On 11/2/21 7:57 AM, Ming Lei wrote:
> On Tue, Nov 02, 2021 at 07:47:44AM -0600, Jens Axboe wrote:
>> On 11/2/21 7:35 AM, Ming Lei wrote:
>>> In case of shared tags and none io sched, batched completion still may
>>> be run into, and hctx->nr_active is accounted when getting driver tag,
>>> so it has to be updated in blk_mq_end_request_batch().
>>>
>>> Otherwise, hctx->nr_active may become same with queue depth, then
>>> hctx_may_queue() always return false, then io hang is caused.
>>>
>>> Fixes the issue by updating the counter in batched way.
>>>
>>> Reported-by: Shinichiro Kawasaki <shinichiro.kawasaki@xxxxxxx>
>>> Fixes: f794f3351f26 ("block: add support for blk_mq_end_request_batch()")
>>> Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx>
>>> ---
>>> block/blk-mq.c | 15 +++++++++++++--
>>> block/blk-mq.h | 12 +++++++++---
>>> 2 files changed, 22 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>> index 07eb1412760b..0dbe75034f61 100644
>>> --- a/block/blk-mq.c
>>> +++ b/block/blk-mq.c
>>> @@ -825,6 +825,7 @@ void blk_mq_end_request_batch(struct io_comp_batch *iob)
>>> struct blk_mq_hw_ctx *cur_hctx = NULL;
>>> struct request *rq;
>>> u64 now = 0;
>>> + int active = 0;
>>>
>>> if (iob->need_ts)
>>> now = ktime_get_ns();
>>> @@ -846,16 +847,26 @@ void blk_mq_end_request_batch(struct io_comp_batch *iob)
>>> rq_qos_done(rq->q, rq);
>>>
>>> if (nr_tags == TAG_COMP_BATCH || cur_hctx != rq->mq_hctx) {
>>> - if (cur_hctx)
>>> + if (cur_hctx) {
>>> + if (active)
>>> + __blk_mq_sub_active_requests(cur_hctx,
>>> + active);
>>> blk_mq_flush_tag_batch(cur_hctx, tags, nr_tags);
>>> + }
>>> nr_tags = 0;
>>> + active = 0;
>>> cur_hctx = rq->mq_hctx;
>>> }
>>> tags[nr_tags++] = rq->tag;
>>> + if (rq->rq_flags & RQF_MQ_INFLIGHT)
>>> + active++;
>>
>> Are there any cases where either none or all of requests have the
>> flag set, and hence active == nr_tags?
>
> none and BLK_MQ_F_TAG_QUEUE_SHARED, and Shinichiro only observed the
> issue on two NSs.
Maybe I wasn't clear enough. What I'm saying is that either all of the
requests will have RQF_MQ_INFLIGHT set, or none of them. Hence active
should be either 0, or == nr_tags.
That's the hypothesis that I wanted to check, because if that's true,
then we can do this in a better way.
--
Jens Axboe
