DMA transfer into sk_buff - cached memory question

[John Feller]

Hello everyone.  Great group here - it's been very helpful.

I am using MontaVista 3.1 (2.4.20-based kernel) on a 405-derivative core
(the 405 within a Xilinx Virtex-II Pro).

We've got a custom FPGA-based Ethernet MAC that stores Rx packets in
internal FPGA memory.  From there the software must copy the packet data
into SDRAM.  I've done the simple "memcpy" thing in the Rx interrupt, but
that obviously takes some time.  So the hardware group has put in a single
DMA controller that I can kick off to do the data transfer in the
background.  This works fine with my U-Boot code, but data caching is not
enabled there.

Now I am looking at implementing this DMA operation in my Linux Ethernet
driver, but am worried about data caching issues with the socket buffer data
pointer I get back from "dev_alloc_skb()".  The DMA engine will obviously
bypass the PowerPC data cache (go directly to SDRAM), so it looks like there
is a potential for mis-match between SDRAM and the cache if that SDRAM
address range is cached before the DMA transfer begins.

It doesn't appear there is a way to force the allocation of the socket
buffer data area as uncached.  Passing "__GFP_DMA" to "alloc_skb()" doesn't
seem to guarantee anything about caching - just that the memory was
allocated from a "DMA" address range.  I'm assuming this doesn't mean
anything on the PowerPC.

Ideally the allocated socket buffer data area would still be cacheable, but
guaranteed to not be in the data cache immediately after it is allocated.
Second-best would be to have the socket buffer data area marked as
non-cacheable in the 405 hardware TLB assigned to it.

Any help/advice would be appreciated.  It seems like this would be a very
common issue, but I have been unable to find the answer anywhere.  Thank you
for your help, and I apologize if I have missed something very obvious.

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/