On Thu, 2017-06-01 at 10:21 +0800, Ming Lei wrote: > On Wed, May 31, 2017 at 03:37:30PM +0000, Bart Van Assche wrote: > > On Wed, 2017-05-31 at 20:37 +0800, Ming Lei wrote: > > > > > > + /* wait until queue is unquiesced */ > > > + wait_event_cmd(q->quiesce_wq, !blk_queue_quiesced(q), > > > + may_sleep ? > > > + srcu_read_unlock(&hctx->queue_rq_srcu, *srcu_idx) : > > > + rcu_read_unlock(), > > > + may_sleep ? > > > + *srcu_idx = srcu_read_lock(&hctx->queue_rq_srcu) : > > > + rcu_read_lock()); > > > + > > > if (q->elevator) > > > goto insert; > > > > What I see is that in this patch a new waitqueue has been introduced > > (quiesce_wq) and also that an explanation of why you think this new waitqueue > > is needed is missing completely. Why is it that you think that the > > synchronize_scru() and synchronize_rcu() calls in blk_mq_quiesce_queue() are > > not sufficient? If this new waitqueue is not needed, please remove that > > waitqueue again. > > OK, the reason is simple, and it is only related with direct issue. > > Under this situation, when the queue is quiesced, we have to > > - insert the current request into sw queue(scheduler queue) > OR > -wait until queue becomes unquiesced like what this patch is doing > > The disadvantage of the 1st way is that we have to consider to run queue > again in blk_mq_unquiesce_queue() for the queued requests during quiescing. > > For the 2nd way(what this patch is doing), one benefit is that application > can avoid to submit I/O to a quiesced queue. Another benefit is that we > needn't to consider to run queue in blk_mq_unquiesce_queue(). But with cost > of one waitqueue, the cost should be cheap, but if you persist on the 1st > approach, I am fine to change to that. Hello Ming, Since the runtime overhead of the alternative approach (insert into queue) is significantly smaller and since it will result in a simpler implementation I prefer the alternative approach. Thanks, Bart.
