On Tue, 03 Jun 2014 11:21:10 +0200, Arnd Bergmann [off-list ref] wrote:

On Tuesday 03 June 2014 09:44:59 Grant Likely wrote:

quoted

quoted

The reason I think allow an ECAM makes sense in ranges is because it allows for a direct IO read/write to CFG space (w/o any mapping) similar to what one would do for MEM space or IO.

I don't think that's right. PCI addresses are defined as follows:
phys.hi cell: npt000ss bbbbbbbb dddddfff rrrrrrrr
phys.mid cell: hhhhhhhh hhhhhhhh hhhhhhhh hhhhhhhh
phys.low cell: llllllll llllllll llllllll llllllll

where 'ddddd' is the device number (0-31) and 'fff' is the function number (0-7)

Going up by one device number or even function number does not result in
contiguious address values:

device 0: 0x00000000 00000000 00000000
device 1: 0x00000800 00000000 00000000
device 2: 0x00001000 00000000 00000000
device 3: 0x00001800 00000000 00000000
...
device 30:0x0000f000 00000000 00000000
device 31:0x0000f800 00000000 00000000

a simple ranges doesn't work transparently because each of those config
ranges needs to be mapped to a 4k block. I think ranges would need to
look like this:

ranges = <0x00000000 0 0  0x0ff00000  0x1000>,
         <0x00000800 0 0  0x0ff01000  0x1000>,
         <0x00001800 0 0  0x0ff02000  0x1000>,
         ...
         <0x0000f000 0 0  0x0ff1e000  0x1000>,
         <0x0000f800 0 0  0x0ff1f000  0x1000>;

(I just hacked the above up; I make no claims to it's accuracy for
actual address values)

But I don't even thing the semantics work there because the address is
encoded in the phys.hi cell, not the phys.low cell. Incrementing by one
does not behaves as most bus addresses work. To actually work properly
we would have needed a way to define a stride of 64bits when
incrementing config space addresses in a ranges mapping.

Thanks for clearing that up. I always suspected it was roughly this
way, but never managed to think it through completely before getting
distracted by something else.

I wonder if the OF definition matches CAM though, if not ECAM, as
CAM is also limited to 256 byte config space per function.

It's the same problem for 256 byte entries. The address values don't
increment nicely and there is a big block of remapping needed.

g.