The current version of Userspace-RCU creates just a single call_rcu
thread to do all the cleaning. The only advantage of that is the
serialization of the call_rcu() calls, but on machine with many CPUs, it
is going to slow down the memory reclamation and not fully utilize all
the CPU cores.
Create per-loop call_rcu_data structure and assign each to the
respective isc_loop. We ignore isc_work threads for now - they will
still use the default call_rcu thread.
Ondřej Surý
2deb9296f5