2.6.19-rc1: vm stuff

The first 3 patches are some minor fixes and rearrangements for the
page allocator and are probably fit to go into -mm.

The next set of 3 patches is another attempt at solving the invalidate
vs pagefault race (this got reintroduced after invalidate_complete_page2
was added, and has always been present for nonlinear mappings). These
are booted and have had some stress testing, but are still at the RFC
stage. Comments?

Nick