On Wed, Oct 16, 2024, Ivan Orlov wrote:
On 10/16/24 22:05, Sean Christopherson wrote:
On Tue, Oct 15, 2024, Ivan Orlov wrote:
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c index c67e448c6ebd..afd785e7f3a3 100644 --- a/arch/x86/kvm/vmx/vmx.c +++ b/arch/x86/kvm/vmx/vmx.c @@ -6550,19 +6550,10 @@ static int __vmx_handle_exit(struct kvm_vcpu *vcpu, fastpath_t exit_fastpath) exit_reason.basic != EXIT_REASON_APIC_ACCESS && exit_reason.basic != EXIT_REASON_TASK_SWITCH && exit_reason.basic != EXIT_REASON_NOTIFY)) {
int ndata = 3;
gpa_t gpa = vmcs_read64(GUEST_PHYSICAL_ADDRESS);bool is_mmio = exit_reason.basic == EXIT_REASON_EPT_MISCONFIG;There's no need for is_mmio, just pass INVALID_GPA when the GPA isn't known.
Ah alright, then we definitely don't need an is_mmio field. I assume we can't do MMIO at GPA=0, right?
Wrong :-)
Then getting rid of `is_mmio` will make distinguishing between vectoring error due to MMIO with GPA=0 and non-mmio vectoring error quite hard for the error reporti
Passing INVALID_GPA into the userspace due to non-mmio vectoring error will change the existing internal.data order, but I can do it if it's fine. Sorry for nitpicking :)
KVM's existing ABI is rather awful, though arguably the intent was that there is no ABI, i.e. that KVM is dumping info to try to be helpful. E.g. the existing behavior is that data[3] contains a GPA only for EPT_MISCONFIG, but for everything else, data[3] contains last_vmentry_cpu.
And because it's so awful, I doubt any userspace actually has code that acts on the layout of data[]. So, I suspect we can do the simple and sane thing, and fill data[3] with -1ull if the GPA is invalid, and then document that that's the behavior (if we're feeling generous).