diff options
author | Sean Christopherson <seanjc@google.com> | 2023-11-04 00:05:41 +0100 |
---|---|---|
committer | Sean Christopherson <seanjc@google.com> | 2023-11-30 21:52:55 +0100 |
commit | fd89499a5151d197ba30f7b801f6d8f4646cf446 (patch) | |
tree | 75a45117d277e89cd58135a43be12047bd8803f9 /arch/x86/kvm/pmu.h | |
parent | KVM: x86/pmu: Update sample period in pmc_write_counter() (diff) | |
download | linux-fd89499a5151d197ba30f7b801f6d8f4646cf446.tar.xz linux-fd89499a5151d197ba30f7b801f6d8f4646cf446.zip |
KVM: x86/pmu: Track emulated counter events instead of previous counter
Explicitly track emulated counter events instead of using the common
counter value that's shared with the hardware counter owned by perf.
Bumping the common counter requires snapshotting the pre-increment value
in order to detect overflow from emulation, and the snapshot approach is
inherently flawed.
Snapshotting the previous counter at every increment assumes that there is
at most one emulated counter event per emulated instruction (or rather,
between checks for KVM_REQ_PMU). That's mostly holds true today because
KVM only emulates (branch) instructions retired, but the approach will
fall apart if KVM ever supports event types that don't have a 1:1
relationship with instructions.
And KVM already has a relevant bug, as handle_invalid_guest_state()
emulates multiple instructions without checking KVM_REQ_PMU, i.e. could
miss an overflow event due to clobbering pmc->prev_counter. Not checking
KVM_REQ_PMU is problematic in both cases, but at least with the emulated
counter approach, the resulting behavior is delayed overflow detection,
as opposed to completely lost detection.
Tracking the emulated count fixes another bug where the snapshot approach
can signal spurious overflow due to incorporating both the emulated count
and perf's count in the check, i.e. if overflow is detected by perf, then
KVM's emulation will also incorrectly signal overflow. Add a comment in
the related code to call out the need to process emulated events *after*
pausing the perf event (big kudos to Mingwei for figuring out that
particular wrinkle).
Cc: Mingwei Zhang <mizhang@google.com>
Cc: Roman Kagan <rkagan@amazon.de>
Cc: Jim Mattson <jmattson@google.com>
Cc: Dapeng Mi <dapeng1.mi@linux.intel.com>
Cc: Like Xu <like.xu.linux@gmail.com>
Reviewed-by: Mingwei Zhang <mizhang@google.com>
Link: https://lore.kernel.org/r/20231103230541.352265-7-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Diffstat (limited to 'arch/x86/kvm/pmu.h')
-rw-r--r-- | arch/x86/kvm/pmu.h | 3 |
1 files changed, 2 insertions, 1 deletions
diff --git a/arch/x86/kvm/pmu.h b/arch/x86/kvm/pmu.h index cae85e550f60..7caeb3d8d4fd 100644 --- a/arch/x86/kvm/pmu.h +++ b/arch/x86/kvm/pmu.h @@ -66,7 +66,8 @@ static inline u64 pmc_read_counter(struct kvm_pmc *pmc) { u64 counter, enabled, running; - counter = pmc->counter; + counter = pmc->counter + pmc->emulated_counter; + if (pmc->perf_event && !pmc->is_paused) counter += perf_event_read_value(pmc->perf_event, &enabled, &running); |