Re: 7.0_RC1 slow boot

[Alan Barrett <apb%netbsd.org@localhost>]

Sorry, I have been neglecting NetBSD for a while.

On Mon, 13 Jul 2015, Eric Haszlakiewicz wrote:
> On July 13, 2015 6:14:47 AM EDT, Arto Huusko <armihu%gmail.com@localhost> wrote:
> > On 12.07.2015 21:02, Martin Husemann wrote:
> > > On Sun, Jul 12, 2015 at 08:47:13PM +0300, Arto Huusko wrote:
> > > > I also tested 7.99.19, and it is just as slow to boot as 7.0_RC1,
> > and
> > > > vmstat numbers are pretty much the same.
> > >
> > > See:
> > > http://mail-index.netbsd.org/tech-userlevel/2015/06/01/msg009153.html
> >
> > That helped, but only about 1-2 seconds, so with that 7.0_RC1 
> > boot is 43 seconds compared to 11 seconds of 6.1.5. But I 
> > looked at other rc changes in 7.0, and reverting revision 1.94 
> > of etc/rc.subr restored the boot speed back to the same level 
> > as in 6.1.5, and number of forks dropped back from over 1000 to 
> > less than 300.
> >
> > At a glance the diffs do look like they should not cause any 
> > forks, but maybe someone should take a closer look.
>
> That diff appears to replace a simple "-n" test with a function 
> that does an eval, a kill and a expr, the last with an embedded 
> command substitution to run ps, so it'll definitely cause more 
> forks.  It's aimed at determining whether some pid is alive, but 
> it seems like an awful lot of work to be doing just for that.

It's not just checking whether a pid is alive, it's checking whether the
PID represents a shell running /etc/rc, to guard against pids being
recycled.  That part is probably unnecessary.

I can also reduce the number of times _have_rc_postprocessor is called.

--apb (Alan Barrett)