Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 25 May 2023 15:30:11 -0700
In-Reply-To: <2d7f7f80-278d-9fcf-cfc4-c433e95d9842@gmail.com>
Mime-Version: 1.0
References: <20230427095333.35038-1-abusse@amazon.com> <2d7f7f80-278d-9fcf-cfc4-c433e95d9842@gmail.com>
Message-ID: <ZG/hc+9/2BraMrZB@google.com>
Subject: Re: [PATCH 0/2] KVM support for Intel PMU v5 fixed function PMC bitmap
From:   Sean Christopherson <seanjc@google.com>
To:     Like Xu <like.xu.linux@gmail.com>
Cc:     Anselm Busse <abusse@amazon.com>, dwmw@amazon.co.uk,
        hborghor@amazon.de, sironi@amazon.de,
        Paolo Bonzini <pbonzini@redhat.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
        Dave Hansen <dave.hansen@linux.intel.com>, x86@kernel.org,
        "H. Peter Anvin" <hpa@zytor.com>, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On Tue, May 16, 2023, Like Xu wrote:
> On 27/4/2023 5:53 pm, Anselm Busse wrote:
> > Starting with v5, the Intel PMU allows to indicate the available fixed
> > function PMCs not only through CPUID.0AH.EDX[4:0] but also through a
> > bit mask in CPUID.0AH.ECX. According to the SDM the OS can consider a
> > fix function PMC i supported for:
> > 
> > CPUID.0AH.ECX[i] || (CPUID.0AH.EDX[4:0] > i)
> 
> Yes, this feature is attractive for virtualization scenarios, and it gives
> flexibility to control which fixed counters are available or not in the
> virtual machine.
> 
> However, currently KVM/x86 also supports Intel PMU V2, so I would expect
> that we will review the enablement code for v3 and v4 first.

Looking at v3, I think we probably want to skip straight to v5.  I don't see a sane
way for KVM to emulate/virtualize AnyThread, which comes in v3 without a separate
CPUID feature flag.  The SDM even calls out that it'd be a mess to deal with in a
virtualized environment.  v5 introduces a CPUID bit to allow deprecating AnyThread,
i.e. would give KVM the ability to advertise a sane vPMU model to userspace.
Amusingly, KVM advertises "edx.split.anythread_deprecated = 1" for v1+, so maybe
we don't even need to do any enabling?  At glance, I don't see any other changes
in v3 that require KVM support.

v4 looks to be an entirely different story than v3 though.  So I agree with Like
that we need to enable v3 and v4 before advertising support for v5.  And KVM *does*
need to actually advertise v5.  Emulating the fixed counter bitmap without a way to
tell userspace about the functionality will create a mess.

TL;DR: If y'all want the shiny features in v5, please enable v3 and v4 first.  I'm
totally fine taking a series to go all the way to v5 (might even be preferred due
to the AnyThread crud), but I don't want to advertise v5 without supporting the
required v3/v4 features.

> Ref: https://lore.kernel.org/kvm/CALMp9eQVnk8gkOpX5AHhaCr8-5Fe=qNuX8PUP1Gv2H5FSYmHSw@mail.gmail.com/

I agree 100% with Jim, the bitmask stuff is firmly v5+.