Currently, with PAE enabled, we require the user:kernel split to occur
on a PMD boundary, so it can only be done in 1GB increments.  There
are 2 reasons for this.  First, kernel_physical_mapping_init()
assumes, when it is initializing the kernel's PMD entries, that they
start at offset 0 inside the PMD.  This is fixed by starting them at 
__pmd_offset(PAGE_OFFSET) instead.

Secondly, secondary SMP cpus require that the trampoline code be
identity mapped (map virtual addresses to the same as physical ones). 
  Right now, this is accomplished by setting the first PGD entry to
be the same as the last.  This is OK, as long as that PGD is
eventually mapping to physical 0x00000000.  My changes above break
that.  So, I allocate another PMD, and use it for the identity
mapping.  The current code is in place to allocate PTE if you're
using PAE without PSE support, but there is nothing to free them.  Any
suggestions for a clean way to do this?

Also, this gets the kernel's pagetables right, but neglects 
userspace's for now.  pgd_alloc() needs to be fixed to allocate 
another PMD, if the split isn't PMD-alighed.
-- 
Dave Hansen
haveblue@us.ibm.com