Re: [PATCH] blk-mq: test QUEUE_FLAG_HCTX_ACTIVE for sbitmap_shared in hctx_may_queue

[John Garry <john.garry@huawei.com>]

On 05/01/2021 02:20, Ming Lei wrote:
> On Mon, Jan 04, 2021 at 10:41:36AM +0000, John Garry wrote:
>> On 27/12/2020 11:34, Ming Lei wrote:
>>> In case of blk_mq_is_sbitmap_shared(), we should test QUEUE_FLAG_HCTX_ACTIVE against
>>> q->queue_flags instead of BLK_MQ_S_TAG_ACTIVE.
>>>
>>> So fix it.
>>>
>>> Cc: John Garry<john.garry@huawei.com>
>>> Cc: Kashyap Desai<kashyap.desai@broadcom.com>
>>> Fixes: f1b49fdc1c64 ("blk-mq: Record active_queues_shared_sbitmap per tag_set for when using shared sbitmap")
>>> Signed-off-by: Ming Lei<ming.lei@redhat.com>
>> Reviewed-by: John Garry<john.garry@huawei.com>
>>
>>> ---
>>>    block/blk-mq.h | 2 +-
>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/block/blk-mq.h b/block/blk-mq.h
>>> index c1458d9502f1..3616453ca28c 100644
>>> --- a/block/blk-mq.h
>>> +++ b/block/blk-mq.h
>>> @@ -304,7 +304,7 @@ 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 (!test_bit(BLK_MQ_S_TAG_ACTIVE, &q->queue_flags))
>>> +		if (!test_bit(QUEUE_FLAG_HCTX_ACTIVE, &q->queue_flags))
>> I wonder how this ever worked properly, as BLK_MQ_S_TAG_ACTIVE is bit index
>> 1, and for q->queue_flags that means QUEUE_FLAG_DYING bit, which I figure is
>> not set normally..
> It always return true, and might just take a bit more CPU especially the tag queue
> depth of magsas_raid and hisi_sas_v3 is quite high.

Hi Ming,

Right, but we actually tested by hacking the host tag queue depth to be 
lower such that we should have tag contention, here is an extract from 
the original series cover letter for my results:

Tag depth 		4000 (default)		260**

Baseline (v5.9-rc1):
none sched:		2094K IOPS		513K
mq-deadline sched:	2145K IOPS		1336K

Final, host_tagset=0 in LLDD *, ***:
none sched:		2120K IOPS		550K
mq-deadline sched:	2121K IOPS		1309K

Final ***:
none sched:		2132K IOPS		1185		
mq-deadline sched:	2145K IOPS		2097	

Maybe my test did not expose the issue. Kashyap also tested this and 
reported the original issue such that we needed this feature, so I'm 
confused.

Thanks,
John