blk-mq: do limited block plug for multiple queue case
authorShaohua Li <shli@fb.com>
Fri, 8 May 2015 17:51:32 +0000 (10:51 -0700)
committerJens Axboe <axboe@fb.com>
Fri, 8 May 2015 20:17:21 +0000 (14:17 -0600)
plug is still helpful for workload with IO merge, but it can be harmful
otherwise especially with multiple hardware queues, as there is
(supposed) no lock contention in this case and plug can introduce
latency. For multiple queues, we do limited plug, eg plug only if there
is request merge. If a request doesn't have merge with following
request, the requet will be dispatched immediately.

V2: check blk_queue_nomerges() as suggested by Jeff.

Cc: Jens Axboe <axboe@fb.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
block/blk-mq.c

index f13d0de42f5372ddc7cf35262b52eeb977e9c3ec..902c2eb9a0e74edddbe1f075cd123eadda659099 100644 (file)
@@ -1224,6 +1224,38 @@ static struct request *blk_mq_map_request(struct request_queue *q,
        return rq;
 }
 
+static int blk_mq_direct_issue_request(struct request *rq)
+{
+       int ret;
+       struct request_queue *q = rq->q;
+       struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q,
+                       rq->mq_ctx->cpu);
+       struct blk_mq_queue_data bd = {
+               .rq = rq,
+               .list = NULL,
+               .last = 1
+       };
+
+       /*
+        * For OK queue, we are done. For error, kill it. Any other
+        * error (busy), just add it to our list as we previously
+        * would have done
+        */
+       ret = q->mq_ops->queue_rq(hctx, &bd);
+       if (ret == BLK_MQ_RQ_QUEUE_OK)
+               return 0;
+       else {
+               __blk_mq_requeue_request(rq);
+
+               if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
+                       rq->errors = -EIO;
+                       blk_mq_end_request(rq, rq->errors);
+                       return 0;
+               }
+               return -1;
+       }
+}
+
 /*
  * Multiple hardware queue variant. This will not use per-process plugs,
  * but will attempt to bypass the hctx queueing if we can go straight to
@@ -1235,6 +1267,8 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
        const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
        struct blk_map_ctx data;
        struct request *rq;
+       unsigned int request_count = 0;
+       struct blk_plug *plug;
 
        blk_queue_bounce(q, &bio);
 
@@ -1243,6 +1277,10 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
                return;
        }
 
+       if (!is_flush_fua && !blk_queue_nomerges(q) &&
+           blk_attempt_plug_merge(q, bio, &request_count))
+               return;
+
        rq = blk_mq_map_request(q, bio, &data);
        if (unlikely(!rq))
                return;
@@ -1253,40 +1291,39 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
                goto run_queue;
        }
 
+       plug = current->plug;
        /*
         * If the driver supports defer issued based on 'last', then
         * queue it up like normal since we can potentially save some
         * CPU this way.
         */
-       if (is_sync && !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
-               struct blk_mq_queue_data bd = {
-                       .rq = rq,
-                       .list = NULL,
-                       .last = 1
-               };
-               int ret;
+       if (((plug && !blk_queue_nomerges(q)) || is_sync) &&
+           !(data.hctx->flags & BLK_MQ_F_DEFER_ISSUE)) {
+               struct request *old_rq = NULL;
 
                blk_mq_bio_to_request(rq, bio);
 
                /*
-                * For OK queue, we are done. For error, kill it. Any other
-                * error (busy), just add it to our list as we previously
-                * would have done
+                * we do limited pluging. If bio can be merged, do merge.
+                * Otherwise the existing request in the plug list will be
+                * issued. So the plug list will have one request at most
                 */
-               ret = q->mq_ops->queue_rq(data.hctx, &bd);
-               if (ret == BLK_MQ_RQ_QUEUE_OK)
-                       goto done;
-               else {
-                       __blk_mq_requeue_request(rq);
-
-                       if (ret == BLK_MQ_RQ_QUEUE_ERROR) {
-                               rq->errors = -EIO;
-                               blk_mq_end_request(rq, rq->errors);
-                               goto done;
+               if (plug) {
+                       if (!list_empty(&plug->mq_list)) {
+                               old_rq = list_first_entry(&plug->mq_list,
+                                       struct request, queuelist);
+                               list_del_init(&old_rq->queuelist);
                        }
-                       blk_mq_insert_request(rq, false, true, true);
+                       list_add_tail(&rq->queuelist, &plug->mq_list);
+               } else /* is_sync */
+                       old_rq = rq;
+               blk_mq_put_ctx(data.ctx);
+               if (!old_rq)
                        return;
-               }
+               if (!blk_mq_direct_issue_request(old_rq))
+                       return;
+               blk_mq_insert_request(old_rq, false, true, true);
+               return;
        }
 
        if (!blk_mq_merge_queue_io(data.hctx, data.ctx, rq, bio)) {
@@ -1299,7 +1336,6 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
 run_queue:
                blk_mq_run_hw_queue(data.hctx, !is_sync || is_flush_fua);
        }
-done:
        blk_mq_put_ctx(data.ctx);
 }