Re: [PATCH] memcg: remove unneeded preempt_disable

Re: [PATCH] memcg: remove unneeded preempt_disable

[Andrew Morton]

(cc linux-arch)

On Wed, 17 Aug 2011 23:50:53 -0700
Greg Thelen <gthelen@google.com> wrote:

> Both mem_cgroup_charge_statistics() and mem_cgroup_move_account() were
> unnecessarily disabling preemption when adjusting per-cpu counters:
>     preempt_disable()
>     __this_cpu_xxx()
>     __this_cpu_yyy()
>     preempt_enable()
> 
> This change does not disable preemption and thus CPU switch is possible
> within these routines.  This does not cause a problem because the total
> of all cpu counters is summed when reporting stats.  Now both
> mem_cgroup_charge_statistics() and mem_cgroup_move_account() look like:
>     this_cpu_xxx()
>     this_cpu_yyy()
> 
> ...
>
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -664,24 +664,20 @@ static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
>  static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
>  					 bool file, int nr_pages)
>  {
> -	preempt_disable();
> -
>  	if (file)
> -		__this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
> +		this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
>  	else
> -		__this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
> +		this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
>  
>  	/* pagein of a big page is an event. So, ignore page size */
>  	if (nr_pages > 0)
> -		__this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
> +		this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
>  	else {
> -		__this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
> +		this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
>  		nr_pages = -nr_pages; /* for event */
>  	}
>  
> -	__this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
> -
> -	preempt_enable();
> +	this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
>  }

On non-x86 architectures this_cpu_add() internally does
preempt_disable() and preempt_enable().  So the patch is a small
optimisation for x86 and a larger deoptimisation for non-x86.

I think I'll apply it, as the call frequency is low (correct?) and the
problem will correct itself as other architectures implement their
atomic this_cpu_foo() operations.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Andrew Morton]

(cc linux-arch)

On Wed, 17 Aug 2011 23:50:53 -0700
Greg Thelen <gthelen@google.com> wrote:

> Both mem_cgroup_charge_statistics() and mem_cgroup_move_account() were
> unnecessarily disabling preemption when adjusting per-cpu counters:
>     preempt_disable()
>     __this_cpu_xxx()
>     __this_cpu_yyy()
>     preempt_enable()
> 
> This change does not disable preemption and thus CPU switch is possible
> within these routines.  This does not cause a problem because the total
> of all cpu counters is summed when reporting stats.  Now both
> mem_cgroup_charge_statistics() and mem_cgroup_move_account() look like:
>     this_cpu_xxx()
>     this_cpu_yyy()
> 
> ...
>
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -664,24 +664,20 @@ static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
>  static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
>  					 bool file, int nr_pages)
>  {
> -	preempt_disable();
> -
>  	if (file)
> -		__this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
> +		this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
>  	else
> -		__this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
> +		this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
>  
>  	/* pagein of a big page is an event. So, ignore page size */
>  	if (nr_pages > 0)
> -		__this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
> +		this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
>  	else {
> -		__this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
> +		this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
>  		nr_pages = -nr_pages; /* for event */
>  	}
>  
> -	__this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
> -
> -	preempt_enable();
> +	this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
>  }

On non-x86 architectures this_cpu_add() internally does
preempt_disable() and preempt_enable().  So the patch is a small
optimisation for x86 and a larger deoptimisation for non-x86.

I think I'll apply it, as the call frequency is low (correct?) and the
problem will correct itself as other architectures implement their
atomic this_cpu_foo() operations.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Greg Thelen]

On Thu, Aug 18, 2011 at 2:40 PM, Andrew Morton
<akpm@linux-foundation.org> wrote:
> (cc linux-arch)
>
> On Wed, 17 Aug 2011 23:50:53 -0700
> Greg Thelen <gthelen@google.com> wrote:
>
>> Both mem_cgroup_charge_statistics() and mem_cgroup_move_account() were
>> unnecessarily disabling preemption when adjusting per-cpu counters:
>>     preempt_disable()
>>     __this_cpu_xxx()
>>     __this_cpu_yyy()
>>     preempt_enable()
>>
>> This change does not disable preemption and thus CPU switch is possible
>> within these routines.  This does not cause a problem because the total
>> of all cpu counters is summed when reporting stats.  Now both
>> mem_cgroup_charge_statistics() and mem_cgroup_move_account() look like:
>>     this_cpu_xxx()
>>     this_cpu_yyy()
>>
>> ...
>>
>> --- a/mm/memcontrol.c
>> +++ b/mm/memcontrol.c
>> @@ -664,24 +664,20 @@ static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
>>  static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
>>                                        bool file, int nr_pages)
>>  {
>> -     preempt_disable();
>> -
>>       if (file)
>> -             __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
>> +             this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
>>       else
>> -             __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
>> +             this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
>>
>>       /* pagein of a big page is an event. So, ignore page size */
>>       if (nr_pages > 0)
>> -             __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
>> +             this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
>>       else {
>> -             __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
>> +             this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
>>               nr_pages = -nr_pages; /* for event */
>>       }
>>
>> -     __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
>> -
>> -     preempt_enable();
>> +     this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
>>  }
>
> On non-x86 architectures this_cpu_add() internally does
> preempt_disable() and preempt_enable().  So the patch is a small
> optimisation for x86 and a larger deoptimisation for non-x86.
>
> I think I'll apply it, as the call frequency is low (correct?) and the
> problem will correct itself as other architectures implement their
> atomic this_cpu_foo() operations.

mem_cgroup_charge_statistics() is a common operation, which is called
on each memcg page charge and uncharge.

The per arch/config effects of this patch:
* non-preemptible kernels: there's no difference before/after this patch.
* preemptible x86: this patch helps by removing an unnecessary
preempt_disable/enable.
* preemptible non-x86: this patch hurts by adding implicit
preempt_disable/enable around each operation.

So I am uncomfortable this patch's unmeasured impact on archs that do
not have atomic this_cpu_foo() operations.  Please drop the patch from
mmotm.  Sorry for the noise.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> 
> I think I'll apply it, as the call frequency is low (correct?) and the
> problem will correct itself as other architectures implement their
> atomic this_cpu_foo() operations. 

Which leads me to wonder, can anything but x86 implement that this_cpu_*
muck? I doubt any of the risk chips can actually do all this.
Maybe Itanic, but then that seems to be dying fast.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, Peter Zijlstra wrote:

> On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> >
> > I think I'll apply it, as the call frequency is low (correct?) and the
> > problem will correct itself as other architectures implement their
> > atomic this_cpu_foo() operations.
>
> Which leads me to wonder, can anything but x86 implement that this_cpu_*
> muck? I doubt any of the risk chips can actually do all this.
> Maybe Itanic, but then that seems to be dying fast.

The cpu needs to have an RMW instruction that does something to a
variable relative to a register that points to the per cpu base.

Thats generally possible. The problem is how expensive the RMW is going to
be.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, Peter Zijlstra wrote:

> On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> >
> > I think I'll apply it, as the call frequency is low (correct?) and the
> > problem will correct itself as other architectures implement their
> > atomic this_cpu_foo() operations.
>
> Which leads me to wonder, can anything but x86 implement that this_cpu_*
> muck? I doubt any of the risk chips can actually do all this.
> Maybe Itanic, but then that seems to be dying fast.

The cpu needs to have an RMW instruction that does something to a
variable relative to a register that points to the per cpu base.

Thats generally possible. The problem is how expensive the RMW is going to
be.

Re: [PATCH] memcg: remove unneeded preempt_disable

[James Bottomley]

On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
> On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> 
> > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> > >
> > > I think I'll apply it, as the call frequency is low (correct?) and the
> > > problem will correct itself as other architectures implement their
> > > atomic this_cpu_foo() operations.
> >
> > Which leads me to wonder, can anything but x86 implement that this_cpu_*
> > muck? I doubt any of the risk chips can actually do all this.
> > Maybe Itanic, but then that seems to be dying fast.
> 
> The cpu needs to have an RMW instruction that does something to a
> variable relative to a register that points to the per cpu base.
> 
> Thats generally possible. The problem is how expensive the RMW is going to
> be.

Risc systems generally don't have a single instruction for this, that's
correct.  Obviously we can do it as a non atomic sequence: read
variable, compute relative, read, modify, write ... but there's
absolutely no point hand crafting that in asm since the compiler can
usually work it out nicely.  And, of course, to have this atomic, we
have to use locks, which ends up being very expensive.

James

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, James Bottomley wrote:

> On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
> > On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> >
> > > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> > > >
> > > > I think I'll apply it, as the call frequency is low (correct?) and the
> > > > problem will correct itself as other architectures implement their
> > > > atomic this_cpu_foo() operations.
> > >
> > > Which leads me to wonder, can anything but x86 implement that this_cpu_*
> > > muck? I doubt any of the risk chips can actually do all this.
> > > Maybe Itanic, but then that seems to be dying fast.
> >
> > The cpu needs to have an RMW instruction that does something to a
> > variable relative to a register that points to the per cpu base.
> >
> > Thats generally possible. The problem is how expensive the RMW is going to
> > be.
>
> Risc systems generally don't have a single instruction for this, that's
> correct.  Obviously we can do it as a non atomic sequence: read
> variable, compute relative, read, modify, write ... but there's
> absolutely no point hand crafting that in asm since the compiler can
> usually work it out nicely.  And, of course, to have this atomic, we
> have to use locks, which ends up being very expensive.

ARM seems to have these LDREX/STREX instructions for that purpose which
seem to be used for generating atomic instructions without lockes. I guess
other RISC architectures have similar means of doing it?




--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, James Bottomley wrote:

> On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
> > On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> >
> > > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> > > >
> > > > I think I'll apply it, as the call frequency is low (correct?) and the
> > > > problem will correct itself as other architectures implement their
> > > > atomic this_cpu_foo() operations.
> > >
> > > Which leads me to wonder, can anything but x86 implement that this_cpu_*
> > > muck? I doubt any of the risk chips can actually do all this.
> > > Maybe Itanic, but then that seems to be dying fast.
> >
> > The cpu needs to have an RMW instruction that does something to a
> > variable relative to a register that points to the per cpu base.
> >
> > Thats generally possible. The problem is how expensive the RMW is going to
> > be.
>
> Risc systems generally don't have a single instruction for this, that's
> correct.  Obviously we can do it as a non atomic sequence: read
> variable, compute relative, read, modify, write ... but there's
> absolutely no point hand crafting that in asm since the compiler can
> usually work it out nicely.  And, of course, to have this atomic, we
> have to use locks, which ends up being very expensive.

ARM seems to have these LDREX/STREX instructions for that purpose which
seem to be used for generating atomic instructions without lockes. I guess
other RISC architectures have similar means of doing it?

Re: [PATCH] memcg: remove unneeded preempt_disable

[James Bottomley]

Christoph Lameter <cl@linux.com> wrote:

>On Thu, 25 Aug 2011, James Bottomley wrote:
>
>> On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
>> > On Thu, 25 Aug 2011, Peter Zijlstra wrote:
>> >
>> > > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
>> > > >
>> > > > I think I'll apply it, as the call frequency is low (correct?)
>and the
>> > > > problem will correct itself as other architectures implement
>their
>> > > > atomic this_cpu_foo() operations.
>> > >
>> > > Which leads me to wonder, can anything but x86 implement that
>this_cpu_*
>> > > muck? I doubt any of the risk chips can actually do all this.
>> > > Maybe Itanic, but then that seems to be dying fast.
>> >
>> > The cpu needs to have an RMW instruction that does something to a
>> > variable relative to a register that points to the per cpu base.
>> >
>> > Thats generally possible. The problem is how expensive the RMW is
>going to
>> > be.
>>
>> Risc systems generally don't have a single instruction for this,
>that's
>> correct.  Obviously we can do it as a non atomic sequence: read
>> variable, compute relative, read, modify, write ... but there's
>> absolutely no point hand crafting that in asm since the compiler can
>> usually work it out nicely.  And, of course, to have this atomic, we
>> have to use locks, which ends up being very expensive.
>
>ARM seems to have these LDREX/STREX instructions for that purpose which
>seem to be used for generating atomic instructions without lockes. I
>guess
>other RISC architectures have similar means of doing it?

Arm isn't really risc.  Most don't.  However even with ldrex/strex you need two instructions for rmw.

James
-- 
Sent from my Android phone with K-9 Mail. Please excuse my brevity and top posting.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[James Bottomley]

Christoph Lameter <cl@linux.com> wrote:

>On Thu, 25 Aug 2011, James Bottomley wrote:
>
>> On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
>> > On Thu, 25 Aug 2011, Peter Zijlstra wrote:
>> >
>> > > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
>> > > >
>> > > > I think I'll apply it, as the call frequency is low (correct?)
>and the
>> > > > problem will correct itself as other architectures implement
>their
>> > > > atomic this_cpu_foo() operations.
>> > >
>> > > Which leads me to wonder, can anything but x86 implement that
>this_cpu_*
>> > > muck? I doubt any of the risk chips can actually do all this.
>> > > Maybe Itanic, but then that seems to be dying fast.
>> >
>> > The cpu needs to have an RMW instruction that does something to a
>> > variable relative to a register that points to the per cpu base.
>> >
>> > Thats generally possible. The problem is how expensive the RMW is
>going to
>> > be.
>>
>> Risc systems generally don't have a single instruction for this,
>that's
>> correct.  Obviously we can do it as a non atomic sequence: read
>> variable, compute relative, read, modify, write ... but there's
>> absolutely no point hand crafting that in asm since the compiler can
>> usually work it out nicely.  And, of course, to have this atomic, we
>> have to use locks, which ends up being very expensive.
>
>ARM seems to have these LDREX/STREX instructions for that purpose which
>seem to be used for generating atomic instructions without lockes. I
>guess
>other RISC architectures have similar means of doing it?

Arm isn't really risc.  Most don't.  However even with ldrex/strex you need two instructions for rmw.

James
-- 
Sent from my Android phone with K-9 Mail. Please excuse my brevity and top posting.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, James Bottomley wrote:

> >ARM seems to have these LDREX/STREX instructions for that purpose which
> >seem to be used for generating atomic instructions without lockes. I
> >guess
> >other RISC architectures have similar means of doing it?
>
> Arm isn't really risc.  Most don't.  However even with ldrex/strex you need two instructions for rmw.

Well then what is "really risc"? RISC is an old beaten down marketing term
AFAICT and ARM claims it too.

Re: [PATCH] memcg: remove unneeded preempt_disable

[James Bottomley]

Christoph Lameter <cl@linux.com> wrote:

>On Thu, 25 Aug 2011, James Bottomley wrote:
>
>> >ARM seems to have these LDREX/STREX instructions for that purpose
>which
>> >seem to be used for generating atomic instructions without lockes. I
>> >guess
>> >other RISC architectures have similar means of doing it?
>>
>> Arm isn't really risc.  Most don't.  However even with ldrex/strex
>you need two instructions for rmw.
>
>Well then what is "really risc"? RISC is an old beaten down marketing
>term
>AFAICT and ARM claims it too.

Reduced Instruction Set Computer.  This is why we're unlikely to have complex atomic instructions: the principle of risc is that you build them up from basic ones.

James 
-- 
Sent from my Android phone with K-9 Mail. Please excuse my brevity and top posting.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 11:31 -0500, Christoph Lameter wrote:
> On Thu, 25 Aug 2011, James Bottomley wrote:
> 
> > On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
> > > On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> > >
> > > > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> > > > >
> > > > > I think I'll apply it, as the call frequency is low (correct?) and the
> > > > > problem will correct itself as other architectures implement their
> > > > > atomic this_cpu_foo() operations.
> > > >
> > > > Which leads me to wonder, can anything but x86 implement that this_cpu_*
> > > > muck? I doubt any of the risk chips can actually do all this.
> > > > Maybe Itanic, but then that seems to be dying fast.
> > >
> > > The cpu needs to have an RMW instruction that does something to a
> > > variable relative to a register that points to the per cpu base.
> > >
> > > Thats generally possible. The problem is how expensive the RMW is going to
> > > be.
> >
> > Risc systems generally don't have a single instruction for this, that's
> > correct.  Obviously we can do it as a non atomic sequence: read
> > variable, compute relative, read, modify, write ... but there's
> > absolutely no point hand crafting that in asm since the compiler can
> > usually work it out nicely.  And, of course, to have this atomic, we
> > have to use locks, which ends up being very expensive.
> 
> ARM seems to have these LDREX/STREX instructions for that purpose which
> seem to be used for generating atomic instructions without lockes. I guess
> other RISC architectures have similar means of doing it?

Even with LL/SC and the CPU base in a register you need to do something
like:

again:
	LL $target-reg, $cpubase-reg + offset
	<foo>
	SC $ret, $target-reg, $cpubase-reg + offset
	if !$ret goto again

Its the +offset that's problematic, it either doesn't exist or is very
limited (a quick look at the MIPS instruction set gives a limit of 64k).

Without the +offset you need:

again:
	$tmp-reg = $cpubase-reg
	$tmp-reg += offset;

	LL $target-reg, $tmp-reg
	<foo>
	SC $ret, $target-reg, $tmp-reg
	if !$ret goto again


Which is wide open to migration races. Also, very often there are
constraints on LL/SC that mandate we use preempt_disable/enable around
its use, which pretty much voids the whole purpose, since if we disable
preemption we might as well just use C (ARM belongs in this class).

It does look POWERPC's lwarx/stwcx is sane enough, although the
instruction reference I found doesn't list what happens if the LL/SC
doesn't use the same effective address or has other loads/stores in
between, if its ok with those and simply fails the SC it should be good.

Still, creating atomic ops for per-cpu ops might be more expensive than
simply doing the preempt-disable/rmw/enable dance, dunno don't know
these archs that well.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 11:31 -0500, Christoph Lameter wrote:
> On Thu, 25 Aug 2011, James Bottomley wrote:
> 
> > On Thu, 2011-08-25 at 10:11 -0500, Christoph Lameter wrote:
> > > On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> > >
> > > > On Thu, 2011-08-18 at 14:40 -0700, Andrew Morton wrote:
> > > > >
> > > > > I think I'll apply it, as the call frequency is low (correct?) and the
> > > > > problem will correct itself as other architectures implement their
> > > > > atomic this_cpu_foo() operations.
> > > >
> > > > Which leads me to wonder, can anything but x86 implement that this_cpu_*
> > > > muck? I doubt any of the risk chips can actually do all this.
> > > > Maybe Itanic, but then that seems to be dying fast.
> > >
> > > The cpu needs to have an RMW instruction that does something to a
> > > variable relative to a register that points to the per cpu base.
> > >
> > > Thats generally possible. The problem is how expensive the RMW is going to
> > > be.
> >
> > Risc systems generally don't have a single instruction for this, that's
> > correct.  Obviously we can do it as a non atomic sequence: read
> > variable, compute relative, read, modify, write ... but there's
> > absolutely no point hand crafting that in asm since the compiler can
> > usually work it out nicely.  And, of course, to have this atomic, we
> > have to use locks, which ends up being very expensive.
> 
> ARM seems to have these LDREX/STREX instructions for that purpose which
> seem to be used for generating atomic instructions without lockes. I guess
> other RISC architectures have similar means of doing it?

Even with LL/SC and the CPU base in a register you need to do something
like:

again:
	LL $target-reg, $cpubase-reg + offset
	<foo>
	SC $ret, $target-reg, $cpubase-reg + offset
	if !$ret goto again

Its the +offset that's problematic, it either doesn't exist or is very
limited (a quick look at the MIPS instruction set gives a limit of 64k).

Without the +offset you need:

again:
	$tmp-reg = $cpubase-reg
	$tmp-reg += offset;

	LL $target-reg, $tmp-reg
	<foo>
	SC $ret, $target-reg, $tmp-reg
	if !$ret goto again


Which is wide open to migration races. Also, very often there are
constraints on LL/SC that mandate we use preempt_disable/enable around
its use, which pretty much voids the whole purpose, since if we disable
preemption we might as well just use C (ARM belongs in this class).

It does look POWERPC's lwarx/stwcx is sane enough, although the
instruction reference I found doesn't list what happens if the LL/SC
doesn't use the same effective address or has other loads/stores in
between, if its ok with those and simply fails the SC it should be good.

Still, creating atomic ops for per-cpu ops might be more expensive than
simply doing the preempt-disable/rmw/enable dance, dunno don't know
these archs that well.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, James Bottomley wrote:

> >Well then what is "really risc"? RISC is an old beaten down marketing
> >term
> >AFAICT and ARM claims it too.
>
> Reduced Instruction Set Computer.  This is why we're unlikely to have
> complex atomic instructions: the principle of risc is that you build
> them up from basic ones.

RISC cpus have instruction to construct complex atomic actions by the cpu
as I have shown before for ARM.

Principles always have exceptions to them.

(That statement in itself is a principle that should have an exception I
guess. But then language often only makes sense when it contains
contradictions.)

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 13:46 -0500, Christoph Lameter wrote:
> 
> RISC cpus have instruction to construct complex atomic actions by the cpu
> as I have shown before for ARM. 

Right, but it only makes sense if the whole thing remains cheaper than
the trivial implementation already available.

For instance, the ARM LL/SC constraints pretty much mandate we do
preempt_disable()/preempt_enable() around them, at which point the point
of doing LL/SC is gone (except maybe for the irqsafe_this_cpu_* stuff).

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 13:46 -0500, Christoph Lameter wrote:
> 
> RISC cpus have instruction to construct complex atomic actions by the cpu
> as I have shown before for ARM. 

Right, but it only makes sense if the whole thing remains cheaper than
the trivial implementation already available.

For instance, the ARM LL/SC constraints pretty much mandate we do
preempt_disable()/preempt_enable() around them, at which point the point
of doing LL/SC is gone (except maybe for the irqsafe_this_cpu_* stuff).

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 20:49 +0200, Peter Zijlstra wrote:
> the ARM LL/SC constraints pretty much mandate we do
> preempt_disable()/preempt_enable() around them, 

My bad, that was MIPS, it states that portable programs should not issue
load/store/prefetch insn inside a LL/SC region or the behaviour is
undefined or somesuch.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 20:49 +0200, Peter Zijlstra wrote:
> the ARM LL/SC constraints pretty much mandate we do
> preempt_disable()/preempt_enable() around them, 

My bad, that was MIPS, it states that portable programs should not issue
load/store/prefetch insn inside a LL/SC region or the behaviour is
undefined or somesuch.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 13:46 -0500, Christoph Lameter wrote:
> 
> RISC cpus have instruction to construct complex atomic actions by the cpu
> as I have shown before for ARM. 

Also, I thought this_cpu thing's were at best locally atomic. If you
make them full blown atomic ops then even __this_cpu ops will have to be
full atomic ops, otherwise:


CPU0			CPU(1)

this_cpu_inc(&foo);	preempt_disable();
			__this_cpu_inc(&foo);
			preempt_enable();

might step on each other's toes.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Christoph Lameter]

On Thu, 25 Aug 2011, Peter Zijlstra wrote:

> Also, I thought this_cpu thing's were at best locally atomic. If you
> make them full blown atomic ops then even __this_cpu ops will have to be
> full atomic ops, otherwise:
>
>
> CPU0			CPU(1)
>
> this_cpu_inc(&foo);	preempt_disable();
> 			__this_cpu_inc(&foo);
> 			preempt_enable();
>
> might step on each other's toes.

They would both have their own instance of "foo". per cpu atomicity is
only one requirement of this_cpu_ops. The other is the ability to relocate
accesses relative to the current per cpu area.

Full blown atomicity is almost a superset of per cpu atomicity but its
only usable if the full atomic instructions can also relocate accesses
relative to some base.

Re: [PATCH] memcg: remove unneeded preempt_disable

[James Bottomley]

On Thu, 2011-08-25 at 13:46 -0500, Christoph Lameter wrote:
> On Thu, 25 Aug 2011, James Bottomley wrote:
> 
> > >Well then what is "really risc"? RISC is an old beaten down marketing
> > >term
> > >AFAICT and ARM claims it too.
> >
> > Reduced Instruction Set Computer.  This is why we're unlikely to have
> > complex atomic instructions: the principle of risc is that you build
> > them up from basic ones.
> 
> RISC cpus have instruction to construct complex atomic actions by the cpu
> as I have shown before for ARM.
> 
> Principles always have exceptions to them.
> 
> (That statement in itself is a principle that should have an exception I
> guess. But then language often only makes sense when it contains
> contradictions.)

We seem to be talking at cross purposes.  I'm not saying a risc system
can't do this ... of course the risc primitives can build into whatever
you want.  To make it atomic, you simply add locking.  What I'm saying
is that open coding asm in a macro makes no sense because the compiler
will do it better from C.  Plus, since the net purpose of this patch is
to require us to lock around each op instead of doing a global lock (or
in this case preempt disable) then you're making us less efficient at
executing it.

Therefore from the risc point of view, most of the this_cpu_xxx
operations are things that we don't really care about except that the
result would be easier to read in C.

James

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 14:19 -0500, Christoph Lameter wrote:
> On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> 
> > Also, I thought this_cpu thing's were at best locally atomic. If you
> > make them full blown atomic ops then even __this_cpu ops will have to be
> > full atomic ops, otherwise:
> >
> >
> > CPU0			CPU(1)
> >
> > this_cpu_inc(&foo);	preempt_disable();
> > 			__this_cpu_inc(&foo);
> > 			preempt_enable();
> >
> > might step on each other's toes.
> 
> They would both have their own instance of "foo". per cpu atomicity is
> only one requirement of this_cpu_ops. The other is the ability to relocate
> accesses relative to the current per cpu area.

Ah, but not if the this_cpu_inc() thing ends up being more than a single
instruction, then you have preemption/migration windows. Only when LL/SC
can deal with SC having a different EA from the LL and supports a big
enough offset could this possibly work.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 14:19 -0500, Christoph Lameter wrote:
> On Thu, 25 Aug 2011, Peter Zijlstra wrote:
> 
> > Also, I thought this_cpu thing's were at best locally atomic. If you
> > make them full blown atomic ops then even __this_cpu ops will have to be
> > full atomic ops, otherwise:
> >
> >
> > CPU0			CPU(1)
> >
> > this_cpu_inc(&foo);	preempt_disable();
> > 			__this_cpu_inc(&foo);
> > 			preempt_enable();
> >
> > might step on each other's toes.
> 
> They would both have their own instance of "foo". per cpu atomicity is
> only one requirement of this_cpu_ops. The other is the ability to relocate
> accesses relative to the current per cpu area.

Ah, but not if the this_cpu_inc() thing ends up being more than a single
instruction, then you have preemption/migration windows. Only when LL/SC
can deal with SC having a different EA from the LL and supports a big
enough offset could this possibly work.

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 12:29 -0700, James Bottomley wrote:
> 
> Therefore from the risc point of view, most of the this_cpu_xxx
> operations are things that we don't really care about except that the
> result would be easier to read in C. 

Right, so the current fallback case is pretty much the optimal case for
the RISC machines, which ends up with generic code being better off not
using it much and instead preferring __this_cpu if there's more than
one.

I mean, its absolutely awesome these things are 1 instruction on x86,
but if we pessimize all other 20-odd architectures its just not cool.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [PATCH] memcg: remove unneeded preempt_disable

[Peter Zijlstra]

On Thu, 2011-08-25 at 12:29 -0700, James Bottomley wrote:
> 
> Therefore from the risc point of view, most of the this_cpu_xxx
> operations are things that we don't really care about except that the
> result would be easier to read in C. 

Right, so the current fallback case is pretty much the optimal case for
the RISC machines, which ends up with generic code being better off not
using it much and instead preferring __this_cpu if there's more than
one.

I mean, its absolutely awesome these things are 1 instruction on x86,
but if we pessimize all other 20-odd architectures its just not cool.