KVM: x86: do not report a vCPU as preempted outside instruction boundaries

If a vCPU is outside guest mode and is scheduled out, it might be in the
process of making a memory access.  A problem occurs if another vCPU uses
the PV TLB flush feature during the period when the vCPU is scheduled
out, and a virtual address has already been translated but has not yet
been accessed, because this is equivalent to using a stale TLB entry.

To avoid this, only report a vCPU as preempted if sure that the guest
is at an instruction boundary.  A rescheduling request will be delivered
to the host physical CPU as an external interrupt, so for simplicity
consider any vmexit *not* instruction boundary except for external
interrupts.

It would in principle be okay to report the vCPU as preempted also
if it is sleeping in kvm_vcpu_block(): a TLB flush IPI will incur the
vmentry/vmexit overhead unnecessarily, and optimistic spinning is
also unlikely to succeed.  However, leave it for later because right
now kvm_vcpu_check_block() is doing memory accesses.  Even
though the TLB flush issue only applies to virtual memory address,
it's very much preferrable to be conservative.

Reported-by: Jann Horn <jannh@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

4 files changed, 28 insertions, 0 deletions

diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 959d66b9be94..3a240a64ac68 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -653,6 +653,7 @@ struct kvm_vcpu_arch {
 	u64 ia32_misc_enable_msr;
 	u64 smbase;
 	u64 smi_count;
+	bool at_instruction_boundary;
 	bool tpr_access_reporting;
 	bool xsaves_enabled;
 	bool xfd_no_write_intercept;
@@ -1300,6 +1301,8 @@ struct kvm_vcpu_stat {
 	u64 nested_run;
 	u64 directed_yield_attempted;
 	u64 directed_yield_successful;
+	u64 preemption_reported;
+	u64 preemption_other;
 	u64 guest_mode;
 };
 
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 478e6ee81d88..921fcb85a9cd 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -4263,6 +4263,8 @@ out:
 
 static void svm_handle_exit_irqoff(struct kvm_vcpu *vcpu)
 {
+	if (to_svm(vcpu)->vmcb->control.exit_code == SVM_EXIT_INTR)
+		vcpu->arch.at_instruction_boundary = true;
 }
 
 static void svm_sched_in(struct kvm_vcpu *vcpu, int cpu)
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index f5aeade623d6..14e01178a753 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -6547,6 +6547,7 @@ static void handle_external_interrupt_irqoff(struct kvm_vcpu *vcpu)
 		return;
 
 	handle_interrupt_nmi_irqoff(vcpu, gate_offset(desc));
+	vcpu->arch.at_instruction_boundary = true;
 }
 
 static void vmx_handle_exit_irqoff(struct kvm_vcpu *vcpu)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index a8bb635cb76b..25a517206c4d 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -296,6 +296,8 @@ const struct _kvm_stats_desc kvm_vcpu_stats_desc[] = {
 	STATS_DESC_COUNTER(VCPU, nested_run),
 	STATS_DESC_COUNTER(VCPU, directed_yield_attempted),
 	STATS_DESC_COUNTER(VCPU, directed_yield_successful),
+	STATS_DESC_COUNTER(VCPU, preemption_reported),
+	STATS_DESC_COUNTER(VCPU, preemption_other),
 	STATS_DESC_ICOUNTER(VCPU, guest_mode)
 };
 
@@ -4625,6 +4627,19 @@ static void kvm_steal_time_set_preempted(struct kvm_vcpu *vcpu)
 	struct kvm_memslots *slots;
 	static const u8 preempted = KVM_VCPU_PREEMPTED;
 
+	/*
+	 * The vCPU can be marked preempted if and only if the VM-Exit was on
+	 * an instruction boundary and will not trigger guest emulation of any
+	 * kind (see vcpu_run).  Vendor specific code controls (conservatively)
+	 * when this is true, for example allowing the vCPU to be marked
+	 * preempted if and only if the VM-Exit was due to a host interrupt.
+	 */
+	if (!vcpu->arch.at_instruction_boundary) {
+		vcpu->stat.preemption_other++;
+		return;
+	}
+
+	vcpu->stat.preemption_reported++;
 	if (!(vcpu->arch.st.msr_val & KVM_MSR_ENABLED))
 		return;
 
@@ -10424,6 +10439,13 @@ static int vcpu_run(struct kvm_vcpu *vcpu)
 	vcpu->arch.l1tf_flush_l1d = true;
 
 	for (;;) {
+		/*
+		 * If another guest vCPU requests a PV TLB flush in the middle
+		 * of instruction emulation, the rest of the emulation could
+		 * use a stale page translation. Assume that any code after
+		 * this point can start executing an instruction.
+		 */
+		vcpu->arch.at_instruction_boundary = false;
 		if (kvm_vcpu_running(vcpu)) {
 			r = vcpu_enter_guest(vcpu);
 		} else {