[PATCH] slub: Do not assert not having lock in removing freed partial

[PATCH] slub: Do not assert not having lock in removing freed partial

[Steven Rostedt]

Vladimir reported the following issue:

Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires
remove_partial() to be called with n->list_lock held, but free_partial()
called from kmem_cache_close() on cache destruction does not follow this
rule, leading to a warning:

  WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0()
  Modules linked in:
  CPU: 0 PID: 2787 Comm: modprobe Tainted: G        W    3.14.0-rc1-mm1+ #1
  Hardware name:
   0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600
   0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000
   ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0
  Call Trace:
   [<ffffffff816d9583>] dump_stack+0x51/0x6e
   [<ffffffff8107c107>] warn_slowpath_common+0x87/0xb0
   [<ffffffff8107c145>] warn_slowpath_null+0x15/0x20
   [<ffffffff811c7fe2>] __kmem_cache_shutdown+0x1b2/0x1f0
   [<ffffffff811908d3>] kmem_cache_destroy+0x43/0xf0
   [<ffffffffa013a123>] xfs_destroy_zones+0x103/0x110 [xfs]
   [<ffffffffa0192b54>] exit_xfs_fs+0x38/0x4e4 [xfs]
   [<ffffffff811036fa>] SyS_delete_module+0x19a/0x1f0
   [<ffffffff816dfcd8>] ? retint_swapgs+0x13/0x1b
   [<ffffffff810d2125>] ? trace_hardirqs_on_caller+0x105/0x1d0
   [<ffffffff81359efe>] ? trace_hardirqs_on_thunk+0x3a/0x3f
   [<ffffffff816e8539>] system_call_fastpath+0x16/0x1b


His solution was to add a spinlock in order to quiet lockdep. Although
there would be no contention to adding the lock, that lock also
requires disabling of interrupts which will have a larger impact on the
system.

Instead of adding a spinlock to a location where it is not needed for
lockdep, make a remove_freed_partial() function that does not test if
the list_lock is held, as no one should have it due to it being freed.

Reported-by: Vladimir Davydov <vdavydov@parallels.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

Index: linux-trace.git/mm/slub.c
===================================================================
--- linux-trace.git.orig/mm/slub.c
+++ linux-trace.git/mm/slub.c
@@ -1530,13 +1530,30 @@ static inline void add_partial(struct km
 		list_add(&page->lru, &n->partial);
 }
 
+static __always_inline void
+__remove_partial(struct kmem_cache_node *n, struct page *page)
+{
+	list_del(&page->lru);
+	n->nr_partial--;
+}
+
 static inline void remove_partial(struct kmem_cache_node *n,
 					struct page *page)
 {
 	lockdep_assert_held(&n->list_lock);
+	__remove_partial(n, page);
+}
 
-	list_del(&page->lru);
-	n->nr_partial--;
+/*
+ * The difference between remove_partial and remove_freed_partial
+ * is that remove_freed_partial happens only on a a freed slab
+ * that should not have anyone accessing it, and thus does not
+ * require the n->list_lock.
+ */
+static inline void remove_freed_partial(struct kmem_cache_node *n,
+					struct page *page)
+{
+	__remove_partial(n, page);
 }
 
 /*
@@ -3195,7 +3212,7 @@ static void free_partial(struct kmem_cac
 
 	list_for_each_entry_safe(page, h, &n->partial, lru) {
 		if (!page->inuse) {
-			remove_partial(n, page);
+			remove_freed_partial(n, page);
 			discard_slab(s, page);
 		} else {
 			list_slab_objects(s, page,

Re: [PATCH] slub: Do not assert not having lock in removing freed partial

[David Rientjes]

On Wed, 5 Feb 2014, Steven Rostedt wrote:

> Vladimir reported the following issue:
> 
> Commit c65c1877bd68 ("slub: use lockdep_assert_held") requires
> remove_partial() to be called with n->list_lock held, but free_partial()
> called from kmem_cache_close() on cache destruction does not follow this
> rule, leading to a warning:
> 
>   WARNING: CPU: 0 PID: 2787 at mm/slub.c:1536 __kmem_cache_shutdown+0x1b2/0x1f0()
>   Modules linked in:
>   CPU: 0 PID: 2787 Comm: modprobe Tainted: G        W    3.14.0-rc1-mm1+ #1
>   Hardware name:
>    0000000000000600 ffff88003ae1dde8 ffffffff816d9583 0000000000000600
>    0000000000000000 ffff88003ae1de28 ffffffff8107c107 0000000000000000
>    ffff880037ab2b00 ffff88007c240d30 ffffea0001ee5280 ffffea0001ee52a0
>   Call Trace:
>    [<ffffffff816d9583>] dump_stack+0x51/0x6e
>    [<ffffffff8107c107>] warn_slowpath_common+0x87/0xb0
>    [<ffffffff8107c145>] warn_slowpath_null+0x15/0x20
>    [<ffffffff811c7fe2>] __kmem_cache_shutdown+0x1b2/0x1f0
>    [<ffffffff811908d3>] kmem_cache_destroy+0x43/0xf0
>    [<ffffffffa013a123>] xfs_destroy_zones+0x103/0x110 [xfs]
>    [<ffffffffa0192b54>] exit_xfs_fs+0x38/0x4e4 [xfs]
>    [<ffffffff811036fa>] SyS_delete_module+0x19a/0x1f0
>    [<ffffffff816dfcd8>] ? retint_swapgs+0x13/0x1b
>    [<ffffffff810d2125>] ? trace_hardirqs_on_caller+0x105/0x1d0
>    [<ffffffff81359efe>] ? trace_hardirqs_on_thunk+0x3a/0x3f
>    [<ffffffff816e8539>] system_call_fastpath+0x16/0x1b
> 
> 
> His solution was to add a spinlock in order to quiet lockdep. Although
> there would be no contention to adding the lock, that lock also
> requires disabling of interrupts which will have a larger impact on the
> system.
> 
> Instead of adding a spinlock to a location where it is not needed for
> lockdep, make a remove_freed_partial() function that does not test if
> the list_lock is held, as no one should have it due to it being freed.
> 
> Reported-by: Vladimir Davydov <vdavydov@parallels.com>
> Signed-off-by: Steven Rostedt <rostedt@goodmis.org>
> 
> Index: linux-trace.git/mm/slub.c
> ===================================================================
> --- linux-trace.git.orig/mm/slub.c
> +++ linux-trace.git/mm/slub.c
> @@ -1530,13 +1530,30 @@ static inline void add_partial(struct km
>  		list_add(&page->lru, &n->partial);
>  }
>  
> +static __always_inline void
> +__remove_partial(struct kmem_cache_node *n, struct page *page)
> +{
> +	list_del(&page->lru);
> +	n->nr_partial--;
> +}
> +
>  static inline void remove_partial(struct kmem_cache_node *n,
>  					struct page *page)
>  {
>  	lockdep_assert_held(&n->list_lock);
> +	__remove_partial(n, page);
> +}
>  
> -	list_del(&page->lru);
> -	n->nr_partial--;
> +/*
> + * The difference between remove_partial and remove_freed_partial
> + * is that remove_freed_partial happens only on a a freed slab

Duplicate "a" there.

> + * that should not have anyone accessing it, and thus does not
> + * require the n->list_lock.
> + */
> +static inline void remove_freed_partial(struct kmem_cache_node *n,
> +					struct page *page)
> +{
> +	__remove_partial(n, page);
>  }
>  
>  /*
> @@ -3195,7 +3212,7 @@ static void free_partial(struct kmem_cac
>  
>  	list_for_each_entry_safe(page, h, &n->partial, lru) {
>  		if (!page->inuse) {
> -			remove_partial(n, page);
> +			remove_freed_partial(n, page);
>  			discard_slab(s, page);
>  		} else {
>  			list_slab_objects(s, page,

We'll want to do something similiar for the add_partial() called from
early_kmem_cache_node_alloc(), right?  It had the added n->list_lock for 
the same reason and is done during early init where nobody else can be 
referencing a kmem_cache_node.

It would probably be better to define these in terms of "partial slabs 
that cannot have anyone else accessing it" rather than "freed slabs".

Re: [PATCH] slub: Do not assert not having lock in removing freed partial

[Steven Rostedt]

On Wed, 5 Feb 2014 16:46:43 -0800 (PST)
David Rientjes <rientjes@google.com> wrote:


> > +/*
> > + * The difference between remove_partial and remove_freed_partial
> > + * is that remove_freed_partial happens only on a a freed slab
> 
> Duplicate "a" there.

oops.

> 
> > + * that should not have anyone accessing it, and thus does not
> > + * require the n->list_lock.
> > + */
> > +static inline void remove_freed_partial(struct kmem_cache_node *n,
> > +					struct page *page)
> > +{
> > +	__remove_partial(n, page);
> >  }
> >  
> >  /*
> > @@ -3195,7 +3212,7 @@ static void free_partial(struct kmem_cac
> >  
> >  	list_for_each_entry_safe(page, h, &n->partial, lru) {
> >  		if (!page->inuse) {
> > -			remove_partial(n, page);
> > +			remove_freed_partial(n, page);
> >  			discard_slab(s, page);
> >  		} else {
> >  			list_slab_objects(s, page,
> 
> We'll want to do something similiar for the add_partial() called from
> early_kmem_cache_node_alloc(), right?  It had the added n->list_lock for 
> the same reason and is done during early init where nobody else can be 
> referencing a kmem_cache_node.
> 
> It would probably be better to define these in terms of "partial slabs 
> that cannot have anyone else accessing it" rather than "freed slabs".

Perhaps then we just use the __remove_partial() and __add_partial()
that does not do the checks. That's common practice to use a "__" to
denote that it's special and usually doesn't require locking.

-- Steve

Re: [PATCH] slub: Do not assert not having lock in removing freed partial

[David Rientjes]

On Wed, 5 Feb 2014, Steven Rostedt wrote:

> > We'll want to do something similiar for the add_partial() called from
> > early_kmem_cache_node_alloc(), right?  It had the added n->list_lock for 
> > the same reason and is done during early init where nobody else can be 
> > referencing a kmem_cache_node.
> > 
> > It would probably be better to define these in terms of "partial slabs 
> > that cannot have anyone else accessing it" rather than "freed slabs".
> 
> Perhaps then we just use the __remove_partial() and __add_partial()
> that does not do the checks. That's common practice to use a "__" to
> denote that it's special and usually doesn't require locking.
> 

Sounds appropriate.  Andrew released a mmotm today so I'm assuming it will 
be in linux-next tomorrow with my {add,remove}_full() patch but it 
shouldn't matter if you're going to be playing with 
{add,remove}_partial().