Skip to contentSkip to conversation

arm64: Support perf event modifiers :G and :H

PATCH Superseded Andrew Murray 2 participants 7 messages Dec 4, 2018 – Dec 5, 2018

Subsystems (4)

v4 of 13 Superseded Latest is v13 →

lore
Share & tools
lore
#0 lore raw

v4 cover arm64: Support perf event modifiers :G and :H

This patchset provides support for perf event modifiers :G and :H which allows for filtering of PMU events between host and guests when used with KVM.

As the underlying hardware cannot distinguish between guest and host context, the performance counters must be stopped and started upon entry/exit to the guest. This is performed at EL2 in a way that minimizes overhead and improves accuracy of recording events that only occur in the requested context.

This has been tested with VHE and non-VHE kernels with a KVM guest.

Changes from v3:

- Remove confusing _only suffix from bitfields in kvm_cpu_context
- Remove unnecessary condition when clearing event bits in disable
- Simplify API of KVM accessors
- Prevent unnecessary setting of pmcnten when guest/host events are the same.

Changes from v2:

- Ensured that exclude_kernel works for guest
- Removed unnecessary exclusion of EL2 with exclude_host on !VHE
- Renamed kvm_clr_set_host_pmu_events to reflect args order
- Added additional information to isb patch

Changes from v1:

- Removed unnecessary exclusion of EL1 with exclude_guest on VHE
- Removed unnecessary isb from existing perf_event.c driver
- Folded perf_event.c patches together
- Added additional information to last patch commit message

Andrew Murray (4):
arm64: arm_pmu: remove unnecessary isb instruction
arm64: KVM: add accessors to track guest/host only counters
arm64: arm_pmu: Add support for exclude_host/exclude_guest attributes
arm64: KVM: Enable support for :G/:H perf event modifiers

 arch/arm64/include/asm/kvm_host.h | 24 +++++++++++++++++++
 arch/arm64/kernel/perf_event.c    | 49 ++++++++++++++++++++++++++++++++-------
 arch/arm64/kvm/hyp/switch.c       | 41 ++++++++++++++++++++++++++++++++
 3 files changed, 106 insertions(+), 8 deletions(-)
Signature · 1 line
2.7.4
#1 lore raw

PATCH v4 1/4 arm64: arm_pmu: remove unnecessary isb instruction

The armv8pmu_enable_event_counter function issues an isb instruction after enabling a pair of counters - this doesn't provide any value and is inconsistent with the armv8pmu_disable_event_counter.

In any case armv8pmu_enable_event_counter is always called with the PMU stopped. Starting the PMU with armv8pmu_start results in an isb instruction being issued prior to writing to PMCR_EL0.

Let's remove the unnecessary isb instruction.

---
 arch/arm64/kernel/perf_event.c | 1 -
 1 file changed, 1 deletion(-)
arch/arm64/kernel/perf_event.c +0 -1
index 8e38d52..de564ae 100644
@@ -652,7 +652,6 @@ static inline void armv8pmu_enable_event_counter(struct perf_event *event)
 	armv8pmu_enable_counter(idx);
 	if (armv8pmu_event_is_chained(event))
 		armv8pmu_enable_counter(idx - 1);
-	isb();
 }
 
 static inline int armv8pmu_disable_counter(int idx)
Signature · 1 line
2.7.4
#2 lore raw

PATCH v4 2/4 arm64: KVM: add accessors to track guest/host only counters

In reply to Andrew Murray

In order to effeciently enable/disable guest/host only perf counters at guest entry/exit we add bitfields to kvm_cpu_context for guest and host events as well as accessors for updating them.

---
 arch/arm64/include/asm/kvm_host.h | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)
arch/arm64/include/asm/kvm_host.h +24 -0
index 1550192..800c87b 100644
@@ -203,6 +203,8 @@ struct kvm_cpu_context {
 	};
 
 	struct kvm_vcpu *__hyp_running_vcpu;
+	u32 events_host;
+	u32 events_guest;
 };
 
 typedef struct kvm_cpu_context kvm_cpu_context_t;
@@ -467,11 +469,33 @@ void kvm_arch_vcpu_load_fp(struct kvm_vcpu *vcpu);
 void kvm_arch_vcpu_ctxsync_fp(struct kvm_vcpu *vcpu);
 void kvm_arch_vcpu_put_fp(struct kvm_vcpu *vcpu);
 
+#define KVM_PMU_EVENTS_HOST	1
+#define KVM_PMU_EVENTS_GUEST	2
+
 #ifdef CONFIG_KVM /* Avoid conflicts with core headers if CONFIG_KVM=n */
 static inline int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
 {
 	return kvm_arch_vcpu_run_map_fp(vcpu);
 }
+static inline void kvm_set_pmu_events(u32 set, int flags)
+{
+	kvm_cpu_context_t *ctx = this_cpu_ptr(&kvm_host_cpu_state);
+
+	if (flags & KVM_PMU_EVENTS_HOST)
+		ctx->events_host |= set;
+	if (flags & KVM_PMU_EVENTS_GUEST)
+		ctx->events_guest |= set;
+}
+static inline void kvm_clr_pmu_events(u32 clr)
+{
+	kvm_cpu_context_t *ctx = this_cpu_ptr(&kvm_host_cpu_state);
+
+	ctx->events_host &= ~clr;
+	ctx->events_guest &= ~clr;
+}
+#else
+static inline void kvm_set_pmu_events(u32 set, int flags) {}
+static inline void kvm_clr_pmu_events(u32 clr) {}
 #endif
 
 static inline void kvm_arm_vhe_guest_enter(void)
Signature · 1 line
2.7.4
#3 lore raw

PATCH v4 3/4 arm64: arm_pmu: Add support for exclude_host/exclude_guest attributes

In reply to Andrew Murray

Add support for the :G and :H attributes in perf by handling the exclude_host/exclude_guest event attributes.

We notify KVM of counters that we wish to be enabled or disabled on guest entry/exit and thus defer from starting or stopping :G events as per the events exclude_host attribute.

With both VHE and non-VHE we switch the counters between host/guest at EL2. We are able to eliminate counters counting host events on the boundaries of guest entry/exit when using :G by filtering out EL2 for exclude_host. However when using :H unless exclude_hv is set on non-VHE then there is a small blackout window at the guest entry/exit where host events are not captured.

---
 arch/arm64/kernel/perf_event.c | 48 ++++++++++++++++++++++++++++++++++++------
 1 file changed, 41 insertions(+), 7 deletions(-)
arch/arm64/kernel/perf_event.c +41 -7
index de564ae..a5bd626 100644
@@ -26,6 +26,7 @@
 
 #include <linux/acpi.h>
 #include <linux/clocksource.h>
+#include <linux/kvm_host.h>
 #include <linux/of.h>
 #include <linux/perf/arm_pmu.h>
 #include <linux/platform_device.h>
@@ -647,11 +648,23 @@ static inline int armv8pmu_enable_counter(int idx)
 
 static inline void armv8pmu_enable_event_counter(struct perf_event *event)
 {
+	struct perf_event_attr *attr = &event->attr;
 	int idx = event->hw.idx;
+	u32 counter_bits = BIT(ARMV8_IDX_TO_COUNTER(idx));
 
-	armv8pmu_enable_counter(idx);
 	if (armv8pmu_event_is_chained(event))
-		armv8pmu_enable_counter(idx - 1);
+		counter_bits |= BIT(ARMV8_IDX_TO_COUNTER(idx - 1));
+
+	if (!attr->exclude_host)
+		kvm_set_pmu_events(counter_bits, KVM_PMU_EVENTS_HOST);
+	if (!attr->exclude_guest)
+		kvm_set_pmu_events(counter_bits, KVM_PMU_EVENTS_GUEST);
+
+	if (!attr->exclude_host) {
+		armv8pmu_enable_counter(idx);
+		if (armv8pmu_event_is_chained(event))
+			armv8pmu_enable_counter(idx - 1);
+	}
 }
 
 static inline int armv8pmu_disable_counter(int idx)
@@ -664,11 +677,20 @@ static inline int armv8pmu_disable_counter(int idx)
 static inline void armv8pmu_disable_event_counter(struct perf_event *event)
 {
 	struct hw_perf_event *hwc = &event->hw;
+	struct perf_event_attr *attr = &event->attr;
 	int idx = hwc->idx;
+	u32 counter_bits = BIT(ARMV8_IDX_TO_COUNTER(idx));
 
 	if (armv8pmu_event_is_chained(event))
-		armv8pmu_disable_counter(idx - 1);
-	armv8pmu_disable_counter(idx);
+		counter_bits |= BIT(ARMV8_IDX_TO_COUNTER(idx - 1));
+
+	kvm_clr_pmu_events(counter_bits);
+
+	if (!attr->exclude_host) {
+		if (armv8pmu_event_is_chained(event))
+			armv8pmu_disable_counter(idx - 1);
+		armv8pmu_disable_counter(idx);
+	}
 }
 
 static inline int armv8pmu_enable_intens(int idx)
@@ -943,16 +965,25 @@ static int armv8pmu_set_event_filter(struct hw_perf_event *event,
 	 * Therefore we ignore exclude_hv in this configuration, since
 	 * there's no hypervisor to sample anyway. This is consistent
 	 * with other architectures (x86 and Power).
+	 *
+	 * To eliminate counting host events on the boundaries of
+	 * guest entry/exit we ensure EL2 is not included in hyp mode
+	 * with !exclude_host.
 	 */
 	if (is_kernel_in_hyp_mode()) {
-		if (!attr->exclude_kernel)
+		if (!attr->exclude_kernel && !attr->exclude_host)
 			config_base |= ARMV8_PMU_INCLUDE_EL2;
 	} else {
-		if (attr->exclude_kernel)
-			config_base |= ARMV8_PMU_EXCLUDE_EL1;
 		if (!attr->exclude_hv)
 			config_base |= ARMV8_PMU_INCLUDE_EL2;
 	}
+
+	/*
+	 * Filter out !VHE kernels and guest kernels
+	 */
+	if (attr->exclude_kernel)
+		config_base |= ARMV8_PMU_EXCLUDE_EL1;
+
 	if (attr->exclude_user)
 		config_base |= ARMV8_PMU_EXCLUDE_EL0;
 
@@ -976,6 +1007,9 @@ static void armv8pmu_reset(void *info)
 		armv8pmu_disable_intens(idx);
 	}
 
+	/* Clear the counters we flip at guest entry/exit */
+	kvm_clr_pmu_events(U32_MAX);
+
 	/*
 	 * Initialize & Reset PMNC. Request overflow interrupt for
 	 * 64 bit cycle counter but cheat in armv8pmu_write_counter().
Signature · 1 line
2.7.4
#4 lore raw

PATCH v4 4/4 arm64: KVM: Enable support for :G/:H perf event modifiers

In reply to Andrew Murray

Enable/disable event counters as appropriate when entering and exiting the guest to enable support for guest or host only event counting.

For both VHE and non-VHE we switch the counters between host/guest at EL2. EL2 is filtered out by the PMU when we are using the :G modifier.

The PMU may be on when we change which counters are enabled however we avoid adding an isb as we instead rely on existing context synchronisation events: the isb in kvm_arm_vhe_guest_exit for VHE and the eret from the hvc in kvm_call_hyp.

---
 arch/arm64/kvm/hyp/switch.c | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)
arch/arm64/kvm/hyp/switch.c +41 -0
index d496ef5..5e03921 100644
@@ -373,6 +373,35 @@ static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu)
 	return true;
 }
 
+static bool __hyp_text __pmu_switch_to_guest(struct kvm_cpu_context *host_ctxt)
+{
+	u32 host = host_ctxt->events_host;
+	u32 guest = host_ctxt->events_guest;
+
+	if (host == guest)
+		return false;
+
+	if (host)
+		write_sysreg(host, pmcntenclr_el0);
+
+	if (guest)
+		write_sysreg(guest, pmcntenset_el0);
+
+	return (host || guest);
+}
+
+static void __hyp_text __pmu_switch_to_host(struct kvm_cpu_context *host_ctxt)
+{
+	u32 host = host_ctxt->events_host;
+	u32 guest = host_ctxt->events_guest;
+
+	if (guest)
+		write_sysreg(guest, pmcntenclr_el0);
+
+	if (host)
+		write_sysreg(host, pmcntenset_el0);
+}
+
 /*
  * Return true when we were able to fixup the guest exit and should return to
  * the guest, false when we should restore the host state and return to the
@@ -488,12 +517,15 @@ int kvm_vcpu_run_vhe(struct kvm_vcpu *vcpu)
 {
 	struct kvm_cpu_context *host_ctxt;
 	struct kvm_cpu_context *guest_ctxt;
+	bool pmu_switch_needed;
 	u64 exit_code;
 
 	host_ctxt = vcpu->arch.host_cpu_context;
 	host_ctxt->__hyp_running_vcpu = vcpu;
 	guest_ctxt = &vcpu->arch.ctxt;
 
+	pmu_switch_needed = __pmu_switch_to_guest(host_ctxt);
+
 	sysreg_save_host_state_vhe(host_ctxt);
 
 	__activate_traps(vcpu);
@@ -524,6 +556,9 @@ int kvm_vcpu_run_vhe(struct kvm_vcpu *vcpu)
 
 	__debug_switch_to_host(vcpu);
 
+	if (pmu_switch_needed)
+		__pmu_switch_to_host(host_ctxt);
+
 	return exit_code;
 }
 
@@ -532,6 +567,7 @@ int __hyp_text __kvm_vcpu_run_nvhe(struct kvm_vcpu *vcpu)
 {
 	struct kvm_cpu_context *host_ctxt;
 	struct kvm_cpu_context *guest_ctxt;
+	bool pmu_switch_needed;
 	u64 exit_code;
 
 	vcpu = kern_hyp_va(vcpu);
@@ -540,6 +576,8 @@ int __hyp_text __kvm_vcpu_run_nvhe(struct kvm_vcpu *vcpu)
 	host_ctxt->__hyp_running_vcpu = vcpu;
 	guest_ctxt = &vcpu->arch.ctxt;
 
+	pmu_switch_needed = __pmu_switch_to_guest(host_ctxt);
+
 	__sysreg_save_state_nvhe(host_ctxt);
 
 	__activate_traps(vcpu);
@@ -586,6 +624,9 @@ int __hyp_text __kvm_vcpu_run_nvhe(struct kvm_vcpu *vcpu)
 	 */
 	__debug_switch_to_host(vcpu);
 
+	if (pmu_switch_needed)
+		__pmu_switch_to_host(host_ctxt);
+
 	return exit_code;
 }
 
Signature · 1 line
2.7.4
#5 lore raw

Re: [PATCH v4 4/4] arm64: KVM: Enable support for :G/:H perf event modifiers

Hi Andrew,

Quoted 24 lines · Andrew Murray

Enable/disable event counters as appropriate when entering and exiting the guest to enable support for guest or host only event counting.

For both VHE and non-VHE we switch the counters between host/guest at EL2. EL2 is filtered out by the PMU when we are using the :G modifier.

The PMU may be on when we change which counters are enabled however we avoid adding an isb as we instead rely on existing context synchronisation events: the isb in kvm_arm_vhe_guest_exit for VHE and the eret from the hvc in kvm_call_hyp.

---
 arch/arm64/kvm/hyp/switch.c | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)
arch/arm64/kvm/hyp/switch.c +0 -0
index d496ef5..5e03921 100644
@@ -373,6 +373,35 @@ static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu)
 	return true;
 }
 

arch/arm64/kvm/hyp/switch.c static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu) · line 385

Quoted 10 lines
arch/arm64/kvm/hyp/switch.c +10 -0
+static bool __hyp_text __pmu_switch_to_guest(struct kvm_cpu_context *host_ctxt)
+{
+	u32 host = host_ctxt->events_host;
+	u32 guest = host_ctxt->events_guest;
+
+	if (host == guest)
+		return false;
+
+	if (host)
+		write_sysreg(host, pmcntenclr_el0);

Nit:
Instead of clearing host counters we could clear non-guest counters:

	u32 clear_events = host_ctxt->events_hosts & ~guest;
		
	if (clear_events)
		write_sysreg(clear_events, pmcntenclr_el0);

Don't know if it makes a lot of difference.

arch/arm64/kvm/hyp/switch.c static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu) · line 390

Quoted 5 lines
arch/arm64/kvm/hyp/switch.c +5 -0
+
+	if (guest)
+		write_sysreg(guest, pmcntenset_el0);
+
+	return (host || guest);

We know this is true. If both (host == 0 && guest == 0) then (host == guest) meaning we returned false.

arch/arm64/kvm/hyp/switch.c static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu) · line 399

Quoted 9 lines
arch/arm64/kvm/hyp/switch.c +9 -0
+}
+
+static void __hyp_text __pmu_switch_to_host(struct kvm_cpu_context *host_ctxt)
+{
+	u32 host = host_ctxt->events_host;
+	u32 guest = host_ctxt->events_guest;
+
+	if (guest)
+		write_sysreg(guest, pmcntenclr_el0);

Nit:
Same as above, we could just clear the counters that are exclusive to the guest.

Cheers,

Signature · 1 line
Julien Thierry
#6 lore raw

Re: [PATCH v4 4/4] arm64: KVM: Enable support for :G/:H perf event modifiers

Quoted 41 lines · Julien Thierry

Hi Andrew,

Quoted 24 lines · Andrew Murray

Enable/disable event counters as appropriate when entering and exiting the guest to enable support for guest or host only event counting.

For both VHE and non-VHE we switch the counters between host/guest at EL2. EL2 is filtered out by the PMU when we are using the :G modifier.

The PMU may be on when we change which counters are enabled however we avoid adding an isb as we instead rely on existing context synchronisation events: the isb in kvm_arm_vhe_guest_exit for VHE and the eret from the hvc in kvm_call_hyp.

---
 arch/arm64/kvm/hyp/switch.c | 41 +++++++++++++++++++++++++++++++++++++++++
 1 file changed, 41 insertions(+)
arch/arm64/kvm/hyp/switch.c +0 -0
index d496ef5..5e03921 100644
@@ -373,6 +373,35 @@ static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu)
 	return true;
 }
 

arch/arm64/kvm/hyp/switch.c static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu) · line 385

Quoted 10 lines
arch/arm64/kvm/hyp/switch.c +10 -0
+static bool __hyp_text __pmu_switch_to_guest(struct kvm_cpu_context *host_ctxt)
+{
+	u32 host = host_ctxt->events_host;
+	u32 guest = host_ctxt->events_guest;
+
+	if (host == guest)
+		return false;
+
+	if (host)
+		write_sysreg(host, pmcntenclr_el0);

Nit:
Instead of clearing host counters we could clear non-guest counters:

Quoted 6 lines
	u32 clear_events = host_ctxt->events_hosts & ~guest;
		
	if (clear_events)
		write_sysreg(clear_events, pmcntenclr_el0);

Don't know if it makes a lot of difference.

In the case where an event is enabled for both host and guest, your suggested change prevents us from unnecessarily clearing and then setting the same bit. Thus is may prevent a call to write_sysreg.

Quoted 9 lines

arch/arm64/kvm/hyp/switch.c static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu) · line 390

Quoted 6 lines
arch/arm64/kvm/hyp/switch.c +5 -0
+
+	if (guest)
+		write_sysreg(guest, pmcntenset_el0);
+
+	return (host || guest);

We know this is true. If both (host == 0 && guest == 0) then (host == guest) meaning we returned false.

Doh.

Quoted 15 lines

arch/arm64/kvm/hyp/switch.c static bool __hyp_text __hyp_switch_fpsimd(struct kvm_vcpu *vcpu) · line 399

Quoted 11 lines
arch/arm64/kvm/hyp/switch.c +9 -0
+}
+
+static void __hyp_text __pmu_switch_to_host(struct kvm_cpu_context *host_ctxt)
+{
+	u32 host = host_ctxt->events_host;
+	u32 guest = host_ctxt->events_guest;
+
+	if (guest)
+		write_sysreg(guest, pmcntenclr_el0);

Nit:
Same as above, we could just clear the counters that are exclusive to the guest.

Thanks,

Andrew Murray

Quoted 5 lines

Cheers,

Signature · 1 line
Julien Thierry