>... the iommu init code uncaches the target area before writing
>the new pagetables into it. ...

Admittedly, this particular scenario is unlikely, but you would
get into trouble if something (e.g., the PROM) had written to a
location that became iopte space, and that particular cache line
was pushed out after the CPU wrote a mapping entry, but before (or
even during, depending on iommu TLB activity) the actual I/O.

Modern cache systems are such fun,

Chris