Re: [PATCH] cfq-prio #2

[Jens Axboe <axboe@suse.de>]

On Tue, Nov 11 2003, Nick Piggin wrote:
> 
> 
> Jens Axboe wrote:
> 
> >On Tue, Nov 11 2003, Nick Piggin wrote:
> >
> >>
> >>Jens Axboe wrote:
> >>
> >>
> >>>Hi,
> >>>
> >>>
> >>>
> >>Hi Jens
> >>
> >>
> >>>@@ -1553,6 +1559,10 @@
> >>>	struct io_context *ioc = get_io_context(gfp_mask);
> >>>
> >>>	spin_lock_irq(q->queue_lock);
> >>>+
> >>>+	if (!elv_may_queue(q, rw))
> >>>+		goto out_lock;
> >>>+
> >>>	if (rl->count[rw]+1 >= q->nr_requests) {
> >>>		/*
> >>>		 * The queue will fill after this allocation, so set it as
> >>>@@ -1566,15 +1576,12 @@
> >>>		}
> >>>	}
> >>>
> >>>-	if (blk_queue_full(q, rw)
> >>>-			&& !ioc_batching(ioc) && !elv_may_queue(q, rw)) {
> >>>
> >>>
> >>I know I hijacked elv_may_queue from you... any chance we could seperate
> >>these so our schedulers can live in peace? ;)
> >>
> >
> >IOW, you completely broke it! I'm just changing it back to the
> >original. When was this done, btw? Just discovered it when updating the
> >patch. Pretty annoying...
> >
> 
> You acked the change actually :P
> I guess it was done in mainline when AS was merged.

Probably missed the semantic change of may_queue.

> >>Maybe my version should be called elv_force_queue?
> >>
> >
> >I just hate to see more of these, really. The original idea for
> >may_queue was just that, may this process queue io or not. We can make
> >it return something else, though, to indicate whether the process must
> >be able to queue. Is it really needed?
> >
> 
> Its quite important. If the queue is full, and AS is waiting for a process
> to submit a request, its got a long wait.
> 
> Maybe a lower limit for per process nr_requests. Ie. you may queue if this
> queue has less than 128 requests _or_ you have less than 8 requests
> outstanding. This would solve my problem. It would also give you a much more
> appropriate scaling for server workloads, I think. Still, thats quite a
> change in behaviour (simple to code though).

That basically belongs inside your may_queue for the io scheduler, imo.

-- 
Jens Axboe