Re: [PATCH] Fix lockdep false positive in add_full()
On Mon, Feb 03, 2014 at 03:31:40PM -0800, David Rientjes wrote: > On Mon, 3 Feb 2014, Paul E. McKenney wrote: > > > Hello! > > > > The add_full() function currently has a lockdep_assert_held() requiring > > that the kmem_cache_node structure's ->list_lock be held. However, > > this lock is not acquired by add_full()'s caller deactivate_slab() > > in the full-node case unless debugging is enabled. Because full nodes > > are accessed only by debugging code, this state of affairs results in > > lockdep false-positive splats like the following: > > > > [ 43.942868] WARNING: CPU: 0 PID: 698 at > > /home/paulmck/public_git/linux-rcu/mm/slub.c:1007 > > deactivate_slab+0x509/0x720() > > [ 43.943016] Modules linked in: > > [ 43.943016] CPU: 0 PID: 698 Comm: torture_onoff Not tainted 3.14.0-rc1+ > > #1 > > [ 43.943016] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > > [ 43.943016] 03ef 88001e3f5ba8 818952ec > > 0046 > > [ 43.943016] 88001e3f5be8 81049517 > > ea784e00 > > [ 43.943016] ea7a9000 0002 > > > > [ 43.943016] Call Trace: > > [ 43.943016] [] dump_stack+0x46/0x58 > > [ 43.943016] [] warn_slowpath_common+0x87/0xb0 > > [ 43.943016] [] warn_slowpath_null+0x15/0x20 > > [ 43.943016] [] deactivate_slab+0x509/0x720 > > [ 43.943016] [] ? slab_cpuup_callback+0x3b/0x100 > > [ 43.943016] [] ? slab_cpuup_callback+0xd2/0x100 > > [ 43.943016] [] slab_cpuup_callback+0xa4/0x100 > > [ 43.943016] [] notifier_call_chain+0x54/0x110 > > [ 43.943016] [] __raw_notifier_call_chain+0x9/0x10 > > [ 43.943016] [] __cpu_notify+0x1b/0x30 > > [ 43.943016] [] cpu_notify_nofail+0x10/0x20 > > [ 43.943016] [] _cpu_down+0x10d/0x2e0 > > [ 43.943016] [] cpu_down+0x30/0x50 > > [ 43.943016] [] torture_onoff+0xd3/0x3c0 > > [ 43.943016] [] ? torture_onoff_stats+0x90/0x90 > > [ 43.943016] [] kthread+0xdf/0x100 > > [ 43.943016] [] ? _raw_spin_unlock_irq+0x2b/0x40 > > [ 43.943016] [] ? flush_kthread_worker+0x130/0x130 > > [ 43.943016] [] ret_from_fork+0x7c/0xb0 > > [ 43.943016] [] ? flush_kthread_worker+0x130/0x130 > > > > This commit therefore does the lockdep check only if debuggging is > > enabled, thus avoiding the false positives. > > > > Signed-off-by: Paul E. McKenney > > This was discussed in http://marc.info/?t=13914579132, what do you > think about the patch in that thread instead? Looks fine to me! I also tried it out and it avoided the splats, as noted in my mail in the other thread, so please feel free to add my Tested-by. Thanx, Paul -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] Fix lockdep false positive in add_full()
On Mon, 3 Feb 2014, Paul E. McKenney wrote: > Hello! > > The add_full() function currently has a lockdep_assert_held() requiring > that the kmem_cache_node structure's ->list_lock be held. However, > this lock is not acquired by add_full()'s caller deactivate_slab() > in the full-node case unless debugging is enabled. Because full nodes > are accessed only by debugging code, this state of affairs results in > lockdep false-positive splats like the following: > > [ 43.942868] WARNING: CPU: 0 PID: 698 at > /home/paulmck/public_git/linux-rcu/mm/slub.c:1007 > deactivate_slab+0x509/0x720() > [ 43.943016] Modules linked in: > [ 43.943016] CPU: 0 PID: 698 Comm: torture_onoff Not tainted 3.14.0-rc1+ #1 > [ 43.943016] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > [ 43.943016] 03ef 88001e3f5ba8 818952ec > 0046 > [ 43.943016] 88001e3f5be8 81049517 > ea784e00 > [ 43.943016] ea7a9000 0002 > > [ 43.943016] Call Trace: > [ 43.943016] [] dump_stack+0x46/0x58 > [ 43.943016] [] warn_slowpath_common+0x87/0xb0 > [ 43.943016] [] warn_slowpath_null+0x15/0x20 > [ 43.943016] [] deactivate_slab+0x509/0x720 > [ 43.943016] [] ? slab_cpuup_callback+0x3b/0x100 > [ 43.943016] [] ? slab_cpuup_callback+0xd2/0x100 > [ 43.943016] [] slab_cpuup_callback+0xa4/0x100 > [ 43.943016] [] notifier_call_chain+0x54/0x110 > [ 43.943016] [] __raw_notifier_call_chain+0x9/0x10 > [ 43.943016] [] __cpu_notify+0x1b/0x30 > [ 43.943016] [] cpu_notify_nofail+0x10/0x20 > [ 43.943016] [] _cpu_down+0x10d/0x2e0 > [ 43.943016] [] cpu_down+0x30/0x50 > [ 43.943016] [] torture_onoff+0xd3/0x3c0 > [ 43.943016] [] ? torture_onoff_stats+0x90/0x90 > [ 43.943016] [] kthread+0xdf/0x100 > [ 43.943016] [] ? _raw_spin_unlock_irq+0x2b/0x40 > [ 43.943016] [] ? flush_kthread_worker+0x130/0x130 > [ 43.943016] [] ret_from_fork+0x7c/0xb0 > [ 43.943016] [] ? flush_kthread_worker+0x130/0x130 > > This commit therefore does the lockdep check only if debuggging is > enabled, thus avoiding the false positives. > > Signed-off-by: Paul E. McKenney This was discussed in http://marc.info/?t=13914579132, what do you think about the patch in that thread instead? -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[PATCH] Fix lockdep false positive in add_full()
Hello!
The add_full() function currently has a lockdep_assert_held() requiring
that the kmem_cache_node structure's ->list_lock be held. However,
this lock is not acquired by add_full()'s caller deactivate_slab()
in the full-node case unless debugging is enabled. Because full nodes
are accessed only by debugging code, this state of affairs results in
lockdep false-positive splats like the following:
[ 43.942868] WARNING: CPU: 0 PID: 698 at
/home/paulmck/public_git/linux-rcu/mm/slub.c:1007 deactivate_slab+0x509/0x720()
[ 43.943016] Modules linked in:
[ 43.943016] CPU: 0 PID: 698 Comm: torture_onoff Not tainted 3.14.0-rc1+ #1
[ 43.943016] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007
[ 43.943016] 03ef 88001e3f5ba8 818952ec
0046
[ 43.943016] 88001e3f5be8 81049517
ea784e00
[ 43.943016] ea7a9000 0002
[ 43.943016] Call Trace:
[ 43.943016] [] dump_stack+0x46/0x58
[ 43.943016] [] warn_slowpath_common+0x87/0xb0
[ 43.943016] [] warn_slowpath_null+0x15/0x20
[ 43.943016] [] deactivate_slab+0x509/0x720
[ 43.943016] [] ? slab_cpuup_callback+0x3b/0x100
[ 43.943016] [] ? slab_cpuup_callback+0xd2/0x100
[ 43.943016] [] slab_cpuup_callback+0xa4/0x100
[ 43.943016] [] notifier_call_chain+0x54/0x110
[ 43.943016] [] __raw_notifier_call_chain+0x9/0x10
[ 43.943016] [] __cpu_notify+0x1b/0x30
[ 43.943016] [] cpu_notify_nofail+0x10/0x20
[ 43.943016] [] _cpu_down+0x10d/0x2e0
[ 43.943016] [] cpu_down+0x30/0x50
[ 43.943016] [] torture_onoff+0xd3/0x3c0
[ 43.943016] [] ? torture_onoff_stats+0x90/0x90
[ 43.943016] [] kthread+0xdf/0x100
[ 43.943016] [] ? _raw_spin_unlock_irq+0x2b/0x40
[ 43.943016] [] ? flush_kthread_worker+0x130/0x130
[ 43.943016] [] ret_from_fork+0x7c/0xb0
[ 43.943016] [] ? flush_kthread_worker+0x130/0x130
This commit therefore does the lockdep check only if debuggging is
enabled, thus avoiding the false positives.
Signed-off-by: Paul E. McKenney
Cc: linux...@kvack.org
Cc: c...@linux-foundation.org
Cc: penb...@kernel.org
Cc: m...@selenic.com
diff --git a/mm/slub.c b/mm/slub.c
index 7e3e0458bce4..6fff4d980b7c 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1004,7 +1004,8 @@ static inline void slab_free_hook(struct kmem_cache *s,
void *x)
static void add_full(struct kmem_cache *s,
struct kmem_cache_node *n, struct page *page)
{
- lockdep_assert_held(>list_lock);
+ if (kmem_cache_debug(s))
+ lockdep_assert_held(>list_lock);
if (!(s->flags & SLAB_STORE_USER))
return;
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[PATCH] Fix lockdep false positive in add_full()
Hello!
The add_full() function currently has a lockdep_assert_held() requiring
that the kmem_cache_node structure's -list_lock be held. However,
this lock is not acquired by add_full()'s caller deactivate_slab()
in the full-node case unless debugging is enabled. Because full nodes
are accessed only by debugging code, this state of affairs results in
lockdep false-positive splats like the following:
[ 43.942868] WARNING: CPU: 0 PID: 698 at
/home/paulmck/public_git/linux-rcu/mm/slub.c:1007 deactivate_slab+0x509/0x720()
[ 43.943016] Modules linked in:
[ 43.943016] CPU: 0 PID: 698 Comm: torture_onoff Not tainted 3.14.0-rc1+ #1
[ 43.943016] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007
[ 43.943016] 03ef 88001e3f5ba8 818952ec
0046
[ 43.943016] 88001e3f5be8 81049517
ea784e00
[ 43.943016] ea7a9000 0002
[ 43.943016] Call Trace:
[ 43.943016] [818952ec] dump_stack+0x46/0x58
[ 43.943016] [81049517] warn_slowpath_common+0x87/0xb0
[ 43.943016] [81049555] warn_slowpath_null+0x15/0x20
[ 43.943016] [8116e679] deactivate_slab+0x509/0x720
[ 43.943016] [8116eebb] ? slab_cpuup_callback+0x3b/0x100
[ 43.943016] [8116ef52] ? slab_cpuup_callback+0xd2/0x100
[ 43.943016] [8116ef24] slab_cpuup_callback+0xa4/0x100
[ 43.943016] [818a4c14] notifier_call_chain+0x54/0x110
[ 43.943016] [81075b79] __raw_notifier_call_chain+0x9/0x10
[ 43.943016] [8104963b] __cpu_notify+0x1b/0x30
[ 43.943016] [81049720] cpu_notify_nofail+0x10/0x20
[ 43.943016] [8188cc5d] _cpu_down+0x10d/0x2e0
[ 43.943016] [8188ce60] cpu_down+0x30/0x50
[ 43.943016] [811205f3] torture_onoff+0xd3/0x3c0
[ 43.943016] [81120520] ? torture_onoff_stats+0x90/0x90
[ 43.943016] [810710df] kthread+0xdf/0x100
[ 43.943016] [818a09cb] ? _raw_spin_unlock_irq+0x2b/0x40
[ 43.943016] [81071000] ? flush_kthread_worker+0x130/0x130
[ 43.943016] [818a983c] ret_from_fork+0x7c/0xb0
[ 43.943016] [81071000] ? flush_kthread_worker+0x130/0x130
This commit therefore does the lockdep check only if debuggging is
enabled, thus avoiding the false positives.
Signed-off-by: Paul E. McKenney paul...@linux.vnet.ibm.com
Cc: linux...@kvack.org
Cc: c...@linux-foundation.org
Cc: penb...@kernel.org
Cc: m...@selenic.com
diff --git a/mm/slub.c b/mm/slub.c
index 7e3e0458bce4..6fff4d980b7c 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -1004,7 +1004,8 @@ static inline void slab_free_hook(struct kmem_cache *s,
void *x)
static void add_full(struct kmem_cache *s,
struct kmem_cache_node *n, struct page *page)
{
- lockdep_assert_held(n-list_lock);
+ if (kmem_cache_debug(s))
+ lockdep_assert_held(n-list_lock);
if (!(s-flags SLAB_STORE_USER))
return;
--
To unsubscribe from this list: send the line unsubscribe linux-kernel in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] Fix lockdep false positive in add_full()
On Mon, 3 Feb 2014, Paul E. McKenney wrote: Hello! The add_full() function currently has a lockdep_assert_held() requiring that the kmem_cache_node structure's -list_lock be held. However, this lock is not acquired by add_full()'s caller deactivate_slab() in the full-node case unless debugging is enabled. Because full nodes are accessed only by debugging code, this state of affairs results in lockdep false-positive splats like the following: [ 43.942868] WARNING: CPU: 0 PID: 698 at /home/paulmck/public_git/linux-rcu/mm/slub.c:1007 deactivate_slab+0x509/0x720() [ 43.943016] Modules linked in: [ 43.943016] CPU: 0 PID: 698 Comm: torture_onoff Not tainted 3.14.0-rc1+ #1 [ 43.943016] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 [ 43.943016] 03ef 88001e3f5ba8 818952ec 0046 [ 43.943016] 88001e3f5be8 81049517 ea784e00 [ 43.943016] ea7a9000 0002 [ 43.943016] Call Trace: [ 43.943016] [818952ec] dump_stack+0x46/0x58 [ 43.943016] [81049517] warn_slowpath_common+0x87/0xb0 [ 43.943016] [81049555] warn_slowpath_null+0x15/0x20 [ 43.943016] [8116e679] deactivate_slab+0x509/0x720 [ 43.943016] [8116eebb] ? slab_cpuup_callback+0x3b/0x100 [ 43.943016] [8116ef52] ? slab_cpuup_callback+0xd2/0x100 [ 43.943016] [8116ef24] slab_cpuup_callback+0xa4/0x100 [ 43.943016] [818a4c14] notifier_call_chain+0x54/0x110 [ 43.943016] [81075b79] __raw_notifier_call_chain+0x9/0x10 [ 43.943016] [8104963b] __cpu_notify+0x1b/0x30 [ 43.943016] [81049720] cpu_notify_nofail+0x10/0x20 [ 43.943016] [8188cc5d] _cpu_down+0x10d/0x2e0 [ 43.943016] [8188ce60] cpu_down+0x30/0x50 [ 43.943016] [811205f3] torture_onoff+0xd3/0x3c0 [ 43.943016] [81120520] ? torture_onoff_stats+0x90/0x90 [ 43.943016] [810710df] kthread+0xdf/0x100 [ 43.943016] [818a09cb] ? _raw_spin_unlock_irq+0x2b/0x40 [ 43.943016] [81071000] ? flush_kthread_worker+0x130/0x130 [ 43.943016] [818a983c] ret_from_fork+0x7c/0xb0 [ 43.943016] [81071000] ? flush_kthread_worker+0x130/0x130 This commit therefore does the lockdep check only if debuggging is enabled, thus avoiding the false positives. Signed-off-by: Paul E. McKenney paul...@linux.vnet.ibm.com This was discussed in http://marc.info/?t=13914579132, what do you think about the patch in that thread instead? -- To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [PATCH] Fix lockdep false positive in add_full()
On Mon, Feb 03, 2014 at 03:31:40PM -0800, David Rientjes wrote: On Mon, 3 Feb 2014, Paul E. McKenney wrote: Hello! The add_full() function currently has a lockdep_assert_held() requiring that the kmem_cache_node structure's -list_lock be held. However, this lock is not acquired by add_full()'s caller deactivate_slab() in the full-node case unless debugging is enabled. Because full nodes are accessed only by debugging code, this state of affairs results in lockdep false-positive splats like the following: [ 43.942868] WARNING: CPU: 0 PID: 698 at /home/paulmck/public_git/linux-rcu/mm/slub.c:1007 deactivate_slab+0x509/0x720() [ 43.943016] Modules linked in: [ 43.943016] CPU: 0 PID: 698 Comm: torture_onoff Not tainted 3.14.0-rc1+ #1 [ 43.943016] Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 [ 43.943016] 03ef 88001e3f5ba8 818952ec 0046 [ 43.943016] 88001e3f5be8 81049517 ea784e00 [ 43.943016] ea7a9000 0002 [ 43.943016] Call Trace: [ 43.943016] [818952ec] dump_stack+0x46/0x58 [ 43.943016] [81049517] warn_slowpath_common+0x87/0xb0 [ 43.943016] [81049555] warn_slowpath_null+0x15/0x20 [ 43.943016] [8116e679] deactivate_slab+0x509/0x720 [ 43.943016] [8116eebb] ? slab_cpuup_callback+0x3b/0x100 [ 43.943016] [8116ef52] ? slab_cpuup_callback+0xd2/0x100 [ 43.943016] [8116ef24] slab_cpuup_callback+0xa4/0x100 [ 43.943016] [818a4c14] notifier_call_chain+0x54/0x110 [ 43.943016] [81075b79] __raw_notifier_call_chain+0x9/0x10 [ 43.943016] [8104963b] __cpu_notify+0x1b/0x30 [ 43.943016] [81049720] cpu_notify_nofail+0x10/0x20 [ 43.943016] [8188cc5d] _cpu_down+0x10d/0x2e0 [ 43.943016] [8188ce60] cpu_down+0x30/0x50 [ 43.943016] [811205f3] torture_onoff+0xd3/0x3c0 [ 43.943016] [81120520] ? torture_onoff_stats+0x90/0x90 [ 43.943016] [810710df] kthread+0xdf/0x100 [ 43.943016] [818a09cb] ? _raw_spin_unlock_irq+0x2b/0x40 [ 43.943016] [81071000] ? flush_kthread_worker+0x130/0x130 [ 43.943016] [818a983c] ret_from_fork+0x7c/0xb0 [ 43.943016] [81071000] ? flush_kthread_worker+0x130/0x130 This commit therefore does the lockdep check only if debuggging is enabled, thus avoiding the false positives. Signed-off-by: Paul E. McKenney paul...@linux.vnet.ibm.com This was discussed in http://marc.info/?t=13914579132, what do you think about the patch in that thread instead? Looks fine to me! I also tried it out and it avoided the splats, as noted in my mail in the other thread, so please feel free to add my Tested-by. Thanx, Paul -- To unsubscribe from this list: send the line unsubscribe linux-kernel in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
