On Mon, Jan 13, 2020 at 08:55:46PM +0200, Adalbert Lazăr wrote:
> On Mon, 13 Jan 2020 09:33:58 -0800, Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote:
> > On Mon, Jan 13, 2020 at 04:10:50PM +0800, Yang Weijiang wrote:
> > > On Fri, Jan 10, 2020 at 10:04:59AM -0800, Sean Christopherson wrote:
> > > > On Thu, Jan 02, 2020 at 02:13:15PM +0800, Yang Weijiang wrote:
> > > > > @@ -3585,7 +3602,30 @@ static bool fast_page_fault(struct kvm_vcpu *vcpu, gva_t gva, int level,
> > > > > if ((error_code & PFERR_WRITE_MASK) &&
> > > > > spte_can_locklessly_be_made_writable(spte))
> > > > > {
> > > > > - new_spte |= PT_WRITABLE_MASK;
> > > > > + /*
> > > > > + * Record write protect fault caused by
> > > > > + * Sub-page Protection, let VMI decide
> > > > > + * the next step.
> > > > > + */
> > > > > + if (spte & PT_SPP_MASK) {
> > > > > + int len = kvm_x86_ops->get_inst_len(vcpu);
> > > >
> > > > There's got to be a better way to handle SPP exits than adding a helper
> > > > to retrieve the instruction length.
> > > >
> > > The fault instruction was skipped by kvm_skip_emulated_instruction()
> > > before, but Paolo suggested leave the re-do or skip option to user-space
> > > to make it flexible for write protection or write tracking, so return
> > > length to user-space.
> >
> > Sorry, my comment was unclear. I have no objection to punting the fault
> > to userspace, it's the mechanics of how it's done that I dislike.
> >
> > Specifically, (a) using run->exit_reason to propagate the SPP exit up the
> > stack, e.g. instead of modifying affected call stacks to play nice with
> > any exit to userspace, (b) assuming ->get_insn_len() will always be
> > accurate, e.g. see the various caveats in skip_emulated_instruction() for
> > both VMX and SVM, and (c) duplicating the state capture code in every
> > location that can encounter a SPP fault.
> >
> > What I'm hoping is that it's possible to modify the call stacks to
> > explicitly propagate an exit to userspace and/or SPP fault, and shove all
> > the state capture into a common location, e.g. handle_ept_violation().
> >
> > Side topic, assuming the userspace VMI is going to be instrospecting the
> > faulting instruction, won't it decode the instruction? I.e. calculate
> > the instruction length anyways?
>
> Indeed, we decode the instruction from userspace. I don't know if the
> instruction length helps other projects. Added Tamas and Mathieu.
>
> In our last VMI API proposal, the breakpoint event had the instruction
> length sent to userspace, but I can't remember why.
INT3 is trap-like, i.e. the VM-Exit occurs after the instruction retires.
It's impossible for software to know how far to unwind RIP without the
instruction length being provided by hardware/KVM, e.g. if the guest is
being silly and prepends ignored prefixes on the INT3.
Self-aware software has a priori knowledge of what's being patched in,
and practically speaking I don't any well-behaved sane software uses
prefixes with INT3, but from a VMM's perspective it's legal and possible.
>
> https://lore.kernel.org/kvm/20190809160047.8319-62-alazar@xxxxxxxxxxxxxxx/
