On Mon, Jun 10, 2019 at 03:00:39PM -0700, Bart Van Assche wrote: > On 6/8/19 1:19 AM, Christoph Hellwig wrote: >> On Tue, Jun 04, 2019 at 11:17:35AM -0700, Bart Van Assche wrote: >>> No code that occurs between blk_mq_get_ctx() and blk_mq_put_ctx() depends >>> on preemption being disabled for its correctness. Since removing the CPU >>> preemption calls does not measurably affect performance, simplify the >>> blk-mq code by removing the blk_mq_put_ctx() function and also by not >>> disabling preemption in blk_mq_get_ctx(). >> >> I like the idea behinds this, but I think it makes some small issues >> we have in the current code even worse. As far as I can tell the idea >> behind this call was that we operate on the same blk_mq_ctx for the >> duration of the I/O submission. Now it should not matter which one, >> that is we don't care if we get preempted, but it should stay the same. > > Hi Christoph, > > Can you clarify this? Isn't the goal of the rq->mq_ctx = data->ctx > assignment in blk_mq_rq_ctx_init() to ensure that the same blk_mq_ctx is > used during I/O submission?
Yes. But we still have additional blk_mq_get_ctx calls that I was concerned about. But looking deeper it seems like the additional ones are just used locally for I/O scheduler merge decisions, and we should be ok even if the context changes due to a preemption between the failed merge and the request allocation. That being said it would still be nice to pass the ctx from __blk_mq_sched_bio_merge to ->bio_merge instead of having to find it again in kyber_bio_merge, but that isn't urgent.
