Re: exit1()/scheduler question.. possible typo?

On Mon, 10 May 2004, Bruce Evans wrote:

> On Sun, 9 May 2004, Julian Elischer wrote:
>
> > in exit1 there is an assertion to test that the exiting process
> > is not init. (proc 1 by tradition..)
> >
> > yes later, nearly at the end we see:
> >
> >         /*
> >          * Allow the scheduler to adjust the priority of the
> >          * parent when a kseg is exiting.
> >          */
> >         if (p->p_pid != 1)
> >                 sched_exit(p->p_pptr, td);
> >
> >
> > firstly, the comment is wrong but, the question comes..
> > "if init can not get here then why have the test?"
>
> ...
> > I get the impression that possibly this should be p->p_pptr->p_ppid
>                                                                p_pid
> > but I don't know enough about ULE to know if that makes sense.
> >
> > (maybe it should be (p->p_pptr != initproc)
>
> My version already used initproc here and elsewhere, but was missing the
> fix for the LHS.

I tried using (p->p_pptr != initproc), but this broke load averages and
thus caused problems like sendmail refusing to run.  We were depending
on the bug to get sched_exit() called unconditionally.  The call must
now always be made, because one of the things that sched_exit() does
for the 4BSD scheduler is decrement the load average counter.

The nearby call to cpu_sched_exit() is also bogus:

% 	cpu_sched_exit(td); /* XXXKSE check if this should be in thread_exit */
% 	/*
% 	 * Allow the scheduler to adjust the priority of the
% 	 * parent when a kseg is exiting.
% 	 */
% 	if (p->p_pid != 1)
% 		sched_exit(p->p_pptr, p);

Scheduling is not cpu-specific, and if it were then it should be done
by sched_exit() calling cpu_sched_exit().  In practice, cpu_sched_exit()
is null for all arches except sparc64.  For sparc64, it does things
related to exiting but unrelated to scheduling.  ISTR you asking about
sparc64's cpu_sched_exit() separately.  On i386's, things related to
exiting are done in cpu_exit() and cpu_thread_exit().  This seem to
be the right places to do such things.  However, things are misplaced
internally there.  The debugger registers are per-thread, so they
should only be reset in cpu_thread_exit().

The first senetence of the nearby comment on the call to thread_exit()
is also bogus:

% 	/*
% 	 * Make sure the scheduler takes this thread out of its tables etc.
% 	 * This will also release this thread's reference to the ucred.
% 	 * Other thread parts to release include pcb bits and such.
% 	 */
% 	thread_exit();

This call to thread_exit() doesn't do anything related to scheduling,
since p->p_numthreads is 1.

Bruce