Here is the second version of the patch. I have used alloc_pages_exact
instead of the complex double array approach.

I still fallback to kmalloc/vmalloc because hotplug can happen quite
some time after boot and we can end up not having enough continuous
pages at that time. 

I am also thinking whether it would make sense to introduce
alloc_pages_exact_node function which would allocate pages from the
given node.

Any thoughts?
---