Re: KCSAN: data-race in __io_cqring_fill_event / io_uring_poll

[Dmitry Vyukov]

On Mon, Aug 10, 2020 at 11:58 AM syzbot
<syzbot+3020eb77f81ef0772fbe@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote:
>
> Hello,
>
> syzbot found the following issue on:
>
> HEAD commit:    86cfccb6 Merge tag 'dlm-5.9' of git://git.kernel.org/pub/s..
> git tree:       upstream
> console output: https://syzkaller.appspot.com/x/log.txt?x=171cf11a900000
> kernel config:  https://syzkaller.appspot.com/x/.config?x=3757fa75ecfde8e0
> dashboard link: https://syzkaller.appspot.com/bug?extid=3020eb77f81ef0772fbe
> compiler:       clang version 11.0.0 (https://github.com/llvm/llvm-project.git ca2dcbd030eadbf0aa9b660efe864ff08af6e18b)
>
> Unfortunately, I don't have any reproducer for this issue yet.
>
> IMPORTANT: if you fix the issue, please add the following tag to the commit:
> Reported-by: syzbot+3020eb77f81ef0772fbe@xxxxxxxxxxxxxxxxxxxxxxxxx
>
> ==================================================================
> BUG: KCSAN: data-race in __io_cqring_fill_event / io_uring_poll
>
> write to 0xffff8880a04325c0 of 4 bytes by task 12088 on cpu 1:
>  io_get_cqring fs/io_uring.c:1282 [inline]
>  __io_cqring_fill_event+0x116/0x430 fs/io_uring.c:1386
>  io_cqring_add_event fs/io_uring.c:1420 [inline]
>  __io_req_complete+0xdb/0x1b0 fs/io_uring.c:1458
>  io_complete_rw_common fs/io_uring.c:2208 [inline]
>  __io_complete_rw+0x2c9/0x2e0 fs/io_uring.c:2289
>  kiocb_done fs/io_uring.c:2533 [inline]
>  io_write fs/io_uring.c:3199 [inline]
>  io_issue_sqe+0x4fb1/0x7140 fs/io_uring.c:5530
>  io_wq_submit_work+0x23e/0x340 fs/io_uring.c:5775
>  io_worker_handle_work+0xa69/0xcf0 fs/io-wq.c:527
>  io_wqe_worker+0x1f2/0x860 fs/io-wq.c:569
>  kthread+0x20d/0x230 kernel/kthread.c:292
>  ret_from_fork+0x1f/0x30 arch/x86/entry/entry_64.S:294
>
> read to 0xffff8880a04325c0 of 4 bytes by task 12086 on cpu 0:
>  io_cqring_events fs/io_uring.c:1937 [inline]
>  io_uring_poll+0x105/0x140 fs/io_uring.c:7751
>  vfs_poll include/linux/poll.h:90 [inline]
>  __io_arm_poll_handler+0x176/0x3f0 fs/io_uring.c:4735
>  io_arm_poll_handler+0x293/0x5c0 fs/io_uring.c:4792
>  __io_queue_sqe+0x413/0x760 fs/io_uring.c:5988
>  io_queue_sqe+0x81/0x2b0 fs/io_uring.c:6060
>  io_submit_sqe+0x333/0x560 fs/io_uring.c:6130
>  io_submit_sqes+0x8c6/0xfc0 fs/io_uring.c:6327
>  __do_sys_io_uring_enter fs/io_uring.c:8036 [inline]
>  __se_sys_io_uring_enter+0x1c2/0x720 fs/io_uring.c:7995
>  __x64_sys_io_uring_enter+0x74/0x80 fs/io_uring.c:7995
>  do_syscall_64+0x39/0x80 arch/x86/entry/common.c:46
>  entry_SYSCALL_64_after_hwframe+0x44/0xa9
>
> Reported by Kernel Concurrency Sanitizer on:
> CPU: 0 PID: 12086 Comm: syz-executor.5 Not tainted 5.8.0-syzkaller #0
> Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
> ==================================================================
>
> ---
> This report is generated by a bot. It may contain errors.
> See https://goo.gl/tpsmEJ for more information about syzbot.
> syzbot engineers can be reached at syzkaller@xxxxxxxxxxxxxxxx.
>
> syzbot will keep track of this issue. See:
> https://goo.gl/tpsmEJ#status for how to communicate with syzbot.

Hi Jens,

I wonder if this read of cached_cq_tail needs READ_ONCE as all
concurrent access seems to be marked in io_uring.c overall.

I also have a concern for a potential bad scenario around
cached_cq_tail. I've seen it in other queues. I don't see that it can
happen in io_uring, but from a very brief look at the code io_uring
may be susceptible, so I thought it's better to ask.
The failure is around ordering of reads of head/tail when deciding if
the queue is empty or not.
Consider: initially queue is empty head==tail==0.
One thread adds an element and sets tail=1 and wakes poller.
Now, before the poller wakes, several threads push/pop from the queue
incrementing both head and tail, maybe also waking poller. But at no
point in time the queue is empty.
Now poller wakes and reads, say, tail=1 (at the time when head=0 and
tail=1 so the queue is not empty) and then reads head=1 (at the time
when head=1 and tail=3 so the queue is not empty again). But now the
poller looks at head==tail==1 and decides that the queue is empty.
This may lead to a missed epoll event and a deadlock.
This can be resolved by acquiring a consistent snapshot of head/tail in poller.
Is there anything that prevents such a scenario in io_uring? Or it's
not an issue for some other reason?