[PATCH v2 1/2] blk-mq: add tagset quiesce interface

Paul E. McKenney paulmck at kernel.org
Tue Oct 18 08:04:06 PDT 2022


On Tue, Oct 18, 2022 at 05:52:06PM +0800, Chao Leng wrote:
> On 2022/10/17 23:21, Paul E. McKenney wrote:
> > On Mon, Oct 17, 2022 at 03:39:06PM +0200, Christoph Hellwig wrote:
> > > On Thu, Oct 13, 2022 at 05:44:49PM +0800, Chao Leng wrote:
> > > > +	rcu = kvmalloc(count * sizeof(*rcu), GFP_KERNEL);
> > > > +	if (rcu) {
> > > > +		list_for_each_entry(q, &set->tag_list, tag_set_list) {
> > > > +			if (blk_queue_noquiesced(q))
> > > > +				continue;
> > > > +
> > > > +			init_rcu_head(&rcu[i].head);
> > > > +			init_completion(&rcu[i].completion);
> > > > +			call_srcu(q->srcu, &rcu[i].head, wakeme_after_rcu);
> > > > +			i++;
> > > > +		}
> > > > +
> > > > +		for (i = 0; i < count; i++) {
> > > > +			wait_for_completion(&rcu[i].completion);
> > > > +			destroy_rcu_head(&rcu[i].head);
> > > > +		}
> > > > +		kvfree(rcu);
> > > > +	} else {
> > > > +		list_for_each_entry(q, &set->tag_list, tag_set_list)
> > > > +			synchronize_srcu(q->srcu);
> > > > +	}
> > > 
> > > Having to allocate a struct rcu_synchronize for each of the potentially
> > > many queues here is a bit sad.
> > > 
> > > Pull just explained the start_poll_synchronize_rcu interfaces at ALPSS
> > > last week, so I wonder if something like that would also be feasible
> > > for SRCU, as that would come in really handy here.
> > 
> > There is start_poll_synchronize_srcu() and poll_state_synchronize_srcu(),
> > but there would need to be an unsigned long for each srcu_struct from
> > which an SRCU grace period was required.  This would be half the size
> > of the "rcu" array above, but still maybe larger than you would like.
> > 
> > The resulting code might look something like this, with "rcu" now being
> > a pointer to unsigned long:
> > 
> > 	rcu = kvmalloc(count * sizeof(*rcu), GFP_KERNEL);
> > 	if (rcu) {
> > 		list_for_each_entry(q, &set->tag_list, tag_set_list) {
> > 			if (blk_queue_noquiesced(q))
> > 				continue;
> > 			rcu[i] = start_poll_synchronize_srcu(q->srcu);
> > 			i++;
> > 		}
> > 
> > 		for (i = 0; i < count; i++)
> > 			if (!poll_state_synchronize_srcu(q->srcu))
> > 				synchronize_srcu(q->srcu);
> synchronize_srcu will restart a new period of grace.

True, but SRCU grace periods normally complete reasonably quickly, so
the synchronize_srcu() might well be faster than the loop, depending on
what the corresponding SRCU readers are doing.

> Maybe it would be better like this:
> 			while (!poll_state_synchronize_srcu(q->srcu, rcu[i]))
> 				schedule_timeout_uninterruptible(1);

Why not try it both ways and see what happens?  Assuming that is, that
the differences matter in this code.

							Thanx, Paul

> > 		kvfree(rcu);
> > 	} else {
> > 		list_for_each_entry(q, &set->tag_list, tag_set_list)
> > 			synchronize_srcu(q->srcu);
> > 	}
> > 
> > Or as Christoph suggested, just have a single srcu_struct for the
> > whole group.
> > 
> > The main reason for having multiple srcu_struct structures is to
> > prevent the readers from one from holding up the updaters from another.
> > Except that by waiting for the multiple grace periods, you are losing
> > that property anyway, correct?  Or is this code waiting on only a small
> > fraction of the srcu_struct structures associated with blk_queue?
> > 
> > 							Thanx, Paul
> > .
> > 



More information about the Linux-nvme mailing list