Currently our code to set up the data structures for a PCI host bridge
and create the mapping for its I/O window assumes that the window
starts at I/O port 0 on the PCI side. If this is not true, we can end
up with I/O port numbers in the resources for PCI devices which will
cause an oops if a driver tries to access them via inb/outb etc.,
because there is no mapping for the corresponding addresses.
Normally the I/O window starts at 0, but there are some situations on
partitioned machines with a hypervisor where the window may not start
at 0.
This fixes the problem by allocating space for the range from 0 to the
end of the I/O window. That is, hose->io_base_virt contains the
virtual address for I/O port 0 on the PCI bus, and thus the assumption
that hose->io_base_virt - pci_io_base is the offset between the
"global" I/O port numbers (those in the PCI device resources) and the
I/O port numbers on the PCI bus is maintained.
For PCI host bridges that are present at boot, we only map the portion
of that range that correspond to the bridge's I/O window. For bridges
added after boot we ioremap the range from 0 to the end of the I/O
window, for now; in fact hot-added bridges should be using
reserve_phb_iospace() and __ioremap_explicit (so they get sensible
global port numbers), but we don't have the infrastructure yet to do
that (basically a free_phb_iospace() routine plus appropriate
locking).
Interestingly, this makes the two arms of the if statement in
get_bus_io_range do almost exactly the same thing; that function could
now be simplified in a further patch.
Signed-off-by: Paul Mackerras <paulus@samba.org>
11fbb00c67