Skip to contentSkip to conversation

KVM: arm64: PMU fixes for nVHE, protected mode

PATCH Applied Oliver Upton 3 participants 11 messages Mar 5, 2024 – Mar 26, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover KVM: arm64: PMU fixes for nVHE, protected mode

Series to address a couple of issues I've hit using the nVHE / protected mode configuration:

- Host-programmed PMU events that run during guest execution do not work if the VM doesn't have a vPMU.

- PMU events do not work at all for non-protected VMs on the protected-mode hypervisor

Oliver Upton (3):
KVM: arm64: pkvm: Actually enable/disable PMU events when running vCPU
KVM: arm64: Fix host-programmed guest events in nVHE
KVM: arm64: Do not disable preemption in kvm_vcpu_pmu_restore_guest()

 arch/arm64/kvm/hyp/nvhe/hyp-main.c | 24 ++++++++++++++++++++++--
 arch/arm64/kvm/pmu.c               |  2 --
 include/kvm/arm_pmu.h              |  2 +-
 3 files changed, 23 insertions(+), 5 deletions(-)

base-commit: 6613476e225e090cc9aad49be7fa504e290dd33d

Signature · 1 line
2.44.0.278.ge034bb2e1d-goog
#1 lore raw

PATCH 1/3 KVM: arm64: pkvm: Actually enable/disable PMU events when running vCPU

flush_hyp_vcpu() does not carry through the PMU event context from the host, regardless of whether the VM is protected or not. This is problematic for two reasons:

- The expectation for non-protected VMs is that the protected-mode hypervisor has feature parity with the 'normal' KVM configuration. However, PMU events programmed to run with the guest do not work.

- The protected-mode hypervisor needs to prevent the host from peering in on the guest. Nothing stops the host from leaving an event enabled straight past guest entry...

Address the both of these issues by priming the hyp vCPU's PMU event context before entering the guest. Take special care to not trust the host in the case of pVMs by referring directly to what hardware says is enabled.

Cc Will Deacon <[email protected]>
Fixes be66e67f1750 ("KVM: arm64: Use the pKVM hyp vCPU structure in handle___kvm_vcpu_run()")
---
 arch/arm64/kvm/hyp/nvhe/hyp-main.c | 24 ++++++++++++++++++++++--
 1 file changed, 22 insertions(+), 2 deletions(-)
arch/arm64/kvm/hyp/nvhe/hyp-main.c +22 -2
index 2385fd03ed87..8621f8383f0c 100644
@@ -23,6 +23,26 @@ DEFINE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
 
 void __kvm_hyp_host_forward_smc(struct kvm_cpu_context *host_ctxt);
 
+static void flush_debug_state(struct pkvm_hyp_vcpu *hyp_vcpu)
+{
+	struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;
+
+	hyp_vcpu->vcpu.arch.mdcr_el2	= host_vcpu->arch.mdcr_el2;
+	hyp_vcpu->vcpu.arch.debug_ptr	= kern_hyp_va(host_vcpu->arch.debug_ptr);
+
+	/*
+	 * The host's PMU context cannot be trusted for protected VMs. Refer to
+	 * hardware to determine which PMCs need to be disabled/enabled on vCPU
+	 * entry/exit, respectively.
+	 */
+	if (kvm_vm_is_protected(kern_hyp_va(host_vcpu->kvm))) {
+		hyp_vcpu->vcpu.arch.pmu.events.events_host = read_sysreg(pmcntenclr_el0);
+		hyp_vcpu->vcpu.arch.pmu.events.events_guest = 0;
+	} else {
+		hyp_vcpu->vcpu.arch.pmu.events = host_vcpu->arch.pmu.events;
+	}
+}
+
 static void flush_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
 {
 	struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;
@@ -35,18 +55,18 @@ static void flush_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
 	hyp_vcpu->vcpu.arch.hw_mmu	= host_vcpu->arch.hw_mmu;
 
 	hyp_vcpu->vcpu.arch.hcr_el2	= host_vcpu->arch.hcr_el2;
-	hyp_vcpu->vcpu.arch.mdcr_el2	= host_vcpu->arch.mdcr_el2;
 	hyp_vcpu->vcpu.arch.cptr_el2	= host_vcpu->arch.cptr_el2;
 
 	hyp_vcpu->vcpu.arch.iflags	= host_vcpu->arch.iflags;
 	hyp_vcpu->vcpu.arch.fp_state	= host_vcpu->arch.fp_state;
 
-	hyp_vcpu->vcpu.arch.debug_ptr	= kern_hyp_va(host_vcpu->arch.debug_ptr);
 	hyp_vcpu->vcpu.arch.host_fpsimd_state = host_vcpu->arch.host_fpsimd_state;
 
 	hyp_vcpu->vcpu.arch.vsesr_el2	= host_vcpu->arch.vsesr_el2;
 
 	hyp_vcpu->vcpu.arch.vgic_cpu.vgic_v3 = host_vcpu->arch.vgic_cpu.vgic_v3;
+
+	flush_debug_state(hyp_vcpu);
 }
 
 static void sync_hyp_vcpu(struct pkvm_hyp_vcpu *hyp_vcpu)
Signature · 1 line
2.44.0.278.ge034bb2e1d-goog
maintainer
#2 lore raw

Re: [PATCH 1/3] KVM: arm64: pkvm: Actually enable/disable PMU events when running vCPU

Quoted 36 lines · Oliver Upton

flush_hyp_vcpu() does not carry through the PMU event context from the host, regardless of whether the VM is protected or not. This is problematic for two reasons:

- The expectation for non-protected VMs is that the protected-mode hypervisor has feature parity with the 'normal' KVM configuration. However, PMU events programmed to run with the guest do not work.

- The protected-mode hypervisor needs to prevent the host from peering in on the guest. Nothing stops the host from leaving an event enabled straight past guest entry...

Address the both of these issues by priming the hyp vCPU's PMU event context before entering the guest. Take special care to not trust the host in the case of pVMs by referring directly to what hardware says is enabled.

Cc Will Deacon <[email protected]>
Fixes be66e67f1750 ("KVM: arm64: Use the pKVM hyp vCPU structure in handle___kvm_vcpu_run()")
---
 arch/arm64/kvm/hyp/nvhe/hyp-main.c | 24 ++++++++++++++++++++++--
 1 file changed, 22 insertions(+), 2 deletions(-)
arch/arm64/kvm/hyp/nvhe/hyp-main.c +3 -0
index 2385fd03ed87..8621f8383f0c 100644
@@ -23,6 +23,26 @@ DEFINE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
 
 void __kvm_hyp_host_forward_smc(struct kvm_cpu_context *host_ctxt);
 
+static void flush_debug_state(struct pkvm_hyp_vcpu *hyp_vcpu)
+{
+	struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;

arch/arm64/kvm/hyp/nvhe/hyp-main.c DEFINE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params); · line 38

Quoted 10 lines
arch/arm64/kvm/hyp/nvhe/hyp-main.c +10 -0
+
+	hyp_vcpu->vcpu.arch.mdcr_el2	= host_vcpu->arch.mdcr_el2;
+	hyp_vcpu->vcpu.arch.debug_ptr	= kern_hyp_va(host_vcpu->arch.debug_ptr);
+
+	/*
+	 * The host's PMU context cannot be trusted for protected VMs. Refer to
+	 * hardware to determine which PMCs need to be disabled/enabled on vCPU
+	 * entry/exit, respectively.
+	 */
+	if (kvm_vm_is_protected(kern_hyp_va(host_vcpu->kvm))) {

This looks wrong. We should never rely on *host* controlled data to make a decision at EL2 in the pKVM case. Use of this helper at EL2 should be constrained to the hyp view of the kvm structure only.

The other thing is that most of this code is going to be thrown away, hopefully soon... (hint, nudge...).

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#3 lore raw

Re: [PATCH 1/3] KVM: arm64: pkvm: Actually enable/disable PMU events when running vCPU

Quoted 10 lines · Marc Zyngier
Quoted 9 lines · Oliver Upton
arch/arm64/kvm/hyp/nvhe/hyp-main.c +9 -0
+static void flush_debug_state(struct pkvm_hyp_vcpu *hyp_vcpu)
+{
+	struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;
+
+	hyp_vcpu->vcpu.arch.mdcr_el2	= host_vcpu->arch.mdcr_el2;
+	hyp_vcpu->vcpu.arch.debug_ptr	= kern_hyp_va(host_vcpu->arch.debug_ptr);
+
+	/*
+	 * The host's PMU context cannot be trusted for protected VMs. Refer to
Quoted 8 lines

arch/arm64/kvm/hyp/nvhe/hyp-main.c DEFINE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params); · line 38

Quoted 4 lines
arch/arm64/kvm/hyp/nvhe/hyp-main.c +4 -0
+	 * hardware to determine which PMCs need to be disabled/enabled on vCPU
+	 * entry/exit, respectively.
+	 */
+	if (kvm_vm_is_protected(kern_hyp_va(host_vcpu->kvm))) {

This looks wrong. We should never rely on *host* controlled data to make a decision at EL2 in the pKVM case. Use of this helper at EL2 should be constrained to the hyp view of the kvm structure only.

I should've spelled this out explicitly. Looking at what we have upstream, there's no hyp-controlled view on a VM's protection (yet).

I had imagined this hilariously wrong expression was a placeholder for pKVM folks, much like we have in early_exit_filter() and kvm_get_exit_handler_array(). But that's no good, and I probably should've added one myself that can be updated with the 'right' thing later on.

Quoted 2 lines

The other thing is that most of this code is going to be thrown away, hopefully soon... (hint, nudge...).

I was hoping Will could provide some insight on whether or not the 'real' pKVM hyp state management carries the same bug. Nothing wrong with fixing what we already have though...

Signature · 2 lines
Thanks,
Oliver
#4 lore raw

Re: [PATCH 1/3] KVM: arm64: pkvm: Actually enable/disable PMU events when running vCPU

Hi,

Quoted 30 lines · Oliver Upton
Quoted 10 lines · Marc Zyngier
Quoted 9 lines · Oliver Upton
arch/arm64/kvm/hyp/nvhe/hyp-main.c +9 -0
+static void flush_debug_state(struct pkvm_hyp_vcpu *hyp_vcpu)
+{
+   struct kvm_vcpu *host_vcpu = hyp_vcpu->host_vcpu;
+
+   hyp_vcpu->vcpu.arch.mdcr_el2    = host_vcpu->arch.mdcr_el2;
+   hyp_vcpu->vcpu.arch.debug_ptr   = kern_hyp_va(host_vcpu->arch.debug_ptr);
+
+   /*
+    * The host's PMU context cannot be trusted for protected VMs. Refer to
Quoted 8 lines

arch/arm64/kvm/hyp/nvhe/hyp-main.c DEFINE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params); · line 38

Quoted 4 lines
arch/arm64/kvm/hyp/nvhe/hyp-main.c +4 -0
+    * hardware to determine which PMCs need to be disabled/enabled on vCPU
+    * entry/exit, respectively.
+    */
+   if (kvm_vm_is_protected(kern_hyp_va(host_vcpu->kvm))) {

This looks wrong. We should never rely on *host* controlled data to make a decision at EL2 in the pKVM case. Use of this helper at EL2 should be constrained to the hyp view of the kvm structure only.

I should've spelled this out explicitly. Looking at what we have upstream, there's no hyp-controlled view on a VM's protection (yet).

I had imagined this hilariously wrong expression was a placeholder for pKVM folks, much like we have in early_exit_filter() and kvm_get_exit_handler_array(). But that's no good, and I probably should've added one myself that can be updated with the 'right' thing later on.

Quoted 6 lines
Quoted 2 lines

The other thing is that most of this code is going to be thrown away, hopefully soon... (hint, nudge...).

I was hoping Will could provide some insight on whether or not the 'real' pKVM hyp state management carries the same bug. Nothing wrong with fixing what we already have though...

Like you said, this is bad/buggy placeholder code. PMU isn't supported for protected guests, and the "real" pKVM code we plan to upstream (soon!) doesn't have this bug.

Thanks,
/fuad

Quoted 4 lines

--
Thanks,
Oliver

#5 lore raw

PATCH 2/3 KVM: arm64: Fix host-programmed guest events in nVHE

In reply to Oliver Upton

Programming PMU events in the host that count during guest execution is a feature supported by perf, e.g.

perf stat -e cpu_cycles:G ./lkvm run

While this works for VHE, the guest/host event bitmaps are not carried through to the hypervisor in the nVHE configuration. Make kvm_pmu_update_vcpu_events() conditional on whether or not _hardware_ supports PMUv3 rather than if the vCPU as vPMU enabled.

Fixes 84d751a019a9 ("KVM: arm64: Pass pmu events to hyp via vcpu")
---
 include/kvm/arm_pmu.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/kvm/arm_pmu.h +1 -1
index 4b9d8fb393a8..df32355e3e38 100644
@@ -86,7 +86,7 @@ void kvm_vcpu_pmu_resync_el0(void);
  */
 #define kvm_pmu_update_vcpu_events(vcpu)				\
 	do {								\
-		if (!has_vhe() && kvm_vcpu_has_pmu(vcpu))		\
+		if (!has_vhe() && kvm_arm_support_pmu_v3())		\
 			vcpu->arch.pmu.events = *kvm_get_pmu_events();	\
 	} while (0)
 
Signature · 1 line
2.44.0.278.ge034bb2e1d-goog
maintainer Reviewed-by on v1 2/3
#6 lore raw

Re: [PATCH 2/3] KVM: arm64: Fix host-programmed guest events in nVHE

Quoted 23 lines · Oliver Upton

Programming PMU events in the host that count during guest execution is a feature supported by perf, e.g.

perf stat -e cpu_cycles:G ./lkvm run

While this works for VHE, the guest/host event bitmaps are not carried through to the hypervisor in the nVHE configuration. Make kvm_pmu_update_vcpu_events() conditional on whether or not _hardware_ supports PMUv3 rather than if the vCPU as vPMU enabled.

Fixes 84d751a019a9 ("KVM: arm64: Pass pmu events to hyp via vcpu")
---
 include/kvm/arm_pmu.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/kvm/arm_pmu.h +0 -0
index 4b9d8fb393a8..df32355e3e38 100644

include/kvm/arm_pmu.h void kvm_vcpu_pmu_resync_el0(void); · line 92

Quoted 8 lines
include/kvm/arm_pmu.h +1 -1
@@ -86,7 +86,7 @@ void kvm_vcpu_pmu_resync_el0(void);
  */
 #define kvm_pmu_update_vcpu_events(vcpu)				\
 	do {								\
-		if (!has_vhe() && kvm_vcpu_has_pmu(vcpu))		\
+		if (!has_vhe() && kvm_arm_support_pmu_v3())		\
 			vcpu->arch.pmu.events = *kvm_get_pmu_events();	\
 	} while (0)
 
	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#7 lore raw

PATCH 3/3 KVM: arm64: Do not disable preemption in kvm_vcpu_pmu_restore_guest()

In reply to Oliver Upton

The caller has already disabled preemption, no need to nest further.

---
 arch/arm64/kvm/pmu.c | 2 --
 1 file changed, 2 deletions(-)
arch/arm64/kvm/pmu.c +0 -2
index a243934c5568..43da27d8087d 100644
@@ -181,14 +181,12 @@ void kvm_vcpu_pmu_restore_guest(struct kvm_vcpu *vcpu)
 	if (!kvm_arm_support_pmu_v3() || !has_vhe())
 		return;
 
-	preempt_disable();
 	pmu = kvm_get_pmu_events();
 	events_guest = pmu->events_guest;
 	events_host = pmu->events_host;
 
 	kvm_vcpu_pmu_enable_el0(events_guest);
 	kvm_vcpu_pmu_disable_el0(events_host);
-	preempt_enable();
 }
 
 /*
Signature · 1 line
2.44.0.278.ge034bb2e1d-goog
maintainer
#8 lore raw

Re: [PATCH 3/3] KVM: arm64: Do not disable preemption in kvm_vcpu_pmu_restore_guest()

Quoted 2 lines · Oliver Upton

The caller has already disabled preemption, no need to nest further.

A quick survey indicates that it doesn't seem to be the case, I'm afraid. access_pmu_evtyper() and access_pmcnten() both seem to be preemptible.

Which probably means that these spots are doing the wrong thing and that we should instead turn it into a KVM_REQ_RESYNC_PMU_EL0 request.

Then the preemption can be actually dropped, and even replaced by an assertion.

Thanks,

	M.
Signature · 1 line
Without deviation from the norm, progress is not possible.
#9 lore raw

Re: [PATCH 3/3] KVM: arm64: Do not disable preemption in kvm_vcpu_pmu_restore_guest()

Quoted 8 lines · Marc Zyngier
Quoted 2 lines · Oliver Upton

The caller has already disabled preemption, no need to nest further.

A quick survey indicates that it doesn't seem to be the case, I'm afraid. access_pmu_evtyper() and access_pmcnten() both seem to be preemptible.

Dammit. Someone forgot to grep...

Quoted 5 lines

Which probably means that these spots are doing the wrong thing and that we should instead turn it into a KVM_REQ_RESYNC_PMU_EL0 request.

Then the preemption can be actually dropped, and even replaced by an assertion.

Agreed, thanks for the suggestion.

Signature · 2 lines
Best,
Oliver
author Applied
#10 lore raw

Re: (subset) [PATCH 0/3] KVM: arm64: PMU fixes for nVHE, protected mode

In reply to Oliver Upton
Quoted 10 lines · Oliver Upton

Series to address a couple of issues I've hit using the nVHE / protected mode configuration:

- Host-programmed PMU events that run during guest execution do not work if the VM doesn't have a vPMU.

- PMU events do not work at all for non-protected VMs on the protected-mode hypervisor

[...]

Just applying the commit that fixes nVHE, I'm happy to ignore the pKVM issue until the 'real' vCPU context changes appear upstream.

Apologies if this email comes through garbled, I can't access my usual workstation right now.

Applied to kvmarm/fixes, thanks!

[2/3] KVM: arm64: Fix host-programmed guest events in nVHE

      commit: e89c928bedd77d181edc2df01cb6672184775140
Signature · 2 lines
Best,
Oliver