Re: [RFC/PATCH 2/2] kernel: don't update load average during snapshot/shutdown

[Nigel Cunningham <nigel@nigel.suspend2.net>]

[-- Attachment #1.1: Type: text/plain, Size: 2492 bytes --]

Hi.

On Fri, 2007-04-27 at 16:52 +0200, Pavel Machek wrote:
> Hi!
> 
> > From: Nigel Cunningham <nigel@nigel.suspend2.net>
> > 
> > As explained by Nigel, Sendmail and some other programs look at the
> > load average and stop delivering when it gets too high.  The CPU
> > intensiveness of suspending to disk pushes the load average up quite
> > high but post-resume we shouldn't really take that into account.  Not
> > updating the load average therefore makes snapshot/shutdown invisible
> > to sendmail and so on.
> 
> Is it kernel problem? We _had_ high load, and load average goes down
> pretty quickly.

Don't make artificial distinctions between kernelspace and userspace.

The load average doesn't go down that quickly, and the intensiveness of
the work can make the load average _really_ high. Remember that you're
probably not seeing this because you do an atomic copy pretty quickly.
If you're compressing and writing 80% of ram to disk before the atomic
copy, what will your load average be at the end of that?

> > @@ -1253,9 +1254,18 @@ static inline void calc_load(unsigned lo
> >  	if (unlikely(count < 0)) {
> >  		active_tasks = count_active_tasks();
> >  		do {
> > -			CALC_LOAD(avenrun[0], EXP_1, active_tasks);
> > -			CALC_LOAD(avenrun[1], EXP_5, active_tasks);
> > -			CALC_LOAD(avenrun[2], EXP_15, active_tasks);
> > +			/*
> > +			 * If we let the load average be updated while
> > +			 * snapshot/shutdown, it will be very high
> > +			 * post resume. Processes such as some MTAs
> > +			 * that stop work while the average is high
> > +			 * will be unnecessarily disrupted.
> > +			 */
> > +			if (likely(!freezer_is_on())) {
> > +				CALC_LOAD(avenrun[0], EXP_1, active_tasks);
> > +				CALC_LOAD(avenrun[1], EXP_5, active_tasks);
> > +				CALC_LOAD(avenrun[2], EXP_15, active_tasks);
> > +			}
> >  			count += LOAD_FREQ;
> >  		} while (count < 0);
> >  	}
> 
> NAK. This slows down regular operation, at it is 30 lines for what
> should have been one (or five).

Count them. It is one line (an if statement) plus seven lines of
comment.

> Just place avenrun[0] = avenrun[1] = avenrun[2] = 0 at strategic place
> if you feel strongly about this. Additional points for using Rafael's
> "suspend done" notifier so that you don't have to modify suspend core.

=0 might not be right. If you did have a high load average prior to
suspending, not delivering email would be the right thing to do.

Nigel

[-- Attachment #1.2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 189 bytes --]

[-- Attachment #2: Type: text/plain, Size: 0 bytes --]