mirror of
http://git.haproxy.org/git/haproxy.git/
synced 2024-12-18 09:24:31 +00:00
c4710e14fe
There has been a USE_MEMCHR option for ages that was mostly never enabled because it was unclear when glibc became faster. A quick look at the code indicates that this arrived with the SSE implementation of memchr() which arrived at commit 093ecf92998de2 between 2.14 and 2.15, so let's automatically turn this on on x86_64 with glibc >= 2.15. This results in ~6GB of logs read per second (20 million lines) and ~2.5GB/s (8 million lines) parsed for errors or status codes classification, or 1 GB/s (3 million lines) for time percentiles.
268 lines
7.8 KiB
C
268 lines
7.8 KiB
C
/*
|
|
* fast fgets() replacement for log parsing
|
|
*
|
|
* Copyright 2000-2012 Willy Tarreau <w@1wt.eu>
|
|
*
|
|
* This library is free software; you can redistribute it and/or
|
|
* modify it under the terms of the GNU Lesser General Public
|
|
* License as published by the Free Software Foundation, version 2.1
|
|
* exclusively.
|
|
*
|
|
* This library is distributed in the hope that it will be useful,
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
* Lesser General Public License for more details.
|
|
*
|
|
* You should have received a copy of the GNU Lesser General Public License
|
|
* along with this library; if not, write to the Free Software Foundation,
|
|
* Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
*
|
|
* This function manages its own buffer and returns a pointer to that buffer
|
|
* in order to avoid expensive memory copies. It also checks for line breaks
|
|
* 32 or 64 bits at a time. It could be improved a lot using mmap() but we
|
|
* would not be allowed to replace trailing \n with zeroes and we would be
|
|
* limited to small log files on 32-bit machines.
|
|
*
|
|
*/
|
|
|
|
#include <stdlib.h>
|
|
#include <string.h>
|
|
#include <stdio.h>
|
|
#include <unistd.h>
|
|
|
|
#ifndef FGETS2_BUFSIZE
|
|
#define FGETS2_BUFSIZE (256*1024)
|
|
#endif
|
|
|
|
/* memchr() is faster in glibc with SSE since commit 093ecf92998de2 */
|
|
#if defined(__x86_64__) && (__GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ >= 15))
|
|
#define USE_MEMCHR
|
|
#endif
|
|
|
|
/* return non-zero if the integer contains at least one zero byte */
|
|
static inline __attribute__((unused)) unsigned int has_zero32(unsigned int x)
|
|
{
|
|
unsigned int y;
|
|
|
|
/* Principle: we want to perform 4 tests on one 32-bit int at once. For
|
|
* this, we have to simulate an SIMD instruction which we don't have by
|
|
* default. The principle is that a zero byte is the only one which
|
|
* will cause a 1 to appear on the upper bit of a byte/word/etc... when
|
|
* we subtract 1. So we can detect a zero byte if a one appears at any
|
|
* of the bits 7, 15, 23 or 31 where it was not. It takes only one
|
|
* instruction to test for the presence of any of these bits, but it is
|
|
* still complex to check for their initial absence. Thus, we'll
|
|
* proceed differently : we first save and clear only those bits, then
|
|
* we check in the final result if one of them is present and was not.
|
|
* The order of operations below is important to save registers and
|
|
* tests. The result is used as a boolean, so the last test must apply
|
|
* on the constant so that it can efficiently be inlined.
|
|
*/
|
|
#if defined(__i386__)
|
|
/* gcc on x86 loves copying registers over and over even on code that
|
|
* simple, so let's do it by hand to prevent it from doing so :-(
|
|
*/
|
|
asm("lea -0x01010101(%0),%1\n"
|
|
"not %0\n"
|
|
"and %1,%0\n"
|
|
: "=a" (x), "=r"(y)
|
|
: "0" (x)
|
|
);
|
|
return x & 0x80808080;
|
|
#else
|
|
y = x - 0x01010101; /* generate a carry */
|
|
x = ~x & y; /* clear the bits that were already set */
|
|
return x & 0x80808080;
|
|
#endif
|
|
}
|
|
|
|
/* return non-zero if the argument contains at least one zero byte. See principle above. */
|
|
static inline __attribute__((unused)) unsigned long long has_zero64(unsigned long long x)
|
|
{
|
|
unsigned long long y;
|
|
|
|
y = x - 0x0101010101010101ULL; /* generate a carry */
|
|
y &= ~x; /* clear the bits that were already set */
|
|
return y & 0x8080808080808080ULL;
|
|
}
|
|
|
|
static inline __attribute__((unused)) unsigned long has_zero(unsigned long x)
|
|
{
|
|
return (sizeof(x) == 8) ? has_zero64(x) : has_zero32(x);
|
|
}
|
|
|
|
/* find a '\n' between <next> and <end>. Warning: may read slightly past <end>.
|
|
* If no '\n' is found, <end> is returned.
|
|
*/
|
|
static char *find_lf(char *next, char *end)
|
|
{
|
|
#if defined USE_MEMCHR
|
|
/* some recent libc use platform-specific optimizations to provide more
|
|
* efficient byte search than below (eg: glibc 2.11 on x86_64).
|
|
*/
|
|
next = memchr(next, '\n', end - next);
|
|
if (!next)
|
|
next = end;
|
|
#else
|
|
if (sizeof(long) == 4) { /* 32-bit system */
|
|
/* this is a speed-up, we read 32 bits at once and check for an
|
|
* LF character there. We stop if found then continue one at a
|
|
* time.
|
|
*/
|
|
while (next < end && (((unsigned long)next) & 3) && *next != '\n')
|
|
next++;
|
|
|
|
/* Now next is multiple of 4 or equal to end. We know we can safely
|
|
* read up to 32 bytes past end if needed because they're allocated.
|
|
*/
|
|
while (next < end) {
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
if (has_zero32(*(unsigned int *)next ^ 0x0A0A0A0A))
|
|
break;
|
|
next += 4;
|
|
}
|
|
}
|
|
else { /* 64-bit system */
|
|
/* this is a speed-up, we read 64 bits at once and check for an
|
|
* LF character there. We stop if found then continue one at a
|
|
* time.
|
|
*/
|
|
if (next <= end) {
|
|
/* max 3 bytes tested here */
|
|
while ((((unsigned long)next) & 3) && *next != '\n')
|
|
next++;
|
|
|
|
/* maybe we have can skip 4 more bytes */
|
|
if ((((unsigned long)next) & 4) && !has_zero32(*(unsigned int *)next ^ 0x0A0A0A0AU))
|
|
next += 4;
|
|
}
|
|
|
|
/* now next is multiple of 8 or equal to end */
|
|
while (next <= (end-68)) {
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
if (has_zero64(*(unsigned long long *)next ^ 0x0A0A0A0A0A0A0A0AULL))
|
|
break;
|
|
next += 8;
|
|
}
|
|
|
|
/* maybe we can skip 4 more bytes */
|
|
if (!has_zero32(*(unsigned int *)next ^ 0x0A0A0A0AU))
|
|
next += 4;
|
|
}
|
|
|
|
/* We finish if needed : if <next> is below <end>, it means we
|
|
* found an LF in one of the 4 following bytes.
|
|
*/
|
|
while (next < end) {
|
|
if (*next == '\n')
|
|
break;
|
|
next++;
|
|
}
|
|
#endif
|
|
return next;
|
|
}
|
|
|
|
const char *fgets2(FILE *stream)
|
|
{
|
|
static char buffer[FGETS2_BUFSIZE + 68]; /* Note: +32 is enough on 32-bit systems */
|
|
static char *end = buffer;
|
|
static char *line = buffer;
|
|
char *next;
|
|
int ret;
|
|
|
|
next = line;
|
|
|
|
while (1) {
|
|
next = find_lf(next, end);
|
|
if (next < end) {
|
|
const char *start = line;
|
|
*next = '\0';
|
|
line = next + 1;
|
|
return start;
|
|
}
|
|
|
|
/* we found an incomplete line. First, let's move the
|
|
* remaining part of the buffer to the beginning, then
|
|
* try to complete the buffer with a new read. We can't
|
|
* rely on <next> anymore because it went past <end>.
|
|
*/
|
|
if (line > buffer) {
|
|
if (end != line)
|
|
memmove(buffer, line, end - line);
|
|
end = buffer + (end - line);
|
|
next = end;
|
|
line = buffer;
|
|
} else {
|
|
if (end == buffer + FGETS2_BUFSIZE)
|
|
return NULL;
|
|
}
|
|
|
|
ret = read(fileno(stream), end, buffer + FGETS2_BUFSIZE - end);
|
|
|
|
if (ret <= 0) {
|
|
if (end == line)
|
|
return NULL;
|
|
|
|
*end = '\0';
|
|
end = line; /* ensure we stop next time */
|
|
return line;
|
|
}
|
|
|
|
end += ret;
|
|
*end = '\n'; /* make parser stop ASAP */
|
|
/* search for '\n' again */
|
|
}
|
|
}
|
|
|
|
#ifdef BENCHMARK
|
|
int main() {
|
|
const char *p;
|
|
unsigned int lines = 0;
|
|
|
|
while ((p=fgets2(stdin)))
|
|
lines++;
|
|
printf("lines=%d\n", lines);
|
|
return 0;
|
|
}
|
|
#endif
|