On 11/4/2023 8:41 PM, Jim Mattson wrote:
On Fri, Nov 3, 2023 at 5:02 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
Don't advertise support to userspace for architectural events that KVM
doesn't support, i.e. for "real" events that aren't listed in
intel_pmu_architectural_events. On current hardware, this effectively
means "don't advertise support for Top Down Slots".
NR_REAL_INTEL_ARCH_EVENTS is only used in intel_hw_event_available().
As discussed (https://lore.kernel.org/kvm/ZUU12-TUR_1cj47u@xxxxxxxxxx/),
intel_hw_event_available() should go away.
Aside from mapping fixed counters to event selector and unit mask
(fixed_pmc_events[]), KVM has no reason to know when a new
architectural event is defined.
Since intel_hw_event_available() would be removed, it looks the enum
intel_pmu_architectural_events and intel_arch_events[] array become
useless. We can directly simply modify current fixed_pmc_events[] array
and use it to store fixed counter events code and umask.
The variable that this change "fixes" is only used to feed
CPUID.0AH:EBX in KVM_GET_SUPPORTED_CPUID, and kvm_pmu_cap.events_mask
is already constructed from what host perf advertises support for.
Mask off the associated "unavailable" bits, as said bits for undefined
events are reserved to zero. Arguably the events _are_ defined, but from
a KVM perspective they might as well not exist, and there's absolutely no
reason to leave useless unavailable bits set.
Fixes: a6c06ed1a60a ("KVM: Expose the architectural performance monitoring CPUID leaf")
Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
---
arch/x86/kvm/vmx/pmu_intel.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/arch/x86/kvm/vmx/pmu_intel.c b/arch/x86/kvm/vmx/pmu_intel.c
index 3316fdea212a..8d545f84dc4a 100644
--- a/arch/x86/kvm/vmx/pmu_intel.c
+++ b/arch/x86/kvm/vmx/pmu_intel.c
@@ -73,6 +73,15 @@ static void intel_init_pmu_capability(void)
int i;
/*
+ * Do not enumerate support for architectural events that KVM doesn't
+ * support. Clear unsupported events "unavailable" bit as well, as
+ * architecturally such bits are reserved to zero.
+ */
+ kvm_pmu_cap.events_mask_len = min(kvm_pmu_cap.events_mask_len,
+ NR_REAL_INTEL_ARCH_EVENTS);
+ kvm_pmu_cap.events_mask &= GENMASK(kvm_pmu_cap.events_mask_len - 1, 0);
+
+ /*
* Perf may (sadly) back a guest fixed counter with a general purpose
* counter, and so KVM must hide fixed counters whose associated
* architectural event are unsupported. On real hardware, this should
--
2.42.0.869.gea05f2083d-goog
