[PATCH 2/2] blk-mq: allow direct dispatch to a driver specific workqueue

[Christoph Hellwig <hch@lst.de>]

We have various block drivers that need to execute long term blocking
operations during I/O submission like file system or network I/O.

Currently these drivers just queue up work to an internal workqueue
from their request_fn.  With blk-mq we can make sure they always get
called on their own workqueue directly for I/O submission by:

 1) adding a flag to prevent inline submission of I/O, and
 2) allowing the driver to pass in a workqueue in the tag_set that
    will be used instead of kblockd.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 block/blk-core.c       |  2 +-
 block/blk-mq.c         | 12 +++++++++---
 block/blk.h            |  1 +
 include/linux/blk-mq.h |  4 ++++
 4 files changed, 15 insertions(+), 4 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index 0421b53..7f7249f 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -61,7 +61,7 @@ struct kmem_cache *blk_requestq_cachep;
 /*
  * Controlling structure to kblockd
  */
-static struct workqueue_struct *kblockd_workqueue;
+struct workqueue_struct *kblockd_workqueue;
 
 void blk_queue_congestion_threshold(struct request_queue *q)
 {
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 22e50a5..3d27d22 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -804,12 +804,13 @@ void blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx, bool async)
 	if (unlikely(test_bit(BLK_MQ_S_STOPPED, &hctx->state)))
 		return;
 
-	if (!async && cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) {
+	if (!async && !(hctx->flags & BLK_MQ_F_WORKQUEUE) &&
+	    cpumask_test_cpu(smp_processor_id(), hctx->cpumask)) {
 		__blk_mq_run_hw_queue(hctx);
 		return;
 	}
 
-	kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
+	queue_delayed_work_on(blk_mq_hctx_next_cpu(hctx), hctx->wq,
 			&hctx->run_work, 0);
 }
 
@@ -908,7 +909,7 @@ static void blk_mq_delay_work_fn(struct work_struct *work)
 
 void blk_mq_delay_queue(struct blk_mq_hw_ctx *hctx, unsigned long msecs)
 {
-	kblockd_schedule_delayed_work_on(blk_mq_hctx_next_cpu(hctx),
+	queue_delayed_work_on(blk_mq_hctx_next_cpu(hctx), hctx->wq,
 			&hctx->delay_work, msecs_to_jiffies(msecs));
 }
 EXPORT_SYMBOL(blk_mq_delay_queue);
@@ -1581,6 +1582,11 @@ static int blk_mq_init_hctx(struct request_queue *q,
 	hctx->flags = set->flags;
 	hctx->cmd_size = set->cmd_size;
 
+	if (set->wq)
+		hctx->wq = set->wq;
+	else
+		hctx->wq = kblockd_workqueue;
+
 	blk_mq_init_cpu_notifier(&hctx->cpu_notifier,
 					blk_mq_hctx_notify, hctx);
 	blk_mq_register_cpu_notifier(&hctx->cpu_notifier);
diff --git a/block/blk.h b/block/blk.h
index 43b0361..fb46ad0 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -25,6 +25,7 @@ struct blk_flush_queue {
 	spinlock_t		mq_flush_lock;
 };
 
+extern struct workqueue_struct *kblockd_workqueue;
 extern struct kmem_cache *blk_requestq_cachep;
 extern struct kmem_cache *request_cachep;
 extern struct kobj_type blk_queue_ktype;
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 5a901d0..ebe4699 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -37,6 +37,8 @@ struct blk_mq_hw_ctx {
 	unsigned int		queue_num;
 	struct blk_flush_queue	*fq;
 
+	struct workqueue_struct	*wq;
+
 	void			*driver_data;
 
 	struct blk_mq_ctxmap	ctx_map;
@@ -64,6 +66,7 @@ struct blk_mq_hw_ctx {
 
 struct blk_mq_tag_set {
 	struct blk_mq_ops	*ops;
+	struct workqueue_struct	*wq;
 	unsigned int		nr_hw_queues;
 	unsigned int		queue_depth;	/* max hw supported */
 	unsigned int		reserved_tags;
@@ -156,6 +159,7 @@ enum {
 	BLK_MQ_F_SG_MERGE	= 1 << 2,
 	BLK_MQ_F_SYSFS_UP	= 1 << 3,
 	BLK_MQ_F_DEFER_ISSUE	= 1 << 4,
+	BLK_MQ_F_WORKQUEUE	= 1 << 5,
 
 	BLK_MQ_S_STOPPED	= 0,
 	BLK_MQ_S_TAG_ACTIVE	= 1,
-- 
1.9.1