> Date: Wed, 8 Jul 2015 16:35:09 +0200 (CEST)
> From: Mark Kettenis <mark.kette...@xs4all.nl>
>
> > Date: Wed, 8 Jul 2015 15:17:46 +0100
> > From: Stuart Henderson <st...@openbsd.org>
> >
> > On 2015/07/08 15:53, Mark Kettenis wrote:
> > > Index: uvm_map.c
> > ..
> > > @@ -2466,8 +2470,7 @@ uvm_map_teardown(struct vm_map *map)
> > > if ((entry = RB_ROOT(&map->addr)) != NULL)
> > > DEAD_ENTRY_PUSH(&dead_entries, entry);
> > > while (entry != NULL) {
> > > - if (waitok)
> > > - uvm_pause();
> > > + sched_pause();
> >
> > ah, slightly different than the one I've tried before which
> > had uvm_pause instead of sched_pause here plus an extra change
> > in uvm_glue.c.
>
> Just a bit of a cleanup though. Not going to make a difference.
>
> > I'll give this a spin on i386 after my current build finishes.
> > FWIW I've been running the previous version on amd64 for ages with
> > very positive results.
>
> I won't stop you ;). But my goal was to trick others into testing
> this diff such that this doesn't hamper the i386 ports builds.
And thanks to Stuart, we have found a (fairly straightforward) lock
order reversal. I'm working on a fix.
Thanks for everybody who tested so far. Feel free to continue
testing, although you might want to be a bit careful on i386 systems
as they are somewhat likely to lock up with this diff.
