[PATCH] mm/slab.c: add node spinlock protect in __cache_free_alien

[PATCH] mm/slab.c: add node spinlock protect in __cache_free_alien

[qiang.zhang]

From: Zhang Qiang <qiang.zhang@windriver.com>

We should add node spinlock protect "n->alien" which may be
assigned to NULL in cpuup_canceled func. cause address access
exception.

Fixes: 18bf854117c6 ("slab: use get_node() and kmem_cache_node() functions")
Signed-off-by: Zhang Qiang <qiang.zhang@windriver.com>
---
 mm/slab.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/mm/slab.c b/mm/slab.c
index a89633603b2d..290523c90b4e 100644
--- a/mm/slab.c
+++ b/mm/slab.c
@@ -759,8 +759,10 @@ static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 
 	n = get_node(cachep, node);
 	STATS_INC_NODEFREES(cachep);
+	spin_lock(&n->list_lock);
 	if (n->alien && n->alien[page_node]) {
 		alien = n->alien[page_node];
+		spin_unlock(&n->list_lock);
 		ac = &alien->ac;
 		spin_lock(&alien->lock);
 		if (unlikely(ac->avail == ac->limit)) {
@@ -769,14 +771,15 @@ static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
 		}
 		ac->entry[ac->avail++] = objp;
 		spin_unlock(&alien->lock);
-		slabs_destroy(cachep, &list);
 	} else {
+		spin_unlock(&n->list_lock);
 		n = get_node(cachep, page_node);
 		spin_lock(&n->list_lock);
 		free_block(cachep, &objp, 1, page_node, &list);
 		spin_unlock(&n->list_lock);
-		slabs_destroy(cachep, &list);
 	}
+
+	slabs_destroy(cachep, &list);
 	return 1;
 }
 
-- 
2.26.2

Re: [PATCH] mm/slab.c: add node spinlock protect in __cache_free_alien

[David Rientjes]

On Tue, 28 Jul 2020, qiang.zhang@windriver.com wrote:

> From: Zhang Qiang <qiang.zhang@windriver.com>
> 
> We should add node spinlock protect "n->alien" which may be
> assigned to NULL in cpuup_canceled func. cause address access
> exception.
> 

Hi, do you have an example NULL pointer dereference where you have hit 
this?

This rather looks like something to fix up in cpuup_canceled() since it's 
currently manipulating the alien cache for the canceled cpu's node.

> Fixes: 18bf854117c6 ("slab: use get_node() and kmem_cache_node() functions")
> Signed-off-by: Zhang Qiang <qiang.zhang@windriver.com>
> ---
>  mm/slab.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/slab.c b/mm/slab.c
> index a89633603b2d..290523c90b4e 100644
> --- a/mm/slab.c
> +++ b/mm/slab.c
> @@ -759,8 +759,10 @@ static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
>  
>  	n = get_node(cachep, node);
>  	STATS_INC_NODEFREES(cachep);
> +	spin_lock(&n->list_lock);
>  	if (n->alien && n->alien[page_node]) {
>  		alien = n->alien[page_node];
> +		spin_unlock(&n->list_lock);
>  		ac = &alien->ac;
>  		spin_lock(&alien->lock);
>  		if (unlikely(ac->avail == ac->limit)) {
> @@ -769,14 +771,15 @@ static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
>  		}
>  		ac->entry[ac->avail++] = objp;
>  		spin_unlock(&alien->lock);
> -		slabs_destroy(cachep, &list);
>  	} else {
> +		spin_unlock(&n->list_lock);
>  		n = get_node(cachep, page_node);
>  		spin_lock(&n->list_lock);
>  		free_block(cachep, &objp, 1, page_node, &list);
>  		spin_unlock(&n->list_lock);
> -		slabs_destroy(cachep, &list);
>  	}
> +
> +	slabs_destroy(cachep, &list);
>  	return 1;
>  }
>  
> -- 
> 2.26.2
> 
> 

回复: [PATCH] mm/slab.c: add node spinlock protect in __cache_free_alien

[Zhang, Qiang]

________________________________________
发件人: David Rientjes <rientjes@google.com>
发送时间: 2020年7月29日 3:46
收件人: Zhang, Qiang
抄送: cl@linux.com; penberg@kernel.org; iamjoonsoo.kim@lge.com; akpm@linux-foundation.org; linux-mm@kvack.org; linux-kernel@vger.kernel.org
主题: Re: [PATCH] mm/slab.c: add node spinlock protect in __cache_free_alien

On Tue, 28 Jul 2020, qiang.zhang@windriver.com wrote:

> From: Zhang Qiang <qiang.zhang@windriver.com>
>
> We should add node spinlock protect "n->alien" which may be
> assigned to NULL in cpuup_canceled func. cause address access
> exception.
>

>Hi, do you have an example NULL pointer dereference where you have hit
>this?

>This rather looks like something to fix up in cpuup_canceled() since it's
>currently manipulating the alien cache for the canceled cpu's node.

yes , it is fix up in cpuup_canceled  it's
currently manipulating the alien cache for the canceled cpu's node which  may be the same as the node being operated on in the __cache_free_alien func.

void cpuup_canceled
{
        n = get_node(cachep, node);
        spin_lock_irq(&n->list_lock);
        ...
        n->alien = NULL;
        spin_unlock_irq(&n->list_lock);
         ....
}

> Fixes: 18bf854117c6 ("slab: use get_node() and kmem_cache_node() functions")
> Signed-off-by: Zhang Qiang <qiang.zhang@windriver.com>
> ---
>  mm/slab.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
>
> diff --git a/mm/slab.c b/mm/slab.c
> index a89633603b2d..290523c90b4e 100644
> --- a/mm/slab.c
> +++ b/mm/slab.c
> @@ -759,8 +759,10 @@ static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
>
>       n = get_node(cachep, node);
>       STATS_INC_NODEFREES(cachep);
> +     spin_lock(&n->list_lock);
>       if (n->alien && n->alien[page_node]) {
>               alien = n->alien[page_node];
> +             spin_unlock(&n->list_lock);
>               ac = &alien->ac;
>               spin_lock(&alien->lock);
>               if (unlikely(ac->avail == ac->limit)) {
> @@ -769,14 +771,15 @@ static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
>               }
>               ac->entry[ac->avail++] = objp;
>               spin_unlock(&alien->lock);
> -             slabs_destroy(cachep, &list);
>       } else {
> +             spin_unlock(&n->list_lock);
>               n = get_node(cachep, page_node);
>               spin_lock(&n->list_lock);
>               free_block(cachep, &objp, 1, page_node, &list);
>               spin_unlock(&n->list_lock);
> -             slabs_destroy(cachep, &list);
>       }
> +
> +     slabs_destroy(cachep, &list);
>       return 1;
>  }
>
> --
> 2.26.2
>
>

Re: 回复: [PATCH] mm/slab.c: add node spinlock protect in __cache_free_alien

[David Rientjes]

On Wed, 29 Jul 2020, Zhang, Qiang wrote:

> > From: Zhang Qiang <qiang.zhang@windriver.com>
> >
> > We should add node spinlock protect "n->alien" which may be
> > assigned to NULL in cpuup_canceled func. cause address access
> > exception.
> >
> 
> >Hi, do you have an example NULL pointer dereference where you have hit
> >this?
> 

If you have a NULL pointer dereference or a GPF that occurred because of 
this, it would be helpful to provide as rationale.

> >This rather looks like something to fix up in cpuup_canceled() since it's
> >currently manipulating the alien cache for the canceled cpu's node.
> 
> yes , it is fix up in cpuup_canceled  it's
> currently manipulating the alien cache for the canceled cpu's node which  may be the same as the node being operated on in the __cache_free_alien func.
> 
> void cpuup_canceled
> {
>         n = get_node(cachep, node);
>         spin_lock_irq(&n->list_lock);
>         ...
>         n->alien = NULL;
>         spin_unlock_irq(&n->list_lock);
>          ....
> }
> 

Right, so the idea is that this should be fixed in cpuup_canceled() 
instead -- why would we invaliate the entire node's alien cache because a 
single cpu failed to come online?