On 22/08/2018 23:52, Sean Christopherson wrote:
> reexecute_instruction(), added by commit a6f177efaa58 ("KVM: Reenter
> guest after emulation failure if due to access to non-mmio address"),
> is intended to return true if and only if emulation failed due to a
> write to a shadowed page, i.e. MMIO emulation that happens to fail
> should never be retried. Set EMULTYPE_NO_REEXECUTE when we know we
> are emulating a (alleged) MMIO access to document that we should not
> re-execute the instruction.
>
> This is purely a documentation update as reexecute_instruction() does
> not re-execute instructions whose faulting GPA corresponds to an error
> or no-slot PFN, e.g. EMULTYPE_NO_REEXECUTE was added by commit
> 991eebf9f8e5 ("KVM: VMX: do not try to reexecute failed instruction
> while emulating invalid guest state") for reasons entirely unrelated
> to MMIO. Case in point, mmio_info_in_cache() always returns false for
> nested VMs, e.g. EMULTYPE_NO_REEXECUTE won't be set if MMIO emulation
> for a nested VM is triggered outside of the handle_mmio_page_fault()
> path.
>
> Signed-off-by: Sean Christopherson <sean.j.christopherson@xxxxxxxxx>
> ---
> arch/x86/kvm/mmu.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index a282321329b5..1c152fd9fc7c 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -5231,7 +5231,7 @@ int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u64 error_code,
> if (unlikely(error_code & PFERR_RSVD_MASK)) {
> r = handle_mmio_page_fault(vcpu, cr2, direct);
> if (r == RET_PF_EMULATE) {
> - emulation_type = 0;
> + emulation_type = EMULTYPE_NO_REEXECUTE;
> goto emulate;
> }
> }
> @@ -5261,7 +5261,7 @@ int kvm_mmu_page_fault(struct kvm_vcpu *vcpu, gva_t cr2, u64 error_code,
> }
>
> if (mmio_info_in_cache(vcpu, cr2, direct))
> - emulation_type = 0;
> + emulation_type = EMULTYPE_NO_REEXECUTE;
> emulate:
> /*
> * On AMD platforms, under certain conditions insn_len may be zero on #NPF.
>
Even better, EMULTYPE_NO_REEXECUTE is never set if EMULTYPE_RETRY is
set, and vice versa. The sole exception is rsm_interception in svm.c
that should actually specify EMULTYPE_NO_REEXECUTE but doesn't.
So it seems to me that they can be unified to a single EMULTYPE_RETRY,
which makes sense since they do basically the same thing---only
EMULTYPE_NO_REEXECUTE is a bit more desperado because decoding has failed.
Paolo
