> Could you please cite the file/function names so we can
> see whether removing the barrier makes sense?
> 
> Thanks,

At a glance, several architectures' implementations of smp_call_function() have 
one or more legitimate atomic_read() busy-waits that shouldn't be using 
CPU-relax.  Some of them do work in the loop.

I'm sure there are plenty more examples that various maintainers could find in 
their own code.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 10:04:24PM -0400, Chris Snook wrote:
>
> >Could you please cite the file/function names so we can
> >see whether removing the barrier makes sense?
> 
> At a glance, several architectures' implementations of smp_call_function() 
> have one or more legitimate atomic_read() busy-waits that shouldn't be 
> using CPU-relax.  Some of them do work in the loop.

Care to name one so we can discuss it?

Thanks,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Chris Snook wrote:
> Herbert Xu wrote:
> 
>> On Thu, Aug 16, 2007 at 03:48:54PM -0400, Chris Snook wrote:
>>
>>>> Can you find an actual atomic_read code snippet there that is
>>>> broken without the volatile modifier?
>>>
>>> A whole bunch of atomic_read uses will be broken without the volatile 
>>> modifier once we start removing barriers that aren't needed if 
>>> volatile behavior is guaranteed.
>>
>>
>> Could you please cite the file/function names so we can
>> see whether removing the barrier makes sense?
>>
>> Thanks,
> 
> 
> At a glance, several architectures' implementations of 
> smp_call_function() have one or more legitimate atomic_read() busy-waits 
> that shouldn't be using CPU-relax.  Some of them do work in the loop.

sh looks like the only one there that would be broken (and that's only
because they don't have a cpu_relax there, but it should be added anyway).
Sure, if we removed volatile from other architectures, it would be wise
to audit arch code because arch maintainers do sometimes make assumptions
about their implementation details... however we can assume most generic
code is safe without volatile.

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> Can you find an actual atomic_read code snippet there that is
> broken without the volatile modifier?

There are some in arch-specific code, for example line 1073 of
arch/mips/kernel/smtc.c.  On mips, cpu_relax() is just barrier(), so
the empty loop body is ok provided that atomic_read actually does the
load each time around the loop.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Fri, Aug 17, 2007 at 03:09:57PM +1000, Paul Mackerras wrote:
> Herbert Xu writes:
> 
> > Can you find an actual atomic_read code snippet there that is
> > broken without the volatile modifier?
> 
> There are some in arch-specific code, for example line 1073 of
> arch/mips/kernel/smtc.c.  On mips, cpu_relax() is just barrier(), so
> the empty loop body is ok provided that atomic_read actually does the
> load each time around the loop.

A barrier() is all you need to force the compiler to reread
the value.

The people advocating volatile in this thread are talking
about code that doesn't use barrier()/cpu_relax().

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Herbert Xu writes:

> On Fri, Aug 17, 2007 at 03:09:57PM +1000, Paul Mackerras wrote:
> > Herbert Xu writes:
> > 
> > > Can you find an actual atomic_read code snippet there that is
> > > broken without the volatile modifier?
> > 
> > There are some in arch-specific code, for example line 1073 of
> > arch/mips/kernel/smtc.c.  On mips, cpu_relax() is just barrier(), so
> > the empty loop body is ok provided that atomic_read actually does the
> > load each time around the loop.
> 
> A barrier() is all you need to force the compiler to reread
> the value.
> 
> The people advocating volatile in this thread are talking
> about code that doesn't use barrier()/cpu_relax().

Did you look at it?  Here it is:

	/* Someone else is initializing in parallel - let 'em finish */
	while (atomic_read(&idle_hook_initialized) < 1000)
		;

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Paul Mackerras wrote:

> Herbert Xu writes:
> 
> > On Fri, Aug 17, 2007 at 03:09:57PM +1000, Paul Mackerras wrote:
> > > Herbert Xu writes:
> > > 
> > > > Can you find an actual atomic_read code snippet there that is
> > > > broken without the volatile modifier?
> > > 
> > > There are some in arch-specific code, for example line 1073 of
> > > arch/mips/kernel/smtc.c.  On mips, cpu_relax() is just barrier(), so
> > > the empty loop body is ok provided that atomic_read actually does the
> > > load each time around the loop.
> > 
> > A barrier() is all you need to force the compiler to reread
> > the value.
> > 
> > The people advocating volatile in this thread are talking
> > about code that doesn't use barrier()/cpu_relax().
> 
> Did you look at it?  Here it is:
> 
> 	/* Someone else is initializing in parallel - let 'em finish */
> 	while (atomic_read(&idle_hook_initialized) < 1000)
> 		;


Honestly, this thread is suffering from HUGE communication gaps.

What Herbert (obviously) meant there was that "this loop could've
been okay _without_ using volatile-semantics-atomic_read() also, if
only it used cpu_relax()".

That does work, because cpu_relax() is _at least_ barrier() on all
archs (on some it also emits some arch-dependent "pause" kind of
instruction).

Now, saying that "MIPS does not have such an instruction so I won't
use cpu_relax() for arch-dependent-busy-while-loops in arch/mips/"
sounds like a wrong argument, because: tomorrow, such arch's _may_
introduce such an instruction, so naturally, at that time we'd
change cpu_relax() appropriately (in reality, we would actually
*re-define* cpu_relax() and ensure that the correct version gets
pulled in depending on whether the callsite code is legacy or only
for the newer such CPUs of said arch, whatever), but loops such as
this would remain un-changed, because they never used cpu_relax()!

OTOH an argument that said the following would've made a stronger case:

"I don't want to use cpu_relax() because that's a full memory
clobber barrier() and I have loop-invariants / other variables
around in that code that I *don't* want the compiler to forget
just because it used cpu_relax(), and hence I will not use
cpu_relax() but instead make my atomic_read() itself have
"volatility" semantics. Not just that, but I will introduce a
cpu_relax_no_barrier() on MIPS, that would be a no-op #define
for now, but which may not be so forever, and continue to use
that in such busy loops."

In general, please read the thread-summary I've tried to do at:
http://lkml.org/lkml/2007/8/17/25
Feel free to continue / comment / correct stuff from there, there's
too much confusion and circular-arguments happening on this thread
otherwise.

[ I might've made an incorrect statement there about
  "volatile" w.r.t. cache on non-x86 archs, I think. ]


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Ilpo Järvinen]

On Thu, 16 Aug 2007, Herbert Xu wrote:

> We've been through that already.  If it's a busy-wait it
> should use cpu_relax. 

I looked around a bit by using some command lines and ended up wondering 
if these are equal to busy-wait case (and should be fixed) or not:

./drivers/telephony/ixj.c
6674:   while (atomic_read(&j->DSPWrite) > 0)
6675-           atomic_dec(&j->DSPWrite);

...besides that, there are couple of more similar cases in the same file 
(with braces)...


-- 
 i.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Ilpo Järvinen wrote:
> I looked around a bit by using some command lines and ended up wondering 
> if these are equal to busy-wait case (and should be fixed) or not:
> 
> ./drivers/telephony/ixj.c
> 6674:   while (atomic_read(&j->DSPWrite) > 0)
> 6675-           atomic_dec(&j->DSPWrite);
> 
> ...besides that, there are couple of more similar cases in the same file 
> (with braces)...

Generally, ixj.c has several occurrences of couples of atomic write and
atomic read which potentially do not do what the author wanted.
-- 
Stefan Richter
-=====-=-=== =--- =----
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Ilpo Järvinen wrote:
> On Thu, 16 Aug 2007, Herbert Xu wrote:
> 
>> We've been through that already.  If it's a busy-wait it
>> should use cpu_relax. 
> 
> I looked around a bit by using some command lines and ended up wondering 
> if these are equal to busy-wait case (and should be fixed) or not:
> 
> ./drivers/telephony/ixj.c
> 6674:   while (atomic_read(&j->DSPWrite) > 0)
> 6675-           atomic_dec(&j->DSPWrite);
> 
> ...besides that, there are couple of more similar cases in the same file 
> (with braces)...

atomic_dec() already has volatile behavior everywhere, so this is 
semantically okay, but this code (and any like it) should be calling 
cpu_relax() each iteration through the loop, unless there's a compelling 
reason not to.  I'll allow that for some hardware drivers (possibly this 
one) such a compelling reason may exist, but hardware-independent core 
subsystems probably have no excuse.

If the maintainer of this code doesn't see a compelling reason to add 
cpu_relax() in this loop, then it should be patched.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Chris Snook wrote:

> atomic_dec() already has volatile behavior everywhere, so this is semantically
> okay, but this code (and any like it) should be calling cpu_relax() each
> iteration through the loop, unless there's a compelling reason not to.  I'll
> allow that for some hardware drivers (possibly this one) such a compelling
> reason may exist, but hardware-independent core subsystems probably have no
> excuse.

No it does not have any volatile semantics. atomic_dec() can be reordered 
at will by the compiler within the current basic unit if you do not add a 
barrier.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 01:20:26PM -0700, Christoph Lameter wrote:
> On Thu, 16 Aug 2007, Chris Snook wrote:
> 
> > atomic_dec() already has volatile behavior everywhere, so this is semantically
> > okay, but this code (and any like it) should be calling cpu_relax() each
> > iteration through the loop, unless there's a compelling reason not to.  I'll
> > allow that for some hardware drivers (possibly this one) such a compelling
> > reason may exist, but hardware-independent core subsystems probably have no
> > excuse.
> 
> No it does not have any volatile semantics. atomic_dec() can be reordered 
> at will by the compiler within the current basic unit if you do not add a 
> barrier.

Yep.  Or you can use atomic_dec_return() instead of using a barrier.

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Thu, Aug 16, 2007 at 06:02:32PM -0700, Paul E. McKenney wrote:
> 
> Yep.  Or you can use atomic_dec_return() instead of using a barrier.

Or you could use smp_mb__{before,after}_atomic_dec.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 09:28:00AM +0800, Herbert Xu wrote:
> On Thu, Aug 16, 2007 at 06:02:32PM -0700, Paul E. McKenney wrote:
> > 
> > Yep.  Or you can use atomic_dec_return() instead of using a barrier.
> 
> Or you could use smp_mb__{before,after}_atomic_dec.

Yep.  That would be an example of a barrier, either in the
atomic_dec() itself or in the smp_mb...().

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Christoph Lameter writes:

> No it does not have any volatile semantics. atomic_dec() can be reordered 
> at will by the compiler within the current basic unit if you do not add a 
> barrier.

Volatile doesn't mean it can't be reordered; volatile means the
accesses can't be eliminated.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Fri, 17 Aug 2007, Paul Mackerras wrote:
>
> Volatile doesn't mean it can't be reordered; volatile means the
> accesses can't be eliminated.

It also does limit re-ordering. 

Of course, since *normal* accesses aren't necessarily limited wrt 
re-ordering, the question then becomes one of "with regard to *what* does 
it limit re-ordering?".

A C compiler that re-orders two different volatile accesses that have a 
sequence point in between them is pretty clearly a buggy compiler. So at a 
minimum, it limits re-ordering wrt other volatiles (assuming sequence 
points exists). It also means that the compiler cannot move it 
speculatively across conditionals, but other than that it's starting to 
get fuzzy.

In general, I'd *much* rather we used barriers. Anything that "depends" on 
volatile is pretty much set up to be buggy. But I'm certainly also willing 
to have that volatile inside "atomic_read/atomic_set()" if it avoids code 
that would otherwise break - ie if it hides a bug.

		Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Linus Torvalds writes:

> In general, I'd *much* rather we used barriers. Anything that "depends" on 
> volatile is pretty much set up to be buggy. But I'm certainly also willing 
> to have that volatile inside "atomic_read/atomic_set()" if it avoids code 
> that would otherwise break - ie if it hides a bug.

The cost of doing so seems to me to be well down in the noise - 44
bytes of extra kernel text on a ppc64 G5 config, and I don't believe
the extra few cycles for the occasional extra load would be measurable
(they should all hit in the L1 dcache).  I don't mind if x86[-64] have
atomic_read/set be nonvolatile and find all the missing barriers, but
for now on powerpc, I think that not having to find those missing
barriers is worth the 0.00076% increase in kernel text size.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Fri, Aug 17, 2007 at 01:43:27PM +1000, Paul Mackerras wrote:
>
> The cost of doing so seems to me to be well down in the noise - 44
> bytes of extra kernel text on a ppc64 G5 config, and I don't believe
> the extra few cycles for the occasional extra load would be measurable
> (they should all hit in the L1 dcache).  I don't mind if x86[-64] have
> atomic_read/set be nonvolatile and find all the missing barriers, but
> for now on powerpc, I think that not having to find those missing
> barriers is worth the 0.00076% increase in kernel text size.

BTW, the sort of missing barriers that triggered this thread
aren't that subtle.  It'll result in a simple lock-up if the
loop condition holds upon entry.  At which point it's fairly
straightforward to find the culprit.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Herbert Xu wrote:

> On Fri, Aug 17, 2007 at 01:43:27PM +1000, Paul Mackerras wrote:
> >
> > The cost of doing so seems to me to be well down in the noise - 44
> > bytes of extra kernel text on a ppc64 G5 config, and I don't believe
> > the extra few cycles for the occasional extra load would be measurable
> > (they should all hit in the L1 dcache).  I don't mind if x86[-64] have
> > atomic_read/set be nonvolatile and find all the missing barriers, but
> > for now on powerpc, I think that not having to find those missing
> > barriers is worth the 0.00076% increase in kernel text size.
> 
> BTW, the sort of missing barriers that triggered this thread
> aren't that subtle.  It'll result in a simple lock-up if the
> loop condition holds upon entry.  At which point it's fairly
> straightforward to find the culprit.

Not necessarily. A barrier-less buggy code such as below:

	atomic_set(&v, 0);

	... /* some initial code */

	while (atomic_read(&v))
		;

	... /* code that MUST NOT be executed unless v becomes non-zero */

(where v->counter is has no volatile access semantics)

could be generated by the compiler to simply *elid* or *do away* with
the loop itself, thereby making the:

"/* code that MUST NOT be executed unless v becomes non-zero */"

to be executed even when v is zero! That is subtle indeed, and causes
no hard lockups.

Granted, the above IS buggy code. But, the stated objective is to avoid
heisenbugs. And we have driver / subsystem maintainers such as Stefan
coming up and admitting that often a lot of code that's written to use
atomic_read() does assume the read will not be elided by the compiler.

See, I agree, "volatility" semantics != what we often want. However, if
what we want is compiler barrier, for only the object under consideration,
"volatility" semantics aren't really "nonsensical" or anything.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:
> 
> On Fri, 17 Aug 2007, Herbert Xu wrote:
> 
> 
>>On Fri, Aug 17, 2007 at 01:43:27PM +1000, Paul Mackerras wrote:
>>
>>BTW, the sort of missing barriers that triggered this thread
>>aren't that subtle.  It'll result in a simple lock-up if the
>>loop condition holds upon entry.  At which point it's fairly
>>straightforward to find the culprit.
> 
> 
> Not necessarily. A barrier-less buggy code such as below:
> 
> 	atomic_set(&v, 0);
> 
> 	... /* some initial code */
> 
> 	while (atomic_read(&v))
> 		;
> 
> 	... /* code that MUST NOT be executed unless v becomes non-zero */
> 
> (where v->counter is has no volatile access semantics)
> 
> could be generated by the compiler to simply *elid* or *do away* with
> the loop itself, thereby making the:
> 
> "/* code that MUST NOT be executed unless v becomes non-zero */"
> 
> to be executed even when v is zero! That is subtle indeed, and causes
> no hard lockups.

Then I presume you mean

while (!atomic_read(&v))
     ;

Which is just the same old infinite loop bug solved with cpu_relax().
These are pretty trivial to audit and fix, and also to debug, I would
think.


> Granted, the above IS buggy code. But, the stated objective is to avoid
> heisenbugs.

Anyway, why are you making up code snippets that are buggy in other
ways in order to support this assertion being made that lots of kernel
code supposedly depends on volatile semantics. Just reference the
actual code.


> And we have driver / subsystem maintainers such as Stefan
> coming up and admitting that often a lot of code that's written to use
> atomic_read() does assume the read will not be elided by the compiler.

So these are broken on i386 and x86-64?

Are they definitely safe on SMP and weakly ordered machines with
just a simple compiler barrier there? Because I would not be
surprised if there are a lot of developers who don't really know
what to assume when it comes to memory ordering issues.

This is not a dig at driver writers: we still have memory ordering
problems in the VM too (and probably most of the subtle bugs in
lockless VM code are memory ordering ones). Let's not make up a
false sense of security and hope that sprinkling volatile around
will allow people to write bug-free lockless code. If a writer
can't be bothered reading API documentation and learning the Linux
memory model, they can still be productive writing safely locked
code.

-- 
SUSE Labs, Novell Inc.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> [...]
> > Granted, the above IS buggy code. But, the stated objective is to avoid
> > heisenbugs.
    ^^^^^^^^^^

> Anyway, why are you making up code snippets that are buggy in other
> ways in order to support this assertion being made that lots of kernel
> code supposedly depends on volatile semantics. Just reference the
> actual code.

Because the point is *not* about existing bugs in kernel code. At some
point Chris Snook (who started this thread) did write that "If I knew
of the existing bugs in the kernel, I would be sending patches for them,
not this series" or something to that effect.

The point is about *author expecations*. If people do expect atomic_read()
(or a variant thereof) to have volatile semantics, why not give them such
a variant?

And by the way, the point is *also* about the fact that cpu_relax(), as
of today, implies a full memory clobber, which is not what a lot of such
loops want. (due to stuff mentioned elsewhere, summarized in that summary)


> > And we have driver / subsystem maintainers such as Stefan
> > coming up and admitting that often a lot of code that's written to use
> > atomic_read() does assume the read will not be elided by the compiler.
                                                             ^^^^^^^^^^^^^

(so it's about compiler barrier expectations only, though I fully agree
that those who're using atomic_t as if it were some magic thing that lets
them write lockless code are sorrily mistaken.)

> So these are broken on i386 and x86-64?

Possibly, but the point is not about existing bugs, as mentioned above.

Some such bugs have been found nonetheless -- reminds me, can somebody
please apply http://www.gossamer-threads.com/lists/linux/kernel/810674 ?


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:
> 
> On Fri, 17 Aug 2007, Nick Piggin wrote:
> 
> 
>>Satyam Sharma wrote:
>>[...]
>>
>>>Granted, the above IS buggy code. But, the stated objective is to avoid
>>>heisenbugs.
> 
>     ^^^^^^^^^^
> 
> 
>>Anyway, why are you making up code snippets that are buggy in other
>>ways in order to support this assertion being made that lots of kernel
>>code supposedly depends on volatile semantics. Just reference the
>>actual code.
> 
> 
> Because the point is *not* about existing bugs in kernel code. At some
> point Chris Snook (who started this thread) did write that "If I knew
> of the existing bugs in the kernel, I would be sending patches for them,
> not this series" or something to that effect.
> 
> The point is about *author expecations*. If people do expect atomic_read()
> (or a variant thereof) to have volatile semantics, why not give them such
> a variant?

Because they should be thinking about them in terms of barriers, over
which the compiler / CPU is not to reorder accesses or cache memory
operations, rather than "special" "volatile" accesses. Linux's whole
memory ordering and locking model is completely geared around the
former.


> And by the way, the point is *also* about the fact that cpu_relax(), as
> of today, implies a full memory clobber, which is not what a lot of such
> loops want. (due to stuff mentioned elsewhere, summarized in that summary)

That's not the point, because as I also mentioned, the logical extention
to Linux's barrier API to handle this is the order(x) macro. Again, not
special volatile accessors.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> > [...]
> > The point is about *author expecations*. If people do expect atomic_read()
> > (or a variant thereof) to have volatile semantics, why not give them such
> > a variant?
> 
> Because they should be thinking about them in terms of barriers, over
> which the compiler / CPU is not to reorder accesses or cache memory
> operations, rather than "special" "volatile" accesses.

This is obviously just a taste thing. Whether to have that forget(x)
barrier as something author should explicitly sprinkle appropriately
in appropriate places in the code by himself or use a primitive that
includes it itself.

I'm not saying "taste matters aren't important" (they are), but I'm really
skeptical if most folks would find the former tasteful.

> > And by the way, the point is *also* about the fact that cpu_relax(), as
> > of today, implies a full memory clobber, which is not what a lot of such
> > loops want. (due to stuff mentioned elsewhere, summarized in that summary)
> 
> That's not the point,

That's definitely the point, why not. This is why "barrier()", being
heavy-handed, is not the best option.

> because as I also mentioned, the logical extention
> to Linux's barrier API to handle this is the order(x) macro. Again, not
> special volatile accessors.

Sure, that forget(x) macro _is_ proposed to be made part of the generic
API. Doesn't explain why not to define/use primitives that has volatility
semantics in itself, though (taste matters apart).

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Nick Piggin]

Satyam Sharma wrote:

>
>On Fri, 17 Aug 2007, Nick Piggin wrote:
>
>
>>Because they should be thinking about them in terms of barriers, over
>>which the compiler / CPU is not to reorder accesses or cache memory
>>operations, rather than "special" "volatile" accesses.
>>
>
>This is obviously just a taste thing. Whether to have that forget(x)
>barrier as something author should explicitly sprinkle appropriately
>in appropriate places in the code by himself or use a primitive that
>includes it itself.
>

That's not obviously just taste to me. Not when the primitive has many
(perhaps, the majority) of uses that do not require said barriers. And
this is not solely about the code generation (which, as Paul says, is
relatively minor even on x86). I prefer people to think explicitly
about barriers in their lockless code.


>I'm not saying "taste matters aren't important" (they are), but I'm really
>skeptical if most folks would find the former tasteful.
>

So I /do/ have better taste than most folks? Thanks! :-)


>>>And by the way, the point is *also* about the fact that cpu_relax(), as
>>>of today, implies a full memory clobber, which is not what a lot of such
>>>loops want. (due to stuff mentioned elsewhere, summarized in that summary)
>>>
>>That's not the point,
>>
>
>That's definitely the point, why not. This is why "barrier()", being
>heavy-handed, is not the best option.
>

That is _not_ the point (of why a volatile atomic_read is good) because 
there
has already been an alternative posted that better conforms with Linux 
barrier
API and is much more widely useful and more usable. If you are so 
worried about
barrier() being too heavyweight, then you're off to a poor start by 
wanting to
add a few K of kernel text by making atomic_read volatile.


>>because as I also mentioned, the logical extention
>>to Linux's barrier API to handle this is the order(x) macro. Again, not
>>special volatile accessors.
>>
>
>Sure, that forget(x) macro _is_ proposed to be made part of the generic
>API. Doesn't explain why not to define/use primitives that has volatility
>semantics in itself, though (taste matters apart).
>

If you follow the discussion.... You were thinking of a reason why the
semantics *should* be changed or added, and I was rebutting your argument
that it must be used when a full barrier() is too heavy (ie. by pointing
out that order() has superior semantics anyway).

Why do I keep repeating the same things? I'll not continue bloating this
thread until a new valid point comes up...

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Nick Piggin wrote:

> Satyam Sharma wrote:
> 
> > On Fri, 17 Aug 2007, Nick Piggin wrote:
> > 
> > > Because they should be thinking about them in terms of barriers, over
> > > which the compiler / CPU is not to reorder accesses or cache memory
> > > operations, rather than "special" "volatile" accesses.
> > 
> > This is obviously just a taste thing. Whether to have that forget(x)
> > barrier as something author should explicitly sprinkle appropriately
> > in appropriate places in the code by himself or use a primitive that
> > includes it itself.
> 
> That's not obviously just taste to me. Not when the primitive has many
> (perhaps, the majority) of uses that do not require said barriers. And
> this is not solely about the code generation (which, as Paul says, is
> relatively minor even on x86).

See, you do *require* people to have to repeat the same things to you!

As has been written about enough times already, and if you followed the
discussion on this thread, I am *not* proposing that atomic_read()'s
semantics be changed to have any extra barriers. What is proposed is a
different atomic_read_xxx() variant thereof, that those can use who do
want that.

Now whether to have a kind of barrier ("volatile", whatever) in the
atomic_read_xxx() itself, or whether to make the code writer himself to
explicitly write the order(x) appropriately in appropriate places in the
code _is_ a matter of taste.


> > That's definitely the point, why not. This is why "barrier()", being
> > heavy-handed, is not the best option.
> 
> That is _not_ the point [...]

Again, you're requiring me to repeat things that were already made evident
on this thread (if you follow it).

This _is_ the point, because a lot of loops out there (too many of them,
I WILL NOT bother citing file_name:line_number) end up having to use a
barrier just because they're using a loop-exit-condition that depends
on a value returned by atomic_read(). It would be good for them if they
used an atomic_read_xxx() primitive that gave these "volatility" semantics
without junking compiler optimizations for other memory references.

> because there has already been an alternative posted

Whether that alternative (explicitly using forget(x), or wrappers thereof,
such as the "order_atomic" you proposed) is better than other alternatives
(such as atomic_read_xxx() which includes the volatility behaviour in
itself) is still open, and precisely what we started discussing just one
mail back.

(The above was also mostly stuff I had to repeated for you, sadly.)

> that better conforms with Linux barrier
> API and is much more widely useful and more usable.

I don't think so.

(Now *this* _is_ the "taste-dependent matter" that I mentioned earlier.)

> If you are so worried
> about
> barrier() being too heavyweight, then you're off to a poor start by wanting to
> add a few K of kernel text by making atomic_read volatile.

Repeating myself, for the N'th time, NO, I DON'T want to make atomic_read
have "volatile" semantics.

> > > because as I also mentioned, the logical extention
> > > to Linux's barrier API to handle this is the order(x) macro. Again, not
> > > special volatile accessors.
> > 
> > Sure, that forget(x) macro _is_ proposed to be made part of the generic
> > API. Doesn't explain why not to define/use primitives that has volatility
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > semantics in itself, though (taste matters apart).
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> If you follow the discussion.... You were thinking of a reason why the
> semantics *should* be changed or added, and I was rebutting your argument
> that it must be used when a full barrier() is too heavy (ie. by pointing
> out that order() has superior semantics anyway).

Amazing. Either you have reading comprehension problems, or else, please
try reading this thread (or at least this sub-thread) again. I don't want
_you_ blaming _me_ for having to repeat things to you all over again.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Fri, 17 Aug 2007, Nick Piggin wrote:
> 
> That's not obviously just taste to me. Not when the primitive has many
> (perhaps, the majority) of uses that do not require said barriers. And
> this is not solely about the code generation (which, as Paul says, is
> relatively minor even on x86). I prefer people to think explicitly
> about barriers in their lockless code.

Indeed.

I think the important issues are:

 - "volatile" itself is simply a badly/weakly defined issue. The semantics 
   of it as far as the compiler is concerned are really not very good, and 
   in practice tends to boil down to "I will generate so bad code that 
   nobody can accuse me of optimizing anything away".

 - "volatile" - regardless of how well or badly defined it is - is purely 
   a compiler thing. It has absolutely no meaning for the CPU itself, so 
   it at no point implies any CPU barriers. As a result, even if the 
   compiler generates crap code and doesn't re-order anything, there's 
   nothing that says what the CPU will do.

 - in other words, the *only* possible meaning for "volatile" is a purely 
   single-CPU meaning. And if you only have a single CPU involved in the 
   process, the "volatile" is by definition pointless (because even 
   without a volatile, the compiler is required to make the C code appear 
   consistent as far as a single CPU is concerned).

So, let's take the example *buggy* code where we use "volatile" to wait 
for other CPU's:

	atomic_set(&var, 0);
	while (!atomic_read(&var))
		/* nothing */;


which generates an endless loop if we don't have atomic_read() imply 
volatile.

The point here is that it's buggy whether the volatile is there or not! 
Exactly because the user expects multi-processing behaviour, but 
"volatile" doesn't actually give any real guarantees about it. Another CPU 
may have done:

	external_ptr = kmalloc(..);
	/* Setup is now complete, inform the waiter */
	atomic_inc(&var);

but the fact is, since the other CPU isn't serialized in any way, the 
"while-loop" (even in the presense of "volatile") doesn't actually work 
right! Whatever the "atomic_read()" was waiting for may not have 
completed, because we have no barriers!

So if "volatile" makes a difference, it is invariably a sign of a bug in 
serialization (the one exception is for IO - we use "volatile" to avoid 
having to use inline asm for IO on x86) - and for "random values" like 
jiffies).

So the question should *not* be whether "volatile" actually fixes bugs. It 
*never* fixes a bug. But what it can do is to hide the obvious ones. In 
other words, adding a volaile in the above kind of situation of 
"atomic_read()" will certainly turn an obvious bug into something that 
works "practically all of the time).

So anybody who argues for "volatile" fixing bugs is fundamentally 
incorrect. It does NO SUCH THING. By arguing that, such people only show 
that you have no idea what they are talking about.

So the only reason to add back "volatile" to the atomic_read() sequence is 
not to fix bugs, but to _hide_ the bugs better. They're still there, they 
are just a lot harder to trigger, and tend to be a lot subtler.

And hey, sometimes "hiding bugs well enough" is ok. In this case, I'd 
argue that we've successfully *not* had the volatile there for eight 
months on x86-64, and that should tell people something. 

(Does _removing_ the volatile fix bugs? No - callers still need to think 
about barriers etc, and lots of people don't. So I'm not claiming that 
removing volatile fixes any bugs either, but I *am* claiming that:

 - removing volatile makes some bugs easier to see (which is mostly a good 
   thing: they were there before, anyway).

 - removing volatile generates better code (which is a good thing, even if 
   it's just 0.1%)

 - removing volatile removes a huge mental *bug* that lots of people seem 
   to have, as shown by this whole thread. Anybody who thinks that 
   "volatile" actually fixes anything has a gaping hole in their head, and 
   we should remove volatile just to make sure that nobody thinks that it 
   means soemthign that it doesn't mean!

In other words, this whole discussion has just convinced me that we should 
*not* add back "volatile" to "atomic_read()" - I was willing to do it for 
practical and "hide the bugs" reasons, but having seen people argue for 
it, thinking that it actually fixes something, I'm now convinced that the 
*last* thing we should do is to encourage that kind of superstitious 
thinking.

"volatile" is like a black cat crossing the road. Sure, it affects 
*something* (at a minimum: before, the black cat was on one side of the 
road, afterwards it is on the other side of the road), but it has no 
bigger and longer-lasting direct affects. 

People who think "volatile" really matters are just fooling themselves.

		Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Friesen]

Linus Torvalds wrote:

>  - in other words, the *only* possible meaning for "volatile" is a purely 
>    single-CPU meaning. And if you only have a single CPU involved in the 
>    process, the "volatile" is by definition pointless (because even 
>    without a volatile, the compiler is required to make the C code appear 
>    consistent as far as a single CPU is concerned).

I assume you mean "except for IO-related code and 'random' values like 
jiffies" as you mention later on?  I assume other values set in 
interrupt handlers would count as "random" from a volatility perspective?

> So anybody who argues for "volatile" fixing bugs is fundamentally 
> incorrect. It does NO SUCH THING. By arguing that, such people only show 
> that you have no idea what they are talking about.

What about reading values modified in interrupt handlers, as in your 
"random" case?  Or is this a bug where the user of atomic_read() is 
invalidly expecting a read each time it is called?

Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arjan van de Ven]

On Fri, 2007-08-17 at 12:50 -0600, Chris Friesen wrote:
> Linus Torvalds wrote:
> 
> >  - in other words, the *only* possible meaning for "volatile" is a purely 
> >    single-CPU meaning. And if you only have a single CPU involved in the 
> >    process, the "volatile" is by definition pointless (because even 
> >    without a volatile, the compiler is required to make the C code appear 
> >    consistent as far as a single CPU is concerned).
> 
> I assume you mean "except for IO-related code and 'random' values like 
> jiffies" as you mention later on?  I assume other values set in 
> interrupt handlers would count as "random" from a volatility perspective?
> 
> > So anybody who argues for "volatile" fixing bugs is fundamentally 
> > incorrect. It does NO SUCH THING. By arguing that, such people only show 
> > that you have no idea what they are talking about.
> 
> What about reading values modified in interrupt handlers, as in your 
> "random" case?  Or is this a bug where the user of atomic_read() is 
> invalidly expecting a read each time it is called?

the interrupt handler case is an SMP case since you do not know
beforehand what cpu your interrupt handler will run on.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 11:54:33AM -0700, Arjan van de Ven wrote:
> 
> On Fri, 2007-08-17 at 12:50 -0600, Chris Friesen wrote:
> > Linus Torvalds wrote:
> > 
> > >  - in other words, the *only* possible meaning for "volatile" is a purely 
> > >    single-CPU meaning. And if you only have a single CPU involved in the 
> > >    process, the "volatile" is by definition pointless (because even 
> > >    without a volatile, the compiler is required to make the C code appear 
> > >    consistent as far as a single CPU is concerned).
> > 
> > I assume you mean "except for IO-related code and 'random' values like 
> > jiffies" as you mention later on?  I assume other values set in 
> > interrupt handlers would count as "random" from a volatility perspective?
> > 
> > > So anybody who argues for "volatile" fixing bugs is fundamentally 
> > > incorrect. It does NO SUCH THING. By arguing that, such people only show 
> > > that you have no idea what they are talking about.
> > 
> > What about reading values modified in interrupt handlers, as in your 
> > "random" case?  Or is this a bug where the user of atomic_read() is 
> > invalidly expecting a read each time it is called?
> 
> the interrupt handler case is an SMP case since you do not know
> beforehand what cpu your interrupt handler will run on.

With the exception of per-CPU variables, yes.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arjan van de Ven]

On Fri, 2007-08-17 at 12:49 -0700, Paul E. McKenney wrote:
> > > What about reading values modified in interrupt handlers, as in your 
> > > "random" case?  Or is this a bug where the user of atomic_read() is 
> > > invalidly expecting a read each time it is called?
> > 
> > the interrupt handler case is an SMP case since you do not know
> > beforehand what cpu your interrupt handler will run on.
> 
> With the exception of per-CPU variables, yes.

if you're spinning waiting for a per-CPU variable to get changed by an
interrupt handler... you have bigger problems than "volatile" ;-)

-- 
if you want to mail me at work (you don't), use arjan (at) linux.intel.com
Test the interaction between Linux and your BIOS via http://www.linuxfirmwarekit.org

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Fri, Aug 17, 2007 at 12:49:00PM -0700, Arjan van de Ven wrote:
> 
> On Fri, 2007-08-17 at 12:49 -0700, Paul E. McKenney wrote:
> > > > What about reading values modified in interrupt handlers, as in your 
> > > > "random" case?  Or is this a bug where the user of atomic_read() is 
> > > > invalidly expecting a read each time it is called?
> > > 
> > > the interrupt handler case is an SMP case since you do not know
> > > beforehand what cpu your interrupt handler will run on.
> > 
> > With the exception of per-CPU variables, yes.
> 
> if you're spinning waiting for a per-CPU variable to get changed by an
> interrupt handler... you have bigger problems than "volatile" ;-)

That would be true, if you were doing that.  But you might instead be
simply making sure that the mainline actions were seen in order by the
interrupt handler.  My current example is the NMI-save rcu_read_lock()
implementation for realtime.  Not the common case, I will admit, but
still real.  ;-)

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Fri, 17 Aug 2007, Chris Friesen wrote:
> 
> I assume you mean "except for IO-related code and 'random' values like
> jiffies" as you mention later on?

Yes. There *are* valid uses for "volatile", but they have remained the 
same for the last few years:
 - "jiffies"
 - internal per-architecture IO implementations that can do them as normal 
   stores.

> I assume other values set in interrupt handlers would count as "random" 
> from a volatility perspective?

I don't really see any valid case. I can imagine that you have your own 
"jiffy" counter in a driver, but what's the point, really? I'd suggest not 
using volatile, and using barriers instead.

> 
> > So anybody who argues for "volatile" fixing bugs is fundamentally 
> > incorrect. It does NO SUCH THING. By arguing that, such people only 
> > show that you have no idea what they are talking about.

> What about reading values modified in interrupt handlers, as in your 
> "random" case?  Or is this a bug where the user of atomic_read() is 
> invalidly expecting a read each time it is called?

Quite frankly, the biggest reason for using "volatile" on jiffies was 
really historic. So even the "random" case is not really a very strong 
one. You'll notice that anybody who is actually careful will be using 
sequence locks for the jiffy accesses, if only because the *full* jiffy 
count is actually a 64-bit value, and so you cannot get it atomically on a 
32-bit architecture even on a single CPU (ie a timer interrupt might 
happen in between reading the low and the high word, so "volatile" is only 
used for the low 32 bits).

So even for jiffies, we actually have:

	extern u64 __jiffy_data jiffies_64;
	extern unsigned long volatile __jiffy_data jiffies;

where the *real* jiffies is not volatile: the volatile one is using linker 
tricks to alias the low 32 bits:

 - arch/i386/kernel/vmlinux.lds.S:

	...
	jiffies = jiffies_64;
	...

and the only reason we do all these games is (a) it works and (b) it's 
legacy.

Note how I do *not* say "(c) it's a good idea".

			Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Linus Torvalds wrote:
> So the only reason to add back "volatile" to the atomic_read() sequence is 
> not to fix bugs, but to _hide_ the bugs better. They're still there, they 
> are just a lot harder to trigger, and tend to be a lot subtler.

What about barrier removal?  With consistent semantics we could optimize a fair 
amount of code.  Whether or not that constitutes "premature" optimization is 
open to debate, but there's no question we could reduce our register wiping in 
some places.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Mon, Aug 20, 2007 at 09:15:11AM -0400, Chris Snook wrote:
> Linus Torvalds wrote:
> >So the only reason to add back "volatile" to the atomic_read() sequence is 
> >not to fix bugs, but to _hide_ the bugs better. They're still there, they 
> >are just a lot harder to trigger, and tend to be a lot subtler.
> 
> What about barrier removal?  With consistent semantics we could optimize a 
> fair amount of code.  Whether or not that constitutes "premature" 
> optimization is open to debate, but there's no question we could reduce our 
> register wiping in some places.

If you've been reading all of Linus's emails you should be
thinking about adding memory barriers, and not removing
compiler barriers.

He's just told you that code of the kind

	while (!atomic_read(cond))
		;

	do_something()

probably needs a memory barrier (not just compiler) so that
do_something() doesn't see stale cache content that occured
before cond flipped.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Herbert Xu wrote:
> On Mon, Aug 20, 2007 at 09:15:11AM -0400, Chris Snook wrote:
>> Linus Torvalds wrote:
>>> So the only reason to add back "volatile" to the atomic_read() sequence is 
>>> not to fix bugs, but to _hide_ the bugs better. They're still there, they 
>>> are just a lot harder to trigger, and tend to be a lot subtler.
>> What about barrier removal?  With consistent semantics we could optimize a 
>> fair amount of code.  Whether or not that constitutes "premature" 
>> optimization is open to debate, but there's no question we could reduce our 
>> register wiping in some places.
> 
> If you've been reading all of Linus's emails you should be
> thinking about adding memory barriers, and not removing
> compiler barriers.
> 
> He's just told you that code of the kind
> 
> 	while (!atomic_read(cond))
> 		;
> 
> 	do_something()
> 
> probably needs a memory barrier (not just compiler) so that
> do_something() doesn't see stale cache content that occured
> before cond flipped.

Such code generally doesn't care precisely when it gets the update, just that 
the update is atomic, and it doesn't loop forever.  Regardless, I'm convinced we 
just need to do it all in assembly.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> Such code generally doesn't care precisely when it gets the update, 
> just that the update is atomic, and it doesn't loop forever.

Yes, it _does_ care that it gets the update _at all_, and preferably
as early as possible.

> Regardless, I'm convinced we just need to do it all in assembly.

So do you want "volatile asm" or "plain asm", for atomic_read()?
The asm version has two ways to go about it too...


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Mon, 20 Aug 2007, Chris Snook wrote:
>
> What about barrier removal?  With consistent semantics we could optimize a
> fair amount of code.  Whether or not that constitutes "premature" optimization
> is open to debate, but there's no question we could reduce our register wiping
> in some places.

Why do people think that barriers are expensive? They really aren't. 
Especially the regular compiler barrier is basically zero cost. Any 
reasonable compiler will just flush the stuff it holds in registers that 
isn't already automatic local variables, and for regular kernel code, that 
tends to basically be nothing at all.

Ie a "barrier()" is likely _cheaper_ than the code generation downside 
from using "volatile".

		Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[David Miller]

From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Mon, 20 Aug 2007 22:46:47 -0700 (PDT)

> Ie a "barrier()" is likely _cheaper_ than the code generation downside 
> from using "volatile".

Assuming GCC were ever better about the code generation badness
with volatile that has been discussed here, I much prefer
we tell GCC "this memory piece changed" rather than "every
piece of memory has changed" which is what the barrier() does.

I happened to have been scanning a lot of assembler lately to
track down a gcc-4.2 miscompilation on sparc64, and the barriers
do hurt quite a bit in some places.  Instead of keeping unrelated
variables around cached in local registers, it reloads everything.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

David Miller wrote:
> From: Linus Torvalds <torvalds@linux-foundation.org>
> Date: Mon, 20 Aug 2007 22:46:47 -0700 (PDT)
> 
>> Ie a "barrier()" is likely _cheaper_ than the code generation downside 
>> from using "volatile".
> 
> Assuming GCC were ever better about the code generation badness
> with volatile that has been discussed here, I much prefer
> we tell GCC "this memory piece changed" rather than "every
> piece of memory has changed" which is what the barrier() does.
> 
> I happened to have been scanning a lot of assembler lately to
> track down a gcc-4.2 miscompilation on sparc64, and the barriers
> do hurt quite a bit in some places.  Instead of keeping unrelated
> variables around cached in local registers, it reloads everything.

Moore's law is definitely working against us here.  Register counts, 
pipeline depths, core counts, and clock multipliers are all increasing 
in the long run.  At some point in the future, barrier() will be 
universally regarded as a hammer too big for most purposes.  Whether or 
not removing it now constitutes premature optimization is arguable, but 
I think we should allow such optimization to happen (or not happen) in 
architecture-dependent code, and provide a consistent API that doesn't 
require the use of such things in arch-independent code where it might 
turn into a totally superfluous performance killer depending on what 
hardware it gets compiled for.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> At some point in the future, barrier() will be universally regarded as 
> a hammer too big for most purposes.  Whether or not removing it now

You can't just remove it, it is needed in some places; you want to
replace it in most places with a more fine-grained "compiler barrier",
I presume?

> constitutes premature optimization is arguable, but I think we should 
> allow such optimization to happen (or not happen) in 
> architecture-dependent code, and provide a consistent API that doesn't 
> require the use of such things in arch-independent code where it might 
> turn into a totally superfluous performance killer depending on what 
> hardware it gets compiled for.

Explicit barrier()s won't be too hard to replace -- but what to do
about the implicit barrier()s in rmb() etc. etc. -- *those* will be
hard to get rid of, if only because it is hard enough to teach driver
authors about how to use those primitives *already*.  It is far from
clear what a good interface like that would look like, anyway.

Probably we should first start experimenting with a forget()-style
micro-barrier (but please, find a better name), and see if a nice
usage pattern shows up that can be turned into an API.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Tue, 21 Aug 2007, Chris Snook wrote:

> David Miller wrote:
> > From: Linus Torvalds <torvalds@linux-foundation.org>
> > Date: Mon, 20 Aug 2007 22:46:47 -0700 (PDT)
> > 
> > > Ie a "barrier()" is likely _cheaper_ than the code generation downside
> > > from using "volatile".
> > 
> > Assuming GCC were ever better about the code generation badness
> > with volatile that has been discussed here, I much prefer
> > we tell GCC "this memory piece changed" rather than "every
> > piece of memory has changed" which is what the barrier() does.
> > 
> > I happened to have been scanning a lot of assembler lately to
> > track down a gcc-4.2 miscompilation on sparc64, and the barriers
> > do hurt quite a bit in some places.  Instead of keeping unrelated
> > variables around cached in local registers, it reloads everything.
> 
> Moore's law is definitely working against us here.  Register counts, pipeline
> depths, core counts, and clock multipliers are all increasing in the long run.
> At some point in the future, barrier() will be universally regarded as a
> hammer too big for most purposes.

I do agree, and the important point to note is that the benefits of a
/lighter/ compiler barrier, such as what David referred to above, _can_
be had without having to do anything with the "volatile" keyword at all.
And such a primitive has already been mentioned/proposed on this thread.


But this is all tangential to the core question at hand -- whether to have
implicit (compiler, possibly "light-weight" of the kind referred above)
barrier semantics in atomic ops that do not have them, or not.

I was lately looking in the kernel for _actual_ code that uses atomic_t
and benefits from the lack of any implicit barrier, with the compiler
being free to cache the atomic_t in a register. Now that often does _not_
happen, because all other ops (implemented in asm with LOCK prefix on x86)
_must_ therefore constrain the atomic_t to memory anyway. So typically all
atomic ops code sequences end up operating on memory.

Then I did locate sched.c:select_nohz_load_balancer() -- it repeatedly
references the same atomic_t object, and the code that I saw generated
(with CC_OPTIMIZE_FOR_SIZE=y) did cache it in a register for a sequence of
instructions. It uses atomic_cmpxchg, thereby not requiring explicit
memory barriers anywhere in the code, and is an example of an atomic_t
user that is safe, and yet benefits from its memory loads/stores being
elided/coalesced by the compiler.


# at this point, %%eax holds num_online_cpus() and
# %%ebx holds cpus_weight(nohz.cpu_mask)
# the variable "cpu" is in %esi

0xc1018e1d:      cmp    %eax,%ebx		# if No.A.
0xc1018e1f:      mov    0xc134d900,%eax		# first atomic_read()
0xc1018e24:      jne    0xc1018e36
0xc1018e26:      cmp    %esi,%eax		# if No.B.
0xc1018e28:      jne    0xc1018e80		# returns with 0
0xc1018e2a:      movl   $0xffffffff,0xc134d900	# atomic_set(-1), and ...
0xc1018e34:      jmp    0xc1018e80		# ... returns with 0
0xc1018e36:      cmp    $0xffffffff,%eax	# if No.C. (NOTE!)
0xc1018e39:      jne    0xc1018e46
0xc1018e3b:      lock cmpxchg %esi,0xc134d900	# atomic_cmpxchg()
0xc1018e43:      inc    %eax
0xc1018e44:      jmp    0xc1018e48
0xc1018e46:      cmp    %esi,%eax		# if No.D. (NOTE!)
0xc1018e48:      jne    0xc1018e80		# if !=, default return 0 (if No.E.)
0xc1018e4a:      jmp    0xc1018e84		# otherwise (==) returns with 1

The above is:

	if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {	/* if No.A. */
		if (atomic_read(&nohz.load_balancer) == cpu)	/* if No.B. */
			atomic_set(&nohz.load_balancer, -1);	/* XXX */
		return 0;
	}
	if (atomic_read(&nohz.load_balancer) == -1) {		/* if No.C. */
		/* make me the ilb owner */
		if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)	/* if No.E. */
			return 1;
	} else if (atomic_read(&nohz.load_balancer) == cpu)	/* if No.D. */
		return 1;
	...
	...
	return 0; /* default return from function */

As you can see, the atomic_read()'s of "if"s Nos. B, C, and D, were _all_
coalesced into a single memory reference "mov    0xc134d900,%eax" at the
top of the function, and then "if"s Nos. C and D simply used the value
from %%eax itself. But that's perfectly safe, such is the logic of this
function. It uses cmpxchg _whenever_ updating the value in the memory
atomic_t and then returns appropriately. The _only_ point that a casual
reader may find racy is that marked /* XXX */ above -- atomic_read()
followed by atomic_set() with no barrier in between. But even that is ok,
because if one thread ever finds that condition to succeed, it is 100%
guaranteed no other thread on any other CPU will find _any_ condition
to be true, thereby avoiding any race in the modification of that value.


BTW it does sound reasonable that a lot of atomic_t users that want a
compiler barrier probably also want a memory barrier. Do we make _that_
implicit too? Quite clearly, making _either_ one of those implicit in
atomic_{read,set} (in any form of implementation -- a forget() macro
based, *(volatile int *)& based, or inline asm based) would end up
harming code such as that cited above.

Lastly, the most obvious reason that should be considered against implicit
barriers in atomic ops is that it isn't "required" -- atomicity does not
imply any barrier after all, and making such a distinction would actually
be a healthy separation that helps people think more clearly when writing
lockless code.

[ But the "authors' expectations" / heisenbugs argument also holds some
  water ... for that, we can have a _variant_ in the API for atomic ops
  that has implicit compiler/memory barriers, to make it easier on those
  who want that behaviour. But let us not penalize code that knows what
  it is doing by changing the default to that, please. ]


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Tue, 21 Aug 2007, Chris Snook wrote:
> 
> Moore's law is definitely working against us here.  Register counts, pipeline
> depths, core counts, and clock multipliers are all increasing in the long run.
> At some point in the future, barrier() will be universally regarded as a
> hammer too big for most purposes.

Note that "barrier()" is purely a compiler barrier. It has zero impact on 
the CPU pipeline itself, and also has zero impact on anything that gcc 
knows isn't visible in memory (ie local variables that don't have their 
address taken), so barrier() really is pretty cheap.

Now, it's possible that gcc messes up in some circumstances, and that the 
memory clobber will cause gcc to also do things like flush local registers 
unnecessarily to their stack slots, but quite frankly, if that happens, 
it's a gcc problem, and I also have to say that I've not seen that myself.

So in a very real sense, "barrier()" will just make sure that there is a 
stronger sequence point for the compiler where things are stable. In most 
cases it has absolutely zero performance impact - apart from the 
-intended- impact of making sure that the compiler doesn't re-order or 
cache stuff around it.

And sure, we could make it more finegrained, and also introduce a 
per-variable barrier, but the fact is, people _already_ have problems with 
thinking about these kinds of things, and adding new abstraction issues 
with subtle semantics is the last thing we want.

So I really think you'd want to show a real example of real code that 
actually gets noticeably slower or bigger.

In removing "volatile", we have shown that. It may not have made a big 
difference on powerpc, but it makes a real difference on x86 - and more 
importantly, it removes something that people clearly don't know how it 
works, and incorrectly expect to just fix bugs.

[ There are *other* barriers - the ones that actually add memory barriers 
  to the CPU - that really can be quite expensive. The good news is that 
  the expense is going down rather than up: both Intel and AMD are not 
  only removing the need for some of them (ie "smp_rmb()" will become a 
  compiler-only barrier), but we're _also_ seeing the whole "pipeline 
  flush" approach go away, and be replaced by the CPU itself actually 
  being better - so even the actual CPU pipeline barriers are getting
  cheaper, not more expensive. ]

For example, did anybody even _test_ how expensive "barrier()" is? Just 
as a lark, I did

	#undef barrier
	#define barrier() do { } while (0)

in kernel/sched.c (which only has three of them in it, but hey, that's 
more than most files), and there were _zero_ code generation downsides. 
One instruction was moved (and a few line numbers changed), so it wasn't 
like the assembly language was identical, but the point is, barrier() 
simply doesn't have the same kinds of downsides that "volatile" has.

(That may not be true on other architectures or in other source files, of 
course. This *does* depend on code generation details. But anybody who 
thinks that "barrier()" is fundamentally expensive is simply incorrect. It 
is *fundamnetally* a no-op).

		Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Friday 17 August 2007 17:48, Linus Torvalds wrote:
> 
> On Fri, 17 Aug 2007, Nick Piggin wrote:
> > 
> > That's not obviously just taste to me. Not when the primitive has many
> > (perhaps, the majority) of uses that do not require said barriers. And
> > this is not solely about the code generation (which, as Paul says, is
> > relatively minor even on x86). I prefer people to think explicitly
> > about barriers in their lockless code.
> 
> Indeed.
> 
> I think the important issues are:
> 
>  - "volatile" itself is simply a badly/weakly defined issue. The semantics 
>    of it as far as the compiler is concerned are really not very good, and 
>    in practice tends to boil down to "I will generate so bad code that 
>    nobody can accuse me of optimizing anything away".
> 
>  - "volatile" - regardless of how well or badly defined it is - is purely 
>    a compiler thing. It has absolutely no meaning for the CPU itself, so 
>    it at no point implies any CPU barriers. As a result, even if the 
>    compiler generates crap code and doesn't re-order anything, there's 
>    nothing that says what the CPU will do.
> 
>  - in other words, the *only* possible meaning for "volatile" is a purely 
>    single-CPU meaning. And if you only have a single CPU involved in the 
>    process, the "volatile" is by definition pointless (because even 
>    without a volatile, the compiler is required to make the C code appear 
>    consistent as far as a single CPU is concerned).
> 
> So, let's take the example *buggy* code where we use "volatile" to wait 
> for other CPU's:
> 
> 	atomic_set(&var, 0);
> 	while (!atomic_read(&var))
> 		/* nothing */;
> 
> 
> which generates an endless loop if we don't have atomic_read() imply 
> volatile.
> 
> The point here is that it's buggy whether the volatile is there or not! 
> Exactly because the user expects multi-processing behaviour, but 
> "volatile" doesn't actually give any real guarantees about it. Another CPU 
> may have done:
> 
> 	external_ptr = kmalloc(..);
> 	/* Setup is now complete, inform the waiter */
> 	atomic_inc(&var);
> 
> but the fact is, since the other CPU isn't serialized in any way, the 
> "while-loop" (even in the presense of "volatile") doesn't actually work 
> right! Whatever the "atomic_read()" was waiting for may not have 
> completed, because we have no barriers!

Why is all this fixation on "volatile"? I don't think
people want "volatile" keyword per se, they want atomic_read(&x) to
_always_ compile into an memory-accessing instruction, not register access.
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arjan van de Ven]

On Sun, 9 Sep 2007 19:02:54 +0100
Denys Vlasenko <vda.linux@googlemail.com> wrote:

> Why is all this fixation on "volatile"? I don't think
> people want "volatile" keyword per se, they want atomic_read(&x) to
> _always_ compile into an memory-accessing instruction, not register
> access.

and ... why is that?
is there any valid, non-buggy code sequence that makes that a
reasonable requirement?

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Sunday 09 September 2007 19:18, Arjan van de Ven wrote:
> On Sun, 9 Sep 2007 19:02:54 +0100
> Denys Vlasenko <vda.linux@googlemail.com> wrote:
> 
> > Why is all this fixation on "volatile"? I don't think
> > people want "volatile" keyword per se, they want atomic_read(&x) to
> > _always_ compile into an memory-accessing instruction, not register
> > access.
> 
> and ... why is that?
> is there any valid, non-buggy code sequence that makes that a
> reasonable requirement?

Well, if you insist on having it again:

Waiting for atomic value to be zero:

        while (atomic_read(&x))
                continue;

gcc may happily convert it into:

        reg = atomic_read(&x);
        while (reg)
                continue;

Expecting every driver writer to remember that atomic_read is not in fact
a "read from memory" is naive. That won't happen. Face it, majority of
driver authors are a bit less talented than Ingo Molnar or Arjan van de Ven ;)
The name of the macro is saying that it's a read.
We are confusing users here.

It's doubly confusing that cpy_relax(), which says _nothing_ about barriers
in its name, is actually a barrier you need to insert here.
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Herbert Xu]

On Mon, Sep 10, 2007 at 11:56:29AM +0100, Denys Vlasenko wrote:
> 
> Expecting every driver writer to remember that atomic_read is not in fact
> a "read from memory" is naive. That won't happen. Face it, majority of
> driver authors are a bit less talented than Ingo Molnar or Arjan van de Ven ;)
> The name of the macro is saying that it's a read.
> We are confusing users here.

For driver authors who're too busy to learn the intricacies
of atomic operations, we have the plain old spin lock which
then lets you use normal data structures such as u32 safely.

Cheers,
-- 
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Kyle Moffett]

On Sep 10, 2007, at 06:56:29, Denys Vlasenko wrote:
> On Sunday 09 September 2007 19:18, Arjan van de Ven wrote:
>> On Sun, 9 Sep 2007 19:02:54 +0100
>> Denys Vlasenko <vda.linux@googlemail.com> wrote:
>>
>>> Why is all this fixation on "volatile"? I don't think people want  
>>> "volatile" keyword per se, they want atomic_read(&x) to _always_  
>>> compile into an memory-accessing instruction, not register access.
>>
>> and ... why is that?  is there any valid, non-buggy code sequence  
>> that makes that a reasonable requirement?
>
> Well, if you insist on having it again:
>
> Waiting for atomic value to be zero:
>
>         while (atomic_read(&x))
>                 continue;
>
> gcc may happily convert it into:
>
>         reg = atomic_read(&x);
>         while (reg)
>                 continue;

Bzzt.  Even if you fixed gcc to actually convert it to a busy loop on  
a memory variable, you STILL HAVE A BUG as it may *NOT* be gcc that  
does the conversion, it may be that the CPU does the caching of the  
memory value.  GCC has no mechanism to do cache-flushes or memory- 
barriers except through our custom inline assembly.  Also, you  
probably want a cpu_relax() in there somewhere to avoid overheating  
the CPU.  Thirdly, on a large system it may take some arbitrarily  
large amount of time for cache-propagation to update the value of the  
variable in your local CPU cache.  Finally, if atomics are based on  
based on spinlock+interrupt-disable then you will sit in a tight busy- 
loop of spin_lock_irqsave()->spin_unlock_irqrestore().  Depending on  
your system's internal model this may practically lock up your core  
because the spin_lock() will take the cacheline for exclusive access  
and doing that in a loop can prevent any other CPU from doing any  
operation on it!  Since your IRQs are disabled you even have a very  
small window that an IRQ will come along and free it up long enough  
for the update to take place.

The earlier code segment of:
> while(atomic_read(&x) > 0)
> 	atomic_dec(&x);
is *completely* buggy because you could very easily have 4 CPUs doing  
this on an atomic variable with a value of 1 and end up with it at  
negative 3 by the time you are done.  Moreover all the alternatives  
are also buggy, with the sole exception of this rather obvious- 
seeming one:
> atomic_set(&x, 0);

You simply CANNOT use an atomic_t as your sole synchronizing  
primitive, it doesn't work!  You virtually ALWAYS want to use an  
atomic_t in the following types of situations:

(A) As an object refcount.  The value is never read except as part of  
an atomic_dec_return().  Why aren't you using "struct kref"?

(B) As an atomic value counter (number of processes, for example).   
Just "reading" the value is racy anyways, if you want to enforce a  
limit or something then use atomic_inc_return(), check the result,  
and use atomic_dec() if it's too big.  If you just want to return the  
statistics then you are going to be instantaneous-point-in-time anyways.

(C) As an optimization value (statistics-like, but exact accuracy  
isn't important).

Atomics are NOT A REPLACEMENT for the proper kernel subsystem, like  
completions, mutexes, semaphores, spinlocks, krefs, etc.  It's not  
useful for synchronization, only for keeping track of simple integer  
RMW values.  Note that atomic_read() and atomic_set() aren't very  
useful RMW primitives (read-nomodify-nowrite and read-set-zero- 
write).  Code which assumes anything else is probably buggy in other  
ways too.

So while I see no real reason for the "volatile" on the atomics, I  
also see no real reason why it's terribly harmful.  Regardless of the  
"volatile" on the operation the CPU is perfectly happy to cache it  
anyways so it doesn't buy you any actual "always-access-memory"  
guarantees.  If you are just interested in it as an optimization you  
could probably just read the properly-aligned integer counter  
directly, an atomic read on most CPUs.

If you really need it to hit main memory *every* *single* *time*  
(Why?  Are you using it instead of the proper kernel subsystem?)   
then you probably need a custom inline assembly helper anyways.

Cheers,
Kyle Moffett

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Monday 10 September 2007 13:22, Kyle Moffett wrote:
> On Sep 10, 2007, at 06:56:29, Denys Vlasenko wrote:
> > On Sunday 09 September 2007 19:18, Arjan van de Ven wrote:
> >> On Sun, 9 Sep 2007 19:02:54 +0100
> >> Denys Vlasenko <vda.linux@googlemail.com> wrote:
> >>
> >>> Why is all this fixation on "volatile"? I don't think people want  
> >>> "volatile" keyword per se, they want atomic_read(&x) to _always_  
> >>> compile into an memory-accessing instruction, not register access.
> >>
> >> and ... why is that?  is there any valid, non-buggy code sequence  
> >> that makes that a reasonable requirement?
> >
> > Well, if you insist on having it again:
> >
> > Waiting for atomic value to be zero:
> >
> >         while (atomic_read(&x))
> >                 continue;
> >
> > gcc may happily convert it into:
> >
> >         reg = atomic_read(&x);
> >         while (reg)
> >                 continue;
> 
> Bzzt.  Even if you fixed gcc to actually convert it to a busy loop on  
> a memory variable, you STILL HAVE A BUG as it may *NOT* be gcc that  
> does the conversion, it may be that the CPU does the caching of the  
> memory value.  GCC has no mechanism to do cache-flushes or memory- 
> barriers except through our custom inline assembly.

CPU can cache the value all right, but it cannot use that cached value
*forever*, it has to react to invalidate cycles on the shared bus
and re-fetch new data.

IOW: atomic_read(&x) which compiles down to memory accessor
will work properly.

> the CPU.  Thirdly, on a large system it may take some arbitrarily  
> large amount of time for cache-propagation to update the value of the  
> variable in your local CPU cache.

Yes, but "arbitrarily large amount of time" is actually measured
in nanoseconds here. Let's say 1000ns max for hundreds of CPUs?

> Also, you   
> probably want a cpu_relax() in there somewhere to avoid overheating  
> the CPU.

Yes, but 
1. CPU shouldn't overheat (in a sense that it gets damaged),
   it will only use more power than needed.
2. cpu_relax() just throttles down my CPU, so it's performance
   optimization only. Wait, it isn't, it's a barrier too.
   Wow, "cpu_relax" is a barrier? How am I supposed to know
   that without reading lkml flamewars and/or header files?

Let's try reading headers. asm-x86_64/processor.h:

#define cpu_relax()   rep_nop()

So, is it a barrier? No clue yet.

/* REP NOP (PAUSE) is a good thing to insert into busy-wait loops. */
static inline void rep_nop(void)
{
        __asm__ __volatile__("rep;nop": : :"memory");
}

Comment explicitly says that it is "a good thing" (doesn't say
that it is mandatory) and says NOTHING about barriers!

Barrier-ness is not mentioned and is hidden in "memory" clobber.

Do you think it's obvious enough for average driver writer?
I think not, especially that it's unlikely for him to even start
suspecting that it is a memory barrier based on the "cpu_relax"
name.

> You simply CANNOT use an atomic_t as your sole synchronizing
> primitive, it doesn't work!  You virtually ALWAYS want to use an  
> atomic_t in the following types of situations:
> 
> (A) As an object refcount.  The value is never read except as part of  
> an atomic_dec_return().  Why aren't you using "struct kref"?
> 
> (B) As an atomic value counter (number of processes, for example).   
> Just "reading" the value is racy anyways, if you want to enforce a  
> limit or something then use atomic_inc_return(), check the result,  
> and use atomic_dec() if it's too big.  If you just want to return the  
> statistics then you are going to be instantaneous-point-in-time anyways.
> 
> (C) As an optimization value (statistics-like, but exact accuracy  
> isn't important).
> 
> Atomics are NOT A REPLACEMENT for the proper kernel subsystem, like  
> completions, mutexes, semaphores, spinlocks, krefs, etc.  It's not  
> useful for synchronization, only for keeping track of simple integer  
> RMW values.  Note that atomic_read() and atomic_set() aren't very  
> useful RMW primitives (read-nomodify-nowrite and read-set-zero- 
> write).  Code which assumes anything else is probably buggy in other  
> ways too.

You are basically trying to educate me how to use atomic properly.
You don't need to do it, as I am (currently) not a driver author.

I am saying that people who are already using atomic_read()
(and who unfortunately did not read your explanation above)
will still sometimes use atomic_read() as a way to read atomic value
*from memory*, and will create nasty heisenbugs for you to debug.
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Monday 10 September 2007 14:38, Denys Vlasenko wrote:
> You are basically trying to educate me how to use atomic properly.
> You don't need to do it, as I am (currently) not a driver author.
> 
> I am saying that people who are already using atomic_read()
> (and who unfortunately did not read your explanation above)
> will still sometimes use atomic_read() as a way to read atomic value
> *from memory*, and will create nasty heisenbugs for you to debug.

static inline int
qla2x00_wait_for_loop_ready(scsi_qla_host_t *ha)
{
        int      return_status = QLA_SUCCESS;
        unsigned long loop_timeout ;
        scsi_qla_host_t *pha = to_qla_parent(ha);

        /* wait for 5 min at the max for loop to be ready */
        loop_timeout = jiffies + (MAX_LOOP_TIMEOUT * HZ);

        while ((!atomic_read(&pha->loop_down_timer) &&
            atomic_read(&pha->loop_state) == LOOP_DOWN) ||
            atomic_read(&pha->loop_state) != LOOP_READY) {
                if (atomic_read(&pha->loop_state) == LOOP_DEAD) {
                        return_status = QLA_FUNCTION_FAILED;
                        break;
                }
                msleep(1000);
                if (time_after_eq(jiffies, loop_timeout)) {
                        return_status = QLA_FUNCTION_FAILED;
                        break;
                }
        }
        return (return_status);
}

Is above correct or buggy? Correct, because msleep is a barrier.
Is it obvious? No.

static void
qla2x00_rst_aen(scsi_qla_host_t *ha)
{
        if (ha->flags.online && !ha->flags.reset_active &&
            !atomic_read(&ha->loop_down_timer) &&
            !(test_bit(ABORT_ISP_ACTIVE, &ha->dpc_flags))) {
                do {
                        clear_bit(RESET_MARKER_NEEDED, &ha->dpc_flags);

                        /*
                         * Issue marker command only when we are going to start
                         * the I/O.
                         */
                        ha->marker_needed = 1;
                } while (!atomic_read(&ha->loop_down_timer) &&
                    (test_bit(RESET_MARKER_NEEDED, &ha->dpc_flags)));
        }
}

Is above correct? I honestly don't know. Correct, because set_bit is
a barrier on _all _memory_? Will it break if set_bit will be changed
to be a barrier only on its operand? Probably yes.

drivers/kvm/kvm_main.c

        while (atomic_read(&completed) != needed) {
                cpu_relax();
                barrier();
        }

Obviously author did not know that cpu_relax is already a barrier.
See why I think driver authors will be confused?

arch/x86_64/kernel/crash.c

static void nmi_shootdown_cpus(void)
{
...
        msecs = 1000; /* Wait at most a second for the other cpus to stop */
        while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
                mdelay(1);
                msecs--;
        }
...
}

Is mdelay(1) a barrier? Yes, because it is a function on x86_64.
Absolutely the same code will be buggy on an arch where
mdelay(1) == udelay(1000), and udelay is implemented
as inline busy-wait.

arch/sparc64/kernel/smp.c

        /* Wait for response */
        while (atomic_read(&data.finished) != cpus)
                cpu_relax();
...later in the same file...
                while (atomic_read(&smp_capture_registry) != ncpus)
                        rmb();

I'm confused. Do we need cpu_relax() or rmb()? Does cpu_relax() imply rmb()?
(No it doesn't). Which of those two while loops needs correcting?
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Linus Torvalds]

On Mon, 10 Sep 2007, Denys Vlasenko wrote:
> 
> static inline int
> qla2x00_wait_for_loop_ready(scsi_qla_host_t *ha)
> {
>         int      return_status = QLA_SUCCESS;
>         unsigned long loop_timeout ;
>         scsi_qla_host_t *pha = to_qla_parent(ha);
> 
>         /* wait for 5 min at the max for loop to be ready */
>         loop_timeout = jiffies + (MAX_LOOP_TIMEOUT * HZ);
> 
>         while ((!atomic_read(&pha->loop_down_timer) &&
>             atomic_read(&pha->loop_state) == LOOP_DOWN) ||
>             atomic_read(&pha->loop_state) != LOOP_READY) {
>                 if (atomic_read(&pha->loop_state) == LOOP_DEAD) {
...
> Is above correct or buggy? Correct, because msleep is a barrier.
> Is it obvious? No.

It's *buggy*. But it has nothing to do with any msleep() in the loop, or 
anything else.

And more importantly, it would be equally buggy even *with* a "volatile" 
atomic_read().

Why is this so hard for people to understand? You're all acting like 
morons.

The reason it is buggy has absolutely nothing to do with whether the read 
is done or not, it has to do with the fact that the CPU may re-order the 
reads *regardless* of whether the read is done in some specific order by 
the compiler ot not! In effect, there is zero ordering between all those 
three reads, and if you don't have memory barriers (or a lock or other 
serialization), that code is buggy.

So stop this idiotic discussion thread already. The above kind of code 
needs memory barriers to be non-buggy. The whole "volatile or not" 
discussion is totally idiotic, and pointless, and anybody who doesn't 
understand that by now needs to just shut up and think about it more, 
rather than make this discussion drag out even further.

The fact is, "volatile" *only* makes things worse. It generates worse 
code, and never fixes any real bugs. This is a *fact*.

			Linus

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Monday 10 September 2007 16:09, Linus Torvalds wrote:
> On Mon, 10 Sep 2007, Denys Vlasenko wrote:
> > static inline int
> > qla2x00_wait_for_loop_ready(scsi_qla_host_t *ha)
> > {
> >         int      return_status = QLA_SUCCESS;
> >         unsigned long loop_timeout ;
> >         scsi_qla_host_t *pha = to_qla_parent(ha);
> > 
> >         /* wait for 5 min at the max for loop to be ready */
> >         loop_timeout = jiffies + (MAX_LOOP_TIMEOUT * HZ);
> > 
> >         while ((!atomic_read(&pha->loop_down_timer) &&
> >             atomic_read(&pha->loop_state) == LOOP_DOWN) ||
> >             atomic_read(&pha->loop_state) != LOOP_READY) {
> >                 if (atomic_read(&pha->loop_state) == LOOP_DEAD) {
> ...
> > Is above correct or buggy? Correct, because msleep is a barrier.
> > Is it obvious? No.
> 
> It's *buggy*. But it has nothing to do with any msleep() in the loop, or 
> anything else.
> 
> And more importantly, it would be equally buggy even *with* a "volatile" 
> atomic_read().

I am not saying that this code is okay, this isn't the point.
(The code is in fact awful for several more reasons).

My point is that people are confused as to what atomic_read()
exactly means, and this is bad. Same for cpu_relax().
First one says "read", and second one doesn't say "barrier".

This is real code from current kernel which demonstrates this:

"I don't know that cpu_relax() is a barrier already":

drivers/kvm/kvm_main.c
        while (atomic_read(&completed) != needed) {
                cpu_relax();
                barrier();
        }

"I think that atomic_read() is a read from memory and therefore
I don't need a barrier":

arch/x86_64/kernel/crash.c
        msecs = 1000; /* Wait at most a second for the other cpus to stop */
        while ((atomic_read(&waiting_for_crash_ipi) > 0) && msecs) {
                mdelay(1);
                msecs--;
        }

Since neither camp seems to give up, I am proposing renaming
them to something less confusing, and make everybody happy.

cpu_relax_barrier()
atomic_value(&x)
atomic_fetch(&x)

I'm not native English speaker, do these sound better?
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Kyle Moffett]

On Sep 10, 2007, at 12:46:33, Denys Vlasenko wrote:
> My point is that people are confused as to what atomic_read()   
> exactly means, and this is bad. Same for cpu_relax().  First one  
> says "read", and second one doesn't say "barrier".

Q&A:

Q:  When is it OK to use atomic_read()?
A:  You are asking the question, so never.

Q:  But I need to check the value of the atomic at this point in time...
A:  Your code is buggy if it needs to do that on an atomic_t for  
anything other than debugging or optimization.  Use either  
atomic_*_return() or a lock and some normal integers.

Q:  "So why can't the atomic_read DTRT magically?"
A:  Because "the right thing" depends on the situation and is usually  
best done with something other than atomic_t.

If somebody can post some non-buggy code which is correctly using  
atomic_read() *and* depends on the compiler generating extra  
nonsensical loads due to "volatile" then the issue *might* be  
reconsidered.  This also includes samples of code which uses  
atomic_read() and needs memory barriers (so that we can fix the buggy  
code, not so we can change atomic_read()).  So far the only code  
samples anybody has posted are buggy regardless of whether or not the  
value and/or accessors are flagged "volatile" or not.  And hey, maybe  
the volatile ops *should* be implemented in inline ASM for future- 
proof-ness, but that's a separate issue.

Cheers,
Kyle Moffett

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Mon, 10 Sep 2007, Linus Torvalds wrote:

> The fact is, "volatile" *only* makes things worse. It generates worse 
> code, and never fixes any real bugs. This is a *fact*.

Yes, lets just drop the volatiles now! We need a patch that gets rid of 
them.... Volunteers?

[PATCH] Document non-semantics of atomic_read() and atomic_set()

[Chris Snook]

From: Chris Snook <csnook@redhat.com>

Unambiguously document the fact that atomic_read() and atomic_set()
do not imply any ordering or memory access, and that callers are
obligated to explicitly invoke barriers as needed to ensure that
changes to atomic variables are visible in all contexts that need
to see them.

Signed-off-by: Chris Snook <csnook@redhat.com>

--- a/Documentation/atomic_ops.txt	2007-07-08 19:32:17.000000000 -0400
+++ b/Documentation/atomic_ops.txt	2007-09-10 19:02:50.000000000 -0400
@@ -12,7 +12,11 @@
 C integer type will fail.  Something like the following should
 suffice:
 
-	typedef struct { volatile int counter; } atomic_t;
+	typedef struct { int counter; } atomic_t;
+
+	Historically, counter has been declared volatile.  This is now
+discouraged.  See Documentation/volatile-considered-harmful.txt for the
+complete rationale.
 
 	The first operations to implement for atomic_t's are the
 initializers and plain reads.
@@ -42,6 +46,22 @@
 
 which simply reads the current value of the counter.
 
+*** WARNING: atomic_read() and atomic_set() DO NOT IMPLY BARRIERS! ***
+
+Some architectures may choose to use the volatile keyword, barriers, or
+inline assembly to guarantee some degree of immediacy for atomic_read()
+and atomic_set().  This is not uniformly guaranteed, and may change in
+the future, so all users of atomic_t should treat atomic_read() and
+atomic_set() as simple C assignment statements that may be reordered or
+optimized away entirely by the compiler or processor, and explicitly
+invoke the appropriate compiler and/or memory barrier for each use case.
+Failure to do so will result in code that may suddenly break when used with
+different architectures or compiler optimizations, or even changes in
+unrelated code which changes how the compiler optimizes the section
+accessing atomic_t variables.
+
+*** YOU HAVE BEEN WARNED! ***
+
 Now, we move onto the actual atomic operation interfaces.
 
 	void atomic_add(int i, atomic_t *v);

Re: [PATCH] Document non-semantics of atomic_read() and atomic_set()

[Paul E. McKenney]

On Mon, Sep 10, 2007 at 07:19:44PM -0400, Chris Snook wrote:
> From: Chris Snook <csnook@redhat.com>
> 
> Unambiguously document the fact that atomic_read() and atomic_set()
> do not imply any ordering or memory access, and that callers are
> obligated to explicitly invoke barriers as needed to ensure that
> changes to atomic variables are visible in all contexts that need
> to see them.

Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

> Signed-off-by: Chris Snook <csnook@redhat.com>
> 
> --- a/Documentation/atomic_ops.txt	2007-07-08 19:32:17.000000000 -0400
> +++ b/Documentation/atomic_ops.txt	2007-09-10 19:02:50.000000000 -0400
> @@ -12,7 +12,11 @@
>  C integer type will fail.  Something like the following should
>  suffice:
> 
> -	typedef struct { volatile int counter; } atomic_t;
> +	typedef struct { int counter; } atomic_t;
> +
> +	Historically, counter has been declared volatile.  This is now
> +discouraged.  See Documentation/volatile-considered-harmful.txt for the
> +complete rationale.
> 
>  	The first operations to implement for atomic_t's are the
>  initializers and plain reads.
> @@ -42,6 +46,22 @@
> 
>  which simply reads the current value of the counter.
> 
> +*** WARNING: atomic_read() and atomic_set() DO NOT IMPLY BARRIERS! ***
> +
> +Some architectures may choose to use the volatile keyword, barriers, or
> +inline assembly to guarantee some degree of immediacy for atomic_read()
> +and atomic_set().  This is not uniformly guaranteed, and may change in
> +the future, so all users of atomic_t should treat atomic_read() and
> +atomic_set() as simple C assignment statements that may be reordered or
> +optimized away entirely by the compiler or processor, and explicitly
> +invoke the appropriate compiler and/or memory barrier for each use case.
> +Failure to do so will result in code that may suddenly break when used with
> +different architectures or compiler optimizations, or even changes in
> +unrelated code which changes how the compiler optimizes the section
> +accessing atomic_t variables.
> +
> +*** YOU HAVE BEEN WARNED! ***
> +
>  Now, we move onto the actual atomic operation interfaces.
> 
>  	void atomic_add(int i, atomic_t *v);

Re: [PATCH] Document non-semantics of atomic_read() and atomic_set()

[Christoph Lameter]

Acked-by: Christoph Lameter <clameter@sgi.com>

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arjan van de Ven]

On Mon, 10 Sep 2007 11:56:29 +0100
Denys Vlasenko <vda.linux@googlemail.com> wrote:

> 
> Well, if you insist on having it again:
> 
> Waiting for atomic value to be zero:
> 
>         while (atomic_read(&x))
>                 continue;
> 

and this I would say is buggy code all the way.

Not from a pure C level semantics, but from a "busy waiting is buggy"
semantics level and a "I'm inventing my own locking" semantics level.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Monday 10 September 2007 15:51, Arjan van de Ven wrote:
> On Mon, 10 Sep 2007 11:56:29 +0100
> Denys Vlasenko <vda.linux@googlemail.com> wrote:
> 
> > 
> > Well, if you insist on having it again:
> > 
> > Waiting for atomic value to be zero:
> > 
> >         while (atomic_read(&x))
> >                 continue;
> > 
> 
> and this I would say is buggy code all the way.
>
> Not from a pure C level semantics, but from a "busy waiting is buggy"
> semantics level and a "I'm inventing my own locking" semantics level.

After inspecting arch/*, I cannot agree with you.
Otherwise almost all major architectures use
"conceptually buggy busy-waiting":

arch/alpha
arch/i386
arch/ia64
arch/m32r
arch/mips
arch/parisc
arch/powerpc
arch/sh
arch/sparc64
arch/um
arch/x86_64

All of the above contain busy-waiting on atomic_read.

Including these loops without barriers:

arch/mips/kernel/smtc.c
			while (atomic_read(&idle_hook_initialized) < 1000)
				;
arch/mips/sgi-ip27/ip27-nmi.c
	while (atomic_read(&nmied_cpus) != num_online_cpus());

[Well maybe num_online_cpus() is a barrier, I didn't check]

arch/sh/kernel/smp.c
	if (wait)
		while (atomic_read(&smp_fn_call.finished) != (nr_cpus - 1));

Bugs?
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Arjan van de Ven]

On Mon, 10 Sep 2007 15:38:23 +0100
Denys Vlasenko <vda.linux@googlemail.com> wrote:

> On Monday 10 September 2007 15:51, Arjan van de Ven wrote:
> > On Mon, 10 Sep 2007 11:56:29 +0100
> > Denys Vlasenko <vda.linux@googlemail.com> wrote:
> > 
> > > 
> > > Well, if you insist on having it again:
> > > 
> > > Waiting for atomic value to be zero:
> > > 
> > >         while (atomic_read(&x))
> > >                 continue;
> > > 
> > 
> > and this I would say is buggy code all the way.
> >
> > Not from a pure C level semantics, but from a "busy waiting is
> > buggy" semantics level and a "I'm inventing my own locking"
> > semantics level.
> 
> After inspecting arch/*, I cannot agree with you.

the arch/ people obviously are allowed to do their own locking stuff...
BECAUSE THEY HAVE TO IMPLEMENT THAT!


the arch maintainers know EXACTLY how their hw behaves (well, we hope)
so they tend to be the exception to many rules in the kernel....

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Stefan Richter]

Nick Piggin wrote:
> Satyam Sharma wrote:
>> And we have driver / subsystem maintainers such as Stefan
>> coming up and admitting that often a lot of code that's written to use
>> atomic_read() does assume the read will not be elided by the compiler.
> 
> So these are broken on i386 and x86-64?

The ieee1394 and firewire subsystems have open, undiagnosed bugs, also
on i386 and x86-64.  But whether there is any bug because of wrong
assumptions about atomic_read among them, I don't know.  I don't know
which assumptions the authors made, I only know that I wasn't aware of
all the properties of atomic_read until now.

> Are they definitely safe on SMP and weakly ordered machines with
> just a simple compiler barrier there? Because I would not be
> surprised if there are a lot of developers who don't really know
> what to assume when it comes to memory ordering issues.
> 
> This is not a dig at driver writers: we still have memory ordering
> problems in the VM too (and probably most of the subtle bugs in
> lockless VM code are memory ordering ones). Let's not make up a
> false sense of security and hope that sprinkling volatile around
> will allow people to write bug-free lockless code. If a writer
> can't be bothered reading API documentation

...or, if there is none, the implementation specification (as in case of
the atomic ops), or, if there is none, the implementation (as in case of
a some infrastructure code here and there)...

> and learning the Linux memory model, they can still be productive
> writing safely locked code.

Provided they are aware that they might not have the full picture of the
lockless primitives.  :-)
-- 
Stefan Richter
-=====-=-=== =--- =---=
http://arcgraph.de/sr/

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> Of course, since *normal* accesses aren't necessarily limited wrt
> re-ordering, the question then becomes one of "with regard to *what* 
> does
> it limit re-ordering?".
>
> A C compiler that re-orders two different volatile accesses that have a
> sequence point in between them is pretty clearly a buggy compiler. So 
> at a
> minimum, it limits re-ordering wrt other volatiles (assuming sequence
> points exists). It also means that the compiler cannot move it
> speculatively across conditionals, but other than that it's starting to
> get fuzzy.

This is actually really well-defined in C, not fuzzy at all.
"Volatile accesses" are a side effect, and no side effects can
be reordered with respect to sequence points.  The side effects
that matter in the kernel environment are: 1) accessing a volatile
object; 2) modifying an object; 3) volatile asm(); 4) calling a
function that does any of these.

We certainly should avoid volatile whenever possible, but "because
it's fuzzy wrt reordering" is not a reason -- all alternatives have
exactly the same issues.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> atomic_dec() already has volatile behavior everywhere, so this is 
>> semantically
>> okay, but this code (and any like it) should be calling cpu_relax() 
>> each
>> iteration through the loop, unless there's a compelling reason not 
>> to.  I'll
>> allow that for some hardware drivers (possibly this one) such a 
>> compelling
>> reason may exist, but hardware-independent core subsystems probably 
>> have no
>> excuse.
>
> No it does not have any volatile semantics. atomic_dec() can be 
> reordered
> at will by the compiler within the current basic unit if you do not 
> add a
> barrier.

"volatile" has nothing to do with reordering.  atomic_dec() writes
to memory, so it _does_ have "volatile semantics", implicitly, as
long as the compiler cannot optimise the atomic variable away
completely -- any store counts as a side effect.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Fri, 17 Aug 2007, Segher Boessenkool wrote:

> > > atomic_dec() already has volatile behavior everywhere, so this is
> > > semantically
> > > okay, but this code (and any like it) should be calling cpu_relax() each
> > > iteration through the loop, unless there's a compelling reason not to.
> > > I'll
> > > allow that for some hardware drivers (possibly this one) such a compelling
> > > reason may exist, but hardware-independent core subsystems probably have
> > > no
> > > excuse.
> > 
> > No it does not have any volatile semantics. atomic_dec() can be reordered
> > at will by the compiler within the current basic unit if you do not add a
> > barrier.
> 
> "volatile" has nothing to do with reordering.

If you're talking of "volatile" the type-qualifier keyword, then
http://lkml.org/lkml/2007/8/16/231 (and sub-thread below it) shows
otherwise.

> atomic_dec() writes
> to memory, so it _does_ have "volatile semantics", implicitly, as
> long as the compiler cannot optimise the atomic variable away
> completely -- any store counts as a side effect.

I don't think an atomic_dec() implemented as an inline "asm volatile"
or one that uses a "forget" macro would have the same re-ordering
guarantees as an atomic_dec() that uses a volatile access cast.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>> No it does not have any volatile semantics. atomic_dec() can be 
>>> reordered
>>> at will by the compiler within the current basic unit if you do not 
>>> add a
>>> barrier.
>>
>> "volatile" has nothing to do with reordering.
>
> If you're talking of "volatile" the type-qualifier keyword, then
> http://lkml.org/lkml/2007/8/16/231 (and sub-thread below it) shows
> otherwise.

I'm not sure what in that mail you mean, but anyway...

Yes, of course, the fact that "volatile" creates a side effect
prevents certain things from being reordered wrt the atomic_dec();
but the atomic_dec() has a side effect *already* so the volatile
doesn't change anything.

>> atomic_dec() writes
>> to memory, so it _does_ have "volatile semantics", implicitly, as
>> long as the compiler cannot optimise the atomic variable away
>> completely -- any store counts as a side effect.
>
> I don't think an atomic_dec() implemented as an inline "asm volatile"
> or one that uses a "forget" macro would have the same re-ordering
> guarantees as an atomic_dec() that uses a volatile access cast.

The "asm volatile" implementation does have exactly the same
reordering guarantees as the "volatile cast" thing, if that is
implemented by GCC in the "obvious" way.  Even a "plain" asm()
will do the same.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Sat, 18 Aug 2007, Segher Boessenkool wrote:

> > > > No it does not have any volatile semantics. atomic_dec() can be
> > > > reordered
> > > > at will by the compiler within the current basic unit if you do not add
> > > > a
> > > > barrier.
> > > 
> > > "volatile" has nothing to do with reordering.
> > 
> > If you're talking of "volatile" the type-qualifier keyword, then
> > http://lkml.org/lkml/2007/8/16/231 (and sub-thread below it) shows
> > otherwise.
> 
> I'm not sure what in that mail you mean, but anyway...
> 
> Yes, of course, the fact that "volatile" creates a side effect
> prevents certain things from being reordered wrt the atomic_dec();
> but the atomic_dec() has a side effect *already* so the volatile
> doesn't change anything.

That's precisely what that sub-thread (read down to the last mail
there, and not the first mail only) shows. So yes, "volatile" does
have something to do with re-ordering (as guaranteed by the C
standard).


> > > atomic_dec() writes
> > > to memory, so it _does_ have "volatile semantics", implicitly, as
> > > long as the compiler cannot optimise the atomic variable away
> > > completely -- any store counts as a side effect.
> > 
> > I don't think an atomic_dec() implemented as an inline "asm volatile"
> > or one that uses a "forget" macro would have the same re-ordering
> > guarantees as an atomic_dec() that uses a volatile access cast.
> 
> The "asm volatile" implementation does have exactly the same
> reordering guarantees as the "volatile cast" thing,

I don't think so.

> if that is
> implemented by GCC in the "obvious" way.  Even a "plain" asm()
> will do the same.

Read the relevant GCC documentation.

[ of course, if the (latest) GCC documentation is *yet again*
  wrong, then alright, not much I can do about it, is there. ]

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>>> atomic_dec() writes
>>>> to memory, so it _does_ have "volatile semantics", implicitly, as
>>>> long as the compiler cannot optimise the atomic variable away
>>>> completely -- any store counts as a side effect.
>>>
>>> I don't think an atomic_dec() implemented as an inline "asm volatile"
>>> or one that uses a "forget" macro would have the same re-ordering
>>> guarantees as an atomic_dec() that uses a volatile access cast.
>>
>> The "asm volatile" implementation does have exactly the same
>> reordering guarantees as the "volatile cast" thing,
>
> I don't think so.

"asm volatile" creates a side effect.  Side effects aren't
allowed to be reordered wrt sequence points.  This is exactly
the same reason as why "volatile accesses" cannot be reordered.

>> if that is
>> implemented by GCC in the "obvious" way.  Even a "plain" asm()
>> will do the same.
>
> Read the relevant GCC documentation.

I did, yes.

> [ of course, if the (latest) GCC documentation is *yet again*
>   wrong, then alright, not much I can do about it, is there. ]

There was (and is) nothing wrong about the "+m" documentation, if
that is what you are talking about.  It could be extended now, to
allow "+m" -- but that takes more than just "fixing" the documentation.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Sat, 18 Aug 2007, Segher Boessenkool wrote:

> > > > > atomic_dec() writes
> > > > > to memory, so it _does_ have "volatile semantics", implicitly, as
> > > > > long as the compiler cannot optimise the atomic variable away
> > > > > completely -- any store counts as a side effect.
> > > > 
> > > > I don't think an atomic_dec() implemented as an inline "asm volatile"
> > > > or one that uses a "forget" macro would have the same re-ordering
> > > > guarantees as an atomic_dec() that uses a volatile access cast.
> > > 
> > > The "asm volatile" implementation does have exactly the same
> > > reordering guarantees as the "volatile cast" thing,
> > 
> > I don't think so.
> 
> "asm volatile" creates a side effect.

Yeah.

> Side effects aren't
> allowed to be reordered wrt sequence points.

Yeah.

> This is exactly
> the same reason as why "volatile accesses" cannot be reordered.

No, the code in that sub-thread I earlier pointed you at *WAS* written
such that there was a sequence point after all the uses of that volatile
access cast macro, and _therefore_ we were safe from re-ordering
(behaviour guaranteed by the C standard).

But you seem to be missing the simple and basic fact that:

	(something_that_has_side_effects || statement)
			!= something_that_is_a_sequence_point

Now, one cannot fantasize that "volatile asms" are also sequence points.
In fact such an argument would be sadly mistaken, because "sequence
points" are defined by the C standard and it'd be horribly wrong to
even _try_ claiming that the C standard knows about "volatile asms".


> > > if that is
> > > implemented by GCC in the "obvious" way.  Even a "plain" asm()
> > > will do the same.
> > 
> > Read the relevant GCC documentation.
> 
> I did, yes.

No, you didn't read:

http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html

Read the bit about the need for artificial dependencies, and the example
given there:

	asm volatile("mtfsf 255,%0" : : "f" (fpenv));
	sum = x + y;

The docs explicitly say the addition can be moved before the "volatile
asm". Hopefully, as you know, (x + y) is an C "expression" and hence
a "sequence point" as defined by the standard. So the "volatile asm"
should've happened before it, right? Wrong.

I know there is also stuff written about "side-effects" there which
_could_ give the same semantic w.r.t. sequence points as the volatile
access casts, but hey, it's GCC's own documentation, you obviously can't
find fault with _me_ if there's wrong stuff written in there. Say that
to GCC ...

See, "volatile" C keyword, for all it's ill-definition and dodgy
semantics, is still at least given somewhat of a treatment in the C
standard (whose quality is ... ummm, sadly not always good and clear,
but unsurprisingly, still about 5,482 orders-of-magnitude times
better than GCC docs). Semantics of "volatile" as applies to inline
asm, OTOH? You're completely relying on the compiler for that ...


> > [ of course, if the (latest) GCC documentation is *yet again*
> >   wrong, then alright, not much I can do about it, is there. ]
> 
> There was (and is) nothing wrong about the "+m" documentation, if
> that is what you are talking about.  It could be extended now, to
> allow "+m" -- but that takes more than just "fixing" the documentation.

No, there was (and is) _everything_ wrong about the "+" documentation as
applies to memory-constrained operands. I don't give a whit if it's
some workaround in their gimplifier, or the other, that makes it possible
to use "+m" (like the current kernel code does). The docs suggest
otherwise, so there's obviously a clear disconnect between the docs and
actual GCC behaviour.


[ You seem to often take issue with _amazingly_ petty and pedantic things,
  by the way :-) ]

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>>> The "asm volatile" implementation does have exactly the same
>>>> reordering guarantees as the "volatile cast" thing,
>>>
>>> I don't think so.
>>
>> "asm volatile" creates a side effect.
>
> Yeah.
>
>> Side effects aren't
>> allowed to be reordered wrt sequence points.
>
> Yeah.
>
>> This is exactly
>> the same reason as why "volatile accesses" cannot be reordered.
>
> No, the code in that sub-thread I earlier pointed you at *WAS* written
> such that there was a sequence point after all the uses of that 
> volatile
> access cast macro, and _therefore_ we were safe from re-ordering
> (behaviour guaranteed by the C standard).

And exactly the same is true for the "asm" version.

> Now, one cannot fantasize that "volatile asms" are also sequence 
> points.

Sure you can do that.  I don't though.

> In fact such an argument would be sadly mistaken, because "sequence
> points" are defined by the C standard and it'd be horribly wrong to
> even _try_ claiming that the C standard knows about "volatile asms".

That's nonsense.  GCC can extend the C standard any way they
bloody well please -- witness the fact that they added an
extra class of side effects...

>>> Read the relevant GCC documentation.
>>
>> I did, yes.
>
> No, you didn't read:
>
> http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
>
> Read the bit about the need for artificial dependencies, and the 
> example
> given there:
>
> 	asm volatile("mtfsf 255,%0" : : "f" (fpenv));
> 	sum = x + y;
>
> The docs explicitly say the addition can be moved before the "volatile
> asm". Hopefully, as you know, (x + y) is an C "expression" and hence
> a "sequence point" as defined by the standard.

The _end of a full expression_ is a sequence point, not every
expression.  And that is irrelevant here anyway.

It is perfectly fine to compute x+y any time before the
assignment; the C compiler is allowed to compute it _after_
the assignment even, if it could figure out how ;-)

x+y does not contain a side effect, you know.

> I know there is also stuff written about "side-effects" there which
> _could_ give the same semantic w.r.t. sequence points as the volatile
> access casts,

s/could/does/

> but hey, it's GCC's own documentation, you obviously can't
> find fault with _me_ if there's wrong stuff written in there. Say that
> to GCC ...

There's nothing wrong there.

> See, "volatile" C keyword, for all it's ill-definition and dodgy
> semantics, is still at least given somewhat of a treatment in the C
> standard (whose quality is ... ummm, sadly not always good and clear,
> but unsurprisingly, still about 5,482 orders-of-magnitude times
> better than GCC docs).

If you find any problems/shortcomings in the GCC documentation,
please file a PR, don't go whine on some unrelated mailing lists.
Thank you.

> Semantics of "volatile" as applies to inline
> asm, OTOH? You're completely relying on the compiler for that ...

Yes, and?  GCC promises the behaviour it has documented.

>>> [ of course, if the (latest) GCC documentation is *yet again*
>>>   wrong, then alright, not much I can do about it, is there. ]
>>
>> There was (and is) nothing wrong about the "+m" documentation, if
>> that is what you are talking about.  It could be extended now, to
>> allow "+m" -- but that takes more than just "fixing" the 
>> documentation.
>
> No, there was (and is) _everything_ wrong about the "+" documentation 
> as
> applies to memory-constrained operands. I don't give a whit if it's
> some workaround in their gimplifier, or the other, that makes it 
> possible
> to use "+m" (like the current kernel code does). The docs suggest
> otherwise, so there's obviously a clear disconnect between the docs and
> actual GCC behaviour.

The documentation simply doesn't say "+m" is allowed.  The code to
allow it was added for the benefit of people who do not read the
documentation.  Documentation for "+m" might get added later if it
is decided this [the code, not the documentation] is a sane thing
to have (which isn't directly obvious).

> [ You seem to often take issue with _amazingly_ petty and pedantic 
> things,
>   by the way :-) ]

If you're talking details, you better get them right.  Handwaving is
fine with me as long as you're not purporting you're not.

And I simply cannot stand false assertions.

You can always ignore me if _you_ take issue with _that_ :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Sat, 18 Aug 2007, Segher Boessenkool wrote:

> > > > > The "asm volatile" implementation does have exactly the same
> > > > > reordering guarantees as the "volatile cast" thing,
> > > > 
> > > > I don't think so.
> > > 
> > > "asm volatile" creates a side effect.
> > 
> > Yeah.
> > 
> > > Side effects aren't
> > > allowed to be reordered wrt sequence points.
> > 
> > Yeah.
> > 
> > > This is exactly
> > > the same reason as why "volatile accesses" cannot be reordered.
> > 
> > No, the code in that sub-thread I earlier pointed you at *WAS* written
> > such that there was a sequence point after all the uses of that volatile
> > access cast macro, and _therefore_ we were safe from re-ordering
> > (behaviour guaranteed by the C standard).
> 
> And exactly the same is true for the "asm" version.
> 
> > Now, one cannot fantasize that "volatile asms" are also sequence points.
> 
> Sure you can do that.  I don't though.
> 
> > In fact such an argument would be sadly mistaken, because "sequence
> > points" are defined by the C standard and it'd be horribly wrong to
> > even _try_ claiming that the C standard knows about "volatile asms".
> 
> That's nonsense.  GCC can extend the C standard any way they
> bloody well please -- witness the fact that they added an
> extra class of side effects...
> 
> > > > Read the relevant GCC documentation.
> > > 
> > > I did, yes.
> > 
> > No, you didn't read:
> > 
> > http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
> > 
> > Read the bit about the need for artificial dependencies, and the example
> > given there:
> > 
> > 	asm volatile("mtfsf 255,%0" : : "f" (fpenv));
> > 	sum = x + y;
> > 
> > The docs explicitly say the addition can be moved before the "volatile
> > asm". Hopefully, as you know, (x + y) is an C "expression" and hence
> > a "sequence point" as defined by the standard.
> 
> The _end of a full expression_ is a sequence point, not every
> expression.  And that is irrelevant here anyway.
> 
> It is perfectly fine to compute x+y any time before the
> assignment; the C compiler is allowed to compute it _after_
> the assignment even, if it could figure out how ;-)
> 
> x+y does not contain a side effect, you know.
> 
> > I know there is also stuff written about "side-effects" there which
> > _could_ give the same semantic w.r.t. sequence points as the volatile
> > access casts,
> 
> s/could/does/
> 
> > but hey, it's GCC's own documentation, you obviously can't
> > find fault with _me_ if there's wrong stuff written in there. Say that
> > to GCC ...
> 
> There's nothing wrong there.
> 
> > See, "volatile" C keyword, for all it's ill-definition and dodgy
> > semantics, is still at least given somewhat of a treatment in the C
> > standard (whose quality is ... ummm, sadly not always good and clear,
> > but unsurprisingly, still about 5,482 orders-of-magnitude times
> > better than GCC docs).
> 
> If you find any problems/shortcomings in the GCC documentation,
> please file a PR, don't go whine on some unrelated mailing lists.
> Thank you.
> 
> > Semantics of "volatile" as applies to inline
> > asm, OTOH? You're completely relying on the compiler for that ...
> 
> Yes, and?  GCC promises the behaviour it has documented.

LOTS there, which obviously isn't correct, but which I'll reply to later,
easier stuff first. (call this "handwaving" if you want, but don't worry,
I /will/ bother myself to reply)


> > > > [ of course, if the (latest) GCC documentation is *yet again*
> > > >   wrong, then alright, not much I can do about it, is there. ]
> > > 
> > > There was (and is) nothing wrong about the "+m" documentation, if
> > > that is what you are talking about.  It could be extended now, to
> > > allow "+m" -- but that takes more than just "fixing" the documentation.
> > 
> > No, there was (and is) _everything_ wrong about the "+" documentation as
> > applies to memory-constrained operands. I don't give a whit if it's
> > some workaround in their gimplifier, or the other, that makes it possible
> > to use "+m" (like the current kernel code does). The docs suggest
> > otherwise, so there's obviously a clear disconnect between the docs and
> > actual GCC behaviour.
> 
> The documentation simply doesn't say "+m" is allowed.  The code to
> allow it was added for the benefit of people who do not read the
> documentation.  Documentation for "+m" might get added later if it
> is decided this [the code, not the documentation] is a sane thing
> to have (which isn't directly obvious).

Huh?

"If the (current) documentation doesn't match up with the (current)
code, then _at least one_ of them has to be (as of current) wrong."

I wonder how could you even try to disagree with that.

And I didn't go whining about this ... you asked me. (I think I'd said
something to the effect of GCC docs are often wrong, which is true,
but probably you feel saying that is "not allowed" on non-gcc lists?)

As for the "PR" you're requesting me to file with GCC for this, that
gcc-patches@ thread did precisely that and more (submitted a patch to
said documentation -- and no, saying "documentation might get added
later" is totally bogus and nonsensical -- documentation exists to
document current behaviour, not past). But come on, this is wholly
petty. I wouldn't have replied, really, if you weren't so provoking.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> The documentation simply doesn't say "+m" is allowed.  The code to
>> allow it was added for the benefit of people who do not read the
>> documentation.  Documentation for "+m" might get added later if it
>> is decided this [the code, not the documentation] is a sane thing
>> to have (which isn't directly obvious).
>
> Huh?
>
> "If the (current) documentation doesn't match up with the (current)
> code, then _at least one_ of them has to be (as of current) wrong."
>
> I wonder how could you even try to disagree with that.

Easy.

The GCC documentation you're referring to is the user's manual.
See the blurb on the first page:

"This manual documents how to use the GNU compilers, as well as their
features and incompatibilities, and how to report bugs.  It corresponds
to GCC version 4.3.0.  The internals of the GNU compilers, including
how to port them to new targets and some information about how to write
front ends for new languages, are documented in a separate manual."

_How to use_.  This documentation doesn't describe in minute detail
everything the compiler does (see the source code for that -- no, it
isn't described in the internals manual either).

If it doesn't tell you how to use "+m", and even tells you _not_ to
use it, maybe that is what it means to say?  It doesn't mean "+m"
doesn't actually do something.  It also doesn't mean it does what
you think it should do.  It might do just that of course.  But treating
writing C code as an empirical science isn't such a smart idea.

> And I didn't go whining about this ... you asked me. (I think I'd said
> something to the effect of GCC docs are often wrong,

No need to guess at what you said, even if you managed to delete
your own mail already, there are plenty of free web-based archives
around.  You said:

> See, "volatile" C keyword, for all it's ill-definition and dodgy
> semantics, is still at least given somewhat of a treatment in the C
> standard (whose quality is ... ummm, sadly not always good and clear,
> but unsurprisingly, still about 5,482 orders-of-magnitude times
> better than GCC docs).

and that to me reads as complaining that the ISO C standard "isn't
very good" and that the GCC documentation is 10**5482 times worse
even.  Which of course is hyperbole and cannot be true.  It also
isn't helpful in any way or form for anyone on this list.  I call
that whining.

> which is true,

Yes, documentation of that size often has shortcomings.  No surprise
there.  However, great effort is made to make it better documentation,
and especially to keep it up to date; if you find any errors or
omissions, please report them.  There are many ways how to do that,
see the GCC homepage.</end-of-marketing-blurb>

> but probably you feel saying that is "not allowed" on non-gcc lists?)

You're allowed to say whatever you want.  Let's have a quote again
shall we?  I said:

> If you find any problems/shortcomings in the GCC documentation,
> please file a PR, don't go whine on some unrelated mailing lists.
> Thank you.

I read that as a friendly request, not a prohibition.  Well maybe
not actually friendly, more a bit angry.  A request, either way.

> As for the "PR"

"Problem report", a bugzilla ticket.  Sorry for using terminology
unknown to you.

> you're requesting me to file with GCC for this, that
> gcc-patches@ thread did precisely that

Actually not -- PRs make sure issues aren't forgotten (although
they might gather dust, sure).  But yes, submitting patches is a
Great Thing(tm).

> and more (submitted a patch to
> said documentation -- and no, saying "documentation might get added
> later" is totally bogus and nonsensical -- documentation exists to
> document current behaviour, not past).

When code like you want to write becomes a supported feature, that
will be reflected in the user manual.  It is completely nonsensical
to expect everything that is *not* a supported feature to be mentioned
there.

> I wouldn't have replied, really, if you weren't so provoking.

Hey, maybe that character trait is good for something, then.
Now to build a business plan around it...


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

[ LOL, you _are_ shockingly petty! ]


On Sat, 18 Aug 2007, Segher Boessenkool wrote:

> > > The documentation simply doesn't say "+m" is allowed.  The code to
> > > allow it was added for the benefit of people who do not read the
> > > documentation.  Documentation for "+m" might get added later if it
> > > is decided this [the code, not the documentation] is a sane thing
> > > to have (which isn't directly obvious).
> > 
> > Huh?
> > 
> > "If the (current) documentation doesn't match up with the (current)
> > code, then _at least one_ of them has to be (as of current) wrong."
> > 
> > I wonder how could you even try to disagree with that.
> 
> Easy.
> 
> The GCC documentation you're referring to is the user's manual.
> See the blurb on the first page:
> 
> "This manual documents how to use the GNU compilers, as well as their
> features and incompatibilities, and how to report bugs.  It corresponds
> to GCC version 4.3.0.  The internals of the GNU compilers, including
> how to port them to new targets and some information about how to write
> front ends for new languages, are documented in a separate manual."
> 
> _How to use_.  This documentation doesn't describe in minute detail
> everything the compiler does (see the source code for that -- no, it
> isn't described in the internals manual either).

Wow, now that's a nice "disclaimer". By your (poor) standards of writing
documentation, one can as well write any factually incorrect stuff that
one wants in a document once you've got such a blurb in place :-)


> If it doesn't tell you how to use "+m", and even tells you _not_ to
> use it, maybe that is what it means to say?  It doesn't mean "+m"
> doesn't actually do something.  It also doesn't mean it does what
> you think it should do.  It might do just that of course.  But treating
> writing C code as an empirical science isn't such a smart idea.

Oh, really? Considering how much is (left out of being) documented, often
one would reasonably have to experimentally see (with testcases) how the
compiler behaves for some given code. Well, at least _I_ do it often
(several others on this list do as well), and I think there's everything
smart about it rather than having to read gcc sources -- I'd be surprised
(unless you have infinite free time on your hands, which does look like
teh case actually) if someone actually prefers reading gcc sources first
to know what/how gcc does something for some given code, rather than
simply write it out, compile and look the generated code (saves time for
those who don't have an infinite amount of it).


> > And I didn't go whining about this ... you asked me. (I think I'd said
> > something to the effect of GCC docs are often wrong,
> 
> No need to guess at what you said, even if you managed to delete
> your own mail already, there are plenty of free web-based archives
> around.  You said:
> 
> > See, "volatile" C keyword, for all it's ill-definition and dodgy
> > semantics, is still at least given somewhat of a treatment in the C
> > standard (whose quality is ... ummm, sadly not always good and clear,
> > but unsurprisingly, still about 5,482 orders-of-magnitude times
> > better than GCC docs).

Try _reading_ what I said there, for a change, dude. I'd originally only
said "unless GCC's docs is yet again wrong" ... then _you_ asked me what,
after which this discussion began and I wrote the above [which I fully
agree with -- so what if I used hyperbole in my sentence (yup, that was
intended, and obviously, exaggeration), am I not even allowed to do that?
Man, you're a Nazi or what ...] I didn't go whining about on my own as
you'd had earlier suggested, until _you_ asked me.

[ Ick, I somehow managed to reply this ... this is such a ...
  *disgustingly* petty argument you made here. ]


> > which is true,
> 
> Yes, documentation of that size often has shortcomings.  No surprise
> there.  However, great effort is made to make it better documentation,
> and especially to keep it up to date; if you find any errors or
> omissions, please report them.  There are many ways how to do that,
> see the GCC homepage.</end-of-marketing-blurb>
                         ^^^^^^^^^^^^^^^^^^^^^^

Looks like you even get paid :-)


> > but probably you feel saying that is "not allowed" on non-gcc lists?)
> 
> [amazingly pointless stuff snipped]
> 
> > As for the "PR"
> > you're requesting me to file with GCC for this, that
> > gcc-patches@ thread did precisely that
> 
> [more amazingly pointless stuff snipped]
> 
> > and more (submitted a patch to
> > said documentation -- and no, saying "documentation might get added
> > later" is totally bogus and nonsensical -- documentation exists to
> > document current behaviour, not past).
> 
> When code like you want to write becomes a supported feature, that
> will be reflected in the user manual.  It is completely nonsensical
> to expect everything that is *not* a supported feature to be mentioned
> there.

What crap. It is _perfectly reasonable_ to expect (current) documentation
to keep up with (current) code behaviour. In fact trying to justify such
a state is completely bogus and nonsensical.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Fri, 17 Aug 2007, Segher Boessenkool wrote:

> "volatile" has nothing to do with reordering.  atomic_dec() writes
> to memory, so it _does_ have "volatile semantics", implicitly, as
> long as the compiler cannot optimise the atomic variable away
> completely -- any store counts as a side effect.

Stores can be reordered. Only x86 has (mostly) implicit write ordering. So 
no atomic_dec has no volatile semantics and may be reordered on a variety 
of processors. Writes to memory may not follow code order on several 
processors.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Mon, Sep 10, 2007 at 11:59:29AM -0700, Christoph Lameter wrote:
> On Fri, 17 Aug 2007, Segher Boessenkool wrote:
> 
> > "volatile" has nothing to do with reordering.  atomic_dec() writes
> > to memory, so it _does_ have "volatile semantics", implicitly, as
> > long as the compiler cannot optimise the atomic variable away
> > completely -- any store counts as a side effect.
> 
> Stores can be reordered. Only x86 has (mostly) implicit write ordering. So 
> no atomic_dec has no volatile semantics and may be reordered on a variety 
> of processors. Writes to memory may not follow code order on several 
> processors.

The one exception to this being the case where process-level code is
communicating to an interrupt handler running on that same CPU -- on
all CPUs that I am aware of, a given CPU always sees its own writes
in order.

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Mon, 10 Sep 2007, Paul E. McKenney wrote:

> The one exception to this being the case where process-level code is
> communicating to an interrupt handler running on that same CPU -- on
> all CPUs that I am aware of, a given CPU always sees its own writes
> in order.

Yes but that is due to the code path effectively continuing in the 
interrupt handler. The cpu makes sure that op codes being executed always 
see memory in a consistent way. The basic ordering problem with out of 
order writes is therefore coming from other processors concurrently 
executing code and holding variables in registers that are modified 
elsewhere. The only solution that I know of are one or the other form of 
barrier.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Mon, Sep 10, 2007 at 02:36:26PM -0700, Christoph Lameter wrote:
> On Mon, 10 Sep 2007, Paul E. McKenney wrote:
> 
> > The one exception to this being the case where process-level code is
> > communicating to an interrupt handler running on that same CPU -- on
> > all CPUs that I am aware of, a given CPU always sees its own writes
> > in order.
> 
> Yes but that is due to the code path effectively continuing in the 
> interrupt handler. The cpu makes sure that op codes being executed always 
> see memory in a consistent way. The basic ordering problem with out of 
> order writes is therefore coming from other processors concurrently 
> executing code and holding variables in registers that are modified 
> elsewhere. The only solution that I know of are one or the other form of 
> barrier.

So we are agreed then -- volatile accesses may be of some assistance when
interacting with interrupt handlers running on the same CPU (presumably
when using per-CPU variables), but are generally useless when sharing
variables among CPUs.  Correct?

							Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> "volatile" has nothing to do with reordering.  atomic_dec() writes
>> to memory, so it _does_ have "volatile semantics", implicitly, as
>> long as the compiler cannot optimise the atomic variable away
>> completely -- any store counts as a side effect.
>
> Stores can be reordered. Only x86 has (mostly) implicit write ordering.
> So no atomic_dec has no volatile semantics

Read again: I said the C "volatile" construct has nothing to do
with CPU memory access reordering.

> and may be reordered on a variety
> of processors. Writes to memory may not follow code order on several
> processors.

The _compiler_ isn't allowed to reorder things here.  Yes, of course
you do need stronger barriers for many purposes, volatile isn't all
that useful you know.


Segher

RE: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Luck, Tony]

>> 6674:   while (atomic_read(&j->DSPWrite) > 0)
>> 6675-           atomic_dec(&j->DSPWrite);
>
> If the maintainer of this code doesn't see a compelling reason to add 
> cpu_relax() in this loop, then it should be patched.

Shouldn't it be just re-written without the loop:

	if ((tmp = atomic_read(&j->DSPWrite)) > 0)
		atomic_sub(&j->DSPWrite, tmp);

Has all the same bugs, but runs much faster :-)

-Tony

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Chris Snook]

Ilpo Järvinen wrote:
> On Thu, 16 Aug 2007, Herbert Xu wrote:
> 
>> We've been through that already.  If it's a busy-wait it
>> should use cpu_relax. 
> 
> I looked around a bit by using some command lines and ended up wondering 
> if these are equal to busy-wait case (and should be fixed) or not:
> 
> ./drivers/telephony/ixj.c
> 6674:   while (atomic_read(&j->DSPWrite) > 0)
> 6675-           atomic_dec(&j->DSPWrite);
> 
> ...besides that, there are couple of more similar cases in the same file 
> (with braces)...

atomic_dec() already has volatile behavior everywhere, so this is 
semantically okay, but this code (and any like it) should be calling 
cpu_relax() each iteration through the loop, unless there's a compelling 
reason not to.  I'll allow that for some hardware drivers (possibly this 
one) such a compelling reason may exist, but hardware-independent core 
subsystems probably have no excuse.

If the maintainer of this code doesn't see a compelling reason not to 
add cpu_relax() in this loop, then it should be patched.

	-- Chris

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Thu, 16 Aug 2007, Paul Mackerras wrote:

> The uses of atomic_read where one might want it to allow caching of
> the result seem to me to fall into 3 categories:
> 
> 1. Places that are buggy because of a race arising from the way it's
>    used.
> 
> 2. Places where there is a race but it doesn't matter because we're
>    doing some clever trick.
> 
> 3. Places where there is some locking in place that eliminates any
>    potential race.
> 
> In case 1, adding volatile won't solve the race, of course, but it's
> hard to argue that we shouldn't do something because it will slow down
> buggy code.  Case 2 is hopefully pretty rare and accompanied by large
> comment blocks, and in those cases caching the result of atomic_read
> explicitly in a local variable would probably make the code clearer.
> And in case 3 there is no reason to use atomic_t at all; we might as
> well just use an int.

In 2 + 3 you may increment the atomic variable in some places. The value 
of the atomic variable may not matter because you only do optimizations.

Checking a atomic_t for a definite state has to involve either
some side conditions (lock only taken if refcount is <= 0 or so) or done 
by changing the state (see f.e. atomic_inc_unless_zero).

> So I don't see any good reason to make the atomic API more complex by
> having "volatile" and "non-volatile" versions of atomic_read.  It
> should just have the "volatile" behaviour.

If you want to make it less complex then drop volatile which causes weird 
side effects without solving any problems as you just pointed out.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul E. McKenney]

On Thu, Aug 16, 2007 at 11:54:54AM -0700, Christoph Lameter wrote:
> On Thu, 16 Aug 2007, Paul Mackerras wrote:
> > So I don't see any good reason to make the atomic API more complex by
> > having "volatile" and "non-volatile" versions of atomic_read.  It
> > should just have the "volatile" behaviour.
> 
> If you want to make it less complex then drop volatile which causes weird 
> side effects without solving any problems as you just pointed out.

The other set of problems are communication between process context
and interrupt/NMI handlers.  Volatile does help here.  And the performance
impact of volatile is pretty near zero, so why have the non-volatile
variant?

						Thanx, Paul

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Paul Mackerras]

Satyam Sharma writes:

> I can't speak for this particular case, but there could be similar code
> examples elsewhere, where we do the atomic ops on an atomic_t object
> inside a higher-level locking scheme that would take care of the kind of
> problem you're referring to here. It would be useful for such or similar
> code if the compiler kept the value of that atomic object in a register.

If there is a higher-level locking scheme then there is no point to
using atomic_t variables.  Atomic_t is specifically for the situation
where multiple CPUs are updating a variable without locking.

Paul.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> I can't speak for this particular case, but there could be similar 
>> code
>> examples elsewhere, where we do the atomic ops on an atomic_t object
>> inside a higher-level locking scheme that would take care of the kind 
>> of
>> problem you're referring to here. It would be useful for such or 
>> similar
>> code if the compiler kept the value of that atomic object in a 
>> register.
>
> If there is a higher-level locking scheme then there is no point to
> using atomic_t variables.  Atomic_t is specifically for the situation
> where multiple CPUs are updating a variable without locking.

And don't forget about the case where it is an I/O device that is
updating the memory (in buffer descriptors or similar).  The driver
needs to do a "volatile" atomic read to get at the most recent version
of that data, which can be important for optimising latency (or 
throughput
even).  There is no other way the kernel can get that info -- doing an
MMIO read is way way too expensive.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>> A volatile default would disable optimizations for atomic_read.
>> atomic_read without volatile would allow for full optimization by the
>> compiler. Seems that this is what one wants in many cases.
>
> Name one such case.
>
> An atomic_read should do a load from memory.  If the programmer puts
> an atomic_read() in the code then the compiler should emit a load for
> it, not re-use a value returned by a previous atomic_read.  I do not
> believe it would ever be useful for the compiler to collapse two
> atomic_read statements into a single load.

An atomic_read() implemented as a "normal" C variable read would
allow that read to be combined with another "normal" read from
that variable.  This could perhaps be marginally useful, although
I'd bet you cannot see it unless counting cycles on a simulator
or counting bits in the binary size.

With an asm() implementation, the compiler can not do this; with
a "volatile" implementation (either volatile variable or volatile-cast),
this invokes undefined behaviour (in both C and GCC).


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Thursday 16 August 2007 00:22, Paul Mackerras wrote:
> Satyam Sharma writes:
> In the kernel we use atomic variables in precisely those situations
> where a variable is potentially accessed concurrently by multiple
> CPUs, and where each CPU needs to see updates done by other CPUs in a
> timely fashion.  That is what they are for.  Therefore the compiler
> must not cache values of atomic variables in registers; each
> atomic_read must result in a load and each atomic_set must result in a
> store.  Anything else will just lead to subtle bugs.

Amen.
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Fri, 24 Aug 2007, Denys Vlasenko wrote:

> On Thursday 16 August 2007 00:22, Paul Mackerras wrote:
> > Satyam Sharma writes:
> > In the kernel we use atomic variables in precisely those situations
> > where a variable is potentially accessed concurrently by multiple
> > CPUs, and where each CPU needs to see updates done by other CPUs in a
> > timely fashion.  That is what they are for.  Therefore the compiler
> > must not cache values of atomic variables in registers; each
> > atomic_read must result in a load and each atomic_set must result in a
> > store.  Anything else will just lead to subtle bugs.
> 
> Amen.

A "timely" fashion? One cannot rely on something like that when coding. 
The visibility of updates is insured by barriers and not by some fuzzy 
notion of "timeliness".

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Denys Vlasenko]

On Friday 24 August 2007 18:15, Christoph Lameter wrote:
> On Fri, 24 Aug 2007, Denys Vlasenko wrote:
> > On Thursday 16 August 2007 00:22, Paul Mackerras wrote:
> > > Satyam Sharma writes:
> > > In the kernel we use atomic variables in precisely those situations
> > > where a variable is potentially accessed concurrently by multiple
> > > CPUs, and where each CPU needs to see updates done by other CPUs in a
> > > timely fashion.  That is what they are for.  Therefore the compiler
> > > must not cache values of atomic variables in registers; each
> > > atomic_read must result in a load and each atomic_set must result in a
> > > store.  Anything else will just lead to subtle bugs.
> >
> > Amen.
>
> A "timely" fashion? One cannot rely on something like that when coding.
> The visibility of updates is insured by barriers and not by some fuzzy
> notion of "timeliness".

But here you do have some notion of time:

	while (atomic_read(&x))
		continue;

"continue when other CPU(s) decrement it down to zero".
If "read" includes an insn which accesses RAM, you will
see "new" value sometime after other CPU decrements it.
"Sometime after" is on the order of nanoseconds here.
It is a valid concept of time, right?

The whole confusion is about whether atomic_read implies
"read from RAM" or not. I am in a camp which thinks it does.
You are in an opposite one.

We just need a less ambiguous name.

What about this:

/**
 * atomic_read - read atomic variable
 * @v: pointer of type atomic_t
 *
 * Atomically reads the value of @v.
 * No compiler barrier implied.
 */
#define atomic_read(v)          ((v)->counter)

+/**
+ * atomic_read_uncached - read atomic variable from memory
+ * @v: pointer of type atomic_t
+ *
+ * Atomically reads the value of @v. This is guaranteed to emit an insn
+ * which accesses memory, atomically. No ordering guarantees!
+ */
+#define atomic_read_uncached(v)  asm_or_volatile_ptr_magic(v)

I was thinking of s/atomic_read/atomic_get/ too, but it implies "taking"
atomic a-la get_cpu()...
--
vda

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Bill Fink]

On Wed, 15 Aug 2007, Satyam Sharma wrote:

> (C)
> $ cat tp3.c
> int a;
> 
> void func(void)
> {
> 	*(volatile int *)&a = 10;
> 	*(volatile int *)&a = 20;
> }
> $ gcc -Os -S tp3.c
> $ cat tp3.s
> ...
> movl    $10, a
> movl    $20, a
> ...

I'm curious about one minor tangential point.  Why, instead of:

	b = *(volatile int *)&a;

why can't this just be expressed as:

	b = (volatile int)a;

Isn't it the contents of a that's volatile, i.e. it's value can change
invisibly to the compiler, and that's why you want to force a read from
memory?  Why do you need the "*(volatile int *)&" construct?

						-Bill

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

Hi Bill,


On Wed, 15 Aug 2007, Bill Fink wrote:

> On Wed, 15 Aug 2007, Satyam Sharma wrote:
> 
> > (C)
> > $ cat tp3.c
> > int a;
> > 
> > void func(void)
> > {
> > 	*(volatile int *)&a = 10;
> > 	*(volatile int *)&a = 20;
> > }
> > $ gcc -Os -S tp3.c
> > $ cat tp3.s
> > ...
> > movl    $10, a
> > movl    $20, a
> > ...
> 
> I'm curious about one minor tangential point.  Why, instead of:
> 
> 	b = *(volatile int *)&a;
> 
> why can't this just be expressed as:
> 
> 	b = (volatile int)a;
> 
> Isn't it the contents of a that's volatile, i.e. it's value can change
> invisibly to the compiler, and that's why you want to force a read from
> memory?  Why do you need the "*(volatile int *)&" construct?

"b = (volatile int)a;" doesn't help us because a cast to a qualified type
has the same effect as a cast to an unqualified version of that type, as
mentioned in 6.5.4:4 (footnote 86) of the standard. Note that "volatile"
is a type-qualifier, not a type itself, so a cast of the _object_ itself
to a qualified-type i.e. (volatile int) would not make the access itself
volatile-qualified.

To serve our purposes, it is necessary for us to take the address of this
(non-volatile) object, cast the resulting _pointer_ to the corresponding
volatile-qualified pointer-type, and then dereference it. This makes that
particular _access_ be volatile-qualified, without the object itself being
such. Also note that the (dereferenced) result is also a valid lvalue and
hence can be used in "*(volatile int *)&a = b;" kind of construction
(which we use for the atomic_set case).


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Thu, 16 Aug 2007, Satyam Sharma wrote:

> Hi Bill,
> 
> 
> On Wed, 15 Aug 2007, Bill Fink wrote:
> 
> > On Wed, 15 Aug 2007, Satyam Sharma wrote:
> > 
> > > (C)
> > > $ cat tp3.c
> > > int a;
> > > 
> > > void func(void)
> > > {
> > > 	*(volatile int *)&a = 10;
> > > 	*(volatile int *)&a = 20;
> > > }
> > > $ gcc -Os -S tp3.c
> > > $ cat tp3.s
> > > ...
> > > movl    $10, a
> > > movl    $20, a
> > > ...
> > 
> > I'm curious about one minor tangential point.  Why, instead of:
> > 
> > 	b = *(volatile int *)&a;
> > 
> > why can't this just be expressed as:
> > 
> > 	b = (volatile int)a;
> > 
> > Isn't it the contents of a that's volatile, i.e. it's value can change
> > invisibly to the compiler, and that's why you want to force a read from
> > memory?  Why do you need the "*(volatile int *)&" construct?
> 
> "b = (volatile int)a;" doesn't help us because a cast to a qualified type
> has the same effect as a cast to an unqualified version of that type, as
> mentioned in 6.5.4:4 (footnote 86) of the standard. Note that "volatile"
> is a type-qualifier, not a type itself, so a cast of the _object_ itself
> to a qualified-type i.e. (volatile int) would not make the access itself
> volatile-qualified.
> 
> To serve our purposes, it is necessary for us to take the address of this
> (non-volatile) object, cast the resulting _pointer_ to the corresponding
> volatile-qualified pointer-type, and then dereference it. This makes that
> particular _access_ be volatile-qualified, without the object itself being
> such. Also note that the (dereferenced) result is also a valid lvalue and
> hence can be used in "*(volatile int *)&a = b;" kind of construction
> (which we use for the atomic_set case).

Here, I should obviously admit that the semantics of *(volatile int *)&
aren't any neater or well-defined in the _language standard_ at all. The
standard does say (verbatim) "precisely what constitutes as access to
object of volatile-qualified type is implementation-defined", but GCC
does help us out here by doing the right thing. Accessing the non-volatile
object there using the volatile-qualified pointer-type cast makes GCC
treat the object stored at that memory address itself as if it were a 
volatile object, thus making the access end up having what we're calling
as "volatility" semantics here.

Honestly, given such confusion, and the propensity of the "volatile"
type-qualifier keyword to be ill-defined (or at least poorly understood,
often inconsistently implemented), I'd (again) express my opinion that it
would be best to avoid its usage, given other alternatives do exist.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

[ Bill tells me in private communication he gets this already, but I
  think it's more complicated than the shoddy explanation I'd made
  earlier so would wish to make this clearer in detail one last time,
  for the benefit of others listening in or reading the archives. ]


On Thu, 16 Aug 2007, Satyam Sharma wrote:

> On Thu, 16 Aug 2007, Satyam Sharma wrote:
> [...]
> > On Wed, 15 Aug 2007, Bill Fink wrote:
> > > [...]
> > > I'm curious about one minor tangential point.  Why, instead of:
> > > 
> > > 	b = *(volatile int *)&a;
> > > 
> > > why can't this just be expressed as:
> > > 
> > > 	b = (volatile int)a;
> > > 
> > > Isn't it the contents of a that's volatile, i.e. it's value can change
> > > invisibly to the compiler, and that's why you want to force a read from
> > > memory?  Why do you need the "*(volatile int *)&" construct?
> > 
> > "b = (volatile int)a;" doesn't help us because a cast to a qualified type
> > has the same effect as a cast to an unqualified version of that type, as
> > mentioned in 6.5.4:4 (footnote 86) of the standard. Note that "volatile"
> > is a type-qualifier, not a type itself, so a cast of the _object_ itself
> > to a qualified-type i.e. (volatile int) would not make the access itself
> > volatile-qualified.

Casts don't produce lvalues, and the cast ((volatile int)a) does not
produce the object-int-a-qualified-as-"volatile" -- in fact, the
result of the above cast is whatever is the _value_ of "int a", with
the access to that object having _already_ taken place, as per the
actual type-qualification of the object (that was originally declared
as being _non-volatile_, in fact). Hence, defining atomic_read() as:

#define atomic_read(v)          ((volatile int)((v)->counter))

would be buggy and not give "volatility" semantics at all, unless the
"counter" object itself isn't volatile-qualified already (which it
isn't).

The result of the cast itself being the _value_ of the int object, and
not the object itself (i.e., not an lvalue), is thereby independent of
type-qualification in that cast itself (it just wouldn't make any
difference), hence the "cast to a qualified type has the same effect
as a cast to an unqualified version of that type" bit in section 6.5.4:4
of the standard.


> > To serve our purposes, it is necessary for us to take the address of this
> > (non-volatile) object, cast the resulting _pointer_ to the corresponding
> > volatile-qualified pointer-type, and then dereference it. This makes that
> > particular _access_ be volatile-qualified, without the object itself being
> > such. Also note that the (dereferenced) result is also a valid lvalue and
> > hence can be used in "*(volatile int *)&a = b;" kind of construction
> > (which we use for the atomic_set case).

Dereferencing using the *(pointer-type-cast)& construct, OTOH, serves
us well:

#define atomic_read(v)          (*(volatile int *)&(v)->counter)

Firstly, note that the cast here being (volatile int *) and not
(int * volatile) qualifies the type of the _object_ being pointed to
by the pointer in question as being volatile-qualified, and not the
pointer itself (6.2.5:27 of the standard, and 6.3.2.3:2 allows us to
convert from a pointer-to-non-volatile-qualified-int to a pointer-to-
volatile-qualified-int, which suits us just fine) -- but note that
the _access_ to that address itself has not yet occurred.

_After_ specifying the memory address as containing a volatile-qualified-
int-type object, (and GCC co-operates as mentioned below), we proceed to
dereference it, which is when the _actual access_ occurs, therefore with
"volatility" semantics this time.

Interesting.


> Here, I should obviously admit that the semantics of *(volatile int *)&
> aren't any neater or well-defined in the _language standard_ at all. The
> standard does say (verbatim) "precisely what constitutes as access to
> object of volatile-qualified type is implementation-defined", but GCC
> does help us out here by doing the right thing. Accessing the non-volatile
> object there using the volatile-qualified pointer-type cast makes GCC
> treat the object stored at that memory address itself as if it were a 
> volatile object, thus making the access end up having what we're calling
> as "volatility" semantics here.
> 
> Honestly, given such confusion, and the propensity of the "volatile"
> type-qualifier keyword to be ill-defined (or at least poorly understood,
> often inconsistently implemented), I'd (again) express my opinion that it
> would be best to avoid its usage, given other alternatives do exist.


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> Here, I should obviously admit that the semantics of *(volatile int *)&
> aren't any neater or well-defined in the _language standard_ at all. 
> The
> standard does say (verbatim) "precisely what constitutes as access to
> object of volatile-qualified type is implementation-defined", but GCC
> does help us out here by doing the right thing.

Where do you get that idea?  GCC manual, section 6.1, "When
is a Volatile Object Accessed?" doesn't say anything of the
kind.  PR33053 and some others.

> Honestly, given such confusion, and the propensity of the "volatile"
> type-qualifier keyword to be ill-defined (or at least poorly 
> understood,
> often inconsistently implemented), I'd (again) express my opinion that 
> it
> would be best to avoid its usage, given other alternatives do exist.

Yeah.  Or we can have an email thread like this every time
someone proposes a patch that uses an atomic variable ;-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Thu, 16 Aug 2007, Segher Boessenkool wrote:

> > Here, I should obviously admit that the semantics of *(volatile int *)&
> > aren't any neater or well-defined in the _language standard_ at all. The
> > standard does say (verbatim) "precisely what constitutes as access to
> > object of volatile-qualified type is implementation-defined", but GCC
> > does help us out here by doing the right thing.
> 
> Where do you get that idea?

Try a testcase (experimentally verify).

> GCC manual, section 6.1, "When
> is a Volatile Object Accessed?" doesn't say anything of the
> kind.

True, "implementation-defined" as per the C standard _is_ supposed to mean
"unspecified behaviour where each implementation documents how the choice
is made". So ok, probably GCC isn't "documenting" this
implementation-defined behaviour which it is supposed to, but can't really
fault them much for this, probably.

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

>>> Here, I should obviously admit that the semantics of *(volatile int 
>>> *)&
>>> aren't any neater or well-defined in the _language standard_ at all. 
>>> The
>>> standard does say (verbatim) "precisely what constitutes as access to
>>> object of volatile-qualified type is implementation-defined", but GCC
>>> does help us out here by doing the right thing.
>>
>> Where do you get that idea?
>
> Try a testcase (experimentally verify).

That doesn't prove anything.  Experiments can only disprove
things.

>> GCC manual, section 6.1, "When
>> is a Volatile Object Accessed?" doesn't say anything of the
>> kind.
>
> True, "implementation-defined" as per the C standard _is_ supposed to 
> mean
> "unspecified behaviour where each implementation documents how the 
> choice
> is made". So ok, probably GCC isn't "documenting" this
> implementation-defined behaviour which it is supposed to, but can't 
> really
> fault them much for this, probably.

GCC _is_ documenting this, namely in this section 6.1.  It doesn't
mention volatile-casted stuff.  Draw your own conclusions.


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Sat, 18 Aug 2007, Segher Boessenkool wrote:

> > > GCC manual, section 6.1, "When
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > is a Volatile Object Accessed?" doesn't say anything of the
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > > kind.
      ^^^^^

> > True, "implementation-defined" as per the C standard _is_ supposed to mean
    ^^^^^

> > "unspecified behaviour where each implementation documents how the choice
> > is made". So ok, probably GCC isn't "documenting" this
                              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

> > implementation-defined behaviour which it is supposed to, but can't really
> > fault them much for this, probably.
> 
> GCC _is_ documenting this, namely in this section 6.1.

(Again totally petty, but) Yes, but ...

> It doesn't
  ^^^^^^^^^^
> mention volatile-casted stuff.  Draw your own conclusions.
  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

... exactly. So that's why I said "GCC isn't documenting _this_".

Man, try _reading_ mails before replying to them ...

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> Note that "volatile"
> is a type-qualifier, not a type itself, so a cast of the _object_ 
> itself
> to a qualified-type i.e. (volatile int) would not make the access 
> itself
> volatile-qualified.

There is no such thing as "volatile-qualified access" defined
anywhere; there only is the concept of a "volatile-qualified
*object*".

> To serve our purposes, it is necessary for us to take the address of 
> this
> (non-volatile) object, cast the resulting _pointer_ to the 
> corresponding
> volatile-qualified pointer-type, and then dereference it. This makes 
> that
> particular _access_ be volatile-qualified, without the object itself 
> being
> such. Also note that the (dereferenced) result is also a valid lvalue 
> and
> hence can be used in "*(volatile int *)&a = b;" kind of construction
> (which we use for the atomic_set case).

There is a quite convincing argument that such an access _is_ an
access to a volatile object; see GCC PR21568 comment #9.  This
probably isn't the last word on the matter though...


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Satyam Sharma]

On Thu, 16 Aug 2007, Segher Boessenkool wrote:

> > Note that "volatile"
> > is a type-qualifier, not a type itself, so a cast of the _object_ itself
> > to a qualified-type i.e. (volatile int) would not make the access itself
> > volatile-qualified.
> 
> There is no such thing as "volatile-qualified access" defined
> anywhere; there only is the concept of a "volatile-qualified
> *object*".

Sure, "volatile-qualified access" was not some standard term I used
there. Just something to mean "an access that would make the compiler
treat the object at that memory as if it were an object with a
volatile-qualified type".

Now the second wording *IS* technically correct, but come on, it's
24 words long whereas the original one was 3 -- and hopefully anybody
reading the shorter phrase *would* have known anyway what was meant,
without having to be pedantic about it :-)


Satyam

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Segher Boessenkool]

> Now the second wording *IS* technically correct, but come on, it's
> 24 words long whereas the original one was 3 -- and hopefully anybody
> reading the shorter phrase *would* have known anyway what was meant,
> without having to be pedantic about it :-)

Well you were talking pretty formal (and detailed) stuff, so
IMHO it's good to have that exactly correct.  Sure it's nicer
to use small words most of the time :-)


Segher

Re: [PATCH 0/24] make atomic_read() behave consistently across all architectures

[Christoph Lameter]

On Wed, 15 Aug 2007, Stefan Richter wrote:

> LDD3 says on page 125:  "The following operations are defined for the
> type [atomic_t] and are guaranteed to be atomic with respect to all
> processors of an SMP computer."
> 
> Doesn't "atomic WRT all processors" require volatility?

Atomic operations only require exclusive access to the cacheline while the 
value is modified.

(no subject)

[arnd]

From arnd@arndb.de Mon Apr 27 16:28:40 2009
References: <20090427142010.587518220@arndb.de>
User-Agent: quilt/0.46-1
Date: Mon, 27 Apr 2009 16:20:16 +0200
From: arnd@arndb.de
To: linux-kernel@vger.kernel.org
Cc: john.williams@petalogix.com,
 monstr@monstr.eu,
 linux-api@vger.kernel.org,
 linux-arch@vger.kernel.org,
 liqin.chen@sunplusct.com,
 Sam Ravnborg <sam@ravnborg.org>,
 Remis Lima Baima <remis.developer@googlemail.com>
Subject: [RFC 06/17] asm-generic: add a complete signal.h
Content-Disposition: inline; filename=add-generic-signal-h.patch
X-Provags-ID: V01U2FsdGVkX1+9fyV5xiB/uv57e+JOHz1p/SO7dXImpty2qjk
 qJlT1Ill8RsfbwF39tmrVDe2R5QILPHgAXX1/8KQ3m3F7IAuZ2
 mpIv26Pn+3mqK9UXgjM9g==

There used to be an asm-generic/signal.h, but not one that
could be used out of the box. This renames the existing file
to asm-generic/signal-defs.h and adds a new asm-generic/signal.h
with common definitions for sigaction and sigaltstack, but
not for the old-style signal handlers.

New architectures should only implement rt-signals and use this
generic header.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Remis Lima Baima <remis.developer@googlemail.com>
---
---
 include/asm-generic/Kbuild   |    1 	1 +	0 -	0 !
 include/asm-generic/signal.h |  131 	131 +	0 -	0 !
 2 files changed, 132 insertions(+)

Index: linux-2.6/include/asm-generic/signal.h
===================================================================
--- /dev/null
+++ linux-2.6/include/asm-generic/signal.h
@@ -0,0 +1,131 @@
+#ifndef __ASM_GENERIC_SIGNAL_H
+#define __ASM_GENERIC_SIGNAL_H
+
+#include <linux/types.h>
+
+#define _NSIG		64
+#define _NSIG_BPW	__BITS_PER_LONG
+#define _NSIG_WORDS	(_NSIG / _NSIG_BPW)
+
+#define SIGHUP		 1
+#define SIGINT		 2
+#define SIGQUIT		 3
+#define SIGILL		 4
+#define SIGTRAP		 5
+#define SIGABRT		 6
+#define SIGIOT		 6
+#define SIGBUS		 7
+#define SIGFPE		 8
+#define SIGKILL		 9
+#define SIGUSR1		10
+#define SIGSEGV		11
+#define SIGUSR2		12
+#define SIGPIPE		13
+#define SIGALRM		14
+#define SIGTERM		15
+#define SIGSTKFLT	16
+#define SIGCHLD		17
+#define SIGCONT		18
+#define SIGSTOP		19
+#define SIGTSTP		20
+#define SIGTTIN		21
+#define SIGTTOU		22
+#define SIGURG		23
+#define SIGXCPU		24
+#define SIGXFSZ		25
+#define SIGVTALRM	26
+#define SIGPROF		27
+#define SIGWINCH	28
+#define SIGIO		29
+#define SIGPOLL		SIGIO
+/*
+#define SIGLOST		29
+*/
+#define SIGPWR		30
+#define SIGSYS		31
+#define	SIGUNUSED	31
+
+/* These should not be considered constants from userland.  */
+#define SIGRTMIN	32
+#ifndef SIGRTMAX
+#define SIGRTMAX	_NSIG
+#endif
+
+/*
+ * SA_FLAGS values:
+ *
+ * SA_ONSTACK indicates that a registered stack_t will be used.
+ * SA_RESTART flag to get restarting signals (which were the default long ago)
+ * SA_NOCLDSTOP flag to turn off SIGCHLD when children stop.
+ * SA_RESETHAND clears the handler when the signal is delivered.
+ * SA_NOCLDWAIT flag on SIGCHLD to inhibit zombies.
+ * SA_NODEFER prevents the current signal from being masked in the handler.
+ *
+ * SA_ONESHOT and SA_NOMASK are the historical Linux names for the Single
+ * Unix names RESETHAND and NODEFER respectively.
+ */
+#define SA_NOCLDSTOP	0x00000001
+#define SA_NOCLDWAIT	0x00000002
+#define SA_SIGINFO	0x00000004
+#define SA_ONSTACK	0x08000000
+#define SA_RESTART	0x10000000
+#define SA_NODEFER	0x40000000
+#define SA_RESETHAND	0x80000000
+
+#define SA_NOMASK	SA_NODEFER
+#define SA_ONESHOT	SA_RESETHAND
+
+/*
+ * New architectures should not define the obsolete
+ *	SA_RESTORER	0x04000000
+ */
+
+/*
+ * sigaltstack controls
+ */
+#define SS_ONSTACK	1
+#define SS_DISABLE	2
+
+#define MINSIGSTKSZ	2048
+#define SIGSTKSZ	8192
+
+#ifndef __ASSEMBLY__
+typedef struct {
+	unsigned long sig[_NSIG_WORDS];
+} sigset_t;
+
+/* not actually used, but required for linux/syscalls.h */
+typedef unsigned long old_sigset_t;
+
+#include <asm-generic/signal-defs.h>
+
+struct sigaction {
+	__sighandler_t sa_handler;
+	unsigned long sa_flags;
+#ifdef SA_RESTORER
+	__sigrestore_t sa_restorer;
+#endif
+	sigset_t sa_mask;		/* mask last for extensibility */
+};
+
+struct k_sigaction {
+	struct sigaction sa;
+};
+
+typedef struct sigaltstack {
+	void __user *ss_sp;
+	int ss_flags;
+	size_t ss_size;
+} stack_t;
+
+#ifdef __KERNEL__
+
+#include <asm/sigcontext.h>
+#undef __HAVE_ARCH_SIG_BITOPS
+
+#define ptrace_signal_deliver(regs, cookie) do { } while (0)
+
+#endif /* __KERNEL__ */
+#endif /* __ASSEMBLY__ */
+
+#endif /* _ASM_GENERIC_SIGNAL_H */
Index: linux-2.6/include/asm-generic/Kbuild
===================================================================
--- linux-2.6.orig/include/asm-generic/Kbuild
+++ linux-2.6/include/asm-generic/Kbuild
@@ -5,6 +5,7 @@ header-y += ioctl.h
 header-y += mman-common.h
 header-y += poll.h
 header-y += signal-defs.h
+header-y += signal.h
 header-y += statfs.h
 header-y += termios.h
 

-- 

Re:

[Geert Uytterhoeven]

[bummer, reply-to-all broken]

On Mon, Apr 27, 2009 at 16:41,  <arnd@arndb.de> wrote:
> Subject: [RFC 06/17] asm-generic: add a complete signal.h

> There used to be an asm-generic/signal.h, but not one that
> could be used out of the box. This renames the existing file
> to asm-generic/signal-defs.h and adds a new asm-generic/signal.h

The rename was done in another patch, but you forgot to update this
patch description?

> with common definitions for sigaction and sigaltstack, but
> not for the old-style signal handlers.
>
> New architectures should only implement rt-signals and use this
> generic header.
>
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> Signed-off-by: Remis Lima Baima <remis.developer@googlemail.com>
> ---
> ---
>  include/asm-generic/Kbuild   |    1    1 +     0 -     0 !
>  include/asm-generic/signal.h |  131    131 +   0 -     0 !

Gr{oetje,eeting}s,

						Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
							    -- Linus Torvalds

(no subject)

[arnd]

From arnd@arndb.de Mon Apr 27 16:28:40 2009
References: <20090427142010.587518220@arndb.de>
User-Agent: quilt/0.46-1
Date: Mon, 27 Apr 2009 16:20:14 +0200
From: arnd@arndb.de
To: linux-kernel@vger.kernel.org
Cc: john.williams@petalogix.com,
 monstr@monstr.eu,
 linux-api@vger.kernel.org,
 linux-arch@vger.kernel.org,
 liqin.chen@sunplusct.com,
 Sam Ravnborg <sam@ravnborg.org>,
 Remis Lima Baima <remis.developer@googlemail.com>
Subject: [RFC 04/17] asm-generic: introduce asm/bitsperlong.h
Content-Disposition: inline; filename=0020-define-__BITS_PER_LONG-64-on-all-64-bit-architectu.patch
X-Provags-ID: V01U2FsdGVkX1/jLb8P+uRl1RcCf5V+QdVp81YPPHSqBrYSMpY
 iDoTcq0iAP5RFAPDuBCPUAkN/qpEfUbUueLJkB/uWbn2OBFAX9
 K5gcqKXdxwJlLU3abFV6g==

This provides a reliable way for asm-generic/types.h and other
files to find out if it is running on a 32 or 64 bit platform.

We cannot use CONFIG_64BIT for this in headers that are included
from user space because CONFIG symbols are not available there.
We also cannot do it inside of asm/types.h because some headers
need the word size but cannot include types.h.

The solution is to introduce a new header <asm/bitsperlong.h>
that defines both __BITS_PER_LONG for user space and
BITS_PER_LONG for usage in the kernel. The asm-generic
version falls back to 32 bit unless the architecture overrides
it, which I did for all 64 bit platforms.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Remis Lima Baima <remis.developer@googlemail.com>
---
 arch/alpha/include/asm/bitsperlong.h      |    8 	8 +	0 -	0 !
 arch/alpha/include/asm/types.h            |    3 	0 +	3 -	0 !
 arch/arm/include/asm/bitsperlong.h        |    1 	1 +	0 -	0 !
 arch/avr32/include/asm/bitsperlong.h      |    1 	1 +	0 -	0 !
 arch/blackfin/include/asm/bitsperlong.h   |    1 	1 +	0 -	0 !
 arch/cris/include/asm/bitsperlong.h       |    1 	1 +	0 -	0 !
 arch/frv/include/asm/bitsperlong.h        |    1 	1 +	0 -	0 !
 arch/h8300/include/asm/bitsperlong.h      |    1 	1 +	0 -	0 !
 arch/ia64/include/asm/bitsperlong.h       |    8 	8 +	0 -	0 !
 arch/ia64/include/asm/types.h             |    7 	0 +	7 -	0 !
 arch/m32r/include/asm/bitsperlong.h       |    1 	1 +	0 -	0 !
 arch/m68k/include/asm/bitsperlong.h       |    1 	1 +	0 -	0 !
 arch/m68knommu/include/asm/bitsperlong.h  |    1 	1 +	0 -	0 !
 arch/microblaze/include/asm/bitsperlong.h |    1 	1 +	0 -	0 !
 arch/mips/include/asm/bitsperlong.h       |    8 	8 +	0 -	0 !
 arch/mips/include/asm/types.h             |    3 	0 +	3 -	0 !
 arch/mn10300/include/asm/bitsperlong.h    |    1 	1 +	0 -	0 !
 arch/parisc/include/asm/bitsperlong.h     |   20 	20 +	0 -	0 !
 arch/parisc/include/asm/types.h           |    8 	0 +	8 -	0 !
 arch/powerpc/include/asm/bitsperlong.h    |   12 	12 +	0 -	0 !
 arch/powerpc/include/asm/types.h          |    9 	0 +	9 -	0 !
 arch/s390/include/asm/bitsperlong.h       |   13 	13 +	0 -	0 !
 arch/s390/include/asm/types.h             |    6 	0 +	6 -	0 !
 arch/sh/include/asm/bitsperlong.h         |    1 	1 +	0 -	0 !
 arch/sparc/include/asm/bitsperlong.h      |   13 	13 +	0 -	0 !
 arch/sparc/include/asm/types.h            |    4 	0 +	4 -	0 !
 arch/x86/include/asm/bitsperlong.h        |   13 	13 +	0 -	0 !
 arch/x86/include/asm/types.h              |    6 	0 +	6 -	0 !
 arch/xtensa/include/asm/bitsperlong.h     |    1 	1 +	0 -	0 !
 include/asm-generic/bitsperlong.h         |   32 	32 +	0 -	0 !
 include/asm-generic/int-l64.h             |    2 	2 +	0 -	0 !
 include/asm-generic/int-ll64.h            |    2 	2 +	0 -	0 !
 32 files changed, 144 insertions(+), 46 deletions(-)

Index: linux-2.6/arch/alpha/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/alpha/include/asm/types.h
+++ linux-2.6/arch/alpha/include/asm/types.h
@@ -25,9 +25,6 @@ typedef unsigned int umode_t;
  * These aren't exported outside the kernel to avoid name space clashes
  */
 #ifdef __KERNEL__
-
-#define BITS_PER_LONG 64
-
 #ifndef __ASSEMBLY__
 
 typedef u64 dma_addr_t;
Index: linux-2.6/arch/ia64/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/ia64/include/asm/types.h
+++ linux-2.6/arch/ia64/include/asm/types.h
@@ -19,10 +19,6 @@
 # define __IA64_UL(x)		(x)
 # define __IA64_UL_CONST(x)	x
 
-# ifdef __KERNEL__
-#  define BITS_PER_LONG 64
-# endif
-
 #else
 # define __IA64_UL(x)		((unsigned long)(x))
 # define __IA64_UL_CONST(x)	x##UL
@@ -34,10 +30,7 @@ typedef unsigned int umode_t;
  */
 # ifdef __KERNEL__
 
-#define BITS_PER_LONG 64
-
 /* DMA addresses are 64-bits wide, in general.  */
-
 typedef u64 dma_addr_t;
 
 # endif /* __KERNEL__ */
Index: linux-2.6/arch/mips/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/mips/include/asm/types.h
+++ linux-2.6/arch/mips/include/asm/types.h
@@ -31,9 +31,6 @@ typedef unsigned short umode_t;
  * These aren't exported outside the kernel to avoid name space clashes
  */
 #ifdef __KERNEL__
-
-#define BITS_PER_LONG _MIPS_SZLONG
-
 #ifndef __ASSEMBLY__
 
 #if (defined(CONFIG_HIGHMEM) && defined(CONFIG_64BIT_PHYS_ADDR)) \
Index: linux-2.6/arch/parisc/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/parisc/include/asm/types.h
+++ linux-2.6/arch/parisc/include/asm/types.h
@@ -14,14 +14,6 @@ typedef unsigned short umode_t;
  */
 #ifdef __KERNEL__
 
-#ifdef CONFIG_64BIT
-#define BITS_PER_LONG 64
-#define SHIFT_PER_LONG 6
-#else
-#define BITS_PER_LONG 32
-#define SHIFT_PER_LONG 5
-#endif
-
 #ifndef __ASSEMBLY__
 
 /* Dma addresses are 32-bits wide.  */
Index: linux-2.6/arch/powerpc/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/powerpc/include/asm/types.h
+++ linux-2.6/arch/powerpc/include/asm/types.h
@@ -40,15 +40,6 @@ typedef struct {
 #endif /* __ASSEMBLY__ */
 
 #ifdef __KERNEL__
-/*
- * These aren't exported outside the kernel to avoid name space clashes
- */
-#ifdef __powerpc64__
-#define BITS_PER_LONG 64
-#else
-#define BITS_PER_LONG 32
-#endif
-
 #ifndef __ASSEMBLY__
 
 typedef __vector128 vector128;
Index: linux-2.6/arch/s390/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/s390/include/asm/types.h
+++ linux-2.6/arch/s390/include/asm/types.h
@@ -28,12 +28,6 @@ typedef __signed__ long saddr_t;
  */
 #ifdef __KERNEL__
 
-#ifndef __s390x__
-#define BITS_PER_LONG 32
-#else
-#define BITS_PER_LONG 64
-#endif
-
 #ifndef __ASSEMBLY__
 
 typedef u64 dma64_addr_t;
Index: linux-2.6/arch/sparc/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/sparc/include/asm/types.h
+++ linux-2.6/arch/sparc/include/asm/types.h
@@ -21,8 +21,6 @@ typedef unsigned short umode_t;
 
 #ifdef __KERNEL__
 
-#define BITS_PER_LONG 64
-
 #ifndef __ASSEMBLY__
 
 /* Dma addresses come in generic and 64-bit flavours.  */
@@ -46,8 +44,6 @@ typedef unsigned short umode_t;
 
 #ifdef __KERNEL__
 
-#define BITS_PER_LONG 32
-
 #ifndef __ASSEMBLY__
 
 typedef u32 dma_addr_t;
Index: linux-2.6/arch/x86/include/asm/types.h
===================================================================
--- linux-2.6.orig/arch/x86/include/asm/types.h
+++ linux-2.6/arch/x86/include/asm/types.h
@@ -14,12 +14,6 @@ typedef unsigned short umode_t;
  */
 #ifdef __KERNEL__
 
-#ifdef CONFIG_X86_32
-# define BITS_PER_LONG 32
-#else
-# define BITS_PER_LONG 64
-#endif
-
 #ifndef __ASSEMBLY__
 
 typedef u64 dma64_addr_t;
Index: linux-2.6/arch/alpha/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/alpha/include/asm/bitsperlong.h
@@ -0,0 +1,8 @@
+#ifndef __ASM_ALPHA_BITSPERLONG_H
+#define __ASM_ALPHA_BITSPERLONG_H
+
+#define __BITS_PER_LONG 64
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_ALPHA_BITSPERLONG_H */
Index: linux-2.6/arch/ia64/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/ia64/include/asm/bitsperlong.h
@@ -0,0 +1,8 @@
+#ifndef __ASM_IA64_BITSPERLONG_H
+#define __ASM_IA64_BITSPERLONG_H
+
+#define __BITS_PER_LONG 64
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_IA64_BITSPERLONG_H */
Index: linux-2.6/arch/mips/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/mips/include/asm/bitsperlong.h
@@ -0,0 +1,8 @@
+#ifndef __ASM_MIPS_BITSPERLONG_H
+#define __ASM_MIPS_BITSPERLONG_H
+
+#define __BITS_PER_LONG _MIPS_SZLONG
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_MIPS_BITSPERLONG_H */
Index: linux-2.6/arch/parisc/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/parisc/include/asm/bitsperlong.h
@@ -0,0 +1,20 @@
+#ifndef __ASM_PARISC_BITSPERLONG_H
+#define __ASM_PARISC_BITSPERLONG_H
+
+/*
+ * using CONFIG_* outside of __KERNEL__ is wrong,
+ * __LP64__ was also removed from headers, so what
+ * is the right approach on parisc?
+ *	-arnd
+ */
+#if (defined(__KERNEL__) && defined(CONFIG_64BIT)) || defined (__LP64__)
+#define __BITS_PER_LONG 64
+#define SHIFT_PER_LONG 6
+#else
+#define __BITS_PER_LONG 32
+#define SHIFT_PER_LONG 5
+#endif
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_PARISC_BITSPERLONG_H */
Index: linux-2.6/arch/powerpc/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/powerpc/include/asm/bitsperlong.h
@@ -0,0 +1,12 @@
+#ifndef __ASM_POWERPC_BITSPERLONG_H
+#define __ASM_POWERPC_BITSPERLONG_H
+
+#if defined(__powerpc64__)
+# define __BITS_PER_LONG 64
+#else
+# define __BITS_PER_LONG 32
+#endif
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_POWERPC_BITSPERLONG_H */
Index: linux-2.6/arch/s390/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/s390/include/asm/bitsperlong.h
@@ -0,0 +1,13 @@
+#ifndef __ASM_S390_BITSPERLONG_H
+#define __ASM_S390_BITSPERLONG_H
+
+#ifndef __s390x__
+#define __BITS_PER_LONG 32
+#else
+#define __BITS_PER_LONG 64
+#endif
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_S390_BITSPERLONG_H */
+
Index: linux-2.6/arch/sparc/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/sparc/include/asm/bitsperlong.h
@@ -0,0 +1,13 @@
+#ifndef __ASM_ALPHA_BITSPERLONG_H
+#define __ASM_ALPHA_BITSPERLONG_H
+
+#if defined(__sparc__) && defined(__arch64__)
+#define __BITS_PER_LONG 64
+#else
+#define __BITS_PER_LONG 32
+#endif
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_ALPHA_BITSPERLONG_H */
+
Index: linux-2.6/arch/x86/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/x86/include/asm/bitsperlong.h
@@ -0,0 +1,13 @@
+#ifndef __ASM_X86_BITSPERLONG_H
+#define __ASM_X86_BITSPERLONG_H
+
+#ifdef __x86_64__
+# define __BITS_PER_LONG 64
+#else
+# define __BITS_PER_LONG 32
+#endif
+
+#include <asm-generic/bitsperlong.h>
+
+#endif /* __ASM_X86_BITSPERLONG_H */
+
Index: linux-2.6/include/asm-generic/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/include/asm-generic/bitsperlong.h
@@ -0,0 +1,32 @@
+#ifndef __ASM_GENERIC_BITS_PER_LONG
+#define __ASM_GENERIC_BITS_PER_LONG
+
+/*
+ * There seems to be no way of detecting this automatically from user
+ * space, so 64 bit architectures should override this in their
+ * bitsperlong.h. In particular, an architecture that supports
+ * both 32 and 64 bit user space must not rely on CONFIG_64BIT
+ * to decide it, but rather check a compiler provided macro.
+ */
+#ifndef __BITS_PER_LONG
+#define __BITS_PER_LONG 32
+#endif
+
+#ifdef __KERNEL__
+
+#ifdef CONFIG_64BIT
+#define BITS_PER_LONG 64
+#else
+#define BITS_PER_LONG 32
+#endif /* CONFIG_64BIT */
+
+/*
+ * FIXME: The check currently breaks x86-64 build, so it's
+ * temporarily disabled. Please fix x86-64 and reenable
+ */
+#if 0 && BITS_PER_LONG != __BITS_PER_LONG
+#error Inconsistent word size. Check asm/bitsperlong.h
+#endif
+
+#endif /* __KERNEL__ */
+#endif /* __ASM_GENERIC_BITS_PER_LONG */
Index: linux-2.6/include/asm-generic/int-l64.h
===================================================================
--- linux-2.6.orig/include/asm-generic/int-l64.h
+++ linux-2.6/include/asm-generic/int-l64.h
@@ -8,6 +8,8 @@
 #ifndef _ASM_GENERIC_INT_L64_H
 #define _ASM_GENERIC_INT_L64_H
 
+#include <asm/bitsperlong.h>
+
 #ifndef __ASSEMBLY__
 /*
  * __xx is ok: it doesn't pollute the POSIX namespace. Use these in the
Index: linux-2.6/include/asm-generic/int-ll64.h
===================================================================
--- linux-2.6.orig/include/asm-generic/int-ll64.h
+++ linux-2.6/include/asm-generic/int-ll64.h
@@ -8,6 +8,8 @@
 #ifndef _ASM_GENERIC_INT_LL64_H
 #define _ASM_GENERIC_INT_LL64_H
 
+#include <asm/bitsperlong.h>
+
 #ifndef __ASSEMBLY__
 /*
  * __xx is ok: it doesn't pollute the POSIX namespace. Use these in the
Index: linux-2.6/arch/arm/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/arm/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/avr32/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/avr32/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/blackfin/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/blackfin/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/cris/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/cris/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/frv/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/frv/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/h8300/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/h8300/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/m32r/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/m32r/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/m68k/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/m68k/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/m68knommu/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/m68knommu/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/microblaze/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/microblaze/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/mn10300/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/mn10300/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/sh/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/sh/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>
Index: linux-2.6/arch/xtensa/include/asm/bitsperlong.h
===================================================================
--- /dev/null
+++ linux-2.6/arch/xtensa/include/asm/bitsperlong.h
@@ -0,0 +1 @@
+#include <asm-generic/bitsperlong.h>

-- 

Re:

[Greg Ungerer]

Hi Arnd,

arnd@arndb.de wrote:
>>From arnd@arndb.de Mon Apr 27 16:28:40 2009
> References: <20090427142010.587518220@arndb.de>
> User-Agent: quilt/0.46-1
> Date: Mon, 27 Apr 2009 16:20:14 +0200
> From: arnd@arndb.de
> To: linux-kernel@vger.kernel.org
> Cc: john.williams@petalogix.com,
>  monstr@monstr.eu,
>  linux-api@vger.kernel.org,
>  linux-arch@vger.kernel.org,
>  liqin.chen@sunplusct.com,
>  Sam Ravnborg <sam@ravnborg.org>,
>  Remis Lima Baima <remis.developer@googlemail.com>
> Subject: [RFC 04/17] asm-generic: introduce asm/bitsperlong.h
> Content-Disposition: inline; filename=0020-define-__BITS_PER_LONG-64-on-all-64-bit-architectu.patch
> X-Provags-ID: V01U2FsdGVkX1/jLb8P+uRl1RcCf5V+QdVp81YPPHSqBrYSMpY
>  iDoTcq0iAP5RFAPDuBCPUAkN/qpEfUbUueLJkB/uWbn2OBFAX9
>  K5gcqKXdxwJlLU3abFV6g==
> 
> This provides a reliable way for asm-generic/types.h and other
> files to find out if it is running on a 32 or 64 bit platform.
> 
> We cannot use CONFIG_64BIT for this in headers that are included
> from user space because CONFIG symbols are not available there.
> We also cannot do it inside of asm/types.h because some headers
> need the word size but cannot include types.h.
> 
> The solution is to introduce a new header <asm/bitsperlong.h>
> that defines both __BITS_PER_LONG for user space and
> BITS_PER_LONG for usage in the kernel. The asm-generic
> version falls back to 32 bit unless the architecture overrides
> it, which I did for all 64 bit platforms.
> 
> Signed-off-by: Arnd Bergmann <arnd@arndb.de>
> Signed-off-by: Remis Lima Baima <remis.developer@googlemail.com>
> ---
>  arch/alpha/include/asm/bitsperlong.h      |    8 	8 +	0 -	0 !
>  arch/alpha/include/asm/types.h            |    3 	0 +	3 -	0 !
>  arch/arm/include/asm/bitsperlong.h        |    1 	1 +	0 -	0 !
>  arch/avr32/include/asm/bitsperlong.h      |    1 	1 +	0 -	0 !
>  arch/blackfin/include/asm/bitsperlong.h   |    1 	1 +	0 -	0 !
>  arch/cris/include/asm/bitsperlong.h       |    1 	1 +	0 -	0 !
>  arch/frv/include/asm/bitsperlong.h        |    1 	1 +	0 -	0 !
>  arch/h8300/include/asm/bitsperlong.h      |    1 	1 +	0 -	0 !
>  arch/ia64/include/asm/bitsperlong.h       |    8 	8 +	0 -	0 !
>  arch/ia64/include/asm/types.h             |    7 	0 +	7 -	0 !
>  arch/m32r/include/asm/bitsperlong.h       |    1 	1 +	0 -	0 !
>  arch/m68k/include/asm/bitsperlong.h       |    1 	1 +	0 -	0 !
>  arch/m68knommu/include/asm/bitsperlong.h  |    1 	1 +	0 -	0 !

All the m68knommu includes have merged into the m68k includes.
There is no arch/m68knommu/include directory any more.

Regards
Greg



>  arch/microblaze/include/asm/bitsperlong.h |    1 	1 +	0 -	0 !
>  arch/mips/include/asm/bitsperlong.h       |    8 	8 +	0 -	0 !
>  arch/mips/include/asm/types.h             |    3 	0 +	3 -	0 !
>  arch/mn10300/include/asm/bitsperlong.h    |    1 	1 +	0 -	0 !
>  arch/parisc/include/asm/bitsperlong.h     |   20 	20 +	0 -	0 !
>  arch/parisc/include/asm/types.h           |    8 	0 +	8 -	0 !
>  arch/powerpc/include/asm/bitsperlong.h    |   12 	12 +	0 -	0 !
>  arch/powerpc/include/asm/types.h          |    9 	0 +	9 -	0 !
>  arch/s390/include/asm/bitsperlong.h       |   13 	13 +	0 -	0 !
>  arch/s390/include/asm/types.h             |    6 	0 +	6 -	0 !
>  arch/sh/include/asm/bitsperlong.h         |    1 	1 +	0 -	0 !
>  arch/sparc/include/asm/bitsperlong.h      |   13 	13 +	0 -	0 !
>  arch/sparc/include/asm/types.h            |    4 	0 +	4 -	0 !
>  arch/x86/include/asm/bitsperlong.h        |   13 	13 +	0 -	0 !
>  arch/x86/include/asm/types.h              |    6 	0 +	6 -	0 !
>  arch/xtensa/include/asm/bitsperlong.h     |    1 	1 +	0 -	0 !
>  include/asm-generic/bitsperlong.h         |   32 	32 +	0 -	0 !
>  include/asm-generic/int-l64.h             |    2 	2 +	0 -	0 !
>  include/asm-generic/int-ll64.h            |    2 	2 +	0 -	0 !
>  32 files changed, 144 insertions(+), 46 deletions(-)


------------------------------------------------------------------------
Greg Ungerer  --  Principal Engineer        EMAIL:     gerg@snapgear.com
SnapGear Group, McAfee                      PHONE:       +61 7 3435 2888
825 Stanley St,                             FAX:         +61 7 3891 3630
Woolloongabba, QLD, 4102, Australia         WEB: http://www.SnapGear.com

RE:

[chau chin]

hello, i have a business for you

Re:

[Young Chang]

My name is Young Chang,i have a business Proposal in the tune of $19.7m for you
to handle with me from my bank Are you interested?






----------------------------------------------------------------

Re:

[Mr. Peter Wong]

Good-Day Friend,

I Mr. Peter Wong, I Need Your Assistance

(unknown),

[David Howells]

[-- Attachment #1: Type: text/plain, Size: 3555 bytes --]


Hi Linus,

Could you pull this branch please?  It contains some fixups for the UAPI stuff.

There are four main parts:

 (1) I found I needed some more fixups in the wake of testing Arm64 (some
     asm/unistd.h files had weird guards that caused problems - mostly in
     arches for which I don't have a compiler) and some __KERNEL__ splitting
     needed to take place in Arm64.

 (2) I found that c6x was missing some __KERNEL__ guards in its asm/signal.h.
     Mark Salter pointed me at a tree with a patch to remove that file
     entirely and use the asm-generic variant instead.  I pulled his tree
     since it also give me a defconfig for c6x to use in testing.

 (3) m68k turned out to have a header installation problem due to it lacking a
     kvm_para.h file.

     The conditional installation bits for linux/kvm_para.h, linux/kvm.h and
     linux/a.out.h weren't very well specified - and didn't work if an arch
     didn't have the asm/ version of that file, but there *was* an
     asm-generic/ version.

     It seems the "ifneq $((wildcard ...),)" for each of those three headers
     in include/kernel/Kbuild is invoked twice during header installation, and
     the second time it matches on the just installed asm-generic/kvm_para.h
     file and thus incorrectly installs linux/kvm_para.h as well.

     Most arches actually have an asm/kvm_para.h, so this wasn't detectable in
     those.

 (4) Fix problems with libfdt.h by reverting the changes to that particular
     header file.

Signed-off-by: David Howells <dhowells@redhat.com>
---
The following changes since commit 612a9aab56a93533e76e3ad91642db7033e03b69:

  Merge branch 'drm-next' of git://people.freedesktop.org/~airlied/linux (2012-10-03 23:29:23 -0700)

are available in the git repository at:


  git://git.infradead.org/users/dhowells/linux-headers.git tags/uapi-prep-20121007

for you to fetch changes up to cb6c88a13060f67ce556a06e279fadf46cc7f244:

  UAPI: Fix libfdt.h's #includes (2012-10-07 00:05:00 +0100)

----------------------------------------------------------------
(from the branch description for uapi-prep local branch)

clone of "master"
UAPI prep branch on 2012-10-07

----------------------------------------------------------------
David Howells (5):
      UAPI: Fix the guards on various asm/unistd.h files
      UAPI: Split compound conditionals containing __KERNEL__ in Arm64
      Merge remote-tracking branch 'c6x/for-linux-next' into uapi-prep
      UAPI: Fix conditional header installation handling (notably kvm_para.h on m68k)
      UAPI: Fix libfdt.h's #includes

Mark Salter (2):
      c6x: make dsk6455 the default config
      c6x: remove c6x signal.h

 arch/arm64/include/asm/hwcap.h      |  4 +++-
 arch/arm64/include/asm/stat.h       |  4 +++-
 arch/arm64/include/asm/unistd.h     |  8 +++-----
 arch/arm64/include/asm/unistd32.h   |  4 ----
 arch/c6x/Makefile                   |  2 ++
 arch/c6x/include/asm/Kbuild         |  1 +
 arch/c6x/include/asm/signal.h       | 17 -----------------
 arch/c6x/include/asm/unistd.h       |  4 ----
 arch/hexagon/include/asm/unistd.h   |  5 -----
 arch/openrisc/include/asm/unistd.h  |  5 -----
 arch/score/include/asm/unistd.h     |  5 -----
 arch/tile/include/asm/unistd.h      |  5 -----
 arch/unicore32/include/asm/unistd.h |  4 ----
 include/asm-generic/unistd.h        |  4 ----
 include/linux/Kbuild                |  9 +++------
 include/linux/libfdt.h              |  4 ++--
 16 files changed, 17 insertions(+), 68 deletions(-)
 delete mode 100644 arch/c6x/include/asm/signal.h

[-- Attachment #2: Type: application/pgp-signature, Size: 827 bytes --]

Re:

[Geert Uytterhoeven]

On Sun, Oct 7, 2012 at 1:15 AM, David Howells <dhowells@redhat.com> wrote:
>  (3) m68k turned out to have a header installation problem due to it lacking a
>      kvm_para.h file.

Sh also.

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds

Re:

[Will Deacon]

On Sun, Oct 07, 2012 at 07:36:20AM +0100, Geert Uytterhoeven wrote:
> On Sun, Oct 7, 2012 at 1:15 AM, David Howells <dhowells@redhat.com> wrote:
> >  (3) m68k turned out to have a header installation problem due to it lacking a
> >      kvm_para.h file.
> 
> Sh also.

and arm64 iirc. It should also affect arm, but we have a horrible dummy
header to get around it (just includes the asm-generic variant).

I posted a fix, but then it got derailed by the wildcarding used to generate
generic headers for kvm (which I was going some way to removing):

  https://lkml.org/lkml/2012/8/2/173

  http://marc.info/?l=linux-kernel&m=134393963216492&w=2

Will

Re:

[Quan Han]

Hello,

Compliments of the day to you and I believe all is well. My name is Mr. Quan Han and I work in bank of china. I have a transaction that I believe will be of mutual benefits to both of us. It involves an investment portfolio worth(eight million,three hundred and seventy thousand USD) which I like to acquire with your help and assistance. 
Yours sincerely,
Quan Han.

Re:

[christain147]

Good day,hoping you read this email and respond to me in good time.I do not intend to solicit for funds but  your time and energy in using my own resources to assist the less privileged.I am medically confined at the moment hence I request your indulgence.
I will give you a comprehensive brief once I hear from you.

Please forward your response to my private email address:
gudworks104@yahoo.com

Thanks and reply.

Robert Grondahl

RE:

[Qin's Yanjun]

How are you today and your family? I require your attention and honest
co-operation about some issues which i will really want to discuss with you
which.  Looking forward to read from you soon.  

Qin's


______________________________

Sky Silk, http://aknet.kz

Re:

[Amos Kalonzo]

Attn:

I am wondering why You haven't respond to my email for some days now.
reference to my client's contract balance payment of (11.7M,USD)
Kindly get back to me for more details.

Best Regards

Amos Kalonzo

Re:

[Amy Riddering]

-- 
Mrs. Amy Riddering contacting you for missionary work and i pray you
will be kind enough to deliver my $7 million donation to the less
privileged ones in your country and God will bless your generation for
doing this humanitarian work.

I am a widow suffering of lung cancer which has damaged my liver and
back bone, i decided to entrust this fund to a God fearing person that
will use it for Charity works since i do not have any child who will
inherit this money after i die. Please i want your sincere reply to
know if you will be able to execute this project, and I will give you
more information on how the fund will be transferred to your bank
account. I am waiting for your reply.

Thanks and God bless you.

[PATCH] lib/find_bit: Add find_prev_*_bit functions.

[Yun Levi]

Inspired find_next_*bit function series, add find_prev_*_bit series.
I'm not sure whether it'll be used right now But, I add these functions
for future usage.

Signed-off-by: Levi Yun <ppbuk5246@gmail.com>
---
 fs/ufs/util.h                     |  24 +++---
 include/asm-generic/bitops/find.h |  69 ++++++++++++++++
 include/asm-generic/bitops/le.h   |  33 ++++++++
 include/linux/bitops.h            |  34 +++++---
 lib/find_bit.c                    | 126 +++++++++++++++++++++++++++++-
 5 files changed, 260 insertions(+), 26 deletions(-)

diff --git a/fs/ufs/util.h b/fs/ufs/util.h
index 4931bec1a01c..7c87c77d10ca 100644
--- a/fs/ufs/util.h
+++ b/fs/ufs/util.h
@@ -2,7 +2,7 @@
 /*
  *  linux/fs/ufs/util.h
  *
- * Copyright (C) 1998
+ * Copyright (C) 1998
  * Daniel Pirkl <daniel.pirkl@email.cz>
  * Charles University, Faculty of Mathematics and Physics
  */
@@ -263,7 +263,7 @@ extern int ufs_prepare_chunk(struct page *page,
loff_t pos, unsigned len);
 /*
  * These functions manipulate ufs buffers
  */
-#define ubh_bread(sb,fragment,size) _ubh_bread_(uspi,sb,fragment,size)
+#define ubh_bread(sb,fragment,size) _ubh_bread_(uspi,sb,fragment,size)
 extern struct ufs_buffer_head * _ubh_bread_(struct
ufs_sb_private_info *, struct super_block *, u64 , u64);
 extern struct ufs_buffer_head * ubh_bread_uspi(struct
ufs_sb_private_info *, struct super_block *, u64, u64);
 extern void ubh_brelse (struct ufs_buffer_head *);
@@ -296,7 +296,7 @@ static inline void *get_usb_offset(struct
ufs_sb_private_info *uspi,
                                   unsigned int offset)
 {
        unsigned int index;
-
+
        index = offset >> uspi->s_fshift;
        offset &= ~uspi->s_fmask;
        return uspi->s_ubh.bh[index]->b_data + offset;
@@ -411,9 +411,9 @@ static inline unsigned _ubh_find_next_zero_bit_(
                offset = 0;
        }
        return (base << uspi->s_bpfshift) + pos - begin;
-}
+}

-static inline unsigned find_last_zero_bit (unsigned char * bitmap,
+static inline unsigned __ubh_find_last_zero_bit(unsigned char * bitmap,
        unsigned size, unsigned offset)
 {
        unsigned bit, i;
@@ -453,7 +453,7 @@ static inline unsigned _ubh_find_last_zero_bit_(
                            size + (uspi->s_bpf - start), uspi->s_bpf)
                        - (uspi->s_bpf - start);
                size -= count;
-               pos = find_last_zero_bit (ubh->bh[base]->b_data,
+               pos = __ubh_find_last_zero_bit(ubh->bh[base]->b_data,
                        start, start - count);
                if (pos > start - count || !size)
                        break;
@@ -461,7 +461,7 @@ static inline unsigned _ubh_find_last_zero_bit_(
                start = uspi->s_bpf;
        }
        return (base << uspi->s_bpfshift) + pos - begin;
-}
+}

 #define ubh_isblockclear(ubh,begin,block)
(!_ubh_isblockset_(uspi,ubh,begin,block))

@@ -483,7 +483,7 @@ static inline int _ubh_isblockset_(struct
ufs_sb_private_info * uspi,
                mask = 0x01 << (block & 0x07);
                return (*ubh_get_addr (ubh, begin + (block >> 3)) &
mask) == mask;
        }
-       return 0;
+       return 0;
 }

 #define ubh_clrblock(ubh,begin,block) _ubh_clrblock_(uspi,ubh,begin,block)
@@ -492,8 +492,8 @@ static inline void _ubh_clrblock_(struct
ufs_sb_private_info * uspi,
 {
        switch (uspi->s_fpb) {
        case 8:
-               *ubh_get_addr (ubh, begin + block) = 0x00;
-               return;
+               *ubh_get_addr (ubh, begin + block) = 0x00;
+               return;
        case 4:
                *ubh_get_addr (ubh, begin + (block >> 1)) &= ~(0x0f <<
((block & 0x01) << 2));
                return;
@@ -531,9 +531,9 @@ static inline void ufs_fragacct (struct
super_block * sb, unsigned blockmap,
 {
        struct ufs_sb_private_info * uspi;
        unsigned fragsize, pos;
-
+
        uspi = UFS_SB(sb)->s_uspi;
-
+
        fragsize = 0;
        for (pos = 0; pos < uspi->s_fpb; pos++) {
                if (blockmap & (1 << pos)) {
diff --git a/include/asm-generic/bitops/find.h
b/include/asm-generic/bitops/find.h
index 9fdf21302fdf..ca18b2ec954c 100644
--- a/include/asm-generic/bitops/find.h
+++ b/include/asm-generic/bitops/find.h
@@ -16,6 +16,20 @@ extern unsigned long find_next_bit(const unsigned
long *addr, unsigned long
                size, unsigned long offset);
 #endif

+#ifndef find_prev_bit
+/**
+ * find_prev_bit - find the prev set bit in a memory region
+ * @addr: The address to base the search on
+ * @offset: The bitnumber to start searching at
+ * @size: The bitmap size in bits
+ *
+ * Returns the bit number for the prev set bit
+ * If no bits are set, returns @size.
+ */
+extern unsigned long find_prev_bit(const unsigned long *addr, unsigned long
+               size, unsigned long offset);
+#endif
+
 #ifndef find_next_and_bit
 /**
  * find_next_and_bit - find the next set bit in both memory regions
@@ -32,6 +46,22 @@ extern unsigned long find_next_and_bit(const
unsigned long *addr1,
                unsigned long offset);
 #endif

+#ifndef find_prev_and_bit
+/**
+ * find_prev_and_bit - find the prev set bit in both memory regions
+ * @addr1: The first address to base the search on
+ * @addr2: The second address to base the search on
+ * @offset: The bitnumber to start searching at
+ * @size: The bitmap size in bits
+ *
+ * Returns the bit number for the prev set bit
+ * If no bits are set, returns @size.
+ */
+extern unsigned long find_prev_and_bit(const unsigned long *addr1,
+               const unsigned long *addr2, unsigned long size,
+               unsigned long offset);
+#endif
+
 #ifndef find_next_zero_bit
 /**
  * find_next_zero_bit - find the next cleared bit in a memory region
@@ -46,6 +76,20 @@ extern unsigned long find_next_zero_bit(const
unsigned long *addr, unsigned
                long size, unsigned long offset);
 #endif

+#ifndef find_prev_zero_bit
+/**
+ * find_prev_zero_bit - find the prev cleared bit in a memory region
+ * @addr: The address to base the search on
+ * @offset: The bitnumber to start searching at
+ * @size: The bitmap size in bits
+ *
+ * Returns the bit number of the prev zero bit
+ * If no bits are zero, returns @size.
+ */
+extern unsigned long find_prev_zero_bit(const unsigned long *addr, unsigned
+               long size, unsigned long offset);
+#endif
+
 #ifdef CONFIG_GENERIC_FIND_FIRST_BIT

 /**
@@ -80,6 +124,31 @@ extern unsigned long find_first_zero_bit(const
unsigned long *addr,

 #endif /* CONFIG_GENERIC_FIND_FIRST_BIT */

+#ifndef find_last_bit
+/**
+ * find_last_bit - find the last set bit in a memory region
+ * @addr: The address to start the search at
+ * @size: The number of bits to search
+ *
+ * Returns the bit number of the last set bit, or size.
+ */
+extern unsigned long find_last_bit(const unsigned long *addr,
+                                  unsigned long size);
+#endif
+
+#ifndef find_last_zero_bit
+/**
+ * find_last_zero_bit - find the last cleared bit in a memory region
+ * @addr: The address to start the search at
+ * @size: The maximum number of bits to search
+ *
+ * Returns the bit number of the first cleared bit.
+ * If no bits are zero, returns @size.
+ */
+extern unsigned long find_last_zero_bit(const unsigned long *addr,
+                                        unsigned long size);
+#endif
+
 /**
  * find_next_clump8 - find next 8-bit clump with set bits in a memory region
  * @clump: location to store copy of found clump
diff --git a/include/asm-generic/bitops/le.h b/include/asm-generic/bitops/le.h
index 188d3eba3ace..d0bd15bc34d9 100644
--- a/include/asm-generic/bitops/le.h
+++ b/include/asm-generic/bitops/le.h
@@ -27,6 +27,24 @@ static inline unsigned long
find_first_zero_bit_le(const void *addr,
        return find_first_zero_bit(addr, size);
 }

+static inline unsigned long find_prev_zero_bit_le(const void *addr,
+               unsigned long size, unsigned long offset)
+{
+       return find_prev_zero_bit(addr, size, offset);
+}
+
+static inline unsigned long find_prev_bit_le(const void *addr,
+               unsigned long size, unsigned long offset)
+{
+       return find_prev_bit(addr, size, offset);
+}
+
+static inline unsigned long find_last_zero_bit_le(const void *addr,
+               unsigned long size)
+{
+       return find_last_zero_bit(addr, size);
+}
+
 #elif defined(__BIG_ENDIAN)

 #define BITOP_LE_SWIZZLE       ((BITS_PER_LONG-1) & ~0x7)
@@ -41,11 +59,26 @@ extern unsigned long find_next_bit_le(const void *addr,
                unsigned long size, unsigned long offset);
 #endif

+#ifndef find_prev_zero_bit_le
+extern unsigned long find_prev_zero_bit_le(const void *addr,
+               unsigned long size, unsigned long offset);
+#endif
+
+#ifndef find_prev_bit_le
+extern unsigned long find_prev_bit_le(const void *addr,
+               unsigned long size, unsigned long offset);
+#endif
+
 #ifndef find_first_zero_bit_le
 #define find_first_zero_bit_le(addr, size) \
        find_next_zero_bit_le((addr), (size), 0)
 #endif

+#ifndef find_last_zero_bit_le
+#define find_last_zero_bit_le(addr, size) \
+       find_prev_zero_bit_le((addr), (size), (size - 1))
+#endif
+
 #else
 #error "Please fix <asm/byteorder.h>"
 #endif
diff --git a/include/linux/bitops.h b/include/linux/bitops.h
index 5b74bdf159d6..124c68929861 100644
--- a/include/linux/bitops.h
+++ b/include/linux/bitops.h
@@ -50,6 +50,28 @@ extern unsigned long __sw_hweight64(__u64 w);
             (bit) < (size);                                    \
             (bit) = find_next_zero_bit((addr), (size), (bit) + 1))

+#define for_each_set_bit_reverse(bit, addr, size) \
+       for ((bit) = find_last_bit((addr), (size));             \
+            (bit) < (size);                                    \
+            (bit) = find_prev_bit((addr), (size), (bit)))
+
+/* same as for_each_set_bit_reverse() but use bit as value to start with */
+#define for_each_set_bit_from_reverse(bit, addr, size) \
+       for ((bit) = find_prev_bit((addr), (size), (bit));      \
+            (bit) < (size);                                    \
+            (bit) = find_prev_bit((addr), (size), (bit - 1)))
+
+#define for_each_clear_bit_reverse(bit, addr, size) \
+       for ((bit) = find_last_zero_bit((addr), (size));        \
+            (bit) < (size);                                    \
+            (bit) = find_prev_zero_bit((addr), (size), (bit)))
+
+/* same as for_each_clear_bit_reverse() but use bit as value to start with */
+#define for_each_clear_bit_from_reverse(bit, addr, size) \
+       for ((bit) = find_prev_zero_bit((addr), (size), (bit)); \
+            (bit) < (size);                                    \
+            (bit) = find_next_zero_bit((addr), (size), (bit - 1)))
+
 /**
  * for_each_set_clump8 - iterate over bitmap for each 8-bit clump with set bits
  * @start: bit offset to start search and to store the current iteration offset
@@ -283,17 +305,5 @@ static __always_inline void __assign_bit(long nr,
volatile unsigned long *addr,
 })
 #endif

-#ifndef find_last_bit
-/**
- * find_last_bit - find the last set bit in a memory region
- * @addr: The address to start the search at
- * @size: The number of bits to search
- *
- * Returns the bit number of the last set bit, or size.
- */
-extern unsigned long find_last_bit(const unsigned long *addr,
-                                  unsigned long size);
-#endif
-
 #endif /* __KERNEL__ */
 #endif
diff --git a/lib/find_bit.c b/lib/find_bit.c
index 4a8751010d59..cbe06abd3d21 100644
--- a/lib/find_bit.c
+++ b/lib/find_bit.c
@@ -69,6 +69,58 @@ static unsigned long _find_next_bit(const unsigned
long *addr1,
 }
 #endif

+#if !defined(find_prev_bit) || !defined(find_prev_zero_bit) ||
         \
+       !defined(find_prev_bit_le) || !defined(find_prev_zero_bit_le)
||        \
+       !defined(find_prev_and_bit)
+/*
+ * This is a common helper function for find_prev_bit, find_prev_zero_bit, and
+ * find_prev_and_bit. The differences are:
+ *  - The "invert" argument, which is XORed with each fetched word before
+ *    searching it for one bits.
+ *  - The optional "addr2", which is anded with "addr1" if present.
+ */
+static unsigned long _find_prev_bit(const unsigned long *addr1,
+               const unsigned long *addr2, unsigned long nbits,
+               unsigned long start, unsigned long invert, unsigned long le)
+{
+       unsigned long tmp, mask;
+
+       if (unlikely(start >= nbits))
+               return nbits;
+
+       tmp = addr1[start / BITS_PER_LONG];
+       if (addr2)
+               tmp &= addr2[start / BITS_PER_LONG];
+       tmp ^= invert;
+
+       /* Handle 1st word. */
+       mask = BITMAP_LAST_WORD_MASK(start + 1);
+       if (le)
+               mask = swab(mask);
+
+       tmp &= mask;
+
+       start = round_down(start, BITS_PER_LONG);
+
+       while (!tmp) {
+               start -= BITS_PER_LONG;
+               if (start >= nbits)
+                       return nbits;
+
+               tmp = addr1[start / BITS_PER_LONG];
+               if (addr2)
+                       tmp &= addr2[start / BITS_PER_LONG];
+               tmp ^= invert;
+       }
+
+       if (le)
+               tmp = swab(tmp);
+
+       return start + __fls(tmp);
+}
+#endif
+
+
 #ifndef find_next_bit
 /*
  * Find the next set bit in a memory region.
@@ -81,6 +133,18 @@ unsigned long find_next_bit(const unsigned long
*addr, unsigned long size,
 EXPORT_SYMBOL(find_next_bit);
 #endif

+#ifndef find_prev_bit
+/*
+ * Find the prev set bit in a memory region.
+ */
+unsigned long find_prev_bit(const unsigned long *addr, unsigned long size,
+                           unsigned long offset)
+{
+       return _find_prev_bit(addr, NULL, size, offset, 0UL, 0);
+}
+EXPORT_SYMBOL(find_prev_bit);
+#endif
+
 #ifndef find_next_zero_bit
 unsigned long find_next_zero_bit(const unsigned long *addr, unsigned long size,
                                 unsigned long offset)
@@ -90,7 +154,16 @@ unsigned long find_next_zero_bit(const unsigned
long *addr, unsigned long size,
 EXPORT_SYMBOL(find_next_zero_bit);
 #endif

-#if !defined(find_next_and_bit)
+#ifndef find_prev_zero_bit
+unsigned long find_prev_zero_bit(const unsigned long *addr, unsigned long size,
+                                unsigned long offset)
+{
+       return _find_prev_bit(addr, NULL, size, offset, ~0UL, 0);
+}
+EXPORT_SYMBOL(find_prev_zero_bit);
+#endif
+
+#ifndef find_next_and_bit
 unsigned long find_next_and_bit(const unsigned long *addr1,
                const unsigned long *addr2, unsigned long size,
                unsigned long offset)
@@ -100,6 +173,16 @@ unsigned long find_next_and_bit(const unsigned long *addr1,
 EXPORT_SYMBOL(find_next_and_bit);
 #endif

+#ifndef find_prev_and_bit
+unsigned long find_prev_and_bit(const unsigned long *addr1,
+               const unsigned long *addr2, unsigned long size,
+               unsigned long offset)
+{
+       return _find_prev_bit(addr1, addr2, size, offset, 0UL, 0);
+}
+EXPORT_SYMBOL(find_prev_and_bit);
+#endif
+
 #ifndef find_first_bit
 /*
  * Find the first set bit in a memory region.
@@ -141,7 +224,7 @@ unsigned long find_last_bit(const unsigned long
*addr, unsigned long size)
 {
        if (size) {
                unsigned long val = BITMAP_LAST_WORD_MASK(size);
-               unsigned long idx = (size-1) / BITS_PER_LONG;
+               unsigned long idx = (size - 1) / BITS_PER_LONG;

                do {
                        val &= addr[idx];
@@ -156,6 +239,27 @@ unsigned long find_last_bit(const unsigned long
*addr, unsigned long size)
 EXPORT_SYMBOL(find_last_bit);
 #endif

+#ifndef find_last_zero_bit
+unsigned long find_last_zero_bit(const unsigned long *addr, unsigned long size)
+{
+       if (size) {
+               unsigned long val = BITMAP_LAST_WORD_MASK(size);
+               unsigned long idx = (size - 1) / BITS_PER_LONG;
+
+               do {
+                       val &= ~addr[idx];
+                       if (val)
+                               return idx * BITS_PER_LONG + __fls(val);
+
+                       val = ~0ul;
+               } while (idx--);
+       }
+
+       return size;
+}
+EXPORT_SYMBOL(find_last_zero_bit);
+#endif
+
 #ifdef __BIG_ENDIAN

 #ifndef find_next_zero_bit_le
@@ -167,6 +271,15 @@ unsigned long find_next_zero_bit_le(const void
*addr, unsigned
 EXPORT_SYMBOL(find_next_zero_bit_le);
 #endif

+#ifndef find_prev_zero_bit_le
+unsigned long find_prev_zero_bit_le(const void *addr, unsigned
+               long size, unsigned long offset)
+{
+       return _find_prev_bit(addr, NULL, size, offset, ~0UL, 1);
+}
+EXPORT_SYMBOL(find_prev_zero_bit_le);
+#endif
+
 #ifndef find_next_bit_le
 unsigned long find_next_bit_le(const void *addr, unsigned
                long size, unsigned long offset)
@@ -176,6 +289,15 @@ unsigned long find_next_bit_le(const void *addr, unsigned
 EXPORT_SYMBOL(find_next_bit_le);
 #endif

+#ifdef find_prev_bit_le
+unsigned long find_prev_bit_le(const void *addr, unsigned
+               long size, unsigned long offset)
+{
+       return _find_prev_bit(addr, NULL, size, offset, 0UL, 1);
+}
+EXPORT_SYMBOL(find_prev_bit_le);
+#endif
+
 #endif /* __BIG_ENDIAN */

 unsigned long find_next_clump8(unsigned long *clump, const unsigned long *addr,
--
2.29.2

Re: [PATCH] lib/find_bit: Add find_prev_*_bit functions.

[Andy Shevchenko]

On Wed, Dec 02, 2020 at 10:10:09AM +0900, Yun Levi wrote:
> Inspired find_next_*bit function series, add find_prev_*_bit series.
> I'm not sure whether it'll be used right now But, I add these functions
> for future usage.

This patch has few issues:
- it has more things than described (should be several patches instead)
- new functionality can be split logically to couple or more pieces as well
- it proposes functionality w/o user (dead code)

-- 
With Best Regards,
Andy Shevchenko

Re: [PATCH] lib/find_bit: Add find_prev_*_bit functions.

[Rasmus Villemoes]

On 02/12/2020 10.47, Andy Shevchenko wrote:
> On Wed, Dec 02, 2020 at 10:10:09AM +0900, Yun Levi wrote:
>> Inspired find_next_*bit function series, add find_prev_*_bit series.
>> I'm not sure whether it'll be used right now But, I add these functions
>> for future usage.
> 
> This patch has few issues:
> - it has more things than described (should be several patches instead)
> - new functionality can be split logically to couple or more pieces as well
> - it proposes functionality w/o user (dead code)

Yeah, the last point means it can't be applied - please submit it again
if and when you have an actual use case. And I'll add

- it lacks extension of the bitmap test module to cover the new
functions (that also wants to be a separate patch).

Rasmus

Re: [PATCH] lib/find_bit: Add find_prev_*_bit functions.

[Yun Levi]

Thanks for kind advice. But I'm so afraid to have questions below:

 > - it proposes functionality w/o user (dead code)
     Actually, I add these series functions to rewrite some of the
resource clean-up routine.
     A typical case is ethtool_set_per_queue_coalesce 's rollback label.
     Could this usage be an actual use case?

 >- it lacks extension of the bitmap test module to cover the new
 > functions (that also wants to be a separate patch).
     I see, then Could I add some of testcase on lib/test_bitops.c for testing?






On Wed, Dec 2, 2020 at 7:04 PM Rasmus Villemoes
<linux@rasmusvillemoes.dk> wrote:
>
> On 02/12/2020 10.47, Andy Shevchenko wrote:
> > On Wed, Dec 02, 2020 at 10:10:09AM +0900, Yun Levi wrote:
> >> Inspired find_next_*bit function series, add find_prev_*_bit series.
> >> I'm not sure whether it'll be used right now But, I add these functions
> >> for future usage.
> >
> > This patch has few issues:
> > - it has more things than described (should be several patches instead)
> > - new functionality can be split logically to couple or more pieces as well
> > - it proposes functionality w/o user (dead code)
>
> Yeah, the last point means it can't be applied - please submit it again
> if and when you have an actual use case. And I'll add
>
> - it lacks extension of the bitmap test module to cover the new
> functions (that also wants to be a separate patch).
>
> Rasmus

(no subject)

[Yun Levi]

On Thu, Dec 3, 2020 at 2:26 AM Yury Norov <yury.norov@gmail.com> wrote:

> Also look at lib/find_bit_benchmark.c
Thanks. I'll see.

> We need find_next_*_bit() because find_first_*_bit() can start searching only at word-aligned
> bits. In the case of find_last_*_bit(), we can start at any bit. So, if my understanding is correct,
> for the purpose of reverse traversing we can go with already existing find_last_bit(),

Thank you. I haven't thought that way.
But I think if we implement reverse traversing using find_last_bit(),
we have a problem.
Suppose the last bit 0, 1, 2, is set.
If we start
    find_last_bit(bitmap, 3) ==> return 2;
    find_last_bit(bitmap, 2) ==> return 1;
    find_last_bit(bitmap, 1) ==> return 0;
    find_last_bit(bitmap, 0) ===> return 0? // here we couldn't
distinguish size 0 input or 0 is set

and the for_each traverse routine prevent above case by returning size
(nbits) using find_next_bit.
So, for compatibility and the same expected return value like next traversing,
I think we need to find_prev_*_bit routine. if my understanding is correct.


>  I think this patch has some good catches. We definitely need to implement
> find_last_zero_bit(), as it is used by fs/ufs, and their local implementation is not optimal.
>
> We also should consider adding reverse traversing macros based on find_last_*_bit(),
> if there are proposed users.

Not only this, I think 'steal_from_bitmap_to_front' can be improved
using ffind_prev_zero_bit
like

diff --git a/fs/btrfs/free-space-cache.c b/fs/btrfs/free-space-cache.c
index af0013d3df63..9debb9707390 100644
--- a/fs/btrfs/free-space-cache.c
+++ b/fs/btrfs/free-space-cache.c
@@ -2372,7 +2372,6 @@ static bool steal_from_bitmap_to_front(struct
btrfs_free_space_ctl *ctl,
  u64 bitmap_offset;
  unsigned long i;
  unsigned long j;
- unsigned long prev_j;
  u64 bytes;

  bitmap_offset = offset_to_bitmap(ctl, info->offset);
@@ -2388,20 +2387,15 @@ static bool steal_from_bitmap_to_front(struct
btrfs_free_space_ctl *ctl,
  return false;

  i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
- j = 0;
- prev_j = (unsigned long)-1;
- for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
- if (j > i)
- break;
- prev_j = j;
- }
- if (prev_j == i)
+ j = find_prev_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
+
+ if (j == i)
  return false;

- if (prev_j == (unsigned long)-1)
+ if (j == BITS_PER_BITMAP)
  bytes = (i + 1) * ctl->unit;
  else
- bytes = (i - prev_j) * ctl->unit;
+ bytes = (i - j) * ctl->unit;

  info->offset -= bytes;
  info->bytes += bytes;

Thanks.

HTH
Levi.

Re:

[Yury Norov]

On Wed, Dec 2, 2020 at 10:22 AM Yun Levi <ppbuk5246@gmail.com> wrote:
>
> On Thu, Dec 3, 2020 at 2:26 AM Yury Norov <yury.norov@gmail.com> wrote:
>
> > Also look at lib/find_bit_benchmark.c
> Thanks. I'll see.
>
> > We need find_next_*_bit() because find_first_*_bit() can start searching only at word-aligned
> > bits. In the case of find_last_*_bit(), we can start at any bit. So, if my understanding is correct,
> > for the purpose of reverse traversing we can go with already existing find_last_bit(),
>
> Thank you. I haven't thought that way.
> But I think if we implement reverse traversing using find_last_bit(),
> we have a problem.
> Suppose the last bit 0, 1, 2, is set.
> If we start
>     find_last_bit(bitmap, 3) ==> return 2;
>     find_last_bit(bitmap, 2) ==> return 1;
>     find_last_bit(bitmap, 1) ==> return 0;
>     find_last_bit(bitmap, 0) ===> return 0? // here we couldn't
> distinguish size 0 input or 0 is set

If you traverse backward and reach bit #0, you're done. No need to continue.

>
> and the for_each traverse routine prevent above case by returning size
> (nbits) using find_next_bit.
> So, for compatibility and the same expected return value like next traversing,
> I think we need to find_prev_*_bit routine. if my understanding is correct.
>
>
> >  I think this patch has some good catches. We definitely need to implement
> > find_last_zero_bit(), as it is used by fs/ufs, and their local implementation is not optimal.
> >
> > We also should consider adding reverse traversing macros based on find_last_*_bit(),
> > if there are proposed users.
>
> Not only this, I think 'steal_from_bitmap_to_front' can be improved
> using ffind_prev_zero_bit
> like
>
> diff --git a/fs/btrfs/free-space-cache.c b/fs/btrfs/free-space-cache.c
> index af0013d3df63..9debb9707390 100644
> --- a/fs/btrfs/free-space-cache.c
> +++ b/fs/btrfs/free-space-cache.c
> @@ -2372,7 +2372,6 @@ static bool steal_from_bitmap_to_front(struct
> btrfs_free_space_ctl *ctl,
>   u64 bitmap_offset;
>   unsigned long i;
>   unsigned long j;
> - unsigned long prev_j;
>   u64 bytes;
>
>   bitmap_offset = offset_to_bitmap(ctl, info->offset);
> @@ -2388,20 +2387,15 @@ static bool steal_from_bitmap_to_front(struct
> btrfs_free_space_ctl *ctl,
>   return false;
>
>   i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
> - j = 0;
> - prev_j = (unsigned long)-1;
> - for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
> - if (j > i)
> - break;
> - prev_j = j;
> - }
> - if (prev_j == i)
> + j = find_prev_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);

This one may be implemented with find_last_zero_bit() as well:

unsigned log j = find_last_zero_bit(bitmap, BITS_PER_BITMAP);
if (j <= i || j >= BITS_PER_BITMAP)
        return false;

I believe the latter version is better because find_last_*_bit() is simpler in
implementation (and partially exists), has less parameters, and therefore
simpler for users, and doesn't introduce functionality duplication.

The only consideration I can imagine to advocate find_prev*() is the performance
advantage in the scenario when we know for sure that first N bits of
bitmap are all
set/clear, and we can bypass traversing that area. But again, in this
case we can pass the
bitmap address with the appropriate offset, and stay with find_last_*()

> +
> + if (j == i)
>   return false;
>
> - if (prev_j == (unsigned long)-1)
> + if (j == BITS_PER_BITMAP)
>   bytes = (i + 1) * ctl->unit;
>   else
> - bytes = (i - prev_j) * ctl->unit;
> + bytes = (i - j) * ctl->unit;
>
>   info->offset -= bytes;
>   info->bytes += bytes;
>
> Thanks.
>
> HTH
> Levi.

(no subject)

[Yun Levi]

On Thu, Dec 3, 2020 at 6:26 AM Yury Norov <yury.norov@gmail.com> wrote:
>
> On Wed, Dec 2, 2020 at 10:22 AM Yun Levi <ppbuk5246@gmail.com> wrote:
> >
> > On Thu, Dec 3, 2020 at 2:26 AM Yury Norov <yury.norov@gmail.com> wrote:
> >
> > > Also look at lib/find_bit_benchmark.c
> > Thanks. I'll see.
> >
> > > We need find_next_*_bit() because find_first_*_bit() can start searching only at word-aligned
> > > bits. In the case of find_last_*_bit(), we can start at any bit. So, if my understanding is correct,
> > > for the purpose of reverse traversing we can go with already existing find_last_bit(),
> >
> > Thank you. I haven't thought that way.
> > But I think if we implement reverse traversing using find_last_bit(),
> > we have a problem.
> > Suppose the last bit 0, 1, 2, is set.
> > If we start
> >     find_last_bit(bitmap, 3) ==> return 2;
> >     find_last_bit(bitmap, 2) ==> return 1;
> >     find_last_bit(bitmap, 1) ==> return 0;
> >     find_last_bit(bitmap, 0) ===> return 0? // here we couldn't
> > distinguish size 0 input or 0 is set
>
> If you traverse backward and reach bit #0, you're done. No need to continue.
I think the case when I consider the this macro like

#define for_each_clear_bit_reverse(bit, addr, size)
    for ((bit) = find_last_zero_bit((addr), (size))
          (bit) < (size);
          (bit) = find_prev_zero_bit((addr), (size), (bit)))

If we implement the above macro only with find_last_zero_bit,
I think there is no way without adding any additional variable to finish loop.
But I don't want to add additional variable to sustain same format
with for_each_clear_bit,
That's why i decide to implement find_prev_*_bit series.

I don't know it's correct thinking or biased. Am I wrong?

>
> >
> > and the for_each traverse routine prevent above case by returning size
> > (nbits) using find_next_bit.
> > So, for compatibility and the same expected return value like next traversing,
> > I think we need to find_prev_*_bit routine. if my understanding is correct.
> >
> >
> > >  I think this patch has some good catches. We definitely need to implement
> > > find_last_zero_bit(), as it is used by fs/ufs, and their local implementation is not optimal.
> > >
> > > We also should consider adding reverse traversing macros based on find_last_*_bit(),
> > > if there are proposed users.
> >
> > Not only this, I think 'steal_from_bitmap_to_front' can be improved
> > using ffind_prev_zero_bit
> > like
> >
> > diff --git a/fs/btrfs/free-space-cache.c b/fs/btrfs/free-space-cache.c
> > index af0013d3df63..9debb9707390 100644
> > --- a/fs/btrfs/free-space-cache.c
> > +++ b/fs/btrfs/free-space-cache.c
> > @@ -2372,7 +2372,6 @@ static bool steal_from_bitmap_to_front(struct
> > btrfs_free_space_ctl *ctl,
> >   u64 bitmap_offset;
> >   unsigned long i;
> >   unsigned long j;
> > - unsigned long prev_j;
> >   u64 bytes;
> >
> >   bitmap_offset = offset_to_bitmap(ctl, info->offset);
> > @@ -2388,20 +2387,15 @@ static bool steal_from_bitmap_to_front(struct
> > btrfs_free_space_ctl *ctl,
> >   return false;
> >
> >   i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
> > - j = 0;
> > - prev_j = (unsigned long)-1;
> > - for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
> > - if (j > i)
> > - break;
> > - prev_j = j;
> > - }
> > - if (prev_j == i)
> > + j = find_prev_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
>
> This one may be implemented with find_last_zero_bit() as well:
>
> unsigned log j = find_last_zero_bit(bitmap, BITS_PER_BITMAP);
> if (j <= i || j >= BITS_PER_BITMAP)
>         return false;
>
Actually, in that code, we don't need to check the bit after i.
Originally, if my understanding is correct, former code tries to find
the last 0 bit before i.
and if all bits are fully set before i, it use next one as i + 1

that's why i think the if condition should be
   if (j >= i)

But above condition couldn't the discern the case when all bits are
fully set before i.
Also, I think we don't need to check the bit after i and In this case,
find_prev_zero_bit which
specifies the start point is clear to show the meaning of the code.


> I believe the latter version is better because find_last_*_bit() is simpler in
> implementation (and partially exists), has less parameters, and therefore
> simpler for users, and doesn't introduce functionality duplication.
>
> The only consideration I can imagine to advocate find_prev*() is the performance
> advantage in the scenario when we know for sure that first N bits of
> bitmap are all
> set/clear, and we can bypass traversing that area. But again, in this
> case we can pass the
> bitmap address with the appropriate offset, and stay with find_last_*()
>
> > +
> > + if (j == i)
> >   return false;
> >
> > - if (prev_j == (unsigned long)-1)
> > + if (j == BITS_PER_BITMAP)
> >   bytes = (i + 1) * ctl->unit;
> >   else
> > - bytes = (i - prev_j) * ctl->unit;
> > + bytes = (i - j) * ctl->unit;
> >
> >   info->offset -= bytes;
> >   info->bytes += bytes;
> >
> > Thanks.
> >
> > HTH
> > Levi.

Thanks but

(no subject)

[Yun Levi]

On Thu, Dec 3, 2020 at 7:51 AM Yun Levi <ppbuk5246@gmail.com> wrote:
>
> On Thu, Dec 3, 2020 at 6:26 AM Yury Norov <yury.norov@gmail.com> wrote:
> >
> > On Wed, Dec 2, 2020 at 10:22 AM Yun Levi <ppbuk5246@gmail.com> wrote:
> > >
> > > On Thu, Dec 3, 2020 at 2:26 AM Yury Norov <yury.norov@gmail.com> wrote:
> > >
> > > > Also look at lib/find_bit_benchmark.c
> > > Thanks. I'll see.
> > >
> > > > We need find_next_*_bit() because find_first_*_bit() can start searching only at word-aligned
> > > > bits. In the case of find_last_*_bit(), we can start at any bit. So, if my understanding is correct,
> > > > for the purpose of reverse traversing we can go with already existing find_last_bit(),
> > >
> > > Thank you. I haven't thought that way.
> > > But I think if we implement reverse traversing using find_last_bit(),
> > > we have a problem.
> > > Suppose the last bit 0, 1, 2, is set.
> > > If we start
> > >     find_last_bit(bitmap, 3) ==> return 2;
> > >     find_last_bit(bitmap, 2) ==> return 1;
> > >     find_last_bit(bitmap, 1) ==> return 0;
> > >     find_last_bit(bitmap, 0) ===> return 0? // here we couldn't
> > > distinguish size 0 input or 0 is set
> >
> > If you traverse backward and reach bit #0, you're done. No need to continue.
> I think the case when I consider the this macro like
>
> #define for_each_clear_bit_reverse(bit, addr, size)
>     for ((bit) = find_last_zero_bit((addr), (size))
>           (bit) < (size);
>           (bit) = find_prev_zero_bit((addr), (size), (bit)))
>
> If we implement the above macro only with find_last_zero_bit,
> I think there is no way without adding any additional variable to finish loop.
> But I don't want to add additional variable to sustain same format
> with for_each_clear_bit,
> That's why i decide to implement find_prev_*_bit series.
>
> I don't know it's correct thinking or biased. Am I wrong?
>
> >
> > >
> > > and the for_each traverse routine prevent above case by returning size
> > > (nbits) using find_next_bit.
> > > So, for compatibility and the same expected return value like next traversing,
> > > I think we need to find_prev_*_bit routine. if my understanding is correct.
> > >
> > >
> > > >  I think this patch has some good catches. We definitely need to implement
> > > > find_last_zero_bit(), as it is used by fs/ufs, and their local implementation is not optimal.
> > > >
> > > > We also should consider adding reverse traversing macros based on find_last_*_bit(),
> > > > if there are proposed users.
> > >
> > > Not only this, I think 'steal_from_bitmap_to_front' can be improved
> > > using ffind_prev_zero_bit
> > > like
> > >
> > > diff --git a/fs/btrfs/free-space-cache.c b/fs/btrfs/free-space-cache.c
> > > index af0013d3df63..9debb9707390 100644
> > > --- a/fs/btrfs/free-space-cache.c
> > > +++ b/fs/btrfs/free-space-cache.c
> > > @@ -2372,7 +2372,6 @@ static bool steal_from_bitmap_to_front(struct
> > > btrfs_free_space_ctl *ctl,
> > >   u64 bitmap_offset;
> > >   unsigned long i;
> > >   unsigned long j;
> > > - unsigned long prev_j;
> > >   u64 bytes;
> > >
> > >   bitmap_offset = offset_to_bitmap(ctl, info->offset);
> > > @@ -2388,20 +2387,15 @@ static bool steal_from_bitmap_to_front(struct
> > > btrfs_free_space_ctl *ctl,
> > >   return false;
> > >
> > >   i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
> > > - j = 0;
> > > - prev_j = (unsigned long)-1;
> > > - for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
> > > - if (j > i)
> > > - break;
> > > - prev_j = j;
> > > - }
> > > - if (prev_j == i)
> > > + j = find_prev_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
> >
> > This one may be implemented with find_last_zero_bit() as well:
> >
> > unsigned log j = find_last_zero_bit(bitmap, BITS_PER_BITMAP);
> > if (j <= i || j >= BITS_PER_BITMAP)
> >         return false;
> >
> Actually, in that code, we don't need to check the bit after i.
> Originally, if my understanding is correct, former code tries to find
> the last 0 bit before i.
> and if all bits are fully set before i, it use next one as i + 1
>
> that's why i think the if condition should be
>    if (j >= i)
>
> But above condition couldn't the discern the case when all bits are
> fully set before i.
> Also, I think we don't need to check the bit after i and In this case,
> find_prev_zero_bit which
> specifies the start point is clear to show the meaning of the code.
>
>
> > I believe the latter version is better because find_last_*_bit() is simpler in
> > implementation (and partially exists), has less parameters, and therefore
> > simpler for users, and doesn't introduce functionality duplication.

I think it's not duplication.
Actually, former you teach me find_first_*_bit should be start word-aligned bit,
But as find_first_*_bit declares it as "size of bitmap" not a start offset.
Though the bitmap size it's word-aligned, it doesn't matter to fine
first bit in the specified size of bitmap (it no, it will return just
size of bitmap)

Likewise, find_last_*_bit is also similar in context.
Fundamentally, it's not a start offset of bitmap but I think it just
size of bitmap.

That's the reason why we need to find_next_*_bit to start at the
specified offset.
In this matter, I think it's better to have find_prev_*_bit.

So, I think we can use both of these functions to be used to achieve a goal.
But, each function has different concept actually that's why I don't
think it's not duplication.

if my understanding is wrong.. Forgive me. and let me know..

Thanks.



> >
> > The only consideration I can imagine to advocate find_prev*() is the performance
> > advantage in the scenario when we know for sure that first N bits of
> > bitmap are all
> > set/clear, and we can bypass traversing that area. But again, in this
> > case we can pass the
> > bitmap address with the appropriate offset, and stay with find_last_*()
> >
> > > +
> > > + if (j == i)
> > >   return false;
> > >
> > > - if (prev_j == (unsigned long)-1)
> > > + if (j == BITS_PER_BITMAP)
> > >   bytes = (i + 1) * ctl->unit;
> > >   else
> > > - bytes = (i - prev_j) * ctl->unit;
> > > + bytes = (i - j) * ctl->unit;
> > >
> > >   info->offset -= bytes;
> > >   info->bytes += bytes;
> > >
> > > Thanks.
> > >
> > > HTH
> > > Levi.
>
> Thanks but

Re:

[Rasmus Villemoes]

On 03/12/2020 02.23, Yun Levi wrote:
> On Thu, Dec 3, 2020 at 7:51 AM Yun Levi <ppbuk5246@gmail.com> wrote:
>>
>> On Thu, Dec 3, 2020 at 6:26 AM Yury Norov <yury.norov@gmail.com> wrote:
>>>
>>> On Wed, Dec 2, 2020 at 10:22 AM Yun Levi <ppbuk5246@gmail.com> wrote:
>>>>
>>>> On Thu, Dec 3, 2020 at 2:26 AM Yury Norov <yury.norov@gmail.com> wrote:
>>>>
>>>>> Also look at lib/find_bit_benchmark.c
>>>> Thanks. I'll see.
>>>>
>>>>> We need find_next_*_bit() because find_first_*_bit() can start searching only at word-aligned
>>>>> bits. In the case of find_last_*_bit(), we can start at any bit. So, if my understanding is correct,
>>>>> for the purpose of reverse traversing we can go with already existing find_last_bit(),
>>>>
>>>> Thank you. I haven't thought that way.
>>>> But I think if we implement reverse traversing using find_last_bit(),
>>>> we have a problem.
>>>> Suppose the last bit 0, 1, 2, is set.
>>>> If we start
>>>>     find_last_bit(bitmap, 3) ==> return 2;
>>>>     find_last_bit(bitmap, 2) ==> return 1;
>>>>     find_last_bit(bitmap, 1) ==> return 0;
>>>>     find_last_bit(bitmap, 0) ===> return 0? // here we couldn't

Either just make the return type of all find_prev/find_last be signed
int and use -1 as the sentinel to indicate "no such position exists", so
the loop condition would be foo >= 0. Or, change the condition from
"stop if we get the size returned" to "only continue if we get something
strictly less than the size we passed in (i.e., something which can
possibly be a valid bit index). In the latter case, both (unsigned)-1
aka UINT_MAX and the actual size value passed work equally well as a
sentinel.

If one uses UINT_MAX, a for_each_bit_reverse() macro would just be
something like

for (i = find_last_bit(bitmap, size); i < size; i =
find_last_bit(bitmap, i))

if one wants to use the size argument as the sentinel, the caller would
have to supply a scratch variable to keep track of the last i value:

for (j = size, i = find_last_bit(bitmap, j); i < j; j = i, i =
find_last_bit(bitmap, j))

which is probably a little less ergonomic.

Rasmus

Re:

[Yun Levi]

> If one uses UINT_MAX, a for_each_bit_reverse() macro would just be
> something like
>
> for (i = find_last_bit(bitmap, size); i < size; i =
> find_last_bit(bitmap, i))
>
> if one wants to use the size argument as the sentinel, the caller would
> have to supply a scratch variable to keep track of the last i value:
>
> for (j = size, i = find_last_bit(bitmap, j); i < j; j = i, i =
> find_last_bit(bitmap, j))
>
> which is probably a little less ergonomic.

Actually Because I want to avoid the modification of return type of
find_last_*_bit for new sentinel,
I add find_prev_*_bit.
the big difference between find_last_bit and find_prev_bit is
   find_last_bit doesn't check the size bit and use sentinel with size.
   but find_prev_bit check the offset bit and use sentinel with size
which passed by another argument.
   So if we use find_prev_bit, we could have a clear iteration if
using find_prev_bit like.

  #define for_each_set_bit_reverse(bit, addr, size) \
      for ((bit) = find_last_bit((addr), (size));    \
            (bit) < (size);                                     \
            (bit) = find_prev_bit((addr), (size), (bit - 1)))

  #define for_each_set_bit_from_reverse(bit, addr, size) \
      for ((bit) = find_prev_bit((addr), (size), (bit)); \
             (bit) < (size);                                           \
             (bit) = find_prev_bit((addr), (size), (bit - 1)))

Though find_prev_*_bit / find_last_*_bit have the same functionality.
But they also have a small difference.
I think this small this small difference doesn't make some of
confusion to user but it help to solve problem
with a simple way (just like the iteration above).

So I think I need, find_prev_*_bit series.

Am I missing anything?

Thanks.

Levi.

On Thu, Dec 3, 2020 at 5:33 PM Rasmus Villemoes
<linux@rasmusvillemoes.dk> wrote:
>
> On 03/12/2020 02.23, Yun Levi wrote:
> > On Thu, Dec 3, 2020 at 7:51 AM Yun Levi <ppbuk5246@gmail.com> wrote:
> >>
> >> On Thu, Dec 3, 2020 at 6:26 AM Yury Norov <yury.norov@gmail.com> wrote:
> >>>
> >>> On Wed, Dec 2, 2020 at 10:22 AM Yun Levi <ppbuk5246@gmail.com> wrote:
> >>>>
> >>>> On Thu, Dec 3, 2020 at 2:26 AM Yury Norov <yury.norov@gmail.com> wrote:
> >>>>
> >>>>> Also look at lib/find_bit_benchmark.c
> >>>> Thanks. I'll see.
> >>>>
> >>>>> We need find_next_*_bit() because find_first_*_bit() can start searching only at word-aligned
> >>>>> bits. In the case of find_last_*_bit(), we can start at any bit. So, if my understanding is correct,
> >>>>> for the purpose of reverse traversing we can go with already existing find_last_bit(),
> >>>>
> >>>> Thank you. I haven't thought that way.
> >>>> But I think if we implement reverse traversing using find_last_bit(),
> >>>> we have a problem.
> >>>> Suppose the last bit 0, 1, 2, is set.
> >>>> If we start
> >>>>     find_last_bit(bitmap, 3) ==> return 2;
> >>>>     find_last_bit(bitmap, 2) ==> return 1;
> >>>>     find_last_bit(bitmap, 1) ==> return 0;
> >>>>     find_last_bit(bitmap, 0) ===> return 0? // here we couldn't
>
> Either just make the return type of all find_prev/find_last be signed
> int and use -1 as the sentinel to indicate "no such position exists", so
> the loop condition would be foo >= 0. Or, change the condition from
> "stop if we get the size returned" to "only continue if we get something
> strictly less than the size we passed in (i.e., something which can
> possibly be a valid bit index). In the latter case, both (unsigned)-1
> aka UINT_MAX and the actual size value passed work equally well as a
> sentinel.
>
> If one uses UINT_MAX, a for_each_bit_reverse() macro would just be
> something like
>
> for (i = find_last_bit(bitmap, size); i < size; i =
> find_last_bit(bitmap, i))
>
> if one wants to use the size argument as the sentinel, the caller would
> have to supply a scratch variable to keep track of the last i value:
>
> for (j = size, i = find_last_bit(bitmap, j); i < j; j = i, i =
> find_last_bit(bitmap, j))
>
> which is probably a little less ergonomic.
>
> Rasmus

Re:

[Yury Norov]

Yun, could you please stop top-posting and excessive trimming in the thread?

On Thu, Dec 3, 2020 at 1:47 AM Yun Levi <ppbuk5246@gmail.com> wrote:
> > Either just make the return type of all find_prev/find_last be signed
> > int and use -1 as the sentinel to indicate "no such position exists", so
> > the loop condition would be foo >= 0. Or, change the condition from
> > "stop if we get the size returned" to "only continue if we get something
> > strictly less than the size we passed in (i.e., something which can
> > possibly be a valid bit index). In the latter case, both (unsigned)-1
> > aka UINT_MAX and the actual size value passed work equally well as a
> > sentinel.
> >
> > If one uses UINT_MAX, a for_each_bit_reverse() macro would just be
> > something like
> >
> > for (i = find_last_bit(bitmap, size); i < size; i =
> > find_last_bit(bitmap, i))
> >
> > if one wants to use the size argument as the sentinel, the caller would
> > have to supply a scratch variable to keep track of the last i value:
> >
> > for (j = size, i = find_last_bit(bitmap, j); i < j; j = i, i =
> > find_last_bit(bitmap, j))
> >
> > which is probably a little less ergonomic.
> >
> > Rasmus

I would prefer to avoid changing the find*bit() semantics. As for now,
if any of find_*_bit()
finds nothing, it returns the size of the bitmap it was passed.
Changing this for
a single function would break the consistency, and may cause problems
for those who
rely on existing behaviour.

Passing non-positive size to find_*_bit() should produce undefined
behaviour, because we cannot dereference a pointer to the bitmap in
this case; this is most probably a sign of a problem on a caller side
anyways.

Let's keep this logic unchanged?

> Actually Because I want to avoid the modification of return type of
> find_last_*_bit for new sentinel,
> I add find_prev_*_bit.
> the big difference between find_last_bit and find_prev_bit is
>    find_last_bit doesn't check the size bit and use sentinel with size.
>    but find_prev_bit check the offset bit and use sentinel with size
> which passed by another argument.
>    So if we use find_prev_bit, we could have a clear iteration if
> using find_prev_bit like.
>
>   #define for_each_set_bit_reverse(bit, addr, size) \
>       for ((bit) = find_last_bit((addr), (size));    \
>             (bit) < (size);                                     \
>             (bit) = find_prev_bit((addr), (size), (bit - 1)))
>
>   #define for_each_set_bit_from_reverse(bit, addr, size) \
>       for ((bit) = find_prev_bit((addr), (size), (bit)); \
>              (bit) < (size);                                           \
>              (bit) = find_prev_bit((addr), (size), (bit - 1)))
>
> Though find_prev_*_bit / find_last_*_bit have the same functionality.
> But they also have a small difference.
> I think this small this small difference doesn't make some of
> confusion to user but it help to solve problem
> with a simple way (just like the iteration above).
>
> So I think I need, find_prev_*_bit series.
>
> Am I missing anything?
>
> Thanks.
>
> Levi.

As you said, find_last_bit() and proposed find_prev_*_bit() have the
same functionality.
If you really want to have find_prev_*_bit(), could you please at
least write it using find_last_bit(), otherwise it would be just a
blottering.

Regarding reverse search, we can probably do like this (not tested,
just an idea):

#define for_each_set_bit_reverse(bit, addr, size) \
    for ((bit) = find_last_bit((addr), (size));    \
          (bit) < (size);                                     \
          (size) = (bit), (bit) = find_last_bit((addr), (bit)))

Thanks,
Yury

Re:

[Willy Tarreau]

On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> Yun, could you please stop top-posting and excessive trimming in the thread?

And re-configure the mail agent to make the "Subject" field appear and
fill it.

Willy

Re:

[Yun Levi]

>On Fri, Dec 4, 2020 at 3:53 AM Willy Tarreau <w@1wt.eu> wrote:
>
> On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > Yun, could you please stop top-posting and excessive trimming in the thread?
>
> And re-configure the mail agent to make the "Subject" field appear and
> fill it.

>On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> Yun, could you please stop top-posting and excessive trimming in the thread?
Sorry to make you uncomfortable... Thanks for advice.

>On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> As you said, find_last_bit() and proposed find_prev_*_bit() have the
> same functionality.
> If you really want to have find_prev_*_bit(), could you please at
> least write it using find_last_bit(), otherwise it would be just a
> blottering.

Actually find_prev_*_bit call _find_prev_bit which is a common helper function
like _find_next_bit.
As you know this function is required to support __BIGEDIAN's little
endian search.
find_prev_bit actually wrapper of _find_prev_bit which have a feature
the find_last_bit.

That makes the semantics difference between find_last_bit and find_prev_bit.
-- specify where you find from and
   In loop, find_last_bit couldn't sustain original size as sentinel
return value
    (we should change the size argument for next searching
     But it means whenever we call, "NOT SET or NOT CLEAR"'s sentinel
return value is changed per call).

Because we should have _find_prev_bit,
I think it's the matter to choose which is better to usein
find_prev_bit (find_last_bit? or _find_prev_bit?)
sustaining find_prev_bit feature (give size as sentinel return, from
where I start).
if my understanding is correct.

In my view, I prefer to use _find_prev_bit like find_next_bit for
integrated format.

But In some of the benchmarking, find_last_bit is better than _find_prev_bit,
here what I tested (look similar but sometimes have some difference).

              Start testing find_bit() with random-filled bitmap
[  +0.001850] find_next_bit:                  842792 ns, 163788 iterations
[  +0.000873] find_prev_bit:                  870914 ns, 163788 iterations
[  +0.000824] find_next_zero_bit:             821959 ns, 163894 iterations
[  +0.000677] find_prev_zero_bit:             676240 ns, 163894 iterations
[  +0.000777] find_last_bit:                  659103 ns, 163788 iterations
[  +0.001822] find_first_bit:                1708041 ns,  16250 iterations
[  +0.000539] find_next_and_bit:              492182 ns,  73871 iterations
[  +0.000001]
              Start testing find_bit() with sparse bitmap
[  +0.000222] find_next_bit:                   13227 ns,    654 iterations
[  +0.000013] find_prev_bit:                   11652 ns,    654 iterations
[  +0.001845] find_next_zero_bit:            1723869 ns, 327028 iterations
[  +0.001538] find_prev_zero_bit:            1355808 ns, 327028 iterations
[  +0.000010] find_last_bit:                    8114 ns,    654 iterations
[  +0.000867] find_first_bit:                 710639 ns,    654 iterations
[  +0.000006] find_next_and_bit:                4273 ns,      1 iterations
[  +0.000004] find_next_and_bit:                3278 ns,      1 iterations

              Start testing find_bit() with random-filled bitmap
[  +0.001784] find_next_bit:                  805553 ns, 164240 iterations
[  +0.000643] find_prev_bit:                  632474 ns, 164240 iterations
[  +0.000950] find_next_zero_bit:             877215 ns, 163442 iterations
[  +0.000664] find_prev_zero_bit:             662339 ns, 163442 iterations
[  +0.000680] find_last_bit:                  602204 ns, 164240 iterations
[  +0.001912] find_first_bit:                1758208 ns,  16408 iterations
[  +0.000760] find_next_and_bit:              531033 ns,  73798 iterations
[  +0.000002]
              Start testing find_bit() with sparse bitmap
[  +0.000203] find_next_bit:                   12468 ns,    656 iterations
[  +0.000205] find_prev_bit:                   10948 ns,    656 iterations
[  +0.001759] find_next_zero_bit:            1579447 ns, 327026 iterations
[  +0.001935] find_prev_zero_bit:            1931961 ns, 327026 iterations
[  +0.000013] find_last_bit:                    9543 ns,    656 iterations
[  +0.000732] find_first_bit:                 562009 ns,    656 iterations
[  +0.000217] find_next_and_bit:                6804 ns,      1 iterations
[  +0.000007] find_next_and_bit:                4367 ns,      1 iterations

Is it better to write find_prev_bit using find_last_bit?
I question again.

Thanks for your great advice, But please forgive my fault and lackness.

HTH.
Levi.

Re:

[Yury Norov]

On Thu, Dec 3, 2020 at 5:36 PM Yun Levi <ppbuk5246@gmail.com> wrote:
>
> >On Fri, Dec 4, 2020 at 3:53 AM Willy Tarreau <w@1wt.eu> wrote:
> >
> > On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > > Yun, could you please stop top-posting and excessive trimming in the thread?
> >
> > And re-configure the mail agent to make the "Subject" field appear and
> > fill it.
>
> >On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > Yun, could you please stop top-posting and excessive trimming in the thread?
> Sorry to make you uncomfortable... Thanks for advice.
>
> >On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > As you said, find_last_bit() and proposed find_prev_*_bit() have the
> > same functionality.
> > If you really want to have find_prev_*_bit(), could you please at
> > least write it using find_last_bit(), otherwise it would be just a
> > blottering.
>
> Actually find_prev_*_bit call _find_prev_bit which is a common helper function
> like _find_next_bit.
> As you know this function is required to support __BIGEDIAN's little
> endian search.
> find_prev_bit actually wrapper of _find_prev_bit which have a feature
> the find_last_bit.
>
> That makes the semantics difference between find_last_bit and find_prev_bit.
> -- specify where you find from and
>    In loop, find_last_bit couldn't sustain original size as sentinel
> return value
>     (we should change the size argument for next searching
>      But it means whenever we call, "NOT SET or NOT CLEAR"'s sentinel
> return value is changed per call).
>
> Because we should have _find_prev_bit,
> I think it's the matter to choose which is better to usein
> find_prev_bit (find_last_bit? or _find_prev_bit?)
> sustaining find_prev_bit feature (give size as sentinel return, from
> where I start).
> if my understanding is correct.
>
> In my view, I prefer to use _find_prev_bit like find_next_bit for
> integrated format.
>
> But In some of the benchmarking, find_last_bit is better than _find_prev_bit,
> here what I tested (look similar but sometimes have some difference).
>
>               Start testing find_bit() with random-filled bitmap
> [  +0.001850] find_next_bit:                  842792 ns, 163788 iterations
> [  +0.000873] find_prev_bit:                  870914 ns, 163788 iterations
> [  +0.000824] find_next_zero_bit:             821959 ns, 163894 iterations
> [  +0.000677] find_prev_zero_bit:             676240 ns, 163894 iterations
> [  +0.000777] find_last_bit:                  659103 ns, 163788 iterations
> [  +0.001822] find_first_bit:                1708041 ns,  16250 iterations
> [  +0.000539] find_next_and_bit:              492182 ns,  73871 iterations
> [  +0.000001]
>               Start testing find_bit() with sparse bitmap
> [  +0.000222] find_next_bit:                   13227 ns,    654 iterations
> [  +0.000013] find_prev_bit:                   11652 ns,    654 iterations
> [  +0.001845] find_next_zero_bit:            1723869 ns, 327028 iterations
> [  +0.001538] find_prev_zero_bit:            1355808 ns, 327028 iterations
> [  +0.000010] find_last_bit:                    8114 ns,    654 iterations
> [  +0.000867] find_first_bit:                 710639 ns,    654 iterations
> [  +0.000006] find_next_and_bit:                4273 ns,      1 iterations
> [  +0.000004] find_next_and_bit:                3278 ns,      1 iterations
>
>               Start testing find_bit() with random-filled bitmap
> [  +0.001784] find_next_bit:                  805553 ns, 164240 iterations
> [  +0.000643] find_prev_bit:                  632474 ns, 164240 iterations
> [  +0.000950] find_next_zero_bit:             877215 ns, 163442 iterations
> [  +0.000664] find_prev_zero_bit:             662339 ns, 163442 iterations
> [  +0.000680] find_last_bit:                  602204 ns, 164240 iterations
> [  +0.001912] find_first_bit:                1758208 ns,  16408 iterations
> [  +0.000760] find_next_and_bit:              531033 ns,  73798 iterations
> [  +0.000002]
>               Start testing find_bit() with sparse bitmap
> [  +0.000203] find_next_bit:                   12468 ns,    656 iterations
> [  +0.000205] find_prev_bit:                   10948 ns,    656 iterations
> [  +0.001759] find_next_zero_bit:            1579447 ns, 327026 iterations
> [  +0.001935] find_prev_zero_bit:            1931961 ns, 327026 iterations
> [  +0.000013] find_last_bit:                    9543 ns,    656 iterations
> [  +0.000732] find_first_bit:                 562009 ns,    656 iterations
> [  +0.000217] find_next_and_bit:                6804 ns,      1 iterations
> [  +0.000007] find_next_and_bit:                4367 ns,      1 iterations
>
> Is it better to write find_prev_bit using find_last_bit?
> I question again.

I answer again. It's better not to write find_prev_bit at all and
learn how to use existing functionality.

Yury

> Thanks for your great advice, But please forgive my fault and lackness.
>
> HTH.
> Levi.

Re:

[Yun Levi]

> I answer again. It's better not to write find_prev_bit at all and
> learn how to use existing functionality.

Thanks for the answer I'll fix and send the patch again :)

On Sat, Dec 5, 2020 at 3:14 AM Yury Norov <yury.norov@gmail.com> wrote:
>
> On Thu, Dec 3, 2020 at 5:36 PM Yun Levi <ppbuk5246@gmail.com> wrote:
> >
> > >On Fri, Dec 4, 2020 at 3:53 AM Willy Tarreau <w@1wt.eu> wrote:
> > >
> > > On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > > > Yun, could you please stop top-posting and excessive trimming in the thread?
> > >
> > > And re-configure the mail agent to make the "Subject" field appear and
> > > fill it.
> >
> > >On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > > Yun, could you please stop top-posting and excessive trimming in the thread?
> > Sorry to make you uncomfortable... Thanks for advice.
> >
> > >On Thu, Dec 03, 2020 at 10:46:25AM -0800, Yury Norov wrote:
> > > As you said, find_last_bit() and proposed find_prev_*_bit() have the
> > > same functionality.
> > > If you really want to have find_prev_*_bit(), could you please at
> > > least write it using find_last_bit(), otherwise it would be just a
> > > blottering.
> >
> > Actually find_prev_*_bit call _find_prev_bit which is a common helper function
> > like _find_next_bit.
> > As you know this function is required to support __BIGEDIAN's little
> > endian search.
> > find_prev_bit actually wrapper of _find_prev_bit which have a feature
> > the find_last_bit.
> >
> > That makes the semantics difference between find_last_bit and find_prev_bit.
> > -- specify where you find from and
> >    In loop, find_last_bit couldn't sustain original size as sentinel
> > return value
> >     (we should change the size argument for next searching
> >      But it means whenever we call, "NOT SET or NOT CLEAR"'s sentinel
> > return value is changed per call).
> >
> > Because we should have _find_prev_bit,
> > I think it's the matter to choose which is better to usein
> > find_prev_bit (find_last_bit? or _find_prev_bit?)
> > sustaining find_prev_bit feature (give size as sentinel return, from
> > where I start).
> > if my understanding is correct.
> >
> > In my view, I prefer to use _find_prev_bit like find_next_bit for
> > integrated format.
> >
> > But In some of the benchmarking, find_last_bit is better than _find_prev_bit,
> > here what I tested (look similar but sometimes have some difference).
> >
> >               Start testing find_bit() with random-filled bitmap
> > [  +0.001850] find_next_bit:                  842792 ns, 163788 iterations
> > [  +0.000873] find_prev_bit:                  870914 ns, 163788 iterations
> > [  +0.000824] find_next_zero_bit:             821959 ns, 163894 iterations
> > [  +0.000677] find_prev_zero_bit:             676240 ns, 163894 iterations
> > [  +0.000777] find_last_bit:                  659103 ns, 163788 iterations
> > [  +0.001822] find_first_bit:                1708041 ns,  16250 iterations
> > [  +0.000539] find_next_and_bit:              492182 ns,  73871 iterations
> > [  +0.000001]
> >               Start testing find_bit() with sparse bitmap
> > [  +0.000222] find_next_bit:                   13227 ns,    654 iterations
> > [  +0.000013] find_prev_bit:                   11652 ns,    654 iterations
> > [  +0.001845] find_next_zero_bit:            1723869 ns, 327028 iterations
> > [  +0.001538] find_prev_zero_bit:            1355808 ns, 327028 iterations
> > [  +0.000010] find_last_bit:                    8114 ns,    654 iterations
> > [  +0.000867] find_first_bit:                 710639 ns,    654 iterations
> > [  +0.000006] find_next_and_bit:                4273 ns,      1 iterations
> > [  +0.000004] find_next_and_bit:                3278 ns,      1 iterations
> >
> >               Start testing find_bit() with random-filled bitmap
> > [  +0.001784] find_next_bit:                  805553 ns, 164240 iterations
> > [  +0.000643] find_prev_bit:                  632474 ns, 164240 iterations
> > [  +0.000950] find_next_zero_bit:             877215 ns, 163442 iterations
> > [  +0.000664] find_prev_zero_bit:             662339 ns, 163442 iterations
> > [  +0.000680] find_last_bit:                  602204 ns, 164240 iterations
> > [  +0.001912] find_first_bit:                1758208 ns,  16408 iterations
> > [  +0.000760] find_next_and_bit:              531033 ns,  73798 iterations
> > [  +0.000002]
> >               Start testing find_bit() with sparse bitmap
> > [  +0.000203] find_next_bit:                   12468 ns,    656 iterations
> > [  +0.000205] find_prev_bit:                   10948 ns,    656 iterations
> > [  +0.001759] find_next_zero_bit:            1579447 ns, 327026 iterations
> > [  +0.001935] find_prev_zero_bit:            1931961 ns, 327026 iterations
> > [  +0.000013] find_last_bit:                    9543 ns,    656 iterations
> > [  +0.000732] find_first_bit:                 562009 ns,    656 iterations
> > [  +0.000217] find_next_and_bit:                6804 ns,      1 iterations
> > [  +0.000007] find_next_and_bit:                4367 ns,      1 iterations
> >
> > Is it better to write find_prev_bit using find_last_bit?
> > I question again.
>
> I answer again. It's better not to write find_prev_bit at all and
> learn how to use existing functionality.
>
> Yury
>
> > Thanks for your great advice, But please forgive my fault and lackness.
> >
> > HTH.
> > Levi.

Re:

[Rasmus Villemoes]

On 03/12/2020 19.46, Yury Norov wrote:

> I would prefer to avoid changing the find*bit() semantics. As for now,
> if any of find_*_bit()
> finds nothing, it returns the size of the bitmap it was passed.

Yeah, we should actually try to fix that, it causes bad code generation.
It's hard, because callers of course do that "if ret == size" check. But
it's really silly that something like find_first_bit needs to do that
"min(i*BPL + __ffs(word), size)" - the caller does a comparison anyway,
that comparison might as well be "ret >= size" rather than "ret ==
size", and then we could get rid of that branch (which min() necessarily
becomes) at the end of find_next_bit.

I haven't dug very deep into this, but I could also imagine the
arch-specific parts of this might become a little easier to do if the
semantics were just "if no such bit, return an indeterminate value >=
the size".

> Changing this for
> a single function would break the consistency, and may cause problems
> for those who
> rely on existing behaviour.

True. But I think it should be possible - I suppose most users are via
the iterator macros, which could all be updated at once. Changing ret ==
size to ret >= size will still work even if the implementations have not
been switched over, so it should be doable.

> 
> Passing non-positive size to find_*_bit() should produce undefined
> behaviour, because we cannot dereference a pointer to the bitmap in
> this case; this is most probably a sign of a problem on a caller side
> anyways.

No, the out-of-line bitmap functions should all handle the case of a
zero-size bitmap sensibly.

Is bitmap full? Yes (all the 0 bits are set).
Is bitmap empty? Yes, (none of the 0 bits are set).
Find the first bit set (returns 0, there's no such bit)

Etc. The static inlines for small_const_nbits do assume that the pointer
can be dereferenced, which is why small_const_nbits was updated to mean
1<=bits<=BITS_PER_LONG rather than just bits<=BITS_PER_LONG.

Rasmus

Re:

[Yury Norov]

On Sat, Dec 5, 2020 at 3:10 AM Rasmus Villemoes
<linux@rasmusvillemoes.dk> wrote:
>
> On 03/12/2020 19.46, Yury Norov wrote:
>
> > I would prefer to avoid changing the find*bit() semantics. As for now,
> > if any of find_*_bit()
> > finds nothing, it returns the size of the bitmap it was passed.
>
> Yeah, we should actually try to fix that, it causes bad code generation.
> It's hard, because callers of course do that "if ret == size" check. But
> it's really silly that something like find_first_bit needs to do that
> "min(i*BPL + __ffs(word), size)" - the caller does a comparison anyway,
> that comparison might as well be "ret >= size" rather than "ret ==
> size", and then we could get rid of that branch (which min() necessarily
> becomes) at the end of find_next_bit.

We didn't do that 5 years ago because it's too invasive and the improvement
is barely measurable, the difference is 2 instructions (on arm64).e.
Has something
changed since that?

20000000000000000 <find_first_bit_better>:
   0:   aa0003e3        mov     x3, x0
   4:   aa0103e0        mov     x0, x1
   8:   b4000181        cbz     x1, 38 <find_first_bit_better+0x38>
   c:   f9400064        ldr     x4, [x3]
  10:   d2800802        mov     x2, #0x40                       // #64
  14:   91002063        add     x3, x3, #0x8
  18:   b40000c4        cbz     x4, 30 <find_first_bit_better+0x30>
  1c:   14000008        b       3c <find_first_bit_better+0x3c>
  20:   f8408464        ldr     x4, [x3], #8
  24:   91010045        add     x5, x2, #0x40
  28:   b50000c4        cbnz    x4, 40 <find_first_bit_better+0x40>
  2c:   aa0503e2        mov     x2, x5
  30:   eb00005f        cmp     x2, x0
  34:   54ffff63        b.cc    20 <find_first_bit_better+0x20>  //
b.lo, b.ul, b.last
  38:   d65f03c0        ret
  3c:   d2800002        mov     x2, #0x0                        // #0
  40:   dac00084        rbit    x4, x4
  44:   dac01084        clz     x4, x4
  48:   8b020080        add     x0, x4, x2
  4c:   d65f03c0        ret

0000000000000050 <find_first_bit_worse>:
  50:   aa0003e4        mov     x4, x0
  54:   aa0103e0        mov     x0, x1
  58:   b4000181        cbz     x1, 88 <find_first_bit_worse+0x38>
  5c:   f9400083        ldr     x3, [x4]
  60:   d2800802        mov     x2, #0x40                       // #64
  64:   91002084        add     x4, x4, #0x8
  68:   b40000c3        cbz     x3, 80 <find_first_bit_worse+0x30>
  6c:   14000008        b       8c <find_first_bit_worse+0x3c>
  70:   f8408483        ldr     x3, [x4], #8
  74:   91010045        add     x5, x2, #0x40
  78:   b50000c3        cbnz    x3, 90 <find_first_bit_worse+0x40>
  7c:   aa0503e2        mov     x2, x5
  80:   eb02001f        cmp     x0, x2
  84:   54ffff68        b.hi    70 <find_first_bit_worse+0x20>  // b.pmore
  88:   d65f03c0        ret
  8c:   d2800002        mov     x2, #0x0                        // #0
  90:   dac00063        rbit    x3, x3
  94:   dac01063        clz     x3, x3
  98:   8b020062        add     x2, x3, x2
  9c:   eb02001f        cmp     x0, x2
  a0:   9a829000        csel    x0, x0, x2, ls  // ls = plast
  a4:   d65f03c0        ret

> I haven't dug very deep into this, but I could also imagine the
> arch-specific parts of this might become a little easier to do if the
> semantics were just "if no such bit, return an indeterminate value >=
> the size".
>
> > Changing this for
> > a single function would break the consistency, and may cause problems
> > for those who
> > rely on existing behaviour.
>
> True. But I think it should be possible - I suppose most users are via
> the iterator macros, which could all be updated at once. Changing ret ==
> size to ret >= size will still work even if the implementations have not
> been switched over, so it should be doable.

Since there's no assembler users for it, we can do just:
#define find_first_bit(bitmap, size)
min(better_find_first_bit((bitmap), (size)), (size))

... and deprecate find_first_bit.

> > Passing non-positive size to find_*_bit() should produce undefined
> > behaviour, because we cannot dereference a pointer to the bitmap in
> > this case; this is most probably a sign of a problem on a caller side
> > anyways.
>
> No, the out-of-line bitmap functions should all handle the case of a
> zero-size bitmap sensibly.

I could be more specific, the behaviour is defined: don't dereference
the address and return undefined value (which now is always 0).

> Is bitmap full? Yes (all the 0 bits are set).
> Is bitmap empty? Yes, (none of the 0 bits are set).
> Find the first bit set (returns 0, there's no such bit)

I can't answer because this object is not a map of bits - there's no room for
bits inside.

> Etc. The static inlines for small_const_nbits do assume that the pointer
> can be dereferenced, which is why small_const_nbits was updated to mean
> 1<=bits<=BITS_PER_LONG rather than just bits<=BITS_PER_LONG.

I don't want to do something like

if (size == 0)
        return -1;

... because it legitimizes this kind of usage and hides problems on
callers' side.
Instead, I'd add WARN_ON(size == 0), but I don't think it's so
critical to bother with it.

Yury

> Rasmus