[PATCH 04/14] bus: mvebu-mbus: Add static window allocation to the DT binding

[Arnd Bergmann]

On Tuesday 11 June 2013 16:58:41 Jason Gunthorpe wrote:
> On Wed, Jun 12, 2013 at 12:34:14AM +0200, Arnd Bergmann wrote:
> 
> > a significant waste of physical address space, because the (per-soc)
> > ranges property has to be set up for the largest possible external
> > device connected to the bus, but the mbus window only needs to cover the
> > device that is actually connected.
> 
> Yes, but perhaps that is a good argument to not put the ranges in the
> dtsi?

Fair enough. In that case not using ranges at all is probably easier
so the board file only needs to update one place.

> > I think we have to walk the entire tree of devices underneath mbus,
> > at least any device node that has a 'reg' property, following children
> > of any device node with a 'ranges' property. We might need to
> > add a variant of of_get_address() that does a partial translation
> > up to a given node (the mbus) instead of all the way to the root.
> 
> Hum, how complex do you think this is? Run down the entire tree,
> translate every regs and maintain a max offset for every target id.

It's the same thing we do for PCI.

> > > then dynamic assignment of that window
> > > makes sense to me. (however, this also looks a bit tricky, how do you
> > > avoid hitting the PCI-E window reservations, for instance?)
> > 
> > The PCI-E window would have 'ranges' but no 'regs', so we automatically
> > skip it.
> 
> The trouble is not skipping PCI-E when parsing the dt. When the
> kernel goes to make a dynamic window allocation it needs to exclude
> the address space reserved for the PCI-E aperture.
> 
> The mbus driver starts before the PCI-E driver, so how can this be
> resolved? Hardwire specialness about PCI-E into mbus?
> Make the PCI-E aperture fully dynamic?
>
> Starting to be pretty complex - for what gain?

I think this can be simplified a lot by making PCI a known special case:

We assign all regular (non-PCI) windows from the top, and whatever is
left by the time PCI comes up can be used for that, starting at the
bottom of the avaliable space.
 
> > > I'm not sure there is a good reason to reject the address map in the
> > > DT?
> > 
> > You mean keep all windows that are listed by the boot loader but
> > reassign the others? I guess it would be simpler but may result in
> > a less optimal address map.
> 
> Sure, but what does a more optimal address map get you?
> 
>  1) More space to map SDRAM - however - the SDRAM map is controlled
>     exclusively by the bootloader and the kernel doesn't have the
>     information or the code to mess with it - so this isn't possible
>  2) More space for the PCI-E aperture - this is entirely controlled by
>     the kernel, but today we are using the pre-set allocation from the
>     DT, so it doesn't matter how tightly the rest of the stuff is packed.
> 
> To me, it feels like alot of complex trouble to do actual dynamic
> allocation, and I don't really see any gain.

We could of course decide to skip the dynamic part in the Linux
implementation for the moment, but I would definitely want to
see the binding written in a way that going fully dynamic can be
done later without changing the binding if we decide we need to
squeeze out more RAM or PCI space.

	Arnd