This patch is from James, I've changed it slightly only.

The problem is that some IOMMU implementations have a maximum limit to the
size of the number of contiguously mappable pages (admittedly, this limit
is mostly in the resource management algorithms rather than the IOMMUs
themselves).

This patch adds this concept to the bio layer via the parameter

BIO_VMERGE_MAX_SIZE

which architectures can define in asm/io.h (if undefined, we assume it to
be infinite, which is current behaviour).

While adding this, I noticed several places where bio was making incorrect
assumptions about virtual mergeability (none of which was a bug: bio was
overestimating rather than underestimating).

- The worst offender was bio_add_page(), which seemed never to check for
  virtual mergeability

- I also fixed blk_hw_contig_segments() not to check the QUEUE_CLUSTER
  flag, and not to check the phys segment boundary.

In order to track the hw segment size across bios, I had to introduce two
extra bio parameters: bi_hw_front_size and bi_hw_back_size which store the
sizes of the front and back hw contiguous segments (and which will be equal
if there's only one hw segment).  When the bio is merged into a request,
these fields are updated with the total hw contig size so they can always
be used to assess if the merger would violate the BIO_VMERGE_MAX_SIZE
parameter.
Signed-Off-By: Jens Axboe <axboe@suse.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>