Re: [PATCH v2 8/9] KVM: x86, SVM: isolate vcpu->arch.dr6 from vmcb->save.dr6

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 v2 8/9] KVM: x86, SVM: isolate vcpu->arch.dr6 from vmcb->save.dr6
Date: Fri, 8 May 2020 11:32:10 -0400	[thread overview]
Message-ID: <20200508153210.GZ228260@xz-x1> (raw)
In-Reply-To: <dd8eb45b-4556-6aaa-0061-11b9124020b1@redhat.com>

On Fri, May 08, 2020 at 12:33:57AM +0200, Paolo Bonzini wrote:
> On 07/05/20 21:28, Peter Xu wrote:
> >> -	svm->vcpu.arch.dr6 = dr6;
> >> +	WARN_ON(svm->vcpu.arch.switch_db_regs & KVM_DEBUGREG_WONT_EXIT);
> >> +	svm->vcpu.arch.dr6 &= ~(DR_TRAP_BITS | DR6_RTM);
> >> +	svm->vcpu.arch.dr6 |= dr6 & ~DR6_FIXED_1;
> > I failed to figure out what the above calculation is going to do... 
> 
> The calculation is merging the cause of the #DB with the guest DR6.
> It's basically the same effect as kvm_deliver_exception_payload. The
> payload has DR6_RTM flipped compared to DR6, so you have the following
> simplfications:
> 
> 	payload = (dr6 ^ DR6_RTM) & ~DR6_FIXED_1;
> 	/* This is kvm_deliver_exception_payload: */
>         vcpu->arch.dr6 &= ~DR_TRAP_BITS;
>         vcpu->arch.dr6 |= DR6_RTM;
> 	/* copy dr6 bits other than RTM */
>         vcpu->arch.dr6 |= payload;
> 	/* copy flipped RTM bit */
>         vcpu->arch.dr6 ^= payload & DR6_RTM;
> 
> ->
> 
> 	payload = (dr6 ^ DR6_RTM) & ~DR6_FIXED_1;
> 	/* clear RTM here, so that we can OR it below */
>         vcpu->arch.dr6 &= ~(DR_TRAP_BITS | DR6_RTM);
> 	/* copy dr6 bits other than RTM */
>         vcpu->arch.dr6 |= payload & ~DR6_RTM;
> 	/* copy flipped RTM bit */
>         vcpu->arch.dr6 |= (payload ^ DR6_RTM) & DR6_RTM;
> 
> ->
> 
> 	/* we can drop the double XOR of DR6_RTM */
> 	dr6 &= ~DR6_FIXED_1;
>         vcpu->arch.dr6 &= ~(DR_TRAP_BITS | DR6_RTM);
>         vcpu->arch.dr6 |= dr6 & ~DR6_RTM;
>         vcpu->arch.dr6 |= dr6 & DR6_RTM;
> 
> ->
> 
> 	/* we can do the two ORs with a single operation */
>         vcpu->arch.dr6 &= ~(DR_TRAP_BITS | DR6_RTM);
>         vcpu->arch.dr6 |= dr6 & ~DR6_FIXED_1;

Oh that's quite some math. :) Thanks Paolo!

Shall we introduce a helper for both kvm_deliver_exception_payload and here
(e.g. kvm_merge_dr6)?  Also, wondering whether this could be a bit easier to
follow by defining:

/*
 * These bits could be kept being set until the next #DB if not
 * explicitly cleared.
 */
#define  DR6_CARRY_OVER_BITS  (DR6_BT | DR6_BS | DR6_BD)

Then the imho above calculation could also become:

    vcpu->arch.dr6 = (vcpu->arch.dr6 & DR6_CARRY_OVER_BITS) | save.dr6;

What do you think?

> 
> > E.g., I
> > think the old "BT|BS|BD" bits in the old arch.dr6 cache will be leftover even
> > if none of them is set in save.dr6, while we shouldn't?
> 
> Those bits should be kept; this is covered for example by the "hw
> breakpoint (test that dr6.BS is not cleared)" testcase in kvm-unit-tests
> x86/debug.c.

Right.  Thanks!

-- 
Peter Xu