Re: PCI resource initialisation

On Sat, 26 Aug 2000, David Monro wrote:

Hi,

I need to make some changes to the way we do our PCI resource allocation
:-)

Currently the code goes like this (in pci.c:pcibios_init()):

        pci_scan_bus(0, &generic_pci_ops, NULL);
        if (ppc_md.pcibios_fixup)
                ppc_md.pcibios_fixup();
        pcibios_claim_resources(&pci_root_buses);

So we find all the devices, hack any we don't like, and then grab all
the resources. This means that we pretty much accept whatever the
firmware provides, which includes leaving any uninitialised resources
alone. This doesn't work properly on IBM 850 machines because the
firmware doesn't initialise any devices it doesn't recognise, and also
(for example) doesn't bother initialising the memory mapped resource of
the built-in ethernet (presumably because they expected people to use
the IO resource instead, which in fact we do).

For PrePboot I took the decision to reassign all PCI resources before te
kernel even ahd a chance to run. It works quite well for me, but I had no
choice since I basically switch from a PreP to a CHRP mapping. Actually
given the variety of hardware we are trying to supprt, I suspect that
pushing a few things into the bootloaders (like reading the OF tree if
available) can only be better.

The 2.4 kernel does provide a function,
pci_assign_unassigned_resources(), which will find pci resources which
haven't been claimed yet and do the appropriate thing with them.

Not always, unfortunately.

As I see it we have a few choices:

do:
	pci_scan_bus(0, &generic_pci_ops, NULL);
        if (ppc_md.pcibios_fixup)
                ppc_md.pcibios_fixup();
        pcibios_claim_resources(&pci_root_buses);
	pci_assign_unassigned_resources();

which should allocate any left alone by the original code, and does
work, but still means we have to hack the original resources (eg on IBM
850s the SCSI hos adapters seem to always get allocated IO port
resources above 0x20000000, which we relocate in a not very safe manner

You mean completely unsafe, it has caused hangs in some cases.

down to 0x01000000). It also assumes we don't need to do any more fixups
for the newly discovered resources.

or we could do:
	pci_scan_bus(0, &generic_pci_ops, NULL);
	pci_assign_unassigned_resources();
        if (ppc_md.pcibios_fixup)
                ppc_md.pcibios_fixup();

which results in the kernel completely reassigning all the PCI
resources. This should mean that the resources all get allocated in
areas that we would like to find them in, and means that the
pcibios_fixup code should be a little simpler. This is my preferred
solution, but I guess might break something.

Note that I can't really do this just for the PReP code; if I put the
pci_assign_unassigned_resources() call in the prep_pcibios_fixup()
function it will completely remap the bus and claim all the resources,
and then pcibios_claim_resources() will walk the tree trying to claim
each resource again and get a conflict on every resource!

Of course we could simply make pcibios_init() another arch-dependant
function - what do people think?

I am fed up with all these arch dependant functions, to the point that
I am convinced that ppc_md should disappear.

On a related, is it true to say that every driver which uses ioremap()
to map memory on cards into virtual space needs to have isa_mem_base
added to the base address before it will work? As far as I can tell
virtually no drivers do this at the moment; I note a rather
dirty-looking hack in chrp_pci.c where we hack the resource entries for
the matrox card so that the driver doesn't need rewriting. Or should we
simply do this for all memory resources on all cards?

Yes, you should do it. Actually I had something like this in my tree, but
it's now disabled since I switch the hardware to a CHRP map.

The resource code is still insufficient, there should be a way to have the
physical address, the bus address and in some cases the kernel virtual
address of a given resource. For I/O space, which is permanently mapped,
the resource should contain a kernel virtual address (the existence of
isa_io_base in itself is an aberration, have you counted how many lwz
instructions in the kernel access it ?). For memory resources, it should
contain a processor physical address which is the input to ioremap.

	Regards,
	Gabriel.


** Sent via the linuxppc-dev mail list. See http://lists.linuxppc.org/