On 8/5/23 02:55, Peter Zijlstra wrote:
+ * Clobbering BP here is mostly ok since GIF will block NMIs and with
+ * the exception of #MC and the kvm_rebooting _ASM_EXTABLE()s below
+ * nothing untoward will happen until BP is restored.
+ *
+ * The kvm_rebooting exceptions should not want to unwind stack, and
+ * while #MV might want to unwind stack, it is ultimately fatal.
+ */
Aside from me not being able to type #MC, I did realize that the
kvm_reboot exception will go outside noinstr code and can hit
tracing/instrumentation and do unwinds from there.
Asynchronously disabling SVM requires an IPI, so kvm_rebooting cannot
change within CLGI/STGI. We can check it after CLGI instead of waiting
for a #GP:
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 956726d867aa..e3755f5eaf81 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -4074,7 +4074,10 @@ static __no_kcsan fastpath_t svm_vcpu_run(struct
kvm_vcpu *vcpu)
if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
x86_spec_ctrl_set_guest(svm->virt_spec_ctrl);
- svm_vcpu_enter_exit(vcpu, spec_ctrl_intercepted);
+ if (unlikely(kvm_rebooting))
+ svm->vmcb->control.exit_code = SVM_EXIT_PAUSE;
+ else
+ svm_vcpu_enter_exit(vcpu, spec_ctrl_intercepted);
if (!static_cpu_has(X86_FEATURE_V_SPEC_CTRL))
x86_spec_ctrl_restore_host(svm->virt_spec_ctrl);
diff --git a/arch/x86/kvm/svm/vmenter.S b/arch/x86/kvm/svm/vmenter.S
index 8e8295e774f0..34641b3a6823 100644
--- a/arch/x86/kvm/svm/vmenter.S
+++ b/arch/x86/kvm/svm/vmenter.S
@@ -270,23 +270,12 @@ SYM_FUNC_START(__svm_vcpu_run)
RESTORE_GUEST_SPEC_CTRL_BODY
RESTORE_HOST_SPEC_CTRL_BODY
-10: cmpb $0, kvm_rebooting
- jne 2b
- ud2
-30: cmpb $0, kvm_rebooting
- jne 4b
- ud2
-50: cmpb $0, kvm_rebooting
- jne 6b
- ud2
-70: cmpb $0, kvm_rebooting
- jne 8b
- ud2
+10: ud2
_ASM_EXTABLE(1b, 10b)
- _ASM_EXTABLE(3b, 30b)
- _ASM_EXTABLE(5b, 50b)
- _ASM_EXTABLE(7b, 70b)
+ _ASM_EXTABLE(3b, 10b)
+ _ASM_EXTABLE(5b, 10b)
+ _ASM_EXTABLE(7b, 10b)
SYM_FUNC_END(__svm_vcpu_run)
Paolo
