From mboxrd@z Thu Jan 1 00:00:00 1970 From: marc.zyngier@arm.com (Marc Zyngier) Date: Tue, 23 Feb 2016 14:14:54 +0000 Subject: [PATCH v12 16/21] KVM: ARM64: Add PMU overflow interrupt routing In-Reply-To: <1456133877-9584-17-git-send-email-zhaoshenglong@huawei.com> References: <1456133877-9584-1-git-send-email-zhaoshenglong@huawei.com> <1456133877-9584-17-git-send-email-zhaoshenglong@huawei.com> Message-ID: <56CC695E.4010805@arm.com> To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On 22/02/16 09:37, Shannon Zhao wrote: > From: Shannon Zhao > > When calling perf_event_create_kernel_counter to create perf_event, > assign a overflow handler. Then when the perf event overflows, set the > corresponding bit of guest PMOVSSET register. If this counter is enabled > and its interrupt is enabled as well, kick the vcpu to sync the > interrupt. > > On VM entry, if there is counter overflowed, inject the interrupt with > the level set to 1. Otherwise, inject the interrupt with the level set > to 0. > > Signed-off-by: Shannon Zhao > Reviewed-by: Marc Zyngier > Reviewed-by: Andrew Jones As I mentioned yesterday, I was trying to pinpoint a performance drop, so I added PMU support to kvmtool (and made it an optional flag). This allowed be to find this: > --- > arch/arm/kvm/arm.c | 2 ++ > include/kvm/arm_pmu.h | 2 ++ > virt/kvm/arm/pmu.c | 47 ++++++++++++++++++++++++++++++++++++++++++++++- > 3 files changed, 50 insertions(+), 1 deletion(-) > > diff --git a/arch/arm/kvm/arm.c b/arch/arm/kvm/arm.c > index dda1959..f54264c 100644 > --- a/arch/arm/kvm/arm.c > +++ b/arch/arm/kvm/arm.c > @@ -28,6 +28,7 @@ > #include > #include > #include > +#include > > #define CREATE_TRACE_POINTS > #include "trace.h" > @@ -577,6 +578,7 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu, struct kvm_run *run) > * non-preemptible context. > */ > preempt_disable(); > + kvm_pmu_flush_hwstate(vcpu); > kvm_timer_flush_hwstate(vcpu); > kvm_vgic_flush_hwstate(vcpu); > > diff --git a/include/kvm/arm_pmu.h b/include/kvm/arm_pmu.h > index 8bc92d1..cf68f9a 100644 > --- a/include/kvm/arm_pmu.h > +++ b/include/kvm/arm_pmu.h > @@ -44,6 +44,7 @@ u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu); > void kvm_pmu_disable_counter(struct kvm_vcpu *vcpu, u64 val); > void kvm_pmu_enable_counter(struct kvm_vcpu *vcpu, u64 val); > void kvm_pmu_overflow_set(struct kvm_vcpu *vcpu, u64 val); > +void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu); > void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val); > void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val); > void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data, > @@ -67,6 +68,7 @@ static inline u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu) > static inline void kvm_pmu_disable_counter(struct kvm_vcpu *vcpu, u64 val) {} > static inline void kvm_pmu_enable_counter(struct kvm_vcpu *vcpu, u64 val) {} > static inline void kvm_pmu_overflow_set(struct kvm_vcpu *vcpu, u64 val) {} > +static inline void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu) {} > static inline void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val) {} > static inline void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val) {} > static inline void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, > diff --git a/virt/kvm/arm/pmu.c b/virt/kvm/arm/pmu.c > index cda869c..6ac52ce 100644 > --- a/virt/kvm/arm/pmu.c > +++ b/virt/kvm/arm/pmu.c > @@ -21,6 +21,7 @@ > #include > #include > #include > +#include > > /** > * kvm_pmu_get_counter_value - get PMU counter value > @@ -181,6 +182,49 @@ void kvm_pmu_overflow_set(struct kvm_vcpu *vcpu, u64 val) > } > > /** > + * kvm_pmu_flush_hwstate - flush pmu state to cpu > + * @vcpu: The vcpu pointer > + * > + * Inject virtual PMU IRQ if IRQ is pending for this cpu. > + */ > +void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu) > +{ > + struct kvm_pmu *pmu = &vcpu->arch.pmu; > + u64 overflow; > + > + if (!kvm_arm_pmu_v3_ready(vcpu)) > + return; > + > + overflow = kvm_pmu_overflow_status(vcpu); > + kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id, pmu->irq_num, !!overflow); It turns out that this single line costs us about 400 cycles on each entry: maz at flakes:~/kvm-ws-tests$ make LKVM=~/kvmtool/lkvm LKVM_ARGS=--pmu PERF=perf_4.3 tests-gicv2 GICv2: do_hvc.bin:5690.17 do_sgi.bin:9395.05 do_sysreg.bin:5912.6 maz at flakes:~/kvm-ws-tests$ make LKVM=~/kvmtool/lkvm PERF=perf_4.3 tests-gicv2 GICv2: do_hvc.bin:5285.02 do_sgi.bin:9131.24 do_sysreg.bin:5563.7 Caching the irq state and only injecting it if it has changed (just like we do for the timer) brings performance back to its previous level: diff --git a/include/kvm/arm_pmu.h b/include/kvm/arm_pmu.h index 176913f..b23e636 100644 --- a/include/kvm/arm_pmu.h +++ b/include/kvm/arm_pmu.h @@ -35,6 +35,7 @@ struct kvm_pmu { int irq_num; struct kvm_pmc pmc[ARMV8_PMU_MAX_COUNTERS]; bool ready; + bool irq_level; }; #define kvm_arm_pmu_v3_ready(v) ((v)->arch.pmu.ready) diff --git a/virt/kvm/arm/pmu.c b/virt/kvm/arm/pmu.c index 904617e..7156f8b 100644 --- a/virt/kvm/arm/pmu.c +++ b/virt/kvm/arm/pmu.c @@ -229,13 +229,17 @@ void kvm_pmu_overflow_set(struct kvm_vcpu *vcpu, u64 val) void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu) { struct kvm_pmu *pmu = &vcpu->arch.pmu; - u64 overflow; + bool overflow; if (!kvm_arm_pmu_v3_ready(vcpu)) return; - overflow = kvm_pmu_overflow_status(vcpu); - kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id, pmu->irq_num, !!overflow); + overflow = !!kvm_pmu_overflow_status(vcpu); + if (pmu->irq_level != overflow) { + pmu->irq_level = overflow; + kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id, + pmu->irq_num, overflow); + } } static inline struct kvm_vcpu *kvm_pmc_to_vcpu(struct kvm_pmc *pmc) Thanks, M. -- Jazz is not dead. It just smells funny...