* [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-02-29 3:27 ` Rafael Aquini
0 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-02-29 3:27 UTC (permalink / raw)
To: linux-mm
Cc: linux-kernel, Randy Dunlap, Christoph Lameter, Pekka Enberg,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
allocator to help on debugging OOM conditions. This patch also adds a new
sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
allocation flag and forces the kernel to report every slab allocation failure.
An example print out looks like this:
<snip page allocator out-of-memory message>
SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
cache: bio-0, object size: 192, order: 0
node0: slabs: 3/3, objs: 60/60, free: 0
Signed-off-by: Rafael Aquini <aquini@redhat.com>
---
Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
include/linux/slab.h | 2 +
kernel/sysctl.c | 9 +++++++
mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
4 files changed, 88 insertions(+), 1 deletions(-)
diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
index 96f0ee8..75bdf91 100644
--- a/Documentation/sysctl/vm.txt
+++ b/Documentation/sysctl/vm.txt
@@ -498,6 +498,29 @@ this is causing problems for your system/application.
==============================================================
+oom_dump_slabs_forced
+
+Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
+the system to print warnings about every allocation failure for the
+slab allocator, and helping on debugging certain OOM conditions.
+The print out is pretty similar, and complements data that is reported by
+the page allocator out-of-memory warning:
+
+<snip page allocator out-of-memory message>
+ SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
+ cache: bio-0, object size: 192, order: 0
+ node0: slabs: 3/3, objs: 60/60, free: 0
+
+If this is set to zero, the default behavior is observed and warnings will only
+be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
+
+When set to non-zero, this information is shown whenever the allocator finds
+itself failing to grant a request, regardless the __GFP_NOWARN flag status.
+
+The default value is 0 (disabled).
+
+==============================================================
+
oom_dump_tasks
Enables a system-wide task dump (excluding kernel threads) to be
diff --git a/include/linux/slab.h b/include/linux/slab.h
index 573c809..ca57021 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h
@@ -353,4 +353,6 @@ static inline void *kzalloc_node(size_t size, gfp_t flags, int node)
void __init kmem_cache_init_late(void);
+/* sysctl */
+extern int sysctl_oom_dump_slabs;
#endif /* _LINUX_SLAB_H */
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index f487f25..71fe8ec 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -1040,6 +1040,15 @@ static struct ctl_table vm_table[] = {
.mode = 0644,
.proc_handler = proc_dointvec,
},
+#if defined(CONFIG_SLAB)
+ {
+ .procname = "oom_dump_slabs_forced",
+ .data = &sysctl_oom_dump_slabs,
+ .maxlen = sizeof(sysctl_oom_dump_slabs),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ },
+#endif
{
.procname = "overcommit_ratio",
.data = &sysctl_overcommit_ratio,
diff --git a/mm/slab.c b/mm/slab.c
index f0bd785..993ca4c 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -168,6 +168,9 @@
SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
#endif
+/* sysctl */
+int sysctl_oom_dump_slabs = 0;
+
/*
* kmem_bufctl_t:
*
@@ -1731,6 +1734,52 @@ static int __init cpucache_init(void)
}
__initcall(cpucache_init);
+static noinline void
+slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
+{
+ struct kmem_list3 *l3;
+ struct slab *slabp;
+ unsigned long flags;
+ int node;
+
+ printk(KERN_WARNING
+ "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
+ nodeid, gfpflags);
+ printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
+ cachep->name, cachep->buffer_size, cachep->gfporder);
+
+ for_each_online_node(node) {
+ unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
+ unsigned long active_slabs = 0, num_slabs = 0;
+
+ l3 = cachep->nodelists[node];
+ if (!l3)
+ continue;
+
+ spin_lock_irqsave(&l3->list_lock, flags);
+ list_for_each_entry(slabp, &l3->slabs_full, list) {
+ active_objs += cachep->num;
+ active_slabs++;
+ }
+ list_for_each_entry(slabp, &l3->slabs_partial, list) {
+ active_objs += slabp->inuse;
+ active_slabs++;
+ }
+ list_for_each_entry(slabp, &l3->slabs_free, list)
+ num_slabs++;
+
+ free_objects += l3->free_objects;
+ spin_unlock_irqrestore(&l3->list_lock, flags);
+
+ num_slabs += active_slabs;
+ num_objs = num_slabs * cachep->num;
+ printk(KERN_WARNING
+ " node%d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
+ node, active_slabs, num_slabs, active_objs, num_objs,
+ free_objects);
+ }
+}
+
/*
* Interface to system's page allocator. No need to hold the cache-lock.
*
@@ -1757,8 +1806,12 @@ static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
flags |= __GFP_RECLAIMABLE;
page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
- if (!page)
+ if (!page) {
+ if ((sysctl_oom_dump_slabs || !(flags & __GFP_NOWARN)) &&
+ printk_ratelimit())
+ slab_out_of_memory(cachep, flags, nodeid);
return NULL;
+ }
nr_pages = (1 << cachep->gfporder);
if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
--
1.7.7.6
^ permalink raw reply related [flat|nested] 14+ messages in thread
* [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-02-29 3:27 ` Rafael Aquini
0 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-02-29 3:27 UTC (permalink / raw)
To: linux-mm
Cc: linux-kernel, Randy Dunlap, Christoph Lameter, Pekka Enberg,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
allocator to help on debugging OOM conditions. This patch also adds a new
sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
allocation flag and forces the kernel to report every slab allocation failure.
An example print out looks like this:
<snip page allocator out-of-memory message>
SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
cache: bio-0, object size: 192, order: 0
node0: slabs: 3/3, objs: 60/60, free: 0
Signed-off-by: Rafael Aquini <aquini@redhat.com>
---
Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
include/linux/slab.h | 2 +
kernel/sysctl.c | 9 +++++++
mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
4 files changed, 88 insertions(+), 1 deletions(-)
diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
index 96f0ee8..75bdf91 100644
--- a/Documentation/sysctl/vm.txt
+++ b/Documentation/sysctl/vm.txt
@@ -498,6 +498,29 @@ this is causing problems for your system/application.
==============================================================
+oom_dump_slabs_forced
+
+Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
+the system to print warnings about every allocation failure for the
+slab allocator, and helping on debugging certain OOM conditions.
+The print out is pretty similar, and complements data that is reported by
+the page allocator out-of-memory warning:
+
+<snip page allocator out-of-memory message>
+ SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
+ cache: bio-0, object size: 192, order: 0
+ node0: slabs: 3/3, objs: 60/60, free: 0
+
+If this is set to zero, the default behavior is observed and warnings will only
+be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
+
+When set to non-zero, this information is shown whenever the allocator finds
+itself failing to grant a request, regardless the __GFP_NOWARN flag status.
+
+The default value is 0 (disabled).
+
+==============================================================
+
oom_dump_tasks
Enables a system-wide task dump (excluding kernel threads) to be
diff --git a/include/linux/slab.h b/include/linux/slab.h
index 573c809..ca57021 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h
@@ -353,4 +353,6 @@ static inline void *kzalloc_node(size_t size, gfp_t flags, int node)
void __init kmem_cache_init_late(void);
+/* sysctl */
+extern int sysctl_oom_dump_slabs;
#endif /* _LINUX_SLAB_H */
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index f487f25..71fe8ec 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -1040,6 +1040,15 @@ static struct ctl_table vm_table[] = {
.mode = 0644,
.proc_handler = proc_dointvec,
},
+#if defined(CONFIG_SLAB)
+ {
+ .procname = "oom_dump_slabs_forced",
+ .data = &sysctl_oom_dump_slabs,
+ .maxlen = sizeof(sysctl_oom_dump_slabs),
+ .mode = 0644,
+ .proc_handler = proc_dointvec,
+ },
+#endif
{
.procname = "overcommit_ratio",
.data = &sysctl_overcommit_ratio,
diff --git a/mm/slab.c b/mm/slab.c
index f0bd785..993ca4c 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -168,6 +168,9 @@
SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
#endif
+/* sysctl */
+int sysctl_oom_dump_slabs = 0;
+
/*
* kmem_bufctl_t:
*
@@ -1731,6 +1734,52 @@ static int __init cpucache_init(void)
}
__initcall(cpucache_init);
+static noinline void
+slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
+{
+ struct kmem_list3 *l3;
+ struct slab *slabp;
+ unsigned long flags;
+ int node;
+
+ printk(KERN_WARNING
+ "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
+ nodeid, gfpflags);
+ printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
+ cachep->name, cachep->buffer_size, cachep->gfporder);
+
+ for_each_online_node(node) {
+ unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
+ unsigned long active_slabs = 0, num_slabs = 0;
+
+ l3 = cachep->nodelists[node];
+ if (!l3)
+ continue;
+
+ spin_lock_irqsave(&l3->list_lock, flags);
+ list_for_each_entry(slabp, &l3->slabs_full, list) {
+ active_objs += cachep->num;
+ active_slabs++;
+ }
+ list_for_each_entry(slabp, &l3->slabs_partial, list) {
+ active_objs += slabp->inuse;
+ active_slabs++;
+ }
+ list_for_each_entry(slabp, &l3->slabs_free, list)
+ num_slabs++;
+
+ free_objects += l3->free_objects;
+ spin_unlock_irqrestore(&l3->list_lock, flags);
+
+ num_slabs += active_slabs;
+ num_objs = num_slabs * cachep->num;
+ printk(KERN_WARNING
+ " node%d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
+ node, active_slabs, num_slabs, active_objs, num_objs,
+ free_objects);
+ }
+}
+
/*
* Interface to system's page allocator. No need to hold the cache-lock.
*
@@ -1757,8 +1806,12 @@ static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
flags |= __GFP_RECLAIMABLE;
page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
- if (!page)
+ if (!page) {
+ if ((sysctl_oom_dump_slabs || !(flags & __GFP_NOWARN)) &&
+ printk_ratelimit())
+ slab_out_of_memory(cachep, flags, nodeid);
return NULL;
+ }
nr_pages = (1 << cachep->gfporder);
if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
--
1.7.7.6
--
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>
^ permalink raw reply related [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
2012-02-29 3:27 ` Rafael Aquini
@ 2012-02-29 7:37 ` Pekka Enberg
-1 siblings, 0 replies; 14+ messages in thread
From: Pekka Enberg @ 2012-02-29 7:37 UTC (permalink / raw)
To: Rafael Aquini
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
On Wed, 29 Feb 2012, Rafael Aquini wrote:
> Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> allocator to help on debugging OOM conditions. This patch also adds a new
> sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> allocation flag and forces the kernel to report every slab allocation failure.
>
> An example print out looks like this:
>
> <snip page allocator out-of-memory message>
> SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> cache: bio-0, object size: 192, order: 0
> node0: slabs: 3/3, objs: 60/60, free: 0
>
> Signed-off-by: Rafael Aquini <aquini@redhat.com>
> ---
> Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
> include/linux/slab.h | 2 +
> kernel/sysctl.c | 9 +++++++
> mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
> 4 files changed, 88 insertions(+), 1 deletions(-)
No SLUB support for this?
> diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
> index 96f0ee8..75bdf91 100644
> --- a/Documentation/sysctl/vm.txt
> +++ b/Documentation/sysctl/vm.txt
> @@ -498,6 +498,29 @@ this is causing problems for your system/application.
>
> ==============================================================
>
> +oom_dump_slabs_forced
> +
> +Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
> +the system to print warnings about every allocation failure for the
> +slab allocator, and helping on debugging certain OOM conditions.
> +The print out is pretty similar, and complements data that is reported by
> +the page allocator out-of-memory warning:
> +
> +<snip page allocator out-of-memory message>
> + SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> + cache: bio-0, object size: 192, order: 0
> + node0: slabs: 3/3, objs: 60/60, free: 0
> +
> +If this is set to zero, the default behavior is observed and warnings will only
> +be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
> +
> +When set to non-zero, this information is shown whenever the allocator finds
> +itself failing to grant a request, regardless the __GFP_NOWARN flag status.
> +
> +The default value is 0 (disabled).
> +
> +==============================================================
> +
Why do you want to add a sysctl for this? That'd be an ABI that we need to
keep around forever.
Is there any reason we shouldn't just enable this unconditionally?
Pekka
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-02-29 7:37 ` Pekka Enberg
0 siblings, 0 replies; 14+ messages in thread
From: Pekka Enberg @ 2012-02-29 7:37 UTC (permalink / raw)
To: Rafael Aquini
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
On Wed, 29 Feb 2012, Rafael Aquini wrote:
> Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> allocator to help on debugging OOM conditions. This patch also adds a new
> sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> allocation flag and forces the kernel to report every slab allocation failure.
>
> An example print out looks like this:
>
> <snip page allocator out-of-memory message>
> SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> cache: bio-0, object size: 192, order: 0
> node0: slabs: 3/3, objs: 60/60, free: 0
>
> Signed-off-by: Rafael Aquini <aquini@redhat.com>
> ---
> Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
> include/linux/slab.h | 2 +
> kernel/sysctl.c | 9 +++++++
> mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
> 4 files changed, 88 insertions(+), 1 deletions(-)
No SLUB support for this?
> diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
> index 96f0ee8..75bdf91 100644
> --- a/Documentation/sysctl/vm.txt
> +++ b/Documentation/sysctl/vm.txt
> @@ -498,6 +498,29 @@ this is causing problems for your system/application.
>
> ==============================================================
>
> +oom_dump_slabs_forced
> +
> +Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
> +the system to print warnings about every allocation failure for the
> +slab allocator, and helping on debugging certain OOM conditions.
> +The print out is pretty similar, and complements data that is reported by
> +the page allocator out-of-memory warning:
> +
> +<snip page allocator out-of-memory message>
> + SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> + cache: bio-0, object size: 192, order: 0
> + node0: slabs: 3/3, objs: 60/60, free: 0
> +
> +If this is set to zero, the default behavior is observed and warnings will only
> +be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
> +
> +When set to non-zero, this information is shown whenever the allocator finds
> +itself failing to grant a request, regardless the __GFP_NOWARN flag status.
> +
> +The default value is 0 (disabled).
> +
> +==============================================================
> +
Why do you want to add a sysctl for this? That'd be an ABI that we need to
keep around forever.
Is there any reason we shouldn't just enable this unconditionally?
Pekka
--
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>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
2012-02-29 7:37 ` Pekka Enberg
@ 2012-02-29 18:16 ` Rafael Aquini
-1 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-02-29 18:16 UTC (permalink / raw)
To: Pekka Enberg
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
On Wed, Feb 29, 2012 at 09:37:23AM +0200, Pekka Enberg wrote:
> On Wed, 29 Feb 2012, Rafael Aquini wrote:
> > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > allocator to help on debugging OOM conditions. This patch also adds a new
> > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > allocation flag and forces the kernel to report every slab allocation failure.
> >
> > An example print out looks like this:
> >
> > <snip page allocator out-of-memory message>
> > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > cache: bio-0, object size: 192, order: 0
> > node0: slabs: 3/3, objs: 60/60, free: 0
> >
> > Signed-off-by: Rafael Aquini <aquini@redhat.com>
> > ---
> > Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
> > include/linux/slab.h | 2 +
> > kernel/sysctl.c | 9 +++++++
> > mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
> > 4 files changed, 88 insertions(+), 1 deletions(-)
>
> No SLUB support for this?
>
SLUB already has its version of slab_out_of_memory. I did not propose the sysctl
knob for slub, however. (If we find the knob useful, I can propose its extention
to slub, later).
> > diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
> > index 96f0ee8..75bdf91 100644
> > --- a/Documentation/sysctl/vm.txt
> > +++ b/Documentation/sysctl/vm.txt
> > @@ -498,6 +498,29 @@ this is causing problems for your system/application.
> >
> > ==============================================================
> >
> > +oom_dump_slabs_forced
> > +
> > +Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
> > +the system to print warnings about every allocation failure for the
> > +slab allocator, and helping on debugging certain OOM conditions.
> > +The print out is pretty similar, and complements data that is reported by
> > +the page allocator out-of-memory warning:
> > +
> > +<snip page allocator out-of-memory message>
> > + SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > + cache: bio-0, object size: 192, order: 0
> > + node0: slabs: 3/3, objs: 60/60, free: 0
> > +
> > +If this is set to zero, the default behavior is observed and warnings will only
> > +be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
> > +
> > +When set to non-zero, this information is shown whenever the allocator finds
> > +itself failing to grant a request, regardless the __GFP_NOWARN flag status.
> > +
> > +The default value is 0 (disabled).
> > +
> > +==============================================================
> > +
>
> Why do you want to add a sysctl for this? That'd be an ABI that we need to
> keep around forever.
>
> Is there any reason we shouldn't just enable this unconditionally?
I was afraid of this code becoming a source of garrulous and scaring warnings
by just ignoring __GFP_NOWARN flag, however, I was also concerned with the
'hiding' effect the flag imposes for certain requests, specially when one is
interested in checking all those requests out. Therefore, I thought a sysctl
knob would be the best option to control the __GFP_NOWARN overriding behavior of
slab_out_of_memory printouts without messing with the allocation flags
themselves, as well as not imposing the need for reboots to start checking all
slab allocation failures out.
Rafael
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-02-29 18:16 ` Rafael Aquini
0 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-02-29 18:16 UTC (permalink / raw)
To: Pekka Enberg
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
On Wed, Feb 29, 2012 at 09:37:23AM +0200, Pekka Enberg wrote:
> On Wed, 29 Feb 2012, Rafael Aquini wrote:
> > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > allocator to help on debugging OOM conditions. This patch also adds a new
> > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > allocation flag and forces the kernel to report every slab allocation failure.
> >
> > An example print out looks like this:
> >
> > <snip page allocator out-of-memory message>
> > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > cache: bio-0, object size: 192, order: 0
> > node0: slabs: 3/3, objs: 60/60, free: 0
> >
> > Signed-off-by: Rafael Aquini <aquini@redhat.com>
> > ---
> > Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
> > include/linux/slab.h | 2 +
> > kernel/sysctl.c | 9 +++++++
> > mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
> > 4 files changed, 88 insertions(+), 1 deletions(-)
>
> No SLUB support for this?
>
SLUB already has its version of slab_out_of_memory. I did not propose the sysctl
knob for slub, however. (If we find the knob useful, I can propose its extention
to slub, later).
> > diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
> > index 96f0ee8..75bdf91 100644
> > --- a/Documentation/sysctl/vm.txt
> > +++ b/Documentation/sysctl/vm.txt
> > @@ -498,6 +498,29 @@ this is causing problems for your system/application.
> >
> > ==============================================================
> >
> > +oom_dump_slabs_forced
> > +
> > +Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
> > +the system to print warnings about every allocation failure for the
> > +slab allocator, and helping on debugging certain OOM conditions.
> > +The print out is pretty similar, and complements data that is reported by
> > +the page allocator out-of-memory warning:
> > +
> > +<snip page allocator out-of-memory message>
> > + SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > + cache: bio-0, object size: 192, order: 0
> > + node0: slabs: 3/3, objs: 60/60, free: 0
> > +
> > +If this is set to zero, the default behavior is observed and warnings will only
> > +be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
> > +
> > +When set to non-zero, this information is shown whenever the allocator finds
> > +itself failing to grant a request, regardless the __GFP_NOWARN flag status.
> > +
> > +The default value is 0 (disabled).
> > +
> > +==============================================================
> > +
>
> Why do you want to add a sysctl for this? That'd be an ABI that we need to
> keep around forever.
>
> Is there any reason we shouldn't just enable this unconditionally?
I was afraid of this code becoming a source of garrulous and scaring warnings
by just ignoring __GFP_NOWARN flag, however, I was also concerned with the
'hiding' effect the flag imposes for certain requests, specially when one is
interested in checking all those requests out. Therefore, I thought a sysctl
knob would be the best option to control the __GFP_NOWARN overriding behavior of
slab_out_of_memory printouts without messing with the allocation flags
themselves, as well as not imposing the need for reboots to start checking all
slab allocation failures out.
Rafael
--
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>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
2012-02-29 7:37 ` Pekka Enberg
@ 2012-02-29 18:45 ` Rafael Aquini
-1 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-02-29 18:45 UTC (permalink / raw)
To: Pekka Enberg
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
On Wed, Feb 29, 2012 at 09:37:23AM +0200, Pekka Enberg wrote:
> On Wed, 29 Feb 2012, Rafael Aquini wrote:
> > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > allocator to help on debugging OOM conditions. This patch also adds a new
> > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > allocation flag and forces the kernel to report every slab allocation failure.
> >
> > An example print out looks like this:
> >
> > <snip page allocator out-of-memory message>
> > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > cache: bio-0, object size: 192, order: 0
> > node0: slabs: 3/3, objs: 60/60, free: 0
> >
> > Signed-off-by: Rafael Aquini <aquini@redhat.com>
> > ---
> > Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
> > include/linux/slab.h | 2 +
> > kernel/sysctl.c | 9 +++++++
> > mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
> > 4 files changed, 88 insertions(+), 1 deletions(-)
>
> No SLUB support for this?
SLUB already has its version of slab_out_of_memory. I did not propose the sysctl
knob for slub, however. (If we find the knob useful, I can propose its extention
to slub, later).
>
> > diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
> > index 96f0ee8..75bdf91 100644
> > --- a/Documentation/sysctl/vm.txt
> > +++ b/Documentation/sysctl/vm.txt
> > @@ -498,6 +498,29 @@ this is causing problems for your system/application.
> >
> > ==============================================================
> >
> > +oom_dump_slabs_forced
> > +
> > +Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
> > +the system to print warnings about every allocation failure for the
> > +slab allocator, and helping on debugging certain OOM conditions.
> > +The print out is pretty similar, and complements data that is reported by
> > +the page allocator out-of-memory warning:
> > +
> > +<snip page allocator out-of-memory message>
> > + SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > + cache: bio-0, object size: 192, order: 0
> > + node0: slabs: 3/3, objs: 60/60, free: 0
> > +
> > +If this is set to zero, the default behavior is observed and warnings will only
> > +be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
> > +
> > +When set to non-zero, this information is shown whenever the allocator finds
> > +itself failing to grant a request, regardless the __GFP_NOWARN flag status.
> > +
> > +The default value is 0 (disabled).
> > +
> > +==============================================================
> > +
>
> Why do you want to add a sysctl for this? That'd be an ABI that we need to
> keep around forever.
>
> Is there any reason we shouldn't just enable this unconditionally?
I was afraid of this code becoming a source of garrulous and scaring warnings
by just ignoring __GFP_NOWARN flag, however, I was also concerned with the
'hiding' effect the flag imposes for certain requests, specially when one is
interested in checking all those requests out. Therefore, I thought a sysctl
knob would be the best option to control the __GFP_NOWARN overriding behavior of
slab_out_of_memory printouts without messing with the allocation flags
themselves, as well as not imposing the need for reboots to start checking all
slab allocation failures out.
Rafael
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-02-29 18:45 ` Rafael Aquini
0 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-02-29 18:45 UTC (permalink / raw)
To: Pekka Enberg
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Matt Mackall, Rik van Riel, Josef Bacik, David Rientjes
On Wed, Feb 29, 2012 at 09:37:23AM +0200, Pekka Enberg wrote:
> On Wed, 29 Feb 2012, Rafael Aquini wrote:
> > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > allocator to help on debugging OOM conditions. This patch also adds a new
> > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > allocation flag and forces the kernel to report every slab allocation failure.
> >
> > An example print out looks like this:
> >
> > <snip page allocator out-of-memory message>
> > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > cache: bio-0, object size: 192, order: 0
> > node0: slabs: 3/3, objs: 60/60, free: 0
> >
> > Signed-off-by: Rafael Aquini <aquini@redhat.com>
> > ---
> > Documentation/sysctl/vm.txt | 23 ++++++++++++++++++
> > include/linux/slab.h | 2 +
> > kernel/sysctl.c | 9 +++++++
> > mm/slab.c | 55 ++++++++++++++++++++++++++++++++++++++++++-
> > 4 files changed, 88 insertions(+), 1 deletions(-)
>
> No SLUB support for this?
SLUB already has its version of slab_out_of_memory. I did not propose the sysctl
knob for slub, however. (If we find the knob useful, I can propose its extention
to slub, later).
>
> > diff --git a/Documentation/sysctl/vm.txt b/Documentation/sysctl/vm.txt
> > index 96f0ee8..75bdf91 100644
> > --- a/Documentation/sysctl/vm.txt
> > +++ b/Documentation/sysctl/vm.txt
> > @@ -498,6 +498,29 @@ this is causing problems for your system/application.
> >
> > ==============================================================
> >
> > +oom_dump_slabs_forced
> > +
> > +Overrides the effects of __GFP_NOWARN page allocation flag, thus forcing
> > +the system to print warnings about every allocation failure for the
> > +slab allocator, and helping on debugging certain OOM conditions.
> > +The print out is pretty similar, and complements data that is reported by
> > +the page allocator out-of-memory warning:
> > +
> > +<snip page allocator out-of-memory message>
> > + SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > + cache: bio-0, object size: 192, order: 0
> > + node0: slabs: 3/3, objs: 60/60, free: 0
> > +
> > +If this is set to zero, the default behavior is observed and warnings will only
> > +be printed out for allocation requests that didn't set the __GFP_NOWARN flag.
> > +
> > +When set to non-zero, this information is shown whenever the allocator finds
> > +itself failing to grant a request, regardless the __GFP_NOWARN flag status.
> > +
> > +The default value is 0 (disabled).
> > +
> > +==============================================================
> > +
>
> Why do you want to add a sysctl for this? That'd be an ABI that we need to
> keep around forever.
>
> Is there any reason we shouldn't just enable this unconditionally?
I was afraid of this code becoming a source of garrulous and scaring warnings
by just ignoring __GFP_NOWARN flag, however, I was also concerned with the
'hiding' effect the flag imposes for certain requests, specially when one is
interested in checking all those requests out. Therefore, I thought a sysctl
knob would be the best option to control the __GFP_NOWARN overriding behavior of
slab_out_of_memory printouts without messing with the allocation flags
themselves, as well as not imposing the need for reboots to start checking all
slab allocation failures out.
Rafael
--
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>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
2012-02-29 3:27 ` Rafael Aquini
@ 2012-03-01 1:26 ` David Rientjes
-1 siblings, 0 replies; 14+ messages in thread
From: David Rientjes @ 2012-03-01 1:26 UTC (permalink / raw)
To: Rafael Aquini
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Pekka Enberg, Matt Mackall, Rik van Riel, Josef Bacik
On Wed, 29 Feb 2012, Rafael Aquini wrote:
> Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> allocator to help on debugging OOM conditions. This patch also adds a new
> sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> allocation flag and forces the kernel to report every slab allocation failure.
>
> An example print out looks like this:
>
> <snip page allocator out-of-memory message>
> SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> cache: bio-0, object size: 192, order: 0
> node0: slabs: 3/3, objs: 60/60, free: 0
>
> Signed-off-by: Rafael Aquini <aquini@redhat.com>
I like it, except for the addition of the sysctl. __GFP_NOWARN is used
for a reason, usually because whatever is allocating memory can gracefully
handle a failure and should not be emitted to the kernel log under any
circumstances.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-03-01 1:26 ` David Rientjes
0 siblings, 0 replies; 14+ messages in thread
From: David Rientjes @ 2012-03-01 1:26 UTC (permalink / raw)
To: Rafael Aquini
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Pekka Enberg, Matt Mackall, Rik van Riel, Josef Bacik
On Wed, 29 Feb 2012, Rafael Aquini wrote:
> Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> allocator to help on debugging OOM conditions. This patch also adds a new
> sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> allocation flag and forces the kernel to report every slab allocation failure.
>
> An example print out looks like this:
>
> <snip page allocator out-of-memory message>
> SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> cache: bio-0, object size: 192, order: 0
> node0: slabs: 3/3, objs: 60/60, free: 0
>
> Signed-off-by: Rafael Aquini <aquini@redhat.com>
I like it, except for the addition of the sysctl. __GFP_NOWARN is used
for a reason, usually because whatever is allocating memory can gracefully
handle a failure and should not be emitted to the kernel log under any
circumstances.
--
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>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
2012-03-01 1:26 ` David Rientjes
@ 2012-03-02 14:33 ` Rafael Aquini
-1 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-03-02 14:33 UTC (permalink / raw)
To: David Rientjes
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Pekka Enberg, Matt Mackall, Rik van Riel, Josef Bacik
On Wed, Feb 29, 2012 at 05:26:07PM -0800, David Rientjes wrote:
> On Wed, 29 Feb 2012, Rafael Aquini wrote:
>
> > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > allocator to help on debugging OOM conditions. This patch also adds a new
> > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > allocation flag and forces the kernel to report every slab allocation failure.
> >
> > An example print out looks like this:
> >
> > <snip page allocator out-of-memory message>
> > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > cache: bio-0, object size: 192, order: 0
> > node0: slabs: 3/3, objs: 60/60, free: 0
> >
> > Signed-off-by: Rafael Aquini <aquini@redhat.com>
>
> I like it, except for the addition of the sysctl. __GFP_NOWARN is used
> for a reason, usually because whatever is allocating memory can gracefully
> handle a failure and should not be emitted to the kernel log under any
> circumstances.
Ok, I'll drop the sysctl part then. Pekka?
David, once again, thanks for your feedback!
Rafael
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-03-02 14:33 ` Rafael Aquini
0 siblings, 0 replies; 14+ messages in thread
From: Rafael Aquini @ 2012-03-02 14:33 UTC (permalink / raw)
To: David Rientjes
Cc: linux-mm, linux-kernel, Randy Dunlap, Christoph Lameter,
Pekka Enberg, Matt Mackall, Rik van Riel, Josef Bacik
On Wed, Feb 29, 2012 at 05:26:07PM -0800, David Rientjes wrote:
> On Wed, 29 Feb 2012, Rafael Aquini wrote:
>
> > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > allocator to help on debugging OOM conditions. This patch also adds a new
> > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > allocation flag and forces the kernel to report every slab allocation failure.
> >
> > An example print out looks like this:
> >
> > <snip page allocator out-of-memory message>
> > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > cache: bio-0, object size: 192, order: 0
> > node0: slabs: 3/3, objs: 60/60, free: 0
> >
> > Signed-off-by: Rafael Aquini <aquini@redhat.com>
>
> I like it, except for the addition of the sysctl. __GFP_NOWARN is used
> for a reason, usually because whatever is allocating memory can gracefully
> handle a failure and should not be emitted to the kernel log under any
> circumstances.
Ok, I'll drop the sysctl part then. Pekka?
David, once again, thanks for your feedback!
Rafael
--
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>
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
2012-03-02 14:33 ` Rafael Aquini
@ 2012-03-05 10:33 ` Pekka Enberg
-1 siblings, 0 replies; 14+ messages in thread
From: Pekka Enberg @ 2012-03-05 10:33 UTC (permalink / raw)
To: Rafael Aquini
Cc: David Rientjes, linux-mm, linux-kernel, Randy Dunlap,
Christoph Lameter, Matt Mackall, Rik van Riel, Josef Bacik
On Fri, 2 Mar 2012, Rafael Aquini wrote:
> On Wed, Feb 29, 2012 at 05:26:07PM -0800, David Rientjes wrote:
> > On Wed, 29 Feb 2012, Rafael Aquini wrote:
> >
> > > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > > allocator to help on debugging OOM conditions. This patch also adds a new
> > > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > > allocation flag and forces the kernel to report every slab allocation failure.
> > >
> > > An example print out looks like this:
> > >
> > > <snip page allocator out-of-memory message>
> > > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > > cache: bio-0, object size: 192, order: 0
> > > node0: slabs: 3/3, objs: 60/60, free: 0
> > >
> > > Signed-off-by: Rafael Aquini <aquini@redhat.com>
> >
> > I like it, except for the addition of the sysctl. __GFP_NOWARN is used
> > for a reason, usually because whatever is allocating memory can gracefully
> > handle a failure and should not be emitted to the kernel log under any
> > circumstances.
>
> Ok, I'll drop the sysctl part then. Pekka?
Yes, please.
^ permalink raw reply [flat|nested] 14+ messages in thread
* Re: [PATCH] mm: SLAB Out-of-memory diagnostics
@ 2012-03-05 10:33 ` Pekka Enberg
0 siblings, 0 replies; 14+ messages in thread
From: Pekka Enberg @ 2012-03-05 10:33 UTC (permalink / raw)
To: Rafael Aquini
Cc: David Rientjes, linux-mm, linux-kernel, Randy Dunlap,
Christoph Lameter, Matt Mackall, Rik van Riel, Josef Bacik
On Fri, 2 Mar 2012, Rafael Aquini wrote:
> On Wed, Feb 29, 2012 at 05:26:07PM -0800, David Rientjes wrote:
> > On Wed, 29 Feb 2012, Rafael Aquini wrote:
> >
> > > Following the example at mm/slub.c, add out-of-memory diagnostics to the SLAB
> > > allocator to help on debugging OOM conditions. This patch also adds a new
> > > sysctl, 'oom_dump_slabs_forced', that overrides the effect of __GFP_NOWARN page
> > > allocation flag and forces the kernel to report every slab allocation failure.
> > >
> > > An example print out looks like this:
> > >
> > > <snip page allocator out-of-memory message>
> > > SLAB: Unable to allocate memory on node 0 (gfp=0x11200)
> > > cache: bio-0, object size: 192, order: 0
> > > node0: slabs: 3/3, objs: 60/60, free: 0
> > >
> > > Signed-off-by: Rafael Aquini <aquini@redhat.com>
> >
> > I like it, except for the addition of the sysctl. __GFP_NOWARN is used
> > for a reason, usually because whatever is allocating memory can gracefully
> > handle a failure and should not be emitted to the kernel log under any
> > circumstances.
>
> Ok, I'll drop the sysctl part then. Pekka?
Yes, please.
--
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>
^ permalink raw reply [flat|nested] 14+ messages in thread
end of thread, other threads:[~2012-03-05 10:33 UTC | newest]
Thread overview: 14+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2012-02-29 3:27 [PATCH] mm: SLAB Out-of-memory diagnostics Rafael Aquini
2012-02-29 3:27 ` Rafael Aquini
2012-02-29 7:37 ` Pekka Enberg
2012-02-29 7:37 ` Pekka Enberg
2012-02-29 18:16 ` Rafael Aquini
2012-02-29 18:16 ` Rafael Aquini
2012-02-29 18:45 ` Rafael Aquini
2012-02-29 18:45 ` Rafael Aquini
2012-03-01 1:26 ` David Rientjes
2012-03-01 1:26 ` David Rientjes
2012-03-02 14:33 ` Rafael Aquini
2012-03-02 14:33 ` Rafael Aquini
2012-03-05 10:33 ` Pekka Enberg
2012-03-05 10:33 ` Pekka Enberg
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.