zs_destroy_pool()->destroy_handle_cache() invoked from
zs_create_pool() can pass a NULL ->handle_cachep pointer
to kmem_cache_destroy(), which will dereference it.
Signed-off-by: Sergey Senozhatsky <[email protected]>
---
mm/zsmalloc.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/mm/zsmalloc.c b/mm/zsmalloc.c
index 33d5126..c766240 100644
--- a/mm/zsmalloc.c
+++ b/mm/zsmalloc.c
@@ -285,7 +285,8 @@ static int create_handle_cache(struct zs_pool *pool)
static void destroy_handle_cache(struct zs_pool *pool)
{
- kmem_cache_destroy(pool->handle_cachep);
+ if (pool->handle_cachep)
+ kmem_cache_destroy(pool->handle_cachep);
}
static unsigned long alloc_handle(struct zs_pool *pool)
--
2.4.2.387.gf86f31a
On Fri, 5 Jun 2015 20:11:30 +0900 Sergey Senozhatsky <[email protected]> wrote:
> zs_destroy_pool()->destroy_handle_cache() invoked from
> zs_create_pool() can pass a NULL ->handle_cachep pointer
> to kmem_cache_destroy(), which will dereference it.
>
That's slightly lacking in details (under what circumstances will it
crash) so I changed it to
: If zs_create_pool()->create_handle_cache()->kmem_cache_create() fails,
: zs_create_pool()->destroy_handle_cache() will dereference the NULL
: pool->handle_cachep.
:
: Modify destroy_handle_cache() to avoid this.
> ...
>
> --- a/mm/zsmalloc.c
> +++ b/mm/zsmalloc.c
> @@ -285,7 +285,8 @@ static int create_handle_cache(struct zs_pool *pool)
>
> static void destroy_handle_cache(struct zs_pool *pool)
> {
> - kmem_cache_destroy(pool->handle_cachep);
> + if (pool->handle_cachep)
> + kmem_cache_destroy(pool->handle_cachep);
> }
>
> static unsigned long alloc_handle(struct zs_pool *pool)
I'll apply this, but... from a bit of grepping I'm estimating that we
have approximately 200 instances of
if (foo)
kmem_cache_destroy(foo);
so obviously kmem_cache_destroy() should be doing the check.
On Mon, Jun 08, 2015 at 01:55:32PM -0700, Andrew Morton wrote:
> On Fri, 5 Jun 2015 20:11:30 +0900 Sergey Senozhatsky <[email protected]> wrote:
>
> > zs_destroy_pool()->destroy_handle_cache() invoked from
> > zs_create_pool() can pass a NULL ->handle_cachep pointer
> > to kmem_cache_destroy(), which will dereference it.
> >
>
> That's slightly lacking in details (under what circumstances will it
> crash) so I changed it to
>
> : If zs_create_pool()->create_handle_cache()->kmem_cache_create() fails,
> : zs_create_pool()->destroy_handle_cache() will dereference the NULL
> : pool->handle_cachep.
> :
> : Modify destroy_handle_cache() to avoid this.
>
>
> > ...
> >
> > --- a/mm/zsmalloc.c
> > +++ b/mm/zsmalloc.c
> > @@ -285,7 +285,8 @@ static int create_handle_cache(struct zs_pool *pool)
> >
> > static void destroy_handle_cache(struct zs_pool *pool)
> > {
> > - kmem_cache_destroy(pool->handle_cachep);
> > + if (pool->handle_cachep)
> > + kmem_cache_destroy(pool->handle_cachep);
> > }
> >
> > static unsigned long alloc_handle(struct zs_pool *pool)
>
> I'll apply this, but... from a bit of grepping I'm estimating that we
> have approximately 200 instances of
>
> if (foo)
> kmem_cache_destroy(foo);
>
> so obviously kmem_cache_destroy() should be doing the check.
Hello, Andrew.
I'm not sure if doing the check in kmem_cache_destroy() is better.
My quick grep for other pool based allocators(ex. mempool, zpool) also
says that they don't check whether passed pool pointer is NULL or not
in destroy function. I think that it's general convention that proper
pool pointer should be passed to pool based function APIs.
Thanks.
On Tue, 9 Jun 2015 09:38:27 +0900 Joonsoo Kim <[email protected]> wrote:
> > > ...
> > >
> > > --- a/mm/zsmalloc.c
> > > +++ b/mm/zsmalloc.c
> > > @@ -285,7 +285,8 @@ static int create_handle_cache(struct zs_pool *pool)
> > >
> > > static void destroy_handle_cache(struct zs_pool *pool)
> > > {
> > > - kmem_cache_destroy(pool->handle_cachep);
> > > + if (pool->handle_cachep)
> > > + kmem_cache_destroy(pool->handle_cachep);
> > > }
> > >
> > > static unsigned long alloc_handle(struct zs_pool *pool)
> >
> > I'll apply this, but... from a bit of grepping I'm estimating that we
> > have approximately 200 instances of
> >
> > if (foo)
> > kmem_cache_destroy(foo);
> >
> > so obviously kmem_cache_destroy() should be doing the check.
>
> Hello, Andrew.
>
> I'm not sure if doing the check in kmem_cache_destroy() is better.
Of course it's better - we have *hundreds* of sites doing something
which could be done at a single site. Where's the advantage in that?
> My quick grep for other pool based allocators(ex. mempool, zpool) also
> says that they don't check whether passed pool pointer is NULL or not
> in destroy function.
Maybe some of those should be converted as well.
On (06/08/15 13:55), Andrew Morton wrote:
[..]
> > zs_destroy_pool()->destroy_handle_cache() invoked from
> > zs_create_pool() can pass a NULL ->handle_cachep pointer
> > to kmem_cache_destroy(), which will dereference it.
> >
>
> That's slightly lacking in details (under what circumstances will it
> crash) so I changed it to
>
> : If zs_create_pool()->create_handle_cache()->kmem_cache_create() fails,
> : zs_create_pool()->destroy_handle_cache() will dereference the NULL
> : pool->handle_cachep.
> :
> : Modify destroy_handle_cache() to avoid this.
>
Oh, sorry I first received "+ zsmalloc-fix-a-null-pointer-dereference-in-
destroy_handle_cache.patch added to -mm tree" message, so I replied
there. fetchmail works somewhat confusing over the last weeks.
> > ...
> >
> > --- a/mm/zsmalloc.c
> > +++ b/mm/zsmalloc.c
> > @@ -285,7 +285,8 @@ static int create_handle_cache(struct zs_pool *pool)
> >
> > static void destroy_handle_cache(struct zs_pool *pool)
> > {
> > - kmem_cache_destroy(pool->handle_cachep);
> > + if (pool->handle_cachep)
> > + kmem_cache_destroy(pool->handle_cachep);
> > }
> >
> > static unsigned long alloc_handle(struct zs_pool *pool)
>
> I'll apply this, but... from a bit of grepping I'm estimating that we
> have approximately 200 instances of
>
> if (foo)
> kmem_cache_destroy(foo);
>
> so obviously kmem_cache_destroy() should be doing the check.
Yes, I thought about this.
A naive grepping gave me 563 occurrences
git grep kmem_cache_destroy | wc -l
563
So I decided to hold this activity. Well, I think I can create this
patch bomb, it's trivial.
-ss
On Fri, Jun 05, 2015 at 08:11:30PM +0900, Sergey Senozhatsky wrote:
> zs_destroy_pool()->destroy_handle_cache() invoked from
> zs_create_pool() can pass a NULL ->handle_cachep pointer
> to kmem_cache_destroy(), which will dereference it.
>
> Signed-off-by: Sergey Senozhatsky <[email protected]>
Thanks, Sergey!