When using call_rcu(), the called function may be delayed quite
significantly, and without a matching rcu_barrier() there's no
way to be sure it has finished.
Therefore, global state that could be gone/freed/reused should
never be touched in the callback.
Fix this in mesh by moving the atomic_dec() into the caller;
that's not really a problem since we already unlinked the path
and it will be destroyed anyway.
This fixes a crash Jouni observed when running certain tests in
a certain order, in which the mesh interface was torn down, the
memory reused for a function pointer (work struct) and running
that then crashed since the pointer had been decremented by 1,
resulting in an invalid instruction byte stream.