Bug #1 (found by Jay Estabrook).
On Alpha, under certain circumstances the firmware may close the IO
window of PCI-to-PCI bridge even if there is IO behind.
This wouldn't be a problem - linux PCI setup code does set up this
window properly, but in addition the firmware clears the IO-enable
bit in the PCI_COMMAND register of the bridge.
Since we don't call pci_enable_* routines for bridges in non-hotplug
path, we end up with disabled IO. Fixed by adding pci_enable_bridges()
to pci_assign_unassigned_resources().
Architectures which don't use the latter, but do use other setup-bus
code (parisc?) also should call pci_enable_bridges() for each root bus.

Bug #2 (closely related to #1).
As it turns out, pci_enable_device() doesn't work for bridges at all,
only for regular devices (header type 0) due to 0x3f mask passed to
pci_enable_device_bars(). The mask should be (1 << PCI_NUM_RESOURCES) - 1.

Bug #3 (quite a few archs, including i386).
pcibios_enable_device() does only check first 6 resources (regardless
of the mask) to decide whether or not to enable IO and MEM.
Bridge resources start at 7.

#2 and #3 affect hotplug. I wonder, has anybody ever tried *bridged*
PCI card behind a hot-plug controller?