Re: [PATCH] xen: Always ask the scheduler to re-place the vcpu when the affinity changes

[George Dunlap <george.dunlap@eu.citrix.com>]

On 04/03/13 13:45, George Dunlap wrote:
> On 04/03/13 12:35, Jan Beulich wrote:
>>>>> On 04.03.13 at 13:19, George Dunlap <george.dunlap@eu.citrix.com> 
>>>>> wrote:
>>> It's probably a good idea to re-evaluate placement whenever the
>>> affinity changes.
>>>
>>> This additionally has the benefit of removing scheduler-specific
>>> exceptions introduced in git c/s e6a6fd63.
>>>
>>> Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
>>> ---
>>>   xen/common/schedule.c |    7 ++++---
>>>   1 file changed, 4 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/xen/common/schedule.c b/xen/common/schedule.c
>>> index de11110..dbef5af 100644
>>> --- a/xen/common/schedule.c
>>> +++ b/xen/common/schedule.c
>>> @@ -613,9 +613,10 @@ int vcpu_set_affinity(struct vcpu *v, const 
>>> cpumask_t
>>> *affinity)
>>>       vcpu_schedule_lock_irq(v);
>>>         cpumask_copy(v->cpu_affinity, affinity);
>>> -    if ( VCPU2OP(v)->sched_id == XEN_SCHEDULER_SEDF ||
>>> -         !cpumask_test_cpu(v->processor, v->cpu_affinity) )
>>> -        set_bit(_VPF_migrating, &v->pause_flags);
>>> +
>>> +    /* Always ask the scheduler to re-evaluate placement
>>> +     * when changing the affinity */
>>> +    set_bit(_VPF_migrating, &v->pause_flags);
>>>         vcpu_schedule_unlock_irq(v);
>> The code right below the context visible here is
>>
>>      if ( test_bit(_VPF_migrating, &v->pause_flags) )
>>      {
>>          vcpu_sleep_nosync(v);
>>          vcpu_migrate(v);
>>      }
>>
>> and I think the conditional could (and should) now be removed.
>
> Yeah, I wasn't sure what to make of that one -- it looked as though it 
> was coded such that someone else might be able to clear _VPF_migrating 
> between releasing the lock and this test.  But if that can happen, 
> it's racy anyway.  vcpu_force_reschedule() has this "set, unlock, 
> re-check" pattern.
>
> It looks like there actually is a way that it could conceivably be 
> cleared: if the vcpu is running on another pcpu, if after we release 
> the lock the vcpu is de-scheduled and context_saved() is called, it 
> will check for _VPF_migrating and call vcpu_migrate(), which can clear 
> the flag.
>
> But that's probably a rare enough occurrence that it's better overall 
> to take the occasional double-migrate.

Hmm -- but thinking it further, it actually seems likely that a 
different double-migrate race will happen:

1. vcpu is running on pcpu A
2. pcpu B runs set_affinity, setting VPF_migrate
3. pcpu B calls vcpu_sleep_nosync
4. pcpu A wakes up and grabs the schedule lock
5. pcpu A notices that VPF_migrate is set, and calls vcpu_migrate()
6. pcpu B calls vcpu_migrate()

Either that, or 6 happens before 4, but 4 still happens before pcpu B 
clears VPF_migrate.

It seems like we should really only call if (!v->is_running || 
v->processor == this_cpu).

Dario, any thoughts?

  -George


>
>  -George