Re: [PATCH 9/9] KVM: VMX: pass correct DR6 for GD userspace exit

From: Peter Xu <peterx@redhat.com>
To: Paolo Bonzini <pbonzini@redhat.com>
Cc: linux-kernel@vger.kernel.org, kvm@vger.kernel.org
Subject: Re: [PATCH 9/9] KVM: VMX: pass correct DR6 for GD userspace exit
Date: Thu, 7 May 2020 14:05:08 -0400	[thread overview]
Message-ID: <20200507180508.GH228260@xz-x1> (raw)
In-Reply-To: <db06ffa7-1e3c-14e5-28b8-5053f4383ecf@redhat.com>

On Thu, May 07, 2020 at 07:42:25PM +0200, Paolo Bonzini wrote:
> On 07/05/20 18:38, Peter Xu wrote:
> > On Thu, May 07, 2020 at 06:21:18PM +0200, Paolo Bonzini wrote:
> >> On 07/05/20 18:18, Peter Xu wrote:
> >>>>  		if (vcpu->guest_debug & KVM_GUESTDBG_USE_HW_BP) {
> >>>> -			vcpu->run->debug.arch.dr6 = vcpu->arch.dr6;
> >>>> +			vcpu->run->debug.arch.dr6 = DR6_BD | DR6_RTM | DR6_FIXED_1;
> >>> After a second thought I'm thinking whether it would be okay to have BS set in
> >>> that test case.  I just remembered there's a test case in the kvm-unit-test
> >>> that checks explicitly against BS leftover as long as dr6 is not cleared
> >>> explicitly by the guest code, while the spec seems to have no explicit
> >>> description on this case.
> >>
> >> Yes, I noticed that test as well.  But I don't like having different
> >> behavior for Intel and AMD, and the Intel behavior is more sensible.
> >> Also...
> > 
> > Do you mean the AMD behavior is more sensible instead? :)
> 
> No, I mean within the context of KVM_EXIT_DEBUG: the Intel behavior is
> to only include the latest debug exception in kvm_run's DR6 field, while
> the AMD behavior would be to include all of them.  This was an
> implementation detail (it happens because Intel sets kvm_run's DR6 from
> the exit qualification of #DB), but it's more sensible too.
> 
> In addition:
> 
> * AMD was completely broken until this week, so the behavior of
> KVM_EXIT_DEBUG is defined de facto by kvm_intel.ko.  Userspace has not
> been required to set DR6 with KVM_SET_GUEST_DEBUG, and since we can
> emulate that on AMD, we should.
> 
> * we have to fix anyway the fact that on AMD a KVM_EXIT_DEBUG is
> clobbering the contents of the guest's DR6
> 
> >>> Intead of above, I'm thinking whether we should allow the userspace to also
> >>> change dr6 with the KVM_SET_GUEST_DEBUG ioctl when they wanted to (right now
> >>> iiuc dr6 from userspace is completely ignored), instead of offering a fake dr6.
> >>> Or to make it simple, maybe we can just check BD bit only?
> >>
> >> ... I'm afraid that this would be a backwards-incompatible change, and
> >> it would require changes in userspace.  If you look at v2, emulating the
> >> Intel behavior in AMD turns out to be self-contained and relatively
> >> elegant (will be better when we finish cleaning up nested SVM).
> > 
> > I'm still trying to read the other patches (I need some more digest because I'm
> > even less familiar with nested...).  I agree that it would be good to keep the
> > same behavior across Intel/AMD.  Actually that also does not violate Intel spec
> > because the AMD one is stricter.
> 
> Again, careful---we're talking about KVM_EXIT_DEBUG, not the #DB exception.

OK I get your point now.  Thanks,

-- 
Peter Xu