From: Like Xu <likexu@xxxxxxxxxxx>
KVM will accumulate an enabled counter for at least INSTRUCTIONS or
BRANCH_INSTRUCTION hw event from any KVM emulated instructions,
generating emulated overflow interrupt on counter overflow, which
in theory should also happen when the PEBS counter overflows but
it currently lacks this part of the underlying support (e.g. through
software injection of records in the irq context or a lazy approach).
In this case, KVM skips the injection of this BUFFER_OVF PMI (effectively
dropping one PEBS record) and let the overflow counter move on. The loss
of a single sample does not introduce a loss of accuracy, but is easily
noticeable for certain specific instructions.
This issue is expected to be addressed along with the issue
of PEBS cross-mapped counters with a slow-path proposal.
Fixes: 79f3e3b58386 ("KVM: x86/pmu: Reprogram PEBS event to emulate guest PEBS counter")
Signed-off-by: Like Xu <likexu@xxxxxxxxxxx>
---
arch/x86/kvm/pmu.c | 16 +++++++++++++---
1 file changed, 13 insertions(+), 3 deletions(-)
diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
index 02f9e4f245bd..390d697efde1 100644
--- a/arch/x86/kvm/pmu.c
+++ b/arch/x86/kvm/pmu.c
@@ -106,9 +106,19 @@ static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
return;
if (pmc->perf_event && pmc->perf_event->attr.precise_ip) {
- /* Indicate PEBS overflow PMI to guest. */
- skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
- (unsigned long *)&pmu->global_status);
+ if (!in_pmi) {
+ /*
+ * TODO: KVM is currently _choosing_ to not generate records
+ * for emulated instructions, avoiding BUFFER_OVF PMI when
+ * there are no records. Strictly speaking, it should be done
+ * as well in the right context to improve sampling accuracy.
+ */
+ 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);
+ }
} else {
__set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
}
--
2.37.1
