Re: [PATCH BUGFIX V3] block, bfq: add requeue-request hook

[Paolo Valente]

> Il giorno 24 feb 2018, alle ore 13:05, Ming Lei <ming.lei@xxxxxxxxxx> ha scritto:
> 
> On Sat, Feb 24, 2018 at 08:54:31AM +0100, Paolo Valente wrote:
>> 
>> 
>>> Il giorno 23 feb 2018, alle ore 17:17, Ming Lei <ming.lei@xxxxxxxxxx> ha scritto:
>>> 
>>> Hi Paolo,
>>> 
>>> On Fri, Feb 23, 2018 at 04:41:36PM +0100, Paolo Valente wrote:
>>>> 
>>>> 
>>>>> Il giorno 23 feb 2018, alle ore 16:07, Ming Lei <ming.lei@xxxxxxxxxx> ha scritto:
>>>>> 
>>>>> Hi Paolo,
>>>>> 
>>>>> On Wed, Feb 07, 2018 at 10:19:20PM +0100, Paolo Valente wrote:
>>>>>> Commit 'a6a252e64914 ("blk-mq-sched: decide how to handle flush rq via
>>>>>> RQF_FLUSH_SEQ")' makes all non-flush re-prepared requests for a device
>>>>>> be re-inserted into the active I/O scheduler for that device. As a
>>>>> 
>>>>> No, this behaviour isn't related with commit a6a252e64914, and
>>>>> it has been there since blk_mq_requeue_request() is introduced.
>>>>> 
>>>> 
>>>> Hi Ming,
>>>> actually, we wrote the above statement after simply following the call
>>>> chain that led to the failure.  In this respect, the change in commit
>>>> a6a252e64914:
>>>> 
>>>> static bool blk_mq_sched_bypass_insert(struct blk_mq_hw_ctx *hctx,
>>>> +                                      bool has_sched,
>>>>                                      struct request *rq)
>>>> {
>>>> -       if (rq->tag == -1) {
>>>> +       /* dispatch flush rq directly */
>>>> +       if (rq->rq_flags & RQF_FLUSH_SEQ) {
>>>> +               spin_lock(&hctx->lock);
>>>> +               list_add(&rq->queuelist, &hctx->dispatch);
>>>> +               spin_unlock(&hctx->lock);
>>>> +               return true;
>>>> +       }
>>>> +
>>>> +       if (has_sched) {
>>>>               rq->rq_flags |= RQF_SORTED;
>>>> -               return false;
>>>> +               WARN_ON(rq->tag != -1);
>>>>       }
>>>> 
>>>> -       /*
>>>> -        * If we already have a real request tag, send directly to
>>>> -        * the dispatch list.
>>>> -        */
>>>> -       spin_lock(&hctx->lock);
>>>> -       list_add(&rq->queuelist, &hctx->dispatch);
>>>> -       spin_unlock(&hctx->lock);
>>>> -       return true;
>>>> +       return false;
>>>> }
>>>> 
>>>> makes blk_mq_sched_bypass_insert return false for all non-flush
>>>> requests.  From that, the anomaly described in our commit follows, for
>>>> bfq any stateful scheduler that waits for the completion of requests
>>>> that passed through it.  I'm elaborating again a little bit on this in
>>>> my replies to your next points below.
>>> 
>>> Before a6a252e64914, follows blk_mq_sched_bypass_insert()
>>> 
>>>      if (rq->tag == -1) {
>>>              rq->rq_flags |= RQF_SORTED;
>>>              return false;
>>> 	   }
>>> 
>>>      /*
>>>       * If we already have a real request tag, send directly to
>>>       * the dispatch list.
>>>       */
>>>      spin_lock(&hctx->lock);
>>>      list_add(&rq->queuelist, &hctx->dispatch);
>>>      spin_unlock(&hctx->lock);
>>>      return true;
>>> 
>>> This function still returns false for all non-flush request, so nothing
>>> changes wrt. this kind of handling.
>>> 
>> 
>> Yep Ming.  I don't have the expertise to tell you why, but the failure
>> in the USB case was caused by an rq that is not flush, but for which
>> rq->tag != -1.  So, the previous version of blk_mq_sched_bypass_insert
>> returned true, and there was not failure, while after commit
>> a6a252e64914 the function returns true and the failure occurs if bfq
>> does not exploit the requeue hook.
>> 
>> You have actually shown it yourself, several months ago, through the
>> simple code instrumentation you made and used to show that bfq was
>> stuck.  And I guess it can still be reproduced very easily, unless
>> something else has changed in blk-mq.
>> 
>> BTW, if you can shed a light on this fact, that would be a great
>> occasion to learn for me.
> 
> The difference should be made by commit 923218f6166a84 (blk-mq: don't
> allocate driver tag upfront for flush rq), which releases driver tag
> before requeuing the request, but that is definitely the correct thing
> to do.
> 

Ok.  We just did a bisection, guided by changes in the
function blk_mq_sched_bypass_insert, as that is the function whose
different return value led to the hang.  And commit a6a252e64914
apparently was the latest commit after which the hang occurs.

Anyway, what matters is just that, from some commit on, a requeued
request that before that commit was not re-inserted into the scheduler
started to be re-inserted.  And this was the trigger of the failure.

>> 
>>>> 
>>>> I don't mean that this change is an error, it simply sends a stateful
>>>> scheduler in an inconsistent state, unless the scheduler properly
>>>> handles the requeue that precedes the re-insertion into the
>>>> scheduler.
>>>> 
>>>> If this clarifies the situation, but there is still some misleading
>>>> statement in the commit, just let me better understand, and I'll be
>>>> glad to rectify it, if possible somehow.
>>>> 
>>>> 
>>>>> And you can see blk_mq_requeue_request() is called by lots of drivers,
>>>>> especially it is often used in error handler, see SCSI's example.
>>>>> 
>>>>>> consequence, I/O schedulers may get the same request inserted again,
>>>>>> even several times, without a finish_request invoked on that request
>>>>>> before each re-insertion.
>>>>>> 
>>>> 
>>>> ...
>>>> 
>>>>>> @@ -5426,7 +5482,8 @@ static struct elevator_type iosched_bfq_mq = {
>>>>>> 	.ops.mq = {
>>>>>> 		.limit_depth		= bfq_limit_depth,
>>>>>> 		.prepare_request	= bfq_prepare_request,
>>>>>> -		.finish_request		= bfq_finish_request,
>>>>>> +		.requeue_request        = bfq_finish_requeue_request,
>>>>>> +		.finish_request		= bfq_finish_requeue_request,
>>>>>> 		.exit_icq		= bfq_exit_icq,
>>>>>> 		.insert_requests	= bfq_insert_requests,
>>>>>> 		.dispatch_request	= bfq_dispatch_request,
>>>>> 
>>>>> This way may not be correct since blk_mq_sched_requeue_request() can be
>>>>> called for one request which won't enter io scheduler.
>>>>> 
>>>> 
>>>> Exactly, there are two cases: requeues that lead to subsequent
>>>> re-insertions, and requeues that don't.  The function
>>>> bfq_finish_requeue_request handles both, and both need to be handled,
>>>> to inform bfq that it has not to wait for the completion of rq any
>>>> longer.
>>>> 
>>>> One special case is when bfq_finish_requeue_request gets invoked even
>>>> if rq has nothing to do with any scheduler.  In that case,
>>>> bfq_finish_requeue_request exists immediately.
>>>> 
>>>> 
>>>>> __blk_mq_requeue_request() is called for two cases:
>>>>> 
>>>>> 	- one is that the requeued request is added to hctx->dispatch, such
>>>>> 	as blk_mq_dispatch_rq_list()
>>>> 
>>>> yes
>>>> 
>>>>> 	- another case is that the request is requeued to io scheduler, such as
>>>>> 	blk_mq_requeue_request().
>>>>> 
>>>> 
>>>> yes
>>>> 
>>>>> For the 1st case, blk_mq_sched_requeue_request() shouldn't be called
>>>>> since it is nothing to do with scheduler,
>>>> 
>>>> No, if that rq has been inserted and then extracted from the scheduler
>>>> through a dispatch_request, then it has.  The scheduler is stateful,
>>>> and keeps state for rq, because it must do so, until a completion or a
>>>> requeue arrive.  In particular, bfq may decide that no other
>>> 
>>> This 'requeue' to hctx->dispatch is invisible to io scheduler, and from
>>> io scheduler view, seems no difference between STS_OK and STS_RESOURCE,
>>> since this request will be submitted to lld automatically by blk-mq
>>> core.
>>> 
>>> Also in legacy path, no such notification to io scheduler too.
>> 
>> ok, still, for reasons that at this point I don't know, in the last
>> 9-10 years it has never been reported a failure caused by a request
>> that has been first dispatched by bfq and then re-inserted into bfq
>> without being completed.  This new event is the problem.
> 
> This mechanism is invented by blk-mq, and I guess it is for avoiding to
> disable irq when acquiring hctx->lock since blk-mq's completion handler
> is done in irq context.
> 
> Legacy's requeue won't re-insert one request to scheduler queue.

ok, that's the root cause of this issue then.

> 
>> 
>> 
>>>> bfq_queues must be served before rq has been completed, because this
>>>> is necessary to preserve its target service guarantees.  If bfq is not
>>>> informed, either through its completion or its requeue hook, then it
>>>> will wait forever.
>>> 
>>> The .finish_request will be called after this request is completed for this
>>> case, either after this io is completed by device, or timeout.
>>> 
>>> Or .requeue_request will be called if the driver need to requeue it to
>>> io scheduler via blk_mq_requeue_request() explicitly.
>>> 
>>> So io scheduler will be made sure to get notified.
>>> 
>> 
>> So, you mean that the scheduler will get notified in a way or the
>> other, if it has a requeue hook defined, right?  Or, simply, the
>> request will be eventually completed, thus unblocking the scheduler?
> 
> Right, once .queue_rq() returns BLK_STS_RESOURCE, this rq is added to
> hctx->dispatch list, and finally it will be completed via driver or
> timeout handler, or it may be re-inserted to io scheduler queue via
> blk_mq_requeue_request() by driver in driver's completion handler.
> 
> So io scheduler will get the notification finally.

Yes, provided that the scheduler has a requeue hook, otherwise is one
case it is not notified before the re-isnertion.  And from this
troubles follow with a stateful scheduler like bfq.

> 
>> 
>> 
>>>> 
>>>>> seems we only need to do that
>>>>> for 2nd case.
>>>>> 
>>>> 
>>>>> So looks we need the following patch:
>>>>> 
>>>>> diff --git a/block/blk-mq.c b/block/blk-mq.c
>>>>> index 23de7fd8099a..a216f3c3c3ce 100644
>>>>> --- a/block/blk-mq.c
>>>>> +++ b/block/blk-mq.c
>>>>> @@ -712,7 +714,6 @@ static void __blk_mq_requeue_request(struct request *rq)
>>>>> 
>>>>> 	trace_block_rq_requeue(q, rq);
>>>>> 	wbt_requeue(q->rq_wb, &rq->issue_stat);
>>>>> -	blk_mq_sched_requeue_request(rq);
>>>>> 
>>>> 
>>>> By doing so, if rq has 'passed' through bfq, then you block again bfq
>>>> forever.
>>> 
>>> Could you explain it a bit why bfq is blocked by this way?
>> 
>> It's a theoretical problem (unfortunately not just a contingency you
>> can work around).  If you want to guarantee that a process doing sync
>> I/O, and with a higher weight than other processes, gets a share of
>> the total throughput proportional to its weight, then you have to idle
>> the device during the service slot of that process.  Otherwise, the
>> pending requests of the other processes will simply slip through
>> *every time* the high-weight process has no pending request because it
>> is blocked by one of its sync request.  The final result will be loss
>> of control on the percentage of requests of the high-weight process
>> that get dispatched w.r.t. to the total number of requests
>> dispatched.  So, failure to provide the process with its due share of
>> the throughput.
>> 
>> I'm working on solutions to compensate for this nasty service
>> constraint, and its effects on throughput, but this is another story.
> 
> As I explained, io scheduler will get notified via either
> .finish_request or .requeue_request finally, so there shouldn't be
> such block if this patch is in.
> 

Well, it depends on which patch is your "this patch" :)

I mean, if you refer to your new patch, then I'm afraid the
answer is "no, exactly because of what you explained".  In fact, in
case of a re-insertion, the scheduler gets notified through its
requeue hook.  But for this mechanism to work, the scheduler must have
a its requeue hook defined.  bfq had no such a hook defined, and our
patch defines exactly it.  So, I don'n see how the re-insertion case
might be handled correctly without our patch.  Maybe I'm just missing
something here.


>> 
>>> 
>>>> 
>>>> Of course, I may be wrong, as I don't have a complete mental model of
>>>> all what blk-mq does around bfq.  But I thin that you can quickly
>>>> check whether a hang occurs again if you add this change.  In
>>>> particular, it should happen in the USB former failure case.
>>> 
>>> USB/BFQ runs well on my parted test with this change, and this test can
>>> trigger the IO hang issue easily on kyber or without your patch of 'block,
>>> bfq: add requeue-request hook'.
>>> 
>> 
>> Great!  According to the facts I reported at the beginning of
>> this email, I guess your patch works because it avoids the
>> nasty case for bfq, i.e., a request re-inserted into bfq without being
>> completed first, right? 
> 
> BFQ still can get notified via .requeue_request when one request is re-inserted
> to BFQ with this patch.
> 
>> And then, about the requeues without
>> re-insertions, everything works because the latter type of requests
>> eventually get a completion.  Is it correct?
> 
> Right.
> 
>> 
>> Sorry for being pedantic, but because of our lack of expertise on
>> these requeueing mechanism, we worked really a lot on this commit, and
>> it would be frustrating if we bumped again into troubles, after
>> removing it.  Apart from that, if this commit is useless, it is
>> perfectly fine for me that it is reverted and replaced by a better
>> solution.  In the end, less code in bfq means less room for bugs ;)
> 
> I agree. We should deal with this issue in a simple/clean way, just like
> the patch of 'block: kyber: fix domain token leak during requeue'.
> 

Our patch "block, bfq: add requeue-request hook" is exactly the bfq
counterpart of the patch 'block: kyber: fix domain token leak during requeue'

> Now could you please let us know if you are fine with the patch of 'blk-mq:
> don't call io sched's .requeue_request when requeueing rq to ->dispatch'?
> 

Absolutely.  Just, in view of what you explained, and of what I wrote
in this reply, such a patch does not seem to make our patch useless.

Thanks,
Paolo


> 
> Thanks,
> Ming