Re: [PATCH 2/2] powerpc - Make the irq reverse mapping radix tree lockless

[Sebastien Dugue <sebastien.dugue@bull.net>]

On Thu, 04 Sep 2008 17:34:03 +1000 Benjamin Herrenschmidt <benh@kernel.crashing.org> wrote:

> 
> > > >   I could not think of anything simple so far and I'm open for suggestions.
> > > 
> > > GFP_KERNEL should not fail, it will just block no ?
> > 
> >   No it won't block and will fail (returns NULL).
> 
> hrm... it used to never fail.. that may have changed. But it will
> definitely block and try very hard to push things out to make space,
> which is the whole point :-)

  Right, but it still can fail - and you're right, we're in trouble already.
Last time I dug into slab code I got lost into the maze :(

> 
> >   I will have to add that back as there is no more fallback.
> 
> Well, the must be one in the case the tree isn't initialized yet,

  

> so if
> there's an allocation failure, you may "de-initialize" it or
> something...

  There's nothing to 'de-initialize' here, or am I missing something?
radix_tree_insert() will return ENOMEM and won't insert anything.

> Or you can fallback if you don't find, as easy, probably
> easier since it shouldn't happen in practice.

  That's what I had in mind.

> 
> > > I don't know if it's worth trying to fire off a new
> > > allocation attempt later, probably not.
> > 
> >   I've been pondering with this lately, but I think that adding a linear
> > lookup fallback should be OK.
> 
> Yup.
> 

  Thanks Ben,

  Sebastien.