On Wed, 3 Jul 2019 at 08:47, Wanpeng Li <kernellwp@xxxxxxxxx> wrote:
>
> On Wed, 3 Jul 2019 at 06:23, Marcelo Tosatti <mtosatti@xxxxxxxxxx> wrote:
> >
> > On Tue, Jul 02, 2019 at 06:38:56PM +0200, Paolo Bonzini wrote:
> > > On 21/06/19 11:39, Wanpeng Li wrote:
> > > > Dedicated instances are currently disturbed by unnecessary jitter due
> > > > to the emulated lapic timers fire on the same pCPUs which vCPUs resident.
> > > > There is no hardware virtual timer on Intel for guest like ARM. Both
> > > > programming timer in guest and the emulated timer fires incur vmexits.
> > > > This patchset tries to avoid vmexit which is incurred by the emulated
> > > > timer fires in dedicated instance scenario.
> > > >
> > > > When nohz_full is enabled in dedicated instances scenario, the unpinned
> > > > timer will be moved to the nearest busy housekeepers after commit
> > > > 9642d18eee2cd (nohz: Affine unpinned timers to housekeepers) and commit
> > > > 444969223c8 ("sched/nohz: Fix affine unpinned timers mess"). However,
> > > > KVM always makes lapic timer pinned to the pCPU which vCPU residents, the
> > > > reason is explained by commit 61abdbe0 (kvm: x86: make lapic hrtimer
> > > > pinned). Actually, these emulated timers can be offload to the housekeeping
> > > > cpus since APICv is really common in recent years. The guest timer interrupt
> > > > is injected by posted-interrupt which is delivered by housekeeping cpu
> > > > once the emulated timer fires.
> > > >
> > > > The host admin should fine tuned, e.g. dedicated instances scenario w/
> > > > nohz_full cover the pCPUs which vCPUs resident, several pCPUs surplus
> > > > for busy housekeeping, disable mwait/hlt/pause vmexits to keep in non-root
> > > > mode, ~3% redis performance benefit can be observed on Skylake server.
> > >
> > > Marcelo,
> > >
> > > does this patch work for you or can you still see the oops?
> >
> > Hi Paolo,
> >
> > No more oopses with kvm/queue. Can you include:
>
> Cool, thanks for the confirm, Marcelo!
>
> >
> > Index: kvm/arch/x86/kvm/lapic.c
> > ===================================================================
> > --- kvm.orig/arch/x86/kvm/lapic.c
> > +++ kvm/arch/x86/kvm/lapic.c
> > @@ -124,8 +124,7 @@ static inline u32 kvm_x2apic_id(struct k
> >
> > bool posted_interrupt_inject_timer(struct kvm_vcpu *vcpu)
> > {
> > - return pi_inject_timer && kvm_vcpu_apicv_active(vcpu) &&
> > - kvm_hlt_in_guest(vcpu->kvm);
> > + return pi_inject_timer && kvm_vcpu_apicv_active(vcpu);
> > }
> > EXPORT_SYMBOL_GPL(posted_interrupt_inject_timer);
> >
> > However, for some reason (hrtimer subsystems responsability) with cyclictest -i 200
> > on the guest, the timer runs on the local CPU:
> >
> > CPU 1/KVM-9454 [003] d..2 881.674196: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d..2 881.674200: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d.h. 881.674387: apic_timer_fn <-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 881.674393: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d..2 881.674395: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d..2 881.674399: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d.h. 881.674586: apic_timer_fn <-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 881.674593: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d..2 881.674595: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d..2 881.674599: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d.h. 881.674787: apic_timer_fn <-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 881.674793: get_nohz_timer_target: get_nohz_timer_target 3->0
> > CPU 1/KVM-9454 [003] d..2 881.674795: get_nohz_timer_target: get_nohz_timer_target 3->0
> >
> > But on boot:
> >
> > CPU 1/KVM-9454 [003] d..2 578.625394: get_nohz_timer_target: get_nohz_timer_target 3->0
> > <idle>-0 [000] d.h1 578.626390: apic_timer_fn <-__hrtimer_run_queues
> > <idle>-0 [000] d.h1 578.626394: apic_timer_fn<-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 578.626401: get_nohz_timer_target: get_nohz_timer_target 3->0
> > <idle>-0 [000] d.h1 578.628397: apic_timer_fn <-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 578.628407: get_nohz_timer_target: get_nohz_timer_target 3->0
> > <idle>-0 [000] d.h1 578.631403: apic_timer_fn <-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 578.631413: get_nohz_timer_target: get_nohz_timer_target 3->0
> > <idle>-0 [000] d.h1 578.635409: apic_timer_fn <-__hrtimer_run_queues
> > CPU 1/KVM-9454 [003] d..2 578.635419: get_nohz_timer_target: get_nohz_timer_target 3->0
> > <idle>-0 [000] d.h1 578.640415: apic_timer_fn <-__hrtimer_run_queues
>
> You have an idle housekeeping cpu(cpu 0), however, most of
> housekeeping cpus will be busy in product environment to avoid to
> waste money. get_nohz_timer_target() will find a busy housekeeping cpu
> but the timer migration will fail if the timer is the first expiring
> timer on the new target(as the comments above the function
> switch_hrtimer_base()). Please try taskset -c 0 stress --cpu 1 on your
> host, you can observe(through /proc/timer_list) apic_timer_fn running
> on cpu 0 most of the time and sporadically on local cpu.
Or if you have a little bigger VM/multiple VMs, the apic_timer_fn from
all virtual lapics will make a housekeeping cpu busy. :)
Regards,
Wanpeng Li
