On Mon, 2008-01-28 at 10:23 -0600, Olof Johansson wrote:

Ok, makes sense.

I was going to protest the hack for >32GB configs, with the motivation
that just using the htab-backed window is way too small for such a
config. However, with 32GB memory and 4K pages, that window is 512MB, so
we should be fine.

Might be a problem with 64K pages tho... Or do we use the same
calculation ?

In addition, on those blades, really the only device that is limited to
32 bits (and thus is forced to use the iommu remapped region) is USB.

Having that described in the patch (or at least in the patch description)
to make it more clear could be good. That, and the fact that the mapping
is offset on <32GB memory machines, and thus not really a 1:1 mapping.

Should be called a "linear" mapping.

Does the cell I/O bridge reflect out accesses to 2-4GB on the bus
again? If not, that could be another place to stick the dynamic range
for large config machines.

On the PCI bus itself, 2-4GB is where MMIO sits.

Ben.