3.6.0 - 2021-02-24
    * UDP DNS threads now idle much more efficiently when there's little to no
      network traffic arriving for them.  Previously all such threads woke up
      and re-executed recvmsg() at least once every ~3.1 seconds.  Now they
      block indefinitely in a ppoll() call once they've been idle for ~257ms or
      longer.  This reduces thread wakeup churn and thus power/cpu efficiency
      waste in the case of unused threads for extra listen addresses.  It also
      reduces the worst corner case for daemon shutdown delays caused by UDP
      threads from ~3.1s to ~257ms.
    * TCP DNS threads now invoke rcu_quiescent_state() more often, which
      can help data reloads complete faster when there are many active TCP
      clients per thread.  Previously they quiesced or went offline from RCU
      once per eventloop iteration, where one iteration could handle up to 1
      pending request per client connection for many connections.  Now they
      also quiesce once for each response sent.
    * SO_REUSEPORT_LB is now used for better UDP and TCP socket loadbalancing
      on BSDs which have this socket option available.
    * Zone and config files are now loaded more efficiently when there are many
      small ones, by switching to a malloc()+read() pattern instead of mmap()
      for files that are smaller than 1MB.  mmap modifies process-global memory
      mappings once per file, which is expensive for a multi-threaded daemon
      loading many small files (especially in parallel!), so it's just not
      worth it unless the files are large-ish.