Re: [PATCH] get_pid() performance fix

[Hubertus Franke <frankeh@watson.ibm.com>]

On Saturday 16 March 2002 12:12 am, OGAWA Hirofumi wrote:
> Hubertus Franke <frankeh@watson.ibm.com> writes:
> > On Friday 15 March 2002 10:16 am, OGAWA Hirofumi wrote:
> > > Whoops! I'm sorry. previous email was the middle of writing.
> > >
> > > Hubertus Franke <frankeh@watson.ibm.com> writes:
> > > > +	if (i == PID_MAP_SIZE) {
> > > > +		if (again) {
> > > > +			/* we didn't find any pid , sweep and try again */
> > > > +			again = 0;
> > > > +			memset(pid_map, 0, PID_MAP_SIZE * sizeof(unsigned long));
> > > > +			last_pid = RESERVED_PIDS;
> > > > +			goto repeat;
> > > > +		}
> > > > +		next_safe = RESERVED_PIDS;
> > > > +		return 0;
> > >
> > > Probably, the bug is here. No bug ....
> > >
> > >
> > >   +	next_safe = RESERVED_PIDS;	/* or 0 */
> > >
> > > > +	read_unlock(&tasklist_lock);
> > > > +	spin_unlock(&lastpid_lock);
> > > > +	return 0;
> > > >  }
> > >
> > > Basically nice, I think.
> > >
> > > BTW, How about using the __set_bit(), find_next_zero_bit(), and
> > > find_next_bit() in get_pid_by_map().
> > >
> > > Thanks for nice work.
> >
> > OGAWA, honestly I only tried testcase 2.
> > But looking at your suggestion its not clear to me whether
> > there is a bug.
> > Remember we need to determine a valid interval [ last_pid .. next_safe ).
> > In the pid_map function, if no pid is available, then
> > [ PID_MAX .. PID_MAX ) will be returned.
> > The other path should also end up with this as well.
> > Could you  point where you see this not happening.
>
> Maybe my point was unclear. Sorry.
>
> Please consider what happens after using up pid. Then,
> get_pid_by_map() returns 0.
>
> And last_pid = 0, next_safe = RESERVED_PIDS.  After it, get_pid()
> returns the values between 0 and RESERVED_PIDS.
>
> And the line which I added is also the same reason.
>
> Regards.

Hirofumi, yes you are right, "the thought was there, but the implementation 
was weak".
Proofs again, all these benchmarks are not enough. I'll make a new patch 
soon and resubmit.
-- 
-- Hubertus Franke  (frankeh@watson.ibm.com)