On 2020-05-10 19:11, Ming Lei wrote:
> One simple solution is to pass BLK_MQ_REQ_PREEMPT to blk_get_request()
> called in blk_mq_resubmit_rq() because at that time freezing wait won't
> return and it is safe to allocate a new request for completing old
> requests originated from inactive hctx.
I don't think that will help. Freezing a request queue starts with a
call of this function:
void blk_freeze_queue_start(struct request_queue *q)
{
mutex_lock(&q->mq_freeze_lock);
if (++q->mq_freeze_depth == 1) {
percpu_ref_kill(&q->q_usage_counter);
mutex_unlock(&q->mq_freeze_lock);
if (queue_is_mq(q))
blk_mq_run_hw_queues(q, false);
} else {
mutex_unlock(&q->mq_freeze_lock);
}
}
>From blk_queue_enter():
const bool pm = flags & BLK_MQ_REQ_PREEMPT;
[ ... ]
if (percpu_ref_tryget_live(&q->q_usage_counter)) {
/*
* The code that increments the pm_only counter is
* responsible for ensuring that that counter is
* globally visible before the queue is unfrozen.
*/
if (pm || !blk_queue_pm_only(q)) {
success = true;
} else {
percpu_ref_put(&q->q_usage_counter);
}
}
In other words, setting the BLK_MQ_REQ_PREEMPT flag only makes a
difference if blk_queue_pm_only(q) == true. Freezing a request queue
involves calling percpu_ref_kill(&q->q_usage_counter). That causes all
future percpu_ref_tryget_live() calls to return false until the queue
has been unfrozen.
Bart.
