Re: [PATCH v3 4/4] mm/vmap: move BUG_ON() check to the unlink_va()

[Uladzislau Rezki]

Hello, Roman!

On Wed, May 29, 2019 at 04:26:43PM +0000, Roman Gushchin wrote:
> On Wed, May 29, 2019 at 03:58:17PM +0200, Uladzislau Rezki wrote:
> > Hello, Roman!
> > 
> > > > Move the BUG_ON()/RB_EMPTY_NODE() check under unlink_va()
> > > > function, it means if an empty node gets freed it is a BUG
> > > > thus is considered as faulty behaviour.
> > > 
> > > It's not exactly clear from the description, why it's better.
> > > 
> > It is rather about if "unlink" happens on unhandled node it is
> > faulty behavior. Something that clearly written in stone. We used
> > to call "unlink" on detached node during merge, but after:
> > 
> > [PATCH v3 3/4] mm/vmap: get rid of one single unlink_va() when merge
> > 
> > it is not supposed to be ever happened across the logic.
> > 
> > >
> > > Also, do we really need a BUG_ON() in either place?
> > > 
> > Historically we used to have the BUG_ON there. We can get rid of it
> > for sure. But in this case, it would be harder to find a head or tail
> > of it when the crash occurs, soon or later.
> > 
> > > Isn't something like this better?
> > > 
> > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > > index c42872ed82ac..2df0e86d6aff 100644
> > > --- a/mm/vmalloc.c
> > > +++ b/mm/vmalloc.c
> > > @@ -1118,7 +1118,8 @@ EXPORT_SYMBOL_GPL(unregister_vmap_purge_notifier);
> > >  
> > >  static void __free_vmap_area(struct vmap_area *va)
> > >  {
> > > -       BUG_ON(RB_EMPTY_NODE(&va->rb_node));
> > > +       if (WARN_ON_ONCE(RB_EMPTY_NODE(&va->rb_node)))
> > > +               return;
> > >
> > I was thinking about WARN_ON_ONCE. The concern was about if the
> > message gets lost due to kernel ring buffer. Therefore i used that.
> > I am not sure if we have something like WARN_ONE_RATELIMIT that
> > would be the best i think. At least it would indicate if a warning
> > happens periodically or not.
> > 
> > Any thoughts?
> 
> Hello, Uladzislau!
> 
> I don't have a strong opinion here. If you're worried about losing the message,
> WARN_ON() should be fine here. I don't think that this event will happen often,
> if at all.
>


If it happens then we are in trouble :) I prefer to keep it here as of now,
later on will see. Anyway, let's keep it and i will update it with:

<snip>
    if (WARN_ON(RB_EMPTY_NODE(&va->rb_node)))
        return;
<snip>

Thank you for the comments!

--
Vlad Rezki