Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:7 as permitted sender) client-ip=2620:137:e000::3:7;
Date:   Mon, 25 Sep 2023 17:03:32 -0700
In-Reply-To: <20230924124410.897646-1-mlevitsk@redhat.com>
Mime-Version: 1.0
References: <20230924124410.897646-1-mlevitsk@redhat.com>
Message-ID: <ZRIf1OPjKV66Y17/@google.com>
Subject: Re: [PATCH v2 0/4] KVM: x86: tracepoint updates
From:   Sean Christopherson <seanjc@google.com>
To:     Maxim Levitsky <mlevitsk@redhat.com>
Cc:     kvm@vger.kernel.org, Dave Hansen <dave.hansen@linux.intel.com>,
        x86@kernel.org, linux-kernel@vger.kernel.org,
        "H. Peter Anvin" <hpa@zytor.com>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Ingo Molnar <mingo@redhat.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Borislav Petkov <bp@alien8.de>
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On Sun, Sep 24, 2023, Maxim Levitsky wrote:
> This patch series is intended to add some selected information
> to the kvm tracepoints to make it easier to gather insights about
> running nested guests.
> 
> This patch series was developed together with a new x86 performance analysis tool
> that I developed recently (https://gitlab.com/maximlevitsky/kvmon)
> which aims to be a better kvm_stat, and allows you at glance
> to see what is happening in a VM, including nesting.

Rather than add more and more tracepoints, I think we should be more thoughtful
about (a) where we place KVM's tracepoints and (b) giving userspace the necessary
hooks to write BPF programs to extract whatever data is needed at any give time.

There's simply no way we can iterate fast enough in KVM tracepoints to adapt to
userspace's debug/monitoring needs.  E.g. if it turns out someone wants detailed
info on hypercalls that use memory or registers beyond ABCD, the new tracepoints
won't help them.

If all KVM tracepoints grab "struct kvm_vcpu" and force VMCS "registers" to be
cached (or decached depending on one's viewpoint), then I think that'll serve 99%
of usecases.  E.g. the vCPU gives a BPF program kvm_vcpu, vcpu_{vmx,svm}, kvm, etc.

trace_kvm_exit is good example, where despite all of the information that is captured
by KVM, it's borderline worthless for CPUID and MSR exits because their interesting
information is held in registers and not captured in the VMCS or VMCB.

There are some on BTF type info issues that I've encountered, but I suspect that's
as much a PEBKAC problem as anything.