[ofa-general] Re: mthca use of dma_sync_single is bogus

[Roland Dreier <rdreier@cisco.com>]

 > > coherent and consistent are synonyms.  It's confusing because there is
 > > pci_alloc_consistent(), which is in general just a wrapper for
 > > dma_alloc_coherent().
 > 
 > According to DMA-mapping.txt they are not. Alpha, M68000 wihtout MMU, PPC,
 > Sparc, Sparc64, V850 have own implementation of pci_alloc_consistent().
 > 
 > Yes, on i386, the pci_alloc_consistent() is just wrapper for
 > dma_alloc_coherent().

Sorry, I was a little confusing.  The implementations may be different
but in general there is no real difference between consistent and
coherent memory.  Using either pci_alloc_consistent() or
dma_alloc_coherent() will exhaust the same small pool of address space
on powerpc 4xx for example.

 > So, what about some #ifdefs ? E.g., allow config option - Xen optimizations?

Seems pretty ugly, especially given that Xen is not upstream.  I think
the Xen tree should just carry such patches, at least until Xen is
merged.  Even then I'm quite dubious about having two code paths for this.

 - R.