On 11/19/22 13:29, Dongli Zhang wrote:
The "perf stat" at the VM side still works even we set "-cpu host,-pmu" in
the QEMU command line. That is, neither "-cpu host,-pmu" nor "-cpu EPYC"
could disable the pmu virtualization in an AMD environment.
We still see below at VM kernel side ...
[ 0.510611] Performance Events: Fam17h+ core perfctr, AMD PMU driver.
... although we expect something like below.
[ 0.596381] Performance Events: PMU not available due to virtualization,
using software events only.
[ 0.600972] NMI watchdog: Perf NMI watchdog permanently disabled
This is because the AMD pmu (v1) does not rely on cpuid to decide if the
pmu virtualization is supported.
We disable KVM_CAP_PMU_CAPABILITY if the 'pmu' is disabled in the vcpu
properties.
Cc: Joe Jin <joe....@oracle.com>
Signed-off-by: Dongli Zhang <dongli.zh...@oracle.com>
---
target/i386/kvm/kvm.c | 17 +++++++++++++++++
1 file changed, 17 insertions(+)
diff --git a/target/i386/kvm/kvm.c b/target/i386/kvm/kvm.c
index 8fec0bc5b5..0b1226ff7f 100644
--- a/target/i386/kvm/kvm.c
+++ b/target/i386/kvm/kvm.c
@@ -137,6 +137,8 @@ static int has_triple_fault_event;
static bool has_msr_mcg_ext_ctl;
+static int has_pmu_cap;
+
static struct kvm_cpuid2 *cpuid_cache;
static struct kvm_cpuid2 *hv_cpuid_cache;
static struct kvm_msr_list *kvm_feature_msrs;
@@ -1725,6 +1727,19 @@ static void kvm_init_nested_state(CPUX86State *env)
void kvm_arch_pre_create_vcpu(CPUState *cs)
{
+ X86CPU *cpu = X86_CPU(cs);
+ int ret;
+
+ if (has_pmu_cap && !cpu->enable_pmu) {
+ ret = kvm_vm_enable_cap(kvm_state, KVM_CAP_PMU_CAPABILITY, 0,
+ KVM_PMU_CAP_DISABLE);
+ if (ret < 0) {
+ error_report("kvm: Failed to disable pmu cap: %s",
+ strerror(-ret));
+ }
+
+ has_pmu_cap = 0;
+ }
}
int kvm_arch_init_vcpu(CPUState *cs)
@@ -2517,6 +2532,8 @@ int kvm_arch_init(MachineState *ms, KVMState *s)
}
}
+ has_pmu_cap = kvm_check_extension(s, KVM_CAP_PMU_CAPABILITY);
+
ret = kvm_get_supported_msrs(s);
if (ret < 0) {
return ret;
This patch is very important in particular.
It boosts performance of any single VMexit
is 13% for AMD. Intel is being measured.
At my opinion v1 of the patch is better that
version 2. We should not introduce any
new capability but disable PMU if we can
while it is disabled according to the configuration.
The discussion about performance improvement
is here
https://lore.kernel.org/lkml/zu2d3f6kc0mdz...@google.com/T/
Den
