When we can't allocate from the preferred allocation type we need fallback to
other allocation types.  This patch determines which allocation types we try
to fallback to in which order.  It also adds a special fallback type that is
designed to minimize the fragmentation caused by fallback between the other
types.

There is an implicit tradeoff being made here between avoiding fragmentation
and satisfying allocations.  This patch aims to keep existing behavior of
satisfying allocations if there is any free memory of any type to satisfy them.
It does a reasonable job of trying to minimize the fragmentation, and certainly
does better than a stock kernel in all situations.

However, it would not be hard to imagine scenarios where a different fallback
algorithm that fails more allocations was able to keep fragmentation down much
better, and on some systems this decreased fragmentation might even be worth
the cost of failing allocations.  Systems doing memory hotplug remove for
example.  This patch is designed so that the static function
fallback_alloc() can be easily replaced with an alternate implementation (under
a config option perhaps) in the future.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
Signed-off-by: Joel Schopp <jschopp@austin.ibm.com>