Re: Race inotify_rm_watch vs umount

[Eric Paris <eparis@redhat.com>]

On Fri, 2011-06-10 at 20:20 +0900, OGAWA Hirofumi wrote:
> Hi,
> 
> I'm looking the race inotify_rm_watch() vs umount(). This race become the
> cause of Oops. You can see the oops at
> 
> https://bugzilla.kernel.org/show_bug.cgi?id=22602
> 
> So, what race?

Ok, I see what you are saying, I'll see what I can do.  I'm a little
scared to call something like iput() under a lock though.  I might be
able to make the bigest lock a mutex and fix this....

I'll add this to my test suite.

-Eric
> 
> umount                               inotify_rm_watch
>   ...                                  fsnotify_destroy_mark()
>                                          fsnotify_destroy_inode_mark()
>                                            /* removed from i_fsnotify_marks */
>     generic_shutdown_super()
>       fsnotify_unmount_inodes()
>         put_super()
>                                            iput()
>                                              iput_final()
>                                                /* this is after put_super() */
> 
> Like above, inotify doesn't guarantee to call final iput() before
> put_super(). With this violation, FS driver can oops.
> 
> Well, so, what are requested for inotify?  We can't simply take
> sb->s_umount in inotify_rm_watch()?
> 
> Any ideas?
> 
> Thanks.