When the VM is under heavy load and a page allocator caller is holding
the BKL there can be several-millisecond scheduling stalls.

So add a cond_resched() at every 32nd page (approx).