Fix performance bug in write_syslog(): the code to preferentially break the

log message at newlines cost O(N^2) for very long messages with few or no
newlines.  For messages in the megabyte range this became the dominant cost.
Per gripe from Achilleas Mantzios.

Patch all the way back, since this is a safe change with no portability
risks.  I am also thinking of increasing PG_SYSLOG_LIMIT, but that should
be done separately.
This commit is contained in:
Tom Lane 2008-07-08 22:17:55 +00:00
parent 1b3dfb041a
commit c3c76c2de3

View File

@ -42,7 +42,7 @@
*
*
* IDENTIFICATION
* $PostgreSQL: pgsql/src/backend/utils/error/elog.c,v 1.178.2.4 2007/07/21 22:12:11 tgl Exp $
* $PostgreSQL: pgsql/src/backend/utils/error/elog.c,v 1.178.2.5 2008/07/08 22:17:55 tgl Exp $
*
*-------------------------------------------------------------------------
*/
@ -1205,6 +1205,7 @@ write_syslog(int level, const char *line)
static unsigned long seq = 0;
int len;
const char *nlpos;
/* Open syslog connection if not done yet */
if (!openlog_done)
@ -1227,17 +1228,17 @@ write_syslog(int level, const char *line)
* fact, it does work around by splitting up messages into smaller pieces.
*
* We divide into multiple syslog() calls if message is too long or if the
* message contains embedded NewLine(s) '\n'.
* message contains embedded newline(s).
*/
len = strlen(line);
if (len > PG_SYSLOG_LIMIT || strchr(line, '\n') != NULL)
nlpos = strchr(line, '\n');
if (len > PG_SYSLOG_LIMIT || nlpos != NULL)
{
int chunk_nr = 0;
while (len > 0)
{
char buf[PG_SYSLOG_LIMIT + 1];
const char *nlpos;
int buflen;
int i;
@ -1246,11 +1247,12 @@ write_syslog(int level, const char *line)
{
line++;
len--;
/* we need to recompute the next newline's position, too */
nlpos = strchr(line, '\n');
continue;
}
/* copy one line, or as much as will fit, to buf */
nlpos = strchr(line, '\n');
if (nlpos != NULL)
buflen = nlpos - line;
else