blk-mq: request initialization optimizations
authorJens Axboe <axboe@fb.com>
Thu, 29 May 2014 17:00:11 +0000 (11:00 -0600)
committerJens Axboe <axboe@fb.com>
Thu, 29 May 2014 17:00:11 +0000 (11:00 -0600)
We currently clear a lot more than we need to, so make that a bit
more clever. Make some of the init dependent on features, like
only setting start_time if we are going to use it.

Signed-off-by: Jens Axboe <axboe@fb.com>
block/blk-mq.c

index f98d977..6160128 100644 (file)
@@ -199,19 +199,12 @@ static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
        rq->q = q;
        rq->mq_ctx = ctx;
        rq->cmd_flags |= rw_flags;
-       rq->cmd_type = 0;
        /* do not touch atomic flags, it needs atomic ops against the timer */
        rq->cpu = -1;
-       rq->__data_len = 0;
-       rq->__sector = (sector_t) -1;
-       rq->bio = NULL;
-       rq->biotail = NULL;
        INIT_HLIST_NODE(&rq->hash);
        RB_CLEAR_NODE(&rq->rb_node);
-       memset(&rq->flush, 0, max(sizeof(rq->flush), sizeof(rq->elv)));
        rq->rq_disk = NULL;
        rq->part = NULL;
-       rq->start_time = jiffies;
 #ifdef CONFIG_BLK_CGROUP
        rq->rl = NULL;
        set_start_time_ns(rq);
@@ -221,23 +214,16 @@ static void blk_mq_rq_ctx_init(struct request_queue *q, struct blk_mq_ctx *ctx,
 #if defined(CONFIG_BLK_DEV_INTEGRITY)
        rq->nr_integrity_segments = 0;
 #endif
-       rq->ioprio = 0;
        rq->special = NULL;
        /* tag was already set */
        rq->errors = 0;
-       memset(rq->__cmd, 0, sizeof(rq->__cmd));
-       rq->cmd = rq->__cmd;
-       rq->cmd_len = BLK_MAX_CDB;
 
        rq->extra_len = 0;
        rq->sense_len = 0;
        rq->resid_len = 0;
        rq->sense = NULL;
 
-       rq->deadline = 0;
        INIT_LIST_HEAD(&rq->timeout_list);
-       rq->timeout = 0;
-       rq->retries = 0;
        rq->end_io = NULL;
        rq->end_io_data = NULL;
        rq->next_rq = NULL;
@@ -449,8 +435,10 @@ static void blk_mq_start_request(struct request *rq, bool last)
         * complete. So be sure to clear complete again when we start
         * the request, otherwise we'll ignore the completion event.
         */
-       set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
-       clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
+       if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
+               set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
+       if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
+               clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 
        if (q->dma_drain_size && blk_rq_bytes(rq)) {
                /*
@@ -1112,7 +1100,11 @@ void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
 static void blk_mq_bio_to_request(struct request *rq, struct bio *bio)
 {
        init_request_from_bio(rq, bio);
-       blk_account_io_start(rq, 1);
+
+       if (blk_do_io_stat(rq)) {
+               rq->start_time = jiffies;
+               blk_account_io_start(rq, 1);
+       }
 }
 
 static inline bool blk_mq_merge_queue_io(struct blk_mq_hw_ctx *hctx,