linux-kernel - Re: [PATCH 11/17] find: micro-optimize for_each_{set,clear}

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]
Message-ID: <20210830121249.2fgyvf47py2tz5s5@pathway.suse.cz>
Date:   Mon, 30 Aug 2021 14:12:49 +0200
From:   Petr Mladek <pmladek@...e.com>
To:     Yury Norov <yury.norov@...il.com>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        linux-kernel@...r.kernel.org, linux-mm@...ck.org,
        linux-arch@...r.kernel.org, linux-kselftest@...r.kernel.org,
        linux-mmc@...r.kernel.org, linux-perf-users@...r.kernel.org,
        kvm@...r.kernel.org,
        "James E.J. Bottomley" <James.Bottomley@...senpartnership.com>,
        Alexander Lobakin <alobakin@...me>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Alexey Klimov <aklimov@...hat.com>,
        Andrea Merello <andrea.merello@...il.com>,
        Andy Shevchenko <andriy.shevchenko@...ux.intel.com>,
        Arnaldo Carvalho de Melo <acme@...hat.com>,
        Arnd Bergmann <arnd@...db.de>, Ben Gardon <bgardon@...gle.com>,
        Benjamin Herrenschmidt <benh@...nel.crashing.org>,
        Brian Cain <bcain@...eaurora.org>,
        Catalin Marinas <catalin.marinas@....com>,
        Christoph Lameter <cl@...ux.com>,
        Daniel Bristot de Oliveira <bristot@...hat.com>,
        David Hildenbrand <david@...hat.com>,
        Dennis Zhou <dennis@...nel.org>,
        Geert Uytterhoeven <geert@...ux-m68k.org>,
        Heiko Carstens <hca@...ux.ibm.com>,
        Ian Rogers <irogers@...gle.com>,
        Ingo Molnar <mingo@...hat.com>,
        Jaegeuk Kim <jaegeuk@...nel.org>,
        Jakub Kicinski <kuba@...nel.org>, Jiri Olsa <jolsa@...hat.com>,
        Joe Perches <joe@...ches.com>, Jonas Bonn <jonas@...thpole.se>,
        Leo Yan <leo.yan@...aro.org>,
        Mark Rutland <mark.rutland@....com>,
        Namhyung Kim <namhyung@...nel.org>,
        Palmer Dabbelt <palmer@...belt.com>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Peter Xu <peterx@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Rasmus Villemoes <linux@...musvillemoes.dk>,
        Rich Felker <dalias@...c.org>,
        Samuel Mendoza-Jonas <sam@...dozajonas.com>,
        Sean Christopherson <seanjc@...gle.com>,
        Sergey Senozhatsky <senozhatsky@...omium.org>,
        Shuah Khan <shuah@...nel.org>,
        Stefan Kristiansson <stefan.kristiansson@...nalahti.fi>,
        Steven Rostedt <rostedt@...dmis.org>,
        Tejun Heo <tj@...nel.org>,
        Thomas Bogendoerfer <tsbogend@...ha.franken.de>,
        Ulf Hansson <ulf.hansson@...aro.org>,
        Will Deacon <will@...nel.org>,
        Wolfram Sang <wsa+renesas@...g-engineering.com>,
        Yoshinori Sato <ysato@...rs.sourceforge.jp>
Subject: Re: [PATCH 11/17] find: micro-optimize for_each_{set,clear}_bit()

On Thu 2021-08-26 14:09:55, Yury Norov wrote:
> On Thu, Aug 26, 2021 at 03:57:13PM +0200, Petr Mladek wrote:
> > On Sat 2021-08-14 14:17:07, Yury Norov wrote:
> > > The macros iterate thru all set/clear bits in a bitmap. They search a
> > > first bit using find_first_bit(), and the rest bits using find_next_bit().
> > > 
> > > Since find_next_bit() is called shortly after find_first_bit(), we can
> > > save few lines of I-cache by not using find_first_bit().
> > 
> > Is this only a speculation or does it fix a real performance problem?
> > 
> > The macro is used like:
> > 
> > 	for_each_set_bit(bit, addr, size) {
> > 		fn(bit);
> > 	}
> > 
> > IMHO, the micro-opimization does not help when fn() is non-trivial.
>  
> The effect is measurable:
> 
> Start testing for_each_bit()
> for_each_set_bit:                15296 ns,   1000 iterations
> for_each_set_bit_from:           15225 ns,   1000 iterations
> 
> Start testing for_each_bit() with cash flushing
> for_each_set_bit:               547626 ns,   1000 iterations
> for_each_set_bit_from:          497899 ns,   1000 iterations
> 
> Refer this:
> 
> https://www.mail-archive.com/dri-devel@lists.freedesktop.org/msg356151.html

I see. The results look convincing on the first look.

But I am still not sure. This patch is basically contradicting many
other patches from this patchset:

  + 5th patch optimizes find_first_and_bit() and proves that it is
    much faster:

    Before (#define find_first_and_bit(...) find_next_and_bit(..., 0):
    Start testing find_bit() with random-filled bitmap
    [  140.291468] find_first_and_bit:           46890919 ns,  32671 iterations
    Start testing find_bit() with sparse bitmap
    [  140.295028] find_first_and_bit:               7103 ns,      1 iterations

    After:
    Start testing find_bit() with random-filled bitmap
    [  162.574907] find_first_and_bit:           25045813 ns,  32846 iterations
    Start testing find_bit() with sparse bitmap
    [  162.578458] find_first_and_bit:               4900 ns,      1 iterations

       => saves 46% in random bitmap
	  saves 31% in sparse bitmap


  + 6th, 7th, and 9th patch makes the code use find_first_bit()
    because it is faster than find_next_bit(mask, size, 0);

  + Now, 11th (this) patch replaces find_first_bit() with
    find_next_bit(mask, size, 0) because find_first_bit()
    makes things slower. It is suspicious at minimum.


By other words. The I-cache could safe 10% in one case.
But find_first_bit() might safe 46% in random case.

Does I-cache cost more than the faster code?

Or was for_each_set_bit() tested only with a bitmap
where find_first_bit() optimization did not help much?

How would for_each_set_bit() work with random bitmap?
How does it work with larger bitmaps?

Best Regards,
Petr