blk-mq: implement new and more efficient tagging scheme

blk-mq currently uses percpu_ida for tag allocation. But that only
works well if the ratio between tag space and number of CPUs is
sufficiently high. For most devices and systems, that is not the
case. The end result if that we either only utilize the tag space
partially, or we end up attempting to fully exhaust it and run
into lots of lock contention with stealing between CPUs. This is
not optimal.

This new tagging scheme is a hybrid bitmap allocator. It uses
two tricks to both be SMP friendly and allow full exhaustion
of the space:

1) We cache the last allocated (or freed) tag on a per blk-mq
   software context basis. This allows us to limit the space
   we have to search. The key element here is not caching it
   in the shared tag structure, otherwise we end up dirtying
   more shared cache lines on each allocate/free operation.

2) The tag space is split into cache line sized groups, and
   each context will start off randomly in that space. Even up
   to full utilization of the space, this divides the tag users
   efficiently into cache line groups, avoiding dirtying the same
   one both between allocators and between allocator and freeer.

This scheme shows drastically better behaviour, both on small
tag spaces but on large ones as well. It has been tested extensively
to show better performance for all the cases blk-mq cares about.

Signed-off-by: Jens Axboe <axboe@fb.com>

5 files changed, 391 insertions, 99 deletions

diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c
index 1f43d6e..467f3a2 100644
--- a/block/blk-mq-tag.c
+++ b/block/blk-mq-tag.c
@@ -1,64 +1,257 @@
 #include <linux/kernel.h>
 #include <linux/module.h>
+#include <linux/random.h>
 
 #include <linux/blk-mq.h>
 #include "blk.h"
 #include "blk-mq.h"
 #include "blk-mq-tag.h"
 
-void blk_mq_wait_for_tags(struct blk_mq_tags *tags, bool reserved)
+void blk_mq_wait_for_tags(struct blk_mq_tags *tags, struct blk_mq_hw_ctx *hctx,
+			  bool reserved)
 {
-	int tag = blk_mq_get_tag(tags, __GFP_WAIT, reserved);
-	blk_mq_put_tag(tags, tag);
+	int tag, zero = 0;
+
+	tag = blk_mq_get_tag(tags, hctx, &zero, __GFP_WAIT, reserved);
+	blk_mq_put_tag(tags, tag, &zero);
+}
+
+static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
+{
+	int i;
+
+	for (i = 0; i < bt->map_nr; i++) {
+		struct blk_mq_bitmap *bm = &bt->map[i];
+		int ret;
+
+		ret = find_first_zero_bit(&bm->word, bm->depth);
+		if (ret < bm->depth)
+			return true;
+	}
+
+	return false;
 }
 
 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
 {
-	return !tags ||
-		percpu_ida_free_tags(&tags->free_tags, nr_cpu_ids) != 0;
+	if (!tags)
+		return true;
+
+	return bt_has_free_tags(&tags->bitmap_tags);
+}
+
+static int __bt_get_word(struct blk_mq_bitmap *bm, unsigned int last_tag)
+{
+	int tag, org_last_tag, end;
+
+	org_last_tag = last_tag = TAG_TO_BIT(last_tag);
+	end = bm->depth;
+	do {
+restart:
+		tag = find_next_zero_bit(&bm->word, end, last_tag);
+		if (unlikely(tag >= end)) {
+			/*
+			 * We started with an offset, start from 0 to
+			 * exhaust the map.
+			 */
+			if (org_last_tag && last_tag) {
+				end = last_tag;
+				last_tag = 0;
+				goto restart;
+			}
+			return -1;
+		}
+		last_tag = tag + 1;
+	} while (test_and_set_bit_lock(tag, &bm->word));
+
+	return tag;
+}
+
+/*
+ * Straight forward bitmap tag implementation, where each bit is a tag
+ * (cleared == free, and set == busy). The small twist is using per-cpu
+ * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
+ * contexts. This enables us to drastically limit the space searched,
+ * without dirtying an extra shared cacheline like we would if we stored
+ * the cache value inside the shared blk_mq_bitmap_tags structure. On top
+ * of that, each word of tags is in a separate cacheline. This means that
+ * multiple users will tend to stick to different cachelines, at least
+ * until the map is exhausted.
+ */
+static int __bt_get(struct blk_mq_bitmap_tags *bt, unsigned int *tag_cache)
+{
+	unsigned int last_tag, org_last_tag;
+	int index, i, tag;
+
+	last_tag = org_last_tag = *tag_cache;
+	index = TAG_TO_INDEX(last_tag);
+
+	for (i = 0; i < bt->map_nr; i++) {
+		tag = __bt_get_word(&bt->map[index], last_tag);
+		if (tag != -1) {
+			tag += index * BITS_PER_LONG;
+			goto done;
+		}
+
+		last_tag = 0;
+		if (++index >= bt->map_nr)
+			index = 0;
+	}
+
+	*tag_cache = 0;
+	return -1;
+
+	/*
+	 * Only update the cache from the allocation path, if we ended
+	 * up using the specific cached tag.
+	 */
+done:
+	if (tag == org_last_tag) {
+		last_tag = tag + 1;
+		if (last_tag >= bt->depth - 1)
+			last_tag = 0;
+
+		*tag_cache = last_tag;
+	}
+
+	return tag;
+}
+
+static inline void bt_index_inc(unsigned int *index)
+{
+	*index = (*index + 1) & (BT_WAIT_QUEUES - 1);
+}
+
+static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
+					 struct blk_mq_hw_ctx *hctx)
+{
+	struct bt_wait_state *bs;
+
+	if (!hctx)
+		return &bt->bs[0];
+
+	bs = &bt->bs[hctx->wait_index];
+	bt_index_inc(&hctx->wait_index);
+	return bs;
 }
 
-static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp)
+static int bt_get(struct blk_mq_bitmap_tags *bt, struct blk_mq_hw_ctx *hctx,
+		  unsigned int *last_tag, gfp_t gfp)
 {
+	struct bt_wait_state *bs;
+	DEFINE_WAIT(wait);
 	int tag;
 
-	tag = percpu_ida_alloc(&tags->free_tags, (gfp & __GFP_WAIT) ?
-			       TASK_UNINTERRUPTIBLE : TASK_RUNNING);
-	if (tag < 0)
-		return BLK_MQ_TAG_FAIL;
-	return tag + tags->nr_reserved_tags;
+	tag = __bt_get(bt, last_tag);
+	if (tag != -1)
+		return tag;
+
+	if (!(gfp & __GFP_WAIT))
+		return -1;
+
+	bs = bt_wait_ptr(bt, hctx);
+	do {
+		bool was_empty;
+
+		was_empty = list_empty(&wait.task_list);
+		prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
+
+		tag = __bt_get(bt, last_tag);
+		if (tag != -1)
+			break;
+
+		if (was_empty)
+			atomic_set(&bs->wait_cnt, bt->wake_cnt);
+
+		io_schedule();
+	} while (1);
+
+	finish_wait(&bs->wait, &wait);
+	return tag;
+}
+
+static unsigned int __blk_mq_get_tag(struct blk_mq_tags *tags,
+				     struct blk_mq_hw_ctx *hctx,
+				     unsigned int *last_tag, gfp_t gfp)
+{
+	int tag;
+
+	tag = bt_get(&tags->bitmap_tags, hctx, last_tag, gfp);
+	if (tag >= 0)
+		return tag + tags->nr_reserved_tags;
+
+	return BLK_MQ_TAG_FAIL;
 }
 
 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_tags *tags,
 					      gfp_t gfp)
 {
-	int tag;
+	int tag, zero = 0;
 
 	if (unlikely(!tags->nr_reserved_tags)) {
 		WARN_ON_ONCE(1);
 		return BLK_MQ_TAG_FAIL;
 	}
 
-	tag = percpu_ida_alloc(&tags->reserved_tags, (gfp & __GFP_WAIT) ?
-			       TASK_UNINTERRUPTIBLE : TASK_RUNNING);
+	tag = bt_get(&tags->breserved_tags, NULL, &zero, gfp);
 	if (tag < 0)
 		return BLK_MQ_TAG_FAIL;
+
 	return tag;
 }
 
-unsigned int blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp, bool reserved)
+unsigned int blk_mq_get_tag(struct blk_mq_tags *tags,
+			    struct blk_mq_hw_ctx *hctx, unsigned int *last_tag,
+			    gfp_t gfp, bool reserved)
 {
 	if (!reserved)
-		return __blk_mq_get_tag(tags, gfp);
+		return __blk_mq_get_tag(tags, hctx, last_tag, gfp);
 
 	return __blk_mq_get_reserved_tag(tags, gfp);
 }
 
+static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
+{
+	int i, wake_index;
+
+	wake_index = bt->wake_index;
+	for (i = 0; i < BT_WAIT_QUEUES; i++) {
+		struct bt_wait_state *bs = &bt->bs[wake_index];
+
+		if (waitqueue_active(&bs->wait)) {
+			if (wake_index != bt->wake_index)
+				bt->wake_index = wake_index;
+
+			return bs;
+		}
+
+		bt_index_inc(&wake_index);
+	}
+
+	return NULL;
+}
+
+static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
+{
+	const int index = TAG_TO_INDEX(tag);
+	struct bt_wait_state *bs;
+
+	clear_bit(TAG_TO_BIT(tag), &bt->map[index].word);
+
+	bs = bt_wake_ptr(bt);
+	if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
+		smp_mb__after_clear_bit();
+		atomic_set(&bs->wait_cnt, bt->wake_cnt);
+		bt_index_inc(&bt->wake_index);
+		wake_up(&bs->wait);
+	}
+}
+
 static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
 {
 	BUG_ON(tag >= tags->nr_tags);
 
-	percpu_ida_free(&tags->free_tags, tag - tags->nr_reserved_tags);
+	bt_clear_tag(&tags->bitmap_tags, tag);
 }
 
 static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
@@ -66,22 +259,41 @@ static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
 {
 	BUG_ON(tag >= tags->nr_reserved_tags);
 
-	percpu_ida_free(&tags->reserved_tags, tag);
+	bt_clear_tag(&tags->breserved_tags, tag);
 }
 
-void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
+void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag,
+		    unsigned int *last_tag)
 {
-	if (tag >= tags->nr_reserved_tags)
-		__blk_mq_put_tag(tags, tag);
-	else
+	if (tag >= tags->nr_reserved_tags) {
+		const int real_tag = tag - tags->nr_reserved_tags;
+
+		__blk_mq_put_tag(tags, real_tag);
+		*last_tag = real_tag;
+	} else
 		__blk_mq_put_reserved_tag(tags, tag);
 }
 
-static int __blk_mq_tag_iter(unsigned id, void *data)
+static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
+			     unsigned long *free_map, unsigned int off)
 {
-	unsigned long *tag_map = data;
-	__set_bit(id, tag_map);
-	return 0;
+	int i;
+
+	for (i = 0; i < bt->map_nr; i++) {
+		struct blk_mq_bitmap *bm = &bt->map[i];
+		int bit = 0;
+
+		do {
+			bit = find_next_zero_bit(&bm->word, bm->depth, bit);
+			if (bit >= bm->depth)
+				break;
+
+			__set_bit(bit + off, free_map);
+			bit++;
+		} while (1);
+
+		off += BITS_PER_LONG;
+	}
 }
 
 void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
@@ -95,21 +307,98 @@ void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
 	if (!tag_map)
 		return;
 
-	percpu_ida_for_each_free(&tags->free_tags, __blk_mq_tag_iter, tag_map);
+	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
 	if (tags->nr_reserved_tags)
-		percpu_ida_for_each_free(&tags->reserved_tags, __blk_mq_tag_iter,
-			tag_map);
+		bt_for_each_free(&tags->breserved_tags, tag_map, 0);
 
 	fn(data, tag_map);
 	kfree(tag_map);
 }
 
+static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
+{
+	unsigned int i, used;
+
+	for (i = 0, used = 0; i < bt->map_nr; i++) {
+		struct blk_mq_bitmap *bm = &bt->map[i];
+
+		used += bitmap_weight(&bm->word, bm->depth);
+	}
+
+	return bt->depth - used;
+}
+
+static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
+			int node, bool reserved)
+{
+	int i;
+
+	/*
+	 * Depth can be zero for reserved tags, that's not a failure
+	 * condition.
+	 */
+	if (depth) {
+		int nr, i, map_depth;
+
+		nr = ALIGN(depth, BITS_PER_LONG) / BITS_PER_LONG;
+		bt->map = kzalloc_node(nr * sizeof(struct blk_mq_bitmap),
+						GFP_KERNEL, node);
+		if (!bt->map)
+			return -ENOMEM;
+
+		bt->map_nr = nr;
+		map_depth = depth;
+		for (i = 0; i < nr; i++) {
+			bt->map[i].depth = min(map_depth, BITS_PER_LONG);
+			map_depth -= BITS_PER_LONG;
+		}
+	}
+
+	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
+	if (!bt->bs) {
+		kfree(bt->map);
+		return -ENOMEM;
+	}
+
+	for (i = 0; i < BT_WAIT_QUEUES; i++)
+		init_waitqueue_head(&bt->bs[i].wait);
+
+	bt->wake_cnt = BT_WAIT_BATCH;
+	if (bt->wake_cnt > depth / 4)
+		bt->wake_cnt = max(1U, depth / 4);
+
+	bt->depth = depth;
+	return 0;
+}
+
+static void bt_free(struct blk_mq_bitmap_tags *bt)
+{
+	kfree(bt->map);
+	kfree(bt->bs);
+}
+
+static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
+						   int node)
+{
+	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
+
+	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
+		goto enomem;
+	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
+		goto enomem;
+
+	return tags;
+enomem:
+	bt_free(&tags->bitmap_tags);
+	kfree(tags);
+	return NULL;
+}
+
 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 				     unsigned int reserved_tags, int node)
 {
 	unsigned int nr_tags, nr_cache;
 	struct blk_mq_tags *tags;
-	int ret;
 
 	if (total_tags > BLK_MQ_TAG_MAX) {
 		pr_err("blk-mq: tag depth too large\n");
@@ -121,72 +410,46 @@ struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
 		return NULL;
 
 	nr_tags = total_tags - reserved_tags;
-	nr_cache = nr_tags / num_possible_cpus();
-
-	if (nr_cache < BLK_MQ_TAG_CACHE_MIN)
-		nr_cache = BLK_MQ_TAG_CACHE_MIN;
-	else if (nr_cache > BLK_MQ_TAG_CACHE_MAX)
-		nr_cache = BLK_MQ_TAG_CACHE_MAX;
+	nr_cache = nr_tags / num_online_cpus();
 
 	tags->nr_tags = total_tags;
 	tags->nr_reserved_tags = reserved_tags;
-	tags->nr_max_cache = nr_cache;
-	tags->nr_batch_move = max(1u, nr_cache / 2);
-
-	ret = __percpu_ida_init(&tags->free_tags, tags->nr_tags -
-				tags->nr_reserved_tags,
-				tags->nr_max_cache,
-				tags->nr_batch_move);
-	if (ret)
-		goto err_free_tags;
-
-	if (reserved_tags) {
-		/*
-		 * With max_cahe and batch set to 1, the allocator fallbacks to
-		 * no cached. It's fine reserved tags allocation is slow.
-		 */
-		ret = __percpu_ida_init(&tags->reserved_tags, reserved_tags,
-				1, 1);
-		if (ret)
-			goto err_reserved_tags;
-	}
 
-	return tags;
-
-err_reserved_tags:
-	percpu_ida_destroy(&tags->free_tags);
-err_free_tags:
-	kfree(tags);
-	return NULL;
+	return blk_mq_init_bitmap_tags(tags, node);
 }
 
 void blk_mq_free_tags(struct blk_mq_tags *tags)
 {
-	percpu_ida_destroy(&tags->free_tags);
-	percpu_ida_destroy(&tags->reserved_tags);
+	bt_free(&tags->bitmap_tags);
+	bt_free(&tags->breserved_tags);
 	kfree(tags);
 }
 
+void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
+{
+	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
+
+	if (depth > 1)
+		*tag = prandom_u32() % (depth - 1);
+	else
+		*tag = 0;
+}
+
 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
 {
 	char *orig_page = page;
-	unsigned int cpu;
+	unsigned int free, res;
 
 	if (!tags)
 		return 0;
 
-	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, batch_move=%u,"
-			" max_cache=%u\n", tags->nr_tags, tags->nr_reserved_tags,
-			tags->nr_batch_move, tags->nr_max_cache);
+	page += sprintf(page, "nr_tags=%u, reserved_tags=%u\n",
+			tags->nr_tags, tags->nr_reserved_tags);
 
-	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n",
-			percpu_ida_free_tags(&tags->free_tags, nr_cpu_ids),
-			percpu_ida_free_tags(&tags->reserved_tags, nr_cpu_ids));
+	free = bt_unused_tags(&tags->bitmap_tags);
+	res = bt_unused_tags(&tags->breserved_tags);
 
-	for_each_possible_cpu(cpu) {
-		page += sprintf(page, "  cpu%02u: nr_free=%u\n", cpu,
-				percpu_ida_free_tags(&tags->free_tags, cpu));
-	}
+	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
 
 	return page - orig_page;
 }
diff --git a/block/blk-mq-tag.h b/block/blk-mq-tag.h
index c8e0645..06d4a2f 100644
--- a/block/blk-mq-tag.h
+++ b/block/blk-mq-tag.h
@@ -1,7 +1,34 @@
 #ifndef INT_BLK_MQ_TAG_H
 #define INT_BLK_MQ_TAG_H
 
-#include <linux/percpu_ida.h>
+enum {
+	BT_WAIT_QUEUES	= 8,
+	BT_WAIT_BATCH	= 8,
+};
+
+struct bt_wait_state {
+	atomic_t wait_cnt;
+	wait_queue_head_t wait;
+} ____cacheline_aligned_in_smp;
+
+#define TAG_TO_INDEX(tag)	((tag) / BITS_PER_LONG)
+#define TAG_TO_BIT(tag)		((tag) & (BITS_PER_LONG - 1))
+
+struct blk_mq_bitmap {
+	unsigned long word;
+	unsigned long depth;
+} ____cacheline_aligned_in_smp;
+
+struct blk_mq_bitmap_tags {
+	unsigned int depth;
+	unsigned int wake_cnt;
+
+	struct blk_mq_bitmap *map;
+	unsigned int map_nr;
+
+	unsigned int wake_index;
+	struct bt_wait_state *bs;
+};
 
 /*
  * Tag address space map.
@@ -9,11 +36,9 @@
 struct blk_mq_tags {
 	unsigned int nr_tags;
 	unsigned int nr_reserved_tags;
-	unsigned int nr_batch_move;
-	unsigned int nr_max_cache;
 
-	struct percpu_ida free_tags;
-	struct percpu_ida reserved_tags;
+	struct blk_mq_bitmap_tags bitmap_tags;
+	struct blk_mq_bitmap_tags breserved_tags;
 
 	struct request **rqs;
 	struct list_head page_list;
@@ -23,12 +48,13 @@ struct blk_mq_tags {
 extern struct blk_mq_tags *blk_mq_init_tags(unsigned int nr_tags, unsigned int reserved_tags, int node);
 extern void blk_mq_free_tags(struct blk_mq_tags *tags);
 
-extern unsigned int blk_mq_get_tag(struct blk_mq_tags *tags, gfp_t gfp, bool reserved);
-extern void blk_mq_wait_for_tags(struct blk_mq_tags *tags, bool reserved);
-extern void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag);
+extern unsigned int blk_mq_get_tag(struct blk_mq_tags *tags, struct blk_mq_hw_ctx *hctx, unsigned int *last_tag, gfp_t gfp, bool reserved);
+extern void blk_mq_wait_for_tags(struct blk_mq_tags *tags, struct blk_mq_hw_ctx *hctx, bool reserved);
+extern void blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag, unsigned int *last_tag);
 extern void blk_mq_tag_busy_iter(struct blk_mq_tags *tags, void (*fn)(void *data, unsigned long *), void *data);
 extern bool blk_mq_has_free_tags(struct blk_mq_tags *tags);
 extern ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page);
+extern void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *last_tag);
 
 enum {
 	BLK_MQ_TAG_CACHE_MIN	= 1,
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 492f49f..9f07a26 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -74,12 +74,13 @@ static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
 }
 
 static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
+					      struct blk_mq_ctx *ctx,
 					      gfp_t gfp, bool reserved)
 {
 	struct request *rq;
 	unsigned int tag;
 
-	tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
+	tag = blk_mq_get_tag(hctx->tags, hctx, &ctx->last_tag, gfp, reserved);
 	if (tag != BLK_MQ_TAG_FAIL) {
 		rq = hctx->tags->rqs[tag];
 		rq->tag = tag;
@@ -246,7 +247,8 @@ static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
 		struct blk_mq_ctx *ctx = blk_mq_get_ctx(q);
 		struct blk_mq_hw_ctx *hctx = q->mq_ops->map_queue(q, ctx->cpu);
 
-		rq = __blk_mq_alloc_request(hctx, gfp & ~__GFP_WAIT, reserved);
+		rq = __blk_mq_alloc_request(hctx, ctx, gfp & ~__GFP_WAIT,
+						reserved);
 		if (rq) {
 			blk_mq_rq_ctx_init(q, ctx, rq, rw);
 			break;
@@ -260,7 +262,7 @@ static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
 			break;
 		}
 
-		blk_mq_wait_for_tags(hctx->tags, reserved);
+		blk_mq_wait_for_tags(hctx->tags, hctx, reserved);
 	} while (1);
 
 	return rq;
@@ -278,6 +280,7 @@ struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
 		blk_mq_put_ctx(rq->mq_ctx);
 	return rq;
 }
+EXPORT_SYMBOL(blk_mq_alloc_request);
 
 struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
 					      gfp_t gfp)
@@ -301,7 +304,7 @@ static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
 	struct request_queue *q = rq->q;
 
 	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
-	blk_mq_put_tag(hctx->tags, tag);
+	blk_mq_put_tag(hctx->tags, tag, &ctx->last_tag);
 	blk_mq_queue_exit(q);
 }
 
@@ -677,11 +680,6 @@ static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 			queued++;
 			continue;
 		case BLK_MQ_RQ_QUEUE_BUSY:
-			/*
-			 * FIXME: we should have a mechanism to stop the queue
-			 * like blk_stop_queue, otherwise we will waste cpu
-			 * time
-			 */
 			list_add(&rq->queuelist, &rq_list);
 			__blk_mq_requeue_request(rq);
 			break;
@@ -873,6 +871,7 @@ static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
 		list_add(&rq->queuelist, &ctx->rq_list);
 	else
 		list_add_tail(&rq->queuelist, &ctx->rq_list);
+
 	blk_mq_hctx_mark_pending(hctx, ctx);
 
 	/*
@@ -1046,7 +1045,7 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
 	if (is_sync)
 		rw |= REQ_SYNC;
 	trace_block_getrq(q, bio, rw);
-	rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
+	rq = __blk_mq_alloc_request(hctx, ctx, GFP_ATOMIC, false);
 	if (likely(rq))
 		blk_mq_rq_ctx_init(q, ctx, rq, rw);
 	else {
@@ -1130,8 +1129,8 @@ EXPORT_SYMBOL(blk_mq_map_queue);
 struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_tag_set *set,
 						   unsigned int hctx_index)
 {
-	return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
-				GFP_KERNEL | __GFP_ZERO, set->numa_node);
+	return kzalloc_node(sizeof(struct blk_mq_hw_ctx), GFP_KERNEL,
+				set->numa_node);
 }
 EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
 
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 1ae364c..97cfab9 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -12,6 +12,8 @@ struct blk_mq_ctx {
 	unsigned int		cpu;
 	unsigned int		index_hw;
 
+	unsigned int		last_tag ____cacheline_aligned_in_smp;
+
 	/* incremented at dispatch time */
 	unsigned long		rq_dispatched[2];
 	unsigned long		rq_merged;
@@ -21,7 +23,7 @@ struct blk_mq_ctx {
 
 	struct request_queue	*queue;
 	struct kobject		kobj;
-};
+} ____cacheline_aligned_in_smp;
 
 void __blk_mq_complete_request(struct request *rq);
 void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async);
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 5bd677e..f83d15f 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -31,10 +31,12 @@ struct blk_mq_hw_ctx {
 
 	void			*driver_data;
 
-	unsigned int		nr_ctx;
-	struct blk_mq_ctx	**ctxs;
 	unsigned int 		nr_ctx_map;
 	unsigned long		*ctx_map;
+	unsigned int		nr_ctx;
+	struct blk_mq_ctx	**ctxs;
+
+	unsigned int		wait_index;
 
 	struct blk_mq_tags	*tags;