Re: [Xen-devel] [PATCH 1/8] cpupools: fix state when downing a CPU failed

[Juergen Gross <jgross@xxxxxxxx>]

On 16/07/18 15:01, Jan Beulich wrote:
>>>> On 16.07.18 at 14:47, <jgross@xxxxxxxx> wrote:
>> On 16/07/18 14:19, Jan Beulich wrote:
>>>>>> On 16.07.18 at 13:47, <jgross@xxxxxxxx> wrote:
>>>> On 16/07/18 11:17, Jan Beulich wrote:
>>>>>>>> On 13.07.18 at 11:02, <jgross@xxxxxxxx> wrote:
>>>>>> On 11/07/18 14:04, Jan Beulich wrote:
>>>>>>> While I've run into the issue with further patches in place which no
>>>>>>> longer guarantee the per-CPU area to start out as all zeros, the
>>>>>>> CPU_DOWN_FAILED processing looks to have the same issue: By not zapping
>>>>>>> the per-CPU cpupool pointer, cpupool_cpu_add()'s (indirect) invocation
>>>>>>> of schedule_cpu_switch() will trigger the "c != old_pool" assertion
>>>>>>> there.
>>>>>>>
>>>>>>> Clearing the field during CPU_DOWN_PREPARE is too early (afaict this
>>>>>>> should not happen before cpu_disable_scheduler()). Clearing it in
>>>>>>> CPU_DEAD and CPU_DOWN_FAILED would be an option, but would take the same
>>>>>>> piece of code twice. Since the field's value shouldn't matter while the
>>>>>>> CPU is offline, simply clear it in CPU_ONLINE and CPU_DOWN_FAILED, but
>>>>>>> only for other than the suspend/resume case (which gets specially
>>>>>>> handled in cpupool_cpu_remove()).
>>>>>>>
>>>>>>> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
>>>>>>> ---
>>>>>>> TBD: I think this would better call schedule_cpu_switch(cpu, NULL) from
>>>>>>>      cpupool_cpu_remove(), but besides that - as per above - likely
>>>>>>>      being too early, that function has further prereqs to be met. It
>>>>>>>      also doesn't look as if cpupool_unassign_cpu_helper() could be used
>>>>>>>      there.
>>>>>>>
>>>>>>> --- a/xen/common/cpupool.c
>>>>>>> +++ b/xen/common/cpupool.c
>>>>>>> @@ -778,6 +778,8 @@ static int cpu_callback(
>>>>>>>      {
>>>>>>>      case CPU_DOWN_FAILED:
>>>>>>>      case CPU_ONLINE:
>>>>>>> +        if ( system_state <= SYS_STATE_active )
>>>>>>> +            per_cpu(cpupool, cpu) = NULL;
>>>>>>>          rc = cpupool_cpu_add(cpu);
>>>>>>
>>>>>> Wouldn't it make more sense to clear the field in cpupool_cpu_add()
>>>>>> which already is testing system_state?
>>>>>
>>>>> Hmm, this may be a matter of taste: I consider the change done here
>>>>> a prereq to calling the function in the first place. As said in the
>>>>> description, I actually think this should come earlier, and it's just that
>>>>> I can't see how to cleanly do so.
>>>
>>> You didn't comment on this one at all, yet it matters for how a v2
>>> is supposed to look like.
>>
>> My comment was thought to address this question, too. cpupool_cpu_add()
>> is handling the special case of resuming explicitly, where the old cpu
>> assignment to a cpupool is kept. So I believe setting
>>   per_cpu(cpupool, cpu) = NULL
>> in the else clause of cpupool_cpu_add() only is better.
> 
> Well, okay then. You're the maintainer.
> 
>>>>>> Modifying the condition in cpupool_cpu_add() to
>>>>>>
>>>>>>   if ( system_state <= SYS_STATE_active )
>>>>>>
>>>>>> at the same time would have the benefit to catch problems in case
>>>>>> suspending cpus is failing during SYS_STATE_suspend (I'd expect
>>>>>> triggering the first ASSERT in schedule_cpu_switch() in this case).
>>>>>
>>>>> You mean the if() there, not the else? If so - how would the "else"
>>>>> body then ever be reached? IOW if anything I could only see the
>>>>> "else" to become "else if ( system_state <= SYS_STATE_active )".
>>>>
>>>> Bad wording on my side.
>>>>
>>>> I should have written "the condition in cpupool_cpu_add() should match
>>>> if ( system_state <= SYS_STATE_active )."
>>>>
>>>> So: "if ( system_state > SYS_STATE_active )", as the test is for the
>>>> other case.
>>>
>>> I'd recommend against this, as someone adding a new SYS_STATE_*
>>> past suspend/resume would quite likely miss this one. The strong
>>> ordering of states imo should only be used for active and lower states.
>>> But yes, I could see the if() there to become suspend || resume to
>>> address the problem you describe.
>>
>> Yes, this would seem to be a better choice here.
>>
>>> Coming back to your DOWN_FAILED consideration: Why are you
>>> thinking this can't happen during suspend? disable_nonboot_cpus()
>>> uses plain cpu_down() after all.
>>
>> Right.
>>
>> DOWN_FAILED is used only once, and that is in cpu_down() after the step
>> CPU_DOWN_PREPARE returned an error. And CPU_DOWN_PREPARE is only used
>> for cpufreq driver where it never returns an error, and for cpupools
>> which don't matter here, as only other components failing at step
>> CPU_DOWN_PREPARE would lead to calling cpupool/DOWN_FAILED.
> 
> What about the stop_machine_run() failure case?

Oh. No idea how I missed that.

So maybe changing the condition in cpupool_cpu_add() should be split out
into a patch of its own in order to be able to backport it?


Juergen


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel