[PATCH 0/3] Use static_call for kvm_pmu_ops

[PATCH 0/3] Use static_call for kvm_pmu_ops

[Like Xu]

Hi,

This is a successor to a previous patch set from Jason Baron. Let's convert
kvm_pmu_ops to use static_call. Shows good performance gains for
a typical perf use case [2] in the guest (results in patch 3/3).

[1] https://lore.kernel.org/lkml/cover.1610680941.git.jbaron@akamai.com/
[2] perf record -e branch-instructions -e branch-misses \
-e cache-misses -e cache-references -e cpu-cycles \
-e instructions ./workload

Thanks,

Like Xu (3):
  KVM: x86: Copy kvm_pmu_ops by value to eliminate layer of indirection
  KVM: x86: Introduce definitions to support static calls for
    kvm_pmu_ops
  KVM: x86: Use static calls to reduce kvm_pmu_ops overhead

 arch/x86/include/asm/kvm-x86-pmu-ops.h | 32 ++++++++++++++++++
 arch/x86/kvm/pmu.c                     | 46 +++++++++++++++-----------
 arch/x86/kvm/pmu.h                     | 19 ++++++++++-
 arch/x86/kvm/vmx/nested.c              |  2 +-
 arch/x86/kvm/x86.c                     |  5 +++
 5 files changed, 83 insertions(+), 21 deletions(-)
 create mode 100644 arch/x86/include/asm/kvm-x86-pmu-ops.h

-- 
2.33.0

[PATCH 1/3] KVM: x86: Copy kvm_pmu_ops by value to eliminate layer of indirection

[Like Xu]

Replace the kvm_pmu_ops pointer in common x86 with an instance of the
struct to save one pointer dereference when invoking functions. Copy the
struct by value to set the ops during kvm_init().

Using kvm_x86_ops.hardware_enable to track whether or not the
ops have been initialized, i.e. a vendor KVM module has been loaded.

Signed-off-by: Like Xu <likexu@tencent.com>
---
 arch/x86/kvm/pmu.c        | 41 +++++++++++++++++++++------------------
 arch/x86/kvm/pmu.h        |  4 +++-
 arch/x86/kvm/vmx/nested.c |  2 +-
 arch/x86/kvm/x86.c        |  3 +++
 4 files changed, 29 insertions(+), 21 deletions(-)

diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
index 0772bad9165c..0db1887137d9 100644
--- a/arch/x86/kvm/pmu.c
+++ b/arch/x86/kvm/pmu.c
@@ -47,6 +47,9 @@
  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] <=> gp counters
  */
 
+struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
+EXPORT_SYMBOL_GPL(kvm_pmu_ops);
+
 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
 {
 	struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
@@ -214,7 +217,7 @@ void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
 			  ARCH_PERFMON_EVENTSEL_CMASK |
 			  HSW_IN_TX |
 			  HSW_IN_TX_CHECKPOINTED))) {
-		config = kvm_x86_ops.pmu_ops->find_arch_event(pmc_to_pmu(pmc),
+		config = kvm_pmu_ops.find_arch_event(pmc_to_pmu(pmc),
 						      event_select,
 						      unit_mask);
 		if (config != PERF_COUNT_HW_MAX)
@@ -268,7 +271,7 @@ void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
 
 	pmc->current_config = (u64)ctrl;
 	pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
-			      kvm_x86_ops.pmu_ops->find_fixed_event(idx),
+			      kvm_pmu_ops.find_fixed_event(idx),
 			      !(en_field & 0x2), /* exclude user */
 			      !(en_field & 0x1), /* exclude kernel */
 			      pmi, false, false);
@@ -277,7 +280,7 @@ EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
 
 void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
 {
-	struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, pmc_idx);
+	struct kvm_pmc *pmc = kvm_pmu_ops.pmc_idx_to_pmc(pmu, pmc_idx);
 
 	if (!pmc)
 		return;
@@ -299,7 +302,7 @@ 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 = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, bit);
+		struct kvm_pmc *pmc = kvm_pmu_ops.pmc_idx_to_pmc(pmu, bit);
 
 		if (unlikely(!pmc || !pmc->perf_event)) {
 			clear_bit(bit, pmu->reprogram_pmi);
@@ -321,7 +324,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 /* check if idx is a valid index to access PMU */
 int kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
 {
-	return kvm_x86_ops.pmu_ops->is_valid_rdpmc_ecx(vcpu, idx);
+	return kvm_pmu_ops.is_valid_rdpmc_ecx(vcpu, idx);
 }
 
 bool is_vmware_backdoor_pmc(u32 pmc_idx)
@@ -371,7 +374,7 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 	if (is_vmware_backdoor_pmc(idx))
 		return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
 
-	pmc = kvm_x86_ops.pmu_ops->rdpmc_ecx_to_pmc(vcpu, idx, &mask);
+	pmc = kvm_pmu_ops.rdpmc_ecx_to_pmc(vcpu, idx, &mask);
 	if (!pmc)
 		return 1;
 
@@ -387,22 +390,22 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
 {
 	if (lapic_in_kernel(vcpu)) {
-		if (kvm_x86_ops.pmu_ops->deliver_pmi)
-			kvm_x86_ops.pmu_ops->deliver_pmi(vcpu);
+		if (kvm_pmu_ops.deliver_pmi)
+			kvm_pmu_ops.deliver_pmi(vcpu);
 		kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
 	}
 }
 
 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
 {
-	return kvm_x86_ops.pmu_ops->msr_idx_to_pmc(vcpu, msr) ||
-		kvm_x86_ops.pmu_ops->is_valid_msr(vcpu, msr);
+	return kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr) ||
+		kvm_pmu_ops.is_valid_msr(vcpu, msr);
 }
 
 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 {
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
-	struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->msr_idx_to_pmc(vcpu, msr);
+	struct kvm_pmc *pmc = kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr);
 
 	if (pmc)
 		__set_bit(pmc->idx, pmu->pmc_in_use);
@@ -410,13 +413,13 @@ static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 
 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 {
-	return kvm_x86_ops.pmu_ops->get_msr(vcpu, msr_info);
+	return kvm_pmu_ops.get_msr(vcpu, msr_info);
 }
 
 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 {
 	kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
-	return kvm_x86_ops.pmu_ops->set_msr(vcpu, msr_info);
+	return kvm_pmu_ops.set_msr(vcpu, msr_info);
 }
 
 /* refresh PMU settings. This function generally is called when underlying
@@ -425,7 +428,7 @@ int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
  */
 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 {
-	kvm_x86_ops.pmu_ops->refresh(vcpu);
+	kvm_pmu_ops.refresh(vcpu);
 }
 
 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
@@ -433,7 +436,7 @@ void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 
 	irq_work_sync(&pmu->irq_work);
-	kvm_x86_ops.pmu_ops->reset(vcpu);
+	kvm_pmu_ops.reset(vcpu);
 }
 
 void kvm_pmu_init(struct kvm_vcpu *vcpu)
@@ -441,7 +444,7 @@ void kvm_pmu_init(struct kvm_vcpu *vcpu)
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 
 	memset(pmu, 0, sizeof(*pmu));
-	kvm_x86_ops.pmu_ops->init(vcpu);
+	kvm_pmu_ops.init(vcpu);
 	init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
 	pmu->event_count = 0;
 	pmu->need_cleanup = false;
@@ -473,14 +476,14 @@ 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 = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, i);
+		pmc = kvm_pmu_ops.pmc_idx_to_pmc(pmu, i);
 
 		if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
 			pmc_stop_counter(pmc);
 	}
 
-	if (kvm_x86_ops.pmu_ops->cleanup)
-		kvm_x86_ops.pmu_ops->cleanup(vcpu);
+	if (kvm_pmu_ops.cleanup)
+		kvm_pmu_ops.cleanup(vcpu);
 
 	bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
 }
diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
index 0e4f2b1fa9fb..b2fe135d395a 100644
--- a/arch/x86/kvm/pmu.h
+++ b/arch/x86/kvm/pmu.h
@@ -17,6 +17,8 @@
 
 #define MAX_FIXED_COUNTERS	3
 
+extern struct kvm_pmu_ops kvm_pmu_ops;
+
 struct kvm_event_hw_type_mapping {
 	u8 eventsel;
 	u8 unit_mask;
@@ -92,7 +94,7 @@ static inline bool pmc_is_fixed(struct kvm_pmc *pmc)
 
 static inline bool pmc_is_enabled(struct kvm_pmc *pmc)
 {
-	return kvm_x86_ops.pmu_ops->pmc_is_enabled(pmc);
+	return kvm_pmu_ops.pmc_is_enabled(pmc);
 }
 
 static inline bool kvm_valid_perf_global_ctrl(struct kvm_pmu *pmu,
diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index b4ee5e9f9e20..1e793e44b5ff 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -4796,7 +4796,7 @@ void nested_vmx_pmu_entry_exit_ctls_update(struct kvm_vcpu *vcpu)
 		return;
 
 	vmx = to_vmx(vcpu);
-	if (kvm_x86_ops.pmu_ops->is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
+	if (kvm_pmu_ops.is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
 		vmx->nested.msrs.entry_ctls_high |=
 				VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL;
 		vmx->nested.msrs.exit_ctls_high |=
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index ac83d873d65b..72d286595012 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -11317,6 +11317,9 @@ int kvm_arch_hardware_setup(void *opaque)
 	memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
 	kvm_ops_static_call_update();
 
+	if (kvm_x86_ops.hardware_enable)
+		memcpy(&kvm_pmu_ops, kvm_x86_ops.pmu_ops, sizeof(kvm_pmu_ops));
+
 	if (!kvm_cpu_cap_has(X86_FEATURE_XSAVES))
 		supported_xss = 0;
 
-- 
2.33.0

Re: [PATCH 1/3] KVM: x86: Copy kvm_pmu_ops by value to eliminate layer of indirection

[Sean Christopherson]

On Wed, Nov 03, 2021, Like Xu wrote:
> Replace the kvm_pmu_ops pointer in common x86 with an instance of the
> struct to save one pointer dereference when invoking functions. Copy the
> struct by value to set the ops during kvm_init().
> 
> Using kvm_x86_ops.hardware_enable to track whether or not the
> ops have been initialized, i.e. a vendor KVM module has been loaded.
> 
> Signed-off-by: Like Xu <likexu@tencent.com>
> ---
>  arch/x86/kvm/pmu.c        | 41 +++++++++++++++++++++------------------
>  arch/x86/kvm/pmu.h        |  4 +++-
>  arch/x86/kvm/vmx/nested.c |  2 +-
>  arch/x86/kvm/x86.c        |  3 +++
>  4 files changed, 29 insertions(+), 21 deletions(-)
> 
> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> index 0772bad9165c..0db1887137d9 100644
> --- a/arch/x86/kvm/pmu.c
> +++ b/arch/x86/kvm/pmu.c
> @@ -47,6 +47,9 @@
>   *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] <=> gp counters
>   */
>  
> +struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
> +EXPORT_SYMBOL_GPL(kvm_pmu_ops);
> +

...

> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> index b4ee5e9f9e20..1e793e44b5ff 100644
> --- a/arch/x86/kvm/vmx/nested.c
> +++ b/arch/x86/kvm/vmx/nested.c
> @@ -4796,7 +4796,7 @@ void nested_vmx_pmu_entry_exit_ctls_update(struct kvm_vcpu *vcpu)
>  		return;
>  
>  	vmx = to_vmx(vcpu);
> -	if (kvm_x86_ops.pmu_ops->is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
> +	if (kvm_pmu_ops.is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {

I would much prefer we export kvm_pmu_is_valid_msr() and go through that for nVMX
than export all of kvm_pmu_ops for this one case.

>  		vmx->nested.msrs.entry_ctls_high |=
>  				VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL;
>  		vmx->nested.msrs.exit_ctls_high |=
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index ac83d873d65b..72d286595012 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -11317,6 +11317,9 @@ int kvm_arch_hardware_setup(void *opaque)
>  	memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
>  	kvm_ops_static_call_update();
>  
> +	if (kvm_x86_ops.hardware_enable)

Huh?  Did you intend this to be?

	if (kvm_x86_ops.pmu_ops)

Either way, I don't see the point, VMX and SVM unconditionally provide the ops.

I would also say land this memcpy() above kvm_ops_static_call_update(), then the
enabling patch can do the static call updates in kvm_ops_static_call_update()
instead of adding another helper.

> +		memcpy(&kvm_pmu_ops, kvm_x86_ops.pmu_ops, sizeof(kvm_pmu_ops));

As part of this change, the pmu_ops should be moved to kvm_x86_init_ops and tagged
as __initdata.  That'll save those precious few bytes, and more importantly make
the original ops unreachable, i.e. make it harder to sneak in post-init modification
bugs.

> +
>  	if (!kvm_cpu_cap_has(X86_FEATURE_XSAVES))
>  		supported_xss = 0;
>  
> -- 
> 2.33.0
> 

Re: [PATCH 1/3] KVM: x86: Copy kvm_pmu_ops by value to eliminate layer of indirection

[Sean Christopherson]

On Fri, Nov 05, 2021, Sean Christopherson wrote:
> On Wed, Nov 03, 2021, Like Xu wrote:
> I would also say land this memcpy() above kvm_ops_static_call_update(), then the
> enabling patch can do the static call updates in kvm_ops_static_call_update()
> instead of adding another helper.

Ugh, kvm_ops_static_call_update() is defined in kvm_host.h.  That's completely
unnecessary, it should have exactly one caller, kvm_arch_hardware_setup().  As a
prep match, move kvm_ops_static_call_update() to x86.c, then it can reference the
pmu ops.

Re: [PATCH 1/3] KVM: x86: Copy kvm_pmu_ops by value to eliminate layer of indirection

[Like Xu]

On 5/11/2021 11:36 pm, Sean Christopherson wrote:
> On Fri, Nov 05, 2021, Sean Christopherson wrote:
>> On Wed, Nov 03, 2021, Like Xu wrote:
>> I would also say land this memcpy() above kvm_ops_static_call_update(), then the
>> enabling patch can do the static call updates in kvm_ops_static_call_update()
>> instead of adding another helper.
> 
> Ugh, kvm_ops_static_call_update() is defined in kvm_host.h.  That's completely
> unnecessary, it should have exactly one caller, kvm_arch_hardware_setup().  As a
> prep match, move kvm_ops_static_call_update() to x86.c, then it can reference the
> pmu ops.

Quite good and thank you.

Re: [PATCH 1/3] KVM: x86: Copy kvm_pmu_ops by value to eliminate layer of indirection

[Like Xu]

On 5/11/2021 11:30 pm, Sean Christopherson wrote:
> On Wed, Nov 03, 2021, Like Xu wrote:
>> Replace the kvm_pmu_ops pointer in common x86 with an instance of the
>> struct to save one pointer dereference when invoking functions. Copy the
>> struct by value to set the ops during kvm_init().
>>
>> Using kvm_x86_ops.hardware_enable to track whether or not the
>> ops have been initialized, i.e. a vendor KVM module has been loaded.
>>
>> Signed-off-by: Like Xu <likexu@tencent.com>
>> ---
>>   arch/x86/kvm/pmu.c        | 41 +++++++++++++++++++++------------------
>>   arch/x86/kvm/pmu.h        |  4 +++-
>>   arch/x86/kvm/vmx/nested.c |  2 +-
>>   arch/x86/kvm/x86.c        |  3 +++
>>   4 files changed, 29 insertions(+), 21 deletions(-)
>>
>> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
>> index 0772bad9165c..0db1887137d9 100644
>> --- a/arch/x86/kvm/pmu.c
>> +++ b/arch/x86/kvm/pmu.c
>> @@ -47,6 +47,9 @@
>>    *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] <=> gp counters
>>    */
>>   
>> +struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
>> +EXPORT_SYMBOL_GPL(kvm_pmu_ops);
>> +
> 
> ...
> 
>> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
>> index b4ee5e9f9e20..1e793e44b5ff 100644
>> --- a/arch/x86/kvm/vmx/nested.c
>> +++ b/arch/x86/kvm/vmx/nested.c
>> @@ -4796,7 +4796,7 @@ void nested_vmx_pmu_entry_exit_ctls_update(struct kvm_vcpu *vcpu)
>>   		return;
>>   
>>   	vmx = to_vmx(vcpu);
>> -	if (kvm_x86_ops.pmu_ops->is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
>> +	if (kvm_pmu_ops.is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
> 
> I would much prefer we export kvm_pmu_is_valid_msr() and go through that for nVMX
> than export all of kvm_pmu_ops for this one case.

Applied. Is it an abuse to export a function for only one case ?

> 
>>   		vmx->nested.msrs.entry_ctls_high |=
>>   				VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL;
>>   		vmx->nested.msrs.exit_ctls_high |=
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index ac83d873d65b..72d286595012 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -11317,6 +11317,9 @@ int kvm_arch_hardware_setup(void *opaque)
>>   	memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
>>   	kvm_ops_static_call_update();
>>   
>> +	if (kvm_x86_ops.hardware_enable)
> 
> Huh?  Did you intend this to be?
> 
> 	if (kvm_x86_ops.pmu_ops)
> 
> Either way, I don't see the point, VMX and SVM unconditionally provide the ops.

Let me drop it.

> 
> I would also say land this memcpy() above kvm_ops_static_call_update(), then the
> enabling patch can do the static call updates in kvm_ops_static_call_update()
> instead of adding another helper.

Both applied.

> 
>> +		memcpy(&kvm_pmu_ops, kvm_x86_ops.pmu_ops, sizeof(kvm_pmu_ops));
> 
> As part of this change, the pmu_ops should be moved to kvm_x86_init_ops and tagged
> as __initdata.  That'll save those precious few bytes, and more importantly make
> the original ops unreachable, i.e. make it harder to sneak in post-init modification
> bugs.

Applied. Opportunistically,

-struct kvm_pmu_ops [amd|intel]_pmu_ops = {
+struct kvm_pmu_ops [amd|intel]_pmu_ops __initdata = {

> 
>> +
>>   	if (!kvm_cpu_cap_has(X86_FEATURE_XSAVES))
>>   		supported_xss = 0;
>>   
>> -- 
>> 2.33.0
>>
> 

[PATCH 2/3] KVM: x86: Introduce definitions to support static calls for kvm_pmu_ops

[Like Xu]

Use static calls to improve kvm_pmu_ops performance. Introduce the
definitions that will be used by a subsequent patch to actualize the
savings. Add a new kvm-x86-pmu-ops.h header that can be used for the
definition of static calls. This header is also intended to be
used to simplify the defition of amd_pmu_ops and intel_pmu_ops.

Like what we did for kvm_x86_ops, 'pmu_ops' can be covered by
static calls in a simlilar manner for insignificant but not
negligible performance impact, especially on older models.

Signed-off-by: Like Xu <likexu@tencent.com>
---
 arch/x86/include/asm/kvm-x86-pmu-ops.h | 32 ++++++++++++++++++++++++++
 arch/x86/kvm/pmu.c                     |  7 ++++++
 arch/x86/kvm/pmu.h                     | 15 ++++++++++++
 3 files changed, 54 insertions(+)
 create mode 100644 arch/x86/include/asm/kvm-x86-pmu-ops.h

diff --git a/arch/x86/include/asm/kvm-x86-pmu-ops.h b/arch/x86/include/asm/kvm-x86-pmu-ops.h
new file mode 100644
index 000000000000..b7713b16d21d
--- /dev/null
+++ b/arch/x86/include/asm/kvm-x86-pmu-ops.h
@@ -0,0 +1,32 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#if !defined(KVM_X86_PMU_OP) || !defined(KVM_X86_PMU_OP_NULL)
+BUILD_BUG_ON(1)
+#endif
+
+/*
+ * KVM_X86_PMU_OP() and KVM_X86_PMU_OP_NULL() are used to
+ * help generate "static_call()"s. They are also intended for use when defining
+ * the amd/intel KVM_X86_PMU_OPs. KVM_X86_PMU_OP() can be used
+ * for those functions that follow the [amd|intel]_func_name convention.
+ * KVM_X86_PMU_OP_NULL() can leave a NULL definition for the
+ * case where there is no definition or a function name that
+ * doesn't match the typical naming convention is supplied.
+ */
+KVM_X86_PMU_OP(find_arch_event);
+KVM_X86_PMU_OP(find_fixed_event);
+KVM_X86_PMU_OP(pmc_is_enabled);
+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);
+KVM_X86_PMU_OP(is_valid_msr);
+KVM_X86_PMU_OP(get_msr);
+KVM_X86_PMU_OP(set_msr);
+KVM_X86_PMU_OP(refresh);
+KVM_X86_PMU_OP(init);
+KVM_X86_PMU_OP(reset);
+KVM_X86_PMU_OP_NULL(deliver_pmi);
+KVM_X86_PMU_OP_NULL(cleanup);
+
+#undef KVM_X86_PMU_OP
+#undef KVM_X86_PMU_OP_NULL
diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
index 0db1887137d9..b6f08c719125 100644
--- a/arch/x86/kvm/pmu.c
+++ b/arch/x86/kvm/pmu.c
@@ -50,6 +50,13 @@
 struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
 EXPORT_SYMBOL_GPL(kvm_pmu_ops);
 
+#define	KVM_X86_PMU_OP(func)	\
+	DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,	\
+				*(((struct kvm_pmu_ops *)0)->func))
+#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
+#include <asm/kvm-x86-pmu-ops.h>
+EXPORT_STATIC_CALL_GPL(kvm_x86_pmu_is_valid_msr);
+
 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
 {
 	struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
index b2fe135d395a..e5550d4acf14 100644
--- a/arch/x86/kvm/pmu.h
+++ b/arch/x86/kvm/pmu.h
@@ -3,6 +3,8 @@
 #define __KVM_X86_PMU_H
 
 #include <linux/nospec.h>
+#include <linux/static_call_types.h>
+#include <linux/static_call.h>
 
 #define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu)
 #define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu))
@@ -45,6 +47,19 @@ struct kvm_pmu_ops {
 	void (*cleanup)(struct kvm_vcpu *vcpu);
 };
 
+#define	KVM_X86_PMU_OP(func)	\
+	DECLARE_STATIC_CALL(kvm_x86_pmu_##func, *(((struct kvm_pmu_ops *)0)->func))
+#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
+#include <asm/kvm-x86-pmu-ops.h>
+
+static inline void kvm_pmu_ops_static_call_update(void)
+{
+#define	KVM_X86_PMU_OP(func)	\
+	static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func)
+#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
+#include <asm/kvm-x86-pmu-ops.h>
+}
+
 static inline u64 pmc_bitmask(struct kvm_pmc *pmc)
 {
 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
-- 
2.33.0

Re: [PATCH 2/3] KVM: x86: Introduce definitions to support static calls for kvm_pmu_ops

[Sean Christopherson]

On Wed, Nov 03, 2021, Like Xu wrote:
> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> index 0db1887137d9..b6f08c719125 100644
> --- a/arch/x86/kvm/pmu.c
> +++ b/arch/x86/kvm/pmu.c
> @@ -50,6 +50,13 @@
>  struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
>  EXPORT_SYMBOL_GPL(kvm_pmu_ops);
>  
> +#define	KVM_X86_PMU_OP(func)	\
> +	DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,	\
> +				*(((struct kvm_pmu_ops *)0)->func))
> +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP

More of a question for the existing code, what's the point of KVM_X86_OP_NULL?
AFAICT, it always resolves to KVM_X86_OP.  Unless there's some magic I'm missing,
I vote we remove KVM_X86_OP_NULL and then not introduce KVM_X86_PMU_OP_NULL.
And I'm pretty sure it's useless, e.g. get_cs_db_l_bits is defined with the NULL
variant, but it's never NULL and its calls aren't guarded with anything.  And if
KVM_X86_OP_NULL is intended to aid in documenting behavior, it's doing a pretty
miserable job of that :-)

> +#include <asm/kvm-x86-pmu-ops.h>
> +EXPORT_STATIC_CALL_GPL(kvm_x86_pmu_is_valid_msr);

I'll double down on my nVMX suggestion so that this export can be avoided.

>  static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
>  {
>  	struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
> diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
> index b2fe135d395a..e5550d4acf14 100644
> --- a/arch/x86/kvm/pmu.h
> +++ b/arch/x86/kvm/pmu.h
> @@ -3,6 +3,8 @@
>  #define __KVM_X86_PMU_H
>  
>  #include <linux/nospec.h>
> +#include <linux/static_call_types.h>
> +#include <linux/static_call.h>
>  
>  #define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu)
>  #define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu))
> @@ -45,6 +47,19 @@ struct kvm_pmu_ops {
>  	void (*cleanup)(struct kvm_vcpu *vcpu);
>  };
>  
> +#define	KVM_X86_PMU_OP(func)	\
> +	DECLARE_STATIC_CALL(kvm_x86_pmu_##func, *(((struct kvm_pmu_ops *)0)->func))
> +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
> +#include <asm/kvm-x86-pmu-ops.h>
> +
> +static inline void kvm_pmu_ops_static_call_update(void)
> +{
> +#define	KVM_X86_PMU_OP(func)	\
> +	static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func)
> +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
> +#include <asm/kvm-x86-pmu-ops.h>
> +}

As alluded to in patch 01, I'd prefer these go in kvm_ops_static_call_update()
to keep the static call magic somewhat contained.

> +
>  static inline u64 pmc_bitmask(struct kvm_pmc *pmc)
>  {
>  	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
> -- 
> 2.33.0
> 

Re: [PATCH 2/3] KVM: x86: Introduce definitions to support static calls for kvm_pmu_ops

[Like Xu]

On 5/11/2021 11:48 pm, Sean Christopherson wrote:
> On Wed, Nov 03, 2021, Like Xu wrote:
>> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
>> index 0db1887137d9..b6f08c719125 100644
>> --- a/arch/x86/kvm/pmu.c
>> +++ b/arch/x86/kvm/pmu.c
>> @@ -50,6 +50,13 @@
>>   struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
>>   EXPORT_SYMBOL_GPL(kvm_pmu_ops);
>>   
>> +#define	KVM_X86_PMU_OP(func)	\
>> +	DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,	\
>> +				*(((struct kvm_pmu_ops *)0)->func))
>> +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
> 
> More of a question for the existing code, what's the point of KVM_X86_OP_NULL?

The comment says:

  * KVM_X86_OP_NULL() can leave a NULL definition for the
  * case where there is no definition or a function name that
  * doesn't match the typical naming convention is supplied.

Does it help ?

> AFAICT, it always resolves to KVM_X86_OP.  Unless there's some magic I'm missing,
> I vote we remove KVM_X86_OP_NULL and then not introduce KVM_X86_PMU_OP_NULL.
> And I'm pretty sure it's useless, e.g. get_cs_db_l_bits is defined with the NULL

This transitions will not be included in the next version. Open to you.

> variant, but it's never NULL and its calls aren't guarded with anything.  And if
> KVM_X86_OP_NULL is intended to aid in documenting behavior, it's doing a pretty
> miserable job of that :-)
> 
>> +#include <asm/kvm-x86-pmu-ops.h>
>> +EXPORT_STATIC_CALL_GPL(kvm_x86_pmu_is_valid_msr);
> 
> I'll double down on my nVMX suggestion so that this export can be avoided.

Fine to me.

> 
>>   static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
>>   {
>>   	struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
>> diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
>> index b2fe135d395a..e5550d4acf14 100644
>> --- a/arch/x86/kvm/pmu.h
>> +++ b/arch/x86/kvm/pmu.h
>> @@ -3,6 +3,8 @@
>>   #define __KVM_X86_PMU_H
>>   
>>   #include <linux/nospec.h>
>> +#include <linux/static_call_types.h>
>> +#include <linux/static_call.h>
>>   
>>   #define vcpu_to_pmu(vcpu) (&(vcpu)->arch.pmu)
>>   #define pmu_to_vcpu(pmu)  (container_of((pmu), struct kvm_vcpu, arch.pmu))
>> @@ -45,6 +47,19 @@ struct kvm_pmu_ops {
>>   	void (*cleanup)(struct kvm_vcpu *vcpu);
>>   };
>>   
>> +#define	KVM_X86_PMU_OP(func)	\
>> +	DECLARE_STATIC_CALL(kvm_x86_pmu_##func, *(((struct kvm_pmu_ops *)0)->func))
>> +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
>> +#include <asm/kvm-x86-pmu-ops.h>
>> +
>> +static inline void kvm_pmu_ops_static_call_update(void)
>> +{
>> +#define	KVM_X86_PMU_OP(func)	\
>> +	static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func)
>> +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
>> +#include <asm/kvm-x86-pmu-ops.h>
>> +}
> 
> As alluded to in patch 01, I'd prefer these go in kvm_ops_static_call_update()
> to keep the static call magic somewhat contained.

Thank and applied.

> 
>> +
>>   static inline u64 pmc_bitmask(struct kvm_pmc *pmc)
>>   {
>>   	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
>> -- 
>> 2.33.0
>>
> 

Re: [PATCH 2/3] KVM: x86: Introduce definitions to support static calls for kvm_pmu_ops

[Sean Christopherson]

On Mon, Nov 08, 2021, Like Xu wrote:
> On 5/11/2021 11:48 pm, Sean Christopherson wrote:
> > On Wed, Nov 03, 2021, Like Xu wrote:
> > > diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> > > index 0db1887137d9..b6f08c719125 100644
> > > --- a/arch/x86/kvm/pmu.c
> > > +++ b/arch/x86/kvm/pmu.c
> > > @@ -50,6 +50,13 @@
> > >   struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
> > >   EXPORT_SYMBOL_GPL(kvm_pmu_ops);
> > > +#define	KVM_X86_PMU_OP(func)	\
> > > +	DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,	\
> > > +				*(((struct kvm_pmu_ops *)0)->func))
> > > +#define	KVM_X86_PMU_OP_NULL	KVM_X86_PMU_OP
> > 
> > More of a question for the existing code, what's the point of KVM_X86_OP_NULL?
> 
> The comment says:
> 
>  * KVM_X86_OP_NULL() can leave a NULL definition for the
>  * case where there is no definition or a function name that
>  * doesn't match the typical naming convention is supplied.
> 
> Does it help ?

No.  I understand the original intent of KVM_X86_OP_NULL, but unless there's some
form of enforcement, it does more harm than good because it can very easily become
stale, e.g. see get_cs_db_l_bits().  I guess "what's the point of KVM_X86_OP_NULL?"
was somewhat of a rhetorical question.

> > AFAICT, it always resolves to KVM_X86_OP.  Unless there's some magic I'm missing,
> > I vote we remove KVM_X86_OP_NULL and then not introduce KVM_X86_PMU_OP_NULL.
> > And I'm pretty sure it's useless, e.g. get_cs_db_l_bits is defined with the NULL

[PATCH 3/3] KVM: x86: Use static calls to reduce kvm_pmu_ops overhead

[Like Xu]

Convert kvm_pmu_ops to use static calls.

Here are the worst sched_clock() nanosecond numbers for the kvm_pmu_ops
functions that is most often called (up to 7 digits of calls) when running
a single perf test case in a guest on an ICX 2.70GHz host (mitigations=on):

		|	legacy	|	static call
------------------------------------------------------------
.pmc_idx_to_pmc	|	10946	|	10047 (8%)
.pmc_is_enabled	|	11291	|	11175 (1%)
.msr_idx_to_pmc	|	13526	|	12346 (8%)
.is_valid_msr	|	10895	|	10484 (3%)

Signed-off-by: Like Xu <likexu@tencent.com>
---
 arch/x86/kvm/pmu.c        | 36 +++++++++++++++++-------------------
 arch/x86/kvm/pmu.h        |  2 +-
 arch/x86/kvm/vmx/nested.c |  2 +-
 arch/x86/kvm/x86.c        |  4 +++-
 4 files changed, 22 insertions(+), 22 deletions(-)

diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
index b6f08c719125..193f925e2064 100644
--- a/arch/x86/kvm/pmu.c
+++ b/arch/x86/kvm/pmu.c
@@ -224,7 +224,7 @@ void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
 			  ARCH_PERFMON_EVENTSEL_CMASK |
 			  HSW_IN_TX |
 			  HSW_IN_TX_CHECKPOINTED))) {
-		config = kvm_pmu_ops.find_arch_event(pmc_to_pmu(pmc),
+		config = static_call(kvm_x86_pmu_find_arch_event)(pmc_to_pmu(pmc),
 						      event_select,
 						      unit_mask);
 		if (config != PERF_COUNT_HW_MAX)
@@ -278,7 +278,7 @@ void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
 
 	pmc->current_config = (u64)ctrl;
 	pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
-			      kvm_pmu_ops.find_fixed_event(idx),
+			      static_call(kvm_x86_pmu_find_fixed_event)(idx),
 			      !(en_field & 0x2), /* exclude user */
 			      !(en_field & 0x1), /* exclude kernel */
 			      pmi, false, false);
@@ -287,7 +287,7 @@ EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
 
 void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
 {
-	struct kvm_pmc *pmc = kvm_pmu_ops.pmc_idx_to_pmc(pmu, pmc_idx);
+	struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, pmc_idx);
 
 	if (!pmc)
 		return;
@@ -309,7 +309,7 @@ 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 = kvm_pmu_ops.pmc_idx_to_pmc(pmu, bit);
+		struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, bit);
 
 		if (unlikely(!pmc || !pmc->perf_event)) {
 			clear_bit(bit, pmu->reprogram_pmi);
@@ -331,7 +331,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 /* check if idx is a valid index to access PMU */
 int kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
 {
-	return kvm_pmu_ops.is_valid_rdpmc_ecx(vcpu, idx);
+	return static_call(kvm_x86_pmu_is_valid_rdpmc_ecx)(vcpu, idx);
 }
 
 bool is_vmware_backdoor_pmc(u32 pmc_idx)
@@ -381,7 +381,7 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 	if (is_vmware_backdoor_pmc(idx))
 		return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
 
-	pmc = kvm_pmu_ops.rdpmc_ecx_to_pmc(vcpu, idx, &mask);
+	pmc = static_call(kvm_x86_pmu_rdpmc_ecx_to_pmc)(vcpu, idx, &mask);
 	if (!pmc)
 		return 1;
 
@@ -397,22 +397,21 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
 {
 	if (lapic_in_kernel(vcpu)) {
-		if (kvm_pmu_ops.deliver_pmi)
-			kvm_pmu_ops.deliver_pmi(vcpu);
+		static_call_cond(kvm_x86_pmu_deliver_pmi)(vcpu);
 		kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
 	}
 }
 
 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
 {
-	return kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr) ||
-		kvm_pmu_ops.is_valid_msr(vcpu, msr);
+	return static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr) ||
+		static_call(kvm_x86_pmu_is_valid_msr)(vcpu, msr);
 }
 
 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 {
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
-	struct kvm_pmc *pmc = kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr);
+	struct kvm_pmc *pmc = static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr);
 
 	if (pmc)
 		__set_bit(pmc->idx, pmu->pmc_in_use);
@@ -420,13 +419,13 @@ static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 
 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 {
-	return kvm_pmu_ops.get_msr(vcpu, msr_info);
+	return static_call(kvm_x86_pmu_get_msr)(vcpu, msr_info);
 }
 
 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 {
 	kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
-	return kvm_pmu_ops.set_msr(vcpu, msr_info);
+	return static_call(kvm_x86_pmu_set_msr)(vcpu, msr_info);
 }
 
 /* refresh PMU settings. This function generally is called when underlying
@@ -435,7 +434,7 @@ int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
  */
 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 {
-	kvm_pmu_ops.refresh(vcpu);
+	static_call(kvm_x86_pmu_refresh)(vcpu);
 }
 
 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
@@ -443,7 +442,7 @@ void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 
 	irq_work_sync(&pmu->irq_work);
-	kvm_pmu_ops.reset(vcpu);
+	static_call(kvm_x86_pmu_reset)(vcpu);
 }
 
 void kvm_pmu_init(struct kvm_vcpu *vcpu)
@@ -451,7 +450,7 @@ void kvm_pmu_init(struct kvm_vcpu *vcpu)
 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 
 	memset(pmu, 0, sizeof(*pmu));
-	kvm_pmu_ops.init(vcpu);
+	static_call(kvm_x86_pmu_init)(vcpu);
 	init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
 	pmu->event_count = 0;
 	pmu->need_cleanup = false;
@@ -483,14 +482,13 @@ 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 = kvm_pmu_ops.pmc_idx_to_pmc(pmu, i);
+		pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
 
 		if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
 			pmc_stop_counter(pmc);
 	}
 
-	if (kvm_pmu_ops.cleanup)
-		kvm_pmu_ops.cleanup(vcpu);
+	static_call_cond(kvm_x86_pmu_cleanup)(vcpu);
 
 	bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
 }
diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
index e5550d4acf14..1818d1371ece 100644
--- a/arch/x86/kvm/pmu.h
+++ b/arch/x86/kvm/pmu.h
@@ -109,7 +109,7 @@ static inline bool pmc_is_fixed(struct kvm_pmc *pmc)
 
 static inline bool pmc_is_enabled(struct kvm_pmc *pmc)
 {
-	return kvm_pmu_ops.pmc_is_enabled(pmc);
+	return static_call(kvm_x86_pmu_pmc_is_enabled)(pmc);
 }
 
 static inline bool kvm_valid_perf_global_ctrl(struct kvm_pmu *pmu,
diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index 1e793e44b5ff..a61661de1f39 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -4796,7 +4796,7 @@ void nested_vmx_pmu_entry_exit_ctls_update(struct kvm_vcpu *vcpu)
 		return;
 
 	vmx = to_vmx(vcpu);
-	if (kvm_pmu_ops.is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
+	if (static_call(kvm_x86_pmu_is_valid_msr)(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
 		vmx->nested.msrs.entry_ctls_high |=
 				VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL;
 		vmx->nested.msrs.exit_ctls_high |=
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 72d286595012..88a3ef809c98 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -11317,8 +11317,10 @@ int kvm_arch_hardware_setup(void *opaque)
 	memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
 	kvm_ops_static_call_update();
 
-	if (kvm_x86_ops.hardware_enable)
+	if (kvm_x86_ops.hardware_enable) {
 		memcpy(&kvm_pmu_ops, kvm_x86_ops.pmu_ops, sizeof(kvm_pmu_ops));
+		kvm_pmu_ops_static_call_update();
+	}
 
 	if (!kvm_cpu_cap_has(X86_FEATURE_XSAVES))
 		supported_xss = 0;
-- 
2.33.0

Re: [PATCH 3/3] KVM: x86: Use static calls to reduce kvm_pmu_ops overhead

[Yao Yuan]

On Wed, Nov 03, 2021 at 03:03:10PM +0800, Like Xu wrote:
> Convert kvm_pmu_ops to use static calls.
>
> Here are the worst sched_clock() nanosecond numbers for the kvm_pmu_ops
> functions that is most often called (up to 7 digits of calls) when running
> a single perf test case in a guest on an ICX 2.70GHz host (mitigations=on):
>
>       |	legacy	|	static call
> ------------------------------------------------------------
> .pmc_idx_to_pmc	|	10946	|	10047 (8%)
> .pmc_is_enabled	|	11291	|	11175 (1%)
> .msr_idx_to_pmc	|	13526	|	12346 (8%)
> .is_valid_msr	|	10895	|	10484 (3%)
>
> Signed-off-by: Like Xu <likexu@tencent.com>
> ---
>  arch/x86/kvm/pmu.c        | 36 +++++++++++++++++-------------------
>  arch/x86/kvm/pmu.h        |  2 +-
>  arch/x86/kvm/vmx/nested.c |  2 +-
>  arch/x86/kvm/x86.c        |  4 +++-
>  4 files changed, 22 insertions(+), 22 deletions(-)
>
> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> index b6f08c719125..193f925e2064 100644
> --- a/arch/x86/kvm/pmu.c
> +++ b/arch/x86/kvm/pmu.c
> @@ -224,7 +224,7 @@ void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
>             ARCH_PERFMON_EVENTSEL_CMASK |
>             HSW_IN_TX |
>             HSW_IN_TX_CHECKPOINTED))) {
> -		config = kvm_pmu_ops.find_arch_event(pmc_to_pmu(pmc),
> +		config = static_call(kvm_x86_pmu_find_arch_event)(pmc_to_pmu(pmc),

Why you need change them into kvm_pmu_ops.XXX then convert
them into static call ? Move the instance definition of
kvm_pmu_ops from patch 1 into patch 3 and then drop patch 1,
will this work ?

>                             event_select,
>                             unit_mask);
>       if (config != PERF_COUNT_HW_MAX)
> @@ -278,7 +278,7 @@ void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
>
>   pmc->current_config = (u64)ctrl;
>   pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
> -               kvm_pmu_ops.find_fixed_event(idx),
> +               static_call(kvm_x86_pmu_find_fixed_event)(idx),
>                 !(en_field & 0x2), /* exclude user */
>                 !(en_field & 0x1), /* exclude kernel */
>                 pmi, false, false);
> @@ -287,7 +287,7 @@ EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
>
>  void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
>  {
> -	struct kvm_pmc *pmc = kvm_pmu_ops.pmc_idx_to_pmc(pmu, pmc_idx);
> +	struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, pmc_idx);
>
>   if (!pmc)
>       return;
> @@ -309,7 +309,7 @@ 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 = kvm_pmu_ops.pmc_idx_to_pmc(pmu, bit);
> +		struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, bit);
>
>       if (unlikely(!pmc || !pmc->perf_event)) {
>           clear_bit(bit, pmu->reprogram_pmi);
> @@ -331,7 +331,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
>  /* check if idx is a valid index to access PMU */
>  int kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
>  {
> -	return kvm_pmu_ops.is_valid_rdpmc_ecx(vcpu, idx);
> +	return static_call(kvm_x86_pmu_is_valid_rdpmc_ecx)(vcpu, idx);
>  }
>
>  bool is_vmware_backdoor_pmc(u32 pmc_idx)
> @@ -381,7 +381,7 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
>   if (is_vmware_backdoor_pmc(idx))
>       return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
>
> -	pmc = kvm_pmu_ops.rdpmc_ecx_to_pmc(vcpu, idx, &mask);
> +	pmc = static_call(kvm_x86_pmu_rdpmc_ecx_to_pmc)(vcpu, idx, &mask);
>   if (!pmc)
>       return 1;
>
> @@ -397,22 +397,21 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
>  void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
>  {
>   if (lapic_in_kernel(vcpu)) {
> -		if (kvm_pmu_ops.deliver_pmi)
> -			kvm_pmu_ops.deliver_pmi(vcpu);
> +		static_call_cond(kvm_x86_pmu_deliver_pmi)(vcpu);
>       kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
>   }
>  }
>
>  bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
>  {
> -	return kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr) ||
> -		kvm_pmu_ops.is_valid_msr(vcpu, msr);
> +	return static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr) ||
> +		static_call(kvm_x86_pmu_is_valid_msr)(vcpu, msr);
>  }
>
>  static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
>  {
>   struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
> -	struct kvm_pmc *pmc = kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr);
> +	struct kvm_pmc *pmc = static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr);
>
>   if (pmc)
>       __set_bit(pmc->idx, pmu->pmc_in_use);
> @@ -420,13 +419,13 @@ static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
>
>  int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>  {
> -	return kvm_pmu_ops.get_msr(vcpu, msr_info);
> +	return static_call(kvm_x86_pmu_get_msr)(vcpu, msr_info);
>  }
>
>  int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>  {
>   kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
> -	return kvm_pmu_ops.set_msr(vcpu, msr_info);
> +	return static_call(kvm_x86_pmu_set_msr)(vcpu, msr_info);
>  }
>
>  /* refresh PMU settings. This function generally is called when underlying
> @@ -435,7 +434,7 @@ int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>   */
>  void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
>  {
> -	kvm_pmu_ops.refresh(vcpu);
> +	static_call(kvm_x86_pmu_refresh)(vcpu);
>  }
>
>  void kvm_pmu_reset(struct kvm_vcpu *vcpu)
> @@ -443,7 +442,7 @@ void kvm_pmu_reset(struct kvm_vcpu *vcpu)
>   struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
>
>   irq_work_sync(&pmu->irq_work);
> -	kvm_pmu_ops.reset(vcpu);
> +	static_call(kvm_x86_pmu_reset)(vcpu);
>  }
>
>  void kvm_pmu_init(struct kvm_vcpu *vcpu)
> @@ -451,7 +450,7 @@ void kvm_pmu_init(struct kvm_vcpu *vcpu)
>   struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
>
>   memset(pmu, 0, sizeof(*pmu));
> -	kvm_pmu_ops.init(vcpu);
> +	static_call(kvm_x86_pmu_init)(vcpu);
>   init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
>   pmu->event_count = 0;
>   pmu->need_cleanup = false;
> @@ -483,14 +482,13 @@ 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 = kvm_pmu_ops.pmc_idx_to_pmc(pmu, i);
> +		pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
>
>       if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
>           pmc_stop_counter(pmc);
>   }
>
> -	if (kvm_pmu_ops.cleanup)
> -		kvm_pmu_ops.cleanup(vcpu);
> +	static_call_cond(kvm_x86_pmu_cleanup)(vcpu);
>
>   bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
>  }
> diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
> index e5550d4acf14..1818d1371ece 100644
> --- a/arch/x86/kvm/pmu.h
> +++ b/arch/x86/kvm/pmu.h
> @@ -109,7 +109,7 @@ static inline bool pmc_is_fixed(struct kvm_pmc *pmc)
>
>  static inline bool pmc_is_enabled(struct kvm_pmc *pmc)
>  {
> -	return kvm_pmu_ops.pmc_is_enabled(pmc);
> +	return static_call(kvm_x86_pmu_pmc_is_enabled)(pmc);
>  }
>
>  static inline bool kvm_valid_perf_global_ctrl(struct kvm_pmu *pmu,
> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> index 1e793e44b5ff..a61661de1f39 100644
> --- a/arch/x86/kvm/vmx/nested.c
> +++ b/arch/x86/kvm/vmx/nested.c
> @@ -4796,7 +4796,7 @@ void nested_vmx_pmu_entry_exit_ctls_update(struct kvm_vcpu *vcpu)
>       return;
>
>   vmx = to_vmx(vcpu);
> -	if (kvm_pmu_ops.is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
> +	if (static_call(kvm_x86_pmu_is_valid_msr)(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
>       vmx->nested.msrs.entry_ctls_high |=
>               VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL;
>       vmx->nested.msrs.exit_ctls_high |=
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 72d286595012..88a3ef809c98 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -11317,8 +11317,10 @@ int kvm_arch_hardware_setup(void *opaque)
>   memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
>   kvm_ops_static_call_update();
>
> -	if (kvm_x86_ops.hardware_enable)
> +	if (kvm_x86_ops.hardware_enable) {
>       memcpy(&kvm_pmu_ops, kvm_x86_ops.pmu_ops, sizeof(kvm_pmu_ops));
> +		kvm_pmu_ops_static_call_update();
> +	}
>
>   if (!kvm_cpu_cap_has(X86_FEATURE_XSAVES))
>       supported_xss = 0;
> --
> 2.33.0
>

Re: [PATCH 3/3] KVM: x86: Use static calls to reduce kvm_pmu_ops overhead

[Like Xu]

On 3/11/2021 8:08 pm, Yao Yuan wrote:
> On Wed, Nov 03, 2021 at 03:03:10PM +0800, Like Xu wrote:
>> Convert kvm_pmu_ops to use static calls.
>>
>> Here are the worst sched_clock() nanosecond numbers for the kvm_pmu_ops
>> functions that is most often called (up to 7 digits of calls) when running
>> a single perf test case in a guest on an ICX 2.70GHz host (mitigations=on):
>>
>>        |	legacy	|	static call
>> ------------------------------------------------------------
>> .pmc_idx_to_pmc	|	10946	|	10047 (8%)
>> .pmc_is_enabled	|	11291	|	11175 (1%)
>> .msr_idx_to_pmc	|	13526	|	12346 (8%)
>> .is_valid_msr	|	10895	|	10484 (3%)
>>
>> Signed-off-by: Like Xu <likexu@tencent.com>
>> ---
>>   arch/x86/kvm/pmu.c        | 36 +++++++++++++++++-------------------
>>   arch/x86/kvm/pmu.h        |  2 +-
>>   arch/x86/kvm/vmx/nested.c |  2 +-
>>   arch/x86/kvm/x86.c        |  4 +++-
>>   4 files changed, 22 insertions(+), 22 deletions(-)
>>
>> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
>> index b6f08c719125..193f925e2064 100644
>> --- a/arch/x86/kvm/pmu.c
>> +++ b/arch/x86/kvm/pmu.c
>> @@ -224,7 +224,7 @@ void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
>>              ARCH_PERFMON_EVENTSEL_CMASK |
>>              HSW_IN_TX |
>>              HSW_IN_TX_CHECKPOINTED))) {
>> -		config = kvm_pmu_ops.find_arch_event(pmc_to_pmu(pmc),
>> +		config = static_call(kvm_x86_pmu_find_arch_event)(pmc_to_pmu(pmc),
> 
> Why you need change them into kvm_pmu_ops.XXX then convert
> them into static call ? Move the instance definition of
> kvm_pmu_ops from patch 1 into patch 3 and then drop patch 1,
> will this work ?

You may git squash all commits to get it *work*.

With reference to afaf0b2f9b801c6eb2278b52d49e6a7d7b659cf1[1],
doing one thing at a time will make things go smoother.

[1] 
https://lore.kernel.org/lkml/20200321202603.19355-7-sean.j.christopherson@intel.com/

> 
>>                              event_select,
>>                              unit_mask);
>>        if (config != PERF_COUNT_HW_MAX)
>> @@ -278,7 +278,7 @@ void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
>>
>>    pmc->current_config = (u64)ctrl;
>>    pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
>> -               kvm_pmu_ops.find_fixed_event(idx),
>> +               static_call(kvm_x86_pmu_find_fixed_event)(idx),
>>                  !(en_field & 0x2), /* exclude user */
>>                  !(en_field & 0x1), /* exclude kernel */
>>                  pmi, false, false);
>> @@ -287,7 +287,7 @@ EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
>>
>>   void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
>>   {
>> -	struct kvm_pmc *pmc = kvm_pmu_ops.pmc_idx_to_pmc(pmu, pmc_idx);
>> +	struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, pmc_idx);
>>
>>    if (!pmc)
>>        return;
>> @@ -309,7 +309,7 @@ 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 = kvm_pmu_ops.pmc_idx_to_pmc(pmu, bit);
>> +		struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, bit);
>>
>>        if (unlikely(!pmc || !pmc->perf_event)) {
>>            clear_bit(bit, pmu->reprogram_pmi);
>> @@ -331,7 +331,7 @@ void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
>>   /* check if idx is a valid index to access PMU */
>>   int kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
>>   {
>> -	return kvm_pmu_ops.is_valid_rdpmc_ecx(vcpu, idx);
>> +	return static_call(kvm_x86_pmu_is_valid_rdpmc_ecx)(vcpu, idx);
>>   }
>>
>>   bool is_vmware_backdoor_pmc(u32 pmc_idx)
>> @@ -381,7 +381,7 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
>>    if (is_vmware_backdoor_pmc(idx))
>>        return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
>>
>> -	pmc = kvm_pmu_ops.rdpmc_ecx_to_pmc(vcpu, idx, &mask);
>> +	pmc = static_call(kvm_x86_pmu_rdpmc_ecx_to_pmc)(vcpu, idx, &mask);
>>    if (!pmc)
>>        return 1;
>>
>> @@ -397,22 +397,21 @@ int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
>>   void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
>>   {
>>    if (lapic_in_kernel(vcpu)) {
>> -		if (kvm_pmu_ops.deliver_pmi)
>> -			kvm_pmu_ops.deliver_pmi(vcpu);
>> +		static_call_cond(kvm_x86_pmu_deliver_pmi)(vcpu);
>>        kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
>>    }
>>   }
>>
>>   bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
>>   {
>> -	return kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr) ||
>> -		kvm_pmu_ops.is_valid_msr(vcpu, msr);
>> +	return static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr) ||
>> +		static_call(kvm_x86_pmu_is_valid_msr)(vcpu, msr);
>>   }
>>
>>   static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
>>   {
>>    struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
>> -	struct kvm_pmc *pmc = kvm_pmu_ops.msr_idx_to_pmc(vcpu, msr);
>> +	struct kvm_pmc *pmc = static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr);
>>
>>    if (pmc)
>>        __set_bit(pmc->idx, pmu->pmc_in_use);
>> @@ -420,13 +419,13 @@ static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
>>
>>   int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>>   {
>> -	return kvm_pmu_ops.get_msr(vcpu, msr_info);
>> +	return static_call(kvm_x86_pmu_get_msr)(vcpu, msr_info);
>>   }
>>
>>   int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>>   {
>>    kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
>> -	return kvm_pmu_ops.set_msr(vcpu, msr_info);
>> +	return static_call(kvm_x86_pmu_set_msr)(vcpu, msr_info);
>>   }
>>
>>   /* refresh PMU settings. This function generally is called when underlying
>> @@ -435,7 +434,7 @@ int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
>>    */
>>   void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
>>   {
>> -	kvm_pmu_ops.refresh(vcpu);
>> +	static_call(kvm_x86_pmu_refresh)(vcpu);
>>   }
>>
>>   void kvm_pmu_reset(struct kvm_vcpu *vcpu)
>> @@ -443,7 +442,7 @@ void kvm_pmu_reset(struct kvm_vcpu *vcpu)
>>    struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
>>
>>    irq_work_sync(&pmu->irq_work);
>> -	kvm_pmu_ops.reset(vcpu);
>> +	static_call(kvm_x86_pmu_reset)(vcpu);
>>   }
>>
>>   void kvm_pmu_init(struct kvm_vcpu *vcpu)
>> @@ -451,7 +450,7 @@ void kvm_pmu_init(struct kvm_vcpu *vcpu)
>>    struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
>>
>>    memset(pmu, 0, sizeof(*pmu));
>> -	kvm_pmu_ops.init(vcpu);
>> +	static_call(kvm_x86_pmu_init)(vcpu);
>>    init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
>>    pmu->event_count = 0;
>>    pmu->need_cleanup = false;
>> @@ -483,14 +482,13 @@ 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 = kvm_pmu_ops.pmc_idx_to_pmc(pmu, i);
>> +		pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
>>
>>        if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
>>            pmc_stop_counter(pmc);
>>    }
>>
>> -	if (kvm_pmu_ops.cleanup)
>> -		kvm_pmu_ops.cleanup(vcpu);
>> +	static_call_cond(kvm_x86_pmu_cleanup)(vcpu);
>>
>>    bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
>>   }
>> diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h
>> index e5550d4acf14..1818d1371ece 100644
>> --- a/arch/x86/kvm/pmu.h
>> +++ b/arch/x86/kvm/pmu.h
>> @@ -109,7 +109,7 @@ static inline bool pmc_is_fixed(struct kvm_pmc *pmc)
>>
>>   static inline bool pmc_is_enabled(struct kvm_pmc *pmc)
>>   {
>> -	return kvm_pmu_ops.pmc_is_enabled(pmc);
>> +	return static_call(kvm_x86_pmu_pmc_is_enabled)(pmc);
>>   }
>>
>>   static inline bool kvm_valid_perf_global_ctrl(struct kvm_pmu *pmu,
>> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
>> index 1e793e44b5ff..a61661de1f39 100644
>> --- a/arch/x86/kvm/vmx/nested.c
>> +++ b/arch/x86/kvm/vmx/nested.c
>> @@ -4796,7 +4796,7 @@ void nested_vmx_pmu_entry_exit_ctls_update(struct kvm_vcpu *vcpu)
>>        return;
>>
>>    vmx = to_vmx(vcpu);
>> -	if (kvm_pmu_ops.is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
>> +	if (static_call(kvm_x86_pmu_is_valid_msr)(vcpu, MSR_CORE_PERF_GLOBAL_CTRL)) {
>>        vmx->nested.msrs.entry_ctls_high |=
>>                VM_ENTRY_LOAD_IA32_PERF_GLOBAL_CTRL;
>>        vmx->nested.msrs.exit_ctls_high |=
>> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
>> index 72d286595012..88a3ef809c98 100644
>> --- a/arch/x86/kvm/x86.c
>> +++ b/arch/x86/kvm/x86.c
>> @@ -11317,8 +11317,10 @@ int kvm_arch_hardware_setup(void *opaque)
>>    memcpy(&kvm_x86_ops, ops->runtime_ops, sizeof(kvm_x86_ops));
>>    kvm_ops_static_call_update();
>>
>> -	if (kvm_x86_ops.hardware_enable)
>> +	if (kvm_x86_ops.hardware_enable) {
>>        memcpy(&kvm_pmu_ops, kvm_x86_ops.pmu_ops, sizeof(kvm_pmu_ops));
>> +		kvm_pmu_ops_static_call_update();
>> +	}
>>
>>    if (!kvm_cpu_cap_has(X86_FEATURE_XSAVES))
>>        supported_xss = 0;
>> --
>> 2.33.0
>>
>