On Wed, Oct 18, 2017 at 3:55 PM, Andrew Jones <drjo...@redhat.com> wrote: > On Wed, Oct 18, 2017 at 03:18:43PM +0200, Christoffer Dall wrote: >> On Wed, Oct 18, 2017 at 02:13:05PM +0200, Andrew Jones wrote: >> > On Sat, Oct 14, 2017 at 09:13:35PM +0200, Christoffer Dall wrote: >> > > On Fri, Sep 29, 2017 at 01:30:41PM +0200, Andrew Jones wrote: >> > > > When the vPMU is in use if a VCPU's perf event overflow handler >> > > > were to fire after the VCPU started waiting, then the wake up >> > > > done by the kvm_vcpu_kick() call in the handler would do nothing, >> > > > as no "pmu overflow" state is checked in kvm_arch_vcpu_runnable(). >> > > > Fix this by checking the IRQ_PENDING VCPU request in runnable(). >> > > > Checking the request also sufficiently covers all the cases that >> > > > kvm_vgic_vcpu_pending_irq() cover, so we can just replace that. >> > > > >> > > > Signed-off-by: Andrew Jones <drjo...@redhat.com> >> > > > --- >> > > > virt/kvm/arm/arm.c | 2 +- >> > > > 1 file changed, 1 insertion(+), 1 deletion(-) >> > > > >> > > > diff --git a/virt/kvm/arm/arm.c b/virt/kvm/arm/arm.c >> > > > index 5bc9b0d2fd0f..725527f491e4 100644 >> > > > --- a/virt/kvm/arm/arm.c >> > > > +++ b/virt/kvm/arm/arm.c >> > > > @@ -423,7 +423,7 @@ int kvm_arch_vcpu_runnable(struct kvm_vcpu *vcpu) >> > > > return !vcpu_should_sleep(vcpu) && >> > > > (vcpu->arch.mp_state != KVM_MP_STATE_HALTED || >> > > > (!!vcpu->arch.irq_lines || >> > > > - kvm_vgic_vcpu_pending_irq(vcpu))); >> > > > + kvm_test_request(KVM_REQ_IRQ_PENDING, vcpu))); >> > > >> > > So what if a VCPU blocks, a device raises an IRQ, the VCPU loops around, >> > > clears the KVM_REQ_IRQ_PENDING flag, enters the VM again, which does >> > > another WFI (for fun), and you end up here again with a pending IRQ but >> > > no KVM_REQ_IRQ_PENDING flag anymore. Doesn't this end up incorrectly >> > > stalling the VCPU? >> > >> > Hmm, I see what you mean. I'm sorry I missed that. >> > >> > > >> > > I don't think that a transient flag will work for a persistent binary >> > > state here. >> > >> > I think we can fix it by adding >> > >> > if (kvm_vgic_vcpu_pending_irq(vcpu)) >> > kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); >> > >> > to kvm_vgic_sync_hwstate(), if the additional overhead would be >> > acceptable. Otherwise we need to find some other way to ensure >> > vPMU irqs unblock the VCPU. >> > >> >> I think you just need to check >> kvm_vgic_vcpu_pending_irq() || >> kvm_test_request(KVM_REQ_IRQ_PENDING, vcpu) >> >> Wouldn't that fix it? > > It would. I was just hoping to eliminate the lock-taking > kvm_vgic_vcpu_pending_irq() call from kvm_arch_vcpu_runnable() > though, as kvm_arch_vcpu_runnable() may be called repeatedly > during halt polling. >
Could we make it so that the KVM_REQ_IRQ_PENDING request only gets cleared when there are really no more pending IRQs for the VCPU then? Getting rid of the lock would be potentially good from a latency point of view, but then again, it's slow to wake up from WFI, and the lock is per-VCPU so not likely to be contended is it? Thanks, -Christoffer _______________________________________________ kvmarm mailing list kvmarm@lists.cs.columbia.edu https://lists.cs.columbia.edu/mailman/listinfo/kvmarm