Re: kernel BUG at fs/pipe.c:LINE!

[Eric Biggers <ebiggers@kernel.org>]

On Wed, Dec 04, 2019 at 09:40:23PM -0800, Eric Biggers wrote:
> David,
> 
> On Sun, Dec 01, 2019 at 10:45:08PM -0800, syzbot wrote:
> > Hello,
> > 
> > syzbot found the following crash on:
> > 
> > HEAD commit:    b94ae8ad Merge tag 'seccomp-v5.5-rc1' of git://git.kernel...
> > git tree:       upstream
> > console output: https://syzkaller.appspot.com/x/log.txt?x=1387ab12e00000
> > kernel config:  https://syzkaller.appspot.com/x/.config?x=ff560c3de405258c
> > dashboard link: https://syzkaller.appspot.com/bug?extid=d37abaade33a934f16f2
> > compiler:       gcc (GCC) 9.0.0 20181231 (experimental)
> > syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=12945c41e00000
> > C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=161e202ee00000
> > 
> > The bug was bisected to:
> > 
> > commit 8cefc107ca54c8b06438b7dc9cc08bc0a11d5b98
> > Author: David Howells <dhowells@redhat.com>
> > Date:   Fri Nov 15 13:30:32 2019 +0000
> > 
> >     pipe: Use head and tail pointers for the ring, not cursor and length
> > 
> > bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=118cce96e00000
> > final crash:    https://syzkaller.appspot.com/x/report.txt?x=138cce96e00000
> > console output: https://syzkaller.appspot.com/x/log.txt?x=158cce96e00000
> > 
> > IMPORTANT: if you fix the bug, please add the following tag to the commit:
> > Reported-by: syzbot+d37abaade33a934f16f2@syzkaller.appspotmail.com
> > Fixes: 8cefc107ca54 ("pipe: Use head and tail pointers for the ring, not
> > cursor and length")
> > 
> > ------------[ cut here ]------------
> > kernel BUG at fs/pipe.c:582!
> 
> This same BUG_ON() crashed my system during normal use, no syzkaller involved at
> all, on mainline 937d6eefc7.  Can you please take a look?  This syzbot report
> has a reproducer so that might be the easiest place to start.
> 
> - Eric

Code is:

static __poll_t
pipe_poll(struct file *filp, poll_table *wait)
{
        __poll_t mask;
        struct pipe_inode_info *pipe = filp->private_data;
        unsigned int head = READ_ONCE(pipe->head);
        unsigned int tail = READ_ONCE(pipe->tail);

        poll_wait(filp, &pipe->wait, wait);

        BUG_ON(pipe_occupancy(head, tail) > pipe->ring_size);

It's not holding the pipe mutex, right?  So 'head', 'tail' and 'ring_size' can
all be changed concurrently, and they aren't read atomically with respect to
each other.

How do you propose to implement poll() correctly with the new head + tail
approach?  Just take the mutex?

- Eric