Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Vitaly Kuznetsov <vkuznets@redhat.com>
To:     Jim Mattson <jmattson@google.com>,
        Paolo Bonzini <pbonzini@redhat.com>
Cc:     kvm list <kvm@vger.kernel.org>,
        Radim =?utf-8?B?S3LEjW3DocWZ?= <rkrcmar@redhat.com>,
        Liran Alon <liran.alon@oracle.com>,
        LKML <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH] x86/kvm/nVMX: tweak shadow fields
In-Reply-To: <CALMp9eQ4493qH4sA4e5UibaHhatEKnztGNBO2s40jc2jXXCdUQ@mail.gmail.com>
References: <20181019141603.15995-1-vkuznets@redhat.com> <31279dfd-d0a1-3720-46a2-52395a124057@redhat.com> <CALMp9eQ4493qH4sA4e5UibaHhatEKnztGNBO2s40jc2jXXCdUQ@mail.gmail.com>
Date:   Mon, 12 Nov 2018 15:39:40 +0100
Message-ID: <87pnvas6kz.fsf@vitty.brq.redhat.com>
MIME-Version: 1.0
Content-Type: text/plain
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Jim Mattson <jmattson@google.com> writes:

> I'm not convinced that the "one size fits all" and "context-free"
> approaches to VMCS shadowing are terribly effective.
>
> For example, we never shadow VMX_INSTRUCTION_INFO, but if we just
> reflected an exit to L1 for which that field is defined, there's
> probably a good chance that L1 will use it. We always shadow
> VM_EXIT_INTR_INFO, but if we didn't just reflect exit reason 0 to L1,
> it's not likely to be read. If the L2 guest is in legacy mode or
> compatibility mode, L1 is much more likely to be interested in the
> contents of the descriptor cache than if the guest is in 64-bit mode.
>
> Some hypervisors write TSC_OFFSET quite frequently. Others rarely.
> Last time I checked (it's been a while), VirtualBox was always
> interested in everything. :-) Kvm, Hyper-V, VMware, VirtualBox,
> Parallels...they all have different patterns, and they change from
> release to release.
>
> Is it worth having a set of VMCS shadowing bitmaps per-vCPU, in order
> to make better use of this feature?

Per CPU or not, to improve the feature we'll probably need some sort of
an 'adaptive' algorithm picking which fields to shadow. I haven't
thought this through, especially read/write shadowing, but we can
probably start with an empty bitmap and later shadow it when we get over
some threshold of vmread/vmwrite exits we enabling shadowing. The
question is when we un-shadow it. For example, we can un-shadow a field
for writing every time we see it was not changed between two exits to L0
(so we're trying to write the same value to vmcs12).

-- 
Vitaly