lkml.org 
[lkml]   [2022]   [Sep]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: stalling IO regression since linux 5.12, through 5.18
From
Date


> Il giorno 1 set 2022, alle ore 10:19, Yu Kuai <yukuai1@huaweicloud.com> ha scritto:
>
> 在 2022/09/01 16:03, Jan Kara 写道:
>> On Thu 01-09-22 15:02:03, Yu Kuai wrote:
>>> Hi, Chris
>>>
>>> 在 2022/08/20 15:00, Ming Lei 写道:
>>>> On Fri, Aug 19, 2022 at 03:20:25PM -0400, Chris Murphy wrote:
>>>>>
>>>>>
>>>>> On Thu, Aug 18, 2022, at 1:24 AM, Ming Lei wrote:
>>>>>> On Thu, Aug 18, 2022 at 12:27:04AM -0400, Chris Murphy wrote:
>>>>>>>
>>>>>>>
>>>>>>> On Thu, Aug 18, 2022, at 12:18 AM, Chris Murphy wrote:
>>>>>>>> On Thu, Aug 18, 2022, at 12:12 AM, Chris Murphy wrote:
>>>>>>>>> On Wed, Aug 17, 2022, at 11:41 PM, Ming Lei wrote:
>>>>>>>>>
>>>>>>>>>> OK, can you post the blk-mq debugfs log after you trigger it on v5.17?
>>>>>>>
>>>>>>> Same boot, 3rd log. But the load is above 300 so I kinda need to sysrq+b soon.
>>>>>>>
>>>>>>> https://drive.google.com/file/d/1375H558kqPTdng439rvG6LuXXWPXLToo/view?usp=sharing
>>>>>>>
>>>>>>
>>>>>> Also please test the following one too:
>>>>>>
>>>>>>
>>>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>>>>> index 5ee62b95f3e5..d01c64be08e2 100644
>>>>>> --- a/block/blk-mq.c
>>>>>> +++ b/block/blk-mq.c
>>>>>> @@ -1991,7 +1991,8 @@ bool blk_mq_dispatch_rq_list(struct blk_mq_hw_ctx
>>>>>> *hctx, struct list_head *list,
>>>>>> if (!needs_restart ||
>>>>>> (no_tag && list_empty_careful(&hctx->dispatch_wait.entry)))
>>>>>> blk_mq_run_hw_queue(hctx, true);
>>>>>> - else if (needs_restart && needs_resource)
>>>>>> + else if (needs_restart && (needs_resource ||
>>>>>> + blk_mq_is_shared_tags(hctx->flags)))
>>>>>> blk_mq_delay_run_hw_queue(hctx, BLK_MQ_RESOURCE_DELAY);
>>>>>>
>>>>>> blk_mq_update_dispatch_busy(hctx, true);
>>>>>>
>>>>>
>>>>>
>>>>> With just this patch on top of 5.17.0, it still hangs. I've captured block debugfs log:
>>>>> https://drive.google.com/file/d/1ic4YHxoL9RrCdy_5FNdGfh_q_J3d_Ft0/view?usp=sharing
>>>>
>>>> The log is similar with before, and the only difference is RESTART not
>>>> set.
>>>>
>>>> Also follows another patch merged to v5.18 and it fixes io stall too, feel free to test it:
>>>>
>>>> 8f5fea65b06d blk-mq: avoid extending delays of active hctx from blk_mq_delay_run_hw_queues
>>>
>>> Have you tried this patch?
>>>
>>> We meet a similar problem in our test, and I'm pretty sure about the
>>> situation at the scene,
>>>
>>> Our test environment:nvme with bfq ioscheduler,
>>>
>>> How io is stalled:
>>>
>>> 1. hctx1 dispatch rq from bfq in service queue, bfqq becomes empty,
>>> dispatch somehow fails and rq is inserted to hctx1->dispatch, new run
>>> work is queued.
>>>
>>> 2. other hctx tries to dispatch rq, however, in service bfqq is
>>> empty, bfq_dispatch_request return NULL, thus
>>> blk_mq_delay_run_hw_queues is called.
>>>
>>> 3. for the problem described in above patch,run work from "hctx1"
>>> can be stalled.
>>>
>>> Above patch should fix this io stall, however, it seems to me bfq do
>>> have some problems that in service bfqq doesn't expire under following
>>> situation:
>>>
>>> 1. dispatched rqs don't complete
>>> 2. no new rq is issued to bfq
>> And I guess:
>> 3. there are requests queued in other bfqqs
>> ?
>
> Yes, of course, other bfqqs still have requests, but current
> implementation have flaws that even if other bfqqs doesn't have
> requests, bfq_asymmetric_scenario() can still return true because
> num_groups_with_pending_reqs > 0. We tried to fix this, however, there
> seems to be some misunderstanding with Paolo, and it's not applied to
> mainline yet...
>

I think this is an unsolved performance issue (being solved patiently
by Yu Kuai), but not a functional flaw. The solution of this issue
would probably solve this stall, but not the essential problem:
refcounting gets broken if reqs disappear for bfq without any
notification.

Thanks,
Paolo

> Thanks,
> Kuai
>> Otherwise I don't see a point in expiring current bfqq because there's
>> nothing bfq could do anyway. But under normal circumstances the request
>> completion should not take so long so I don't think it would be really
>> worth it to implement some special mechanism for this in bfq.
>> Honza

\
 
 \ /
  Last update: 2022-09-06 11:51    [W:0.253 / U:0.264 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site