On 6/17/25 07:33, Bart Van Assche wrote: > Prepare for allocating a request from a specific hctx by making > blk_mq_submit_bio() allocate a request later. > > The performance impact of this patch on the hot path is small: if a > request is cached, one percpu_ref_get(&q->q_usage_counter) call and one > percpu_ref_put(&q->q_usage_counter) call are added to the hot path. Numbers ? The change is forcing a queue enter for all BIOs because you remove the cached request optimization. So I am not sure it is the impact is that small if you have a very fast storage device. -- Damien Le Moal Western Digital Research
