Date:   Fri, 17 Feb 2023 08:35:36 -0800
In-Reply-To: <2c2f77a3-1d77-0d88-991a-60dcdc370ea8@grsecurity.net>
Mime-Version: 1.0
References: <20230213163351.30704-1-minipli@grsecurity.net>
 <Y+pt5MGR+EjLH4qQ@google.com> <13deaeb6-dfb2-224c-0aa3-5546ad426f63@grsecurity.net>
 <Y+5okhlB4rkXjKWS@google.com> <2c2f77a3-1d77-0d88-991a-60dcdc370ea8@grsecurity.net>
Message-ID: <Y++s2LvnBxydxhVM@google.com>
Subject: Re: [PATCH 0/5] KVM: Put struct kvm_vcpu on a diet
From:   Sean Christopherson <seanjc@google.com>
To:     Mathias Krause <minipli@grsecurity.net>
Cc:     kvm@vger.kernel.org, linux-kernel@vger.kernel.org,
        Paolo Bonzini <pbonzini@redhat.com>
Content-Type: text/plain; charset="us-ascii"
Precedence: bulk

On Fri, Feb 17, 2023, Mathias Krause wrote:
> On 16.02.23 18:32, Sean Christopherson wrote:
> > I'm not necessarily opposed to such aggressive optimization, but the ROI is likely
> > very, very low.  For optimized workloads, there simply aren't very many VM-Exits,
> > e.g. the majority of exits on a modern CPU are due to timer ticks.  And even those
> > will hopefully be eliminiated in the not-too-distant future, e.g. by having hardware
> > virtualize the TSC deadline timer, and by moving to a vCPU scheduling scheme that
> > allows for a tickless host.
> 
> Well, for guests running grsecurity kernels, there's also the CR0.WP
> toggling triggering VMEXITs, which happens a lot! -- at least until
> something along the lines of [1] gets merged *hint ;)*

Ha!  It's high on my todo list for 6.4, catching up on other stuff at the moment.

That series is also _exactly_ why the ROI for aggressive cache line optimization
is low.  The better long term answer is almost always to avoid the VM-Exit in the
first place, or failing that, to handle the exit in a fastpath.  Sometimes it takes
a few years, e.g. to get necessary hardware support, but x86 virtualization is fast
approaching the point where anything remotely performance critical is handled entirely
within the guest.