Earlier cleanup efforts of the ppc64 IOMMU code have mostly been targeted
at simplifying the allocation schemes and modularising things for the
various platforms.  The IOMMU init functions are still a mess.  This is an
attempt to clean them up and make them somewhat easier to follow.

The new rules are:

1. iommu_init_early_<arch> is called before any PCI/VIO init is done
2. The pcibios fixup routines will call the iommu_{bus,dev}_setup functions
   appropriately as devices are added.

TCE space allocation has changed somewhat:

* On LPARs, nothing is really different. ibm,dma-window properties are still
  used to determine table sizes.
* On pSeries SMP-mode (non-LPAR), the full TCE space per PHB is split up
  in 256MB chunks, each handed out to one child bus/slot as needed. This
  makes current max 7 child buses per PHB, something we're currently below
  on all machine models I'm aware of.
* Exception to the above: Pre-POWER4 machines with Python PHBs have a full
  GB of DMA space allocated at the PHB level, since there are no EADS-level
  tables on such systems.
* PowerMac and Maple still work like before: all buses/slots share one table.
* VIO works like before, ibm,my-dma-window is used like before.
* iSeries has not been touched much at all, besides the changed unit of
  the it_size variable in struct iommu_table.

Other things changed:
* Powermac and maple PCI/IOMMU inits have been changed a bit to conform to
  the new init structure
* pci_dma_direct.c has been renamed pci_direct_iommu.c to match
  pci_iommu.c (see separate patch)
* Likewise, a couple of the pci direct init functions have been renamed.
Signed-off-by: Olof Johansson <olof@austin.ibm.com>
Acked-by: Paul Mackerras <paulus@samba.org>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>