David Mosberger wrote:

>  consistency = coherency + ordering

That's one of the things I meant when I said the docs weren't
quite there yet ... the dma_*() API docs don't address how
to achieve the ordering, or even mention that it's an issue.

Plus, there are definitions of coherency that include event
ordering; it's not clear which definition is intended.


Anyway, see the attached patch.  It goes on top of 2.6.4-bk2
and passes my unlink tests (the ones that haven't been run
much on recent kernels!) on several different OHCI hosts.

Other issues may be lurking, but this seems to fix a handful
of nasties that I could reproduce.

- Dave