Re: 2.6.32-rc6 BUG at mm/slab.c:2869!

[Eric Paris <eparis@redhat.com>]

On Fri, 2009-08-21 at 10:28 +0200, Vegard Nossum wrote:
> 2009/8/21 Vegard Nossum <vegard.nossum@gmail.com>:
> > 2009/8/21 Bob Copeland <me@bobcopeland.com>:
> >> On Thu, Aug 20, 2009 at 02:02:49PM +0200, Vegard Nossum wrote:
> >>> > I'll try that and kmemcheck next.
> >>>
> >>> Hm, I'm afraid kmemcheck gives some known false positives related to
> >>> bitfields in ext4 code, so in the case that something turned up, it
> >>> might be hard to distinguish it from those false positives.
> >>
> >> Well I didn't get anything from ext4 so far.  I did hit one with
> >> fsnotify:
> >>
> >> WARNING: kmemcheck: Caught 32-bit read from freed memory (f34a443c)
> >> eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee008a06f700011000
> >>  a a a a a a a a a a a a a a a a a a a a a a a a f f f f f f f f
> >>                                                         ^
> >>
> >> Pid: 2745, comm: fsck.ext4 Not tainted (2.6.31-rc6 #2) MacBook1,1
> >> EIP: 0060:[<c10f3656>] EFLAGS: 00010217 CPU: 0
> >> EIP is at inotify_handle_event+0x76/0xc0
> >> EAX: f34a443c EBX: f34a4438 ECX: 00000000 EDX: f6732000
> >> ESI: f6559764 EDI: 00000000 EBP: f6733f0c ESP: c1527450
> >>  DS: 007b ES: 007b FS: 00d8 GS: 0033 SS: 0068
> >> CR0: 8005003b CR2: f6c046d4 CR3: 367fb000 CR4: 000026d0
> >> DR0: 00000000 DR1: 00000000 DR2: 00000000 DR3: 00000000
> >> DR6: ffff4ff0 DR7: 00000400
> >>  [<c10f0d78>] fsnotify+0xa8/0x130
> >>  [<c10c5e11>] __fput+0xb1/0x1e0
> >>  [<c10c5f55>] fput+0x15/0x20
> >>  [<c10c2ca7>] filp_close+0x47/0x80
> >>  [<c10c2d54>] sys_close+0x74/0xc0
> >>  [<c1002ec8>] sysenter_do_call+0x12/0x36
> >>  [<ffffffff>] 0xffffffff
> >>
> >> I think that is list_empty() here where %eax is list_head
> >> and event_list->next is the read location... which definitely
> >> doesn't look like a pointer, if I'm reading it correctly.
> >
> > I think f34a443c is a valid pointer. On my machine, at least:
> >
> > [    0.004000]     lowmem  : 0xc0000000 - 0xf73fe000   ( 883 MB)
> >
> >>
> >> inotify_fsnotify.o:
> >>
> >>        /* did event_priv get attached? */
> >>        if (list_empty(&fsn_event_priv->event_list))
> >>  143:   8d 43 04                lea    0x4(%ebx),%eax
> >>
> >>        event_priv = kmem_cache_alloc(event_priv_cachep, GFP_KERNEL);
> >>        if (unlikely(!event_priv))
> >>                return -ENOMEM;
> >>
> >>        fsn_event_priv = &event_priv->fsnotify_event_priv_data;
> >>  146:   39 43 04                cmp    %eax,0x4(%ebx)     <=== read here
> >>  149:   74 1d                   je     168 <inotify_handle_event+0x98>
> >
> > I can see somewhat of a race, I think:
> >
> > 1. userspace calls inotify_read(), where we wait for something to happen:
> >
> > 249         while (1) {
> > 250                 prepare_to_wait(&group->notification_waitq, &wait,
> > TASK_INTERRUPTIBLE);
> > 251
> > 252                 mutex_lock(&group->notification_mutex);
> > 253                 kevent = get_one_event(group, count);
> > 254                 mutex_unlock(&group->notification_mutex);
> >
> > 2. an event occurs, and inotify_handle_event() calls
> > fsnotify_add_notify_event():
> >
> >  64         ret = fsnotify_add_notify_event(group, event, fsn_event_priv);
> >  65         /* EEXIST is not an error */
> >  66         if (ret == -EEXIST)
> >  67                 ret = 0;
> >
> > 3. fsnotify_add_notify_event() adds the fsn_event_priv to the event,
> > and adds the event to the group, and finally wakes up anybody who is
> > waiting on &group->notification_waitq:
> >
> > 230         fsnotify_get_event(event);
> > 231         list_add_tail(&holder->event_list, list);
> > 232         if (priv)
> > 233                 list_add_tail(&priv->event_list, &event->private_data_list);
> > 234         spin_unlock(&event->lock);
> > 235         mutex_unlock(&group->notification_mutex);
> > 236
> > 237         wake_up(&group->notification_waitq);
> >
> > 4. inotify_read() wakes up and frees the event:
> >
> > 253                 kevent = get_one_event(group, count);
> >
> > 5. inotify_handle_event() now dereferences the event_priv pointer,
> > which was already freed:
> >
> >  69         /* did event_priv get attached? */
> >  70         if (list_empty(&fsn_event_priv->event_list))
> >
> >
> > I think that's it. Any thoughts? I put Eric Paris on Cc.
> 
> I guess it was fixed by this recently posted patch:
> 
> http://osdir.com/ml/linux-kernel/2009-08/msg05185.html
> 
> Was kmemcheck by any chance used to discover this race in the first place? ;-)

No, it was found by Linus' stellar eye.  I haven't tried kmemcheck since
my last report that I couldn't get a vmware server machine to boot with
kmemcheck=1

http://osdir.com/ml/linux-kernel/2009-08/msg03797.html

I'll give it another shot today.

-Eric