cleanup and fixes for updating nr_requests

[PATCH for-6.18/block 08/10] blk-mq: fix potential deadlock while nr_requests grown

Posted by Yu Kuai 5 months ago

From: Yu Kuai <yukuai3@huawei.com>

Allocate and free sched_tags while queue is freezed can deadlock[1],
this is a long term problem, hence allocate memory before freezing
queue and free memory after queue is unfreezed.

[1] https://lore.kernel.org/all/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/
Fixes: e3a2b3f931f5 ("blk-mq: allow changing of queue depth through sysfs")

Signed-off-by: Yu Kuai <yukuai3@huawei.com>
---
 block/blk-mq.c    | 22 ++++++++++------------
 block/blk-mq.h    |  5 ++++-
 block/blk-sysfs.c | 25 +++++++++++++++++--------
 3 files changed, 31 insertions(+), 21 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 82fa81036115..d85afbb9f031 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -4917,11 +4917,13 @@ void blk_mq_free_tag_set(struct blk_mq_tag_set *set)
 }
 EXPORT_SYMBOL(blk_mq_free_tag_set);
 
-int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
+struct elevator_tags *blk_mq_update_nr_requests(struct request_queue *q,
+						struct elevator_tags *et,
+						unsigned int nr)
 {
 	struct blk_mq_tag_set *set = q->tag_set;
+	struct elevator_tags *old_et = NULL;
 	struct blk_mq_hw_ctx *hctx;
-	int ret = 0;
 	unsigned long i;
 
 	blk_mq_quiesce_queue(q);
@@ -4946,23 +4948,19 @@ int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr)
 				nr - hctx->sched_tags->nr_reserved_tags);
 		}
 	} else {
-		queue_for_each_hw_ctx(q, hctx, i) {
-			if (!hctx->sched_tags)
-				continue;
-			blk_mq_tag_update_depth(hctx, &hctx->sched_tags, nr);
-			if (ret)
-				goto out;
-		}
+		queue_for_each_hw_ctx(q, hctx, i)
+			hctx->sched_tags = et->tags[i];
+
+		old_et = q->elevator->et;
+		q->elevator->et = et;
 	}
 
 	q->nr_requests = nr;
 	if (q->elevator && q->elevator->type->ops.depth_updated)
 		q->elevator->type->ops.depth_updated(q);
 
-out:
 	blk_mq_unquiesce_queue(q);
-
-	return ret;
+	return old_et;
 }
 
 /*
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 731f4578d9a8..6c9d03625ba1 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -6,6 +6,7 @@
 #include "blk-stat.h"
 
 struct blk_mq_tag_set;
+struct elevator_tags;
 
 struct blk_mq_ctxs {
 	struct kobject kobj;
@@ -45,7 +46,9 @@ void blk_mq_submit_bio(struct bio *bio);
 int blk_mq_poll(struct request_queue *q, blk_qc_t cookie, struct io_comp_batch *iob,
 		unsigned int flags);
 void blk_mq_exit_queue(struct request_queue *q);
-int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
+struct elevator_tags *blk_mq_update_nr_requests(struct request_queue *q,
+						struct elevator_tags *tags,
+						unsigned int nr);
 void blk_mq_wake_waiters(struct request_queue *q);
 bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx *hctx, struct list_head *,
 			     bool);
diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
index 7ea15bf68b4b..a0a7ebad378f 100644
--- a/block/blk-sysfs.c
+++ b/block/blk-sysfs.c
@@ -64,11 +64,12 @@ static ssize_t queue_requests_show(struct gendisk *disk, char *page)
 static ssize_t
 queue_requests_store(struct gendisk *disk, const char *page, size_t count)
 {
-	unsigned long nr;
-	int ret, err;
-	unsigned int memflags;
 	struct request_queue *q = disk->queue;
 	struct blk_mq_tag_set *set = q->tag_set;
+	struct elevator_tags *et = NULL;
+	unsigned int memflags;
+	unsigned long nr;
+	int ret;
 
 	ret = queue_var_store(&nr, page, count);
 	if (ret < 0)
@@ -90,16 +91,24 @@ queue_requests_store(struct gendisk *disk, const char *page, size_t count)
 		goto unlock;
 	}
 
+	if (q->elevator && nr > q->elevator->et->nr_requests) {
+		/* allocate memory before freezing queue to prevent deadlock */
+		et = blk_mq_alloc_sched_tags(set, q->nr_hw_queues, nr);
+		if (!et) {
+			ret = -ENOMEM;
+			goto unlock;
+		}
+	}
+
 	memflags = blk_mq_freeze_queue(q);
 	mutex_lock(&q->elevator_lock);
-
-	err = blk_mq_update_nr_requests(disk->queue, nr);
-	if (err)
-		ret = err;
-
+	et = blk_mq_update_nr_requests(q, et, nr);
 	mutex_unlock(&q->elevator_lock);
 	blk_mq_unfreeze_queue(q, memflags);
 
+	if (et)
+		blk_mq_free_sched_tags(et, set);
+
 unlock:
 	up_write(&set->update_nr_hwq_lock);
 	return ret;
-- 
2.39.2

Re: [PATCH for-6.18/block 08/10] blk-mq: fix potential deadlock while nr_requests grown

Posted by Yu Kuai 5 months ago

Hi,

在 2025/09/08 14:15, Yu Kuai 写道:
> @@ -90,16 +91,24 @@ queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>   		goto unlock;
>   	}
>   
> +	if (q->elevator && nr > q->elevator->et->nr_requests) {

While rebasing v2, I found that I should also add non-shared checking
here, because from blk_mq_alloc_shced_tags(), et->nr_requests is not set
to MAX_SCHED_RQ, and we don't want to allocate memory for shared case.

I'll fix this as well in v2.

Thanks,
Kuai

Re: [PATCH for-6.18/block 08/10] blk-mq: fix potential deadlock while nr_requests grown

Posted by Nilay Shroff 5 months ago


On 9/8/25 11:45 AM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Allocate and free sched_tags while queue is freezed can deadlock[1],
> this is a long term problem, hence allocate memory before freezing
> queue and free memory after queue is unfreezed.
> 
> [1] https://lore.kernel.org/all/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/
> Fixes: e3a2b3f931f5 ("blk-mq: allow changing of queue depth through sysfs")
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>

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

Re: [PATCH for-6.18/block 08/10] blk-mq: fix potential deadlock while nr_requests grown

Posted by Nilay Shroff 5 months ago


On 9/8/25 11:45 AM, Yu Kuai wrote:
> From: Yu Kuai <yukuai3@huawei.com>
> 
> Allocate and free sched_tags while queue is freezed can deadlock[1],
> this is a long term problem, hence allocate memory before freezing
> queue and free memory after queue is unfreezed.
> 
> [1] https://lore.kernel.org/all/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/
> Fixes: e3a2b3f931f5 ("blk-mq: allow changing of queue depth through sysfs")
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
[...]
[...]

> diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
> index 7ea15bf68b4b..a0a7ebad378f 100644
> --- a/block/blk-sysfs.c
> +++ b/block/blk-sysfs.c
> @@ -64,11 +64,12 @@ static ssize_t queue_requests_show(struct gendisk *disk, char *page)
>  static ssize_t
>  queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>  {
> -	unsigned long nr;
> -	int ret, err;
> -	unsigned int memflags;
>  	struct request_queue *q = disk->queue;
>  	struct blk_mq_tag_set *set = q->tag_set;
> +	struct elevator_tags *et = NULL;
> +	unsigned int memflags;
> +	unsigned long nr;
> +	int ret;
>  
>  	ret = queue_var_store(&nr, page, count);
>  	if (ret < 0)
> @@ -90,16 +91,24 @@ queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>  		goto unlock;
>  	}
>  
> +	if (q->elevator && nr > q->elevator->et->nr_requests) {
> +		/* allocate memory before freezing queue to prevent deadlock */
> +		et = blk_mq_alloc_sched_tags(set, q->nr_hw_queues, nr);
> +		if (!et) {
> +			ret = -ENOMEM;
> +			goto unlock;
> +		}
> +	}
> +
I think we should add a comment above explaining why is it safe 
to access q->elevator without holding ->elevator_lock.

Thanks,
--Nilay

Re: [PATCH for-6.18/block 08/10] blk-mq: fix potential deadlock while nr_requests grown

Posted by Yu Kuai 5 months ago

Hi,

在 2025/09/09 14:39, Nilay Shroff 写道:
> 
> 
> On 9/8/25 11:45 AM, Yu Kuai wrote:
>> From: Yu Kuai <yukuai3@huawei.com>
>>
>> Allocate and free sched_tags while queue is freezed can deadlock[1],
>> this is a long term problem, hence allocate memory before freezing
>> queue and free memory after queue is unfreezed.
>>
>> [1] https://lore.kernel.org/all/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/
>> Fixes: e3a2b3f931f5 ("blk-mq: allow changing of queue depth through sysfs")
>>
>> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> [...]
> [...]
> 
>> diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
>> index 7ea15bf68b4b..a0a7ebad378f 100644
>> --- a/block/blk-sysfs.c
>> +++ b/block/blk-sysfs.c
>> @@ -64,11 +64,12 @@ static ssize_t queue_requests_show(struct gendisk *disk, char *page)
>>   static ssize_t
>>   queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>>   {
>> -	unsigned long nr;
>> -	int ret, err;
>> -	unsigned int memflags;
>>   	struct request_queue *q = disk->queue;
>>   	struct blk_mq_tag_set *set = q->tag_set;
>> +	struct elevator_tags *et = NULL;
>> +	unsigned int memflags;
>> +	unsigned long nr;
>> +	int ret;
>>   
>>   	ret = queue_var_store(&nr, page, count);
>>   	if (ret < 0)
>> @@ -90,16 +91,24 @@ queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>>   		goto unlock;
>>   	}
>>   
>> +	if (q->elevator && nr > q->elevator->et->nr_requests) {
>> +		/* allocate memory before freezing queue to prevent deadlock */
>> +		et = blk_mq_alloc_sched_tags(set, q->nr_hw_queues, nr);
>> +		if (!et) {
>> +			ret = -ENOMEM;
>> +			goto unlock;
>> +		}
>> +	}
>> +
> I think we should add a comment above explaining why is it safe
> to access q->elevator without holding ->elevator_lock.
> 

I already access q->elevator to check input nr from patch 4, and that's
why I add comments to explain switching elevator is serialized, is this
enough?

Thanks,
Kuai

> Thanks,
> --Nilay
> .
>

Re: [PATCH for-6.18/block 08/10] blk-mq: fix potential deadlock while nr_requests grown

Posted by Nilay Shroff 5 months ago


On 9/9/25 1:07 PM, Yu Kuai wrote:
> Hi,
> 
> 在 2025/09/09 14:39, Nilay Shroff 写道:
>>
>>
>> On 9/8/25 11:45 AM, Yu Kuai wrote:
>>> From: Yu Kuai <yukuai3@huawei.com>
>>>
>>> Allocate and free sched_tags while queue is freezed can deadlock[1],
>>> this is a long term problem, hence allocate memory before freezing
>>> queue and free memory after queue is unfreezed.
>>>
>>> [1] https://lore.kernel.org/all/0659ea8d-a463-47c8-9180-43c719e106eb@linux.ibm.com/
>>> Fixes: e3a2b3f931f5 ("blk-mq: allow changing of queue depth through sysfs")
>>>
>>> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
>> [...]
>> [...]
>>
>>> diff --git a/block/blk-sysfs.c b/block/blk-sysfs.c
>>> index 7ea15bf68b4b..a0a7ebad378f 100644
>>> --- a/block/blk-sysfs.c
>>> +++ b/block/blk-sysfs.c
>>> @@ -64,11 +64,12 @@ static ssize_t queue_requests_show(struct gendisk *disk, char *page)
>>>   static ssize_t
>>>   queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>>>   {
>>> -    unsigned long nr;
>>> -    int ret, err;
>>> -    unsigned int memflags;
>>>       struct request_queue *q = disk->queue;
>>>       struct blk_mq_tag_set *set = q->tag_set;
>>> +    struct elevator_tags *et = NULL;
>>> +    unsigned int memflags;
>>> +    unsigned long nr;
>>> +    int ret;
>>>         ret = queue_var_store(&nr, page, count);
>>>       if (ret < 0)
>>> @@ -90,16 +91,24 @@ queue_requests_store(struct gendisk *disk, const char *page, size_t count)
>>>           goto unlock;
>>>       }
>>>   +    if (q->elevator && nr > q->elevator->et->nr_requests) {
>>> +        /* allocate memory before freezing queue to prevent deadlock */
>>> +        et = blk_mq_alloc_sched_tags(set, q->nr_hw_queues, nr);
>>> +        if (!et) {
>>> +            ret = -ENOMEM;
>>> +            goto unlock;
>>> +        }
>>> +    }
>>> +
>> I think we should add a comment above explaining why is it safe
>> to access q->elevator without holding ->elevator_lock.
>>
> 
> I already access q->elevator to check input nr from patch 4, and that's
> why I add comments to explain switching elevator is serialized, is this
> enough?
> 
yes in patch 04/10 you moved the ->elevator_lock after the 
usual sanity checks. However when we run those sanity checks
or the code in this patch where we have to access q->elevator,
it's good to add a comment here in the code (not in commit).
For reference, you may check blk_mq_alloc_sched_tags_batch.
I think similar comment may be added here as well.

Thanks,
--Nilay