[PATCH 2/2] lib/btree: Fix possible NULL pointer dereference

[PATCH 2/2] lib/btree: Fix possible NULL pointer dereference

[Denis Kirjanov <kirjanov@gmail.com]

mempool_alloc can return null in atomic case.

Signed-off-by: Denis Kirjanov <kirjanov@gmail.com>
---
diff --git a/lib/btree.c b/lib/btree.c
index 41859a8..542c904 100644
--- a/lib/btree.c
+++ b/lib/btree.c
@@ -95,7 +94,8 @@ static unsigned long *btree_node_alloc(struct btree_head *head, gfp_t gfp)
 	unsigned long *node;
 
 	node = mempool_alloc(head->mempool, gfp);
-	memset(node, 0, NODESIZE);
+	if (likely(node))
+		memset(node, 0, NODESIZE);
 	return node;
 }
 

Re: [PATCH 2/2] lib/btree: Fix possible NULL pointer dereference

[Andrew Morton]

On Thu, 13 May 2010 01:20:27 +0400
"Denis Kirjanov <kirjanov@gmail.com" <kirjanov@gmail.com> wrote:

> mempool_alloc can return null in atomic case.
> 
> Signed-off-by: Denis Kirjanov <kirjanov@gmail.com>
> ---
> diff --git a/lib/btree.c b/lib/btree.c
> index 41859a8..542c904 100644
> --- a/lib/btree.c
> +++ b/lib/btree.c
> @@ -95,7 +94,8 @@ static unsigned long *btree_node_alloc(struct btree_head *head, gfp_t gfp)
>  	unsigned long *node;
>  
>  	node = mempool_alloc(head->mempool, gfp);
> -	memset(node, 0, NODESIZE);
> +	if (likely(node))
> +		memset(node, 0, NODESIZE);
>  	return node;
>  }

hm, why is btree.c using mempools?  mempools are only appropriate when
it is known that objects will become available if the allocating task
simply waits for a while.  Typically, things like BIOs and
request-structs.  Simply waiting for the disk to complete some IO will
cause some objects to be returned to the mempool.

If waiting-and-doing-nothing fails to cause objects to be returned to
the pool then the mempool code can lock up.

Re: [PATCH 2/2] lib/btree: Fix possible NULL pointer dereference

[Jörn Engel]

On Thu, 13 May 2010 13:19:07 -0700, Andrew Morton wrote:
> On Thu, 13 May 2010 01:20:27 +0400
> "Denis Kirjanov <kirjanov@gmail.com" <kirjanov@gmail.com> wrote:
> 
> > mempool_alloc can return null in atomic case.
> > 
> > Signed-off-by: Denis Kirjanov <kirjanov@gmail.com>
> > ---
> > diff --git a/lib/btree.c b/lib/btree.c
> > index 41859a8..542c904 100644
> > --- a/lib/btree.c
> > +++ b/lib/btree.c
> > @@ -95,7 +94,8 @@ static unsigned long *btree_node_alloc(struct btree_head *head, gfp_t gfp)
> >  	unsigned long *node;
> >  
> >  	node = mempool_alloc(head->mempool, gfp);
> > -	memset(node, 0, NODESIZE);
> > +	if (likely(node))
> > +		memset(node, 0, NODESIZE);
> >  	return node;
> >  }
> 
> hm, why is btree.c using mempools?  mempools are only appropriate when
> it is known that objects will become available if the allocating task
> simply waits for a while.  Typically, things like BIOs and
> request-structs.  Simply waiting for the disk to complete some IO will
> cause some objects to be returned to the mempool.

For the current caller (logfs), that is a fairly accurate description.

> If waiting-and-doing-nothing fails to cause objects to be returned to
> the pool then the mempool code can lock up.

True.  And I am not 100% sure logfs is bug-free in that respect.  One
item on my todo list is to add some sort of mempool_prefill() that
either ensures pool->curr_nr == pool->min_nr or returns -ENOMEM.  That
would allow logfs start some writeback and wait for the flash, when
necessary.

Jörn

-- 
When in doubt, punt.  When somebody actually complains, go back and fix it...
The 90% solution is a good thing.
-- Rob Landley