There are a lot of places in Lustre where we unnecessarily use more than one atomic operation where one will do. The compiler can't fix this for us, since usage of atomics is largely exempt from optimization. In a few places, this has a clearly measurable cost (lu_object_get can use atomic_inc_return for the ASSERT, and this nets a few % improvement on single process reads). Some of the changes in the osc page code may also give some measurable benefits.

This cleanup is not intended to catch every bad use of atomics (I looked primarily at asserts), but should be a good start.