Re: [parisc-linux-cvs] grundler

[Philipp Rumpf <prumpf@suse.de>]

> Update of /home/cvs/parisc/linux/include/asm-parisc
> In directory puffin.external.hp.com:/tmp/cvs-serv4014/linux/include/asm-parisc
> 
> Modified Files:
> 	pci.h 
> Log Message:

> Update GSCtoPCI (Dino) bus adapter support. Have most of the design
> in place and what's my "first cut" implementation.

I don't want to sound aggressive, but I really do not see the point in most
of your changes.

Let's go through them:

 - renamed pci_{port,config,bus}_ops to {port,config,bus}_ops.

waste of struct namespace should asm/pci.h be included accidentally
(furthermore it strikes me as strange to have pci_bios_ops, but simply bus_ops.

 - typedef unsigned int irq_t;

the type of irqs is defined to be int (not unsigned int) in ABIs we cannot
change if we ever want to see a merged parisc kernel.

 - renamed PCI_NUM_BUS to MAX_BUSSES.

I think we really should restrict ourselves to exporting symbols, macros and
other names starting with PCI_ and pci_.  (I am aware this may be because it
took me way too long to synch the trees and that this just reflects the rest
of your code).

> Marked lots of stuff with "TODO" - I'll keep whacking at these.

 - "** TODO: I don't think we need the "bus" parameter. "dev""

ignoring your comment style ;), it really strikes me as a good idea to keep
the interface in place to support native-to-PCI bridges with more than one
bus on them.  It might be a good idea for those not to pass the "relative"
bus number to those instead of the absolute one, but that is just a change
in one spot as opposed to changing an interface and digging through several
drivers.

> Main issues to still be resolved are:
> o GSC device claim - how does dino driver learn about GSC devices

For that we need Alex's code, but it looks very much like we will have a list
of GSC devices like we currently have for PCI on systems where that is the
native bus.  Then, you will use gsc_find_device(u16 hversion); and be happy.

> o Registration and handling of interrupts (request_irq() support)

None.  The current idea is dino.c registers a region of 32 "virtual" IRQs (i.e.
Linux IRQ cookies, not EIRR bits) and uses do_irq_mask to handle them.  Also,
pcibios_fixup will put the right irq cookie numbers in the pci_devices list.

> o How drivers will use I/O port space accessor functions.
>   (assumes we won't change them to use MMIO space)

basically,

u8 inb(u32 address) {
	return pci_ops[(address>>16)].port.inb(
		pci_devs[(address>>16],
		(u8) (address>>16) / pci_rel_bus[(address>>16)],
		(u16) adress);
}

that looks slow but I hope for the "port accesses are slow anyway" argument
to win.

	Philipp Rumpf