Re: Work on PCI support :-)

[Radosław Kujawa <radoslaw.kujawa%c0ff33.net@localhost>]

Hi again.

As you have already (probably) seen, I'm working on a driver for PCI bus 
present on CyberVisionPPC and BlizzardVisionPPC.

The basics are working - It is now possible to access the configuration 
registers through PCI MI code, but I'm still facing a few problems, which I am 
not sure how to solve:

- Implementation of bus_space_read_4 and write_4 is missing. It is used often 
in the PCI MI code, and most importantly in pm2fb.c. Ofcourse the only proper 
solution is to implement the missing functions functions, but I feel a bit lost 
with all these dbsdr/dbsdw/dbsm/dbss/bsr/bsw/bswm/bsrm/wtf...
- PCI MI code needs bus_space_mmap. For now I've just faked it, but it should 
be implemented to allow framebuffer mmaping. Not sure how to do it. I've looked 
atari's implementation and it seems not that complicated, but not don't know if 
it will work the same way on amiga port.
- PCI MI code needs bus_dma support, for now I've just included m68k/bus_dma.h 
in pci_machdep.h.
- CVPPC/BVPPC are very tricky to detect. Newer versions of the firmware 
autoconfigure the card like this:

Board (phase 5): Prod=8512/101($2140/$65) (@$FFFE0000 64K)
Board (phase 5): Prod=8512/101($2140/$65) (@$FFFA0000 64K)
Board (phase 5): Prod=8512/101($2140/$65) (@$FFFC0000 128K)
Board (phase 5): Prod=8512/101($2140/$65)
(@$E0000000, size RESERVED, subsize same)
Board (phase 5): Prod=8512/101($2140/$65)
(@$E0800000, size RESERVED, subsize same)
Board (phase 5): Prod=8512/101($2140/$65) (@$E1000000 128K)

The same vendor/product ID is used 6 times. Also, it is possible that the last 
entry is different on some configurations (EF000000 instead of E1000000). 
Unfortunately, some firmware versions do not create autoconfig entries at all.

- The above leads to another problem. PCI device resources are configured in 
the E0000000 - EFFFFFFF area. Doing VA=PA mapping of this space immediately 
leads to a panic, due to running out of Zorro I/O address space. I suspect that 
I should create an extent(9) for this space and not map it directly?

- VA=PA mapping of FFFA0000-FFFE1000 space is now done with zbusmap, which is 
kinda wrong, because I can't really use information from autoconfig. 
Temporarily I had to rip zbusmap out of zbus.c and make it available to both 
zbus.c and p5pb.c. I really dislike this solution.

- The pm2fb needs prop entries like "width", "height", "depth", "is_console". 
It is possible to fake these entries in p5pb driver, or modify the pm2fb to 
support operation when these entries are absent. I didn't look at pm2fb too 
closely yet.

- I was not able to figure out how interrupt stuff works. Fortunately, it seems 
that pm2fb does not need it. Same with DMA support, hardware is probably 
capable of doing DMA. 

Any suggestions on how to deal with these problems would be appreciated. All 
information related to programming the Phase5 PCI bridge was obtained thorugh 
reverse engineering, so I may be wrong here and there.

--
Best regards,
Radosław Kujawa