Re: INFO: task hung in fanotify_handle_event

[Jan Kara <jack@suse.cz>]

Hi Dmirty!

On Mon 15-10-18 14:29:14, Dmitry Vyukov wrote:
> On Mon, Oct 15, 2018 at 2:15 PM, Jan Kara <jack@suse.cz> wrote:
> > Hello,
> >
> > On Mon 15-10-18 04:32:02, syzbot wrote:
> >> syzbot found the following crash on:
> >>
> >> HEAD commit:    90ad18418c2d Merge git://git.kernel.org/pub/scm/linux/kern..
> >> git tree:       upstream
> >> console output: https://syzkaller.appspot.com/x/log.txt?x=17f1776e400000
> >> kernel config:  https://syzkaller.appspot.com/x/.config?x=88e9a8a39dc0be2d
> >> dashboard link: https://syzkaller.appspot.com/bug?extid=29143581b0ded3213e99
> >> compiler:       gcc (GCC) 8.0.1 20180413 (experimental)
> >> syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=123459d6400000
> >>
> >> IMPORTANT: if you fix the bug, please add the following tag to the commit:
> >> Reported-by: syzbot+29143581b0ded3213e99@syzkaller.appspotmail.com
> >
> > Syzbot has apparently generated fanotify watch for FAN_OPEN_PERM event and
> > then the process got stuck waiting for userspace to respond to that event -
> > which never happened. So everything works as designed here - the process
> > placing FAN_OPEN_PERM mark is responsible for replying to the generated
> > events as all opens hang waiting for responses. That's why the
> > functionality is behind CAP_SYS_ADMIN after all... Could we fix syzbot to
> > actually generate replies for these events?
> 
> Is there a reliable way to kill such processes?
> Or admins are never supposed to kill any root processes and have not
> bugs whatsoever? :)

Currently the wait is not killable but yes, we want to make it killable
exactly because of userspace bugs :). But it is non-trivial because
currently the waker has also other responsibilities and all that stuff has
to be cleaned up when handling killed wait. Konstantin Khlebnikov was
working on that so I might need to prod him.

> syzkaller probably capable of generating replies in some cases, but
> unfortunately it can't work this way. It's practically not possible to
> ensure that it will always generate a proper reply and it will be
> actually delivered and the process won't be killed in the middle, or
> another thread won't crash or call exit_group concurrently, etc. The
> thing either needs to be reliable, work without any but's and be
> reliably killable, or it's not suitable for stress testing.
> If there is no reliable way to kill it, I think we need to disable
> FAN_OPEN_PERM entirely.

Understood. Then just disable FAN_OPEN_PERM & FAN_ACCESS_PERM for now.

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR