Re: sound: use-after-free in snd_timer_interrupt

[Takashi Iwai]

On Wed, 13 Jan 2016 21:48:47 +0100,
Dmitry Vyukov wrote:
> 
> On Wed, Jan 13, 2016 at 9:30 PM, Takashi Iwai <tiwai@xxxxxxx> wrote:
> > On Wed, 13 Jan 2016 20:41:30 +0100,
> > Dmitry Vyukov wrote:
> >>
> >> On Wed, Jan 13, 2016 at 8:30 PM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> >> > On Wed, Jan 13, 2016 at 8:05 PM, Takashi Iwai <tiwai@xxxxxxx> wrote:
> >> >> On Wed, 13 Jan 2016 19:34:36 +0100,
> >> >> Dmitry Vyukov wrote:
> >> >>>
> >> >>> On Wed, Jan 13, 2016 at 5:53 PM, Takashi Iwai <tiwai@xxxxxxx> wrote:
> >> >>> > This and your other relevant reports seem pointing the race of timer
> >> >>> > ioctls.  Although snd_timer_close() itself calls snd_timer_stop(),
> >> >>> > there is no other protection against the concurrent execution.
> >> >>> >
> >> >>> > If my guess is correct, a simplistic fix like below should work.  It
> >> >>> > basically serializes the timer ioctl by using a new mutex (and
> >> >>> > replacing the old tread_sem mutex).  They are no longtime blocking
> >> >>> > calls, so this shouldn't be a big problem.  But certainly there can be
> >> >>> > a less intrusive way to paper over this if this really matters.
> >> >>> >
> >> >>> > In this case for timer.c, I'd leave the final decision rather to
> >> >>> > Jaroslav.  Jaroslav, what do you think?
> >> >>>
> >> >>>
> >> >>> After applying this patch I still see the following WARNINGS:
> >> >>>
> >> >>> ------------[ cut here ]------------
> >> >>> WARNING: CPU: 2 PID: 30398 at lib/list_debug.c:53 __list_del_entry+0x10b/0x1e0()
> >> >>> list_del corruption, ffff880032d933b0->next is LIST_POISON1 (dead000000000100)
> >> >>> Modules linked in:
> >> >>> CPU: 2 PID: 30398 Comm: syz-executor Not tainted 4.4.0+ #241
> >> >>> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS Bochs 01/01/2011
> >> >>>  00000000ffffffff ffff8800627778d8 ffffffff82926eed ffff880062777948
> >> >>>  ffff880061c2af80 ffffffff8660b640 ffff880062777918 ffffffff81350c89
> >> >>>  ffffffff8298e77b ffffed000c4eef25 ffffffff8660b640 0000000000000035
> >> >>> Call Trace:
> >> >>>  [<     inline     >] __dump_stack lib/dump_stack.c:15
> >> >>>  [<ffffffff82926eed>] dump_stack+0x6f/0xa2 lib/dump_stack.c:50
> >> >>>  [<ffffffff81350c89>] warn_slowpath_common+0xd9/0x140 kernel/panic.c:483
> >> >>>  [<ffffffff81350d99>] warn_slowpath_fmt+0xa9/0xd0 kernel/panic.c:495
> >> >>>  [<ffffffff8298e77b>] __list_del_entry+0x10b/0x1e0 lib/list_debug.c:51
> >> >>>  [<     inline     >] list_del_init include/linux/list.h:145
> >> >>>  [<ffffffff84ebd199>] _snd_timer_stop+0x119/0x450 sound/core/timer.c:501
> >> >>
> >> >> This is
> >> >>
> >> >>         list_del_init(&timeri->active_list);
> >> >>
> >> >> right?  Possibly the following oneliner covers it?
> >> >
> >> > Yes, that is this line.
> >> > Yes, these two patches fix use-after-frees and GPFs.
> >> >
> >> > Tested-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
> >>
> >>
> >> I've re-tested the programs that I reported. But when I started the
> >> fuzzer again I hit a similar use-after-free in snd_timer_interrupt:
> >
> > Is it the result with all patches, i.e. four patches (two for
> > sequencer and two for timer)?
> 
> Yes, with 4 recent patches.

OK, then this might be a possible race at the current snd_timer_stop()
implementation.  There is no sync action there, so the ISR might be
still alive after snd_timer_close() call.  Or might be another race.
This pattern looks a bit different, as it's involved with hrtimer.

I'll take a look at it tomorrow.


thanks,

Takashi