Re: [RFC][PATCH 1/7] rbtree: Add generic add and find helpers

[Juri Lelli <juri.lelli@redhat.com>]

On 30/04/20 00:51, Michel Lespinasse wrote:
> On Thu, Apr 30, 2020 at 12:28 AM Juri Lelli <juri.lelli@redhat.com> wrote:
> > > --- a/include/linux/rbtree.h
> > > +++ b/include/linux/rbtree.h
> > > @@ -141,12 +141,18 @@ static inline void rb_insert_color_cache
> > >       rb_insert_color(node, &root->rb_root);
> > >  }
> > >
> > > -static inline void rb_erase_cached(struct rb_node *node,
> > > +static inline bool rb_erase_cached(struct rb_node *node,
> > >                                  struct rb_root_cached *root)
> > >  {
> > > -     if (root->rb_leftmost == node)
> > > +     bool leftmost = false;
> > > +
> > > +     if (root->rb_leftmost == node) {
> > >               root->rb_leftmost = rb_next(node);
> >
> > Think we need
> >
> >  if (root->rb_leftmost)
> >
> > > +             leftmost = true;
> >
> > DEADLINE crashes w/o that.
> 
> I think Peter's code is correct; after removing the only node in an
> rbtree rb_leftmost should be NULL.

Indeed, I've only got the idea that Peter was thinking of using
rb_erase_cached return value as an indication that new rb_leftmost is
not NULL (and for example perform an update in DEADLINE earliest_dl).

I also had the impression that DEADLINE is actually the only consumer of
that return value and so we were able to define its semantic.

> The issue appears to be in dequeue_pushable_dl_task unconditionally
> dereferencing the pointer returned by rb_first_cached(), which may be
> NULL. I'm not sure what the correct behavior is though, i.e. what
> dl_rq->earliest_dl.next should be set to if the rbtree ends up empty.
> Current code (before Peter's changes) preserves the existing
> dl_rq->earliest_dl.next value in that case, which seems very weird to
> me (and worthy of a comment if it's correct).

But, yeah. Fixing things in DEADLINE code works for me as well. We could
reset it to 0 (initial value), but I now actually wonder if places where
that is consumed are actually OK. Different discussion, though. :-)

Thanks,

Juri