[patch v2 0/7] blk-mq: introduce new queue attribute async_depth

[patch v2 0/7] blk-mq: introduce new queue attribute async_depth

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

Changes in v2:
 - keep limit_depth() method for kyber and mq-deadline in patch 3;
 - add description about sysfs api change for kyber and mq-deadline;
 - improve documentation in patch 7;
 - add review tag for patch 1;

Background and motivation:

At first, we test a performance regression from 5.10 to 6.6 in
downstream kernel(described in patch 5), the regression is related to
async_depth in mq-dealine.

While trying to fix this regression, Bart suggests add a new attribute
to request_queue, and I think this is a good idea because all elevators
have similar logical, however only mq-deadline allow user to configure
async_depth.

patch 1-3 add new queue attribute async_depth;
patch 4 convert kyber to use request_queue->async_depth;
patch 5 covnert mq-dedaline to use request_queue->async_depth, also the
performance regression will be fixed;
patch 6 convert bfq to use request_queue->async_depth;

Yu Kuai (7):
  block: convert nr_requests to unsigned int
  blk-mq-sched: unify elevators checking for async requests
  blk-mq: add a new queue sysfs attribute async_depth
  kyber: covert to use request_queue->async_depth
  mq-deadline: covert to use request_queue->async_depth
  block, bfq: convert to use request_queue->async_depth
  blk-mq: add documentation for new queue attribute async_dpeth

 Documentation/ABI/stable/sysfs-block | 34 +++++++++++++++
 block/bfq-iosched.c                  | 45 ++++++++-----------
 block/blk-core.c                     |  1 +
 block/blk-mq-sched.h                 |  5 +++
 block/blk-mq.c                       | 64 +++++++++++++++++-----------
 block/blk-sysfs.c                    | 47 ++++++++++++++++++++
 block/elevator.c                     |  1 +
 block/kyber-iosched.c                | 33 +++-----------
 block/mq-deadline.c                  | 39 +++--------------
 include/linux/blkdev.h               |  3 +-
 10 files changed, 157 insertions(+), 115 deletions(-)

-- 
2.39.2

[patch v2 1/7] block: convert nr_requests to unsigned int

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

This value represents the number of requests for elevator tags, or drivers
tags if elevator is none. The max value for elevator tags is 2048, and
in drivers at most 16 bits is used for tag.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>
---
 include/linux/blkdev.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 066e5309bd45..02c006fb94c5 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -541,7 +541,7 @@ struct request_queue {
 	/*
 	 * queue settings
 	 */
-	unsigned long		nr_requests;	/* Max # of requests */
+	unsigned int		nr_requests;	/* Max # of requests */
 
 #ifdef CONFIG_BLK_INLINE_ENCRYPTION
 	struct blk_crypto_profile *crypto_profile;
-- 
2.39.2

[patch v2 2/7] blk-mq-sched: unify elevators checking for async requests

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

bfq and mq-deadline consider sync writes as async requests and only
resver tags for sync reads by async_depth, however, kyber doesn't
consider sync writes as async requests for now.

Consider the case there are lots of dirty pages, and user use fsync to
flush dirty pages. In this case sched_tags can be exhausted by sync writes
and sync reads can stuck waiting for tag. Hence let kyber follow what
mq-deadline and bfq did, and unify async requests checking for all
elevators.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/bfq-iosched.c   | 2 +-
 block/blk-mq-sched.h  | 5 +++++
 block/kyber-iosched.c | 2 +-
 block/mq-deadline.c   | 2 +-
 4 files changed, 8 insertions(+), 3 deletions(-)

diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
index 4a8d3d96bfe4..63452e791f98 100644
--- a/block/bfq-iosched.c
+++ b/block/bfq-iosched.c
@@ -697,7 +697,7 @@ static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 	unsigned int limit, act_idx;
 
 	/* Sync reads have full depth available */
-	if (op_is_sync(opf) && !op_is_write(opf))
+	if (blk_mq_sched_sync_request(opf))
 		limit = data->q->nr_requests;
 	else
 		limit = bfqd->async_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
diff --git a/block/blk-mq-sched.h b/block/blk-mq-sched.h
index 8e21a6b1415d..ae747f9053c7 100644
--- a/block/blk-mq-sched.h
+++ b/block/blk-mq-sched.h
@@ -103,4 +103,9 @@ static inline void blk_mq_set_min_shallow_depth(struct request_queue *q,
 						depth);
 }
 
+static inline bool blk_mq_sched_sync_request(blk_opf_t opf)
+{
+	return op_is_sync(opf) && !op_is_write(opf);
+}
+
 #endif
diff --git a/block/kyber-iosched.c b/block/kyber-iosched.c
index 18efd6ef2a2b..cf243a457175 100644
--- a/block/kyber-iosched.c
+++ b/block/kyber-iosched.c
@@ -544,7 +544,7 @@ static void kyber_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 	 * We use the scheduler tags as per-hardware queue queueing tokens.
 	 * Async requests can be limited at this stage.
 	 */
-	if (!op_is_sync(opf)) {
+	if (!blk_mq_sched_sync_request(opf)) {
 		struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
 
 		data->shallow_depth = kqd->async_depth;
diff --git a/block/mq-deadline.c b/block/mq-deadline.c
index 3e741d33142d..592dd853f6e5 100644
--- a/block/mq-deadline.c
+++ b/block/mq-deadline.c
@@ -492,7 +492,7 @@ static void dd_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 	struct deadline_data *dd = data->q->elevator->elevator_data;
 
 	/* Do not throttle synchronous reads. */
-	if (op_is_sync(opf) && !op_is_write(opf))
+	if (blk_mq_sched_sync_request(opf))
 		return;
 
 	/*
-- 
2.39.2

Re: [patch v2 2/7] blk-mq-sched: unify elevators checking for async requests

[Nilay Shroff]

On 10/9/25 1:16 PM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> bfq and mq-deadline consider sync writes as async requests and only
> resver tags for sync reads by async_depth, however, kyber doesn't
> consider sync writes as async requests for now.
> 
> Consider the case there are lots of dirty pages, and user use fsync to
> flush dirty pages. In this case sched_tags can be exhausted by sync writes
> and sync reads can stuck waiting for tag. Hence let kyber follow what
> mq-deadline and bfq did, and unify async requests checking for all
> elevators.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

Looks good to me:
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>

Re: [patch v2 2/7] blk-mq-sched: unify elevators checking for async requests

[Bart Van Assche]

On 10/9/25 12:46 AM, Yu Kuai wrote:
> +static inline bool blk_mq_sched_sync_request(blk_opf_t opf)
> +{
> +	return op_is_sync(opf) && !op_is_write(opf);
> +}

The "sched" part in the function name suggests that this function
schedules something while it only tests something. Maybe
"blk_mq_is_sync_read()" is a better function name?

Thanks,

Bart.

Re: [patch v2 2/7] blk-mq-sched: unify elevators checking for async requests

[Yu Kuai]

Hi,

在 2025/10/10 1:02, Bart Van Assche 写道:
> On 10/9/25 12:46 AM, Yu Kuai wrote:
>> +static inline bool blk_mq_sched_sync_request(blk_opf_t opf)
>> +{
>> +    return op_is_sync(opf) && !op_is_write(opf);
>> +}
>
> The "sched" part in the function name suggests that this function
> schedules something while it only tests something. Maybe
> "blk_mq_is_sync_read()" is a better function name?

This is just a prefix to indicate that this helper is user for scheduler,
just like lots of other helpers in blk-mq-sched.h

I'll prefer to keep this name if you don't mind :)

Thanks,
Kuai

>
> Thanks,
>
> Bart.
>

[patch v2 3/7] blk-mq: add a new queue sysfs attribute async_depth

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

Add a new field async_depth to request_queue and related APIs, this is
currently not used, following patches will convert elevators to use
this instead of internal async_depth.

Also factor out a helper blk_mq_limit_depth() to make code cleaner.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/blk-core.c       |  1 +
 block/blk-mq.c         | 64 +++++++++++++++++++++++++-----------------
 block/blk-sysfs.c      | 47 +++++++++++++++++++++++++++++++
 block/elevator.c       |  1 +
 include/linux/blkdev.h |  1 +
 5 files changed, 89 insertions(+), 25 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index dd39ff651095..76df70cfc103 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -463,6 +463,7 @@ struct request_queue *blk_alloc_queue(struct queue_limits *lim, int node_id)
 	fs_reclaim_release(GFP_KERNEL);
 
 	q->nr_requests = BLKDEV_DEFAULT_RQ;
+	q->async_depth = BLKDEV_DEFAULT_RQ;
 
 	return q;
 
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 09f579414161..16b335723610 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -491,6 +491,38 @@ __blk_mq_alloc_requests_batch(struct blk_mq_alloc_data *data)
 	return rq_list_pop(data->cached_rqs);
 }
 
+static void blk_mq_limit_depth(struct blk_mq_alloc_data *data)
+{
+	struct elevator_mq_ops *ops;
+
+	/* If elevator is none, don't limit requests */
+	if (!data->q->elevator) {
+		blk_mq_tag_busy(data->hctx);
+		return;
+	}
+
+	data->rq_flags |= RQF_SCHED_TAGS;
+
+	/*
+	 * Flush/passthrough requests are special and go directly to the
+	 * dispatch list, they don't have limit.
+	 */
+	if ((data->cmd_flags & REQ_OP_MASK) == REQ_OP_FLUSH ||
+	    blk_op_is_passthrough(data->cmd_flags))
+		return;
+
+	WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
+	data->rq_flags |= RQF_USE_SCHED;
+
+	/*
+	 * By default, sync requests have no limit, and async requests is
+	 * limited to async_depth.
+	 */
+	ops = &data->q->elevator->type->ops;
+	if (ops->limit_depth)
+		ops->limit_depth(data->cmd_flags, data);
+}
+
 static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
 {
 	struct request_queue *q = data->q;
@@ -509,31 +541,7 @@ static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data)
 	data->ctx = blk_mq_get_ctx(q);
 	data->hctx = blk_mq_map_queue(data->cmd_flags, data->ctx);
 
-	if (q->elevator) {
-		/*
-		 * All requests use scheduler tags when an I/O scheduler is
-		 * enabled for the queue.
-		 */
-		data->rq_flags |= RQF_SCHED_TAGS;
-
-		/*
-		 * Flush/passthrough requests are special and go directly to the
-		 * dispatch list.
-		 */
-		if ((data->cmd_flags & REQ_OP_MASK) != REQ_OP_FLUSH &&
-		    !blk_op_is_passthrough(data->cmd_flags)) {
-			struct elevator_mq_ops *ops = &q->elevator->type->ops;
-
-			WARN_ON_ONCE(data->flags & BLK_MQ_REQ_RESERVED);
-
-			data->rq_flags |= RQF_USE_SCHED;
-			if (ops->limit_depth)
-				ops->limit_depth(data->cmd_flags, data);
-		}
-	} else {
-		blk_mq_tag_busy(data->hctx);
-	}
-
+	blk_mq_limit_depth(data);
 	if (data->flags & BLK_MQ_REQ_RESERVED)
 		data->rq_flags |= RQF_RESV;
 
@@ -4605,6 +4613,7 @@ int blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	spin_lock_init(&q->requeue_lock);
 
 	q->nr_requests = set->queue_depth;
+	q->async_depth = set->queue_depth;
 
 	blk_mq_init_cpu_queues(q, set->nr_hw_queues);
 	blk_mq_map_swqueue(q);
@@ -4971,6 +4980,11 @@ struct elevator_tags *blk_mq_update_nr_requests(struct request_queue *q,
 		q->elevator->et = et;
 	}
 
+	/*
+	 * Preserve relative value, both nr and async_depth are at most 16 bit
+	 * value, no need to worry about overflow.
+	 */
+	q->async_depth = max(q->async_depth * nr / q->nr_requests, 1);
 	q->nr_requests = nr;
 	if (q->elevator && q->elevator->type->ops.depth_updated)
 		q->elevator->type->ops.depth_updated(q);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 76c47fe9b8d6..9553cc022c7e 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -127,6 +127,51 @@ queue_requests_store(struct gendisk *disk, const char *page, size_t count)
 	return ret;
 }
 
+static ssize_t queue_async_depth_show(struct gendisk *disk, char *page)
+{
+	ssize_t ret;
+
+	mutex_lock(&disk->queue->elevator_lock);
+	ret = queue_var_show(disk->queue->async_depth, page);
+	mutex_unlock(&disk->queue->elevator_lock);
+	return ret;
+}
+
+static ssize_t
+queue_async_depth_store(struct gendisk *disk, const char *page, size_t count)
+{
+	struct request_queue *q = disk->queue;
+	unsigned int memflags;
+	unsigned long nr;
+	int ret;
+
+	if (!queue_is_mq(q))
+		return -EINVAL;
+
+	ret = queue_var_store(&nr, page, count);
+	if (ret < 0)
+		return ret;
+
+	if (nr == 0)
+		return -EINVAL;
+
+	memflags = blk_mq_freeze_queue(q);
+	mutex_lock(&q->elevator_lock);
+
+	if (q->elevator) {
+		q->async_depth = min(q->nr_requests, nr);
+		if (q->elevator->type->ops.depth_updated)
+			q->elevator->type->ops.depth_updated(q);
+	} else {
+		ret = -EINVAL;
+	}
+
+	mutex_unlock(&q->elevator_lock);
+	blk_mq_unfreeze_queue(q, memflags);
+
+	return ret;
+}
+
 static ssize_t queue_ra_show(struct gendisk *disk, char *page)
 {
 	ssize_t ret;
@@ -542,6 +587,7 @@ static struct queue_sysfs_entry _prefix##_entry = {	\
 }
 
 QUEUE_RW_ENTRY(queue_requests, "nr_requests");
+QUEUE_RW_ENTRY(queue_async_depth, "async_depth");
 QUEUE_RW_ENTRY(queue_ra, "read_ahead_kb");
 QUEUE_LIM_RW_ENTRY(queue_max_sectors, "max_sectors_kb");
 QUEUE_LIM_RO_ENTRY(queue_max_hw_sectors, "max_hw_sectors_kb");
@@ -764,6 +810,7 @@ static struct attribute *blk_mq_queue_attrs[] = {
 	 */
 	&elv_iosched_entry.attr,
 	&queue_requests_entry.attr,
+	&queue_async_depth_entry.attr,
 #ifdef CONFIG_BLK_WBT
 	&queue_wb_lat_entry.attr,
 #endif
diff --git a/block/elevator.c b/block/elevator.c
index e2ebfbf107b3..8f510cb881ba 100644
--- a/block/elevator.c
+++ b/block/elevator.c
@@ -601,6 +601,7 @@ static int elevator_switch(struct request_queue *q, struct elv_change_ctx *ctx)
 		blk_queue_flag_clear(QUEUE_FLAG_SQ_SCHED, q);
 		q->elevator = NULL;
 		q->nr_requests = q->tag_set->queue_depth;
+		q->async_depth = q->tag_set->queue_depth;
 	}
 	blk_add_trace_msg(q, "elv switch: %s", ctx->name);
 
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 02c006fb94c5..1d470ac71c64 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -542,6 +542,7 @@ struct request_queue {
 	 * queue settings
 	 */
 	unsigned int		nr_requests;	/* Max # of requests */
+	unsigned int		async_depth;	/* Max # of async requests */
 
 #ifdef CONFIG_BLK_INLINE_ENCRYPTION
 	struct blk_crypto_profile *crypto_profile;
-- 
2.39.2

Re: [patch v2 3/7] blk-mq: add a new queue sysfs attribute async_depth

[Nilay Shroff]

On 10/9/25 1:16 PM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Add a new field async_depth to request_queue and related APIs, this is
> currently not used, following patches will convert elevators to use
> this instead of internal async_depth.
> 
> Also factor out a helper blk_mq_limit_depth() to make code cleaner.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

Looks good to me:
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>

Re: [patch v2 3/7] blk-mq: add a new queue sysfs attribute async_depth

[Bart Van Assche]

On 10/9/25 12:46 AM, Yu Kuai wrote:
> +static ssize_t queue_async_depth_show(struct gendisk *disk, char *page)
> +{
> +	ssize_t ret;
> +
> +	mutex_lock(&disk->queue->elevator_lock);
> +	ret = queue_var_show(disk->queue->async_depth, page);
> +	mutex_unlock(&disk->queue->elevator_lock);
> +	return ret;
> +}

Functions like the above can be simplified by using guard(mutex)(...) or
scoped_guard(mutex, ...).

Thanks,

Bart.

Re: [patch v2 3/7] blk-mq: add a new queue sysfs attribute async_depth

[Yu Kuai]

Hi,

在 2025/10/10 1:05, Bart Van Assche 写道:
> On 10/9/25 12:46 AM, Yu Kuai wrote:
>> +static ssize_t queue_async_depth_show(struct gendisk *disk, char *page)
>> +{
>> +    ssize_t ret;
>> +
>> +    mutex_lock(&disk->queue->elevator_lock);
>> +    ret = queue_var_show(disk->queue->async_depth, page);
>> +    mutex_unlock(&disk->queue->elevator_lock);
>> +    return ret;
>> +}
>
> Functions like the above can be simplified by using guard(mutex)(...) or
> scoped_guard(mutex, ...).

Yeah, sounds good. I'm still not used to this.

Thanks,
Kuai

>
> Thanks,
>
> Bart.
>

[patch v2 4/7] kyber: covert to use request_queue->async_depth

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

Instead of the internal async_depth, remove kqd->async_depth and related
helpers.

Noted elevator attribute async_depth is now removed, queue attribute
with the same name is used instead.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/kyber-iosched.c | 33 +++++----------------------------
 1 file changed, 5 insertions(+), 28 deletions(-)

diff --git a/block/kyber-iosched.c b/block/kyber-iosched.c
index cf243a457175..469dad8a981c 100644
--- a/block/kyber-iosched.c
+++ b/block/kyber-iosched.c
@@ -47,9 +47,8 @@ enum {
 	 * asynchronous requests, we reserve 25% of requests for synchronous
 	 * operations.
 	 */
-	KYBER_ASYNC_PERCENT = 75,
+	KYBER_DEFAULT_ASYNC_PERCENT = 75,
 };
-
 /*
  * Maximum device-wide depth for each scheduling domain.
  *
@@ -157,9 +156,6 @@ struct kyber_queue_data {
 	 */
 	struct sbitmap_queue domain_tokens[KYBER_NUM_DOMAINS];
 
-	/* Number of allowed async requests. */
-	unsigned int async_depth;
-
 	struct kyber_cpu_latency __percpu *cpu_latency;
 
 	/* Timer for stats aggregation and adjusting domain tokens. */
@@ -401,10 +397,7 @@ static struct kyber_queue_data *kyber_queue_data_alloc(struct request_queue *q)
 
 static void kyber_depth_updated(struct request_queue *q)
 {
-	struct kyber_queue_data *kqd = q->elevator->elevator_data;
-
-	kqd->async_depth = q->nr_requests * KYBER_ASYNC_PERCENT / 100U;
-	blk_mq_set_min_shallow_depth(q, kqd->async_depth);
+	blk_mq_set_min_shallow_depth(q, q->async_depth);
 }
 
 static int kyber_init_sched(struct request_queue *q, struct elevator_queue *eq)
@@ -421,6 +414,7 @@ static int kyber_init_sched(struct request_queue *q, struct elevator_queue *eq)
 
 	eq->elevator_data = kqd;
 	q->elevator = eq;
+	q->async_depth = q->nr_requests * KYBER_DEFAULT_ASYNC_PERCENT / 100;
 	kyber_depth_updated(q);
 
 	return 0;
@@ -540,15 +534,8 @@ static void rq_clear_domain_token(struct kyber_queue_data *kqd,
 
 static void kyber_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 {
-	/*
-	 * We use the scheduler tags as per-hardware queue queueing tokens.
-	 * Async requests can be limited at this stage.
-	 */
-	if (!blk_mq_sched_sync_request(opf)) {
-		struct kyber_queue_data *kqd = data->q->elevator->elevator_data;
-
-		data->shallow_depth = kqd->async_depth;
-	}
+	if (!blk_mq_sched_sync_request(opf))
+		data->shallow_depth = data->q->async_depth;
 }
 
 static bool kyber_bio_merge(struct request_queue *q, struct bio *bio,
@@ -944,15 +931,6 @@ KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_DISCARD, discard)
 KYBER_DEBUGFS_DOMAIN_ATTRS(KYBER_OTHER, other)
 #undef KYBER_DEBUGFS_DOMAIN_ATTRS
 
-static int kyber_async_depth_show(void *data, struct seq_file *m)
-{
-	struct request_queue *q = data;
-	struct kyber_queue_data *kqd = q->elevator->elevator_data;
-
-	seq_printf(m, "%u\n", kqd->async_depth);
-	return 0;
-}
-
 static int kyber_cur_domain_show(void *data, struct seq_file *m)
 {
 	struct blk_mq_hw_ctx *hctx = data;
@@ -978,7 +956,6 @@ static const struct blk_mq_debugfs_attr kyber_queue_debugfs_attrs[] = {
 	KYBER_QUEUE_DOMAIN_ATTRS(write),
 	KYBER_QUEUE_DOMAIN_ATTRS(discard),
 	KYBER_QUEUE_DOMAIN_ATTRS(other),
-	{"async_depth", 0400, kyber_async_depth_show},
 	{},
 };
 #undef KYBER_QUEUE_DOMAIN_ATTRS
-- 
2.39.2

Re: [patch v2 4/7] kyber: covert to use request_queue->async_depth

[Nilay Shroff]

On 10/9/25 1:16 PM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Instead of the internal async_depth, remove kqd->async_depth and related
> helpers.
> 
> Noted elevator attribute async_depth is now removed, queue attribute
> with the same name is used instead.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

Looks good to me:
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>

[patch v2 5/7] mq-deadline: covert to use request_queue->async_depth

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

In downstream kernel, we test with mq-deadline with many fio workloads, and
we found a performance regression after commit 39823b47bbd4
("block/mq-deadline: Fix the tag reservation code") with following test:

[global]
rw=randread
direct=1
ramp_time=1
ioengine=libaio
iodepth=1024
numjobs=24
bs=1024k
group_reporting=1
runtime=60

[job1]
filename=/dev/sda

Root cause is that mq-deadline now support configuring async_depth,
although the default value is nr_request, however the minimal value is
1, hence min_shallow_depth is set to 1, causing wake_batch to be 1. For
consequence, sbitmap_queue will be waken up after each IO instead of
8 IO.

In this test case, sda is HDD and max_sectors is 128k, hence each
submitted 1M io will be splited into 8 sequential 128k requests, however
due to there are 24 jobs and total tags are exhausted, the 8 requests are
unlikely to be dispatched sequentially, and changing wake_batch to 1
will make this much worse, accounting blktrace D stage, the percentage
of sequential io is decreased from 8% to 0.8%.

Fix this problem by converting to request_queue->async_depth, where
min_shallow_depth is set each time async_depth is updated.

Noted elevator attribute async_depth is now removed, queue attribute
with the same name is used instead.

Fixes: 39823b47bbd4 ("block/mq-deadline: Fix the tag reservation code")
Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/mq-deadline.c | 39 +++++----------------------------------
 1 file changed, 5 insertions(+), 34 deletions(-)

diff --git a/block/mq-deadline.c b/block/mq-deadline.c
index 592dd853f6e5..87c7b7378397 100644
--- a/block/mq-deadline.c
+++ b/block/mq-deadline.c
@@ -98,7 +98,6 @@ struct deadline_data {
 	int fifo_batch;
 	int writes_starved;
 	int front_merges;
-	u32 async_depth;
 	int prio_aging_expire;
 
 	spinlock_t lock;
@@ -483,32 +482,16 @@ static struct request *dd_dispatch_request(struct blk_mq_hw_ctx *hctx)
 	return rq;
 }
 
-/*
- * Called by __blk_mq_alloc_request(). The shallow_depth value set by this
- * function is used by __blk_mq_get_tag().
- */
 static void dd_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
 {
-	struct deadline_data *dd = data->q->elevator->elevator_data;
-
-	/* Do not throttle synchronous reads. */
-	if (blk_mq_sched_sync_request(opf))
-		return;
-
-	/*
-	 * Throttle asynchronous requests and writes such that these requests
-	 * do not block the allocation of synchronous requests.
-	 */
-	data->shallow_depth = dd->async_depth;
+	if (!blk_mq_sched_sync_request(opf))
+		data->shallow_depth = data->q->async_depth;
 }
 
-/* Called by blk_mq_update_nr_requests(). */
+/* Called by blk_mq_init_sched() and blk_mq_update_nr_requests(). */
 static void dd_depth_updated(struct request_queue *q)
 {
-	struct deadline_data *dd = q->elevator->elevator_data;
-
-	dd->async_depth = q->nr_requests;
-	blk_mq_set_min_shallow_depth(q, 1);
+	blk_mq_set_min_shallow_depth(q, q->async_depth);
 }
 
 static void dd_exit_sched(struct elevator_queue *e)
@@ -573,6 +556,7 @@ static int dd_init_sched(struct request_queue *q, struct elevator_queue *eq)
 	blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
 
 	q->elevator = eq;
+	q->async_depth = q->nr_requests;
 	dd_depth_updated(q);
 	return 0;
 }
@@ -758,7 +742,6 @@ SHOW_JIFFIES(deadline_write_expire_show, dd->fifo_expire[DD_WRITE]);
 SHOW_JIFFIES(deadline_prio_aging_expire_show, dd->prio_aging_expire);
 SHOW_INT(deadline_writes_starved_show, dd->writes_starved);
 SHOW_INT(deadline_front_merges_show, dd->front_merges);
-SHOW_INT(deadline_async_depth_show, dd->async_depth);
 SHOW_INT(deadline_fifo_batch_show, dd->fifo_batch);
 #undef SHOW_INT
 #undef SHOW_JIFFIES
@@ -788,7 +771,6 @@ STORE_JIFFIES(deadline_write_expire_store, &dd->fifo_expire[DD_WRITE], 0, INT_MA
 STORE_JIFFIES(deadline_prio_aging_expire_store, &dd->prio_aging_expire, 0, INT_MAX);
 STORE_INT(deadline_writes_starved_store, &dd->writes_starved, INT_MIN, INT_MAX);
 STORE_INT(deadline_front_merges_store, &dd->front_merges, 0, 1);
-STORE_INT(deadline_async_depth_store, &dd->async_depth, 1, INT_MAX);
 STORE_INT(deadline_fifo_batch_store, &dd->fifo_batch, 0, INT_MAX);
 #undef STORE_FUNCTION
 #undef STORE_INT
@@ -802,7 +784,6 @@ static const struct elv_fs_entry deadline_attrs[] = {
 	DD_ATTR(write_expire),
 	DD_ATTR(writes_starved),
 	DD_ATTR(front_merges),
-	DD_ATTR(async_depth),
 	DD_ATTR(fifo_batch),
 	DD_ATTR(prio_aging_expire),
 	__ATTR_NULL
@@ -889,15 +870,6 @@ static int deadline_starved_show(void *data, struct seq_file *m)
 	return 0;
 }
 
-static int dd_async_depth_show(void *data, struct seq_file *m)
-{
-	struct request_queue *q = data;
-	struct deadline_data *dd = q->elevator->elevator_data;
-
-	seq_printf(m, "%u\n", dd->async_depth);
-	return 0;
-}
-
 static int dd_queued_show(void *data, struct seq_file *m)
 {
 	struct request_queue *q = data;
@@ -1007,7 +979,6 @@ static const struct blk_mq_debugfs_attr deadline_queue_debugfs_attrs[] = {
 	DEADLINE_NEXT_RQ_ATTR(write2),
 	{"batching", 0400, deadline_batching_show},
 	{"starved", 0400, deadline_starved_show},
-	{"async_depth", 0400, dd_async_depth_show},
 	{"dispatch0", 0400, .seq_ops = &deadline_dispatch0_seq_ops},
 	{"dispatch1", 0400, .seq_ops = &deadline_dispatch1_seq_ops},
 	{"dispatch2", 0400, .seq_ops = &deadline_dispatch2_seq_ops},
-- 
2.39.2

Re: [patch v2 5/7] mq-deadline: covert to use request_queue->async_depth

[Nilay Shroff]

On 10/9/25 1:16 PM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>

> Fix this problem by converting to request_queue->async_depth, where
> min_shallow_depth is set each time async_depth is updated.
> 
> Noted elevator attribute async_depth is now removed, queue attribute
> with the same name is used instead.
> 
> Fixes: 39823b47bbd4 ("block/mq-deadline: Fix the tag reservation code")
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

Looks good to me:
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>

Re: [patch v2 5/7] mq-deadline: covert to use request_queue->async_depth

[Bart Van Assche]

On 10/9/25 12:46 AM, Yu Kuai wrote:
> In downstream kernel, we test with mq-deadline with many fio workloads, and
> we found a performance regression after commit 39823b47bbd4
> ("block/mq-deadline: Fix the tag reservation code") with following test:
> 
> [global]
> rw=randread
> direct=1
> ramp_time=1
> ioengine=libaio
> iodepth=1024
> numjobs=24
> bs=1024k
> group_reporting=1
> runtime=60
> 
> [job1]
> filename=/dev/sda
> 
> Root cause is that mq-deadline now support configuring async_depth,
> although the default value is nr_request, however the minimal value is
> 1, hence min_shallow_depth is set to 1, causing wake_batch to be 1. For
> consequence, sbitmap_queue will be waken up after each IO instead of
> 8 IO.
> 
> In this test case, sda is HDD and max_sectors is 128k, hence each
> submitted 1M io will be splited into 8 sequential 128k requests, however
> due to there are 24 jobs and total tags are exhausted, the 8 requests are
> unlikely to be dispatched sequentially, and changing wake_batch to 1
> will make this much worse, accounting blktrace D stage, the percentage
> of sequential io is decreased from 8% to 0.8%.
> 
> Fix this problem by converting to request_queue->async_depth, where
> min_shallow_depth is set each time async_depth is updated.
> 
> Noted elevator attribute async_depth is now removed, queue attribute
> with the same name is used instead.

Reviewed-by: Bart Van Assche <bvanassche@acm.org>

[patch v2 6/7] block, bfq: convert to use request_queue->async_depth

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

The default limits is unchanged, and user can configure async_depth now.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/bfq-iosched.c | 43 +++++++++++++++++--------------------------
 1 file changed, 17 insertions(+), 26 deletions(-)

diff --git a/block/bfq-iosched.c b/block/bfq-iosched.c
index 63452e791f98..9e0eee9aba5c 100644
--- a/block/bfq-iosched.c
+++ b/block/bfq-iosched.c
@@ -7112,39 +7112,29 @@ void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
 static void bfq_depth_updated(struct request_queue *q)
 {
 	struct bfq_data *bfqd = q->elevator->elevator_data;
-	unsigned int nr_requests = q->nr_requests;
+	unsigned int async_depth = q->async_depth;
 
 	/*
-	 * In-word depths if no bfq_queue is being weight-raised:
-	 * leaving 25% of tags only for sync reads.
+	 * By default:
+	 *  - sync reads are not limited
+	 * If bfqq is not being weight-raised:
+	 *  - sync writes are limited to 75%(async depth default value)
+	 *  - async IO are limited to 50%
+	 * If bfqq is being weight-raised:
+	 *  - sync writes are limited to ~37%
+	 *  - async IO are limited to ~18
 	 *
-	 * In next formulas, right-shift the value
-	 * (1U<<bt->sb.shift), instead of computing directly
-	 * (1U<<(bt->sb.shift - something)), to be robust against
-	 * any possible value of bt->sb.shift, without having to
-	 * limit 'something'.
+	 * If request_queue->async_depth is updated by user, all limit are
+	 * updated relatively.
 	 */
-	/* no more than 50% of tags for async I/O */
-	bfqd->async_depths[0][0] = max(nr_requests >> 1, 1U);
-	/*
-	 * no more than 75% of tags for sync writes (25% extra tags
-	 * w.r.t. async I/O, to prevent async I/O from starving sync
-	 * writes)
-	 */
-	bfqd->async_depths[0][1] = max((nr_requests * 3) >> 2, 1U);
+	bfqd->async_depths[0][1] = async_depth;
+	bfqd->async_depths[0][0] = max(async_depth * 2 / 3, 1U);
+	bfqd->async_depths[1][1] = max(async_depth >> 1, 1U);
+	bfqd->async_depths[1][0] = max(async_depth >> 2, 1U);
 
 	/*
-	 * In-word depths in case some bfq_queue is being weight-
-	 * raised: leaving ~63% of tags for sync reads. This is the
-	 * highest percentage for which, in our tests, application
-	 * start-up times didn't suffer from any regression due to tag
-	 * shortage.
+	 * Due to cgroup qos, the allowed request for bfqq might be 1
 	 */
-	/* no more than ~18% of tags for async I/O */
-	bfqd->async_depths[1][0] = max((nr_requests * 3) >> 4, 1U);
-	/* no more than ~37% of tags for sync writes (~20% extra tags) */
-	bfqd->async_depths[1][1] = max((nr_requests * 6) >> 4, 1U);
-
 	blk_mq_set_min_shallow_depth(q, 1);
 }
 
@@ -7365,6 +7355,7 @@ static int bfq_init_queue(struct request_queue *q, struct elevator_queue *eq)
 	blk_queue_flag_set(QUEUE_FLAG_DISABLE_WBT_DEF, q);
 	wbt_disable_default(q->disk);
 	blk_stat_enable_accounting(q);
+	q->async_depth = (q->nr_requests * 3) >> 2;
 
 	return 0;
 
-- 
2.39.2

Re: [patch v2 6/7] block, bfq: convert to use request_queue->async_depth

[Nilay Shroff]

On 10/9/25 1:16 PM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> The default limits is unchanged, and user can configure async_depth now.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

Looks good to me:
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>

[patch v2 7/7] blk-mq: add documentation for new queue attribute async_dpeth

[Yu Kuai]

From: Yu Kuai <yukuai3@huawei.com>

Explain the attribute and the default value in different case.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 Documentation/ABI/stable/sysfs-block | 34 ++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)

diff --git a/Documentation/ABI/stable/sysfs-block b/Documentation/ABI/stable/sysfs-block
index 0ed10aeff86b..aa1e94169666 100644
--- a/Documentation/ABI/stable/sysfs-block
+++ b/Documentation/ABI/stable/sysfs-block
@@ -609,6 +609,40 @@ Description:
 		enabled, and whether tags are shared.
 
 
+What:		/sys/block/<disk>/queue/async_depth
+Date:		August 2025
+Contact:	linux-block@vger.kernel.org
+Description:
+		[RW] Controls how many asynchronous requests may be allocated in the
+		block layer. The value is always capped at nr_requests.
+
+		When no elevator is active (none):
+		- async_depth is always equal to nr_requests.
+
+		For bfq scheduler:
+		- By default, async_depth is set to 75% of nr_requests.
+		  Internal limits are then derived from this value:
+		  * Sync writes: limited to async_depth (≈75% of nr_requests).
+		  * Async I/O: limited to ~2/3 of async_depth (≈50% of nr_requests).
+
+		  If a bfq_queue is weight-raised:
+		  * Sync writes: limited to ~1/2 of async_depth (≈37% of nr_requests).
+		  * Async I/O: limited to ~1/4 of async_depth (≈18% of nr_requests).
+
+		- If the user writes a custom value to async_depth, BFQ will recompute
+		  these limits proportionally based on the new value.
+
+		For Kyber:
+		- By default async_depth is set to 75% of nr_requests.
+		- If the user writes a custom value to async_depth, then it override the
+		  default and directly control the limit for writes and async I/O.
+
+		For mq-deadline:
+		- By default async_depth is set to nr_requests.
+		- If the user writes a custom value to async_depth, then it override the
+		  default and directly control the limit for writes and async I/O.
+
+
 What:		/sys/block/<disk>/queue/nr_zones
 Date:		November 2018
 Contact:	Damien Le Moal <damien.lemoal@wdc.com>
-- 
2.39.2

Re: [patch v2 7/7] blk-mq: add documentation for new queue attribute async_dpeth

[Nilay Shroff]

On 10/9/25 1:16 PM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Explain the attribute and the default value in different case.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

Looks good to me:
Reviewed-by: Nilay Shroff <nilay@linux.ibm.com>