linux-kernel - Question Regarding ERMS memcpy

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-next>] [day] [month] [year] [list]

Message-ID: <e3e805ee-dc81-55f3-46e6-e3c7430096c3@deltatee.com>
Date:   Sat, 4 Mar 2017 13:08:15 -0700
From:   Logan Gunthorpe <logang@...tatee.com>
To:     Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>,
        "H. Peter Anvin" <hpa@...or.com>, Tony Luck <tony.luck@...el.com>,
        Borislav Petkov <bp@...e.de>, Al Viro <viro@...iv.linux.org.uk>
Cc:     the arch/x86 maintainers <x86@...nel.org>,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>
Subject: Question Regarding ERMS memcpy

Hi,

I'm trying to chase down a performance issue with a driver I'm working
on that does a repeated memcpy_fromio of about 1KB from a PCI device. I
made a small change from a fixed size copy to a variable size only to be
surprised with a performance decrease of about 1/3.

I've looked through the code and discovered this is due to switching out
__builtin_memcpy with __memcpy when the length is not constant. This
makes sense and I've now been looking into the inner workings of
memcpy_64.S.

The CPU I'm testing on is a Sandy Bridge and according to /proc/cpuinfo,
it does _not_ have the erms bit and does have the rep_good bit. So I
expect to be using the "rep movsq" version of memcpy. However, I've done
some testing with my driver by hacking in specific memcpy
implementations and I've found the following results:

__builtin_memcpy w/const length: 85KB/s
memcpy_fromio: 26kB/s
__builtin_memcpy: 26kB/s
memcpy_movsq: 126kB/s
memcpy_erms: 26kB/s

Thus, based on these performance numbers, it almost seems like my
platform is using the erms version when it probably shouldn't be.

So my question is: how do I find out what version of memcpy my actual
machine is using and fix it if it is wrong?

I'm running with a 4.10.0 kernel and I've attached my cpuinfo.

Thanks for any insights,

Logan

View attachment "cpuinfo.txt" of type "text/plain" (7488 bytes)