From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-fsdevel-owner@vger.kernel.org>
Received: from mail-db5eur01on0131.outbound.protection.outlook.com ([104.47.2.131]:20448
        "EHLO EUR01-DB5-obe.outbound.protection.outlook.com"
        rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP
        id S1752322AbeAOHFB (ORCPT <rfc822;linux-fsdevel@vger.kernel.org>);
        Mon, 15 Jan 2018 02:05:01 -0500
Date: Sun, 14 Jan 2018 23:04:41 -0800
From: Andrei Vagin <avagin@virtuozzo.com>
To: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrei Vagin <avagin@openvz.org>, linux-fsdevel@vger.kernel.org,
        Alexey Dobriyan <adobriyan@gmail.com>,
        KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Subject: Re: [PATCH 1/2] procfs: add seq_put_hex_ll to speed up /proc/pid/maps
Message-ID: <20180115070440.GA23941@outlook.office365.com>
References: <20180112185812.7710-1-avagin@openvz.org>
 <20180112153304.f4a7dfbae2942e3fdd93eab9@linux-foundation.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=koi8-r
Content-Disposition: inline
In-Reply-To: <20180112153304.f4a7dfbae2942e3fdd93eab9@linux-foundation.org>
Sender: linux-fsdevel-owner@vger.kernel.org
List-ID: <linux-fsdevel.vger.kernel.org>

On Fri, Jan 12, 2018 at 03:33:04PM -0800, Andrew Morton wrote:
> On Fri, 12 Jan 2018 10:58:11 -0800 Andrei Vagin <avagin@openvz.org> wrote:
> 
> > seq_put_hex_ll() prints a number in hexadecimal notation and works
> > faster than seq_printf().
> > 
> > ...
> >
> > --- a/fs/seq_file.c
> > +++ b/fs/seq_file.c
> > @@ -670,6 +670,26 @@ void seq_puts(struct seq_file *m, const char *s)
> >  }
> >  EXPORT_SYMBOL(seq_puts);
> >  
> > +static inline void seq_put_delimeter(struct seq_file *m, const char *delimiter)
> > +{
> > +	int len;
> > +
> > +	if (!delimiter || !delimiter[0])
> > +		return;
> > +
> > +	if (delimiter[1] == 0)
> > +		return seq_putc(m, delimiter[0]);
> > +
> > +	len = strlen(delimiter);
> > +	if (m->count + len >= m->size) {
> > +		seq_set_overflow(m);
> > +		return;
> > +	}
> > +
> > +	memcpy(m->buf + m->count, delimiter, len);
> > +	m->count += len;
> > +}
> 
> Can we please have a nice comment describing this function's role and
> behaviour?

seq_put_decimal_* and seq_put_hex_ll prints a string before printing a
number. Originaly it was just one symbol, it is probably a reason why it
is called delimeter.

I added an optimization for a case when delimiter is one symbol, and
found that it sinificantly affect perfomance (about 13% for
/proc/pid/maps):

Without this optimization:
[root@fc24 ~]# time python test.py 

real	0m9.105s
user	0m2.200s
sys	0m6.901s

With this optimization:
[root@fc24 ~]# time python test.py 

real	0m8.097s
user	0m1.994s
sys	0m6.102s


If inline is replaced by noinline
[root@fc24 ~]# time python test.py 

real	0m8.263s
user	0m2.058s
sys	0m6.200s

[root@fc24 ~]# cat test.py 
#!/usr/bin/env python2

num = 0
with open("/proc/1/maps") as f:
        for x in xrange(100000):
                data = f.read()
                f.seek(0, 0)

Andrew, thank you for the review, I will send a fixed patch soon.

> 
> I don't think the `inline' is needed or desirable - gcc can figure that
> out, and with three callsites a `noinline' would be more justified!
> 
> That `return seq_putc(...)' will generate a warning in some situations
> - seq_putc() returns void.  Let's split it into 'seq_putc(...);
> return;' please.
> 
> > +/**
> > + * seq_put_hex_ll - put a number in hexadecimal notation
> > + * @m: seq_file identifying the buffer to which data should be written
> > + * @delimiter: a string which is printed before the number
> > + * @v: the number
> > + * @width: a minimum field width
> > + *
> > + * seq_put_hex_ll(m, "", v, 8) is equal to seq_printf(m, "0x08llx", v)
> > + *
> > + * This routine is very quick when you show lots of numbers.
> > + * In usual cases, it will be better to use seq_printf(). It's easier to read.
> > + */
> > +void seq_put_hex_ll(struct seq_file *m, const char *delimiter,
> > +				unsigned long long v, int width)
> > +{
> > +	int i, len;
> > +
> > +	seq_put_delimeter(m, delimiter);
> > +
> > +	len = (sizeof(v) * 8 - __builtin_clzll(v) + 3) / 4;
> > +
> > +	if (unlikely(len == 0))
> > +		len = 1;
> > +
> > +	if (len < width)
> > +		len = width;
> > +
> > +	if (m->count + len > m->size)
> > +		goto overflow;
> > +
> > +	for (i = len - 1; i >= 0; i--) {
> > +		m->buf[m->count + i] = hex_asc[0xf & v];
> > +		v = v >> 4;
> > +	}
> > +	m->count += len;
> > +	return;
> > +overflow:
> > +	seq_set_overflow(m);
> > +}
> 
> I don't think we need the goto.  Just do "seq_set_overflow(m); return;".
>