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: <20200902125935.20646-1-graf@amazon.com> <20200902125935.20646-2-graf@amazon.com>
 <CAAAPnDFGD8+5KBCLKERrH0hajHEwU9UdEEGqp3RZu3Lws+5rmw@mail.gmail.com> <186ccace-2fad-3db3-0848-cd272b1a64ba@amazon.com>
In-Reply-To: <186ccace-2fad-3db3-0848-cd272b1a64ba@amazon.com>
From:   Aaron Lewis <aaronlewis@google.com>
Date:   Thu, 17 Sep 2020 12:38:36 -0700
Message-ID: <CAAAPnDFxR8yeB0sq4ZMRoZRO4QycZsBiKzaShGwMWE_0RM6Aow@mail.gmail.com>
Subject: Re: [PATCH v6 1/7] KVM: x86: Deflect unknown MSR accesses to user space
To:     Alexander Graf <graf@amazon.com>
Cc:     Paolo Bonzini <pbonzini@redhat.com>,
        Jonathan Corbet <corbet@lwn.net>,
        Sean Christopherson <sean.j.christopherson@intel.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>,
        KarimAllah Raslan <karahmed@amazon.de>,
        Dan Carpenter <dan.carpenter@oracle.com>,
        kvm list <kvm@vger.kernel.org>, linux-doc@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

> >> +The "reason" field specifies why the MSR trap occurred. User space will only
> >> +receive MSR exit traps when a particular reason was requested during through
> >> +ENABLE_CAP. Currently valid exit reasons are:
> >> +
> >> +       KVM_MSR_EXIT_REASON_INVAL - access to invalid MSRs or reserved bits
> >
> >
> > Can we also have ENOENT?
> >          KVM_MSR_EXIT_REASON_ENOENT - Unknown MSR
>
> I tried to add that at first, but it gets tricky really fast. Why should
> user space have a vested interest in differentiating between "MSR is not
> implemented" and "MSR is guarded by a CPUID flag and thus not handled"
> or "MSR is guarded by a CAP"?
>
> The more details we reveal, the more likely we're to break ABI
> compatibility.
>

I don't suspect we will ever have a stable ABI here, whether we split
the two error values or not.  But there could be value in it.
Consider an MSR that raises #GP if any bit in the high dword is set.
KVM version 0 knows nothing about this MSR, but KVM version 1
implements it properly.  Assuming ignore_msrs=0 and error codes:
EINVAL (invalid argument, should raise a #GP) and ENOTSUP (seems like
a better name than ENOENT, Unknown MSR).

With Just EINVAL: KVM version 0 will always exit to userspace if
EINVAL is requested (it needs help). KVM version 1 will exit to
userspace only for illegal accesses if EINVAL is requested (but it
doesn't really need help).
With EINVAL and ENOTSUP: KVM version 0 will always exit to userspace
if ENOTSUP is requested (it needs help). KVM version 1 will not exit
to userspace if ENOTSUP is requested.

If you want to implement ignore_msrs in userspace it seems much easier
with the second approach, and I think all you'd have to do is return
-ENOTSUP from kvm_msr_ignored_check() instead of returning 1.