Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20210424004645.3950558-1-seanjc@google.com> <20210424004645.3950558-43-seanjc@google.com>
 <e2974b79-a6e5-81be-2adb-456f114391da@redhat.com> <YKwomNuTEwgf4Xt0@google.com>
In-Reply-To: <YKwomNuTEwgf4Xt0@google.com>
From:   Jim Mattson <jmattson@google.com>
Date:   Mon, 24 May 2021 15:48:04 -0700
Message-ID: <CALMp9eSsOw0=n4-rn5B1A_T9nYBB0UkXWQ+oOJNx6ammfJ6Q-A@mail.gmail.com>
Subject: Re: [PATCH 42/43] KVM: VMX: Drop VMWRITEs to zero fields at vCPU RESET
To:     Sean Christopherson <seanjc@google.com>
Cc:     Paolo Bonzini <pbonzini@redhat.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Joerg Roedel <joro@8bytes.org>, kvm list <kvm@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, May 24, 2021 at 3:28 PM Sean Christopherson <seanjc@google.com> wrote:
>
> On Mon, May 24, 2021, Paolo Bonzini wrote:
> > On 24/04/21 02:46, Sean Christopherson wrote:
> > > Don't waste time writing zeros via VMWRITE during vCPU RESET, the VMCS
> > > is zero allocated.
> >
> > Is this guaranteed to be valid, or could the VMCS in principle use some
> > weird encoding? (Like it does for the access rights, even though this does
> > not matter for this patch).
>
> Phooey.  In principle, the CPU can do whatever it wants, e.g. the SDM states that
> software should never write to the data portion of the VMCS under any circumstance.
>
> In practice, I would be flabbergasted if Intel ever ships a CPU that doesn't play
> nice with zero initiazing the VMCS via software writes.  I'd bet dollars to
> donuts that KVM isn't the only software that relies on that behavior.

It's not just Intel. It's any manufacturer of physical or virtual CPUs
that implement VT-x. Non-architected behavior isn't guaranteed.

> That said, I'm not against switching to VMWRITE for everything, but regardless
> of which route we choose, we should commit to one or the other.  I.e. double down
> on memset() and bet that Intel won't break KVM, or replace the memset() in
> alloc_vmcs_cpu() with a sequence that writes all known (possible?) fields.  The
> current approach of zeroing the memory in software but initializing _some_ fields
> is the worst option, e.g. I highly doubt vmcs01 and vmcs02 do VMWRITE(..., 0) on
> the same fields.

The memset should probably be dropped, unless it is there to prevent
information leakage. However, it is not necessary to VMWRITE all known
(or possible) fields--just those that aren't guarded by an enable bit.