The old consistent memory allocator, which sat behind
dma_coherent_alloc() and pci_consistent_alloc() was completely unable
to handle allocations from interrupt context because we traditionally
used ioremap, which in turn:
  - allocates memory using GFP_KERNEL for the vm_struct
    and the page tables themselves.
  - calls get_vm_area, which uses write_lock, and therefore
    is unsafe to call from interrupt context.

In order to address this issue, a new consistent_alloc() which avoids
the above issues has been implemented.  Essentially, we set aside
a section of the kernel VM space, and pre-allocate page tables to
cover this area.  We allocate "consistent" memory within this region.

The handling of the allocation is designed to be generic; it should
be possible to replace the vmalloc(), ioremap() and module_alloc()
without too much hastle, but that would clearly be a 2.7 thing at
this stage.