blk-mq: support per-distpatch_queue flush machinery

This patch supports to run one single flush machinery for
each blk-mq dispatch queue, so that:

- current init_request and exit_request callbacks can
cover flush request too, then the buggy copying way of
initializing flush request's pdu can be fixed

- flushing performance gets improved in case of multi hw-queue

In fio sync write test over virtio-blk(4 hw queues, ioengine=sync,
iodepth=64, numjobs=4, bs=4K), it is observed that througput gets
increased a lot over my test environment:
	- throughput: +70% in case of virtio-blk over null_blk
	- throughput: +30% in case of virtio-blk over SSD image

The multi virtqueue feature isn't merged to QEMU yet, and patches for
the feature can be found in below tree:

	git://kernel.ubuntu.com/ming/qemu.git  	v2.1.0-mq.4

And simply passing 'num_queues=4 vectors=5' should be enough to
enable multi queue(quad queue) feature for QEMU virtio-blk.

Suggested-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>

6 files changed, 60 insertions, 39 deletions

diff --git a/block/blk-core.c b/block/blk-core.c
index b1dd4e086740..e1c2775c7597 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -704,7 +704,7 @@ blk_init_allocated_queue(struct request_queue *q, request_fn_proc *rfn,
 	if (!q)
 		return NULL;
 
-	q->fq = blk_alloc_flush_queue(q);
+	q->fq = blk_alloc_flush_queue(q, NUMA_NO_NODE, 0);
 	if (!q->fq)
 		return NULL;
 
diff --git a/block/blk-flush.c b/block/blk-flush.c
index 004d95e4098e..20badd7b9d1b 100644
--- a/block/blk-flush.c
+++ b/block/blk-flush.c
@@ -305,8 +305,15 @@ static bool blk_kick_flush(struct request_queue *q, struct blk_flush_queue *fq)
 	fq->flush_pending_idx ^= 1;
 
 	blk_rq_init(q, flush_rq);
-	if (q->mq_ops)
-		blk_mq_clone_flush_request(flush_rq, first_rq);
+
+	/*
+	 * Borrow tag from the first request since they can't
+	 * be in flight at the same time.
+	 */
+	if (q->mq_ops) {
+		flush_rq->mq_ctx = first_rq->mq_ctx;
+		flush_rq->tag = first_rq->tag;
+	}
 
 	flush_rq->cmd_type = REQ_TYPE_FS;
 	flush_rq->cmd_flags = WRITE_FLUSH | REQ_FLUSH_SEQ;
@@ -480,22 +487,22 @@ int blkdev_issue_flush(struct block_device *bdev, gfp_t gfp_mask,
 }
 EXPORT_SYMBOL(blkdev_issue_flush);
 
-struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q)
+struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
+		int node, int cmd_size)
 {
 	struct blk_flush_queue *fq;
 	int rq_sz = sizeof(struct request);
 
-	fq = kzalloc(sizeof(*fq), GFP_KERNEL);
+	fq = kzalloc_node(sizeof(*fq), GFP_KERNEL, node);
 	if (!fq)
 		goto fail;
 
 	if (q->mq_ops) {
 		spin_lock_init(&fq->mq_flush_lock);
-		rq_sz = round_up(rq_sz + q->tag_set->cmd_size,
-				cache_line_size());
+		rq_sz = round_up(rq_sz + cmd_size, cache_line_size());
 	}
 
-	fq->flush_rq = kzalloc(rq_sz, GFP_KERNEL);
+	fq->flush_rq = kzalloc_node(rq_sz, GFP_KERNEL, node);
 	if (!fq->flush_rq)
 		goto fail_rq;
 
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 53b6def12fc4..4e7a31466139 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -281,26 +281,6 @@ void blk_mq_free_request(struct request *rq)
 	__blk_mq_free_request(hctx, ctx, rq);
 }
 
-/*
- * Clone all relevant state from a request that has been put on hold in
- * the flush state machine into the preallocated flush request that hangs
- * off the request queue.
- *
- * For a driver the flush request should be invisible, that's why we are
- * impersonating the original request here.
- */
-void blk_mq_clone_flush_request(struct request *flush_rq,
-		struct request *orig_rq)
-{
-	struct blk_mq_hw_ctx *hctx =
-		orig_rq->q->mq_ops->map_queue(orig_rq->q, orig_rq->mq_ctx->cpu);
-
-	flush_rq->mq_ctx = orig_rq->mq_ctx;
-	flush_rq->tag = orig_rq->tag;
-	memcpy(blk_mq_rq_to_pdu(flush_rq), blk_mq_rq_to_pdu(orig_rq),
-		hctx->cmd_size);
-}
-
 inline void __blk_mq_end_request(struct request *rq, int error)
 {
 	blk_account_io_done(rq);
@@ -1516,12 +1496,20 @@ static void blk_mq_exit_hctx(struct request_queue *q,
 		struct blk_mq_tag_set *set,
 		struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
 {
+	unsigned flush_start_tag = set->queue_depth;
+
 	blk_mq_tag_idle(hctx);
 
+	if (set->ops->exit_request)
+		set->ops->exit_request(set->driver_data,
+				       hctx->fq->flush_rq, hctx_idx,
+				       flush_start_tag + hctx_idx);
+
 	if (set->ops->exit_hctx)
 		set->ops->exit_hctx(hctx, hctx_idx);
 
 	blk_mq_unregister_cpu_notifier(&hctx->cpu_notifier);
+	blk_free_flush_queue(hctx->fq);
 	kfree(hctx->ctxs);
 	blk_mq_free_bitmap(&hctx->ctx_map);
 }
@@ -1556,6 +1544,7 @@ static int blk_mq_init_hctx(struct request_queue *q,
 		struct blk_mq_hw_ctx *hctx, unsigned hctx_idx)
 {
 	int node;
+	unsigned flush_start_tag = set->queue_depth;
 
 	node = hctx->numa_node;
 	if (node == NUMA_NO_NODE)
@@ -1594,8 +1583,23 @@ static int blk_mq_init_hctx(struct request_queue *q,
 	    set->ops->init_hctx(hctx, set->driver_data, hctx_idx))
 		goto free_bitmap;
 
+	hctx->fq = blk_alloc_flush_queue(q, hctx->numa_node, set->cmd_size);
+	if (!hctx->fq)
+		goto exit_hctx;
+
+	if (set->ops->init_request &&
+	    set->ops->init_request(set->driver_data,
+				   hctx->fq->flush_rq, hctx_idx,
+				   flush_start_tag + hctx_idx, node))
+		goto free_fq;
+
 	return 0;
 
+ free_fq:
+	kfree(hctx->fq);
+ exit_hctx:
+	if (set->ops->exit_hctx)
+		set->ops->exit_hctx(hctx, hctx_idx);
  free_bitmap:
 	blk_mq_free_bitmap(&hctx->ctx_map);
  free_ctxs:
@@ -1862,16 +1866,10 @@ struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
 
 	blk_mq_add_queue_tag_set(set, q);
 
-	q->fq = blk_alloc_flush_queue(q);
-	if (!q->fq)
-		goto err_hw_queues;
-
 	blk_mq_map_swqueue(q);
 
 	return q;
 
-err_hw_queues:
-	blk_mq_exit_hw_queues(q, set, set->nr_hw_queues);
 err_hw:
 	blk_cleanup_queue(q);
 err_hctxs:
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 718cffc4c678..e8f38a36c625 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -517,10 +517,10 @@ static void blk_release_queue(struct kobject *kobj)
 	if (q->queue_tags)
 		__blk_queue_free_tags(q);
 
-	blk_free_flush_queue(q->fq);
-
 	if (q->mq_ops)
 		blk_mq_free_queue(q);
+	else
+		blk_free_flush_queue(q->fq);
 
 	blk_trace_shutdown(q);
 
diff --git a/block/blk.h b/block/blk.h
index 7ecdd8517e69..43b036185712 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -2,6 +2,8 @@
 #define BLK_INTERNAL_H
 
 #include <linux/idr.h>
+#include <linux/blk-mq.h>
+#include "blk-mq.h"
 
 /* Amount of time in which a process may batch requests */
 #define BLK_BATCH_TIME	(HZ/50UL)
@@ -31,7 +33,14 @@ extern struct ida blk_queue_ida;
 static inline struct blk_flush_queue *blk_get_flush_queue(
 		struct request_queue *q, struct blk_mq_ctx *ctx)
 {
-	return q->fq;
+	struct blk_mq_hw_ctx *hctx;
+
+	if (!q->mq_ops)
+		return q->fq;
+
+	hctx = q->mq_ops->map_queue(q, ctx->cpu);
+
+	return hctx->fq;
 }
 
 static inline void __blk_get_queue(struct request_queue *q)
@@ -39,8 +48,9 @@ static inline void __blk_get_queue(struct request_queue *q)
 	kobject_get(&q->kobj);
 }
 
-struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q);
-void blk_free_flush_queue(struct blk_flush_queue *fq);
+struct blk_flush_queue *blk_alloc_flush_queue(struct request_queue *q,
+		int node, int cmd_size);
+void blk_free_flush_queue(struct blk_flush_queue *q);
 
 int blk_init_rl(struct request_list *rl, struct request_queue *q,
 		gfp_t gfp_mask);
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 325349559fb0..02c5d950f444 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -4,6 +4,7 @@
 #include <linux/blkdev.h>
 
 struct blk_mq_tags;
+struct blk_flush_queue;
 
 struct blk_mq_cpu_notifier {
 	struct list_head list;
@@ -34,6 +35,7 @@ struct blk_mq_hw_ctx {
 
 	struct request_queue	*queue;
 	unsigned int		queue_num;
+	struct blk_flush_queue	*fq;
 
 	void			*driver_data;
 
@@ -119,6 +121,10 @@ struct blk_mq_ops {
 	/*
 	 * Called for every command allocated by the block layer to allow
 	 * the driver to set up driver specific data.
+	 *
+	 * Tag greater than or equal to queue_depth is for setting up
+	 * flush request.
+	 *
 	 * Ditto for exit/teardown.
 	 */
 	init_request_fn		*init_request;