[PATCH V4 3/3] blk-mq: issue request directly for blk_insert_cloned_request

[Ming Lei <ming.lei@xxxxxxxxxx>]

blk_insert_cloned_request() is called in fast path of dm-rq driver, and
in this function we append request to hctx->dispatch_list of the underlying
queue directly.

1) This way isn't efficient enough because hctx lock is always required

2) With blk_insert_cloned_request(), we bypass underlying queue's IO scheduler
totally, and depend on DM rq driver to do IO schedule completely. But DM
rq driver can't get underlying queue's dispatch feedback at all, and this
information is extreamly useful for IO merge. Without that IO merge can't
be done basically by blk-mq, and causes very bad sequential IO performance.

This patch makes use of blk_mq_try_issue_directly() to dispatch rq to
underlying queue and provides disptch result to dm-rq and blk-mq, and
improves the above situations very much.

With this patch, seqential IO is improved by 3X in my test over
mpath/virtio-scsi.

Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
---
 block/blk-mq.c | 29 ++++++++++++++++++++++++++---
 1 file changed, 26 insertions(+), 3 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 1654c9c284d8..ce3965f271e3 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -1730,6 +1730,11 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 	blk_qc_t new_cookie;
 	blk_status_t ret = BLK_STS_OK;
 	bool run_queue = true;
+	/*
+	 * This function is called from blk_insert_cloned_request() if
+	 * 'cookie' is NULL, and for dispatching this request only.
+	 */
+	bool dispatch_only = !cookie;
 
 	/* RCU or SRCU read lock is needed before checking quiesced flag */
 	if (blk_mq_hctx_stopped(hctx) || blk_queue_quiesced(q)) {
@@ -1737,10 +1742,17 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 		goto insert;
 	}
 
-	if (q->elevator)
+	if (q->elevator && !dispatch_only)
 		goto insert;
 
 	ret = __blk_mq_issue_req(hctx, rq, &new_cookie);
+	if (dispatch_only) {
+		if (ret == BLK_STS_AGAIN)
+			return BLK_STS_RESOURCE;
+		if (ret == BLK_STS_RESOURCE)
+			__blk_mq_requeue_request(rq);
+		return ret;
+	}
 	if (ret == BLK_STS_AGAIN)
 		goto insert;
 
@@ -1763,8 +1775,11 @@ static blk_status_t __blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 	}
 
 insert:
-	blk_mq_sched_insert_request(rq, false, run_queue, false,
-					hctx->flags & BLK_MQ_F_BLOCKING);
+	if (!dispatch_only)
+		blk_mq_sched_insert_request(rq, false, run_queue, false,
+				hctx->flags & BLK_MQ_F_BLOCKING);
+	else
+		blk_mq_request_bypass_insert(rq, run_queue);
 	return ret;
 }
 
@@ -1784,6 +1799,14 @@ static blk_status_t blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 	return ret;
 }
 
+blk_status_t blk_mq_request_direct_issue(struct request *rq)
+{
+	struct blk_mq_ctx *ctx = rq->mq_ctx;
+	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(rq->q, ctx->cpu);
+
+	return blk_mq_try_issue_directly(hctx, rq, NULL);
+}
+
 static blk_qc_t blk_mq_make_request(struct request_queue *q, struct bio *bio)
 {
 	const int is_sync = op_is_sync(bio->bi_opf);
-- 
2.9.5