netdev - Re: Optimising csum

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]

Date:   Tue, 22 Nov 2022 17:24:51 +0100
From:   Willy Tarreau <w@....eu>
To:     David Laight <David.Laight@...LAB.COM>
Cc:     "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
        "x86@...nel.org" <x86@...nel.org>, Arnd Bergmann <arnd@...db.de>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...hat.com>,
        "dave.hansen@...ux.intel.com" <dave.hansen@...ux.intel.com>
Subject: Re: Optimising csum_fold()

On Tue, Nov 22, 2022 at 01:08:23PM +0000, David Laight wrote:
> There are currently 20 copies of csum_fold(), some in C some in assembler.
> The default C version (in asm-generic/checksum.h) is pretty horrid.
> Some of the asm versions (including x86 and x86-64) aren't much better.
> 
> There are 3 pretty good C versions:
>   1:	(~sum - rol32(sum, 16)) >> 16
>   2:  ~(sum + rol32(sum, 16)) >> 16
>   3:  (u16)~((sum + rol32(sum, 16)) >> 16)
> All three are (usually) 4 arithmetic instructions.
> 
> The first two have the advantage that the high bits are zero.
> Relevant when the value is being checked rather than set.
> 
> The first one can generate better instruction scheduling (the rotate
> and invert can be executed in the same clock).
> 
> The 3rd one saves an instruction on arm, but may need masking.
> (I've not compiled an arm kernel to see how often that happens.)
> 
> The only architectures where (I think) the current asm code is better
> than the C above are sparc and sparc64.
> Sparc doesn't have a rotate instruction, but does have a carry flag.
> This makes the current asm version one instruction shorter.
> 
> For architectures like mips and risc-v which have neither rotate
> instructions nor carry flags the C is as good as the current asm.
> The rotate is 3 instructions - the same as the extra cmp+add.
> 
> Changing everything to use [1] would improve quite a few architectures
> while only adding 1 clock to some paths in arm/arm64 and sparc.
> 
> Unfortunately it is all currently a mess.
> Most architectures don't include asm-generic/checksum.h at all.
> 
> Thoughts?

Then why not just have one version per arch, the most efficient one,
and use it everywhere ? The simple fact that we're discussing the
tradeoffs means that if we don't want to compromise performance here
(which I assume to be the case), then it needs to be per-arch and
that's all. At least that's the way I understand it.

Regards,
Willy