linux-kernel - Re: [PATCH RFC 0/7] x86: convert ticketlocks to C and remove duplicate code

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 for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20110625101146.GB19097@elte.hu>
Date:	Sat, 25 Jun 2011 12:11:46 +0200
From:	Ingo Molnar <mingo@...e.hu>
To:	Jeremy Fitzhardinge <jeremy@...p.org>
Cc:	Peter Zijlstra <peterz@...radead.org>,
	"H. Peter Anvin" <hpa@...or.com>,
	the arch/x86 maintainers <x86@...nel.org>,
	Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
	Nick Piggin <npiggin@...nel.dk>,
	Jeremy Fitzhardinge <jeremy.fitzhardinge@...rix.com>
Subject: Re: [PATCH RFC 0/7] x86: convert ticketlocks to C and remove
 duplicate code


* Jeremy Fitzhardinge <jeremy@...p.org> wrote:

>  2. With NR_CPUS < 256 the ticket size is 8 bits.  The compiler doesn't
>     use the same trick as the hand-coded asm to directly compare the high
>     and low bytes in the word, but does a bit of extra shuffling around.
>     However, the Intel optimisation guide and several x86 experts have
>     opined that its best to avoid the high-byte operations anyway, since
>     they will cause a partial word stall, and the gcc-generated code should
>     be better.
> 
>     Overall the compiler-generated code is very similar to the hand-coded
>     versions, with the partial byte operations being the only significant
>     difference. (Curiously, gcc does generate a high-byte compare for me
>     in trylock, so it can if it wants to.)
> 
> I've been running with this code in place for several months on 4 core
> systems without any problems.

Please do measurements both in terms of disassembly based instruction 
count(s) in the fastpath(s) (via looking at the before/after 
disassembly) and actual cycle, instruction and branch counts (via 
perf measurements).

> I couldn't measure a consistent performance difference between the two
> implemenations; there seemed to be +/- ~1% +/-, which is the level of
> variation I see from simply recompiling the kernel with slightly
> different code alignment.

Then you've done the micro-cost measurements the wrong way - we can 
and do detect much finer effects than 1%, see the methods used in 
this commit for example:

  c8b281161dfa: sched: Increase SCHED_LOAD_SCALE resolution

Please also ensure that the cold-cache behavior is fairly measured 
via hot-cache benchmarks (that is not always guaranteed).

Thanks,

	Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/