perf/x86/intel: Fix guest vPMU warning on hybrid CPUs
Subsystems (3)
PATCH 1/1 perf/x86/intel: Fix guest vPMU warning on hybrid CPUs
From: Kan Liang <[email protected]>
The below error can be observed in a Linux guest, when the hypervisor is on a hybrid machine.
[ 0.118214] unchecked MSR access error: WRMSR to 0x38f (tried to write 0x00011000f0000003f) at rIP: 0xffffffff83082124 (native_write_msr+0x4/0x30) [ 0.118949] Call Trace: [ 0.119092] <TASK> [ 0.119215] ? __intel_pmu_enable_all.constprop.0+0x88/0xe0 [ 0.119533] intel_pmu_enable_all+0x15/0x20 [ 0.119778] x86_pmu_enable+0x17c/0x320
The current perf wrongly assumes that the perf metrics feature is always enabled on p-core. It unconditionally enables the feature to workaround the unreliable enumeration of the PERF_CAPABILITIES MSR. The assumption is safe to bare metal. However, KVM doesn't support the perf metrics feature yet. Setting the corresponding bit triggers MSR access error in a guest.
Only unconditionally enable the core specific PMU feature for bare metal on ADL and RPL, which includes the perf metrics on p-core and PEBS-via-PT on e-core.
For the future platforms, perf doesn't need to hardcode the PMU feature. The per-core PMU features can be enumerated by the enhanced PERF_CAPABILITIES MSR and CPUID leaf 0x23. There is no such issue.
---
Based on my limit knowledge regarding KVM and guest, I use the HYPERVISOR bit to tell whether it's a guest. But I'm not sure whether it's reliable. Please let me know if there is a better way. Thanks.
arch/x86/events/intel/core.c | 19 +++++++++++++++---- 1 file changed, 15 insertions(+), 4 deletions(-)
arch/x86/events/intel/core.c +15 -4
@@ -6459,8 +6459,17 @@ __init int intel_pmu_init(void)__EVENT_CONSTRAINT (0 , (1ULL << pmu->num_counters) -1 ,0 , pmu->num_counters,0 ,0 ); pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; - pmu->intel_cap.perf_metrics =1 ; - pmu->intel_cap.pebs_output_pt_available =0 ; +/* +* The capability bits are not reliable on ADL and RPL. +* For bare metal, it's safe to assume that some features +* are always enabled, e.g., the perf metrics on p-core, +* but we cannot do the same assumption for a hypervisor. +* Only update the core specific PMU feature for bare metal. +*/ +if (!boot_cpu_has (X86_FEATURE_HYPERVISOR )) { + pmu->intel_cap.perf_metrics =1 ; + pmu->intel_cap.pebs_output_pt_available =0 ; + }memcpy (pmu->hw_cache_event_ids, spr_hw_cache_event_ids,sizeof (pmu->hw_cache_event_ids));memcpy (pmu->hw_cache_extra_regs, spr_hw_cache_extra_regs,sizeof (pmu->hw_cache_extra_regs));
@@ -6480,8 +6489,10 @@ __init int intel_pmu_init(void)__EVENT_CONSTRAINT (0 , (1ULL << pmu->num_counters) -1 ,0 , pmu->num_counters,0 ,0 ); pmu->intel_cap.capabilities = x86_pmu.intel_cap.capabilities; - pmu->intel_cap.perf_metrics =0 ; - pmu->intel_cap.pebs_output_pt_available =1 ; +if (!boot_cpu_has (X86_FEATURE_HYPERVISOR )) { + pmu->intel_cap.perf_metrics =0 ; + pmu->intel_cap.pebs_output_pt_available =1 ; + }memcpy (pmu->hw_cache_event_ids, glp_hw_cache_event_ids,sizeof (pmu->hw_cache_event_ids));memcpy (pmu->hw_cache_extra_regs, tnt_hw_cache_extra_regs,sizeof (pmu->hw_cache_extra_regs));
Signature · 1 line
2.35.1