blk-mq: allow hardware queue to get more tag while sharing a tag set

hulk inclusion
category: performance
bugzilla: https://gitee.com/openeuler/kernel/issues/I4S8DW

---------------------------

When sharing a tag set, if most disks are issuing small amount of IO, and
only a few is issuing a large amount of IO. Current approach is to limit
the max amount of tags a disk can get equally to the average of total
tags. Thus the few heavy load disk can't get enough tags while many tags
are still free in the tag set.

We add 'pending_queues' in blk_mq_tag_set to count how many queues can't
get driver tag. Thus if this value is zero, there is no need to limit
the max number of available tags.

On the other hand, if a queue doesn't issue IO, the 'active_queues' will
not be decreased in a period of time(request timeout), thus a lot of tags
will not be available because max number of available tags is set to
max(total tags / active_queues, 4). Thus we decreased it when
'nr_active' is 0.

This functionality is enabled by default, to disable it, add
"blk_mq.unfair_dtag=0" to boot cmd.
Signed-off-by: NYu Kuai <yukuai3@huawei.com>
Reviewed-by: NHou Tao <houtao1@huawei.com>
Signed-off-by: NZheng Zengkai <zhengzengkai@huawei.com>

block/blk-mq-debugfs.c

@@ -453,6 +453,8 @@ static void blk_mq_debugfs_tags_show(struct seq_file *m,
 	seq_printf(m, "nr_reserved_tags=%u\n", tags->nr_reserved_tags);
 	seq_printf(m, "active_queues=%d\n",
 		   atomic_read(&tags->active_queues));
+	seq_printf(m, "pending_queues=%d\n",
+		   atomic_read(&tags->pending_queues));
 
 	seq_puts(m, "\nbitmap_tags:\n");
 	sbitmap_queue_show(tags->bitmap_tags, m);

block/blk-mq-tag.c

@@ -73,6 +73,40 @@ void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
 	blk_mq_tag_wakeup_all(tags, false);
 }
 
+void __blk_mq_dtag_busy(struct blk_mq_hw_ctx *hctx)
+{
+	if (blk_mq_is_sbitmap_shared(hctx->flags)) {
+		struct request_queue *q = hctx->queue;
+		struct blk_mq_tag_set *set = q->tag_set;
+
+		if (!test_bit(QUEUE_FLAG_HCTX_WAIT, &q->queue_flags) &&
+		    !test_and_set_bit(QUEUE_FLAG_HCTX_WAIT, &q->queue_flags))
+			atomic_inc(&set->pending_queues_shared_sbitmap);
+	} else {
+		if (!test_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state) &&
+		    !test_and_set_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state))
+			atomic_inc(&hctx->tags->pending_queues);
+	}
+}
+
+void __blk_mq_dtag_idle(struct blk_mq_hw_ctx *hctx)
+{
+	struct blk_mq_tags *tags = hctx->tags;
+	struct request_queue *q = hctx->queue;
+	struct blk_mq_tag_set *set = q->tag_set;
+
+	if (blk_mq_is_sbitmap_shared(hctx->flags)) {
+		if (!test_and_clear_bit(QUEUE_FLAG_HCTX_WAIT,
+					&q->queue_flags))
+			return;
+		atomic_dec(&set->pending_queues_shared_sbitmap);
+	} else {
+		if (!test_and_clear_bit(BLK_MQ_S_DTAG_WAIT, &hctx->state))
+			return;
+		atomic_dec(&tags->pending_queues);
+	}
+}
+
 static int __blk_mq_get_tag(struct blk_mq_alloc_data *data,
 			    struct sbitmap_queue *bt)
 {
@@ -111,8 +145,11 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 	if (tag != BLK_MQ_NO_TAG)
 		goto found_tag;
 
-	if (data->flags & BLK_MQ_REQ_NOWAIT)
+	if (data->flags & BLK_MQ_REQ_NOWAIT) {
+		if (!data->q->elevator)
+			blk_mq_dtag_busy(data->hctx);
 		return BLK_MQ_NO_TAG;
+	}
 
 	ws = bt_wait_ptr(bt, data->hctx);
 	do {
@@ -139,6 +176,8 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
 		if (tag != BLK_MQ_NO_TAG)
 			break;
 
+		if (!data->q->elevator)
+			blk_mq_dtag_busy(data->hctx);
 		bt_prev = bt;
 		io_schedule();
 

block/blk-mq-tag.h

@@ -4,6 +4,7 @@
 
 #include <linux/kabi.h>
 
+extern bool mq_unfair_dtag;
 /*
  * Tag address space map.
  */
@@ -12,6 +13,11 @@ struct blk_mq_tags {
 	unsigned int nr_reserved_tags;
 
 	atomic_t active_queues;
+	/*
+	 * If multiple queues share a tag set, pending_queues record the
+	 * number of queues that can't get driver tag.
+	 */
+	atomic_t pending_queues;
 
 	struct sbitmap_queue *bitmap_tags;
 	struct sbitmap_queue *breserved_tags;
@@ -73,8 +79,11 @@ enum {
 	BLK_MQ_TAG_MAX		= BLK_MQ_NO_TAG - 1,
 };
 
-extern bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *);
-extern void __blk_mq_tag_idle(struct blk_mq_hw_ctx *);
+extern bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx);
+extern void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx);
+extern void __blk_mq_dtag_busy(struct blk_mq_hw_ctx *hctx);
+extern void __blk_mq_dtag_idle(struct blk_mq_hw_ctx *hctx);
+
 
 static inline bool blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
 {
@@ -92,6 +101,22 @@ static inline void blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
 	__blk_mq_tag_idle(hctx);
 }
 
+static inline void blk_mq_dtag_busy(struct blk_mq_hw_ctx *hctx)
+{
+	if (!(mq_unfair_dtag && (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)))
+		return;
+
+	__blk_mq_dtag_busy(hctx);
+}
+
+static inline void blk_mq_dtag_idle(struct blk_mq_hw_ctx *hctx)
+{
+	if (!(mq_unfair_dtag && (hctx->flags & BLK_MQ_F_TAG_QUEUE_SHARED)))
+		return;
+
+	__blk_mq_dtag_idle(hctx);
+}
+
 static inline bool blk_mq_tag_is_reserved(struct blk_mq_tags *tags,
 					  unsigned int tag)
 {

block/blk-mq.c

@@ -41,6 +41,9 @@
 #include "blk-mq-sched.h"
 #include "blk-rq-qos.h"
 
+bool mq_unfair_dtag = true;
+module_param_named(unfair_dtag, mq_unfair_dtag, bool, 0444);
+
 static DEFINE_PER_CPU(struct list_head, blk_cpu_done);
 
 static void blk_mq_poll_stats_start(struct request_queue *q);
@@ -535,8 +538,13 @@ void blk_mq_free_request(struct request *rq)
 	}
 
 	ctx->rq_completed[rq_is_sync(rq)]++;
-	if (rq->rq_flags & RQF_MQ_INFLIGHT)
+	if (rq->rq_flags & RQF_MQ_INFLIGHT) {
 		__blk_mq_dec_active_requests(hctx);
+		if (mq_unfair_dtag && !__blk_mq_active_requests(hctx)) {
+			blk_mq_tag_idle(hctx);
+			blk_mq_dtag_idle(hctx);
+		}
+	}
 
 	if (unlikely(laptop_mode && !blk_rq_is_passthrough(rq)))
 		laptop_io_completion(q->backing_dev_info);
@@ -1003,8 +1011,10 @@ static void blk_mq_timeout_work(struct work_struct *work)
 		 */
 		queue_for_each_hw_ctx(q, hctx, i) {
 			/* the hctx may be unmapped, so check it here */
-			if (blk_mq_hw_queue_mapped(hctx))
+			if (blk_mq_hw_queue_mapped(hctx)) {
 				blk_mq_tag_idle(hctx);
+				blk_mq_dtag_idle(hctx);
+			}
 		}
 	}
 	blk_queue_exit(q);
@@ -1109,8 +1119,10 @@ static bool __blk_mq_get_driver_tag(struct request *rq)
 	}
 
 	tag = __sbitmap_queue_get(bt);
-	if (tag == BLK_MQ_NO_TAG)
+	if (tag == BLK_MQ_NO_TAG) {
+		blk_mq_dtag_busy(rq->mq_hctx);
 		return false;
+	}
 
 	rq->tag = tag + tag_offset;
 	return true;
@@ -2711,8 +2723,10 @@ static void blk_mq_exit_hctx(struct request_queue *q,
 {
 	struct request *flush_rq = hctx->fq->flush_rq;
 
-	if (blk_mq_hw_queue_mapped(hctx))
+	if (blk_mq_hw_queue_mapped(hctx)) {
 		blk_mq_tag_idle(hctx);
+		blk_mq_dtag_idle(hctx);
+	}
 
 	blk_mq_clear_flush_rq_mapping(set->tags[hctx_idx],
 			set->queue_depth, flush_rq);
@@ -3033,6 +3047,7 @@ static void queue_set_hctx_shared(struct request_queue *q, bool shared)
 			hctx->flags |= BLK_MQ_F_TAG_QUEUE_SHARED;
 		} else {
 			blk_mq_tag_idle(hctx);
+			blk_mq_dtag_idle(hctx);
 			hctx->flags &= ~BLK_MQ_F_TAG_QUEUE_SHARED;
 		}
 	}
@@ -3589,6 +3604,7 @@ int blk_mq_alloc_tag_set(struct blk_mq_tag_set *set)
 
 	if (blk_mq_is_sbitmap_shared(set->flags)) {
 		atomic_set(&set->active_queues_shared_sbitmap, 0);
+		atomic_set(&set->pending_queues_shared_sbitmap, 0);
 
 		if (blk_mq_init_shared_sbitmap(set, set->flags)) {
 			ret = -ENOMEM;

block/blk-mq.h

@@ -316,10 +316,15 @@ static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
 		struct request_queue *q = hctx->queue;
 		struct blk_mq_tag_set *set = q->tag_set;
 
+		if (mq_unfair_dtag &&
+		    !atomic_read(&set->pending_queues_shared_sbitmap))
+			return true;
 		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
 			return true;
 		users = atomic_read(&set->active_queues_shared_sbitmap);
 	} else {
+		if (mq_unfair_dtag && !atomic_read(&hctx->tags->pending_queues))
+			return true;
 		if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
 			return true;
 		users = atomic_read(&hctx->tags->active_queues);

include/linux/blk-mq.h

@@ -266,6 +266,7 @@ struct blk_mq_tag_set {
 	unsigned int		flags;
 	void			*driver_data;
 	atomic_t		active_queues_shared_sbitmap;
+	atomic_t		pending_queues_shared_sbitmap;
 
 	struct sbitmap_queue	__bitmap_tags;
 	struct sbitmap_queue	__breserved_tags;
@@ -442,6 +443,9 @@ enum {
 	/* hw queue is inactive after all its CPUs become offline */
 	BLK_MQ_S_INACTIVE	= 3,
 
+	/* hw queue is waiting for driver tag */
+	BLK_MQ_S_DTAG_WAIT	= 4,
+
 	BLK_MQ_MAX_DEPTH	= 10240,
 
 	BLK_MQ_CPU_WORK_BATCH	= 8,

include/linux/blkdev.h

@@ -640,6 +640,8 @@ struct request_queue {
 #define QUEUE_FLAG_RQ_ALLOC_TIME 27	/* record rq->alloc_time_ns */
 #define QUEUE_FLAG_HCTX_ACTIVE	28	/* at least one blk-mq hctx is active */
 #define QUEUE_FLAG_NOWAIT       29	/* device supports NOWAIT */
+/*at least one blk-mq hctx can't get driver tag */
+#define QUEUE_FLAG_HCTX_WAIT	30
 
 #define QUEUE_FLAG_MQ_DEFAULT	((1 << QUEUE_FLAG_IO_STAT) |		\
 				 (1 << QUEUE_FLAG_SAME_COMP) |		\