Skip to contentSkip to conversation

KVM: vPMU: Use atomic bit operations for global_status

PATCH In review Mingwei Zhang 2 participants 5 messages Sep 11, 2023 – Sep 12, 2023

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

PATCH 1/1 KVM: vPMU: Use atomic bit operations for global_status

Use atomic bit operations for pmu->global_status because it may suffer from race conditions between emulated overflow in KVM vPMU and PEBS overflow in host PMI handler.

Fixes f331601c65ad ("KVM: x86/pmu: Don't generate PEBS records for emulated instructions")
---
 arch/x86/kvm/pmu.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/kvm/pmu.c +3 -3
index edb89b51b383..00b48f25afdb 100644
@@ -117,11 +117,11 @@ static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
 			skip_pmi = true;
 		} else {
 			/* Indicate PEBS overflow PMI to guest. */
-			skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
-						      (unsigned long *)&pmu->global_status);
+			skip_pmi = test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
+						    (unsigned long *)&pmu->global_status);
 		}
 	} else {
-		__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
+		set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
 	}
 
 	if (!pmc->intr || skip_pmi)

base-commit: e2013f46ee2e721567783557c301e5c91d0b74ff

Signature · 1 line
2.42.0.283.g2d96d420d3-goog
#1 lore raw
Quoted 3 lines · Mingwei Zhang

Use atomic bit operations for pmu->global_status because it may suffer from race conditions between emulated overflow in KVM vPMU and PEBS overflow in host PMI handler.

Only if the host PMI occurs on a different pCPU, and if that can happen don't we have a much larger problem?

Quoted 30 lines
Fixes f331601c65ad ("KVM: x86/pmu: Don't generate PEBS records for emulated instructions")
---
 arch/x86/kvm/pmu.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/kvm/pmu.c +3 -3
index edb89b51b383..00b48f25afdb 100644
@@ -117,11 +117,11 @@ static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
 			skip_pmi = true;
 		} else {
 			/* Indicate PEBS overflow PMI to guest. */
-			skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
-						      (unsigned long *)&pmu->global_status);
+			skip_pmi = test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
+						    (unsigned long *)&pmu->global_status);
 		}
 	} else {
-		__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
+		set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
 	}
 
 	if (!pmc->intr || skip_pmi)

base-commit: e2013f46ee2e721567783557c301e5c91d0b74ff

Signature · 1 line
2.42.0.283.g2d96d420d3-goog
#2 lore raw
Quoted 8 lines · Sean Christopherson
Quoted 3 lines · Mingwei Zhang

Use atomic bit operations for pmu->global_status because it may suffer from race conditions between emulated overflow in KVM vPMU and PEBS overflow in host PMI handler.

Only if the host PMI occurs on a different pCPU, and if that can happen don't we have a much larger problem?

Why on different pCPU? For vPMU, I think there is always contention between the vCPU thread and the host PMI handler running on the same pCPU, no?

So, in that case, anything that __kvm_perf_overflow(..., in_pmi=true) touches on struct kvm_pmu will potentially race with the functions like reprogram_counter() -> __kvm_perf_overflow(..., in_pmi=false).

-Mingwei

Quoted 31 lines
Quoted 31 lines
Fixes f331601c65ad ("KVM: x86/pmu: Don't generate PEBS records for emulated instructions")
---
 arch/x86/kvm/pmu.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
arch/x86/kvm/pmu.c +3 -3
index edb89b51b383..00b48f25afdb 100644
@@ -117,11 +117,11 @@ static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
                      skip_pmi = true;
              } else {
                      /* Indicate PEBS overflow PMI to guest. */
-                     skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
-                                                   (unsigned long *)&pmu->global_status);
+                     skip_pmi = test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
+                                                 (unsigned long *)&pmu->global_status);
              }
      } else {
-             __set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
+             set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
      }
 
      if (!pmc->intr || skip_pmi)

base-commit: e2013f46ee2e721567783557c301e5c91d0b74ff --
2.42.0.283.g2d96d420d3-goog

#3 lore raw
Quoted 7 lines · Mingwei Zhang
Quoted 6 lines · Sean Christopherson
Quoted 3 lines · Mingwei Zhang

Use atomic bit operations for pmu->global_status because it may suffer from race conditions between emulated overflow in KVM vPMU and PEBS overflow in host PMI handler.

Quoted 6 lines
Quoted 2 lines

Only if the host PMI occurs on a different pCPU, and if that can happen don't we have a much larger problem?

Why on different pCPU? For vPMU, I think there is always contention between the vCPU thread and the host PMI handler running on the same pCPU, no?

A non-atomic instruction can't be interrupted by an NMI, or any other event, so I don't see how switching to atomic operations fixes anything unless the NMI comes in on a different pCPU.

#4 lore raw
Quoted 12 lines · Sean Christopherson
Quoted 10 lines · Mingwei Zhang
Quoted 8 lines · Sean Christopherson
Quoted 3 lines · Mingwei Zhang

Use atomic bit operations for pmu->global_status because it may suffer from race conditions between emulated overflow in KVM vPMU and PEBS overflow in host PMI handler.

Only if the host PMI occurs on a different pCPU, and if that can happen don't we have a much larger problem?

Quoted 7 lines
Quoted 3 lines

Why on different pCPU? For vPMU, I think there is always contention between the vCPU thread and the host PMI handler running on the same pCPU, no?

A non-atomic instruction can't be interrupted by an NMI, or any other event, so I don't see how switching to atomic operations fixes anything unless the NMI comes in on a different pCPU.

You are right. I realize that. The race condition has to happen concurrently from two different pCPUs. This happens to pmu->reprogram_nmi but not pmu->global_status. The critical stuff we care about should be re-entrancy issues for __kvm_perf_overflow() and some state maintenance issues like avoiding duplicate NMI injection.

That concludes that we don't need this change.

Thanks
-Mingwei