Re: Switch per cpu partial page support off for debugging

Re: Switch per cpu partial page support off for debugging

[Pekka Enberg]

On Wed, 23 Nov 2011, Christoph Lameter wrote:
> Eric saw an issue with accounting of slabs during validation. Its not
> possible to determine accurately how many per cpu partial slabs exist at
> any time so this switches off per cpu partial pages during debug.
>
> Subject: Switch per cpu partial page support off for debugging
>
> Otherwise we have accounting issues.
>
> Signed-off-by: Christoph Lameter <cl@linux.com>
>
>
> ---
> mm/slub.c |    4 +++-
> 1 file changed, 3 insertions(+), 1 deletion(-)
>
> Index: linux-2.6/mm/slub.c
> ===================================================================
> --- linux-2.6.orig/mm/slub.c	2011-11-22 11:57:21.000000000 -0600
> +++ linux-2.6/mm/slub.c	2011-11-22 11:57:55.000000000 -0600
> @@ -3027,7 +3027,9 @@ static int kmem_cache_open(struct kmem_c
> 	 *    per node list when we run out of per cpu objects. We only fetch 50%
> 	 *    to keep some capacity around for frees.
> 	 */
> -	if (s->size >= PAGE_SIZE)
> +	if (kmem_cache_debug(s))
> +		s->cpu_partial = 0;
> +	else if (s->size >= PAGE_SIZE)
> 		s->cpu_partial = 2;
> 	else if (s->size >= 1024)
> 		s->cpu_partial = 6;

Looks OK to me. CC'ing Eric and David.

Re: Switch per cpu partial page support off for debugging

[Eric Dumazet]

Le lundi 12 décembre 2011 à 09:16 +0200, Pekka Enberg a écrit :
> On Wed, 23 Nov 2011, Christoph Lameter wrote:
> > Eric saw an issue with accounting of slabs during validation. Its not
> > possible to determine accurately how many per cpu partial slabs exist at
> > any time so this switches off per cpu partial pages during debug.
> >
> > Subject: Switch per cpu partial page support off for debugging
> >
> > Otherwise we have accounting issues.
> >
> > Signed-off-by: Christoph Lameter <cl@linux.com>
> >
> >
> > ---
> > mm/slub.c |    4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > Index: linux-2.6/mm/slub.c
> > ===================================================================
> > --- linux-2.6.orig/mm/slub.c	2011-11-22 11:57:21.000000000 -0600
> > +++ linux-2.6/mm/slub.c	2011-11-22 11:57:55.000000000 -0600
> > @@ -3027,7 +3027,9 @@ static int kmem_cache_open(struct kmem_c
> > 	 *    per node list when we run out of per cpu objects. We only fetch 50%
> > 	 *    to keep some capacity around for frees.
> > 	 */
> > -	if (s->size >= PAGE_SIZE)
> > +	if (kmem_cache_debug(s))
> > +		s->cpu_partial = 0;
> > +	else if (s->size >= PAGE_SIZE)
> > 		s->cpu_partial = 2;
> > 	else if (s->size >= 1024)
> > 		s->cpu_partial = 6;
> 
> Looks OK to me. CC'ing Eric and David.

This seems fine to me, thanks.

Acked-by: Eric Dumazet <eric.dumazet@gmail.com>

Re: Switch per cpu partial page support off for debugging

[David Rientjes]

On Mon, 12 Dec 2011, Pekka Enberg wrote:

> > Eric saw an issue with accounting of slabs during validation. Its not
> > possible to determine accurately how many per cpu partial slabs exist at
> > any time so this switches off per cpu partial pages during debug.
> > 
> > Subject: Switch per cpu partial page support off for debugging
> > 
> > Otherwise we have accounting issues.
> > 
> > Signed-off-by: Christoph Lameter <cl@linux.com>
> > 
> > 
> > ---
> > mm/slub.c |    4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> > 
> > Index: linux-2.6/mm/slub.c
> > ===================================================================
> > --- linux-2.6.orig/mm/slub.c	2011-11-22 11:57:21.000000000 -0600
> > +++ linux-2.6/mm/slub.c	2011-11-22 11:57:55.000000000 -0600
> > @@ -3027,7 +3027,9 @@ static int kmem_cache_open(struct kmem_c
> > 	 *    per node list when we run out of per cpu objects. We only fetch
> > 50%
> > 	 *    to keep some capacity around for frees.
> > 	 */
> > -	if (s->size >= PAGE_SIZE)
> > +	if (kmem_cache_debug(s))
> > +		s->cpu_partial = 0;
> > +	else if (s->size >= PAGE_SIZE)
> > 		s->cpu_partial = 2;
> > 	else if (s->size >= 1024)
> > 		s->cpu_partial = 6;

I think this deserves a comment since it's not clear why it's being done 
neither in the code or by the changelog.  Also, you'd want a check for 
kmem_cache_debug() when storing to /sys/kernel/slab/cache/cpu_partial as 
well, otherwise this could easily be broken from userspace again.  (It 
also seems like Documentation/ABI/testing/sysfs-kernel-slab got left out 
of an update when that was added, too.)

Re: Switch per cpu partial page support off for debugging

[David Rientjes]

On Wed, 14 Dec 2011, David Rientjes wrote:

> > > Eric saw an issue with accounting of slabs during validation. Its not
> > > possible to determine accurately how many per cpu partial slabs exist at
> > > any time so this switches off per cpu partial pages during debug.
> > > 
> > > Subject: Switch per cpu partial page support off for debugging
> > > 
> > > Otherwise we have accounting issues.
> > > 
> > > Signed-off-by: Christoph Lameter <cl@linux.com>
> > > 
> > > 
> > > ---
> > > mm/slub.c |    4 +++-
> > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > > 
> > > Index: linux-2.6/mm/slub.c
> > > ===================================================================
> > > --- linux-2.6.orig/mm/slub.c	2011-11-22 11:57:21.000000000 -0600
> > > +++ linux-2.6/mm/slub.c	2011-11-22 11:57:55.000000000 -0600
> > > @@ -3027,7 +3027,9 @@ static int kmem_cache_open(struct kmem_c
> > > 	 *    per node list when we run out of per cpu objects. We only fetch
> > > 50%
> > > 	 *    to keep some capacity around for frees.
> > > 	 */
> > > -	if (s->size >= PAGE_SIZE)
> > > +	if (kmem_cache_debug(s))
> > > +		s->cpu_partial = 0;
> > > +	else if (s->size >= PAGE_SIZE)
> > > 		s->cpu_partial = 2;
> > > 	else if (s->size >= 1024)
> > > 		s->cpu_partial = 6;
> 
> I think this deserves a comment since it's not clear why it's being done 
> neither in the code or by the changelog.  Also, you'd want a check for 
> kmem_cache_debug() when storing to /sys/kernel/slab/cache/cpu_partial as 
> well, otherwise this could easily be broken from userspace again.  (It 
> also seems like Documentation/ABI/testing/sysfs-kernel-slab got left out 
> of an update when that was added, too.)
> 

This patch is merged in slab/urgent without any changes from my comment 
above; since userspace can alter cpu_partial directly via 
/sys/kernel/slab/cache/cpu_partial which would reintroduce the 
"accounting issues", shouldn't that return -EINVAL when debug is enabled?

Re: Switch per cpu partial page support off for debugging

[Pekka Enberg]

On Fri, 2012-01-06 at 09:37 -0800, David Rientjes wrote:
> On Wed, 14 Dec 2011, David Rientjes wrote:
> 
> > > > Eric saw an issue with accounting of slabs during validation. Its not
> > > > possible to determine accurately how many per cpu partial slabs exist at
> > > > any time so this switches off per cpu partial pages during debug.
> > > > 
> > > > Subject: Switch per cpu partial page support off for debugging
> > > > 
> > > > Otherwise we have accounting issues.
> > > > 
> > > > Signed-off-by: Christoph Lameter <cl@linux.com>
> > > > 
> > > > 
> > > > ---
> > > > mm/slub.c |    4 +++-
> > > > 1 file changed, 3 insertions(+), 1 deletion(-)
> > > > 
> > > > Index: linux-2.6/mm/slub.c
> > > > ===================================================================
> > > > --- linux-2.6.orig/mm/slub.c	2011-11-22 11:57:21.000000000 -0600
> > > > +++ linux-2.6/mm/slub.c	2011-11-22 11:57:55.000000000 -0600
> > > > @@ -3027,7 +3027,9 @@ static int kmem_cache_open(struct kmem_c
> > > > 	 *    per node list when we run out of per cpu objects. We only fetch
> > > > 50%
> > > > 	 *    to keep some capacity around for frees.
> > > > 	 */
> > > > -	if (s->size >= PAGE_SIZE)
> > > > +	if (kmem_cache_debug(s))
> > > > +		s->cpu_partial = 0;
> > > > +	else if (s->size >= PAGE_SIZE)
> > > > 		s->cpu_partial = 2;
> > > > 	else if (s->size >= 1024)
> > > > 		s->cpu_partial = 6;
> > 
> > I think this deserves a comment since it's not clear why it's being done 
> > neither in the code or by the changelog.  Also, you'd want a check for 
> > kmem_cache_debug() when storing to /sys/kernel/slab/cache/cpu_partial as 
> > well, otherwise this could easily be broken from userspace again.  (It 
> > also seems like Documentation/ABI/testing/sysfs-kernel-slab got left out 
> > of an update when that was added, too.)
> > 
> 
> This patch is merged in slab/urgent without any changes from my comment 
> above; since userspace can alter cpu_partial directly via 
> /sys/kernel/slab/cache/cpu_partial which would reintroduce the 
> "accounting issues", shouldn't that return -EINVAL when debug is enabled?

I think I applied the patch before your comments. EINVAL sounds
reasonable to me. Christoph?

			Pekka

Re: Switch per cpu partial page support off for debugging

[Christoph Lameter]

On Sun, 8 Jan 2012, Pekka Enberg wrote:

> I think I applied the patch before your comments. EINVAL sounds
> reasonable to me. Christoph?

Sounds all fine to me. Hope I get around to it this week.

[patch slab/urgent] disallow changing cpu_partial from userspace for debug caches

[David Rientjes]

For caches with debugging enabled, "slub: Switch per cpu partial page 
support off for debugging" changes cpu_partial to 0.  It shouldn't be 
tunable from userspace for such caches, otherwise the same accounting 
issues arise during validation.

This patch disallows tuning /sys/kernel/slab/cache/cpu_partial to be non-
zero for caches with debugging enabled.

Signed-off-by: David Rientjes <rientjes@google.com>
---
 mm/slub.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

diff --git a/mm/slub.c b/mm/slub.c
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -4639,6 +4639,8 @@ static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
 	err = strict_strtoul(buf, 10, &objects);
 	if (err)
 		return err;
+	if (objects && kmem_cache_debug(s))
+		return -EINVAL;
 
 	s->cpu_partial = objects;
 	flush_all(s);

Re: [patch slab/urgent] disallow changing cpu_partial from userspace for debug caches

[Christoph Lameter]

On Mon, 9 Jan 2012, David Rientjes wrote:

> This patch disallows tuning /sys/kernel/slab/cache/cpu_partial to be non-
> zero for caches with debugging enabled.

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