blk-mq: new multi-queue block IO queueing mechanism

Linux currently has two models for block devices:

- The classic request_fn based approach, where drivers use struct
  request units for IO. The block layer provides various helper
  functionalities to let drivers share code, things like tag
  management, timeout handling, queueing, etc.

- The "stacked" approach, where a driver squeezes in between the
  block layer and IO submitter. Since this bypasses the IO stack,
  driver generally have to manage everything themselves.

With drivers being written for new high IOPS devices, the classic
request_fn based driver doesn't work well enough. The design dates
back to when both SMP and high IOPS was rare. It has problems with
scaling to bigger machines, and runs into scaling issues even on
smaller machines when you have IOPS in the hundreds of thousands
per device.

The stacked approach is then most often selected as the model
for the driver. But this means that everybody has to re-invent
everything, and along with that we get all the problems again
that the shared approach solved.

This commit introduces blk-mq, block multi queue support. The
design is centered around per-cpu queues for queueing IO, which
then funnel down into x number of hardware submission queues.
We might have a 1:1 mapping between the two, or it might be
an N:M mapping. That all depends on what the hardware supports.

blk-mq provides various helper functions, which include:

- Scalable support for request tagging. Most devices need to
  be able to uniquely identify a request both in the driver and
  to the hardware. The tagging uses per-cpu caches for freed
  tags, to enable cache hot reuse.

- Timeout handling without tracking request on a per-device
  basis. Basically the driver should be able to get a notification,
  if a request happens to fail.

- Optional support for non 1:1 mappings between issue and
  submission queues. blk-mq can redirect IO completions to the
  desired location.

- Support for per-request payloads. Drivers almost always need
  to associate a request structure with some driver private
  command structure. Drivers can tell blk-mq this at init time,
  and then any request handed to the driver will have the
  required size of memory associated with it.

- Support for merging of IO, and plugging. The stacked model
  gets neither of these. Even for high IOPS devices, merging
  sequential IO reduces per-command overhead and thus
  increases bandwidth.

For now, this is provided as a potential 3rd queueing model, with
the hope being that, as it matures, it can replace both the classic
and stacked model. That would get us back to having just 1 real
model for block devices, leaving the stacked approach to dm/md
devices (as it was originally intended).

Contributions in this patch from the following people:

Shaohua Li <shli@fusionio.com>
Alexander Gordeev <agordeev@redhat.com>
Christoph Hellwig <hch@infradead.org>
Mike Christie <michaelc@cs.wisc.edu>
Matias Bjorling <m@bjorling.me>
Jeff Moyer <jmoyer@redhat.com>

Acked-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

1 files changed, 1480 insertions, 0 deletions

diff --git a/block/blk-mq.c b/block/blk-mq.c
new file mode 100644
index 00000000000..f21ec964e41
--- /dev/null
+++ b/block/blk-mq.c
@@ -0,0 +1,1480 @@
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/backing-dev.h>
+#include <linux/bio.h>
+#include <linux/blkdev.h>
+#include <linux/mm.h>
+#include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/workqueue.h>
+#include <linux/smp.h>
+#include <linux/llist.h>
+#include <linux/list_sort.h>
+#include <linux/cpu.h>
+#include <linux/cache.h>
+#include <linux/sched/sysctl.h>
+#include <linux/delay.h>
+
+#include <trace/events/block.h>
+
+#include <linux/blk-mq.h>
+#include "blk.h"
+#include "blk-mq.h"
+#include "blk-mq-tag.h"
+
+static DEFINE_MUTEX(all_q_mutex);
+static LIST_HEAD(all_q_list);
+
+static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx);
+
+DEFINE_PER_CPU(struct llist_head, ipi_lists);
+
+static struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
+					   unsigned int cpu)
+{
+	return per_cpu_ptr(q->queue_ctx, cpu);
+}
+
+/*
+ * This assumes per-cpu software queueing queues. They could be per-node
+ * as well, for instance. For now this is hardcoded as-is. Note that we don't
+ * care about preemption, since we know the ctx's are persistent. This does
+ * mean that we can't rely on ctx always matching the currently running CPU.
+ */
+static struct blk_mq_ctx *blk_mq_get_ctx(struct request_queue *q)
+{
+	return __blk_mq_get_ctx(q, get_cpu());
+}
+
+static void blk_mq_put_ctx(struct blk_mq_ctx *ctx)
+{
+	put_cpu();
+}
+
+/*
+ * Check if any of the ctx's have pending work in this hardware queue
+ */
+static bool blk_mq_hctx_has_pending(struct blk_mq_hw_ctx *hctx)
+{
+	unsigned int i;
+
+	for (i = 0; i < hctx->nr_ctx_map; i++)
+		if (hctx->ctx_map[i])
+			return true;
+
+	return false;
+}
+
+/*
+ * Mark this ctx as having pending work in this hardware queue
+ */
+static void blk_mq_hctx_mark_pending(struct blk_mq_hw_ctx *hctx,
+				     struct blk_mq_ctx *ctx)
+{
+	if (!test_bit(ctx->index_hw, hctx->ctx_map))
+		set_bit(ctx->index_hw, hctx->ctx_map);
+}
+
+static struct request *blk_mq_alloc_rq(struct blk_mq_hw_ctx *hctx, gfp_t gfp,
+				       bool reserved)
+{
+	struct request *rq;
+	unsigned int tag;
+
+	tag = blk_mq_get_tag(hctx->tags, gfp, reserved);
+	if (tag != BLK_MQ_TAG_FAIL) {
+		rq = hctx->rqs[tag];
+		rq->tag = tag;
+
+		return rq;
+	}
+
+	return NULL;
+}
+
+static int blk_mq_queue_enter(struct request_queue *q)
+{
+	int ret;
+
+	__percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
+	smp_wmb();
+	/* we have problems to freeze the queue if it's initializing */
+	if (!blk_queue_bypass(q) || !blk_queue_init_done(q))
+		return 0;
+
+	__percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
+
+	spin_lock_irq(q->queue_lock);
+	ret = wait_event_interruptible_lock_irq(q->mq_freeze_wq,
+		!blk_queue_bypass(q), *q->queue_lock);
+	/* inc usage with lock hold to avoid freeze_queue runs here */
+	if (!ret)
+		__percpu_counter_add(&q->mq_usage_counter, 1, 1000000);
+	spin_unlock_irq(q->queue_lock);
+
+	return ret;
+}
+
+static void blk_mq_queue_exit(struct request_queue *q)
+{
+	__percpu_counter_add(&q->mq_usage_counter, -1, 1000000);
+}
+
+/*
+ * Guarantee no request is in use, so we can change any data structure of
+ * the queue afterward.
+ */
+static void blk_mq_freeze_queue(struct request_queue *q)
+{
+	bool drain;
+
+	spin_lock_irq(q->queue_lock);
+	drain = !q->bypass_depth++;
+	queue_flag_set(QUEUE_FLAG_BYPASS, q);
+	spin_unlock_irq(q->queue_lock);
+
+	if (!drain)
+		return;
+
+	while (true) {
+		s64 count;
+
+		spin_lock_irq(q->queue_lock);
+		count = percpu_counter_sum(&q->mq_usage_counter);
+		spin_unlock_irq(q->queue_lock);
+
+		if (count == 0)
+			break;
+		blk_mq_run_queues(q, false);
+		msleep(10);
+	}
+}
+
+static void blk_mq_unfreeze_queue(struct request_queue *q)
+{
+	bool wake = false;
+
+	spin_lock_irq(q->queue_lock);
+	if (!--q->bypass_depth) {
+		queue_flag_clear(QUEUE_FLAG_BYPASS, q);
+		wake = true;
+	}
+	WARN_ON_ONCE(q->bypass_depth < 0);
+	spin_unlock_irq(q->queue_lock);
+	if (wake)
+		wake_up_all(&q->mq_freeze_wq);
+}
+
+bool blk_mq_can_queue(struct blk_mq_hw_ctx *hctx)
+{
+	return blk_mq_has_free_tags(hctx->tags);
+}
+EXPORT_SYMBOL(blk_mq_can_queue);
+
+static void blk_mq_rq_ctx_init(struct blk_mq_ctx *ctx, struct request *rq,
+			       unsigned int rw_flags)
+{
+	rq->mq_ctx = ctx;
+	rq->cmd_flags = rw_flags;
+	ctx->rq_dispatched[rw_is_sync(rw_flags)]++;
+}
+
+static struct request *__blk_mq_alloc_request(struct blk_mq_hw_ctx *hctx,
+					      gfp_t gfp, bool reserved)
+{
+	return blk_mq_alloc_rq(hctx, gfp, reserved);
+}
+
+static struct request *blk_mq_alloc_request_pinned(struct request_queue *q,
+						   int rw, gfp_t gfp,
+						   bool reserved)
+{
+	struct request *rq;
+
+	do {
+		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);
+		if (rq) {
+			blk_mq_rq_ctx_init(ctx, rq, rw);
+			break;
+		} else if (!(gfp & __GFP_WAIT))
+			break;
+
+		blk_mq_put_ctx(ctx);
+		__blk_mq_run_hw_queue(hctx);
+		blk_mq_wait_for_tags(hctx->tags);
+	} while (1);
+
+	return rq;
+}
+
+struct request *blk_mq_alloc_request(struct request_queue *q, int rw, gfp_t gfp)
+{
+	struct request *rq;
+
+	if (blk_mq_queue_enter(q))
+		return NULL;
+
+	rq = blk_mq_alloc_request_pinned(q, rw, gfp, false);
+	blk_mq_put_ctx(rq->mq_ctx);
+	return rq;
+}
+
+struct request *blk_mq_alloc_reserved_request(struct request_queue *q, int rw,
+					      gfp_t gfp)
+{
+	struct request *rq;
+
+	if (blk_mq_queue_enter(q))
+		return NULL;
+
+	rq = blk_mq_alloc_request_pinned(q, rw, gfp, true);
+	blk_mq_put_ctx(rq->mq_ctx);
+	return rq;
+}
+EXPORT_SYMBOL(blk_mq_alloc_reserved_request);
+
+/*
+ * Re-init and set pdu, if we have it
+ */
+static void blk_mq_rq_init(struct blk_mq_hw_ctx *hctx, struct request *rq)
+{
+	blk_rq_init(hctx->queue, rq);
+
+	if (hctx->cmd_size)
+		rq->special = blk_mq_rq_to_pdu(rq);
+}
+
+static void __blk_mq_free_request(struct blk_mq_hw_ctx *hctx,
+				  struct blk_mq_ctx *ctx, struct request *rq)
+{
+	const int tag = rq->tag;
+	struct request_queue *q = rq->q;
+
+	blk_mq_rq_init(hctx, rq);
+	blk_mq_put_tag(hctx->tags, tag);
+
+	blk_mq_queue_exit(q);
+}
+
+void blk_mq_free_request(struct request *rq)
+{
+	struct blk_mq_ctx *ctx = rq->mq_ctx;
+	struct blk_mq_hw_ctx *hctx;
+	struct request_queue *q = rq->q;
+
+	ctx->rq_completed[rq_is_sync(rq)]++;
+
+	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+	__blk_mq_free_request(hctx, ctx, rq);
+}
+
+static void blk_mq_bio_endio(struct request *rq, struct bio *bio, int error)
+{
+	if (error)
+		clear_bit(BIO_UPTODATE, &bio->bi_flags);
+	else if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
+		error = -EIO;
+
+	if (unlikely(rq->cmd_flags & REQ_QUIET))
+		set_bit(BIO_QUIET, &bio->bi_flags);
+
+	/* don't actually finish bio if it's part of flush sequence */
+	if (!(rq->cmd_flags & REQ_FLUSH_SEQ))
+		bio_endio(bio, error);
+}
+
+void blk_mq_complete_request(struct request *rq, int error)
+{
+	struct bio *bio = rq->bio;
+	unsigned int bytes = 0;
+
+	trace_block_rq_complete(rq->q, rq);
+
+	while (bio) {
+		struct bio *next = bio->bi_next;
+
+		bio->bi_next = NULL;
+		bytes += bio->bi_size;
+		blk_mq_bio_endio(rq, bio, error);
+		bio = next;
+	}
+
+	blk_account_io_completion(rq, bytes);
+
+	if (rq->end_io)
+		rq->end_io(rq, error);
+	else
+		blk_mq_free_request(rq);
+
+	blk_account_io_done(rq);
+}
+
+void __blk_mq_end_io(struct request *rq, int error)
+{
+	if (!blk_mark_rq_complete(rq))
+		blk_mq_complete_request(rq, error);
+}
+
+#if defined(CONFIG_SMP) && defined(CONFIG_USE_GENERIC_SMP_HELPERS)
+
+/*
+ * Called with interrupts disabled.
+ */
+static void ipi_end_io(void *data)
+{
+	struct llist_head *list = &per_cpu(ipi_lists, smp_processor_id());
+	struct llist_node *entry, *next;
+	struct request *rq;
+
+	entry = llist_del_all(list);
+
+	while (entry) {
+		next = entry->next;
+		rq = llist_entry(entry, struct request, ll_list);
+		__blk_mq_end_io(rq, rq->errors);
+		entry = next;
+	}
+}
+
+static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
+			  struct request *rq, const int error)
+{
+	struct call_single_data *data = &rq->csd;
+
+	rq->errors = error;
+	rq->ll_list.next = NULL;
+
+	/*
+	 * If the list is non-empty, an existing IPI must already
+	 * be "in flight". If that is the case, we need not schedule
+	 * a new one.
+	 */
+	if (llist_add(&rq->ll_list, &per_cpu(ipi_lists, ctx->cpu))) {
+		data->func = ipi_end_io;
+		data->flags = 0;
+		__smp_call_function_single(ctx->cpu, data, 0);
+	}
+
+	return true;
+}
+#else /* CONFIG_SMP && CONFIG_USE_GENERIC_SMP_HELPERS */
+static int ipi_remote_cpu(struct blk_mq_ctx *ctx, const int cpu,
+			  struct request *rq, const int error)
+{
+	return false;
+}
+#endif
+
+/*
+ * End IO on this request on a multiqueue enabled driver. We'll either do
+ * it directly inline, or punt to a local IPI handler on the matching
+ * remote CPU.
+ */
+void blk_mq_end_io(struct request *rq, int error)
+{
+	struct blk_mq_ctx *ctx = rq->mq_ctx;
+	int cpu;
+
+	if (!ctx->ipi_redirect)
+		return __blk_mq_end_io(rq, error);
+
+	cpu = get_cpu();
+
+	if (cpu == ctx->cpu || !cpu_online(ctx->cpu) ||
+	    !ipi_remote_cpu(ctx, cpu, rq, error))
+		__blk_mq_end_io(rq, error);
+
+	put_cpu();
+}
+EXPORT_SYMBOL(blk_mq_end_io);
+
+static void blk_mq_start_request(struct request *rq)
+{
+	struct request_queue *q = rq->q;
+
+	trace_block_rq_issue(q, rq);
+
+	/*
+	 * Just mark start time and set the started bit. Due to memory
+	 * ordering, we know we'll see the correct deadline as long as
+	 * REQ_ATOMIC_STARTED is seen.
+	 */
+	rq->deadline = jiffies + q->rq_timeout;
+	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
+}
+
+static void blk_mq_requeue_request(struct request *rq)
+{
+	struct request_queue *q = rq->q;
+
+	trace_block_rq_requeue(q, rq);
+	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
+}
+
+struct blk_mq_timeout_data {
+	struct blk_mq_hw_ctx *hctx;
+	unsigned long *next;
+	unsigned int *next_set;
+};
+
+static void blk_mq_timeout_check(void *__data, unsigned long *free_tags)
+{
+	struct blk_mq_timeout_data *data = __data;
+	struct blk_mq_hw_ctx *hctx = data->hctx;
+	unsigned int tag;
+
+	 /* It may not be in flight yet (this is where
+	 * the REQ_ATOMIC_STARTED flag comes in). The requests are
+	 * statically allocated, so we know it's always safe to access the
+	 * memory associated with a bit offset into ->rqs[].
+	 */
+	tag = 0;
+	do {
+		struct request *rq;
+
+		tag = find_next_zero_bit(free_tags, hctx->queue_depth, tag);
+		if (tag >= hctx->queue_depth)
+			break;
+
+		rq = hctx->rqs[tag++];
+
+		if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
+			continue;
+
+		blk_rq_check_expired(rq, data->next, data->next_set);
+	} while (1);
+}
+
+static void blk_mq_hw_ctx_check_timeout(struct blk_mq_hw_ctx *hctx,
+					unsigned long *next,
+					unsigned int *next_set)
+{
+	struct blk_mq_timeout_data data = {
+		.hctx		= hctx,
+		.next		= next,
+		.next_set	= next_set,
+	};
+
+	/*
+	 * Ask the tagging code to iterate busy requests, so we can
+	 * check them for timeout.
+	 */
+	blk_mq_tag_busy_iter(hctx->tags, blk_mq_timeout_check, &data);
+}
+
+static void blk_mq_rq_timer(unsigned long data)
+{
+	struct request_queue *q = (struct request_queue *) data;
+	struct blk_mq_hw_ctx *hctx;
+	unsigned long next = 0;
+	int i, next_set = 0;
+
+	queue_for_each_hw_ctx(q, hctx, i)
+		blk_mq_hw_ctx_check_timeout(hctx, &next, &next_set);
+
+	if (next_set)
+		mod_timer(&q->timeout, round_jiffies_up(next));
+}
+
+/*
+ * Reverse check our software queue for entries that we could potentially
+ * merge with. Currently includes a hand-wavy stop count of 8, to not spend
+ * too much time checking for merges.
+ */
+static bool blk_mq_attempt_merge(struct request_queue *q,
+				 struct blk_mq_ctx *ctx, struct bio *bio)
+{
+	struct request *rq;
+	int checked = 8;
+
+	list_for_each_entry_reverse(rq, &ctx->rq_list, queuelist) {
+		int el_ret;
+
+		if (!checked--)
+			break;
+
+		if (!blk_rq_merge_ok(rq, bio))
+			continue;
+
+		el_ret = blk_try_merge(rq, bio);
+		if (el_ret == ELEVATOR_BACK_MERGE) {
+			if (bio_attempt_back_merge(q, rq, bio)) {
+				ctx->rq_merged++;
+				return true;
+			}
+			break;
+		} else if (el_ret == ELEVATOR_FRONT_MERGE) {
+			if (bio_attempt_front_merge(q, rq, bio)) {
+				ctx->rq_merged++;
+				return true;
+			}
+			break;
+		}
+	}
+
+	return false;
+}
+
+void blk_mq_add_timer(struct request *rq)
+{
+	__blk_add_timer(rq, NULL);
+}
+
+/*
+ * Run this hardware queue, pulling any software queues mapped to it in.
+ * Note that this function currently has various problems around ordering
+ * of IO. In particular, we'd like FIFO behaviour on handling existing
+ * items on the hctx->dispatch list. Ignore that for now.
+ */
+static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
+{
+	struct request_queue *q = hctx->queue;
+	struct blk_mq_ctx *ctx;
+	struct request *rq;
+	LIST_HEAD(rq_list);
+	int bit, queued;
+
+	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
+		return;
+
+	hctx->run++;
+
+	/*
+	 * Touch any software queue that has pending entries.
+	 */
+	for_each_set_bit(bit, hctx->ctx_map, hctx->nr_ctx) {
+		clear_bit(bit, hctx->ctx_map);
+		ctx = hctx->ctxs[bit];
+		BUG_ON(bit != ctx->index_hw);
+
+		spin_lock(&ctx->lock);
+		list_splice_tail_init(&ctx->rq_list, &rq_list);
+		spin_unlock(&ctx->lock);
+	}
+
+	/*
+	 * If we have previous entries on our dispatch list, grab them
+	 * and stuff them at the front for more fair dispatch.
+	 */
+	if (!list_empty_careful(&hctx->dispatch)) {
+		spin_lock(&hctx->lock);
+		if (!list_empty(&hctx->dispatch))
+			list_splice_init(&hctx->dispatch, &rq_list);
+		spin_unlock(&hctx->lock);
+	}
+
+	/*
+	 * Delete and return all entries from our dispatch list
+	 */
+	queued = 0;
+
+	/*
+	 * Now process all the entries, sending them to the driver.
+	 */
+	while (!list_empty(&rq_list)) {
+		int ret;
+
+		rq = list_first_entry(&rq_list, struct request, queuelist);
+		list_del_init(&rq->queuelist);
+		blk_mq_start_request(rq);
+
+		/*
+		 * Last request in the series. Flag it as such, this
+		 * enables drivers to know when IO should be kicked off,
+		 * if they don't do it on a per-request basis.
+		 *
+		 * Note: the flag isn't the only condition drivers
+		 * should do kick off. If drive is busy, the last
+		 * request might not have the bit set.
+		 */
+		if (list_empty(&rq_list))
+			rq->cmd_flags |= REQ_END;
+
+		ret = q->mq_ops->queue_rq(hctx, rq);
+		switch (ret) {
+		case BLK_MQ_RQ_QUEUE_OK:
+			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;
+		default:
+			pr_err("blk-mq: bad return on queue: %d\n", ret);
+			rq->errors = -EIO;
+		case BLK_MQ_RQ_QUEUE_ERROR:
+			blk_mq_end_io(rq, rq->errors);
+			break;
+		}
+
+		if (ret == BLK_MQ_RQ_QUEUE_BUSY)
+			break;
+	}
+
+	if (!queued)
+		hctx->dispatched[0]++;
+	else if (queued < (1 << (BLK_MQ_MAX_DISPATCH_ORDER - 1)))
+		hctx->dispatched[ilog2(queued) + 1]++;
+
+	/*
+	 * Any items that need requeuing? Stuff them into hctx->dispatch,
+	 * that is where we will continue on next queue run.
+	 */
+	if (!list_empty(&rq_list)) {
+		spin_lock(&hctx->lock);
+		list_splice(&rq_list, &hctx->dispatch);
+		spin_unlock(&hctx->lock);
+	}
+}
+
+void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
+{
+	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->flags)))
+		return;
+
+	if (!async)
+		__blk_mq_run_hw_queue(hctx);
+	else {
+		struct request_queue *q = hctx->queue;
+
+		kblockd_schedule_delayed_work(q, &hctx->delayed_work, 0);
+	}
+}
+
+void blk_mq_run_queues(struct request_queue *q, bool async)
+{
+	struct blk_mq_hw_ctx *hctx;
+	int i;
+
+	queue_for_each_hw_ctx(q, hctx, i) {
+		if ((!blk_mq_hctx_has_pending(hctx) &&
+		    list_empty_careful(&hctx->dispatch)) ||
+		    test_bit(BLK_MQ_S_STOPPED, &hctx->flags))
+			continue;
+
+		blk_mq_run_hw_queue(hctx, async);
+	}
+}
+EXPORT_SYMBOL(blk_mq_run_queues);
+
+void blk_mq_stop_hw_queue(struct blk_mq_hw_ctx *hctx)
+{
+	cancel_delayed_work(&hctx->delayed_work);
+	set_bit(BLK_MQ_S_STOPPED, &hctx->state);
+}
+EXPORT_SYMBOL(blk_mq_stop_hw_queue);
+
+void blk_mq_start_hw_queue(struct blk_mq_hw_ctx *hctx)
+{
+	clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
+	__blk_mq_run_hw_queue(hctx);
+}
+EXPORT_SYMBOL(blk_mq_start_hw_queue);
+
+void blk_mq_start_stopped_hw_queues(struct request_queue *q)
+{
+	struct blk_mq_hw_ctx *hctx;
+	int i;
+
+	queue_for_each_hw_ctx(q, hctx, i) {
+		if (!test_bit(BLK_MQ_S_STOPPED, &hctx->state))
+			continue;
+
+		clear_bit(BLK_MQ_S_STOPPED, &hctx->state);
+		blk_mq_run_hw_queue(hctx, true);
+	}
+}
+EXPORT_SYMBOL(blk_mq_start_stopped_hw_queues);
+
+static void blk_mq_work_fn(struct work_struct *work)
+{
+	struct blk_mq_hw_ctx *hctx;
+
+	hctx = container_of(work, struct blk_mq_hw_ctx, delayed_work.work);
+	__blk_mq_run_hw_queue(hctx);
+}
+
+static void __blk_mq_insert_request(struct blk_mq_hw_ctx *hctx,
+				    struct request *rq)
+{
+	struct blk_mq_ctx *ctx = rq->mq_ctx;
+
+	list_add_tail(&rq->queuelist, &ctx->rq_list);
+	blk_mq_hctx_mark_pending(hctx, ctx);
+
+	/*
+	 * We do this early, to ensure we are on the right CPU.
+	 */
+	blk_mq_add_timer(rq);
+}
+
+void blk_mq_insert_request(struct request_queue *q, struct request *rq,
+			   bool run_queue)
+{
+	struct blk_mq_hw_ctx *hctx;
+	struct blk_mq_ctx *ctx, *current_ctx;
+
+	ctx = rq->mq_ctx;
+	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+
+	if (rq->cmd_flags & (REQ_FLUSH | REQ_FUA)) {
+		blk_insert_flush(rq);
+	} else {
+		current_ctx = blk_mq_get_ctx(q);
+
+		if (!cpu_online(ctx->cpu)) {
+			ctx = current_ctx;
+			hctx = q->mq_ops->map_queue(q, ctx->cpu);
+			rq->mq_ctx = ctx;
+		}
+		spin_lock(&ctx->lock);
+		__blk_mq_insert_request(hctx, rq);
+		spin_unlock(&ctx->lock);
+
+		blk_mq_put_ctx(current_ctx);
+	}
+
+	if (run_queue)
+		__blk_mq_run_hw_queue(hctx);
+}
+EXPORT_SYMBOL(blk_mq_insert_request);
+
+/*
+ * This is a special version of blk_mq_insert_request to bypass FLUSH request
+ * check. Should only be used internally.
+ */
+void blk_mq_run_request(struct request *rq, bool run_queue, bool async)
+{
+	struct request_queue *q = rq->q;
+	struct blk_mq_hw_ctx *hctx;
+	struct blk_mq_ctx *ctx, *current_ctx;
+
+	current_ctx = blk_mq_get_ctx(q);
+
+	ctx = rq->mq_ctx;
+	if (!cpu_online(ctx->cpu)) {
+		ctx = current_ctx;
+		rq->mq_ctx = ctx;
+	}
+	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+
+	/* ctx->cpu might be offline */
+	spin_lock(&ctx->lock);
+	__blk_mq_insert_request(hctx, rq);
+	spin_unlock(&ctx->lock);
+
+	blk_mq_put_ctx(current_ctx);
+
+	if (run_queue)
+		blk_mq_run_hw_queue(hctx, async);
+}
+
+static void blk_mq_insert_requests(struct request_queue *q,
+				     struct blk_mq_ctx *ctx,
+				     struct list_head *list,
+				     int depth,
+				     bool from_schedule)
+
+{
+	struct blk_mq_hw_ctx *hctx;
+	struct blk_mq_ctx *current_ctx;
+
+	trace_block_unplug(q, depth, !from_schedule);
+
+	current_ctx = blk_mq_get_ctx(q);
+
+	if (!cpu_online(ctx->cpu))
+		ctx = current_ctx;
+	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+
+	/*
+	 * preemption doesn't flush plug list, so it's possible ctx->cpu is
+	 * offline now
+	 */
+	spin_lock(&ctx->lock);
+	while (!list_empty(list)) {
+		struct request *rq;
+
+		rq = list_first_entry(list, struct request, queuelist);
+		list_del_init(&rq->queuelist);
+		rq->mq_ctx = ctx;
+		__blk_mq_insert_request(hctx, rq);
+	}
+	spin_unlock(&ctx->lock);
+
+	blk_mq_put_ctx(current_ctx);
+
+	blk_mq_run_hw_queue(hctx, from_schedule);
+}
+
+static int plug_ctx_cmp(void *priv, struct list_head *a, struct list_head *b)
+{
+	struct request *rqa = container_of(a, struct request, queuelist);
+	struct request *rqb = container_of(b, struct request, queuelist);
+
+	return !(rqa->mq_ctx < rqb->mq_ctx ||
+		 (rqa->mq_ctx == rqb->mq_ctx &&
+		  blk_rq_pos(rqa) < blk_rq_pos(rqb)));
+}
+
+void blk_mq_flush_plug_list(struct blk_plug *plug, bool from_schedule)
+{
+	struct blk_mq_ctx *this_ctx;
+	struct request_queue *this_q;
+	struct request *rq;
+	LIST_HEAD(list);
+	LIST_HEAD(ctx_list);
+	unsigned int depth;
+
+	list_splice_init(&plug->mq_list, &list);
+
+	list_sort(NULL, &list, plug_ctx_cmp);
+
+	this_q = NULL;
+	this_ctx = NULL;
+	depth = 0;
+
+	while (!list_empty(&list)) {
+		rq = list_entry_rq(list.next);
+		list_del_init(&rq->queuelist);
+		BUG_ON(!rq->q);
+		if (rq->mq_ctx != this_ctx) {
+			if (this_ctx) {
+				blk_mq_insert_requests(this_q, this_ctx,
+							&ctx_list, depth,
+							from_schedule);
+			}
+
+			this_ctx = rq->mq_ctx;
+			this_q = rq->q;
+			depth = 0;
+		}
+
+		depth++;
+		list_add_tail(&rq->queuelist, &ctx_list);
+	}
+
+	/*
+	 * If 'this_ctx' is set, we know we have entries to complete
+	 * on 'ctx_list'. Do those.
+	 */
+	if (this_ctx) {
+		blk_mq_insert_requests(this_q, this_ctx, &ctx_list, depth,
+				       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);
+}
+
+static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
+{
+	struct blk_mq_hw_ctx *hctx;
+	struct blk_mq_ctx *ctx;
+	const int is_sync = rw_is_sync(bio->bi_rw);
+	const int is_flush_fua = bio->bi_rw & (REQ_FLUSH | REQ_FUA);
+	int rw = bio_data_dir(bio);
+	struct request *rq;
+	unsigned int use_plug, request_count = 0;
+
+	/*
+	 * If we have multiple hardware queues, just go directly to
+	 * one of those for sync IO.
+	 */
+	use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);
+
+	blk_queue_bounce(q, &bio);
+
+	if (use_plug && blk_attempt_plug_merge(q, bio, &request_count))
+		return;
+
+	if (blk_mq_queue_enter(q)) {
+		bio_endio(bio, -EIO);
+		return;
+	}
+
+	ctx = blk_mq_get_ctx(q);
+	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+
+	trace_block_getrq(q, bio, rw);
+	rq = __blk_mq_alloc_request(hctx, GFP_ATOMIC, false);
+	if (likely(rq))
+		blk_mq_rq_ctx_init(ctx, rq, rw);
+	else {
+		blk_mq_put_ctx(ctx);
+		trace_block_sleeprq(q, bio, rw);
+		rq = blk_mq_alloc_request_pinned(q, rw, __GFP_WAIT|GFP_ATOMIC,
+							false);
+		ctx = rq->mq_ctx;
+		hctx = q->mq_ops->map_queue(q, ctx->cpu);
+	}
+
+	hctx->queued++;
+
+	if (unlikely(is_flush_fua)) {
+		blk_mq_bio_to_request(rq, bio);
+		blk_mq_put_ctx(ctx);
+		blk_insert_flush(rq);
+		goto run_queue;
+	}
+
+	/*
+	 * A task plug currently exists. Since this is completely lockless,
+	 * utilize that to temporarily store requests until the task is
+	 * either done or scheduled away.
+	 */
+	if (use_plug) {
+		struct blk_plug *plug = current->plug;
+
+		if (plug) {
+			blk_mq_bio_to_request(rq, bio);
+			if (list_empty(&plug->list))
+				trace_block_plug(q);
+			else if (request_count >= BLK_MAX_REQUEST_COUNT) {
+				blk_flush_plug_list(plug, false);
+				trace_block_plug(q);
+			}
+			list_add_tail(&rq->queuelist, &plug->mq_list);
+			blk_mq_put_ctx(ctx);
+			return;
+		}
+	}
+
+	spin_lock(&ctx->lock);
+
+	if ((hctx->flags & BLK_MQ_F_SHOULD_MERGE) &&
+	    blk_mq_attempt_merge(q, ctx, bio))
+		__blk_mq_free_request(hctx, ctx, rq);
+	else {
+		blk_mq_bio_to_request(rq, bio);
+		__blk_mq_insert_request(hctx, rq);
+	}
+
+	spin_unlock(&ctx->lock);
+	blk_mq_put_ctx(ctx);
+
+	/*
+	 * For a SYNC request, send it to the hardware immediately. For an
+	 * ASYNC request, just ensure that we run it later on. The latter
+	 * allows for merging opportunities and more efficient dispatching.
+	 */
+run_queue:
+	blk_mq_run_hw_queue(hctx, !is_sync || is_flush_fua);
+}
+
+/*
+ * Default mapping to a software queue, since we use one per CPU.
+ */
+struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q, const int cpu)
+{
+	return q->queue_hw_ctx[q->mq_map[cpu]];
+}
+EXPORT_SYMBOL(blk_mq_map_queue);
+
+struct blk_mq_hw_ctx *blk_mq_alloc_single_hw_queue(struct blk_mq_reg *reg,
+						   unsigned int hctx_index)
+{
+	return kmalloc_node(sizeof(struct blk_mq_hw_ctx),
+				GFP_KERNEL | __GFP_ZERO, reg->numa_node);
+}
+EXPORT_SYMBOL(blk_mq_alloc_single_hw_queue);
+
+void blk_mq_free_single_hw_queue(struct blk_mq_hw_ctx *hctx,
+				 unsigned int hctx_index)
+{
+	kfree(hctx);
+}
+EXPORT_SYMBOL(blk_mq_free_single_hw_queue);
+
+static void blk_mq_hctx_notify(void *data, unsigned long action,
+			       unsigned int cpu)
+{
+	struct blk_mq_hw_ctx *hctx = data;
+	struct blk_mq_ctx *ctx;
+	LIST_HEAD(tmp);
+
+	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
+		return;
+
+	/*
+	 * Move ctx entries to new CPU, if this one is going away.
+	 */
+	ctx = __blk_mq_get_ctx(hctx->queue, cpu);
+
+	spin_lock(&ctx->lock);
+	if (!list_empty(&ctx->rq_list)) {
+		list_splice_init(&ctx->rq_list, &tmp);
+		clear_bit(ctx->index_hw, hctx->ctx_map);
+	}
+	spin_unlock(&ctx->lock);
+
+	if (list_empty(&tmp))
+		return;
+
+	ctx = blk_mq_get_ctx(hctx->queue);
+	spin_lock(&ctx->lock);
+
+	while (!list_empty(&tmp)) {
+		struct request *rq;
+
+		rq = list_first_entry(&tmp, struct request, queuelist);
+		rq->mq_ctx = ctx;
+		list_move_tail(&rq->queuelist, &ctx->rq_list);
+	}
+
+	blk_mq_hctx_mark_pending(hctx, ctx);
+
+	spin_unlock(&ctx->lock);
+	blk_mq_put_ctx(ctx);
+}
+
+static void blk_mq_init_hw_commands(struct blk_mq_hw_ctx *hctx,
+				    void (*init)(void *, struct blk_mq_hw_ctx *,
+					struct request *, unsigned int),
+				    void *data)
+{
+	unsigned int i;
+
+	for (i = 0; i < hctx->queue_depth; i++) {
+		struct request *rq = hctx->rqs[i];
+
+		init(data, hctx, rq, i);
+	}
+}
+
+void blk_mq_init_commands(struct request_queue *q,
+			  void (*init)(void *, struct blk_mq_hw_ctx *,
+					struct request *, unsigned int),
+			  void *data)
+{
+	struct blk_mq_hw_ctx *hctx;
+	unsigned int i;
+
+	queue_for_each_hw_ctx(q, hctx, i)
+		blk_mq_init_hw_commands(hctx, init, data);
+}
+EXPORT_SYMBOL(blk_mq_init_commands);
+
+static void blk_mq_free_rq_map(struct blk_mq_hw_ctx *hctx)
+{
+	struct page *page;
+
+	while (!list_empty(&hctx->page_list)) {
+		page = list_first_entry(&hctx->page_list, struct page, list);
+		list_del_init(&page->list);
+		__free_pages(page, page->private);
+	}
+
+	kfree(hctx->rqs);
+
+	if (hctx->tags)
+		blk_mq_free_tags(hctx->tags);
+}
+
+static size_t order_to_size(unsigned int order)
+{
+	size_t ret = PAGE_SIZE;
+
+	while (order--)
+		ret *= 2;
+
+	return ret;
+}
+
+static int blk_mq_init_rq_map(struct blk_mq_hw_ctx *hctx,
+			      unsigned int reserved_tags, int node)
+{
+	unsigned int i, j, entries_per_page, max_order = 4;
+	size_t rq_size, left;
+
+	INIT_LIST_HEAD(&hctx->page_list);
+
+	hctx->rqs = kmalloc_node(hctx->queue_depth * sizeof(struct request *),
+					GFP_KERNEL, node);
+	if (!hctx->rqs)
+		return -ENOMEM;
+
+	/*
+	 * rq_size is the size of the request plus driver payload, rounded
+	 * to the cacheline size
+	 */
+	rq_size = round_up(sizeof(struct request) + hctx->cmd_size,
+				cache_line_size());
+	left = rq_size * hctx->queue_depth;
+
+	for (i = 0; i < hctx->queue_depth;) {
+		int this_order = max_order;
+		struct page *page;
+		int to_do;
+		void *p;
+
+		while (left < order_to_size(this_order - 1) && this_order)
+			this_order--;
+
+		do {
+			page = alloc_pages_node(node, GFP_KERNEL, this_order);
+			if (page)
+				break;
+			if (!this_order--)
+				break;
+			if (order_to_size(this_order) < rq_size)
+				break;
+		} while (1);
+
+		if (!page)
+			break;
+
+		page->private = this_order;
+		list_add_tail(&page->list, &hctx->page_list);
+
+		p = page_address(page);
+		entries_per_page = order_to_size(this_order) / rq_size;
+		to_do = min(entries_per_page, hctx->queue_depth - i);
+		left -= to_do * rq_size;
+		for (j = 0; j < to_do; j++) {
+			hctx->rqs[i] = p;
+			blk_mq_rq_init(hctx, hctx->rqs[i]);
+			p += rq_size;
+			i++;
+		}
+	}
+
+	if (i < (reserved_tags + BLK_MQ_TAG_MIN))
+		goto err_rq_map;
+	else if (i != hctx->queue_depth) {
+		hctx->queue_depth = i;
+		pr_warn("%s: queue depth set to %u because of low memory\n",
+					__func__, i);
+	}
+
+	hctx->tags = blk_mq_init_tags(hctx->queue_depth, reserved_tags, node);
+	if (!hctx->tags) {
+err_rq_map:
+		blk_mq_free_rq_map(hctx);
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+
+static int blk_mq_init_hw_queues(struct request_queue *q,
+				 struct blk_mq_reg *reg, void *driver_data)
+{
+	struct blk_mq_hw_ctx *hctx;
+	unsigned int i, j;
+
+	/*
+	 * Initialize hardware queues
+	 */
+	queue_for_each_hw_ctx(q, hctx, i) {
+		unsigned int num_maps;
+		int node;
+
+		node = hctx->numa_node;
+		if (node == NUMA_NO_NODE)
+			node = hctx->numa_node = reg->numa_node;
+
+		INIT_DELAYED_WORK(&hctx->delayed_work, blk_mq_work_fn);
+		spin_lock_init(&hctx->lock);
+		INIT_LIST_HEAD(&hctx->dispatch);
+		hctx->queue = q;
+		hctx->queue_num = i;
+		hctx->flags = reg->flags;
+		hctx->queue_depth = reg->queue_depth;
+		hctx->cmd_size = reg->cmd_size;
+
+		blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
+						blk_mq_hctx_notify, hctx);
+		blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
+
+		if (blk_mq_init_rq_map(hctx, reg->reserved_tags, node))
+			break;
+
+		/*
+		 * Allocate space for all possible cpus to avoid allocation in
+		 * runtime
+		 */
+		hctx->ctxs = kmalloc_node(nr_cpu_ids * sizeof(void *),
+						GFP_KERNEL, node);
+		if (!hctx->ctxs)
+			break;
+
+		num_maps = ALIGN(nr_cpu_ids, BITS_PER_LONG) / BITS_PER_LONG;
+		hctx->ctx_map = kzalloc_node(num_maps * sizeof(unsigned long),
+						GFP_KERNEL, node);
+		if (!hctx->ctx_map)
+			break;
+
+		hctx->nr_ctx_map = num_maps;
+		hctx->nr_ctx = 0;
+
+		if (reg->ops->init_hctx &&
+		    reg->ops->init_hctx(hctx, driver_data, i))
+			break;
+	}
+
+	if (i == q->nr_hw_queues)
+		return 0;
+
+	/*
+	 * Init failed
+	 */
+	queue_for_each_hw_ctx(q, hctx, j) {
+		if (i == j)
+			break;
+
+		if (reg->ops->exit_hctx)
+			reg->ops->exit_hctx(hctx, j);
+
+		blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
+		blk_mq_free_rq_map(hctx);
+		kfree(hctx->ctxs);
+	}
+
+	return 1;
+}
+
+static void blk_mq_init_cpu_queues(struct request_queue *q,
+				   unsigned int nr_hw_queues)
+{
+	unsigned int i;
+
+	for_each_possible_cpu(i) {
+		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
+		struct blk_mq_hw_ctx *hctx;
+
+		memset(__ctx, 0, sizeof(*__ctx));
+		__ctx->cpu = i;
+		spin_lock_init(&__ctx->lock);
+		INIT_LIST_HEAD(&__ctx->rq_list);
+		__ctx->queue = q;
+
+		/* If the cpu isn't online, the cpu is mapped to first hctx */
+		hctx = q->mq_ops->map_queue(q, i);
+		hctx->nr_ctx++;
+
+		if (!cpu_online(i))
+			continue;
+
+		/*
+		 * Set local node, IFF we have more than one hw queue. If
+		 * not, we remain on the home node of the device
+		 */
+		if (nr_hw_queues > 1 && hctx->numa_node == NUMA_NO_NODE)
+			hctx->numa_node = cpu_to_node(i);
+	}
+}
+
+static void blk_mq_map_swqueue(struct request_queue *q)
+{
+	unsigned int i;
+	struct blk_mq_hw_ctx *hctx;
+	struct blk_mq_ctx *ctx;
+
+	queue_for_each_hw_ctx(q, hctx, i) {
+		hctx->nr_ctx = 0;
+	}
+
+	/*
+	 * Map software to hardware queues
+	 */
+	queue_for_each_ctx(q, ctx, i) {
+		/* If the cpu isn't online, the cpu is mapped to first hctx */
+		hctx = q->mq_ops->map_queue(q, i);
+		ctx->index_hw = hctx->nr_ctx;
+		hctx->ctxs[hctx->nr_ctx++] = ctx;
+	}
+}
+
+struct request_queue *blk_mq_init_queue(struct blk_mq_reg *reg,
+					void *driver_data)
+{
+	struct blk_mq_hw_ctx **hctxs;
+	struct blk_mq_ctx *ctx;
+	struct request_queue *q;
+	int i;
+
+	if (!reg->nr_hw_queues ||
+	    !reg->ops->queue_rq || !reg->ops->map_queue ||
+	    !reg->ops->alloc_hctx || !reg->ops->free_hctx)
+		return ERR_PTR(-EINVAL);
+
+	if (!reg->queue_depth)
+		reg->queue_depth = BLK_MQ_MAX_DEPTH;
+	else if (reg->queue_depth > BLK_MQ_MAX_DEPTH) {
+		pr_err("blk-mq: queuedepth too large (%u)\n", reg->queue_depth);
+		reg->queue_depth = BLK_MQ_MAX_DEPTH;
+	}
+
+	if (reg->queue_depth < (reg->reserved_tags + BLK_MQ_TAG_MIN))
+		return ERR_PTR(-EINVAL);
+
+	ctx = alloc_percpu(struct blk_mq_ctx);
+	if (!ctx)
+		return ERR_PTR(-ENOMEM);
+
+	hctxs = kmalloc_node(reg->nr_hw_queues * sizeof(*hctxs), GFP_KERNEL,
+			reg->numa_node);
+
+	if (!hctxs)
+		goto err_percpu;
+
+	for (i = 0; i < reg->nr_hw_queues; i++) {
+		hctxs[i] = reg->ops->alloc_hctx(reg, i);
+		if (!hctxs[i])
+			goto err_hctxs;
+
+		hctxs[i]->numa_node = NUMA_NO_NODE;
+		hctxs[i]->queue_num = i;
+	}
+
+	q = blk_alloc_queue_node(GFP_KERNEL, reg->numa_node);
+	if (!q)
+		goto err_hctxs;
+
+	q->mq_map = blk_mq_make_queue_map(reg);
+	if (!q->mq_map)
+		goto err_map;
+
+	setup_timer(&q->timeout, blk_mq_rq_timer, (unsigned long) q);
+	blk_queue_rq_timeout(q, 30000);
+
+	q->nr_queues = nr_cpu_ids;
+	q->nr_hw_queues = reg->nr_hw_queues;
+
+	q->queue_ctx = ctx;
+	q->queue_hw_ctx = hctxs;
+
+	q->mq_ops = reg->ops;
+
+	blk_queue_make_request(q, blk_mq_make_request);
+	blk_queue_rq_timed_out(q, reg->ops->timeout);
+	if (reg->timeout)
+		blk_queue_rq_timeout(q, reg->timeout);
+
+	blk_mq_init_flush(q);
+	blk_mq_init_cpu_queues(q, reg->nr_hw_queues);
+
+	if (blk_mq_init_hw_queues(q, reg, driver_data))
+		goto err_hw;
+
+	blk_mq_map_swqueue(q);
+
+	mutex_lock(&all_q_mutex);
+	list_add_tail(&q->all_q_node, &all_q_list);
+	mutex_unlock(&all_q_mutex);
+
+	return q;
+err_hw:
+	kfree(q->mq_map);
+err_map:
+	blk_cleanup_queue(q);
+err_hctxs:
+	for (i = 0; i < reg->nr_hw_queues; i++) {
+		if (!hctxs[i])
+			break;
+		reg->ops->free_hctx(hctxs[i], i);
+	}
+	kfree(hctxs);
+err_percpu:
+	free_percpu(ctx);
+	return ERR_PTR(-ENOMEM);
+}
+EXPORT_SYMBOL(blk_mq_init_queue);
+
+void blk_mq_free_queue(struct request_queue *q)
+{
+	struct blk_mq_hw_ctx *hctx;
+	int i;
+
+	queue_for_each_hw_ctx(q, hctx, i) {
+		cancel_delayed_work_sync(&hctx->delayed_work);
+		kfree(hctx->ctx_map);
+		kfree(hctx->ctxs);
+		blk_mq_free_rq_map(hctx);
+		blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
+		if (q->mq_ops->exit_hctx)
+			q->mq_ops->exit_hctx(hctx, i);
+		q->mq_ops->free_hctx(hctx, i);
+	}
+
+	free_percpu(q->queue_ctx);
+	kfree(q->queue_hw_ctx);
+	kfree(q->mq_map);
+
+	q->queue_ctx = NULL;
+	q->queue_hw_ctx = NULL;
+	q->mq_map = NULL;
+
+	mutex_lock(&all_q_mutex);
+	list_del_init(&q->all_q_node);
+	mutex_unlock(&all_q_mutex);
+}
+EXPORT_SYMBOL(blk_mq_free_queue);
+
+/* Basically redo blk_mq_init_queue with queue frozen */
+static void __cpuinit blk_mq_queue_reinit(struct request_queue *q)
+{
+	blk_mq_freeze_queue(q);
+
+	blk_mq_update_queue_map(q->mq_map, q->nr_hw_queues);
+
+	/*
+	 * redo blk_mq_init_cpu_queues and blk_mq_init_hw_queues. FIXME: maybe
+	 * we should change hctx numa_node according to new topology (this
+	 * involves free and re-allocate memory, worthy doing?)
+	 */
+
+	blk_mq_map_swqueue(q);
+
+	blk_mq_unfreeze_queue(q);
+}
+
+static int __cpuinit blk_mq_queue_reinit_notify(struct notifier_block *nb,
+		unsigned long action, void *hcpu)
+{
+	struct request_queue *q;
+
+	/*
+	 * Before new mapping is established, hotadded cpu might already start
+	 * handling requests. This doesn't break anything as we map offline
+	 * CPUs to first hardware queue. We will re-init queue below to get
+	 * optimal settings.
+	 */
+	if (action != CPU_DEAD && action != CPU_DEAD_FROZEN &&
+	    action != CPU_ONLINE && action != CPU_ONLINE_FROZEN)
+		return NOTIFY_OK;
+
+	mutex_lock(&all_q_mutex);
+	list_for_each_entry(q, &all_q_list, all_q_node)
+		blk_mq_queue_reinit(q);
+	mutex_unlock(&all_q_mutex);
+	return NOTIFY_OK;
+}
+
+static int __init blk_mq_init(void)
+{
+	unsigned int i;
+
+	for_each_possible_cpu(i)
+		init_llist_head(&per_cpu(ipi_lists, i));
+
+	blk_mq_cpu_init();
+
+	/* Must be called after percpu_counter_hotcpu_callback() */
+	hotcpu_notifier(blk_mq_queue_reinit_notify, -10);
+
+	return 0;
+}
+subsys_initcall(blk_mq_init);