open(O_DIRECT) on a tmpfs?

open(O_DIRECT) on a tmpfs?

[Michael Tokarev]

I wonder why open() with O_DIRECT (for example) bit set is
disallowed on a tmpfs (again, for example) filesystem,
returning EINVAL.

Yes, the question may seems strange a bit, because of two
somewhat conflicting reasons.  First, there's no reason to
use O_DIRECT with tmpfs in a first place, because tmpfs does
not have backing store at all, so there's no place to do
direct writes to.  But on another hand, again due to the very
nature of tmpfs, there's no reason not to allow O_DIRECT
open and just ignore it, -- exactly because there's no
backing store for this filesystem.

Why I'm asking is:  Currently I'm trying to evaluate a disk
subsystem for large loaded database (currently running with
Oracle, but there's no reason not to try Mysql or Postgres -
the stuff below equally applies to any database).

Almost any database uses two different I/O patterns for two
different kinds of files.  They are - regular data files, with
mostly random relatively large-block I/O, and control+redolog
files, which are small and receives very many relatively small
updates.

The same two kinds of load (large random I/O and small I/O)
applies to any journalling filesystem too, and even to linux
software raid devices.

I was thinking about trying to place those small redolog files
which receives alot of small updates to a battery-backed RAM.
The reason is simple: with fast I/O subsystem (composed of many
spindles, nicely distributed and so on), those redo-log files,
which can not be distributed, becomes real bottleneck.

But since such devices - battery-backed RAM - are relatively
expensive, I want to see how it works BEFORE buying a real
device.  So I just placed the redo-log files into a tmpfs,
because that's the most close "alternative", and tried to
start a database.  And it failed.

Failed because it rightly tries to open all the files with
O_DIRECT flag set, including control and redolog files.  And
tmpfs returns EINVAL.

Ok, I was able to work around this.. "issue" by creating a
loop device on a file residing on a tmpfs, creating a filesystem
on it and placing my files there.

But the original question remains...  Why tmpfs and similar
filesystems disallows O_DIRECT opens?

Thanks.

/mjt

Re: open(O_DIRECT) on a tmpfs?

[Hugh Dickins]

On Thu, 4 Jan 2007, Michael Tokarev wrote:
> I wonder why open() with O_DIRECT (for example) bit set is
> disallowed on a tmpfs (again, for example) filesystem,
> returning EINVAL.

Because it would be (a very small amount of) work and bloat to
support O_DIRECT on tmpfs; because that work didn't seem useful;
and because the nature of tmpfs (completely in page cache) is at
odds with the nature of O_DIRECT (completely avoiding page cache),
so it would seem misleading to support it.

You have a valid view, that we should not forbid what can easily be
allowed; and a valid (experimental) use for O_DIRECT on tmpfs; and
a valid alternative perception, that the nature of tmpfs is already
direct, so O_DIRECT should be allowed as a no-op upon it.

On the other hand, I'm glad that you've found a good workaround,
using loop, and suspect that it's appropriate that you should have
to use such a workaround: if the app cares so much that it insists
on O_DIRECT succeeding (for the ordering and persistence of its
metadata), would it be right for tmpfs to deceive it?

I'm inclined to stick with the status quo;
but could be persuaded by a chorus behind you.

Hugh

p.s.  You said "O_DIRECT (for example)" - what other open
flag do you think tmpfs should support which it does not?

Re: open(O_DIRECT) on a tmpfs?

[Bill Davidsen]

Hugh Dickins wrote:
> On Thu, 4 Jan 2007, Michael Tokarev wrote:
>> I wonder why open() with O_DIRECT (for example) bit set is
>> disallowed on a tmpfs (again, for example) filesystem,
>> returning EINVAL.
> 
> Because it would be (a very small amount of) work and bloat to
> support O_DIRECT on tmpfs; because that work didn't seem useful;
> and because the nature of tmpfs (completely in page cache) is at
> odds with the nature of O_DIRECT (completely avoiding page cache),
> so it would seem misleading to support it.
> 
> You have a valid view, that we should not forbid what can easily be
> allowed; and a valid (experimental) use for O_DIRECT on tmpfs; and
> a valid alternative perception, that the nature of tmpfs is already
> direct, so O_DIRECT should be allowed as a no-op upon it.

It does seem odd to require that every application using O_DIRECT would 
have to contain code to make it work with tmpfs, or that the admin would 
have to jump through a hoop and introduce (slight) overhead to bypass 
the problem, when the implementation is mostly to stop disallowing 
something which would currently work if allowed.

> 
> On the other hand, I'm glad that you've found a good workaround,
> using loop, and suspect that it's appropriate that you should have
> to use such a workaround: if the app cares so much that it insists
> on O_DIRECT succeeding (for the ordering and persistence of its
> metadata), would it be right for tmpfs to deceive it?

In many cases the use of O_DIRECT is purely to avoid impact on cache 
used by other applications. An application which writes a large quantity 
of data will have less impact on other applications by using O_DIRECT, 
assuming that the data will not be read from cache due to application 
pattern or the data being much larger than physical memory.
> 
> I'm inclined to stick with the status quo;
> but could be persuaded by a chorus behind you.

This isn't impacting me directly, but I can imagine some applications I 
have written, which currently use O_DIRECT, failing if someone chose the 
put a control file on tmpfs. I may be missing some benefit from 
restricting O_DIRECT, feel free to point it out.
> 
> Hugh
> 
> p.s.  You said "O_DIRECT (for example)" - what other open
> flag do you think tmpfs should support which it does not?


-- 
bill davidsen <davidsen@tmr.com>
   CTO TMR Associates, Inc
   Doing interesting things with small computers since 1979

Re: open(O_DIRECT) on a tmpfs?

[Hugh Dickins]

On Thu, 4 Jan 2007, Bill Davidsen wrote:
> 
> In many cases the use of O_DIRECT is purely to avoid impact on cache used by
> other applications. An application which writes a large quantity of data will
> have less impact on other applications by using O_DIRECT, assuming that the
> data will not be read from cache due to application pattern or the data being
> much larger than physical memory.

I see that as a good argument _not_ to allow O_DIRECT on tmpfs,
which inevitably impacts cache, even if O_DIRECT were requested.

But I'd also expect any app requesting O_DIRECT in that way, as a caring
citizen, to fall back to going without O_DIRECT when it's not supported.

Hugh

Re: open(O_DIRECT) on a tmpfs?

[Peter Staubach]

Hugh Dickins wrote:
> On Thu, 4 Jan 2007, Bill Davidsen wrote:
>   
>> In many cases the use of O_DIRECT is purely to avoid impact on cache used by
>> other applications. An application which writes a large quantity of data will
>> have less impact on other applications by using O_DIRECT, assuming that the
>> data will not be read from cache due to application pattern or the data being
>> much larger than physical memory.
>>     
>
> I see that as a good argument _not_ to allow O_DIRECT on tmpfs,
> which inevitably impacts cache, even if O_DIRECT were requested.
>
> But I'd also expect any app requesting O_DIRECT in that way, as a caring
> citizen, to fall back to going without O_DIRECT when it's not supported.

I suppose that one could also argue that the backing store for tmpfs
is the memory itself and thus, O_DIRECT could or should be supported.

    Thanx...

       ps

Re: open(O_DIRECT) on a tmpfs?

[Bill Davidsen]

Peter Staubach wrote:
> Hugh Dickins wrote:
>> On Thu, 4 Jan 2007, Bill Davidsen wrote:
>>  
>>> In many cases the use of O_DIRECT is purely to avoid impact on cache 
>>> used by
>>> other applications. An application which writes a large quantity of 
>>> data will
>>> have less impact on other applications by using O_DIRECT, assuming 
>>> that the
>>> data will not be read from cache due to application pattern or the 
>>> data being
>>> much larger than physical memory.
>>>     
>>
>> I see that as a good argument _not_ to allow O_DIRECT on tmpfs,
>> which inevitably impacts cache, even if O_DIRECT were requested.
>>
>> But I'd also expect any app requesting O_DIRECT in that way, as a caring
>> citizen, to fall back to going without O_DIRECT when it's not supported.
>
> I suppose that one could also argue that the backing store for tmpfs
> is the memory itself and thus, O_DIRECT could or should be supported. 

I suspect that many applications don't try to distinguish an open error 
beyond pass/fail. If the application actually tried to correct errors, 
like creating missing directories, it might, but if the error is going 
to be reported to the user and treated as fatal there's probably no 
logic to tell "can't do it" from "could if you asked the right way."

I always thought the difference between Linux and Windows was the "big 
brother" attitude. If someone wants to use O_DIRECT and tmpfs, and the 
system can allow it, why have code to block it because someone thinks 
they know better how the users should do things.

-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979

RE: open(O_DIRECT) on a tmpfs?

[Hua Zhong]

> I see that as a good argument _not_ to allow O_DIRECT on 
> tmpfs, which inevitably impacts cache, even if O_DIRECT were 
> requested.
> 
> But I'd also expect any app requesting O_DIRECT in that way, 
> as a caring citizen, to fall back to going without O_DIRECT 
> when it's not supported.

According to "man 2 open" on my system:

       O_DIRECT
              Try to minimize cache effects of the I/O to and from this file.
              In  general  this will degrade performance, but it is useful in
              special situations, such as  when  applications  do  their  own
              caching.  File I/O is done directly to/from user space buffers.
              The I/O is synchronous, i.e., at the completion of the  read(2)
              or write(2) system call, data is guaranteed to have been trans-
              ferred.  Under Linux 2.4 transfer sizes, and the  alignment  of
              user  buffer and file offset must all be multiples of the logi-
              cal block size of the file system. Under Linux 2.6 alignment to
              512-byte boundaries suffices.
              A semantically similar interface for block devices is described
              in raw(8).

This says nothing about (probably disk based) persistent backing store. I don't see why tmpfs has to conflict with it.

So I'd argue that it makes more sense to support O_DIRECT on tmpfs as the memory IS the backing store.

And EINVAL isn't even a very specific error.

Hua

RE: open(O_DIRECT) on a tmpfs?

[Hugh Dickins]

On Thu, 4 Jan 2007, Hua Zhong wrote:
> 
> So I'd argue that it makes more sense to support O_DIRECT
> on tmpfs as the memory IS the backing store.

A few more voices in favour and I'll be persuaded.  Perhaps I'm
out of date: when O_DIRECT came in, just a few filesystems supported
it, and it was perfectly normal for open O_DIRECT to be failed; but
I wouldn't want tmpfs to stand out now as a lone obstacle.

Christoph, what's your take on this?

Hugh

Re: open(O_DIRECT) on a tmpfs?

[Mark Lord]

Hugh Dickins wrote:
> On Thu, 4 Jan 2007, Hua Zhong wrote:
>> So I'd argue that it makes more sense to support O_DIRECT
>> on tmpfs as the memory IS the backing store.
> 
> A few more voices in favour and I'll be persuaded. 

I see no reason to restrict it as is currently done.

Policy belongs in userspace, not in the kernel,
so long as the code impact is miniscule.

Cheers

RE: open(O_DIRECT) on a tmpfs?

[Chen, Kenneth W]

Hugh Dickins wrote on Thursday, January 04, 2007 11:14 AM
> On Thu, 4 Jan 2007, Hua Zhong wrote:
> > So I'd argue that it makes more sense to support O_DIRECT
> > on tmpfs as the memory IS the backing store.
> 
> A few more voices in favour and I'll be persuaded.  Perhaps I'm
> out of date: when O_DIRECT came in, just a few filesystems supported
> it, and it was perfectly normal for open O_DIRECT to be failed; but
> I wouldn't want tmpfs to stand out now as a lone obstacle.

Maybe a bit hackish, all we need is to have an empty .direct_IO method
in shmem_aops to make __dentry_open() to pass the O_DIRECT check.  The
following patch adds 40 bytes to kernel text on x86-64.  An even more
hackish but zero cost route is to make .direct_IO variable non-zero via
a cast of -1 or some sort (that is probably ugly as hell).


diff -Nurp linus-2.6.git/mm/shmem.c linus-2.6.git.ken/mm/shmem.c
--- linus-2.6.git/mm/shmem.c	2006-12-27 19:06:11.000000000 -0800
+++ linus-2.6.git.ken/mm/shmem.c	2007-01-04 21:03:14.000000000 -0800
@@ -2314,10 +2314,18 @@ static void destroy_inodecache(void)
 	kmem_cache_destroy(shmem_inode_cachep);
 }
 
+ssize_t shmem_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
+			loff_t offset, unsigned long nr_segs)
+{
+	/* dummy direct_IO function.  Not to be executed */
+	BUG();
+}
+
 static const struct address_space_operations shmem_aops = {
 	.writepage	= shmem_writepage,
 	.set_page_dirty	= __set_page_dirty_nobuffers,
 #ifdef CONFIG_TMPFS
+	.direct_IO	= shmem_direct_IO,
 	.prepare_write	= shmem_prepare_write,
 	.commit_write	= simple_commit_write,
 #endif

Re: open(O_DIRECT) on a tmpfs?

[Helge Hafting]

Hugh Dickins wrote:
> On Thu, 4 Jan 2007, Hua Zhong wrote:
>   
>> So I'd argue that it makes more sense to support O_DIRECT
>> on tmpfs as the memory IS the backing store.
>>     
>
> A few more voices in favour and I'll be persuaded.  Perhaps I'm
> out of date: when O_DIRECT came in, just a few filesystems supported
> it, and it was perfectly normal for open O_DIRECT to be failed; but
> I wouldn't want tmpfs to stand out now as a lone obstacle.
>   
Having tmpfs suppoting O_DIRECT makes sense.
For me, O_DIRECT says "write directly to the device
and don't return till its done."  Which is what tmpfs
always do anyway.

The support could probably be as simple as ignoring
the flag entirely, mask it away in open() or something like that.


Arguments about "O_DIRECT says don't cache it and tmpfs
_is_ the cache" don't work.  O_DIRECT says "write straight
to the device" and the device just happens to be pagecache
memory.  The tmpfs file sure isn't cached elsewhere in
addition to its tmpfs pages.

Helge Hafting

Re: open(O_DIRECT) on a tmpfs?

[Jesper Juhl]

On 04/01/07, Hua Zhong <hzhong@gmail.com> wrote:
> > I see that as a good argument _not_ to allow O_DIRECT on
> > tmpfs, which inevitably impacts cache, even if O_DIRECT were
> > requested.
> >
> > But I'd also expect any app requesting O_DIRECT in that way,
> > as a caring citizen, to fall back to going without O_DIRECT
> > when it's not supported.
>
> According to "man 2 open" on my system:
>
>        O_DIRECT
>               Try to minimize cache effects of the I/O to and from this file.
>               In  general  this will degrade performance, but it is useful in
>               special situations, such as  when  applications  do  their  own
>               caching.  File I/O is done directly to/from user space buffers.
>               The I/O is synchronous, i.e., at the completion of the  read(2)
>               or write(2) system call, data is guaranteed to have been trans-
>               ferred.  Under Linux 2.4 transfer sizes, and the  alignment  of
>               user  buffer and file offset must all be multiples of the logi-
>               cal block size of the file system. Under Linux 2.6 alignment to
>               512-byte boundaries suffices.
>               A semantically similar interface for block devices is described
>               in raw(8).
>
> This says nothing about (probably disk based) persistent backing store. I don't see why tmpfs has to conflict with it.
>
> So I'd argue that it makes more sense to support O_DIRECT on tmpfs as the memory IS the backing store.
>

I'd agree.  O_DIRECT means data will go direct to backing store, so if
RAM *is* the backing store as in the tmpfs case, then I see why
O_DIRECT should fail for it...

I often use tmpfs when I want to test new setups - it's easy to get
rid of again and it's fast during testing. Why shouldn't I be able to
test apps that use O_DIRECT this way?

-- 
Jesper Juhl <jesper.juhl@gmail.com>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html

Re: open(O_DIRECT) on a tmpfs?

[Jesper Juhl]

On 05/01/07, Jesper Juhl <jesper.juhl@gmail.com> wrote:
> On 04/01/07, Hua Zhong <hzhong@gmail.com> wrote:
> > > I see that as a good argument _not_ to allow O_DIRECT on
> > > tmpfs, which inevitably impacts cache, even if O_DIRECT were
> > > requested.
> > >
> > > But I'd also expect any app requesting O_DIRECT in that way,
> > > as a caring citizen, to fall back to going without O_DIRECT
> > > when it's not supported.
> >
> > According to "man 2 open" on my system:
> >
> >        O_DIRECT
> >               Try to minimize cache effects of the I/O to and from this file.
> >               In  general  this will degrade performance, but it is useful in
> >               special situations, such as  when  applications  do  their  own
> >               caching.  File I/O is done directly to/from user space buffers.
> >               The I/O is synchronous, i.e., at the completion of the  read(2)
> >               or write(2) system call, data is guaranteed to have been trans-
> >               ferred.  Under Linux 2.4 transfer sizes, and the  alignment  of
> >               user  buffer and file offset must all be multiples of the logi-
> >               cal block size of the file system. Under Linux 2.6 alignment to
> >               512-byte boundaries suffices.
> >               A semantically similar interface for block devices is described
> >               in raw(8).
> >
> > This says nothing about (probably disk based) persistent backing store. I don't see why tmpfs has to conflict with it.
> >
> > So I'd argue that it makes more sense to support O_DIRECT on tmpfs as the memory IS the backing store.
> >
>
> I'd agree.  O_DIRECT means data will go direct to backing store, so if
> RAM *is* the backing store as in the tmpfs case, then I see why
> O_DIRECT should fail for it...
>
Whoops, that should of course have read " then I *DON'T* see why
O_DIRECT should fail" .

-- 
Jesper Juhl <jesper.juhl@gmail.com>
Don't top-post  http://www.catb.org/~esr/jargon/html/T/top-post.html
Plain text mails only, please      http://www.expita.com/nomime.html

Re: open(O_DIRECT) on a tmpfs?

[Denis Vlasenko]

On Thursday 04 January 2007 17:19, Bill Davidsen wrote:
> Hugh Dickins wrote:
> In many cases the use of O_DIRECT is purely to avoid impact on cache 
> used by other applications. An application which writes a large quantity 
> of data will have less impact on other applications by using O_DIRECT, 
> assuming that the data will not be read from cache due to application 
> pattern or the data being much larger than physical memory.

But O_DIRECT is _not_ about cache. At least I think it was not about
cache initially, it was more about DMAing data directly from/to
application address space to/from disks, saving memcpy's and double
allocations. Why do you think it has that special alignment requirements?
Are they cache related? Not at all!

After that people started adding unrelated semantics on it -
"oh, we use O_DIRECT in our database code and it pushes EVERYTHING
else out of cache. This is bad. Let's overload O_DIRECT to also mean
'do not pollute the cache'. Here's the patch".

DB people from certain well-known commercial DB have zero coding
taste. No wonder their binaries are nearly 100 MB (!!!) in size...

In all fairness, O_DIRECT's direct-DMA makes is easier to implement
"do-not-cache-me" than to do it for generic read()/write()
(just because O_DIRECT is (was?) using different code path,
not integrated into VM cache machinery that much).

But _conceptually_ "direct DMAing" and "do-not-cache-me"
are orthogonal, right?

That's why we also have bona fide fadvise and madvise
with FADV_DONTNEED/MADV_DONTNEED:

http://www.die.net/doc/linux/man/man2/fadvise.2.html
http://www.die.net/doc/linux/man/man2/madvise.2.html

_This_ is the proper way to say "do not cache me".

I think tmpfs should just ignore O_DIRECT bit.
That won't require much coding.
--
vda

Re: open(O_DIRECT) on a tmpfs?

[Nick Piggin]

Denis Vlasenko wrote:
> On Thursday 04 January 2007 17:19, Bill Davidsen wrote:
> 
>>Hugh Dickins wrote:
>>In many cases the use of O_DIRECT is purely to avoid impact on cache 
>>used by other applications. An application which writes a large quantity 
>>of data will have less impact on other applications by using O_DIRECT, 
>>assuming that the data will not be read from cache due to application 
>>pattern or the data being much larger than physical memory.
> 
> 
> But O_DIRECT is _not_ about cache. At least I think it was not about
> cache initially, it was more about DMAing data directly from/to
> application address space to/from disks, saving memcpy's and double
> allocations. Why do you think it has that special alignment requirements?
> Are they cache related? Not at all!

I don't know whether that is the case. The two issues are related -- the
IO is be done zero-copy because there is no cache involved, and due to
there being no cache, there are alignment restrictions.

I think IRIX might have implemented O_DIRECT first, and although the
semantics are a bit vague, I think it has always been to do zero copy
IO _and_ to bypass cache (ie. no splice-like tricks).

> After that people started adding unrelated semantics on it -
> "oh, we use O_DIRECT in our database code and it pushes EVERYTHING
> else out of cache. This is bad. Let's overload O_DIRECT to also mean
> 'do not pollute the cache'. Here's the patch".

It is because they already do their own caching, so going through
another, dumber, cache of same or less size (the pagecache) is useless.
fadvise does not change that.

That said, tmpfs's page are not really a cache (except when they are
swapcache, but let's not complicate things). So O_DIRECT on tmpfs
may not exactly be wrong.

-- 
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com 

Re: open(O_DIRECT) on a tmpfs?

[Bill Davidsen]

Denis Vlasenko wrote:
> On Thursday 04 January 2007 17:19, Bill Davidsen wrote:
>   
>> Hugh Dickins wrote:
>> In many cases the use of O_DIRECT is purely to avoid impact on cache 
>> used by other applications. An application which writes a large quantity 
>> of data will have less impact on other applications by using O_DIRECT, 
>> assuming that the data will not be read from cache due to application 
>> pattern or the data being much larger than physical memory.
>>     
>
> But O_DIRECT is _not_ about cache. At least I think it was not about
> cache initially, it was more about DMAing data directly from/to
> application address space to/from disks, saving memcpy's and double
> allocations. Why do you think it has that special alignment requirements?
> Are they cache related? Not at all!
>   
I'm not sure I can see how you find "don't use cache" not cache related. 
Saving the resources needed for cache would seem to obviously leave them 
for other processes.
> After that people started adding unrelated semantics on it -
> "oh, we use O_DIRECT in our database code and it pushes EVERYTHING
> else out of cache. This is bad. Let's overload O_DIRECT to also mean
> 'do not pollute the cache'. Here's the patch".
>   
Did O_DIRECT ever use cache in some way? Doing DMA directly out of user 
space would seem to avoid using cache unless code was actually added to 
write to cache as well as disk, since the data isn't needed in any buffer.
> DB people from certain well-known commercial DB have zero coding
> taste. No wonder their binaries are nearly 100 MB (!!!) in size...
>
> In all fairness, O_DIRECT's direct-DMA makes is easier to implement
> "do-not-cache-me" than to do it for generic read()/write()
> (just because O_DIRECT is (was?) using different code path,
> not integrated into VM cache machinery that much).
>
> But _conceptually_ "direct DMAing" and "do-not-cache-me"
> are orthogonal, right?
>   
In the sense that you must do DMA or use cache, yes.
> That's why we also have bona fide fadvise and madvise
> with FADV_DONTNEED/MADV_DONTNEED:
>
> http://www.die.net/doc/linux/man/man2/fadvise.2.html
> http://www.die.net/doc/linux/man/man2/madvise.2.html
>
> _This_ is the proper way to say "do not cache me".
>   
But none of those advisories says how to cache or not, only what the 
expected behavior will be. So FADV_NOREUSE does not control cache use, 
it simply allows the system to make assumptions. If I still had the load 
which generated my cache problems I would try both methods while doing a 
large data copy, and see if the end result was similar. In theory 
NOREUSE "could be" more efficient of disk, but also use a lot of cache 
depending  on the implementation.

One of the problems with RAID-5 and large data is that you can read it a 
lot faster than you can write it (in most cases), resulting in filling 
the cache with data from one process. Perhaps a scheduler tunable for 
allowed queued disk data would help with this, but copying a TB data set 
has a very bad effect on other i/o.
> I think tmpfs should just ignore O_DIRECT bit.
> That won't require much coding.

Since tmpfs is useful for testing programs, this would have an actual 
user benefit.

-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979

Re: open(O_DIRECT) on a tmpfs?

[Denis Vlasenko]

On Friday 05 January 2007 17:20, Bill Davidsen wrote:
> Denis Vlasenko wrote:
> > But O_DIRECT is _not_ about cache. At least I think it was not about
> > cache initially, it was more about DMAing data directly from/to
> > application address space to/from disks, saving memcpy's and double
> > allocations. Why do you think it has that special alignment requirements?
> > Are they cache related? Not at all!

> I'm not sure I can see how you find "don't use cache" not cache related. 
> Saving the resources needed for cache would seem to obviously leave them 
> for other processes.

I feel that word "direct" has nothing to do with caching (or lack thereof).
"Direct" means that I want to avoid extra allocations and memcpy:

	write(fd, hugebuf, 100*1024*1024);

Here application uses 100 megs for hugebuf, and if it is not sufficiently
aligned, even smartest kernel in this universe cannot DMA this data
to disk. No way. So it needs to allocate ANOTHER, aligned buffer,
memcpy the data (completely flushing L1 and L2 dcaches), and DMA it
from there. Thus we use twice as much RAM as we really need, and do
a lot of mostly pointless memory moves! And worse, application cannot
even detect it - it works, it's just slow and eats a lot of RAM and CPU.

That's where O_DIRECT helps. When app wants to avoid that, it opens fd
with O_DIRECT. App in effect says: "I *do* want to avoid extra shuffling,
because I will write huge amounts of data in big blocks."

> > But _conceptually_ "direct DMAing" and "do-not-cache-me"
> > are orthogonal, right?
>
> In the sense that you must do DMA or use cache, yes.

Let's say I implemented a heuristic in my cp command:
if source file is indeed a regular file and it is
larger than 128K, allocate aligned 128K buffer
and try to copy it using O_DIRECT i/o.

Then I use this "enhanced" cp command to copy a large directory
recursively, and then I run grep on that directory.

Can you explain why cp shouldn't cache the data it just wrote?
I *am* going to use it shortly thereafter!

> > That's why we also have bona fide fadvise and madvise
> > with FADV_DONTNEED/MADV_DONTNEED:
> >
> > http://www.die.net/doc/linux/man/man2/fadvise.2.html
> > http://www.die.net/doc/linux/man/man2/madvise.2.html
> >
> > _This_ is the proper way to say "do not cache me".
>
> But none of those advisories says how to cache or not, only what the 
> expected behavior will be. So FADV_NOREUSE does not control cache use, 
> it simply allows the system to make assumptions.

Exactly. If you don't need the data, Just let kernel know that.
When you use O_DIRECT, you are saying "I want direct DMA to disk without
extra copying". With fadvise(FADV_DONTNEED) you are saying
"do not expect access in the near future" == "do not try to optimize
for possible accesses in near future" == "do not cache"!.

Again: with O_DIRECT:

write(fd, hugebuf, 100*1024*1024);

kernel _has _difficulty_ caching these data, simply because
data isn't copied into kernel pages anyway, and if user will
continue to use hugebuf after write(), kernel simply cannot
cache that data - it _hasn't_ the data.

But if user will unmap the hugebuf? What then? Should kernel
forget that data in these pages is in effect a cached data from
the file being written to? Not necessarily.

Four years ago Linus wrote an email about it:

http://lkml.org/lkml/2002/5/11/58

btw, as an Oracle DBA on my day job, I completely agree
with Linus on the "deranged monkey" comparison in that mail...
--
vda

Re: open(O_DIRECT) on a tmpfs?

[Bill Davidsen]

Denis Vlasenko wrote:
> On Friday 05 January 2007 17:20, Bill Davidsen wrote:
>   
>> Denis Vlasenko wrote:
>>     
>>> But O_DIRECT is _not_ about cache. At least I think it was not about
>>> cache initially, it was more about DMAing data directly from/to
>>> application address space to/from disks, saving memcpy's and double
>>> allocations. Why do you think it has that special alignment requirements?
>>> Are they cache related? Not at all!
>>>       
>
>   
>> I'm not sure I can see how you find "don't use cache" not cache related. 
>> Saving the resources needed for cache would seem to obviously leave them 
>> for other processes.
>>     
>
> I feel that word "direct" has nothing to do with caching (or lack thereof).
> "Direct" means that I want to avoid extra allocations and memcpy:
>
> 	write(fd, hugebuf, 100*1024*1024);
>
> Here application uses 100 megs for hugebuf, and if it is not sufficiently
> aligned, even smartest kernel in this universe cannot DMA this data
> to disk. No way. So it needs to allocate ANOTHER, aligned buffer,
> memcpy the data (completely flushing L1 and L2 dcaches), and DMA it
> from there. Thus we use twice as much RAM as we really need, and do
> a lot of mostly pointless memory moves! And worse, application cannot
> even detect it - it works, it's just slow and eats a lot of RAM and CPU.
>
> That's where O_DIRECT helps. When app wants to avoid that, it opens fd
> with O_DIRECT. App in effect says: "I *do* want to avoid extra shuffling,
> because I will write huge amounts of data in big blocks."
>
>   
>>> But _conceptually_ "direct DMAing" and "do-not-cache-me"
>>> are orthogonal, right?
>>>       
>> In the sense that you must do DMA or use cache, yes.
>>     
>
> Let's say I implemented a heuristic in my cp command:
> if source file is indeed a regular file and it is
> larger than 128K, allocate aligned 128K buffer
> and try to copy it using O_DIRECT i/o.
>
> Then I use this "enhanced" cp command to copy a large directory
> recursively, and then I run grep on that directory.
>
> Can you explain why cp shouldn't cache the data it just wrote?
> I *am* going to use it shortly thereafter!
>
>   
>>> That's why we also have bona fide fadvise and madvise
>>> with FADV_DONTNEED/MADV_DONTNEED:
>>>
>>> http://www.die.net/doc/linux/man/man2/fadvise.2.html
>>> http://www.die.net/doc/linux/man/man2/madvise.2.html
>>>
>>> _This_ is the proper way to say "do not cache me".
>>>       
>> But none of those advisories says how to cache or not, only what the 
>> expected behavior will be. So FADV_NOREUSE does not control cache use, 
>> it simply allows the system to make assumptions.
>>     
>
> Exactly. If you don't need the data, Just let kernel know that.
> When you use O_DIRECT, you are saying "I want direct DMA to disk without
> extra copying". With fadvise(FADV_DONTNEED) you are saying
> "do not expect access in the near future" == "do not try to optimize
> for possible accesses in near future" == "do not cache"!.
>   
As long as "don't cache" doesn't imply "don't buffer." In the case of a 
large copy or other large single-file write (8.5GB backup DVDs come to 
mind), the desired behavior is to buffer if possible, start writing 
immediately (data will not change in buffer), and release the buffer as 
soon as write is complete. That doesn't seem to be the current 
interpretation of DONTNEED. Or O_DIRECT either, I agree.
> Again: with O_DIRECT:
>
> write(fd, hugebuf, 100*1024*1024);
>
> kernel _has _difficulty_ caching these data, simply because
> data isn't copied into kernel pages anyway, and if user will
> continue to use hugebuf after write(), kernel simply cannot
> cache that data - it _hasn't_ the data.
>   
In linux if you point the gun at your foot and pull the trigger it goes 
bang. I have no problem with that.
> But if user will unmap the hugebuf? What then? Should kernel
> forget that data in these pages is in effect a cached data from
> the file being written to? Not necessarily.
>   
Why should the kernel make an effort to remember? Incompetence, like 
virtue, is its own reward.
> Four years ago Linus wrote an email about it:
>
> http://lkml.org/lkml/2002/5/11/58
>
> btw, as an Oracle DBA on my day job, I completely agree
> with Linus on the "deranged monkey" comparison in that mail...
The problem with the suggested Linux implementation is that it's 
complex, and currently would move a lot of the logic into user space, in 
code which is probably not portable, or might tickle bad behavior on 
other systems.

Around 2.4.16 (or an -aa variant) I tried code to track writes per file, 
and if some number of bytes had been written to a file without a read or 
seek, any buffered blocks were queued to be written. This got around the 
behavior of generating data until memory was full, then writing it all 
out and having the disk very busy. It was just a proof of concept, but 
it did spread the disk writes to a more constant load and more 
consistent response to other i/o. There doesn't seem to be an easy 
tunable to do this, probably because the need isn't all that common.

-- 
bill davidsen <davidsen@tmr.com>
  CTO TMR Associates, Inc
  Doing interesting things with small computers since 1979

Re: open(O_DIRECT) on a tmpfs?

[Michael Tokarev]

Hugh Dickins wrote:
> On Thu, 4 Jan 2007, Michael Tokarev wrote:
>> I wonder why open() with O_DIRECT (for example) bit set is
>> disallowed on a tmpfs (again, for example) filesystem,
>> returning EINVAL.
[]
> p.s.  You said "O_DIRECT (for example)" - what other open
> flag do you think tmpfs should support which it does not?

Well.  Somehow I was under an impression O_SYNC behaves the
same as O_DIRECT on a tmpfs.  But I was wrong - tmpfs permits
O_SYNC opens just fine.  Strange thing to do having in mind
its behaviour with O_DIRECT - to me it's inconsistent ;)
But that's it - looks like only O_DIRECT is "mishandled"
(which is not a big deal obviously).

Thanks for your time!

/mjt

Re: open(O_DIRECT) on a tmpfs?

[Bodo Eggert]

Michael Tokarev <mjt@tls.msk.ru> wrote:

> I wonder why open() with O_DIRECT (for example) bit set is
> disallowed on a tmpfs (again, for example) filesystem,
> returning EINVAL.
> 
> Yes, the question may seems strange a bit, because of two
> somewhat conflicting reasons.  First, there's no reason to
> use O_DIRECT with tmpfs in a first place, because tmpfs does
> not have backing store at all, so there's no place to do
> direct writes to.  But on another hand, again due to the very
> nature of tmpfs, there's no reason not to allow O_DIRECT
> open and just ignore it, -- exactly because there's no
> backing store for this filesystem.

I'm using a tmpfs as a mostly-ramdisk, that is I've set up a large swap
partition in case I need the RAM instead of using it for a filesystem.
Therefore it will sometimes have a backing store.

OTOH, ramfs does not have this property (the cache is the backing store),
so it would make sense to allow it at least there.

BTW: Maybe you could use a ramdisk instead of the loop-on-tmpfs.
-- 
Ich danke GMX dafür, die Verwendung meiner Adressen mittels per SPF
verbreiteten Lügen zu sabotieren.

http://david.woodhou.se/why-not-spf.html