On 12/20/18 11:01 AM, Bart Van Assche wrote:

On Thu, 2018-12-20 at 06:07 -0700, Jens Axboe wrote:

quoted

On 12/20/18 6:02 AM, Jens Axboe wrote:

quoted

quoted

I'm afraid this cannot work.

The 'tags' here could be the hctx->sched_tags, but what we need to
clear is hctx->tags->rqs[].

You are right, of course, a bit too quick on the trigger. This one
should work better, and also avoids that silly quadratic loop. I don't
think we need the tag == -1 check, but probably best to be safe.

Sent out the wrong one, here it is. Bart, if you can reproduce, can you
give it a whirl?

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 2de972857496..fc04bb648f18 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -2025,7 +2025,7 @@ void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
 {
 	struct page *page;
 
-	if (tags->rqs && set->ops->exit_request) {
+	if (tags->rqs) {
 		int i;
 
 		for (i = 0; i < tags->nr_tags; i++) {

@@ -2033,8 +2033,14 @@ void blk_mq_free_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
 
 			if (!rq)
 				continue;
-			set->ops->exit_request(set, rq, hctx_idx);
+			if (set->ops->exit_request)
+				set->ops->exit_request(set, rq, hctx_idx);
 			tags->static_rqs[i] = NULL;
+
+			if (rq->tag == -1)
+				continue;
+			if (set->tags[hctx_idx]->rqs[rq->tag] == rq)
+				set->tags[hctx_idx]->rqs[rq->tag] = NULL;
 		}
 	}
 

@@ -2113,6 +2119,7 @@ static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
 			return ret;
 	}
 
+	rq->tag = -1;
 	WRITE_ONCE(rq->state, MQ_RQ_IDLE);
 	return 0;
 }

Hi Jens,

Are you sure this is sufficient?

No, I don't think it is.

My understanding is that
blk_mq_queue_tag_busy_iter() iterates over all tags in the tag set. So if the
request queue on which part_in_flight() is called and the request queue for
which blk_mq_free_rqs() is called share their tag set then part_in_flight()
and blk_mq_free_rqs() can run concurrently. That can cause ugly race
conditions. Do you think it would be a good idea to modify the inflight
accounting code such that it only considers the requests of a single request
queue instead of all requests for a given tag set?

That would of course solve it, the question is how to do it. One option
would be to have ->rqs[] be:

struct rq_entry {
	struct request_queue *q;
	struct request *rq;
};

instead of just a request, since then you could check the queue without
having to dereference the request. The current race is inherent in that
we set ->rqs[] AFTER having acquired the tag, so there's a window where
you could find a stale entry. That's not normally an issue since
requests are persistent, but for shared tag maps and queues disappearing
it can pose a problem.


-- 
Jens Axboe