Re: [RESEND PATCH v10 06/10] vmx: spp: Set up SPP paging table at vmentry/vmexit

From: Sean Christopherson <sean.j.christopherson@intel.com>
To: Yang Weijiang <weijiang.yang@intel.com>
Cc: kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
	pbonzini@redhat.com, jmattson@google.com,
	yu.c.zhang@linux.intel.com, alazar@bitdefender.com,
	edwin.zhai@intel.com
Subject: Re: [RESEND PATCH v10 06/10] vmx: spp: Set up SPP paging table at vmentry/vmexit
Date: Mon, 13 Jan 2020 09:33:58 -0800	[thread overview]
Message-ID: <20200113173358.GC1175@linux.intel.com> (raw)
In-Reply-To: <20200113081050.GF12253@local-michael-cet-test.sh.intel.com>

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?