KVM: x86/pmu: Explicitly check NMI from guest to reducee false positives

Explicitly check that the source of external interrupt is indeed an NMI
in kvm_arch_pmi_in_guest(), which reduces perf-kvm false positive samples
(host samples labelled as guest samples) generated by perf/core NMI mode
if an NMI arrives after VM-Exit, but before kvm_after_interrupt():

 # test: perf-record + cpu-cycles:HP (which collects host-only precise samples)
 # Symbol                                   Overhead       sys       usr  guest sys  guest usr
 # .......................................  ........  ........  ........  .........  .........
 #
 # Before:
   [g] entry_SYSCALL_64                       24.63%     0.00%     0.00%     24.63%      0.00%
   [g] syscall_return_via_sysret              23.23%     0.00%     0.00%     23.23%      0.00%
   [g] files_lookup_fd_raw                     6.35%     0.00%     0.00%      6.35%      0.00%
 # After:
   [k] perf_adjust_freq_unthr_context         57.23%    57.23%     0.00%      0.00%      0.00%
   [k] __vmx_vcpu_run                          4.09%     4.09%     0.00%      0.00%      0.00%
   [k] vmx_update_host_rsp                     3.17%     3.17%     0.00%      0.00%      0.00%

In the above case, perf records the samples labelled '[g]', the RIPs behind
the weird samples are actually being queried by perf_instruction_pointer()
after determining whether it's in GUEST state or not, and here's the issue:

If VM-Exit is caused by a non-NMI interrupt (such as hrtimer_interrupt) and
at least one PMU counter is enabled on host, the kvm_arch_pmi_in_guest()
will remain true (KVM_HANDLING_IRQ is set) until kvm_before_interrupt().

During this window, if a PMI occurs on host (since the KVM instructions on
host are being executed), the control flow, with the help of the host NMI
context, will be transferred to perf/core to generate performance samples,
thus perf_instruction_pointer() and perf_guest_get_ip() is called.

Since kvm_arch_pmi_in_guest() only checks if there is an interrupt, it may
cause perf/core to mistakenly assume that the source RIP of the host NMI
belongs to the guest world and use perf_guest_get_ip() to get the RIP of
a vCPU that has already exited by a non-NMI interrupt.

Error samples are recorded and presented to the end-user via perf-report.
Such false positive samples could be eliminated by explicitly determining
if the exit reason is KVM_HANDLING_NMI.

Note that when VM-exit is indeed triggered by PMI and before HANDLING_NMI
is cleared, it's also still possible that another PMI is generated on host.
Also for perf/core timer mode, the false positives are still possible since
those non-NMI sources of interrupts are not always being used by perf/core.

For events that are host-only, perf/core can and should eliminate false
positives by checking event->attr.exclude_guest, i.e. events that are
configured to exclude KVM guests should never fire in the guest.

Events that are configured to count host and guest are trickier, perhaps
impossible to handle with 100% accuracy?  And regardless of what accuracy
is provided by perf/core, improving KVM's accuracy is cheap and easy, with
no real downsides.

Fixes: dd60d217062f ("KVM: x86: Fix perf timer mode IP reporting")
Signed-off-by: Like Xu <likexu@tencent.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20231206032054.55070-1-likexu@tencent.com
[sean: massage changelog, squash !!in_nmi() fixup from Like]
Signed-off-by: Sean Christopherson <seanjc@google.com>

1 files changed, 0 insertions, 6 deletions

diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
index 2f7e19166658..4dc38092d599 100644
--- a/arch/x86/kvm/x86.h
+++ b/arch/x86/kvm/x86.h
@@ -431,12 +431,6 @@ static inline bool kvm_notify_vmexit_enabled(struct kvm *kvm)
 	return kvm->arch.notify_vmexit_flags & KVM_X86_NOTIFY_VMEXIT_ENABLED;
 }
 
-enum kvm_intr_type {
-	/* Values are arbitrary, but must be non-zero. */
-	KVM_HANDLING_IRQ = 1,
-	KVM_HANDLING_NMI,
-};
-
 static __always_inline void kvm_before_interrupt(struct kvm_vcpu *vcpu,
 						 enum kvm_intr_type intr)
 {