Finn Thain wrote:
Here is the latest sonic driver update, which I've been rattling on about
for some time now. There are several reasons for this patch,
- Adopt the DMA API (jazzsonic, macsonic & core driver)
- Adopt the driver model (macsonic)
This part was cribbed from jazzsonic. As a consequence, macsonic once
again works as a module. Driver model is also used by the DMA calls.
- Support 16 bit cards (macsonic & core driver, also affects jazzsonic)
This code was adapted from the mac68k linux 2.2 kernel, where it has
languished for a long time.
- Support more mac cards (macsonic)
Also from mac68k repo.
- Zero-copy buffer handling (core driver)
Provides a nice performance improvement. The new algorithm incidentally
helped to replace the old Jazz DMA code.
A nice consequence of this patch is that one can now diff jazzsonic.c and
macsonic.c and get a comprehensible result. Maybe that will make them
easier to maintain.
The patch has been tested on a variety of macs (several 32-bit quadra
built-in NICs, a 16-bit LC PDS NIC and a 16-bit comm-slot NIC), but only
compile tested for MIPS Jazz.
The DMA API has been implemented for m68k by Roman Zippel, and his patch
is now required for macsonic to function. There is also a patch available
which implements that API for Jazz machines (by Thomas I believe).
It would be nice if someone could test this on a little endian Jazz
machine.
Comments and criticism welcomed.
-f
Signed-off-by: Finn Thain <fthain@xxxxxxxxxxxxxxxxxxx>
Patch looks OK to me. Comments:
1) Either Geert or Ralf can merge this, with my ACK.
2) Would be nice to get it tested on the machines you list as untested.
3) [possible problem in driver, not your changes] I wonder if
IRQ_HANDLED is ever returned for shared interrupts? I don't know enough
about the platform interrupt architecture to answer this question.
4) Remove casts to/from void. This is especially noticable in all the
casts of the netdev_priv() return value.
5) If it doesn't cause too much patch noise, consider using enums rather
than #defines, for numeric constants. This gives the compiler more type
information and makes the symbols visible in a debugger. This is a
-maintainer preference- issue overall, so don't sweat it if you disagree.
