On Wed, Apr 30, 2008 at 05:34:56PM +0200, Joakim Tjernlund wrote:
On Wed, 2008-04-30 at 10:21 -0500, Scott Wood wrote:
On Wed, Apr 30, 2008 at 05:11:09PM +0200, Joakim Tjernlund wrote:
Yet the in_bex()/out_bex() functions in PowerPC linux uses sync and all SOC drivers are encouraged to use them. What a waste :(
sync is needed in some of the cases, to sync I/O accesses with DMA buffer accesses. Ideally, we could trust the driver writers to put synchronization in where needed, but it seems Linux has too much x86 heritage for that.
Perhaps, is sync needed in this case for non-smp too? or is eieio enough?
Yes, sync is needed -- eieio doesn't order between stores to cacheable memory and stores to cache-inhibited memory.
Anyway, just have a look at ucc_geth
Do I have to? :-)
and you will see that most such accesses are just about getting the endian right.
If you mean the descriptor accesses, ordering is relevant there as well.
There should at least be raw alternatives, though...
There need be a get-the-endian-right-but-no-sync.
Agreed. There's cpu_to_be32, etc., but that doesn't fit well with load/store endian-swapping instructions.
After all 2.4 managed well without using the in/out be() functions.
I see in/out_be32() in 2.4... and the significant chunks of code that use volatile pointers instead, I wouldn't consider to be managing "well"; they just happen to work most of the time.
-Scott