A recent patch attempted to enable more efficient memory usage by using
only 2kB descriptors for jumbo frames.  The method used to implement this
has since been commented upon as "illegal" and in recent kernels even
causes a BUG when receiving ip fragments while using jumbo frames.
This patch simply goes back to the way things were.  We expect some
complaints due to order 3 allocations failing to come back due to this
change.
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>