| Message ID | 1479316160-8567-1-git-send-email-wei@redhat.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Wed, Nov 16, 2016 at 11:09:20AM -0600, Wei Huang wrote: > KVM calls kvm_pmu_set_counter_event_type() when PMCCFILTR is configured. > But this function can't deals with PMCCFILTR correctly because the evtCount > bits of PMCCFILTR, which is reserved 0, conflits with the SW_INCR event > type of other PMXEVTYPER<n> registers. To fix it, when eventsel == 0, this > function shouldn't return immediately; instead it needs to check further > if select_idx is ARMV8_PMU_CYCLE_IDX. > > Another issue is that KVM shouldn't copy the eventsel bits of PMCCFILTER > blindly to attr.config. Instead it ought to convert the request to the > "cpu cycle" event type (i.e. 0x11). > > To support this patch and to prevent duplicated definitions, a limited > set of ARMv8 perf event types were relocated from perf_event.c to > asm/perf_event.h. > > Signed-off-by: Wei Huang <wei@redhat.com> > --- > arch/arm64/include/asm/perf_event.h | 10 +++++++++- > arch/arm64/kernel/perf_event.c | 10 +--------- > virt/kvm/arm/pmu.c | 8 +++++--- > 3 files changed, 15 insertions(+), 13 deletions(-) Acked-by: Will Deacon <will.deacon@arm.com> I'm assuming this will go via kvm-arm. Will -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 17/11/16 13:30, Will Deacon wrote: > On Wed, Nov 16, 2016 at 11:09:20AM -0600, Wei Huang wrote: >> KVM calls kvm_pmu_set_counter_event_type() when PMCCFILTR is configured. >> But this function can't deals with PMCCFILTR correctly because the evtCount >> bits of PMCCFILTR, which is reserved 0, conflits with the SW_INCR event >> type of other PMXEVTYPER<n> registers. To fix it, when eventsel == 0, this >> function shouldn't return immediately; instead it needs to check further >> if select_idx is ARMV8_PMU_CYCLE_IDX. >> >> Another issue is that KVM shouldn't copy the eventsel bits of PMCCFILTER >> blindly to attr.config. Instead it ought to convert the request to the >> "cpu cycle" event type (i.e. 0x11). >> >> To support this patch and to prevent duplicated definitions, a limited >> set of ARMv8 perf event types were relocated from perf_event.c to >> asm/perf_event.h. >> >> Signed-off-by: Wei Huang <wei@redhat.com> >> --- >> arch/arm64/include/asm/perf_event.h | 10 +++++++++- >> arch/arm64/kernel/perf_event.c | 10 +--------- >> virt/kvm/arm/pmu.c | 8 +++++--- >> 3 files changed, 15 insertions(+), 13 deletions(-) > > Acked-by: Will Deacon <will.deacon@arm.com> > > I'm assuming this will go via kvm-arm. Yup, I'll take it. Thanks. M.
diff --git a/arch/arm64/include/asm/perf_event.h b/arch/arm64/include/asm/perf_event.h index 2065f46..38b6a2b 100644 --- a/arch/arm64/include/asm/perf_event.h +++ b/arch/arm64/include/asm/perf_event.h @@ -46,7 +46,15 @@ #define ARMV8_PMU_EVTYPE_MASK 0xc800ffff /* Mask for writable bits */ #define ARMV8_PMU_EVTYPE_EVENT 0xffff /* Mask for EVENT bits */ -#define ARMV8_PMU_EVTYPE_EVENT_SW_INCR 0 /* Software increment event */ +/* + * PMUv3 event types: required events + */ +#define ARMV8_PMUV3_PERFCTR_SW_INCR 0x00 +#define ARMV8_PMUV3_PERFCTR_L1D_CACHE_REFILL 0x03 +#define ARMV8_PMUV3_PERFCTR_L1D_CACHE 0x04 +#define ARMV8_PMUV3_PERFCTR_BR_MIS_PRED 0x10 +#define ARMV8_PMUV3_PERFCTR_CPU_CYCLES 0x11 +#define ARMV8_PMUV3_PERFCTR_BR_PRED 0x12 /* * Event filters for PMUv3 diff --git a/arch/arm64/kernel/perf_event.c b/arch/arm64/kernel/perf_event.c index a9310a6..57ae9d9 100644 --- a/arch/arm64/kernel/perf_event.c +++ b/arch/arm64/kernel/perf_event.c @@ -31,17 +31,9 @@ /* * ARMv8 PMUv3 Performance Events handling code. - * Common event types. + * Common event types (some are defined in asm/perf_event.h). */ -/* Required events. */ -#define ARMV8_PMUV3_PERFCTR_SW_INCR 0x00 -#define ARMV8_PMUV3_PERFCTR_L1D_CACHE_REFILL 0x03 -#define ARMV8_PMUV3_PERFCTR_L1D_CACHE 0x04 -#define ARMV8_PMUV3_PERFCTR_BR_MIS_PRED 0x10 -#define ARMV8_PMUV3_PERFCTR_CPU_CYCLES 0x11 -#define ARMV8_PMUV3_PERFCTR_BR_PRED 0x12 - /* At least one of the following is required. */ #define ARMV8_PMUV3_PERFCTR_INST_RETIRED 0x08 #define ARMV8_PMUV3_PERFCTR_INST_SPEC 0x1B diff --git a/virt/kvm/arm/pmu.c b/virt/kvm/arm/pmu.c index 6e9c40e..69ccce3 100644 --- a/virt/kvm/arm/pmu.c +++ b/virt/kvm/arm/pmu.c @@ -305,7 +305,7 @@ void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val) continue; type = vcpu_sys_reg(vcpu, PMEVTYPER0_EL0 + i) & ARMV8_PMU_EVTYPE_EVENT; - if ((type == ARMV8_PMU_EVTYPE_EVENT_SW_INCR) + if ((type == ARMV8_PMUV3_PERFCTR_SW_INCR) && (enable & BIT(i))) { reg = vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) + 1; reg = lower_32_bits(reg); @@ -379,7 +379,8 @@ void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data, eventsel = data & ARMV8_PMU_EVTYPE_EVENT; /* Software increment event does't need to be backed by a perf event */ - if (eventsel == ARMV8_PMU_EVTYPE_EVENT_SW_INCR) + if (eventsel == ARMV8_PMUV3_PERFCTR_SW_INCR && + select_idx != ARMV8_PMU_CYCLE_IDX) return; memset(&attr, 0, sizeof(struct perf_event_attr)); @@ -391,7 +392,8 @@ void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data, attr.exclude_kernel = data & ARMV8_PMU_EXCLUDE_EL1 ? 1 : 0; attr.exclude_hv = 1; /* Don't count EL2 events */ attr.exclude_host = 1; /* Don't count host events */ - attr.config = eventsel; + attr.config = (select_idx == ARMV8_PMU_CYCLE_IDX) ? + ARMV8_PMUV3_PERFCTR_CPU_CYCLES : eventsel; counter = kvm_pmu_get_counter_value(vcpu, select_idx); /* The initial sample period (overflow count) of an event. */
KVM calls kvm_pmu_set_counter_event_type() when PMCCFILTR is configured. But this function can't deals with PMCCFILTR correctly because the evtCount bits of PMCCFILTR, which is reserved 0, conflits with the SW_INCR event type of other PMXEVTYPER<n> registers. To fix it, when eventsel == 0, this function shouldn't return immediately; instead it needs to check further if select_idx is ARMV8_PMU_CYCLE_IDX. Another issue is that KVM shouldn't copy the eventsel bits of PMCCFILTER blindly to attr.config. Instead it ought to convert the request to the "cpu cycle" event type (i.e. 0x11). To support this patch and to prevent duplicated definitions, a limited set of ARMv8 perf event types were relocated from perf_event.c to asm/perf_event.h. Signed-off-by: Wei Huang <wei@redhat.com> --- arch/arm64/include/asm/perf_event.h | 10 +++++++++- arch/arm64/kernel/perf_event.c | 10 +--------- virt/kvm/arm/pmu.c | 8 +++++--- 3 files changed, 15 insertions(+), 13 deletions(-)