[sheepdog] [PATCH v4] sheep: avoid the OOM killer

[Hitoshi Mitake]

At Fri, 05 Jul 2013 19:24:59 +0900,
Hitoshi Mitake wrote:
> 
> At Fri, 5 Jul 2013 14:58:33 +0800,
> Liu Yuan wrote:
> > 
> > On Fri, Jul 05, 2013 at 03:40:45PM +0900, MORITA Kazutaka wrote:
> > > At Fri,  5 Jul 2013 15:30:10 +0900,
> > > Hitoshi Mitake wrote:
> > > > 
> > > > +int disable_oom_killing(void)
> > > > +{
> > > > +	int fd, ret = 0;
> > > > +	char path[PATH_MAX], score_str[8];
> > > > +
> > > > +	/*
> > > > +	 * Processes of Linux which wrote a string representation of
> > > > +	 * OOM_SCORE_ADJ_MIN to /proc/<their pid>/oom_score_adj are excluded by
> > > > +	 * the target list of the OOM killer certainly.
> > > > +	 */
> > > > +
> > > > +	snprintf(path, PATH_MAX, "/proc/%d/oom_score_adj", getpid());
> > > 
> > > Using /proc/self/oom_score_adj is simpler.
> > > 
> > > > +	fd = open(path, O_WRONLY);
> > > > +	if (fd < 0) {
> > > > +		sd_eprintf("opening %s failed, %m", path);
> > > > +		ret = -1;
> > > > +		goto out;
> > > > +	}
> > > > +
> > > > +	snprintf(score_str, 8, "%d\n", OOM_SCORE_ADJ_MIN);
> > > 
> > > Use sizeof(score_str).
> > > 
> > > > +
> > > > +	if (xwrite(fd, score_str, strlen(score_str)) != 6) {
> > > 
> > > Please don't use a magic number.  OOM_SCORE_ADJ_MIN is -16 on my
> > > environment and this code doesn't work.
> > > 
> > > BTW, I wonder if we should set oom_score_adj inside the program.  I
> > > don't how other programs disable OOM, but, perhaps, should we set this
> > > in /etc/init.d/sheepdog?
> > > 
> > 
> > I guess no one want to be oom killed and this might be better decided by admin.
> > Disable it by sheep would be too intrusive.
> > 
> > I think put it init.d/sheepdog is okay but we'd better introduce sheepdog.conf.
> 
> As we discussed at IRC meeting, sheepdog.conf will be a useful
> thing. But setting oom_score_adj is a bother work because we need to
> grep a PID of a sheep process and every sheep has its logger. Doing it
> inside sheep is much simpler.

On the second thought, it seems that setting oom_score_adj or oom_adj
from outer processes would be easy. I agree with doing it in the init
script.

Thanks,
Hitoshi