Re: [PATCH] block/mq-deadline: Speed up the dispatch of low-priority requests

[Bart Van Assche]

On 8/26/21 7:40 AM, Zhen Lei wrote:
> lock protection needs to be added only in
> dd_finish_request(), which is unlikely to cause significant performance
> side effects.

Not sure the above is correct. Every new atomic instruction has a measurable
performance overhead. But I guess in this case that overhead is smaller than
the time needed to sum 128 per-CPU variables.

> Tested on my 128-core board with two ssd disks.
> fio bs=4k rw=read iodepth=128 cpus_allowed=0-95 <others>
> Before:
> [183K/0/0 iops]
> [172K/0/0 iops]
> 
> After:
> [258K/0/0 iops]
> [258K/0/0 iops]

Nice work!

> Fixes: fb926032b320 ("block/mq-deadline: Prioritize high-priority requests")

Shouldn't the Fixes: tag be used only for patches that modify functionality?
I'm not sure it is appropriate to use this tag for performance improvements.

>  struct deadline_data {
> @@ -277,9 +278,9 @@ deadline_move_request(struct deadline_data *dd, struct dd_per_prio *per_prio,
>  }
>  
>  /* Number of requests queued for a given priority level. */
> -static u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
> +static __always_inline u32 dd_queued(struct deadline_data *dd, enum dd_prio prio)
>  {
> -	return dd_sum(dd, inserted, prio) - dd_sum(dd, completed, prio);
> +	return dd->per_prio[prio].nr_queued;
>  }

Please leave out "__always_inline". Modern compilers are smart enough to
inline this function without using the "inline" keyword.

> @@ -711,6 +712,8 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
>  
>  	prio = ioprio_class_to_prio[ioprio_class];
>  	dd_count(dd, inserted, prio);
> +	per_prio = &dd->per_prio[prio];
> +	per_prio->nr_queued++;
>  
>  	if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
>  		blk_mq_free_requests(&free);

I think the above is wrong - nr_queued should not be incremented if the
request is merged into another request. Please move the code that increments
nr_queued past the above if-statement.

Thanks,

Bart.