Initial version of PCI access library

[Christoph Hellwig]

On Mon, Feb 20, 2006 at 11:42:38AM -0800, Ian Romanick wrote:
> > I think the most important thing about the driver interface is that
> > it should be a callback scheme where the core calls the driver for
> > every found devices instead of the driver calling some kind of iterator
> > code.  From experience with Linux kernel drivers the latter leads to
> > rather horrible spaghetti code, not to mention that such a scheme
> > doesn't work at all for dyanmic devices additions or removals.
> 
> I agree 100%, but SOMETHING still has to iterate over the list of
> devices.  If the devices aren't doing it, then the X server is.

Yes.  But if only the server does it the way they are iterated can
be changed arbitrary or even be different for different platforms.

> My current understanding is that right now each driver has a probe
> function.  The server calls that probe function to ask the driver which
> devices it can handle.  This includes PCI, AGP, ISA, SBUS, and Zorro
> devices.  The driver marks devices as "claimed" and does some per-device
> initialization.  Based on discussions from the past couple weeks, I
> believe that this is backwards.
> 
> For PCI, AGP, and other busses with enumeratable device IDs (PCMCIA?
> USB?),

sbus (although davem prefers to enumerate them in kernelspac based on
      his last mail)
mca
eisa
isapnp

although I doubt anything but pci(,-X,-express) and agp which is
identical to PCI at the device enumeration level really matters
a lot for X currently.  The Linux kernel has some generic infrastucture
used by the various bus drivers to share code for this, but IMHO that
would be complete overkill for X right now.
currenty.

> the driver should have a list of devices IDs and a probe
> function.  If the server finds a device that is in the driver's list, it
> calls the driver's probe function with a pointer to the device.  That
> driver now owns that device.

Yes.

> We'll need another probe function for those older busses.  Fortunately,
> most drivers don't support any devices on those types of busses, so code
> to support that entry point won't be needed very often.

Can't we just keep the old code for those busses for now?  In the Linux
kernel a driver can use modern style PCI probing while still doing it's
old and awkward unstructured poke and see ISA probing.