> On Jul 24, 2023, at 12:00 PM, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>
> On Mon, Jul 24, 2023 at 09:36:02AM -0400, Joel Fernandes wrote:
>>> On Sun, Jul 23, 2023 at 11:35 PM Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
>>>
>>> On Mon, Jul 24, 2023 at 12:32:57AM +0000, Joel Fernandes wrote:
>>>> On Sun, Jul 23, 2023 at 10:19:27AM -0700, Paul E. McKenney wrote:
>>>>> On Sun, Jul 23, 2023 at 10:50:26AM -0400, Joel Fernandes wrote:
>>>>>>
>>>>>>
>>>>>> On 7/22/23 13:27, Paul E. McKenney wrote:
>>>>>> [..]
>>>>>>>
>>>>>>> OK, if this kernel is non-preemptible, you are not running TREE03,
>>>>>>> correct?
>>>>>>>
>>>>>>>> Next plan of action is to get sched_waking stack traces since I have a
>>>>>>>> very reliable repro of this now.
>>>>>>>
>>>>>>> Too much fun! ;-)
>>>>>>
>>>>>> For TREE07 issue, it is actually the schedule_timeout_interruptible(1)
>>>>>> in stutter_wait() that is beating up the CPU0 for 4 seconds.
>>>>>>
>>>>>> This is very similar to the issue I fixed in New year in d52d3a2bf408
>>>>>> ("torture: Fix hang during kthread shutdown phase")
>>>>>
>>>>> Agreed, if there are enough kthreads, and all the kthreads are on a
>>>>> single CPU, this could consume that CPU.
>>>>>
>>>>>> Adding a cond_resched() there also did not help.
>>>>>>
>>>>>> I think the issue is the stutter thread fails to move spt forward
>>>>>> because it does not get CPU time. But spt == 1 should be very brief
>>>>>> AFAIU. I was wondering if we could set that to RT.
>>>>>
>>>>> Or just use a single hrtimer-based wait for each kthread?
>>>>
>>>> [Joel]
>>>> Yes this might be better, but there's still the issue that spt may not be set
>>>> back to 0 in some future release where the thread gets starved.
>>>
>>> But if each thread knows the absolute time at which the current stutter
>>> period is supposed to end, there should not be any need for the spt
>>> variable, correct?
>>
>> Yes.
>>
>>>>>> But also maybe the following will cure it like it did for the shutdown
>>>>>> issue, giving the stutter thread just enough CPU time to move spt forward.
>>>>>>
>>>>>> Now I am trying the following and will let it run while I go do other
>>>>>> family related things. ;)
>>>>>
>>>>> Good point, if this avoids the problem, that gives a strong indication
>>>>> that your hypothesis on the root cause is correct.
>>>>
>>>> [Joel]
>>>> And the TREE07 issue is gone with that change!
>> [...]
>>>> Let me know what you think, thanks!
>>>
>>> If we can make the stutter kthread set an absolute time for the current
>>> stutter period to end, then we should be able to simplify the code quite
>>> a bit and get rid of the CPU consumption entirely. (Give or take the
>>> possible need for a given thread to check whether it was erroneously
>>> awakened early.)
>>>
>>> But what specifically did you have in mind?
>>
>> I was thinking of a 2 counter approach storing the absolute time. Use
>> an alternative counter for different stuttering sessions. But yes,
>> generally I agree with the absolute time idea. What do you think Paul?
>>
>> Do we want to just do the simpler schedule_timeout at HZ / 20 to keep stable
>> green, and do the absolute-time approach for mainline? That might be better
>> from a process PoV. But I think stable requires patches to be upstream. Greg?
>>
>> I will try to send out patches this week to discuss this, thanks,
>
> Heh!!!
>
> Me, I was just thinking of mainline. ;-)
Turns out it is simple enough for both mainline and stable :-).
Will test more and send it out soon.
Thanks,
- Joel
>
> Thanx, Paul
