io_uring: add a memory barrier before atomic_read
authorZhengyuan Liu <liuzhengyuan@kylinos.cn>
Thu, 18 Jul 2019 12:44:00 +0000 (20:44 +0800)
committerJens Axboe <axboe@kernel.dk>
Thu, 18 Jul 2019 17:09:09 +0000 (11:09 -0600)
commitc0e48f9dea9129aa11bec3ed13803bcc26e96e49
tree0953d43a651c2bf057b97cfef00ad604c18a5588
parentac38297f7038cd5b80d66f8809c7bbf5b70031f3
io_uring: add a memory barrier before atomic_read

There is a hang issue while using fio to do some basic test. The issue
can be easily reproduced using the below script:

        while true
        do
                fio  --ioengine=io_uring  -rw=write -bs=4k -numjobs=1 \
                     -size=1G -iodepth=64 -name=uring   --filename=/dev/zero
        done

After several minutes (or more), fio would block at
io_uring_enter->io_cqring_wait in order to waiting for previously
committed sqes to be completed and can't return to user anymore until
we send a SIGTERM to fio. After receiving SIGTERM, fio hangs at
io_ring_ctx_wait_and_kill with a backtrace like this:

        [54133.243816] Call Trace:
        [54133.243842]  __schedule+0x3a0/0x790
        [54133.243868]  schedule+0x38/0xa0
        [54133.243880]  schedule_timeout+0x218/0x3b0
        [54133.243891]  ? sched_clock+0x9/0x10
        [54133.243903]  ? wait_for_completion+0xa3/0x130
        [54133.243916]  ? _raw_spin_unlock_irq+0x2c/0x40
        [54133.243930]  ? trace_hardirqs_on+0x3f/0xe0
        [54133.243951]  wait_for_completion+0xab/0x130
        [54133.243962]  ? wake_up_q+0x70/0x70
        [54133.243984]  io_ring_ctx_wait_and_kill+0xa0/0x1d0
        [54133.243998]  io_uring_release+0x20/0x30
        [54133.244008]  __fput+0xcf/0x270
        [54133.244029]  ____fput+0xe/0x10
        [54133.244040]  task_work_run+0x7f/0xa0
        [54133.244056]  do_exit+0x305/0xc40
        [54133.244067]  ? get_signal+0x13b/0xbd0
        [54133.244088]  do_group_exit+0x50/0xd0
        [54133.244103]  get_signal+0x18d/0xbd0
        [54133.244112]  ? _raw_spin_unlock_irqrestore+0x36/0x60
        [54133.244142]  do_signal+0x34/0x720
        [54133.244171]  ? exit_to_usermode_loop+0x7e/0x130
        [54133.244190]  exit_to_usermode_loop+0xc0/0x130
        [54133.244209]  do_syscall_64+0x16b/0x1d0
        [54133.244221]  entry_SYSCALL_64_after_hwframe+0x49/0xbe

The reason is that we had added a req to ctx->pending_async at the very
end, but it didn't get a chance to be processed. How could this happen?

        fio#cpu0                                        wq#cpu1

        io_add_to_prev_work                    io_sq_wq_submit_work

          atomic_read() <<< 1

                                                  atomic_dec_return() << 1->0
                                                  list_empty();    <<< true;

          list_add_tail()
          atomic_read() << 0 or 1?

As atomic_ops.rst states, atomic_read does not guarantee that the
runtime modification by any other thread is visible yet, so we must take
care of that with a proper implicit or explicit memory barrier.

This issue was detected with the help of Jackie's <liuyun01@kylinos.cn>

Fixes: 31b515106428 ("io_uring: allow workqueue item to handle multiple buffered requests")
Signed-off-by: Zhengyuan Liu <liuzhengyuan@kylinos.cn>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
fs/io_uring.c