Bart Van Assche wrote:

On Wed, 2018-05-16 at 17:16 +0200, Dmitry Vyukov wrote:

quoted

On Wed, May 16, 2018 at 4:56 PM, Bart Van Assche [off-list ref] wrote:

quoted

On Wed, 2018-05-16 at 22:05 +0900, Tetsuo Handa wrote:

quoted

diff --git a/block/blk-core.c b/block/blk-core.c
index 85909b4..59e2496 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c

@@ -951,10 +951,10 @@ int blk_queue_enter(struct request_queue *q, blk_mq_req_flags_t flags)
              smp_rmb();

              wait_event(q->mq_freeze_wq,
-                        (atomic_read(&q->mq_freeze_depth) == 0 &&
-                         (preempt || !blk_queue_preempt_only(q))) ||
+                        atomic_read(&q->mq_freeze_depth) ||
+                        (preempt || !blk_queue_preempt_only(q)) ||
                         blk_queue_dying(q));
-             if (blk_queue_dying(q))
+             if (atomic_read(&q->mq_freeze_depth) || blk_queue_dying(q))
                      return -ENODEV;
      }
 }

That change looks wrong to me.

Hi Bart,

Why does it look wrong to you?

Because that change conflicts with the purpose of queue freezing and also because
that change would inject I/O errors in code paths that shouldn't inject I/O errors.

But waiting there until atomic_read(&q->mq_freeze_depth) becomes 0 is causing deadlock.
wait_event() never returns is a bug. I think we should not wait for q->mq_freeze_depth.