Re: [PATCH] bfq: do try insert merge before bfq_init_rq() call

[Jan Kara]

On Thu 13-10-22 21:53:21, Yuwei Guan wrote:
> It's useless to do bfq_init_rq(rq), if the rq can do merge first.
> 
> In the patch 5f550ede5edf8, it moved to bfq_init_rq() before
> blk_mq_sched_try_insert_merge(), but it's pointless,
> as the fifo_time of next is not set yet,
> and !list_empty(&next->queuelist) is 0, so it does not
> need to reposition rq's fifo_time.
> 
> And for the "hash lookup, try again" situation, as follow,
> bfq_requests_merged() call can work normally.
> 
> blk_mq_sched_try_insert_merge
>   elv_attempt_insert_merge
>     elv_rqhash_find
> 
> Signed-off-by: Yuwei Guan <Yuwei.Guan@xxxxxxxxxxxxx>

OK, after some thinking I agree. How much testing has this patch got?
Because I'd like to verify we didn't overlook something.

							Honza

> ---
>  block/bfq-iosched.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
> index 7ea427817f7f..9845370a701c 100644
> --- a/block/bfq-iosched.c
> +++ b/block/bfq-iosched.c
> @@ -6147,7 +6147,7 @@ static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
>  		bfqg_stats_update_legacy_io(q, rq);
>  #endif
>  	spin_lock_irq(&bfqd->lock);
> -	bfqq = bfq_init_rq(rq);
> +
>  	if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
>  		spin_unlock_irq(&bfqd->lock);
>  		blk_mq_free_requests(&free);
> @@ -6156,6 +6156,7 @@ static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
>  
>  	trace_block_rq_insert(rq);
>  
> +	bfqq = bfq_init_rq(rq);
>  	if (!bfqq || at_head) {
>  		if (at_head)
>  			list_add(&rq->queuelist, &bfqd->dispatch);
> -- 
> 2.34.1
> 
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR