On Wed, 2018-05-16 at 22:05 +0900, Tetsuo Handa wrote: > One ore more threads are waiting for q->mq_freeze_depth to become 0. But the > thread who incremented q->mq_freeze_depth at blk_freeze_queue_start(q) from > blk_freeze_queue() is waiting at blk_mq_freeze_queue_wait(). Therefore, > atomic_read(&q->mq_freeze_depth) == 0 condition for wait_event() in > blk_queue_enter() will never be satisfied. But what does that wait_event() > want to do? Isn't "start freezing" a sort of blk_queue_dying(q) == true? > Since percpu_ref_tryget_live(&q->q_usage_counter) failed and the queue is > about to be frozen, shouldn't we treat atomic_read(&q->mq_freeze_depth) != 0 > as if blk_queue_dying(q) == true? That is, something like below: > > 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. Additionally, I think that you are looking in the wrong direction. Since blk_mq_freeze_queue_wait() and blk_queue_enter() work fine for all block drivers except the loop driver I think that you should have a closer look at how the loop driver uses this block layer functionality. Thanks, Bart.
