Re: [REGRESSION] ioprio performance hangs, bisected

[Christoph Hellwig <hch@lst.de>]

On Mon, Nov 25, 2024 at 05:16:39PM +0000, Chris Bainbridge wrote:
> I did a bit of debugging.

Thanks, this was extremely helpful!

mq-deadlink not only looks at the priority in the submission path,
but also in the completion path, which is rather unexpected.  Now
for drivers that consume bios, req->bio will eventually become
NULL before the completion.

Fortunately fixing this is not only easy but also improves the
code in mq-deadline.  Can you test the patch below?

diff --git a/block/mq-deadline.c b/block/mq-deadline.c
index acdc28756d9d..91b3789f710e 100644
--- a/block/mq-deadline.c
+++ b/block/mq-deadline.c
@@ -685,10 +685,9 @@ static void dd_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
 
 	prio = ioprio_class_to_prio[ioprio_class];
 	per_prio = &dd->per_prio[prio];
-	if (!rq->elv.priv[0]) {
+	if (!rq->elv.priv[0])
 		per_prio->stats.inserted++;
-		rq->elv.priv[0] = (void *)(uintptr_t)1;
-	}
+	rq->elv.priv[0] = per_prio;
 
 	if (blk_mq_sched_try_insert_merge(q, rq, free))
 		return;
@@ -753,18 +752,14 @@ static void dd_prepare_request(struct request *rq)
  */
 static void dd_finish_request(struct request *rq)
 {
-	struct request_queue *q = rq->q;
-	struct deadline_data *dd = q->elevator->elevator_data;
-	const u8 ioprio_class = dd_rq_ioclass(rq);
-	const enum dd_prio prio = ioprio_class_to_prio[ioprio_class];
-	struct dd_per_prio *per_prio = &dd->per_prio[prio];
+	struct dd_per_prio *per_prio = rq->elv.priv[0];
 
 	/*
 	 * The block layer core may call dd_finish_request() without having
 	 * called dd_insert_requests(). Skip requests that bypassed I/O
 	 * scheduling. See also blk_mq_request_bypass_insert().
 	 */
-	if (rq->elv.priv[0])
+	if (per_prio)
 		atomic_inc(&per_prio->stats.completed);
 }