[PATCH v2 19/28] ARM: Add fixed PCI i/o mapping

[arnd@arndb.de (Arnd Bergmann)]

On Wednesday 29 February 2012, Russell King - ARM Linux wrote:
> On Wed, Feb 29, 2012 at 10:43:13PM +0000, Arnd Bergmann wrote:
> > On Wednesday 29 February 2012, Rob Herring wrote:
> > > No, then the mapping will fail. I do need to double check all the
> > > mappings and make sure there is no overlap but we only care on the PCI
> > > platforms I converted.
> > > 
> > > Also, it should be noted this shrinks the i/o space on some platforms.
> > > dove and kirkwood had 1MB x 2 buses and now have 64KB per bus. ixp2000 I
> > > think has 32MB with a note saying they need "a lot". Is there really a
> > > use for lots of i/o space?
> > 
> > Given that each PCI bus only wires up the lower 64k, I'd say no ;-)
> 
> That's incorrect.  There's no "wires up" option there.  Addresses and
> data are transferred over a common set of 32 wires on 32-bit PCI, and
> the bus has separate 'address' and 'data' phases.  If you only wire up
> 16 of those lines, then things aren't going to work (because you won't
> be able to access bytes 2,3 of each word.)
> 
> However, there's a separate issue here:
> (1) do peripherals decode the upper 16 bits programmed into their IO BARs?
>     Some do.
> (2) are host PCI controllers capable of generating IO accesses in the
>     lower 64k bus address region?

Very interesting, I had no idea and when I looked up the PCI spec, I
found this PCI-3.0, section 6.2.5.1:

"The upper 16 bits of the I/O Base Address register may be hardwired to
 zero for devices intended for 16-bit I/O systems, such as PC compatibles.
 However, a full 32-bit decode of I/O addresses must still be done."

So it is indeed valid to have a larger I/O space, although in practice
I would expect this not to get used because in an x86 PC it is impossible
to issue an instruction that accesses an I/O port beyond 0xffff.

> I think while I was fiddling with an IOP platform years ago that it did
> not work with PCI setup to only use the lower 64k bus address region
> for IO accesses.  So I'd suggest a certain amount of care is required
> here.

Ok. I've looked up the manual for IOP134x and it's very clear there that
it has only two 64k I/O windows but no more (developer manual, sections
2.2.2.1 and 3.3.2.1). Limiting the virtual space to 64k would result in
seeing only one of the buses, but 1 MB is enough if we map both buses
in there.

Similarly, the IOP333 defines one outbound I/O space window of a fixed
length of 64kb (processor developer's manual section 3.10.33) and won't
need more than that.

Any idea which IOP you might have seen this on?

> > My guess is that they used 1MB ranges in order to benefit from section
> > mapping. In case of ixp2000, I only see 64k in the resource:
> > 
> > static struct resource ixp2000_pci_io_space = {
> >         .start  = 0x00010000,
> >         .end    = 0x0001ffff,
> >         .flags  = IORESOURCE_IO,
> >         .name   = "PCI I/O Space"
> > };
> > 
> > The part that I don't understand here is why the resource starts at
> > 64k and is another 64k in size. I think we need to double-check this
> > in order to be sure whether we have to put the pci io space into the
> > first or the second 64k chunk of the new mapping area.
> > 
> > Hmm, I guess you meant ixp23xx, not ixp2000, which indeed has
> > 
> > static struct resource ixp23xx_pci_io_space = {
> >         .start  = 0x00000100,
> >         .end    = 0x01ffffff,
> >         .flags  = IORESOURCE_IO,
> >         .name   = "PCI I/O Space"
> > };
> > 
> > This seems to be done just for simplicity in the implementation,
> > to keep all parts of the PCI controller 32MB aligned, I can't
> > see any real technical reason why it would be useful othewise.
> 
> It's certainly a weird way to avoid ISA addresses.  Note that these
> will control the set of bus addresses which get assigned within PCI.

Are you referring to ipx2000 or the ixp23xx here?

	Arnd