KVM: x86/svm: Clear reserved bits written to PerfEvtSeln MSRs
authorJim Mattson <jmattson@google.com>
Sat, 26 Feb 2022 23:41:31 +0000 (15:41 -0800)
committerPaolo Bonzini <pbonzini@redhat.com>
Sat, 2 Apr 2022 09:34:45 +0000 (05:34 -0400)
AMD EPYC CPUs never raise a #GP for a WRMSR to a PerfEvtSeln MSR. Some
reserved bits are cleared, and some are not. Specifically, on
Zen3/Milan, bits 19 and 42 are not cleared.

When emulating such a WRMSR, KVM should not synthesize a #GP,
regardless of which bits are set. However, undocumented bits should
not be passed through to the hardware MSR. So, rather than checking
for reserved bits and synthesizing a #GP, just clear the reserved
bits.

This may seem pedantic, but since KVM currently does not support the
"Host/Guest Only" bits (41:40), it is necessary to clear these bits
rather than synthesizing #GP, because some popular guests (e.g Linux)
will set the "Host Only" bit even on CPUs that don't support
EFER.SVME, and they don't expect a #GP.

For example,

root@Ubuntu1804:~# perf stat -e r26 -a sleep 1

 Performance counter stats for 'system wide':

                 0      r26

       1.001070977 seconds time elapsed

Feb 23 03:59:58 Ubuntu1804 kernel: [  405.379957] unchecked MSR access error: WRMSR to 0xc0010200 (tried to write 0x0000020000130026) at rIP: 0xffffffff9b276a28 (native_write_msr+0x8/0x30)
Feb 23 03:59:58 Ubuntu1804 kernel: [  405.379958] Call Trace:
Feb 23 03:59:58 Ubuntu1804 kernel: [  405.379963]  amd_pmu_disable_event+0x27/0x90

Fixes: ca724305a2b0 ("KVM: x86/vPMU: Implement AMD vPMU code for KVM")
Reported-by: Lotus Fenn <lotusf@google.com>
Signed-off-by: Jim Mattson <jmattson@google.com>
Reviewed-by: Like Xu <likexu@tencent.com>
Reviewed-by: David Dunn <daviddunn@google.com>
Message-Id: <20220226234131.2167175-1-jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
arch/x86/kvm/svm/pmu.c

index e4163f7..24eb935 100644 (file)
@@ -262,12 +262,10 @@ static int amd_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
        /* MSR_EVNTSELn */
        pmc = get_gp_pmc_amd(pmu, msr, PMU_TYPE_EVNTSEL);
        if (pmc) {
-               if (data == pmc->eventsel)
-                       return 0;
-               if (!(data & pmu->reserved_bits)) {
+               data &= ~pmu->reserved_bits;
+               if (data != pmc->eventsel)
                        reprogram_gp_counter(pmc, data);
-                       return 0;
-               }
+               return 0;
        }
 
        return 1;