When pci_assign_all_busses is set, the common code will re-assign all
the bus numbers, which we still need to do one various pmacs, at least
until XFree/XOrg understands PCI domains properly.  However, this
process triggers a bug if during this renumbering, on a given segment, a
P2P bridge not yet renumbered has conflicting downstream bus range with
a P2P bridge just renumbered.  The probing will just behave randomly and
either miss devices, show duplicates, or just lockup in some
circumstances.  This is typically triggered on Apple Xserve G4s who have
at least 2 P2P bridges built-in on the same segment of the first PCI
(non-AGP) domain

The "workaround" we use on pmac that was in my tree for a while but got
"lost" during the big merge a while ago is to offset the bus numbers
between domains by 16.  This avoids the above collision scenario in all
practical cases.

The long term solution of course is to stop renumbering and make sure
that all drivers are domain safe (there may be a couple issue remaining
in the code that matches PCI / OF devices on ppc32) and XFree/XOrg has
to be fixed.
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>