Now that brlocks loop over NR_CPUS, on SMP every br_lock/br_unlock
results in the acquire/release of 32 locks.  This incs/decs the
preempt_count by 32.

Since we only have 7 bits now for actually storing the lock depth, we
cannot nest but 3 locks deep.  I doubt we ever acquire three brlocks
concurrently, but it is still a concern.

Attached patch disables/enables preemption explicitly once and only
once for each lock/unlock.  This is also an optimization as it
removes 31 incs, decs, and conditionals. :)

Problem reported by Andrew Morton.