lkml.org 
[lkml]   [2021]   [Jul]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [tip: sched/core] sched/core: Initialize the idle task with preemption disabled
From


On 08/07/2021 00:14, Guenter Roeck wrote:
> On 7/7/21 5:11 AM, Valentin Schneider wrote:
>> On 07/07/21 14:03, Frederic Weisbecker wrote:
>>> On Wed, Jul 07, 2021 at 12:55:20AM +0100, Valentin Schneider wrote:
>>>> Thanks for the report.
>>>>
>>>> So somehow the init task ends up with a non-zero preempt_count()? Per
>>>> FORK_PREEMPT_COUNT we should exit __ret_from_fork() with a zero
>>>> count, are
>>>> you hitting the WARN_ONCE() in finish_task_switch()?
>>>>
>>>> Does CONFIG_DEBUG_PREEMPT=y yield anything interesting?
>>>>
>>>> I can't make sense of this right now, but it's a bit late :) I'll
>>>> grab some
>>>> toolchain+qemu tomorrow and go poke at it (and while at it I need to
>>>> do the
>>>> same with powerpc).
>>>
>>> One possible issue is that s390's init_idle_preempt_count() doesn't
>>> apply on the
>>> target idle task but on the _current_ CPU. And since smp_init() ->
>>> idle_threads_init() is actually called remotely, we are overwriting
>>> the current
>>> CPU preempt_count() instead of the target one.
>>
>> Indeed, this becomes quite obvious when tracing the preemption count
>> changes. This also means that s390 relied on the idle_thread_get()
>> from the
>> hotplug machinery to properly setup the preempt count, rather than
>> init_idle_preempt_count() - which is quite yuck.
>>
>> I'll write a patch for that and likely one for powerpc.
>>
>
> Can you reproduce the problem with a powerpc qemu emulation ?
> If so, how do you reproduce it there ? Reason for asking is that I don't
> see
> the problem with any of my powerpc emulations, and I would like to add test
> case(s) if possible.

I can reproduce the problem on powerpc easily - qemu with "-smp 2" does it.


--
Alexey

\
 
 \ /
  Last update: 2021-07-07 16:36    [W:0.102 / U:0.316 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site