[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

[dm-devel] Re: [PATCH 02/18] io-controller: Common flat fair queuing code in elevaotor layer



On Fri, May 22 2009, Vivek Goyal wrote:
> On Fri, May 22, 2009 at 02:43:04PM +0800, Gui Jianfeng wrote:
> > Vivek Goyal wrote:
> > ...
> > > +/* A request got completed from io_queue. Do the accounting. */
> > > +void elv_ioq_completed_request(struct request_queue *q, struct request *rq)
> > > +{
> > > +	const int sync = rq_is_sync(rq);
> > > +	struct io_queue *ioq = rq->ioq;
> > > +	struct elv_fq_data *efqd = &q->elevator->efqd;
> > > +
> > > +	if (!elv_iosched_fair_queuing_enabled(q->elevator))
> > > +		return;
> > > +
> > > +	elv_log_ioq(efqd, ioq, "complete");
> > > +
> > > +	elv_update_hw_tag(efqd);
> > > +
> > > +	WARN_ON(!efqd->rq_in_driver);
> > > +	WARN_ON(!ioq->dispatched);
> > > +	efqd->rq_in_driver--;
> > > +	ioq->dispatched--;
> > > +
> > > +	if (sync)
> > > +		ioq->last_end_request = jiffies;
> > > +
> > > +	/*
> > > +	 * If this is the active queue, check if it needs to be expired,
> > > +	 * or if we want to idle in case it has no pending requests.
> > > +	 */
> > > +
> > > +	if (elv_active_ioq(q->elevator) == ioq) {
> > > +		if (elv_ioq_slice_new(ioq)) {
> > > +			elv_ioq_set_prio_slice(q, ioq);
> > 
> >   Hi Vivek,
> > 
> >   Would you explain a bit why slice_end should be set when first request completes.
> >   Why not set it just when an ioq gets active?
> >   
> 
> Hi Gui,
> 
> I have kept the behavior same as CFQ. I guess reason behind this is that
> when a new queue is scheduled in, first request completion might take more
> time as head of the disk might be quite a distance away (due to previous
> queue) and one probably does not want to charge the new queue for that
> first seek time. That's the reason we start the queue slice when first
> request has completed.

That's exactly why CFQ does it that way. And not just for the seek
itself, but if have eg writes issued before the switch to a new queue,
it's not fair to charge the potential cache writeout happening ahead of
the read to that new queue. So I'd definitely recommend keeping this
behaviour, as you have.

-- 
Jens Axboe


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]