This function is called a lot.  Every brk().  The atomic_add() against a
global counter hurts on large SMP machines.

The patch simply reduces the rate at which that atomic operation is
performed, by accumulating a per-cpu count which is spilled into the global
counter when the local counter overflows.

It trades off efficiency for a little inaccuracy.

I tried various implementations involving kmalloc_percpu() and open-coded
per-cpu arrays in a generic "per-cpu counter" thing.  They all were
surprisingly sucky - the additional cache misses involved in walking the more
complex data structures really showed up.