On Fri, 2019-10-04 at 09:42 +0200, Paolo Bonzini wrote:
> On 03/10/19 23:23, Rick Edgecombe wrote:
> > + if (!vcpu->arch.gva_available)
> > + return 0;
>
> Please return RET_PF_* constants, RET_PF_EMULATE here.
Ok.
> > + if (error_code & PFERR_WRITE_MASK)
> > + fault_error_code |= X86_PF_WRITE;
> > +
> > + fault.vector = PF_VECTOR;
> > + fault.error_code_valid = true;
> > + fault.error_code = fault_error_code;
> > + fault.nested_page_fault = false;
> > + fault.address = vcpu->arch.gva_val;
> > + fault.async_page_fault = true;
>
> Not an async page fault.
Right.
> > + kvm_inject_page_fault(vcpu, &fault);
> > +
> > + return 1;
>
> Here you would return RET_PF_RETRY - you've injected the page fault and
> all that's left to do is reenter execution of the vCPU.
>
> [...]
>
> > + if (unlikely(vcpu->arch.xo_fault)) {
> > + /*
> > + * If not enough information to inject the fault,
> > + * emulate to figure it out and emulate the PF.
> > + */
> > + if (!try_inject_exec_only_pf(vcpu, error_code))
> > + return RET_PF_EMULATE;
> > +
> > + return 1;
> > + }
>
> Returning 1 is wrong, it's also RET_PF_EMULATE. If you change
> try_inject_exec_only_pf return values to RET_PF_*, you can simply return
> the value of try_inject_exec_only_pf(vcpu, error_code).
Oh right! I must have broken this at some point. Thanks.
> That said, I wonder if it's better to just handle this in
> handle_ept_violation. Basically, if bits 5:3 of the exit qualification
> are 100 you can bypass the whole mmu.c page fault handling and just
> inject an exec-only page fault.
>
> Thanks,
>
> Paolo
Hmm, that could be cleaner. I'll see how it fits together when I fix the nested
case, since some of that logic looks to be in mmu.c.
Thanks,
Rick
