From: Yu Kuai <yukuai3@huawei.com>

Currently, both mq-deadline and bfq have global spin lock that will be
grabbed inside elevator methods like dispatch_request, insert_requests,
and bio_merge. And the global lock is the main reason mq-deadline and
bfq can't scale very well.

While dispatching request, blk_mq_get_disatpch_budget() and
blk_mq_get_driver_tag() must be called, and they are not ready to be called
inside elevator methods, hence introduce a new method like
dispatch_requests is not possible.

Hence introduce a new high level elevator lock, currently it is protecting
dispatch_request only. Following patches will convert mq-deadline and bfq
to use this lock and finally support request batch dispatching by calling
the method multiple time while holding the lock.

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/blk-mq-sched.c | 6 ++++++
 block/elevator.c     | 1 +
 block/elevator.h     | 4 ++--
 3 files changed, 9 insertions(+), 2 deletions(-)

diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c
index 55a0fd105147..7911fae75ce4 100644
--- a/block/blk-mq-sched.c
+++ b/block/blk-mq-sched.c
@@ -98,6 +98,7 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 		max_dispatch = hctx->queue->nr_requests;
 
 	do {
+		bool sq_sched = blk_queue_sq_sched(q);
 		struct request *rq;
 		int budget_token;
 
@@ -113,7 +114,12 @@ static int __blk_mq_do_dispatch_sched(struct blk_mq_hw_ctx *hctx)
 		if (budget_token < 0)
 			break;
 
+		if (sq_sched)
+			spin_lock(&e->lock);
 		rq = e->type->ops.dispatch_request(hctx);
+		if (sq_sched)
+			spin_unlock(&e->lock);
+
 		if (!rq) {
 			blk_mq_put_dispatch_budget(q, budget_token);
 			/*
diff --git a/block/elevator.c b/block/elevator.c
index 88f8f36bed98..45303af0ca73 100644
--- a/block/elevator.c
+++ b/block/elevator.c
@@ -144,6 +144,7 @@ struct elevator_queue *elevator_alloc(struct request_queue *q,
 	eq->type = e;
 	kobject_init(&eq->kobj, &elv_ktype);
 	mutex_init(&eq->sysfs_lock);
+	spin_lock_init(&eq->lock);
 	hash_init(eq->hash);
 
 	return eq;
diff --git a/block/elevator.h b/block/elevator.h
index a07ce773a38f..cbbac4f7825c 100644
--- a/block/elevator.h
+++ b/block/elevator.h
@@ -110,12 +110,12 @@ struct request *elv_rqhash_find(struct request_queue *q, sector_t offset);
 /*
  * each queue has an elevator_queue associated with it
  */
-struct elevator_queue
-{
+struct elevator_queue {
 	struct elevator_type *type;
 	void *elevator_data;
 	struct kobject kobj;
 	struct mutex sysfs_lock;
+	spinlock_t lock;
 	unsigned long flags;
 	DECLARE_HASHTABLE(hash, ELV_HASH_BITS);
 };
-- 
2.39.2

On 7/30/25 10:22, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Currently, both mq-deadline and bfq have global spin lock that will be
> grabbed inside elevator methods like dispatch_request, insert_requests,
> and bio_merge. And the global lock is the main reason mq-deadline and
> bfq can't scale very well.
> 
> While dispatching request, blk_mq_get_disatpch_budget() and
> blk_mq_get_driver_tag() must be called, and they are not ready to be called
> inside elevator methods, hence introduce a new method like
> dispatch_requests is not possible.
> 
> Hence introduce a new high level elevator lock, currently it is protecting
> dispatch_request only. Following patches will convert mq-deadline and bfq
> to use this lock and finally support request batch dispatching by calling
> the method multiple time while holding the lock.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> ---
>   block/blk-mq-sched.c | 6 ++++++
>   block/elevator.c     | 1 +
>   block/elevator.h     | 4 ++--
>   3 files changed, 9 insertions(+), 2 deletions(-)
> 
Reviewed-by: Hannes Reinecke <hare@suse.de>

Cheers,

Hannes
-- 
Dr. Hannes Reinecke                  Kernel Storage Architect
hare@suse.de                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich

On 7/30/25 1:22 AM, Yu Kuai wrote:
> +		if (sq_sched)
> +			spin_lock(&e->lock);
>   		rq = e->type->ops.dispatch_request(hctx);
> +		if (sq_sched)
> +			spin_unlock(&e->lock);

The above will confuse static analyzers. Please change it into something
like the following:

if (blk_queue_sq_sched(q)) {
	spin_lock(&e->lock);
	rq = e->type->ops.dispatch_request(hctx);
	spin_unlock(&e->lock);
} else {
	rq = e->type->ops.dispatch_request(hctx);
}

Otherwise this patch looks good to me.

Thanks,

Bart.

Hi

在 2025/7/31 1:19, Bart Van Assche 写道:
> On 7/30/25 1:22 AM, Yu Kuai wrote:
>> +        if (sq_sched)
>> +            spin_lock(&e->lock);
>>           rq = e->type->ops.dispatch_request(hctx);
>> +        if (sq_sched)
>> +            spin_unlock(&e->lock);
>
> The above will confuse static analyzers. Please change it into something
> like the following:
>
> if (blk_queue_sq_sched(q)) {
>     spin_lock(&e->lock);
>     rq = e->type->ops.dispatch_request(hctx);
>     spin_unlock(&e->lock);
> } else {
>     rq = e->type->ops.dispatch_request(hctx);
> }
>
> Otherwise this patch looks good to me.
Ok, thanks for the review, will change in the next version.
Kuai

>
> Thanks,
>
> Bart.
>