On Sat, Jun 25, 2016 at 04:24:47PM +0200, Peter Zijlstra wrote: > On Sat, Jun 25, 2016 at 01:42:03PM -0400, Pan Xinhui wrote: > > An over-committed guest with more vCPUs than pCPUs has a heavy overload > > in osq_lock(). > > > > This is because vCPU A hold the osq lock and yield out, vCPU B wait > > per_cpu node->locked to be set. IOW, vCPU B wait vCPU A to run and > > unlock the osq lock. Even there is need_resched(), it did not help on > > such scenario. > > > > To fix such bad issue, add a threshold in one while-loop of osq_lock(). > > The value of threshold is somehow equal to SPIN_THRESHOLD. > > Blergh, virt ... > > So yes, lock holder preemption sucks. You would also want to limit the > immediate spin on owner. > > Also; I really hate these random number spin-loop thresholds. > > Is it at all possible to get feedback from your LPAR stuff that the vcpu > was preempted? Because at that point we can add do something like: >
Good point!
>
> int vpc = vcpu_preempt_count();
>
> ...
>
> for (;;) {
>
> /* the big spin loop */
>
> if (need_resched() || vpc != vcpu_preempt_count())
So on PPC, we have lppaca::yield_count to detect when an vcpu is
preempted, if the yield_count is even, the vcpu is running, otherwise it
is preempted(__spin_yield() is a user of this).
Therefore it makes more sense we
if (need_resched() || vcpu_is_preempted(old))
here, and implement vcpu_is_preempted() on PPC as
bool vcpu_is_preempted(int cpu)
{
return !!(be32_to_cpu(lppaca_of(cpu).yield_count) & 1)
}
Thoughts?
Regards,
Boqun
> /* bail */
>
> }
>
>
> With a default implementation like:
>
> static inline int vcpu_preempt_count(void)
> {
> return 0;
> }
>
> So the compiler can make it all go away.
>
>
> But on virt muck it would stop spinning the moment the vcpu gets
> preempted, which is the right moment I'm thinking.
signature.asc
Description: PGP signature
