On Fri, Aug 09, 2013 at 01:17:37PM -0700, Nicholas A. Bellinger wrote:

On Fri, 2013-08-09 at 21:15 +0200, Alexander Gordeev wrote:
Mmmm, I'm able to reproduce over here with ahci + scsi-mq, and it
appears to be a bug related with using sdev->sdev_md_req.queue_depth=1,
that ends up causing the blkdev_issue_flush() to wait forever because
blk_mq_wait_for_tags() never ends up getting the single tag back for the
WRITE_FLUSH bio -> SYNCHRONIZE_CACHE cdb.

It turns out this way - blkdev_issue_flush() claims the only tag, submits
the bio and waits for the completion. But because blk_mq_make_request()
does not mark any context in blk_mq_hw_ctx::ctx_map (nor enslists the request
into blk_mq_ctx::rq_list) it never gets processed from blk_mq_work_fn->
__blk_mq_run_hw_queue() and blkdev_issue_flush() waits endlessly. All other
requests are just waiting for the tag availability as result.

[...]

Bumping queue_depth=2 seems to work-around the issue, but AFAICT it's a
genuine tag starvation bug with queue_depth=1 and WRITE_FLUSH..

If I try to hack and force __blk_mq_run_hw_queue() to process the request...

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 6fc1df3..c22b6f66 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c

@@ -889,9 +962,12 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
 	hctx->queued++;
 
 	if (unlikely(is_flush_fua)) {
+		list_add(&rq->queuelist, &hctx->dispatch);
 		blk_mq_bio_to_request(q, rq, bio);
 		blk_mq_put_ctx(ctx);
 		blk_insert_flush(rq);
 		goto run_queue;
 	}

... I get a kernel BUG at drivers/scsi/scsi_lib.c:1233

	BUG_ON(!req->nr_phys_segments);

IOW I am not sure how to proceed.

--nab

-- 
Regards,
Alexander Gordeev
agordeev@redhat.com