[PATCH v2 19/27] pci: PCIe driver for Marvell Armada 370/XP systems

[Jason Gunthorpe]

On Wed, Feb 06, 2013 at 06:18:52PM +0100, Thomas Petazzoni wrote:
> Dear Jason Gunthorpe,
> 
> On Wed, 6 Feb 2013 10:09:03 -0700, Jason Gunthorpe wrote:
> 
> > Link 0 translates like:
> > 
> > - Linux driver does inl(0x4)
> > - ARM layer converts that into a read from PCI_IO_VIRT_BASE + 0x4
> > - The CPU TLB converts that into a read from CPU physical
> >   0xc0000000 + 0x4
> > - The MBUS window remap register converts that into a read from IO
> >   space 0x4
> > - The address 0x4 is placed in the PCI-E IO transaction of link 0
> > 
> > Link 1 translates like:
> > 
> > - Linux driver does inl(0x1004)
> > - ARM layer converts that into a read from PCI_IO_VIRT_BASE + 0x1004
> > - The CPU TLB converts that into a read from CPU physical
> >   0xc0000000 + 0x11004 (ie the mbus window for the link 1)
> > - The MBUS window remap register converts that into a read from IO
> >   space 0x1004
> > - The address 0x1004 is placed in the PCI-E IO transaction of link 1
> 
> This last step is exactly what I thought would not work. If the PCIe
> device has say 64 bytes of I/O space, then this 0x1004 PCI-E
> transaction will be out of bounds, no?

No.. PCI end devices are required to decode all 32 bits of address,
less the bits requires for their allocation. So a device with 64 bytes
of IO will match bits 31:6 and then use bits 5:0 for the internal
register.

So a full 32 bit address is technically fine, according to the spec,
however:
 - The 32 bit decode is an optional feature on bridges
 - Some devices are broken because x86 only uses the low 64k.

So for best compatibility it is ideal to put everything in the low
64k.

However, IO space really should not be used by anything except very
legacy devices, so if the MMU business is a hassle for some reason I'd
just go with the 64k aligned method.

> > > Can't this be solved using the window_alignement() hook we've been
> > > discussing separately? Just like we teach the Linux PCI core about our
> > > alignment requirements of 64K for the I/O regions, we could teach it
> > > about our alignment requirement on memory regions as well. No?
> > 
> > Hopefully :) As long as it can adjust the start and length you should
> > be fine.
> 
> Why would you need to adjust the length? If Linux allocates a 2 MB
> resource on a 1 MB boundary, we simply increase the start address to
> the next 2 MB boundary, and that's it. Why would the length need to
> change?

Well, lets say 3MB is the example. A 3mb region needs to fit inside a
4mb MBUS window. If you align the start to 4mb then the pci-e core
needs to know that it can't use the extra 1mb covered by the mbus
window. mbus windows must not overlap.

Adjusting the bridge window length to be 4mb communicates that dead
space to the PCI core, and presumably this shows up in lspci and
whatnot.

I suppose if you align the end to 4mb (thus creating the gap before,
not after) things should work out OK, but the information that the gap
is routed to a specific PCI link is lost..

Jason