[Devel] [PATCH 1/1] x86/kvm: skip PMU event trigger when guest PMU is disabled

[Denis V. Lunev]

kvm_skip_emulated_instruction() calls kvm_pmu_trigger_event() on every
emulated-instruction skip (notably every CPUID exit) to feed the PMU
INSTRUCTIONS_RETIRED counter. The function does a stack bitmap allocation,
a bitmap_and against pmu->global_ctrl, indirect calls to pmc_event_is_allowed()
and cpl_is_matched(), and an empty kvm_for_each_pmc() walk before returning
when nothing is programmed. On guests with PMU disabled at the per-VM level
(<pmu state='off'/>), all of this work is dead-weight and runs on every
VMEXIT.

Guard the call with a check on kvm->arch.enable_pmu. The unlikely() hint
keeps the hot exit path straight-line and places the (rarely-taken) call
out-of-line for better I-cache density.

Measured on at_cpu_cpuid (-k 0, VSTOR-131579 methodology), guest PMU off,
on top of a Nova-side default-PMU-off VM config:

  AMD Zen2 (epyc01, EPYC 7302):     +2.7%, ~57 cyc/VMEXIT saved
  Intel Ice Lake-SP (hw-hdd01):     +0.6%, ~26 cyc/VMEXIT saved

The AMD bias is consistent with the per-call retpoline + unret + Safe-RET
tax on Zen2/Zen3 with default mitigations: removing an entire call/return
sequence from the AMD hot path saves more than the same removal on Intel
with Enhanced/Auto IBRS. Same "retpoline fingerprint" identified in
VSTOR-130475's branch-mispredict-rate analysis.

https://virtuozzo.atlassian.net/browse/VSTOR-132543
https://virtuozzo.atlassian.net/browse/VSTOR-132471

Feature: kvm
Signed-off-by: Denis V. Lunev <den at openvz.org>
---
 arch/x86/kvm/x86.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index a68259e4f66a..aa349adf7b90 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8877,7 +8877,8 @@ int kvm_skip_emulated_instruction(struct kvm_vcpu *vcpu)
 	if (unlikely(!r))
 		return 0;
 
-	kvm_pmu_trigger_event(vcpu, PERF_COUNT_HW_INSTRUCTIONS);
+	if (unlikely(vcpu->kvm->arch.enable_pmu))
+		kvm_pmu_trigger_event(vcpu, PERF_COUNT_HW_INSTRUCTIONS);
 
 	/*
 	 * rflags is the old, "raw" value of the flags.  The new value has
-- 
2.51.0