Date:	Mon, 7 Mar 2016 15:52:06 -0800
From:	Alexander Duyck <alexander.duyck@...il.com>
To:	Tom Herbert <tom@...bertland.com>
Cc:	David Laight <David.Laight@...lab.com>,
	Linus Torvalds <torvalds@...ux-foundation.org>,
	"davem@...emloft.net" <davem@...emloft.net>,
	"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
	"tglx@...utronix.de" <tglx@...utronix.de>,
	"mingo@...hat.com" <mingo@...hat.com>,
	"hpa@...or.com" <hpa@...or.com>, "x86@...nel.org" <x86@...nel.org>,
	"kernel-team@...com" <kernel-team@...com>
Subject: Re: [PATCH v5 net-next] net: Implement fast csum_partial for x86_64

On Mon, Mar 7, 2016 at 9:33 AM, Tom Herbert <tom@...bertland.com> wrote:
> On Mon, Mar 7, 2016 at 5:56 AM, David Laight <David.Laight@...lab.com> wrote:
>> From: Alexander Duyck
>>  ...
>>> Actually probably the easiest way to go on x86 is to just replace the
>>> use of len with (len >> 6) and use decl or incl instead of addl or
>>> subl, and lea instead of addq for the buff address.  None of those
>>> instructions effect the carry flag as this is how such loops were
>>> intended to be implemented.
>>>
>>> I've been doing a bit of testing and that seems to work without
>>> needing the adcq until after you exit the loop, but doesn't give that
>>> much of a gain in speed for dropping the instruction from the
>>> hot-path.  I suspect we are probably memory bottle-necked already in
>>> the loop so dropping an instruction or two doesn't gain you much.
>>
>> Right, any superscalar architecture gives you some instructions
>> 'for free' if they can execute at the same time as those on the
>> critical path (in this case the memory reads and the adc).
>> This is why loop unrolling can be pointless.
>>
>> So the loop:
>> 10:     addc %rax,(%rdx,%rcx,8)
>>         inc %rcx
>>         jnz 10b
>> could easily be as fast as anything that doesn't use the 'new'
>> instructions that use the overflow flag.
>> That loop might be measurable faster for aligned buffers.
>
> Tested by replacing the unrolled loop in my patch with just:
>
> if (len >= 8) {
>                 asm("clc\n\t"
>                     "0: adcq (%[src],%%rcx,8),%[res]\n\t"
>                     "decl %%ecx\n\t"
>                     "jge 0b\n\t"
>                     "adcq $0, %[res]\n\t"
>                             : [res] "=r" (result)
>                             : [src] "r" (buff), "[res]" (result), "c"
> ((len >> 3) - 1));
> }
>
> This seems to be significantly slower:
>
> 1400 bytes: 797 nsecs vs. 202 nsecs
> 40 bytes: 6.5 nsecs vs. 26.8 nsecs

You still need the loop unrolling as the decl and jge have some
overhead.  You can't just get rid of it with a single call in a tight
loop but it should improve things.  The gain from what I have seen
ends up being minimal though.  I haven't really noticed all that much
in my tests anyway.

I have been doing some testing and the penalty for an unaligned
checksum can get pretty big if the data-set is big enough.  I was
messing around and tried doing a checksum over 32K minus some offset
and was seeing a penalty of about 200 cycles per 64K frame.

One thought I had is that we may want to look into making an inline
function that we can call for compile-time defined lengths less than
64.  Maybe call it something like __csum_partial and we could then use
that in place of csum_partial for all those headers that are a fixed
length that we pull such as UDP, VXLAN, Ethernet, and the rest.  Then
we might be able to look at taking care of alignment for csum_partial
which will improve the skb_checksum() case without impacting the
header pulling cases as much since that code would be inlined
elsewhere.

- Alex

Powered by blists - more mailing lists