On 11/3/21 9:41 AM, Jens Axboe wrote: > On 11/3/21 9:16 AM, Ming Lei wrote: >> On Wed, Nov 03, 2021 at 09:10:17AM -0600, Jens Axboe wrote: >>> On 11/3/21 9:03 AM, Jens Axboe wrote: >>>> On 11/3/21 8:57 AM, Ming Lei wrote: >>>>> On Wed, Nov 03, 2021 at 09:59:02PM +0800, Yi Zhang wrote: >>>>>> On Wed, Nov 3, 2021 at 7:59 PM Jens Axboe <axboe@xxxxxxxxx> wrote: >>>>>>> >>>>>>> On 11/2/21 9:54 PM, Jens Axboe wrote: >>>>>>>> On Nov 2, 2021, at 9:52 PM, Ming Lei <ming.lei@xxxxxxxxxx> wrote: >>>>>>>>> >>>>>>>>> On Tue, Nov 02, 2021 at 09:21:10PM -0600, Jens Axboe wrote: >>>>>>>>>>> On 11/2/21 8:21 PM, Yi Zhang wrote: >>>>>>>>>>>>> >>>>>>>>>>>>> Can either one of you try with this patch? Won't fix anything, but it'll >>>>>>>>>>>>> hopefully shine a bit of light on the issue. >>>>>>>>>>>>> >>>>>>>>>>> Hi Jens >>>>>>>>>>> >>>>>>>>>>> Here is the full log: >>>>>>>>>> >>>>>>>>>> Thanks! I think I see what it could be - can you try this one as well, >>>>>>>>>> would like to confirm that the condition I think is triggering is what >>>>>>>>>> is triggering. >>>>>>>>>> >>>>>>>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c >>>>>>>>>> index 07eb1412760b..81dede885231 100644 >>>>>>>>>> --- a/block/blk-mq.c >>>>>>>>>> +++ b/block/blk-mq.c >>>>>>>>>> @@ -2515,6 +2515,8 @@ void blk_mq_submit_bio(struct bio *bio) >>>>>>>>>> if (plug && plug->cached_rq) { >>>>>>>>>> rq = rq_list_pop(&plug->cached_rq); >>>>>>>>>> INIT_LIST_HEAD(&rq->queuelist); >>>>>>>>>> + WARN_ON_ONCE(q->elevator && !(rq->rq_flags & RQF_ELV)); >>>>>>>>>> + WARN_ON_ONCE(!q->elevator && (rq->rq_flags & RQF_ELV)); >>>>>>>>>> } else { >>>>>>>>>> struct blk_mq_alloc_data data = { >>>>>>>>>> .q = q, >>>>>>>>>> @@ -2535,6 +2537,8 @@ void blk_mq_submit_bio(struct bio *bio) >>>>>>>>>> bio_wouldblock_error(bio); >>>>>>>>>> goto queue_exit; >>>>>>>>>> } >>>>>>>>>> + WARN_ON_ONCE(q->elevator && !(rq->rq_flags & RQF_ELV)); >>>>>>>>>> + WARN_ON_ONCE(!q->elevator && (rq->rq_flags & RQF_ELV)); >>>>>>>>> >>>>>>>>> Hello Jens, >>>>>>>>> >>>>>>>>> I guess the issue could be the following code run without grabbing >>>>>>>>> ->q_usage_counter from blk_mq_alloc_request() and blk_mq_alloc_request_hctx(). >>>>>>>>> >>>>>>>>> .rq_flags = q->elevator ? RQF_ELV : 0, >>>>>>>>> >>>>>>>>> then elevator is switched to real one from none, and check on q->elevator >>>>>>>>> becomes not consistent. >>>>>>>> >>>>>>>> Indeed, that’s where I was going with this. I have a patch, testing it >>>>>>>> locally but it’s getting late. Will send it out tomorrow. The nice >>>>>>>> benefit is that it allows dropping the weird ref get on plug flush, >>>>>>>> and batches getting the refs as well. >>>>>>> >>>>>>> Yi/Steffen, can you try pulling this into your test kernel: >>>>>>> >>>>>>> git://git.kernel.dk/linux-block for-next >>>>>>> >>>>>>> and see if it fixes the issue for you. Thanks! >>>>>> >>>>>> It still can be reproduced with the latest linux-block/for-next, here is the log >>>>>> >>>>>> fab2914e46eb (HEAD, new/for-next) Merge branch 'for-5.16/drivers' into for-next >>>>> >>>>> Hi Yi, >>>>> >>>>> Please try the following change: >>>>> >>>>> >>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c >>>>> index e1e64964a31b..eb634a9c61ff 100644 >>>>> --- a/block/blk-mq.c >>>>> +++ b/block/blk-mq.c >>>>> @@ -494,7 +494,6 @@ struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, >>>>> .q = q, >>>>> .flags = flags, >>>>> .cmd_flags = op, >>>>> - .rq_flags = q->elevator ? RQF_ELV : 0, >>>>> .nr_tags = 1, >>>>> }; >>>>> struct request *rq; >>>>> @@ -504,6 +503,7 @@ struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, >>>>> if (ret) >>>>> return ERR_PTR(ret); >>>>> >>>>> + data.rq_flags = q->elevator ? RQF_ELV : 0, >>>>> rq = __blk_mq_alloc_requests(&data); >>>>> if (!rq) >>>>> goto out_queue_exit; >>>>> @@ -524,7 +524,6 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q, >>>>> .q = q, >>>>> .flags = flags, >>>>> .cmd_flags = op, >>>>> - .rq_flags = q->elevator ? RQF_ELV : 0, >>>>> .nr_tags = 1, >>>>> }; >>>>> u64 alloc_time_ns = 0; >>>>> @@ -551,6 +550,7 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q, >>>>> ret = blk_queue_enter(q, flags); >>>>> if (ret) >>>>> return ERR_PTR(ret); >>>>> + data.rq_flags = q->elevator ? RQF_ELV : 0, >>>> >>>> Don't think that will compile, but I guess the point is that we can't do >>>> this assignment before queue enter, in case we're in the midst of >>>> switching schedulers. Which is indeed a valid concern. >>> >>> Something like the below. Maybe? On top of the for-next that was already >>> pulled in. >>> >>> >>> diff --git a/block/blk-mq.c b/block/blk-mq.c >>> index b01e05e02277..121f1898d529 100644 >>> --- a/block/blk-mq.c >>> +++ b/block/blk-mq.c >>> @@ -433,9 +433,11 @@ static struct request *__blk_mq_alloc_requests(struct blk_mq_alloc_data *data) >>> if (data->cmd_flags & REQ_NOWAIT) >>> data->flags |= BLK_MQ_REQ_NOWAIT; >>> >>> - if (data->rq_flags & RQF_ELV) { >>> + if (q->elevator) { >>> struct elevator_queue *e = q->elevator; >>> >>> + data->rq_flags |= RQF_ELV; >>> + >>> /* >>> * Flush/passthrough requests are special and go directly to the >>> * dispatch list. Don't include reserved tags in the >>> @@ -494,7 +496,6 @@ struct request *blk_mq_alloc_request(struct request_queue *q, unsigned int op, >>> .q = q, >>> .flags = flags, >>> .cmd_flags = op, >>> - .rq_flags = q->elevator ? RQF_ELV : 0, >>> .nr_tags = 1, >>> }; >>> struct request *rq; >>> @@ -524,7 +525,6 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q, >>> .q = q, >>> .flags = flags, >>> .cmd_flags = op, >>> - .rq_flags = q->elevator ? RQF_ELV : 0, >>> .nr_tags = 1, >>> }; >>> u64 alloc_time_ns = 0; >>> @@ -565,6 +565,8 @@ struct request *blk_mq_alloc_request_hctx(struct request_queue *q, >>> >>> if (!q->elevator) >>> blk_mq_tag_busy(data.hctx); >>> + else >>> + data.rq_flags |= RQF_ELV; >>> >>> ret = -EWOULDBLOCK; >>> tag = blk_mq_get_tag(&data); >>> @@ -2560,7 +2562,6 @@ void blk_mq_submit_bio(struct bio *bio) >>> .q = q, >>> .nr_tags = 1, >>> .cmd_flags = bio->bi_opf, >>> - .rq_flags = q->elevator ? RQF_ELV : 0, >>> }; >> >> The above patch looks fine. >> >> BTW, 9ede85cb670c ("block: move queue enter logic into >> blk_mq_submit_bio()") moves the queue enter into blk_mq_submit_bio(), >> which seems dangerous, especially blk_mq_sched_bio_merge() needs >> hctx/ctx which requires q_usage_counter to be grabbed. > > I think the best solution is to enter just for that as well, and just > retain that enter state. I'll update the patch, there's some real fixes > in there too for the batched alloc. Will post them later today. Is it needed, though? As far as I can tell, it's only needed persistently for having the IO inflight, otherwise if the premise is that the queue can just go away, we're in trouble before that too. And I don't think that's the case. -- Jens Axboe