ceph源码分析--Monitor tick

最新推荐文章于 2023-07-27 11:35:30 发布

原创最新推荐文章于 2023-07-27 11:35:30 发布 · 848 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#cpeh

Ceph 专栏收录该内容

14 篇文章

订阅专栏

本文详细介绍了Ceph集群中Monitor的Tick机制实现原理。包括Monitor初始化时如何启动Tick线程，Tick线程如何定期调用PaxosService的tick函数进行状态检查与更新，并涉及会话管理和健康检查日志的更新等内容。

刚入职的时候曾经定位过一个ceph集群时钟回调8小时的单子。后来投入到了项目中。
最近同事问我osd向monitor定时report的处理过程，也涉及到了这一块的内容，而且
正式入职答辩的时候，主管问，那这个tick线程是定时调用的吗？当时被问懵了，因为自己并没有深入的前后看到这块。事情总要有头有尾，遗留的问题总要解决，写这篇博客的目的就是回答当时的遗留问题。究竟monitor的这个tick是怎么跑的？

Monitor的tick线程启动
同时调用PaxosService的tick函数

int Monitor::init()

int Monitor::init()
{
  dout(2) << "init" << dendl;
  Mutex::Locker l(lock);

  finisher.start();

  // start ticker
  timer.init();
  new_tick();
  ···
}

void Monitor::new_tick()

void Monitor::new_tick()
{
  timer.add_event_after(g_conf->mon_tick_interval, new C_MonContext(this, [this](int) {
    tick();
      }));
}

时间间隔为5s

OPTION(mon_tick_interval, OPT_INT)

void Monitor::tick()

void Monitor::tick()
{
  // ok go.
  dout(11) << "tick" << dendl;
  const utime_t now = ceph_clock_now();

  // Check if we need to emit any delayed health check updated messages
  if (is_leader()) {
    const auto min_period = g_conf->get_val<int64_t>(
                              "mon_health_log_update_period");
    for (auto& svc : paxos_service) {
      auto health = svc->get_health_checks();

      for (const auto &i : health.checks) {
        const std::string &code = i.first;
        const std::string &summary = i.second.summary;
        const health_status_t severity = i.second.severity;

        auto status_iter = health_check_log_times.find(code);
        if (status_iter == health_check_log_times.end()) {
          continue;
        }

        auto &log_status = status_iter->second;
        bool const changed = log_status.last_message != summary
                             || log_status.severity != severity;

        if (changed && now - log_status.updated_at > min_period) {
          log_status.last_message = summary;
          log_status.updated_at = now;
          log_status.severity = severity;

          ostringstream ss;
          ss << "Health check update: " << summary << " (" << code << ")";
          clog->health(severity) << ss.str();
        }
      }
    }
  }


  for (vector<PaxosService*>::iterator p = paxos_service.begin(); p != paxos_service.end(); ++p) {
    (*p)->tick();
    (*p)->maybe_trim();
  }

  // trim sessions
  {
    Mutex::Locker l(session_map_lock);
    auto p = session_map.sessions.begin();

    bool out_for_too_long = (!exited_quorum.is_zero() &&
                 now > (exited_quorum + 2*g_conf->mon_lease));

    while (!p.end()) {
      MonSession *s = *p;
      ++p;

      // don't trim monitors
      if (s->inst.name.is_mon())
    continue;

      if (s->session_timeout < now && s->con) {
    // check keepalive, too
    s->session_timeout = s->con->get_last_keepalive();
    s->session_timeout += g_conf->mon_session_timeout;
      }
      if (s->session_timeout < now) {
    dout(10) << " trimming session " << s->con << " " << s->inst
         << " (timeout " << s->session_timeout
         << " < now " << now << ")" << dendl;
      } else if (out_for_too_long) {
    // boot the client Session because we've taken too long getting back in
    dout(10) << " trimming session " << s->con << " " << s->inst
         << " because we've been out of quorum too long" << dendl;
      } else {
    continue;
      }

      s->con->mark_down();
      remove_session(s);
      logger->inc(l_mon_session_trim);
    }
  }
  sync_trim_providers();

  if (!maybe_wait_for_quorum.empty()) {
    finish_contexts(g_ceph_context, maybe_wait_for_quorum);
  }

  if (is_leader() && paxos->is_active() && fingerprint.is_zero()) {
    // this is only necessary on upgraded clusters.
    MonitorDBStore::TransactionRef t = paxos->get_pending_transaction();
    prepare_new_fingerprint(t);
    paxos->trigger_propose();
  }
  //new_tick()
  new_tick();
}