Re: [PATCH 0/1] Recurse when searching for empty slots in resources trees

[Andrew Patterson <andrew.patterson@hp.com>]

On Tue, 2009-06-16 at 16:05 -0700, Linus Torvalds wrote:
> 
> On Tue, 16 Jun 2009, Andrew Patterson wrote:
> > 
> > That is at least one problem. I initially tried reparenting this stuff.
> > That is what got backed out in
> > http://thread.gmane.org/gmane.linux.kernel/768526/
> 
> Well, aren't we in the exact same situation still? Ie the problem (as 
> Matthew claims) is:
> 
>   'Basically it was that we came across a machine with the opposite 
>    problem -- that we found a parent after we found a child (and claimed 
>    the child's resources), and had no way to insert the parent's region 
>    above the child's region.  Alex's machine finds the child after the 
>    parent and needs to insert the child's resource inside the parent's 
>    resource.'
> 
> and the problem is that anything that isn't explicitly aware of the 
> topology is always going to be potentially confused about things like 
> this, since it's not clear at which level you want to find or add a 
> resource.
> 
> > > But you fix it by making find_resource always go as deep as it can (if I 
> > > read the code correctly).
> > 
> > Well, just deep enough.
> 
> Ok, color me confused now. When is "as deep as it can" different from your 
> "just deep enough"?

Maybe confusion on what is meant by 'as deep as'.  My patch continues
until it doesn't find a conflict including checking sub-children and
stops as soon as an appropriate resource is found that does not
conflict.  Perhaps we mean the same thing.

> 
> > Is there a reason that find_resources should stop at the roots immediate
> > child/sibling.  It seems like a bug to me. Hence this patch.
> 
> Well, find_resource() found room for a resource. So it returns it. The 
> point is, your patch returns another - equally valid one.

I am confused.  The existing code will return a conflict and bomb out.

> 
> Now, I'm not saying that your patch is wrong, but I _am_ worried that it 
> (once more) changes some random heuristic when we have two choices, and it 
> just makes it choose the other choice.

Agreed.

> 
> We've had those kinds of situations before. The thread you point to is an 
> exact case of this. My point is that I'd rather try to _avoid_ any 
> ambiguous cases, and try to solve it properly at a higher PCI level, where 
> the ambiguity doesn't exist any more (because we'd explicitly take the 
> actual bus topology into account).
> So your patch may fix a bug, but I'm pretty sure I've seen a patch from 
> Ivan that should _also_ fix it, and that I would expect to do it not by 
> just tweaking a fundamentally ambiguous case.
> 

OK. I would be happy to test Ivan's patch.

> 		Linus
> 
-- 
Andrew Patterson
Hewlett-Packard