On Mon, Mar 01, 2021, Kai Huang wrote:
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 50810d471462..df8e338267aa 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -1570,12 +1570,18 @@ static int vmx_rtit_ctl_check(struct kvm_vcpu *vcpu,
> u64 data)
>
> static bool vmx_can_emulate_instruction(struct kvm_vcpu *vcpu, void *insn,
> int insn_len)
> {
> + if (to_vmx(vcpu)->exit_reason.enclave_mode) {
> + kvm_queue_exception(vcpu, UD_VECTOR);
Rereading my own code, I think it would be a good idea to add a comment here
explaining that injecting #UD is technically wrong, but avoids giving guest
userspace an easy way to DoS the guest. The EPT misconfig is a good example;
guest userspace could have executed a simple MOV <reg>, <mem> instruction, in
which case injecting a #UD is bizarre behavior. But, the alternative is exiting
to userspace with KVM_INTERNAL_ERROR_EMULATION, which is all but guaranteed to
kill the guest.
If KVM, specifically handle_emulation_failure(), ever gains a more sophisticated
mechanism for handling userspace emulation errors, this should be updated too.
/*
* Emulation of instructions in SGX enclaves is impossible as RIP does
* not point tthe failing instruction, and even if it did, the code
* stream is inaccessible. Inject #UD instead of exiting to userspace
* so that guest userspace can't DoS the guest simply by triggering
* emulation (enclaves are CPL3 only).
*/
> + return false;
> + }
> return true;
> }
...
> @@ -5384,6 +5415,9 @@ static int handle_ept_misconfig(struct kvm_vcpu *vcpu)
> {
> gpa_t gpa;
>
> + if (!vmx_can_emulate_instruction(vcpu, NULL, 0))
> + return 1;
> +
> /*
> * A nested guest cannot optimize MMIO vmexits, because we have an
> * nGPA here instead of the required GPA.
> --
> 2.29.2
>
