linux-kernel - Re: [PATCH] KVM: selftests: Make hyperv

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <2a20d49f5a18ecd3a76a96dec01ad592a59bdc3f.camel@redhat.com>
Date:   Tue, 07 Jun 2022 11:07:58 +0300
From:   Maxim Levitsky <mlevitsk@...hat.com>
To:     Sean Christopherson <seanjc@...gle.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>
Cc:     kvm@...r.kernel.org, Paolo Bonzini <pbonzini@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] KVM: selftests: Make hyperv_clock selftest more stable

On Wed, 2022-06-01 at 16:06 +0000, Sean Christopherson wrote:
> On Wed, Jun 01, 2022, Vitaly Kuznetsov wrote:
> > hyperv_clock doesn't always give a stable test result, especially
> > with
> > AMD CPUs. The test compares Hyper-V MSR clocksource (acquired
> > either
> > with rdmsr() from within the guest or KVM_GET_MSRS from the host)
> > against rdtsc(). To increase the accuracy, increase the measured
> > delay
> > (done with nop loop) by two orders of magnitude and take the mean
> > rdtsc()
> > value before and after rdmsr()/KVM_GET_MSRS.
> 
> Rather than "fixing" the test by reducing the impact of noise, can we
> first try
> to reduce the noise itself?  E.g. pin the test to a single CPU, redo
> the measurement

Pinning is a good idea overall, however IMHO should not be done in 
all KVM selftests, as vCPU migration itself can be source of bugs.


> if the test is interrupted (/proc/interrupts?), etc...  Bonus points
This is not feasable IMHO - timer interrupt alone can fire at rate of
1000 interrupts/s. Just while reading /proc/interurpts you probably get
few of interrupts.


> if that can
> be implemented as a helper or pair of helpers so that other tests
> that want to
> measure latency/time don't need to reinvent the wheel.
> 

Best regards,
	Maxim Levitsky