blk-mq: fix performance regression with shared tags
authorJens Axboe <axboe@kernel.dk>
Tue, 20 Jun 2017 23:56:13 +0000 (17:56 -0600)
committerJens Axboe <axboe@kernel.dk>
Wed, 21 Jun 2017 16:17:49 +0000 (10:17 -0600)
If we have shared tags enabled, then every IO completion will trigger
a full loop of every queue belonging to a tag set, and every hardware
queue for each of those queues, even if nothing needs to be done.
This causes a massive performance regression if you have a lot of
shared devices.

Instead of doing this huge full scan on every IO, add an atomic
counter to the main queue that tracks how many hardware queues have
been marked as needing a restart. With that, we can avoid looking for
restartable queues, if we don't have to.

Max reports that this restores performance. Before this patch, 4K
IOPS was limited to 22-23K IOPS. With the patch, we are running at
950-970K IOPS.

Fixes: 6d8c6c0f97ad ("blk-mq: Restart a single queue if tag sets are shared")
Reported-by: Max Gurtovoy <maxg@mellanox.com>
Tested-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Tested-by: Bart Van Assche <bart.vanassche@wdc.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-mq-sched.c
block/blk-mq-sched.h
block/blk-mq.c
include/linux/blkdev.h

index 1f5b692526ae1a7199ee9bbaef305c4b0a42e696..0ded5e846335667406d58ce08e8439360baeb312 100644 (file)
@@ -68,6 +68,45 @@ static void blk_mq_sched_assign_ioc(struct request_queue *q,
                __blk_mq_sched_assign_ioc(q, rq, bio, ioc);
 }
 
+/*
+ * Mark a hardware queue as needing a restart. For shared queues, maintain
+ * a count of how many hardware queues are marked for restart.
+ */
+static void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
+{
+       if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
+               return;
+
+       if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
+               struct request_queue *q = hctx->queue;
+
+               if (!test_and_set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
+                       atomic_inc(&q->shared_hctx_restart);
+       } else
+               set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
+}
+
+static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
+{
+       if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
+               return false;
+
+       if (hctx->flags & BLK_MQ_F_TAG_SHARED) {
+               struct request_queue *q = hctx->queue;
+
+               if (test_and_clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
+                       atomic_dec(&q->shared_hctx_restart);
+       } else
+               clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
+
+       if (blk_mq_hctx_has_pending(hctx)) {
+               blk_mq_run_hw_queue(hctx, true);
+               return true;
+       }
+
+       return false;
+}
+
 struct request *blk_mq_sched_get_request(struct request_queue *q,
                                         struct bio *bio,
                                         unsigned int op,
@@ -266,18 +305,6 @@ static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
        return true;
 }
 
-static bool blk_mq_sched_restart_hctx(struct blk_mq_hw_ctx *hctx)
-{
-       if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state)) {
-               clear_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
-               if (blk_mq_hctx_has_pending(hctx)) {
-                       blk_mq_run_hw_queue(hctx, true);
-                       return true;
-               }
-       }
-       return false;
-}
-
 /**
  * list_for_each_entry_rcu_rr - iterate in a round-robin fashion over rcu list
  * @pos:    loop cursor.
@@ -309,6 +336,13 @@ void blk_mq_sched_restart(struct blk_mq_hw_ctx *const hctx)
        unsigned int i, j;
 
        if (set->flags & BLK_MQ_F_TAG_SHARED) {
+               /*
+                * If this is 0, then we know that no hardware queues
+                * have RESTART marked. We're done.
+                */
+               if (!atomic_read(&queue->shared_hctx_restart))
+                       return;
+
                rcu_read_lock();
                list_for_each_entry_rcu_rr(q, queue, &set->tag_list,
                                           tag_set_list) {
index edafb5383b7bbdedfd5365ed38f9a5c373ec96ab..5007edece51aced038d3db8f0adbc722c49e3d38 100644 (file)
@@ -115,15 +115,6 @@ static inline bool blk_mq_sched_has_work(struct blk_mq_hw_ctx *hctx)
        return false;
 }
 
-/*
- * Mark a hardware queue as needing a restart.
- */
-static inline void blk_mq_sched_mark_restart_hctx(struct blk_mq_hw_ctx *hctx)
-{
-       if (!test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
-               set_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
-}
-
 static inline bool blk_mq_sched_needs_restart(struct blk_mq_hw_ctx *hctx)
 {
        return test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state);
index bb66c96850b18cb419b0e44aab1894169352f9af..958cedaff8b829ceb4c724dbf1c6f6d30d883aeb 100644 (file)
@@ -2103,20 +2103,30 @@ static void blk_mq_map_swqueue(struct request_queue *q,
        }
 }
 
+/*
+ * Caller needs to ensure that we're either frozen/quiesced, or that
+ * the queue isn't live yet.
+ */
 static void queue_set_hctx_shared(struct request_queue *q, bool shared)
 {
        struct blk_mq_hw_ctx *hctx;
        int i;
 
        queue_for_each_hw_ctx(q, hctx, i) {
-               if (shared)
+               if (shared) {
+                       if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
+                               atomic_inc(&q->shared_hctx_restart);
                        hctx->flags |= BLK_MQ_F_TAG_SHARED;
-               else
+               } else {
+                       if (test_bit(BLK_MQ_S_SCHED_RESTART, &hctx->state))
+                               atomic_dec(&q->shared_hctx_restart);
                        hctx->flags &= ~BLK_MQ_F_TAG_SHARED;
+               }
        }
 }
 
-static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set, bool shared)
+static void blk_mq_update_tag_set_depth(struct blk_mq_tag_set *set,
+                                       bool shared)
 {
        struct request_queue *q;
 
index b74a3edcb3da82903568981a5b49fbbf1f4269cb..1ddd36bd2173b98e925eabdf083a796bfcabdd07 100644 (file)
@@ -391,6 +391,8 @@ struct request_queue {
        int                     nr_rqs[2];      /* # allocated [a]sync rqs */
        int                     nr_rqs_elvpriv; /* # allocated rqs w/ elvpriv */
 
+       atomic_t                shared_hctx_restart;
+
        struct blk_queue_stats  *stats;
        struct rq_wb            *rq_wb;