It looks like what's happening is when dm_dispatch_clone_request
dispatches the "clone" I/O request to the underlying (real) device
from the multipath device, the scsi driver can (often under load)
return BLK_MQ_RQ_QUEUE_DEV_BUSY. dm_dispatch_clone_request doesn't
have that as an exception the way it does BLK_MQ_RQ_QUEUE_BUSY and
so it calls dm_complete_request which propagates
the BLK_MQ_RQ_QUEUE_DEV_BUSY error code up the stack resulting
in multipath_end_io calling fail_path and failing the path because
there is an error value set.

diff --git a/drivers/md/dm-rq.c b/drivers/md/dm-rq.c
index 02da1e65..e4f58472 100644
--- a/drivers/md/dm-rq.c
+++ b/drivers/md/dm-rq.c
@@ -477,7 +477,8 @@ static int dm_dispatch_clone_request(struct request *clone, struct request *rq)
 
 	clone->start_time = jiffies;
 	r = blk_insert_cloned_request(clone->q, clone);
-	if (r != BLK_MQ_RQ_QUEUE_OK && r != BLK_MQ_RQ_QUEUE_BUSY)
+	if (r != BLK_MQ_RQ_QUEUE_OK && r != BLK_MQ_RQ_QUEUE_BUSY &&
+	    r != BLK_MQ_RQ_QUEUE_DEV_BUSY)
 		/* must complete clone in terms of original request */
 		dm_complete_request(rq, r);
 	return r;
@@ -661,7 +662,7 @@ check_again:
 		trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
 				     blk_rq_pos(rq));
 		ret = dm_dispatch_clone_request(clone, rq);
-		if (ret == BLK_MQ_RQ_QUEUE_BUSY) {
+		if (ret == BLK_MQ_RQ_QUEUE_BUSY || ret == BLK_MQ_RQ_QUEUE_DEV_BUSY) {
 			blk_rq_unprep_clone(clone);
 			tio->ti->type->release_clone_rq(clone);
 			tio->clone = NULL;