2.4.20: problem with "ps -olstart"

2.4.20: problem with "ps -olstart"

[Eyal Lebedinsky]

I have a program that needs to check if a process is still running.
It issues a ps and saves the pid and start time. It then uses this
info to compare against a later "ps --pid".

The monitored process is long-running and unlikely to restart at the
same second and with the same pid. I am trying to go better than
just using a pid sentry (I also compare the cmd to be really sure).

I see a different start time returned on different calls. An example
is attached below. This is a show stopper for me. Is this a known
problem? Does it have a solution?

This is vanilla (my build) 2.4.20 on i386.

$ ps
  PID TTY          TIME CMD
  906 pts/0    00:00:01 bash
 3026 pts/0    00:00:00 sh
 8254 pts/0    00:00:00 ps
$ while true ; do ps --pid "3026" -olstart,cmd --no-headers ; done
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:11 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:12 2003 sh
Thu Mar 27 22:03:11 2003 sh

--
Eyal Lebedinsky (eyal@eyal.emu.id.au) <http://samba.org/eyal/>

Re: 2.4.20: problem with "ps -olstart"

[Andries Brouwer]

On Sat, Mar 29, 2003 at 01:08:58PM +1100, Eyal Lebedinsky wrote:

> I see a different start time returned on different calls. An example
> is attached below. This is a show stopper for me. Is this a known
> problem? Does it have a solution?
> 
> This is vanilla (my build) 2.4.20 on i386.
> 
> $ while true ; do ps --pid "3026" -olstart,cmd --no-headers ; done
> Thu Mar 27 22:03:11 2003 sh
> Thu Mar 27 22:03:11 2003 sh
> Thu Mar 27 22:03:12 2003 sh
> Thu Mar 27 22:03:11 2003 sh

Look at your ps source. There are many incarnations of ps,
but perhaps you'll find something like

	seconds_since_boot = uptime(0,0);
	seconds_since_1970 = time(NULL);
	time_of_boot = seconds_since_1970 - seconds_since_boot;
	start = time_of_boot + pp->start_time/Hertz;

The interplay of rounding and truncating you see here
results in what you see. Instead of using ps you might try
a tiny utility that reads the start time directly.

Andries

Re: 2.4.20: problem with "ps -olstart"

[Eyal Lebedinsky]

Andries Brouwer wrote:
> 
> On Sat, Mar 29, 2003 at 01:08:58PM +1100, Eyal Lebedinsky wrote:
> 
> > I see a different start time returned on different calls. An example
> > is attached below. This is a show stopper for me. Is this a known
> > problem? Does it have a solution?
> >
> > This is vanilla (my build) 2.4.20 on i386.
> >
> > $ while true ; do ps --pid "3026" -olstart,cmd --no-headers ; done
> > Thu Mar 27 22:03:11 2003 sh
> > Thu Mar 27 22:03:11 2003 sh
> > Thu Mar 27 22:03:12 2003 sh
> > Thu Mar 27 22:03:11 2003 sh
> 
> Look at your ps source. There are many incarnations of ps,
> but perhaps you'll find something like
> 
>         seconds_since_boot = uptime(0,0);
>         seconds_since_1970 = time(NULL);
>         time_of_boot = seconds_since_1970 - seconds_since_boot;
>         start = time_of_boot + pp->start_time/Hertz;
> 
> The interplay of rounding and truncating you see here
> results in what you see. Instead of using ps you might try
> a tiny utility that reads the start time directly.

OK, I can see how ps may be doing a bad job there. It can do
better, e.g. from I can identify a process using something like:
	grep btime /proc/stat
		time machine booted (secs since 1970 probably)
	cut -d' ' -f22 </proc/$pid/stat
		process start time (jiffies since boot)
and if I really want I can convert start_time to seconds and have
a value (the sum of the two) that does not change.

>From a user program, and not depending on /proc, I do not see
how I can access the kernel 'xtime' and 'task->start_time'.

--
Eyal Lebedinsky (eyal@eyal.emu.id.au) <http://samba.org/eyal/>