Atomics are not needed for accesses synchronized by a mutex.
There is one unlocked read of the pointer for which the atomic
load doesn't do anything (in terms of limiting races).
Atomics remain for accesses under the NET_EPOCH to provide
READ_ONCE semantics.