Re: [Bugme-new] [Bug 12754] New: inotify doesn't free memory allocated to watches

[Al Viro <viro@ZenIV.linux.org.uk>]

On Tue, Feb 24, 2009 at 03:40:23PM -0800, Andrew Morton wrote:
> ./Documentation/filesystems/inotify.txt
> ./kernel/audit.c
> ./kernel/audit_tree.c
> ./fs/notify/inotify/inotify_user.c
> ./fs/notify/inotify/inotify.c
> ./include/linux/inotify.h
> 
> I assume it's inotify_dev_queue_event()?
> 
>         if (mask & IN_IGNORED || w->mask & IN_ONESHOT)
>                 put_inotify_watch(w); /* final put */

Duh.  Nevermind, I'd misparsed the report.  If we are talking about
inotify_user.c (and not some new client), then we are back to "let's
see what gets leaked"...

Actually, looking at inotify_user.c, we seem to be doing something rather
fishy.  Look: event gets triggered, we pick the watch, get inotify_device
(inotify_user-specific stuff) from it, grap mutex on it (dev->ev_mutex)
and drop reference to inotify_watch.  Which happily triggers ->destroy_watch,
which does put_inotify_dev().  Which is
        if (atomic_dec_and_test(&dev->count)) {
                atomic_dec(&dev->user->inotify_devs);
                free_uid(dev->user);
                kfree(dev);
        }
What's to stop that from happening when we'd been holding the last reference
to that sucker?  kfree() while holding a mutex inside the structure being
freed is not nice...