[PATCH] percpu_counter: change inaccurate comment

[PATCH] percpu_counter: change inaccurate comment

[Eric Dumazet]

percpu_counter used to be huge objects, they are not anymore,
thanks to fine alloc_percpu() granularity.

We now consume 4 bytes per possible cpu.

Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
---
 include/linux/percpu_counter.h |    3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/linux/percpu_counter.h b/include/linux/percpu_counter.h
index 8a7d510..5e2cb5c 100644
--- a/include/linux/percpu_counter.h
+++ b/include/linux/percpu_counter.h
@@ -3,7 +3,8 @@
 /*
  * A simple "approximate counter" for use in ext2 and ext3 superblocks.
  *
- * WARNING: these things are HUGE.  4 kbytes per counter on 32-way P4.
+ * WARNING: these things are big :  4 bytes per possible cpu per counter.
+ * For a 64 cpus machine: 256 bytes + sizeof(struct percpu_counter)
  */
 
 #include <linux/spinlock.h>

Re: [PATCH] percpu_counter: change inaccurate comment

[Christoph Lameter]

On Thu, 7 Oct 2010, Eric Dumazet wrote:

> percpu_counter used to be huge objects, they are not anymore,
> thanks to fine alloc_percpu() granularity.
>
> We now consume 4 bytes per possible cpu.

Ahh I did not notice that they switched.... Then we may also want the
following patch to improve performance. It would even be better if we
could do a this_cpu_add(fbc->counters, amount) there but the 64 bit size
fo count looks suspiciously like you are expecting overflows beyond ints
max size.


Subject: percpu_counter: Use this_cpu_ptr instead of per_cpu_ptr

this_cpu_ptr avoids an array lookup and can use the percpu
offset of the local cpu directly.

Signed-off-by: Christoph Lameter <cl@linux.com>

---
 lib/percpu_counter.c |    6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

Index: linux-2.6/lib/percpu_counter.c
===================================================================
--- linux-2.6.orig/lib/percpu_counter.c	2010-10-07 14:36:39.000000000 -0500
+++ linux-2.6/lib/percpu_counter.c	2010-10-07 14:38:43.000000000 -0500
@@ -30,9 +30,9 @@ void __percpu_counter_add(struct percpu_
 {
 	s64 count;
 	s32 *pcount;
-	int cpu = get_cpu();

-	pcount = per_cpu_ptr(fbc->counters, cpu);
+	preempt_disable();
+	pcount = this_cpu_ptr(fbc->counters);
 	count = *pcount + amount;
 	if (count >= batch || count <= -batch) {
 		spin_lock(&fbc->lock);
@@ -42,7 +42,7 @@ void __percpu_counter_add(struct percpu_
 	} else {
 		*pcount = count;
 	}
-	put_cpu();
+	preempt_enable();
 }
 EXPORT_SYMBOL(__percpu_counter_add);

Re: [PATCH] percpu_counter: change inaccurate comment

[Eric Dumazet]

Le jeudi 07 octobre 2010 à 14:42 -0500, Christoph Lameter a écrit :
> On Thu, 7 Oct 2010, Eric Dumazet wrote:
> 
> > percpu_counter used to be huge objects, they are not anymore,
> > thanks to fine alloc_percpu() granularity.
> >
> > We now consume 4 bytes per possible cpu.
> 
> Ahh I did not notice that they switched.... Then we may also want the
> following patch to improve performance. It would even be better if we
> could do a this_cpu_add(fbc->counters, amount) there but the 64 bit size
> fo count looks suspiciously like you are expecting overflows beyond ints
> max size.
> 

Yes, this 64bit interface is very strange.

> 
> Subject: percpu_counter: Use this_cpu_ptr instead of per_cpu_ptr
> 
> this_cpu_ptr avoids an array lookup and can use the percpu
> offset of the local cpu directly.
> 
> Signed-off-by: Christoph Lameter <cl@linux.com>
> 
> ---
>  lib/percpu_counter.c |    6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> Index: linux-2.6/lib/percpu_counter.c
> ===================================================================
> --- linux-2.6.orig/lib/percpu_counter.c	2010-10-07 14:36:39.000000000 -0500
> +++ linux-2.6/lib/percpu_counter.c	2010-10-07 14:38:43.000000000 -0500
> @@ -30,9 +30,9 @@ void __percpu_counter_add(struct percpu_
>  {
>  	s64 count;
>  	s32 *pcount;
> -	int cpu = get_cpu();
> 
> -	pcount = per_cpu_ptr(fbc->counters, cpu);
> +	preempt_disable();
> +	pcount = this_cpu_ptr(fbc->counters);
>  	count = *pcount + amount;
>  	if (count >= batch || count <= -batch) {
>  		spin_lock(&fbc->lock);
> @@ -42,7 +42,7 @@ void __percpu_counter_add(struct percpu_
>  	} else {
>  		*pcount = count;
>  	}
> -	put_cpu();
> +	preempt_enable();
>  }
>  EXPORT_SYMBOL(__percpu_counter_add);
> 

If enclosed by preempt_disable()/preempt_enable(), maybe we could use
__this_cpu_ptr() ?

Re: [PATCH] percpu_counter: change inaccurate comment

[Christoph Lameter]

On Thu, 7 Oct 2010, Eric Dumazet wrote:

> If enclosed by preempt_disable()/preempt_enable(), maybe we could use
> __this_cpu_ptr() ?

The only difference between __this_cpu_ptr and this_cpu_ptr is that
this_cpu_ptr checks that preempt was disabled. __this_cpu_ptr allows use
even without preempt. Preempt must be disabled here so the use of
this_cpu_ptr is appropriate.

Re: [PATCH] percpu_counter: change inaccurate comment

[Eric Dumazet]

Le jeudi 07 octobre 2010 à 15:13 -0500, Christoph Lameter a écrit :
> On Thu, 7 Oct 2010, Eric Dumazet wrote:
> 
> > If enclosed by preempt_disable()/preempt_enable(), maybe we could use
> > __this_cpu_ptr() ?
> 
> The only difference between __this_cpu_ptr and this_cpu_ptr is that
> this_cpu_ptr checks that preempt was disabled. __this_cpu_ptr allows use
> even without preempt. Preempt must be disabled here so the use of
> this_cpu_ptr is appropriate.
> 
> 

Thats not how I read the thing.

In both variants, preemption _must_ be disabled, its only the context
that can tell how sure we are...

<quote>

commit 7340a0b15280c

__this_cpu_ptr  -> Do not check for preemption context
this_cpu_ptr    -> Check preemption context

</quote>

If preemption was enabled, both pointers would not be very useful...

We use __this_cpu_ptr() in contexts where cpu _cannot_ change under us,
(we just disabled preemption one line above), so its not necessary to
perform the check.

vi +316 include/linux/percpu.h

#define _this_cpu_generic_to_op(pcp, val, op)                   \
do {                                                            \
        preempt_disable();                                      \
        *__this_cpu_ptr(&(pcp)) op val;                         \
        preempt_enable();                                       \
} while (0)

...

#define __this_cpu_generic_to_op(pcp, val, op)                  \
do {                                                            \
        *__this_cpu_ptr(&(pcp)) op val;                         \
} while (0)

Re: [PATCH] percpu_counter: change inaccurate comment

[Christoph Lameter]

On Thu, 7 Oct 2010, Eric Dumazet wrote:

> >
> > The only difference between __this_cpu_ptr and this_cpu_ptr is that
> > this_cpu_ptr checks that preempt was disabled. __this_cpu_ptr allows use
> > even without preempt. Preempt must be disabled here so the use of
> > this_cpu_ptr is appropriate.
> >
> >
>
> Thats not how I read the thing.
>
> In both variants, preemption _must_ be disabled, its only the context
> that can tell how sure we are...

Correct. One checks the other one does not check.

> <quote>
>
> commit 7340a0b15280c
>
> __this_cpu_ptr  -> Do not check for preemption context
> this_cpu_ptr    -> Check preemption context
>
> </quote>
>
> If preemption was enabled, both pointers would not be very useful...
>
> We use __this_cpu_ptr() in contexts where cpu _cannot_ change under us,
> (we just disabled preemption one line above), so its not necessary to
> perform the check.

Someone could screw up the code in the future. Thats what its there for.