pids_free double free.

pids_free double free.

[Dave Jones]

One of my debian boxes got a systemd update. After rebooting,
I started seeing a use-after-free trace, followed by a lockup.

I have two slightly different traces from separate boots,
which may give some clue as to how it's getting free'd in two ways..

http://codemonkey.org.uk/junk/IMG_0474.jpg
http://codemonkey.org.uk/junk/IMG_0476.jpg

This isn't new, I booted back to 4.3, and hit slab debugging warnings
from the same code.  (At the least it was broken differently)

The WARN_ON referenced in the 2nd trace is this..

static void pids_cancel(struct pids_cgroup *pids, int num)
{
        /*  
         * A negative count (or overflow for that matter) is invalid,
         * and indicates a bug in the `pids` controller proper.
         */
        WARN_ON_ONCE(atomic64_add_negative(-num, &pids->counter));
}


	Dave

Re: pids_free double free.

[Tejun Heo]

Hello, Dave.

On Thu, Nov 19, 2015 at 11:18:36PM -0500, Dave Jones wrote:
> One of my debian boxes got a systemd update. After rebooting,
> I started seeing a use-after-free trace, followed by a lockup.
> 
> I have two slightly different traces from separate boots,
> which may give some clue as to how it's getting free'd in two ways..
> 
> http://codemonkey.org.uk/junk/IMG_0474.jpg
> http://codemonkey.org.uk/junk/IMG_0476.jpg
> 
> This isn't new, I booted back to 4.3, and hit slab debugging warnings
> from the same code.  (At the least it was broken differently)
> 
> The WARN_ON referenced in the 2nd trace is this..
> 
> static void pids_cancel(struct pids_cgroup *pids, int num)
> {
>         /*  
>          * A negative count (or overflow for that matter) is invalid,
>          * and indicates a bug in the `pids` controller proper.
>          */
>         WARN_ON_ONCE(atomic64_add_negative(-num, &pids->counter));
> }

I *think* I know what's going on.  Zombies don't pin the css so if the
cgroup gets removed, the free path may end up operating on a css which
has already been freed.  Will look more into it.

Thanks.

-- 
tejun