Attached is a patch that should apply against Bjorn's 2.4 bk
tree.  Changes:

 - Fixes a potential silent data corruption because the pdir
   cache on the chipset set was getting improperly flushed.
   It's very hard to hit given the allocation mechanism for the
   pdir, but it is possible.

 - Demangles scatter-gather pointers.  The driver was taking far
   too much liberty in changing these.  A known MCA path caused
   by this is SCSI retries, which passed back in a pre-mangled
   scatter-gather array.

 - Adds an option for keeping the entire iommu pdir valid.  This
   is really just some debugging code that I added along the way.
   In the event that you've installed a device that aggressively
   tries to prefetch, you may get an MCA if it prefetches beyond
   it's pdir entry.  New boxes shouldn't see much of this because
   PCI will disconnect.  Older systems w/ < rev 3.0 LBAs might
   see issues.  Turning this on, makes all unused pdir entries
   point to a spill page that contains poisoned data. (off by
   default)

 - Removes platform_pci_dma_address().  With the extra scatterlist
   entries, it seems reasonable for DMA engines to store the
   address in dma_address.  pci_dma_length is now just a macro for
   sg->dma_length, this feels cleaner and reduces complexity for
   DMA engines that try to coalesce.  I believe it's even a
   benefit for swiotlb.  The sn pci_dma interface likely needs
   some touchups by those that know the interface in this area.
   (thanks to Grant Grundler for helping out here)

   One issue I've seen, that's simply because the scatterlist is
bigger now, 64k pages and IDE don't get along.  ide-dma tries
to kmalloc a _very_ large scatterlist.  The algorithm factors in
PAGE_SIZE, but it only seems to produce a reasonable value if you
have 4k pages.  It goes ballistic at 64k.  Replacing the kmalloc/
kfree w/ a vmalloc/vfree solves the problem, but seems like it's
only a bandaid.

        Alex

--
Alex Williamson                                  Linux Development Lab
alex_williamson@hp.com                                 Hewlett Packard
970-898-9173                                          Fort Collins, CO