Skip to contentSkip to conversation

KVM: x86/pmu: Optimize triggering of emulated events

PATCH Applied Sean Christopherson 3 participants 15 messages Nov 10, 2023 – Feb 3, 2024

Subsystems (3)
lore
Share & tools
lore

Sean Christopherson

author maintainer
#0 lore raw

cover KVM: x86/pmu: Optimize triggering of emulated events

Optimize code used by, or which impacts, kvm_pmu_trigger_event() to try and make a dent in the overhead of emulating PMU events in software, which is quite noticeable due to it kicking in anytime the guest has a vPMU and KVM is skipping an instruction.

Note, Jim has a proposal/idea[*] (that I supported) to make kvm_pmu_trigger_event() even more performant. I opted not to do that as it's a bit more invasive, and I started chewing on this not so much because I care _that_ much about performance, but because it irritates me that the PMU code makes things way harder than they need to be.

Note #2, this applies on top of my other two PMU series:

https://lore.kernel.org/all/[email protected]
https://lore.kernel.org/all/[email protected]

Those series fix actual functional issues, i.e. I'll definitely apply them first (there's definitely no rush on this one).

[*] https://lore.kernel.org/all/CALMp9eQGqqo66fQGwFJMc3y+9XdUrL7ageE8kvoAOV6NJGfJpw@mail.gmail.com

Sean Christopherson (10):
KVM: x86/pmu: Zero out PMU metadata on AMD if PMU is disabled
KVM: x86/pmu: Add common define to capture fixed counters offset
KVM: x86/pmu: Move pmc_idx => pmc translation helper to common code
KVM: x86/pmu: Snapshot and clear reprogramming bitmap before
reprogramming
KVM: x86/pmu: Add macros to iterate over all PMCs given a bitmap
KVM: x86/pmu: Process only enabled PMCs when emulating events in
software
KVM: x86/pmu: Snapshot event selectors that KVM emulates in software
KVM: x86/pmu: Expand the comment about what bits are check emulating
events
KVM: x86/pmu: Check eventsel first when emulating (branch) insns
retired
KVM: x86/pmu: Avoid CPL lookup if PMC enabline for USER and KERNEL is
the same

 arch/x86/include/asm/kvm-x86-pmu-ops.h |   1 -
 arch/x86/include/asm/kvm_host.h        |   1 +
 arch/x86/kvm/pmu.c                     | 143 +++++++++++++++----------
 arch/x86/kvm/pmu.h                     |  52 ++++++++-
 arch/x86/kvm/svm/pmu.c                 |   7 +-
 arch/x86/kvm/vmx/nested.c              |   2 +-
 arch/x86/kvm/vmx/pmu_intel.c           |  44 ++------
 arch/x86/kvm/x86.c                     |   6 +-
 8 files changed, 154 insertions(+), 102 deletions(-)

base-commit: ef1883475d4a24d8eaebb84175ed46206a688103

Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#1 lore raw

PATCH 1/10 KVM: x86/pmu: Zero out PMU metadata on AMD if PMU is disabled

Move the purging of common PMU metadata from intel_pmu_refresh() to kvm_pmu_refresh(), and invoke the vendor refresh() hook if and only if the VM is supposed to have a vPMU.

KVM already denies access to the PMU based on kvm->arch.enable_pmu, as get_gp_pmc_amd() returns NULL for all PMCs in that case, i.e. KVM already violates AMD's architecture by not virtualizing a PMU (kernels have long since learned to not panic when the PMU is unavailable). But configuring the PMU as if it were enabled causes unwanted side effects, e.g. calls to kvm_pmu_trigger_event() waste an absurd number of cycles due to the all_valid_pmc_idx bitmap being non-zero.

Fixes b1d66dad65dc ("KVM: x86/svm: Add module param to control PMU virtualization")
---
 arch/x86/kvm/pmu.c           | 20 ++++++++++++++++++--
 arch/x86/kvm/vmx/pmu_intel.c | 16 ++--------------
 2 files changed, 20 insertions(+), 16 deletions(-)
arch/x86/kvm/pmu.c +18 -2
index 1b74a29ed250..b52bab7dc422 100644
@@ -739,6 +739,8 @@ static void kvm_pmu_reset(struct kvm_vcpu *vcpu)
  */
 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 {
+	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
+
 	if (KVM_BUG_ON(kvm_vcpu_has_run(vcpu), vcpu->kvm))
 		return;
 
@@ -748,8 +750,22 @@ void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 	 */
 	kvm_pmu_reset(vcpu);
 
-	bitmap_zero(vcpu_to_pmu(vcpu)->all_valid_pmc_idx, X86_PMC_IDX_MAX);
-	static_call(kvm_x86_pmu_refresh)(vcpu);
+	pmu->version = 0;
+	pmu->nr_arch_gp_counters = 0;
+	pmu->nr_arch_fixed_counters = 0;
+	pmu->counter_bitmask[KVM_PMC_GP] = 0;
+	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
+	pmu->reserved_bits = 0xffffffff00200000ull;
+	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
+	pmu->global_ctrl_mask = ~0ull;
+	pmu->global_status_mask = ~0ull;
+	pmu->fixed_ctr_ctrl_mask = ~0ull;
+	pmu->pebs_enable_mask = ~0ull;
+	pmu->pebs_data_cfg_mask = ~0ull;
+	bitmap_zero(pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX);
+
+	if (vcpu->kvm->arch.enable_pmu)
+		static_call(kvm_x86_pmu_refresh)(vcpu);
 }
 
 void kvm_pmu_init(struct kvm_vcpu *vcpu)
arch/x86/kvm/vmx/pmu_intel.c +2 -14
index c3a841d8df27..0d2fd9fdcf4b 100644
@@ -463,19 +463,6 @@ static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
 	u64 counter_mask;
 	int i;
 
-	pmu->nr_arch_gp_counters = 0;
-	pmu->nr_arch_fixed_counters = 0;
-	pmu->counter_bitmask[KVM_PMC_GP] = 0;
-	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
-	pmu->version = 0;
-	pmu->reserved_bits = 0xffffffff00200000ull;
-	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
-	pmu->global_ctrl_mask = ~0ull;
-	pmu->global_status_mask = ~0ull;
-	pmu->fixed_ctr_ctrl_mask = ~0ull;
-	pmu->pebs_enable_mask = ~0ull;
-	pmu->pebs_data_cfg_mask = ~0ull;
-
 	memset(&lbr_desc->records, 0, sizeof(lbr_desc->records));
 
 	/*
@@ -487,8 +474,9 @@ static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
 		return;
 
 	entry = kvm_find_cpuid_entry(vcpu, 0xa);
-	if (!entry || !vcpu->kvm->arch.enable_pmu)
+	if (!entry)
 		return;
+
 	eax.full = entry->eax;
 	edx.full = entry->edx;
 
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#2 lore raw

PATCH 2/10 KVM: x86/pmu: Add common define to capture fixed counters offset

In reply to Sean Christopherson

Add a common define to "officially" solidify KVM's split of counters, i.e. to commit to using bits 31:0 to track general purpose counters and bits 63:32 to track fixed counters (which only Intel supports). KVM already bleeds this behavior all over common PMU code, and adding a KVM- defined macro allows clarifying that the value is a _base_, as oppposed to the _flag_ that is used to access fixed PMCs via RDPMC (which perf confusingly calls INTEL_PMC_FIXED_RDPMC_BASE).

No functional change intended.

---
 arch/x86/kvm/pmu.c           |  8 ++++----
 arch/x86/kvm/pmu.h           |  4 +++-
 arch/x86/kvm/vmx/pmu_intel.c | 12 ++++++------
 3 files changed, 13 insertions(+), 11 deletions(-)
arch/x86/kvm/pmu.c +4 -4
index b52bab7dc422..714fa6dd912e 100644
@@ -67,7 +67,7 @@ static const struct x86_cpu_id vmx_pebs_pdist_cpu[] = {
  *        all perf counters (both gp and fixed). The mapping relationship
  *        between pmc and perf counters is as the following:
  *        * Intel: [0 .. KVM_INTEL_PMC_MAX_GENERIC-1] <=> gp counters
- *                 [INTEL_PMC_IDX_FIXED .. INTEL_PMC_IDX_FIXED + 2] <=> fixed
+ *                 [KVM_FIXED_PMC_BASE_IDX .. KVM_FIXED_PMC_BASE_IDX + 2] <=> fixed
  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] and, for families 15H
  *          and later, [0 .. AMD64_NUM_COUNTERS_CORE-1] <=> gp counters
  */
@@ -411,7 +411,7 @@ static bool is_gp_event_allowed(struct kvm_x86_pmu_event_filter *f,
 static bool is_fixed_event_allowed(struct kvm_x86_pmu_event_filter *filter,
 				   int idx)
 {
-	int fixed_idx = idx - INTEL_PMC_IDX_FIXED;
+	int fixed_idx = idx - KVM_FIXED_PMC_BASE_IDX;
 
 	if (filter->action == KVM_PMU_EVENT_DENY &&
 	    test_bit(fixed_idx, (ulong *)&filter->fixed_counter_bitmap))
@@ -465,7 +465,7 @@ static void reprogram_counter(struct kvm_pmc *pmc)
 
 	if (pmc_is_fixed(pmc)) {
 		fixed_ctr_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl,
-						  pmc->idx - INTEL_PMC_IDX_FIXED);
+						  pmc->idx - KVM_FIXED_PMC_BASE_IDX);
 		if (fixed_ctr_ctrl & 0x1)
 			eventsel |= ARCH_PERFMON_EVENTSEL_OS;
 		if (fixed_ctr_ctrl & 0x2)
@@ -831,7 +831,7 @@ static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 		select_user = config & ARCH_PERFMON_EVENTSEL_USR;
 	} else {
 		config = fixed_ctrl_field(pmc_to_pmu(pmc)->fixed_ctr_ctrl,
-					  pmc->idx - INTEL_PMC_IDX_FIXED);
+					  pmc->idx - KVM_FIXED_PMC_BASE_IDX);
 		select_os = config & 0x1;
 		select_user = config & 0x2;
 	}
arch/x86/kvm/pmu.h +3 -1
index 87ecf22f5b25..7ffa4f1dedb0 100644
@@ -18,6 +18,8 @@
 #define VMWARE_BACKDOOR_PMC_REAL_TIME		0x10001
 #define VMWARE_BACKDOOR_PMC_APPARENT_TIME	0x10002
 
+#define KVM_FIXED_PMC_BASE_IDX INTEL_PMC_IDX_FIXED
+
 struct kvm_pmu_ops {
 	struct kvm_pmc *(*pmc_idx_to_pmc)(struct kvm_pmu *pmu, int pmc_idx);
 	struct kvm_pmc *(*rdpmc_ecx_to_pmc)(struct kvm_vcpu *vcpu,
@@ -130,7 +132,7 @@ static inline bool pmc_speculative_in_use(struct kvm_pmc *pmc)
 
 	if (pmc_is_fixed(pmc))
 		return fixed_ctrl_field(pmu->fixed_ctr_ctrl,
-					pmc->idx - INTEL_PMC_IDX_FIXED) & 0x3;
+					pmc->idx - KVM_FIXED_PMC_BASE_IDX) & 0x3;
 
 	return pmc->eventsel & ARCH_PERFMON_EVENTSEL_ENABLE;
 }
arch/x86/kvm/vmx/pmu_intel.c +6 -6
index 0d2fd9fdcf4b..61252bb733c4 100644
@@ -42,18 +42,18 @@ static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
 
 		pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i);
 
-		__set_bit(INTEL_PMC_IDX_FIXED + i, pmu->pmc_in_use);
+		__set_bit(KVM_FIXED_PMC_BASE_IDX + i, pmu->pmc_in_use);
 		kvm_pmu_request_counter_reprogram(pmc);
 	}
 }
 
 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
 {
-	if (pmc_idx < INTEL_PMC_IDX_FIXED) {
+	if (pmc_idx < KVM_FIXED_PMC_BASE_IDX) {
 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
 				  MSR_P6_EVNTSEL0);
 	} else {
-		u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
+		u32 idx = pmc_idx - KVM_FIXED_PMC_BASE_IDX;
 
 		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
 	}
@@ -508,7 +508,7 @@ static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++)
 		pmu->fixed_ctr_ctrl_mask &= ~(0xbull << (i * 4));
 	counter_mask = ~(((1ull << pmu->nr_arch_gp_counters) - 1) |
-		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED));
+		(((1ull << pmu->nr_arch_fixed_counters) - 1) << KVM_FIXED_PMC_BASE_IDX));
 	pmu->global_ctrl_mask = counter_mask;
 
 	/*
@@ -552,7 +552,7 @@ static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
 			pmu->reserved_bits &= ~ICL_EVENTSEL_ADAPTIVE;
 			for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
 				pmu->fixed_ctr_ctrl_mask &=
-					~(1ULL << (INTEL_PMC_IDX_FIXED + i * 4));
+					~(1ULL << (KVM_FIXED_PMC_BASE_IDX + i * 4));
 			}
 			pmu->pebs_data_cfg_mask = ~0xff00000full;
 		} else {
@@ -578,7 +578,7 @@ static void intel_pmu_init(struct kvm_vcpu *vcpu)
 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
 		pmu->fixed_counters[i].vcpu = vcpu;
-		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
+		pmu->fixed_counters[i].idx = i + KVM_FIXED_PMC_BASE_IDX;
 		pmu->fixed_counters[i].current_config = 0;
 		pmu->fixed_counters[i].eventsel = intel_get_fixed_pmc_eventsel(i);
 	}
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#3 lore raw

PATCH 3/10 KVM: x86/pmu: Move pmc_idx => pmc translation helper to common code

In reply to Sean Christopherson

Add a common helper for *internal* PMC lookups, and delete the ops hook and Intel's implementation. Keep AMD's implementation, but rename it to amd_pmu_get_pmc() to make it somewhat more obvious that it's suited for both KVM-internal and guest-initiated lookups.

Because KVM tracks all counters in a single bitmap, getting a counter when iterating over a bitmap, e.g. of all valid PMCs, requires a small amount of math, that while simple, isn't super obvious and doesn't use the same semantics as PMC lookups from RDPMC! Although AMD doesn't support fixed counters, the common PMU code still behaves as if there a split, the high half of which just happens to always be empty.

Opportunstically add a comment to explain both what is going on, and why KVM uses a single bitmap, e.g. the boilerplate for iterating over separate bitmaps could be done via macros, so it's not (just) about deduplicating code.

---
 arch/x86/include/asm/kvm-x86-pmu-ops.h |  1 -
 arch/x86/kvm/pmu.c                     |  8 +++----
 arch/x86/kvm/pmu.h                     | 29 +++++++++++++++++++++++++-
 arch/x86/kvm/svm/pmu.c                 |  7 +++----
 arch/x86/kvm/vmx/pmu_intel.c           | 15 +------------
 5 files changed, 36 insertions(+), 24 deletions(-)
arch/x86/include/asm/kvm-x86-pmu-ops.h +0 -1
index d7eebee4450c..e5e7f036587f 100644
@@ -12,7 +12,6 @@ BUILD_BUG_ON(1)
  * a NULL definition, for example if "static_call_cond()" will be used
  * at the call sites.
  */
-KVM_X86_PMU_OP(pmc_idx_to_pmc)
 KVM_X86_PMU_OP(rdpmc_ecx_to_pmc)
 KVM_X86_PMU_OP(msr_idx_to_pmc)
 KVM_X86_PMU_OP(is_valid_rdpmc_ecx)
arch/x86/kvm/pmu.c +4 -4
index 714fa6dd912e..6ee05ad35f55 100644
@@ -505,7 +505,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 	int bit;
 
 	for_each_set_bit(bit, pmu->reprogram_pmi, X86_PMC_IDX_MAX) {
-		struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, bit);
+		struct kvm_pmc *pmc = kvm_pmc_idx_to_pmc(pmu, bit);
 
 		if (unlikely(!pmc)) {
 			clear_bit(bit, pmu->reprogram_pmi);
@@ -715,7 +715,7 @@ static void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 	bitmap_zero(pmu->reprogram_pmi, X86_PMC_IDX_MAX);
 
 	for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
-		pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
+		pmc = kvm_pmc_idx_to_pmc(pmu, i);
 		if (!pmc)
 			continue;
 
@@ -791,7 +791,7 @@ void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
 		      pmu->pmc_in_use, X86_PMC_IDX_MAX);
 
 	for_each_set_bit(i, bitmask, X86_PMC_IDX_MAX) {
-		pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
+		pmc = kvm_pmc_idx_to_pmc(pmu, i);
 
 		if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
 			pmc_stop_counter(pmc);
@@ -846,7 +846,7 @@ void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
 	int i;
 
 	for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
-		pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
+		pmc = kvm_pmc_idx_to_pmc(pmu, i);
 
 		if (!pmc || !pmc_event_is_allowed(pmc))
 			continue;
arch/x86/kvm/pmu.h +28 -1
index 7ffa4f1dedb0..2235772a495b 100644
@@ -4,6 +4,8 @@
 
 #include <linux/nospec.h>
 
+#include <asm/kvm_host.h>
+
 #define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu)
 #define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu))
 #define pmc_to_pmu(pmc)   (&(pmc)->vcpu->arch.pmu)
@@ -21,7 +23,6 @@
 #define KVM_FIXED_PMC_BASE_IDX INTEL_PMC_IDX_FIXED
 
 struct kvm_pmu_ops {
-	struct kvm_pmc *(*pmc_idx_to_pmc)(struct kvm_pmu *pmu, int pmc_idx);
 	struct kvm_pmc *(*rdpmc_ecx_to_pmc)(struct kvm_vcpu *vcpu,
 		unsigned int idx, u64 *mask);
 	struct kvm_pmc *(*msr_idx_to_pmc)(struct kvm_vcpu *vcpu, u32 msr);
@@ -56,6 +57,32 @@ static inline bool kvm_pmu_has_perf_global_ctrl(struct kvm_pmu *pmu)
 	return pmu->version > 1;
 }
 
+/*
+ * KVM tracks all counters in 64-bit bitmaps, with general purpose counters
+ * mapped to bits 31:0 and fixed counters mapped to 63:32, e.g. fixed counter 0
+ * is tracked internally via index 32.  On Intel, (AMD doesn't support fixed
+ * counters), this mirrors how fixed counters are mapped to PERF_GLOBAL_CTRL
+ * and similar MSRs, i.e. tracking fixed counters at base index 32 reduces the
+ * amounter of boilerplate needed to iterate over PMCs *and* simplifies common
+ * enabling/disable/reset operations.
+ *
+ * WARNING!  This helper is only for lookups that are initiated by KVM, it is
+ * NOT safe for guest lookups, e.g. will do the wrong thing if passed a raw
+ * ECX value from RDPMC (fixed counters are accessed by setting bit 30 in ECX
+ * for RDPMC, not by adding 32 to the fixed counter index).
+ */
+static inline struct kvm_pmc *kvm_pmc_idx_to_pmc(struct kvm_pmu *pmu, int idx)
+{
+	if (idx < pmu->nr_arch_gp_counters)
+		return &pmu->gp_counters[idx];
+
+	idx -= KVM_FIXED_PMC_BASE_IDX;
+	if (idx >= 0 && idx < pmu->nr_arch_fixed_counters)
+		return &pmu->fixed_counters[idx];
+
+	return NULL;
+}
+
 static inline u64 pmc_bitmask(struct kvm_pmc *pmc)
 {
 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
arch/x86/kvm/svm/pmu.c +3 -4
index 1fafc46f61c9..b6c1d1c3f204 100644
@@ -25,7 +25,7 @@ enum pmu_type {
 	PMU_TYPE_EVNTSEL,
 };
 
-static struct kvm_pmc *amd_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
+static struct kvm_pmc *amd_pmu_get_pmc(struct kvm_pmu *pmu, int pmc_idx)
 {
 	unsigned int num_counters = pmu->nr_arch_gp_counters;
 
@@ -70,7 +70,7 @@ static inline struct kvm_pmc *get_gp_pmc_amd(struct kvm_pmu *pmu, u32 msr,
 		return NULL;
 	}
 
-	return amd_pmc_idx_to_pmc(pmu, idx);
+	return amd_pmu_get_pmc(pmu, idx);
 }
 
 static bool amd_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
@@ -84,7 +84,7 @@ static bool amd_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
 static struct kvm_pmc *amd_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
 	unsigned int idx, u64 *mask)
 {
-	return amd_pmc_idx_to_pmc(vcpu_to_pmu(vcpu), idx);
+	return amd_pmu_get_pmc(vcpu_to_pmu(vcpu), idx);
 }
 
 static struct kvm_pmc *amd_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
@@ -226,7 +226,6 @@ static void amd_pmu_init(struct kvm_vcpu *vcpu)
 }
 
 struct kvm_pmu_ops amd_pmu_ops __initdata = {
-	.pmc_idx_to_pmc = amd_pmc_idx_to_pmc,
 	.rdpmc_ecx_to_pmc = amd_rdpmc_ecx_to_pmc,
 	.msr_idx_to_pmc = amd_msr_idx_to_pmc,
 	.is_valid_rdpmc_ecx = amd_is_valid_rdpmc_ecx,
arch/x86/kvm/vmx/pmu_intel.c +1 -14
index 61252bb733c4..4254411be467 100644
@@ -47,18 +47,6 @@ static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
 	}
 }
 
-static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
-{
-	if (pmc_idx < KVM_FIXED_PMC_BASE_IDX) {
-		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
-				  MSR_P6_EVNTSEL0);
-	} else {
-		u32 idx = pmc_idx - KVM_FIXED_PMC_BASE_IDX;
-
-		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
-	}
-}
-
 static u32 intel_rdpmc_get_masked_idx(struct kvm_pmu *pmu, u32 idx)
 {
 	/*
@@ -710,7 +698,7 @@ void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
 
 	for_each_set_bit(bit, (unsigned long *)&pmu->global_ctrl,
 			 X86_PMC_IDX_MAX) {
-		pmc = intel_pmc_idx_to_pmc(pmu, bit);
+		pmc = kvm_pmc_idx_to_pmc(pmu, bit);
 
 		if (!pmc || !pmc_speculative_in_use(pmc) ||
 		    !pmc_is_globally_enabled(pmc) || !pmc->perf_event)
@@ -727,7 +715,6 @@ void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
 }
 
 struct kvm_pmu_ops intel_pmu_ops __initdata = {
-	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#4 lore raw

PATCH 4/10 KVM: x86/pmu: Snapshot and clear reprogramming bitmap before reprogramming

In reply to Sean Christopherson

Refactor the handling of the reprogramming bitmap to snapshot and clear to-be-processed bits before doing the reprogramming, and then explicitly set bits for PMCs that need to be reprogrammed (again). This will allow adding a macro to iterate over all valid PMCs without having to add special handling for the reprogramming bit, which (a) can have bits set for non-existent PMCs and (b) needs to clear such bits to avoid wasting cycles in perpetuity.

Note, the existing behavior of clearing bits after reprogramming does NOT have a race with kvm_vm_ioctl_set_pmu_event_filter(). Setting a new PMU filter synchronizes SRCU _before_ setting the bitmap, i.e. guarantees that the vCPU isn't in the middle of reprogramming with a stale filter prior to setting the bitmap.

---
 arch/x86/include/asm/kvm_host.h |  1 +
 arch/x86/kvm/pmu.c              | 52 ++++++++++++++++++---------------
 2 files changed, 30 insertions(+), 23 deletions(-)
arch/x86/include/asm/kvm_host.h +1 -0
index d8bc9ba88cfc..22ba24d0fd4f 100644
@@ -535,6 +535,7 @@ struct kvm_pmc {
 #define KVM_PMC_MAX_FIXED	3
 #define MSR_ARCH_PERFMON_FIXED_CTR_MAX	(MSR_ARCH_PERFMON_FIXED_CTR0 + KVM_PMC_MAX_FIXED - 1)
 #define KVM_AMD_PMC_MAX_GENERIC	6
+
 struct kvm_pmu {
 	u8 version;
 	unsigned nr_arch_gp_counters;
arch/x86/kvm/pmu.c +29 -23

86 lines not on this page · raw message

Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#5 lore raw

PATCH 5/10 KVM: x86/pmu: Add macros to iterate over all PMCs given a bitmap

In reply to Sean Christopherson

Add and use kvm_for_each_pmc() to dedup a variety of open coded for-loops that iterate over valid PMCs given a bitmap (and because seeing checkpatch whine about bad macro style is always amusing).

No functional change intended.

---
 arch/x86/kvm/pmu.c           | 26 +++++++-------------------
 arch/x86/kvm/pmu.h           |  6 ++++++
 arch/x86/kvm/vmx/pmu_intel.c |  7 ++-----
 3 files changed, 15 insertions(+), 24 deletions(-)
arch/x86/kvm/pmu.c +7 -19
index ee921b24d9e4..0e2175170038 100644
@@ -493,6 +493,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 {
 	DECLARE_BITMAP(bitmap, X86_PMC_IDX_MAX);
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
+	struct kvm_pmc *pmc;
 	int bit;
 
 	bitmap_copy(bitmap, pmu->reprogram_pmi, X86_PMC_IDX_MAX);
@@ -505,12 +506,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 	BUILD_BUG_ON(sizeof(bitmap) != sizeof(atomic64_t));
 	atomic64_andnot(*(s64 *)bitmap, &pmu->__reprogram_pmi);
 
-	for_each_set_bit(bit, bitmap, X86_PMC_IDX_MAX) {
-		struct kvm_pmc *pmc = kvm_pmc_idx_to_pmc(pmu, bit);
-
-		if (unlikely(!pmc))
-			continue;
-
+	kvm_for_each_pmc(pmu, pmc, bit, bitmap) {
 		/*
 		 * If reprogramming fails, e.g. due to contention, re-set the
 		 * regprogram bit set, i.e. opportunistically try again on the
@@ -720,11 +716,7 @@ static void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 
 	bitmap_zero(pmu->reprogram_pmi, X86_PMC_IDX_MAX);
 
-	for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
-		pmc = kvm_pmc_idx_to_pmc(pmu, i);
-		if (!pmc)
-			continue;
-
+	kvm_for_each_pmc(pmu, pmc, i, pmu->all_valid_pmc_idx) {
 		pmc_stop_counter(pmc);
 		pmc->counter = 0;
 		pmc->emulated_counter = 0;
@@ -796,10 +788,8 @@ void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
 	bitmap_andnot(bitmask, pmu->all_valid_pmc_idx,
 		      pmu->pmc_in_use, X86_PMC_IDX_MAX);
 
-	for_each_set_bit(i, bitmask, X86_PMC_IDX_MAX) {
-		pmc = kvm_pmc_idx_to_pmc(pmu, i);
-
-		if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
+	kvm_for_each_pmc(pmu, pmc, i, bitmask) {
+		if (pmc->perf_event && !pmc_speculative_in_use(pmc))
 			pmc_stop_counter(pmc);
 	}
 
@@ -851,10 +841,8 @@ void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
 	struct kvm_pmc *pmc;
 	int i;
 
-	for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
-		pmc = kvm_pmc_idx_to_pmc(pmu, i);
-
-		if (!pmc || !pmc_event_is_allowed(pmc))
+	kvm_for_each_pmc(pmu, pmc, i, pmu->all_valid_pmc_idx) {
+		if (!pmc_event_is_allowed(pmc))
 			continue;
 
 		/* Ignore checks for edge detect, pin control, invert and CMASK bits */
arch/x86/kvm/pmu.h +6 -0
index 2235772a495b..cb62a4e44849 100644
@@ -83,6 +83,12 @@ static inline struct kvm_pmc *kvm_pmc_idx_to_pmc(struct kvm_pmu *pmu, int idx)
 	return NULL;
 }
 
+#define kvm_for_each_pmc(pmu, pmc, i, bitmap)			\
+	for_each_set_bit(i, bitmap, X86_PMC_IDX_MAX)		\
+		if (!(pmc = kvm_pmc_idx_to_pmc(pmu, i)))	\
+			continue;				\
+		else						\
+
 static inline u64 pmc_bitmask(struct kvm_pmc *pmc)
 {
 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
arch/x86/kvm/vmx/pmu_intel.c +2 -5
index 4254411be467..ee3e122d3617 100644
@@ -696,11 +696,8 @@ void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
 	struct kvm_pmc *pmc = NULL;
 	int bit, hw_idx;
 
-	for_each_set_bit(bit, (unsigned long *)&pmu->global_ctrl,
-			 X86_PMC_IDX_MAX) {
-		pmc = kvm_pmc_idx_to_pmc(pmu, bit);
-
-		if (!pmc || !pmc_speculative_in_use(pmc) ||
+	kvm_for_each_pmc(pmu, pmc, bit, (unsigned long *)&pmu->global_ctrl) {
+		if (!pmc_speculative_in_use(pmc) ||
 		    !pmc_is_globally_enabled(pmc) || !pmc->perf_event)
 			continue;
 
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#6 lore raw

PATCH 6/10 KVM: x86/pmu: Process only enabled PMCs when emulating events in software

In reply to Sean Christopherson

Mask off disabled counters based on PERF_GLOBAL_CTRL *before* iterating over PMCs to emulate (branch) instruction required events in software. In the common case where the guest isn't utilizing the PMU, pre-checking for enabled counters turns a relatively expensive search into a few AND uops and a Jcc.

Sadly, PMUs without PERF_GLOBAL_CTRL, e.g. most existing AMD CPUs, are out of luck as there is no way to check that a PMC isn't being used without checking the PMC's event selector.

Cc Konstantin Khorenko <[email protected]>
---
 arch/x86/kvm/pmu.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
arch/x86/kvm/pmu.c +10 -1
index 0e2175170038..488d21024a92 100644
@@ -837,11 +837,20 @@ static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 
 void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
 {
+	DECLARE_BITMAP(bitmap, X86_PMC_IDX_MAX);
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 	struct kvm_pmc *pmc;
 	int i;
 
-	kvm_for_each_pmc(pmu, pmc, i, pmu->all_valid_pmc_idx) {
+	BUILD_BUG_ON(sizeof(pmu->global_ctrl) * BITS_PER_BYTE != X86_PMC_IDX_MAX);
+
+	if (!kvm_pmu_has_perf_global_ctrl(pmu))
+		bitmap_copy(bitmap, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX);
+	else if (!bitmap_and(bitmap, pmu->all_valid_pmc_idx,
+			     (unsigned long *)&pmu->global_ctrl, X86_PMC_IDX_MAX))
+		return;
+
+	kvm_for_each_pmc(pmu, pmc, i, bitmap) {
 		if (!pmc_event_is_allowed(pmc))
 			continue;
 
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#7 lore raw

PATCH 7/10 KVM: x86/pmu: Snapshot event selectors that KVM emulates in software

In reply to Sean Christopherson

Snapshot the event selectors for the events that KVM emulates in software, which is currently instructions retired and branch instructions retired. The event selectors a tied to the underlying CPU, i.e. are constant for a given platform even though perf doesn't manage the mappings as such.

Getting the event selectors from perf isn't exactly cheap, especially if mitigations are enabled, as at least one indirect call is involved.

Snapshot the values in KVM instead of optimizing perf as working with the raw event selectors will be required if KVM ever wants to emulate events that aren't part of perf's uABI, i.e. that don't have an "enum perf_hw_id" entry.

---
 arch/x86/kvm/pmu.c        | 17 ++++++++---------
 arch/x86/kvm/pmu.h        | 13 ++++++++++++-
 arch/x86/kvm/vmx/nested.c |  2 +-
 arch/x86/kvm/x86.c        |  6 +++---
 4 files changed, 24 insertions(+), 14 deletions(-)
arch/x86/kvm/pmu.c +8 -9
index 488d21024a92..45cb8b2a024b 100644
@@ -29,6 +29,9 @@
 struct x86_pmu_capability __read_mostly kvm_pmu_cap;
 EXPORT_SYMBOL_GPL(kvm_pmu_cap);
 
+struct kvm_pmu_emulated_event_selectors __read_mostly kvm_pmu_eventsel;
+EXPORT_SYMBOL_GPL(kvm_pmu_eventsel);
+
 /* Precise Distribution of Instructions Retired (PDIR) */
 static const struct x86_cpu_id vmx_pebs_pdir_cpu[] = {
 	X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, NULL),
@@ -809,13 +812,6 @@ static void kvm_pmu_incr_counter(struct kvm_pmc *pmc)
 	kvm_pmu_request_counter_reprogram(pmc);
 }
 
-static inline bool eventsel_match_perf_hw_id(struct kvm_pmc *pmc,
-	unsigned int perf_hw_id)
-{
-	return !((pmc->eventsel ^ perf_get_hw_event_config(perf_hw_id)) &
-		AMD64_RAW_EVENT_MASK_NB);
-}
-
 static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 {
 	bool select_os, select_user;
@@ -835,7 +831,7 @@ static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 	return (static_call(kvm_x86_get_cpl)(pmc->vcpu) == 0) ? select_os : select_user;
 }
 
-void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
+void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 eventsel)
 {
 	DECLARE_BITMAP(bitmap, X86_PMC_IDX_MAX);
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
@@ -855,7 +851,10 @@ void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
 			continue;
 
 		/* Ignore checks for edge detect, pin control, invert and CMASK bits */
-		if (eventsel_match_perf_hw_id(pmc, perf_hw_id) && cpl_is_matched(pmc))
+		if ((pmc->eventsel ^ eventsel) & AMD64_RAW_EVENT_MASK_NB)
+			continue;
+
+		if (cpl_is_matched(pmc))
 			kvm_pmu_incr_counter(pmc);
 	}
 }
arch/x86/kvm/pmu.h +12 -1
index cb62a4e44849..9dc5f549c98c 100644
@@ -22,6 +22,11 @@
 
 #define KVM_FIXED_PMC_BASE_IDX INTEL_PMC_IDX_FIXED
 
+struct kvm_pmu_emulated_event_selectors {
+	u64 INSTRUCTIONS_RETIRED;
+	u64 BRANCH_INSTRUCTIONS_RETIRED;
+};
+
 struct kvm_pmu_ops {
 	struct kvm_pmc *(*rdpmc_ecx_to_pmc)(struct kvm_vcpu *vcpu,
 		unsigned int idx, u64 *mask);
@@ -171,6 +176,7 @@ static inline bool pmc_speculative_in_use(struct kvm_pmc *pmc)
 }
 
 extern struct x86_pmu_capability kvm_pmu_cap;
+extern struct kvm_pmu_emulated_event_selectors kvm_pmu_eventsel;
 
 static inline void kvm_init_pmu_capability(const struct kvm_pmu_ops *pmu_ops)
 {
@@ -212,6 +218,11 @@ static inline void kvm_init_pmu_capability(const struct kvm_pmu_ops *pmu_ops)
 					  pmu_ops->MAX_NR_GP_COUNTERS);
 	kvm_pmu_cap.num_counters_fixed = min(kvm_pmu_cap.num_counters_fixed,
 					     KVM_PMC_MAX_FIXED);
+
+	kvm_pmu_eventsel.INSTRUCTIONS_RETIRED =
+		perf_get_hw_event_config(PERF_COUNT_HW_INSTRUCTIONS);
+	kvm_pmu_eventsel.BRANCH_INSTRUCTIONS_RETIRED =
+		perf_get_hw_event_config(PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
 }
 
 static inline void kvm_pmu_request_counter_reprogram(struct kvm_pmc *pmc)
@@ -259,7 +270,7 @@ void kvm_pmu_init(struct kvm_vcpu *vcpu);
 void kvm_pmu_cleanup(struct kvm_vcpu *vcpu);
 void kvm_pmu_destroy(struct kvm_vcpu *vcpu);
 int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp);
-void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id);
+void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 eventsel);
 
 bool is_vmware_backdoor_pmc(u32 pmc_idx);
 
arch/x86/kvm/vmx/nested.c +1 -1
index c5ec0ef51ff7..cf985085467b 100644
@@ -3564,7 +3564,7 @@ static int nested_vmx_run(struct kvm_vcpu *vcpu, bool launch)
 		return 1;
 	}
 
-	kvm_pmu_trigger_event(vcpu, PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
+	kvm_pmu_trigger_event(vcpu, kvm_pmu_eventsel.BRANCH_INSTRUCTIONS_RETIRED);
 
 	if (CC(evmptrld_status == EVMPTRLD_VMFAIL))
 		return nested_vmx_failInvalid(vcpu);
arch/x86/kvm/x86.c +3 -3
index efbf52a9dc83..9d9b5f9e4b28 100644
@@ -8839,7 +8839,7 @@ int kvm_skip_emulated_instruction(struct kvm_vcpu *vcpu)
 	if (unlikely(!r))
 		return 0;
 
-	kvm_pmu_trigger_event(vcpu, PERF_COUNT_HW_INSTRUCTIONS);
+	kvm_pmu_trigger_event(vcpu, kvm_pmu_eventsel.INSTRUCTIONS_RETIRED);
 
 	/*
 	 * rflags is the old, "raw" value of the flags.  The new value has
@@ -9152,9 +9152,9 @@ int x86_emulate_instruction(struct kvm_vcpu *vcpu, gpa_t cr2_or_gpa,
 		 */
 		if (!ctxt->have_exception ||
 		    exception_type(ctxt->exception.vector) == EXCPT_TRAP) {
-			kvm_pmu_trigger_event(vcpu, PERF_COUNT_HW_INSTRUCTIONS);
+			kvm_pmu_trigger_event(vcpu, kvm_pmu_eventsel.INSTRUCTIONS_RETIRED);
 			if (ctxt->is_branch)
-				kvm_pmu_trigger_event(vcpu, PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
+				kvm_pmu_trigger_event(vcpu, kvm_pmu_eventsel.BRANCH_INSTRUCTIONS_RETIRED);
 			kvm_rip_write(vcpu, ctxt->eip);
 			if (r && (ctxt->tf || (vcpu->guest_debug & KVM_GUESTDBG_SINGLESTEP)))
 				r = kvm_vcpu_do_singlestep(vcpu);
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#8 lore raw

PATCH 8/10 KVM: x86/pmu: Expand the comment about what bits are check emulating events

In reply to Sean Christopherson

Expand the comment about what bits are and aren't checked when emulating PMC events in software. As pointed out by Jim, AMD's mask includes bits 35:32, which on Intel overlap with the IN_TX and IN_TXCP bits (32 and 33) as well as reserved bits (34 and 45).

Checking The IN_TX* bits is actually correct, as it's safe to assert that the vCPU can't be in an HLE/RTM transaction if KVM is emulating an instruction, i.e. KVM *shouldn't count if either of those bits is set.

For the reserved bits, KVM is has equal odds of being right if Intel adds new behavior, i.e. ignoring them is just as likely to be correct as checking them.

Opportunistically explain *why* the other flags aren't checked.

---
 arch/x86/kvm/pmu.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)
arch/x86/kvm/pmu.c +14 -1
index 45cb8b2a024b..ba561849fd3e 100644
@@ -850,7 +850,20 @@ void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 eventsel)
 		if (!pmc_event_is_allowed(pmc))
 			continue;
 
-		/* Ignore checks for edge detect, pin control, invert and CMASK bits */
+		/*
+		 * Ignore checks for edge detect (all events currently emulated
+		 * but KVM are always rising edges), pin control (unsupported
+		 * by modern CPUs), and counter mask and its invert flag (KVM
+		 * doesn't emulate multiple events in a single clock cycle).
+		 *
+		 * Note, the uppermost nibble of AMD's mask overlaps Intel's
+		 * IN_TX (bit 32) and IN_TXCP (bit 33), as well as two reserved
+		 * bits (bits 35:34).  Checking the "in HLE/RTM transaction"
+		 * flags is correct as the vCPU can't be in a transaction if
+		 * KVM is emulating an instruction.  Checking the reserved bits
+		 * might be wrong if they are defined in the future, but so
+		 * could ignoring them, so do the simple thing for now.
+		 */
 		if ((pmc->eventsel ^ eventsel) & AMD64_RAW_EVENT_MASK_NB)
 			continue;
 
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#9 lore raw

PATCH 9/10 KVM: x86/pmu: Check eventsel first when emulating (branch) insns retired

In reply to Sean Christopherson

When triggering events, i.e. emulating PMC events in software, check for a matching event selector before checking the event is allowed. The "is allowed" check *might* be cheap, but it could also be very costly, e.g. if userspace has defined a large PMU event filter. The event selector check on the other hand is all but guaranteed to be <10 uops, e.g. looks something like:

0xffffffff8105e615 <+5>: movabs $0xf0000ffff,%rax

   0xffffffff8105e61f <+15>:	xor    %rdi,%rsi
   0xffffffff8105e622 <+18>:	test   %rax,%rsi
   0xffffffff8105e625 <+21>:	sete   %al
---
 arch/x86/kvm/pmu.c | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)
arch/x86/kvm/pmu.c +3 -6
index ba561849fd3e..a5ea729b16f2 100644
@@ -847,9 +847,6 @@ void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 eventsel)
 		return;
 
 	kvm_for_each_pmc(pmu, pmc, i, bitmap) {
-		if (!pmc_event_is_allowed(pmc))
-			continue;
-
 		/*
 		 * Ignore checks for edge detect (all events currently emulated
 		 * but KVM are always rising edges), pin control (unsupported
@@ -864,11 +861,11 @@ void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 eventsel)
 		 * might be wrong if they are defined in the future, but so
 		 * could ignoring them, so do the simple thing for now.
 		 */
-		if ((pmc->eventsel ^ eventsel) & AMD64_RAW_EVENT_MASK_NB)
+		if (((pmc->eventsel ^ eventsel) & AMD64_RAW_EVENT_MASK_NB) ||
+		    !pmc_event_is_allowed(pmc) || !cpl_is_matched(pmc))
 			continue;
 
-		if (cpl_is_matched(pmc))
-			kvm_pmu_incr_counter(pmc);
+		kvm_pmu_incr_counter(pmc);
 	}
 }
 EXPORT_SYMBOL_GPL(kvm_pmu_trigger_event);
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#10 lore raw

PATCH 10/10 KVM: x86/pmu: Avoid CPL lookup if PMC enabline for USER and KERNEL is the same

In reply to Sean Christopherson

Don't bother querying the CPL if a PMC is (not) counting for both USER and KERNEL, i.e. if the end result is guaranteed to be the same regardless of the CPL. Querying the CPL on Intel requires a VMREAD, i.e. isn't free, and a single CMP+Jcc is cheap.

---
 arch/x86/kvm/pmu.c | 7 +++++++
 1 file changed, 7 insertions(+)
arch/x86/kvm/pmu.c +7 -0
index a5ea729b16f2..231604d6dc86 100644
@@ -828,6 +828,13 @@ static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 		select_user = config & 0x2;
 	}
 
+	/*
+	 * Skip the CPL lookup, which isn't free on Intel, if the result will
+	 * be the same regardless of the CPL.
+	 */
+	if (select_os == select_user)
+		return select_os;
+
 	return (static_call(kvm_x86_get_cpl)(pmc->vcpu) == 0) ? select_os : select_user;
 }
 
Signature · 1 line
2.42.0.869.gea05f2083d-goog
author maintainer
#12 lore raw
Quoted 11 lines · Dongli Zhang

Hi Sean,

Quoted 2 lines · Sean Christopherson

base-commit: ef1883475d4a24d8eaebb84175ed46206a688103

May I have a silly question?

May I have the tree that the commit is the base? I do not find it in kvm-x86.

https://github.com/kvm-x86/linux/commit/ef1883475d4a24d8eaebb84175ed46206a688103

It's kvm-x86/next (which I failed to explicitly call out), plus two PMU series as mentioned in the cover letter.

https://lore.kernel.org/all/[email protected]
https://lore.kernel.org/all/[email protected]

#13 lore raw
Quoted 7 lines · Sean Christopherson
Quoted 6 lines · Dongli Zhang

Hi Sean,

Quoted 1 line · Sean Christopherson

base-commit: ef1883475d4a24d8eaebb84175ed46206a688103

May I have a silly question?

Quoted 8 lines
Quoted 3 lines

May I have the tree that the commit is the base? I do not find it in kvm-x86.

https://github.com/kvm-x86/linux/commit/ef1883475d4a24d8eaebb84175ed46206a688103

It's kvm-x86/next (which I failed to explicitly call out), plus two PMU series as mentioned in the cover letter.

   https://lore.kernel.org/all/[email protected]
   https://lore.kernel.org/all/[email protected]

Hi Sean,

i have tested my testcase on your patchset, results look impressive, thank you for your work! Note, for AMD node PMU totally disabled case gave a bit worse results on kvm-next kernel vs stock latest mainstream.

The difference is tiny though.

*AMD node*: CPU: AMD Zen 3 (three!): AMD EPYC 7443P 24-Core Processor
-----------------------------------------------------------------------------------------
| Kernel | CPUID rate |
-----------------------------------------------------------------------------------------

| stock ms 6.6.0+ (commit 305230142ae0) 			| 1360250		|
| stock ms 6.6.0+ + kvm.enable_pmu=0				| 1542894 (+13.4%)	|
| kvm-next ms 6.6.0+ + patches from Sean + kvm.enable_pmu=1	| 1498864		|
| kvm-next ms 6.6.0+ + patches from Sean + kvm.enable_pmu=0	| 1526396 (+1.84%)	|

-----------------------------------------------------------------------------------------

*Intel node*: CPU: Intel(R) Xeon(R) E-2136 CPU @ 3.30GHz
-----------------------------------------------------------------------------------------
| Kernel | CPUID rate |
-----------------------------------------------------------------------------------------

| stock ms 6.6.0+						| 1431608		|
| stock ms 6.6.0+ + kvm.enable_pmu=0				| 1553839 (+8.5%)	|
| kvm-next ms 6.6.0+ + patches from Sean + kvm.enable_pmu=1	| 1559365		|
| kvm-next ms 6.6.0+ + patches from Sean + kvm.enable_pmu=0	| 1582281 (+1.5%)	|

-----------------------------------------------------------------------------------------

Note: in order to disable PMU completely i used "kvm" module "enable_pmu=0" option (not used KVM_PMU_CAP_DISABLEthis time).

Hope that helps.

--
Best regards,

Konstantin Khorenko,
Virtuozzo Linux Kernel Team

author maintainer Applied
#14 lore raw
In reply to Sean Christopherson
Quoted 5 lines · Sean Christopherson

Optimize code used by, or which impacts, kvm_pmu_trigger_event() to try and make a dent in the overhead of emulating PMU events in software, which is quite noticeable due to it kicking in anytime the guest has a vPMU and KVM is skipping an instruction.

Quoted 7 lines

Note, Jim has a proposal/idea[*] (that I supported) to make kvm_pmu_trigger_event() even more performant. I opted not to do that as it's a bit more invasive, and I started chewing on this not so much because I care _that_ much about performance, but because it irritates me that the PMU code makes things way harder than they need to be.

[...]

Applied to kvm-x86 pmu, thanks!

[01/10] KVM: x86/pmu: Zero out PMU metadata on AMD if PMU is disabled

        https://github.com/kvm-x86/linux/commit/f933b88e2015
[02/10] KVM: x86/pmu: Add common define to capture fixed counters offset
        https://github.com/kvm-x86/linux/commit/be6b067dae15
[03/10] KVM: x86/pmu: Move pmc_idx => pmc translation helper to common code
        https://github.com/kvm-x86/linux/commit/b31880ca2f41
[04/10] KVM: x86/pmu: Snapshot and clear reprogramming bitmap before reprogramming
        https://github.com/kvm-x86/linux/commit/004a0aa56ede
[05/10] KVM: x86/pmu: Add macros to iterate over all PMCs given a bitmap
        https://github.com/kvm-x86/linux/commit/e5a65d4f723a
[06/10] KVM: x86/pmu: Process only enabled PMCs when emulating events in software
        https://github.com/kvm-x86/linux/commit/d2b321ea9380
[07/10] KVM: x86/pmu: Snapshot event selectors that KVM emulates in software
        https://github.com/kvm-x86/linux/commit/f19063b1ca05
[08/10] KVM: x86/pmu: Expand the comment about what bits are check emulating events
        https://github.com/kvm-x86/linux/commit/afda2d7666f8
[09/10] KVM: x86/pmu: Check eventsel first when emulating (branch) insns retired
        https://github.com/kvm-x86/linux/commit/e35529fb4ac9
[10/10] KVM: x86/pmu: Avoid CPL lookup if PMC enabline for USER and KERNEL is the same
        https://github.com/kvm-x86/linux/commit/83bdfe04c968

--
https://github.com/kvm-x86/linux/tree/next