linux-kernel - Re: [PATCH v2 2/4] KVM: Add paravirt remote TLB flush

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: <20171115095425.2hsgpfomdmdru7ke@hirez.programming.kicks-ass.net>
Date:   Wed, 15 Nov 2017 10:54:25 +0100
From:   Peter Zijlstra <peterz@...radead.org>
To:     Wanpeng Li <kernellwp@...il.com>
Cc:     "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        kvm <kvm@...r.kernel.org>, Radim Kr??m???? <rkrcmar@...hat.com>,
        Wanpeng Li <wanpeng.li@...mail.com>,
        Paolo Bonzini <pbonzini@...hat.com>
Subject: Re: [PATCH v2 2/4] KVM: Add paravirt remote TLB flush

On Wed, Nov 15, 2017 at 04:43:32PM +0800, Wanpeng Li wrote:
> Hi Peterz,
> 
> I found big performance difference as I discuss with you several days ago.
> 
> ebizzy -M
>                         vanilla    static/local cpumask     per-cpu cpumask
>  8 vCPUs       10152            10083                          10117
> 16 vCPUs        1224              4866                          10008
> 24 vCPUs        1109              3871                            9928
> 32 vCPUs        1025              3375                            9811
> 
> In addition, I can observe ~50% perf top time is occupied by
> smp_call_function_many(), ~30% perf top time is occupied by
> call_function_interrupt() in the guest when running ebizzy for
> static/local cpumask variable. However, I almost can't observe these
> IPI stuffs after changing to per-cpu variable. Any opinions?

That doesn't really make sense.. :/

So a single static variable is broken (multiple CPUs can call
flush_tlb_others() concurrently and overwrite each others masks). But I
don't see why a per-cpu variable would be much slower than an on-stack
variable.