blk-mq: use array manage hctx map instead of xarray

After commit 4e5cc99 ("blk-mq: manage hctx map via xarray"), we use
an xarray instead of array to store hctx, but in poll mode, each time
in blk_mq_poll, we need use xa_load to find corresponding hctx, this
introduce some costs. In my test, xa_load may cost 3.8% cpu.

This patch revert previous change, eliminates the overhead of xa_load
and can result in a 3% performance improvement.

Signed-off-by: Fengnan Chang <changfengnan@bytedance.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>

Author: poluo

block/blk-mq-tag.c

@@ -499,7 +499,7 @@ void blk_mq_queue_tag_busy_iter(struct request_queue *q, busy_tag_iter_fn *fn,
 	int srcu_idx;
 
 	/*
-	 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and hctx_table
+	 * __blk_mq_update_nr_hw_queues() updates nr_hw_queues and queue_hw_ctx
 	 * while the queue is frozen. So we can use q_usage_counter to avoid
 	 * racing with it.
 	 */

block/blk-mq.c

@@ -730,7 +730,7 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q,
 	 * If not tell the caller that it should skip this queue.
 	 */
 	ret = -EXDEV;
-	data.hctx = xa_load(&q->hctx_table, hctx_idx);
+	data.hctx = q->queue_hw_ctx[hctx_idx];
 	if (!blk_mq_hw_queue_mapped(data.hctx))
 		goto out_queue_exit;
 	cpu = cpumask_first_and(data.hctx->cpumask, cpu_online_mask);
@@ -3946,8 +3946,6 @@ static void blk_mq_exit_hctx(struct request_queue *q,
 			blk_free_flush_queue_callback);
 	hctx->fq = NULL;
 
-	xa_erase(&q->hctx_table, hctx_idx);
-
 	spin_lock(&q->unused_hctx_lock);
 	list_add(&hctx->hctx_list, &q->unused_hctx_list);
 	spin_unlock(&q->unused_hctx_lock);
@@ -3989,14 +3987,8 @@ static int blk_mq_init_hctx(struct request_queue *q,
 				hctx->numa_node))
 		goto exit_hctx;
 
-	if (xa_insert(&q->hctx_table, hctx_idx, hctx, GFP_KERNEL))
-		goto exit_flush_rq;
-
 	return 0;
 
- exit_flush_rq:
-	if (set->ops->exit_request)
-		set->ops->exit_request(set, hctx->fq->flush_rq, hctx_idx);
  exit_hctx:
 	if (set->ops->exit_hctx)
 		set->ops->exit_hctx(hctx, hctx_idx);
@@ -4385,7 +4377,7 @@ void blk_mq_release(struct request_queue *q)
 		kobject_put(&hctx->kobj);
 	}
 
-	xa_destroy(&q->hctx_table);
+	kfree(q->queue_hw_ctx);
 
 	/*
 	 * release .mq_kobj and sw queue's kobject now because
@@ -4529,26 +4521,44 @@ static struct blk_mq_hw_ctx *blk_mq_alloc_and_init_hctx(
 static void __blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
 				     struct request_queue *q)
 {
-	struct blk_mq_hw_ctx *hctx;
-	unsigned long i, j;
+	int i, j, end;
+	struct blk_mq_hw_ctx **hctxs = q->queue_hw_ctx;
+
+	if (q->nr_hw_queues < set->nr_hw_queues) {
+		struct blk_mq_hw_ctx **new_hctxs;
+
+		new_hctxs = kcalloc_node(set->nr_hw_queues,
+				       sizeof(*new_hctxs), GFP_KERNEL,
+				       set->numa_node);
+		if (!new_hctxs)
+			return;
+		if (hctxs)
+			memcpy(new_hctxs, hctxs, q->nr_hw_queues *
+			       sizeof(*hctxs));
+		q->queue_hw_ctx = new_hctxs;
+		kfree(hctxs);
+		hctxs = new_hctxs;
+	}
 
 	for (i = 0; i < set->nr_hw_queues; i++) {
 		int old_node;
 		int node = blk_mq_get_hctx_node(set, i);
-		struct blk_mq_hw_ctx *old_hctx = xa_load(&q->hctx_table, i);
+		struct blk_mq_hw_ctx *old_hctx = hctxs[i];
 
 		if (old_hctx) {
 			old_node = old_hctx->numa_node;
 			blk_mq_exit_hctx(q, set, old_hctx, i);
 		}
 
-		if (!blk_mq_alloc_and_init_hctx(set, q, i, node)) {
+		hctxs[i] = blk_mq_alloc_and_init_hctx(set, q, i, node);
+		if (!hctxs[i]) {
 			if (!old_hctx)
 				break;
 			pr_warn("Allocate new hctx on node %d fails, fallback to previous one on node %d\n",
 					node, old_node);
-			hctx = blk_mq_alloc_and_init_hctx(set, q, i, old_node);
-			WARN_ON_ONCE(!hctx);
+			hctxs[i] = blk_mq_alloc_and_init_hctx(set, q, i,
+					old_node);
+			WARN_ON_ONCE(!hctxs[i]);
 		}
 	}
 	/*
@@ -4557,13 +4567,21 @@ static void __blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
 	 */
 	if (i != set->nr_hw_queues) {
 		j = q->nr_hw_queues;
+		end = i;
 	} else {
 		j = i;
+		end = q->nr_hw_queues;
 		q->nr_hw_queues = set->nr_hw_queues;
 	}
 
-	xa_for_each_start(&q->hctx_table, j, hctx, j)
-		blk_mq_exit_hctx(q, set, hctx, j);
+	for (; j < end; j++) {
+		struct blk_mq_hw_ctx *hctx = hctxs[j];
+
+		if (hctx) {
+			blk_mq_exit_hctx(q, set, hctx, j);
+			hctxs[j] = NULL;
+		}
+	}
 }
 
 static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
@@ -4599,8 +4617,6 @@ int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	INIT_LIST_HEAD(&q->unused_hctx_list);
 	spin_lock_init(&q->unused_hctx_lock);
 
-	xa_init(&q->hctx_table);
-
 	blk_mq_realloc_hw_ctxs(set, q);
 	if (!q->nr_hw_queues)
 		goto err_hctxs;
@@ -5187,7 +5203,7 @@ int blk_mq_poll(struct request_queue *q, blk_qc_t cookie,
 {
 	if (!blk_mq_can_poll(q))
 		return 0;
-	return blk_hctx_poll(q, xa_load(&q->hctx_table, cookie), iob, flags);
+	return blk_hctx_poll(q, q->queue_hw_ctx[cookie], iob, flags);
 }
 
 int blk_rq_poll(struct request *rq, struct io_comp_batch *iob,

block/blk-mq.h

@@ -84,7 +84,7 @@ static inline struct blk_mq_hw_ctx *blk_mq_map_queue_type(struct request_queue *
 							  enum hctx_type type,
 							  unsigned int cpu)
 {
-	return xa_load(&q->hctx_table, q->tag_set->map[type].mq_map[cpu]);
+	return q->queue_hw_ctx[q->tag_set->map[type].mq_map[cpu]];
 }
 
 static inline enum hctx_type blk_mq_get_hctx_type(blk_opf_t opf)

include/linux/blk-mq.h

@@ -1016,7 +1016,8 @@ static inline void *blk_mq_rq_to_pdu(struct request *rq)
 }
 
 #define queue_for_each_hw_ctx(q, hctx, i)				\
-	xa_for_each(&(q)->hctx_table, (i), (hctx))
+	for ((i) = 0; (i) < (q)->nr_hw_queues &&			\
+	     ({ hctx = (q)->queue_hw_ctx[i]; 1; }); (i)++)
 
 #define hctx_for_each_ctx(hctx, ctx, i)					\
 	for ((i) = 0; (i) < (hctx)->nr_ctx &&				\

include/linux/blkdev.h

@@ -503,7 +503,7 @@ struct request_queue {
 
 	/* hw dispatch queues */
 	unsigned int		nr_hw_queues;
-	struct xarray		hctx_table;
+	struct blk_mq_hw_ctx	**queue_hw_ctx;
 
 	struct percpu_ref	q_usage_counter;
 	struct lock_class_key	io_lock_cls_key;