linux-kernel - Re: x86 TLB flushing: INVPCID vs. deferred CR3 write

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] [day] [month] [year] [list]

Date:   Wed, 6 Dec 2017 18:33:13 +0100
From:   Ingo Molnar <mingo@...nel.org>
To:     Dave Hansen <dave.hansen@...el.com>
Cc:     the arch/x86 maintainers <x86@...nel.org>,
        Andy Lutomirski <luto@...nel.org>,
        LKML <linux-kernel@...r.kernel.org>,
        Linux-MM <linux-mm@...ck.org>,
        "Kleen, Andi" <andi.kleen@...el.com>,
        "Chen, Tim C" <tim.c.chen@...el.com>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        Peter Zijlstra <peterz@...radead.org>
Subject: Re: x86 TLB flushing: INVPCID vs. deferred CR3 write

* Dave Hansen <dave.hansen@...el.com> wrote:

> tl;dr: Kernels with pagetable isolation using INVPCID compile kernels
> 0.58% faster than using the deferred CR3 write.  This tends to say that
> we should leave things as-is and keep using INVPCID, but it's far from
> definitive.

Agreed, thanks for the detailed testing!

> If folks have better ideas for a test methodology, or specific workloads or 
> hardware where you want to see this tested, please speak up.

I had a look at the numbers and it all looks valid and good to me too - it's also 
the intuitive result IMHO.

I suspect there might be synthetic cache-hot workloads where the +330 cycles cost 
of INVPCID is higher than that of the extra TLB miss costs of a CR3 flush - but we 
do know that this offset is constant, while the cost of flushing all TLBs ever 
increases with the future increases of the TLB cache.

Thanks,

	Ingo