ceph mon选举流程

标签:active   option   环境   数据   form   orm   默认   paxos   his   

本章主要介绍mon选举的过程。

选举过程

整个选举过程,在Elector类中实现。此类之中实现了一个election_epoch:

当这个election_epoch为偶数的时候,表示处于稳定状态,为奇数的时候,表示还在选举过程中,mon leader 还未定。

以下是自己环境的mon情况,当前 "election_epoch": 120,说明目前处于稳定状态,没有在选举。

[root@ceph1 ~]#  ceph quorum_status | json_reformat 
{
    "election_epoch": 120,
    "quorum": [
        0,
        1,
        2
    ],
    "quorum_names": [
        "ceph2",
        "ceph1",
        "ceph3"
    ],
    "quorum_leader_name": "ceph2",
    "monmap": {
        "epoch": 1,
        "fsid": "5cd6b9b7-2ebd-47b1-a1d2-7362bb077fa3",
        "modified": "2020-08-04 19:55:21.613792",
        "created": "2020-08-04 19:55:21.613792",
        "features": {
            "persistent": [
                "kraken",
                "luminous"
            ],
            "optional": [

            ]
        },
        "mons": [
            {
                "rank": 0,
                "name": "ceph2",
                "addr": "192.168.111.14:6789/0",
                "public_addr": "192.168.111.14:6789/0"
            },
            {
                "rank": 1,
                "name": "ceph1",
                "addr": "192.168.111.20:6789/0",
                "public_addr": "192.168.111.20:6789/0"
            },
            {
                "rank": 2,
                "name": "ceph3",
                "addr": "192.168.111.34:6789/0",
                "public_addr": "192.168.111.34:6789/0"
            }
        ]
    }
}
[root@ceph1 ~]# 

选举过程关键变量

a.acked_me
/**
 * Set containing all those that acked our proposal to become the Leader.
 *
 * If we are acked by everyone in the MonMap, we will declare
 * victory.  Also note each peer‘s feature set.
 */
//发起选举本端,经过选举,其他所有mon 返回ack消息,承认选自己为leader
map<int, elector_info_t> acked_me;

b.leader_acked
/**
 * Indicates who we have acked
 */
本端已经acked,承认别人是leader
int     leader_acked;

/*
默认值-1,代表从未承认别人是leader。
如果defer 了mon.1,那么leader_acked=1.
Elector::defer(int who)函数中进行赋值
*/

c.electing_me
/**
 * Indicates if we are the ones being elected.
 *
 * We always attempt to be the one being elected if we are the ones starting
 * the election. If we are not the ones that started it, we will only attempt
 * to be elected if we think we might have a chance (i.e., the other guy‘s
 * rank is lower than ours).
 */

是否选取本端作为leader,选举发起端一发起,该值就置位true
bool     electing_me;

d.epoch
/**
 * Latest epoch we‘ve seen.
 *
 * @remarks if its value is odd, we‘re electing; if it‘s even, then we‘re
 *        stable.
 */
//选举的版本,奇数为选举状态,偶数为稳定状态
epoch_t epoch;

以下开始介绍选举过程:

选举整体流程图

ceph mon选举流程

阶段1:mon端发起选举

注意,这里leader端和peon端都可以发起选举。

void Monitor::start_election()
{
  dout(10) << "start_election" << dendl;
  wait_for_paxos_write();
  _reset();
  state = STATE_ELECTING;//设置状态

  logger->inc(l_mon_num_elections);
  logger->inc(l_mon_election_call);

  clog->info() << "mon." << name << " calling monitor election";
  elector.call_election();
}

  void call_election() {
    start();
  }

void Elector::start()
{
  dout(5) << "start -- can i be leader?" << dendl;
  acked_me.clear();
  init();

  // start by trying to elect me  /*从稳定态进入选举态,需要将版本号从偶数往上抬,抬成奇数*/
  if (epoch % 2 == 0) {
    bump_epoch(epoch+1);  // odd == election cycle
  } else {
    // do a trivial db write just to ensure it is writeable.
    auto t(std::make_shared<MonitorDBStore::Transaction>());
    t->put(Monitor::MONITOR_NAME, "election_writeable_test", rand());
    int r = mon->store->apply_transaction(t);
    assert(r >= 0);
  }
  start_stamp = ceph_clock_now();
  electing_me = true;
  acked_me[mon->rank].cluster_features = CEPH_FEATURES_ALL;
  acked_me[mon->rank].mon_features = ceph::features::mon::get_supported();
  mon->collect_metadata(&acked_me[mon->rank].metadata);
  leader_acked = -1;

  // bcast to everyone else  /*向每一个成员广播消息,提议开始重新选举*/
  for (unsigned i=0; i<mon->monmap->size(); ++i) {
    if ((int)i == mon->rank) continue;
    MMonElection *m =
      new MMonElection(MMonElection::OP_PROPOSE, epoch, mon->monmap);// OP_PROPOSE
    m->mon_features = ceph::features::mon::get_supported();
    mon->messenger->send_message(m, mon->monmap->get_inst(i)); //发送选举信号
  }
  /*
  *注意,如果所有的人都承认自己leader地位,那么可以宣布获胜。但是有些情况下,无法等到所有的回应。
  *比如某个ceph-mon进程已经不在了,是不可能得到其承认的。为了防止出现这种情况下,在通知其他节点
  *选自己的start函数设置了定时器“
  */
  reset_timer();
}  

当选举超时后,会进入expire函数。

void Elector::expire()
{
  dout(5) << "election timer expired" << dendl;

  // did i win? 从这里可以看到只要有超过一半的monitor 回复,仍然认为当前节点获胜为leader节点,否则就调用mon->bootstrap() 重新开始选举
  if (electing_me &&
      acked_me.size() > (unsigned)(mon->monmap->size() / 2)) {
    // i win
    victory();
  } else {
    // whoever i deferred to didn‘t declare victory quickly enough.
    if (mon->has_ever_joined)
      start();
    else
      mon->bootstrap();
  }
}

阶段2:选举消息调度

消息发送后,会根据不同的消息类型,经过以下流程进行处理。注意这里一次选举过程,可能要经过N次以下处理流程。

bool ms_dispatch(Message *m) ->Monitor::_ms_dispatch(Message *m)-> Monitor::dispatch_op -> Elector::dispatch(MonOpRequestRef op) ->handle_propose| handle_ack|handle_victory|handle_nak
void Elector::dispatch(MonOpRequestRef op)
{
      //选举的地址如果是非monmap中的地址,那么就退出
      if (!mon->monmap->contains(em->get_source_addr())) {
        dout(1) << "discarding election message: " << em->get_source_addr()
                << " not in my monmap " << *mon->monmap << dendl;
        return;
      }
...
            switch (em->op) {
            case MMonElection::OP_PROPOSE://发起选举提议消息
              handle_propose(op);
              return;
            }

            if (em->epoch < epoch) {
              dout(5) << "old epoch, dropping" << dendl;
              break;
            }

            switch (em->op) {
            case MMonElection::OP_ACK://返回ack消息到发起端
              handle_ack(op);
              return;
            case MMonElection::OP_VICTORY://选举成功,leader端发起消息
              handle_victory(op);//peon端进行处理
              return;
            case MMonElection::OP_NAK:
               handle_nak(op);
               return;
}

阶段3:非发起端处理

//处理发起propose的mon端发过来的消息。
void Elector::handle_propose(MonOpRequestRef op)
{
  op->mark_event("elector:handle_propose");
  MMonElection *m = static_cast<MMonElection*>(op->get_req());
  dout(5) << "handle_propose from " << m->get_source() << dendl;
  int from = m->get_source().num();

  assert(m->epoch % 2 == 1); // election
  uint64_t required_features = mon->get_required_features();
  mon_feature_t required_mon_features = mon->get_required_mon_features();

  dout(10) << __func__ << " required features " << required_features
           << " " << required_mon_features
           << ", peer features " << m->get_connection()->get_features()
           << " " << m->mon_features
           << dendl;

  if ((required_features ^ m->get_connection()->get_features()) &
      required_features) {
    dout(5) << " ignoring propose from mon" << from
        << " without required features" << dendl;
    nak_old_peer(op);
    return;
  } else if (!m->mon_features.contains_all(required_mon_features)) {
    // all the features in ‘required_mon_features‘ not in ‘m->mon_features‘
    mon_feature_t missing = required_mon_features.diff(m->mon_features);
    dout(5) << " ignoring propose from mon." << from
            << " without required mon_features " << missing
            << dendl;
    nak_old_peer(op);
  } else if (m->epoch > epoch) {
    bump_epoch(m->epoch);//正常选举流程
  } else if (m->epoch < epoch) {//说明是较老的propose
    // got an "old" propose,
    if (epoch % 2 == 0 &&    // in a non-election cycle
    mon->quorum.count(from) == 0) {  // from someone outside the quorum
      // a mon just started up, call a new election so they can rejoin!
      dout(5) << " got propose from old epoch, quorum is " << mon->quorum 
          << ", " << m->get_source() << " must have just started" << dendl;
      // we may be active; make sure we reset things in the monitor appropriately.
      mon->start_election();
    } else {
      dout(5) << " ignoring old propose" << dendl;
      return;
    }
  }

  if (mon->rank < from) {//如果从未收到过更强者(rank更小者)发来的选举请求,调用start_election,给所有成员发消息,让他们选自己为mon leader
    // i would win over them.
    if (leader_acked >= 0) {        // we already acked someone 自己曾经认过怂,消息来源的mon还不如自己,直接不理
      assert(leader_acked < from);  // and they still win, of course
      dout(5) << "no, we already acked " << leader_acked << dendl;
    } else {
      // wait, i should win!
      /*注意,electing_me记录了自己是否发出过选我为leader的请求
       *如果先后收到两个弱小者发来的选举请求,处理第一个的时候,本节点已经发出了选自己当leader的请求,
       *当第二个弱者消息到来的时候,没必要再发送选自己当leader的请求*/
      if (!electing_me) {
    mon->start_election();
      }
    }
  } else {//赞同其他人做leader
    // they would win over me
    if (leader_acked < 0 ||      // haven‘t acked anyone yet, or /*从未承认过别人,从未认过怂*/ 
    leader_acked > from ||   // they would win over who you did ack, or /*虽然承认过别人,认过怂,无奈这次来的更强大,所以还是得认怂,承认它*/
    leader_acked == from) {  // this is the guy we‘re already deferring to
      defer(from); /*defer函数的作用是认可对方可以当leader*/
    } else {//leader_acked < from
      // ignore them! /*曾经认可过更强者,不可能向不够强的mon发送认可,不理*/
      dout(5) << "no, we already acked " << leader_acked << dendl;
    }
  }
}

void Elector::bump_epoch(epoch_t e) 
{
  dout(10) << "bump_epoch " << epoch << " to " << e << dendl;
  assert(epoch <= e);
  epoch = e;//epoch 奇数
  auto t(std::make_shared<MonitorDBStore::Transaction>());
  t->put(Monitor::MONITOR_NAME, "election_epoch", epoch);
  mon->store->apply_transaction(t);

  mon->join_election();

  // clear up some state
  electing_me = false;//重置选举状态
  acked_me.clear();//重置选举状态
}

上述handle_propose函数是判断选举条件的核心。已经注释了关键的信息。

大家可能会发现这么个现象,mon发起选举,有时所有mon中都会有选举自己的信息(见下),有时候部分日志中有,部分日志中没有。这个其实和发起选举的序号大小有关。

mon.ceph2 calling monitor election

  • 本端mon序号 < mon发起端,才有可能选举自己;
  • 本端mon序号 > mon发起端,不可能选举自己;
  if (mon->rank < from) {
    // i would win over them.
    if (leader_acked >= 0) {//说明认怂过,此时也不可能选举自己
      assert(leader_acked < from);  // and they still win, of course
      dout(5) << "no, we already acked " << leader_acked << dendl;
    } else {
      // wait, i should win!
      if (!electing_me) {
    mon->start_election();//本端发起选举
      }
    }
  }

阶段3:非发起端向发起端认怂

此阶段,关键处理函数为defer(),该函数可能被调用多次。如果本端mon defer过序号更小的mon,当再来次小的mon,就不再调用defer() 了。

  if (mon->rank < from) {
  ...
  } else {//赞同其他人做leader
    // they would win over me
    if (leader_acked < 0 ||      // haven‘t acked anyone yet, or /*从未承认过别人,从未认过怂*/ 
    leader_acked > from ||   // they would win over who you did ack, or /*虽然承认过别人,认过怂,无奈这次来的更强大,所以还是得认怂,承认它*/
    leader_acked == from) {  // this is the guy we‘re already deferring to
      defer(from); /*defer函数的作用是认可对方可以当leader*/
    } else {//leader_acked < from
      // ignore them! /*曾经认可过更强者,不可能向不够强的mon发送认可,不理*/
      dout(5) << "no, we already acked " << leader_acked << dendl;
    }
void Elector::defer(int who)
{
  dout(5) << "defer to " << who << dendl;

  if (electing_me) { /*注意,认怂就要清零,即使自己曾经要求别人选过自己*/
    // drop out
    acked_me.clear();
    electing_me = false;
  }

  // ack them
  leader_acked = who;
  ack_stamp = ceph_clock_now();
  MMonElection *m = new MMonElection(MMonElection::OP_ACK, epoch, mon->monmap);// OP_ACK /*发送OP_ACK承认对方可以当leader*/
  m->mon_features = ceph::features::mon::get_supported();
  mon->collect_metadata(&m->metadata);

  // This field is unused completely in luminous, but jewel uses it to
  // determine whether we are a dumpling mon due to some crufty old
  // code.  It only needs to see this buffer non-empty, so put
  // something useless there.
  m->sharing_bl = mon->get_local_commands_bl(mon->get_required_mon_features());

  mon->messenger->send_message(m, mon->monmap->get_inst(who));//发送消息

  // set a timer
  reset_timer(1.0);  // give the leader some extra time to declare victory
}

阶段4:发起端处理非发起端的ack消息

基于阶段3,如果非发起端认怂后,会发送OP_ACK消息,那么发起端要进行对应消息的处理。主要处理函数为 Elector::handle_ack(),一般发生在leader端(临时优先级更高的mon端)。

void Elector::handle_ack(MonOpRequestRef op)
{
  op->mark_event("elector:handle_ack");
  MMonElection *m = static_cast<MMonElection*>(op->get_req());
  dout(5) << "handle_ack from " << m->get_source() << dendl;
  int from = m->get_source().num();

  assert(m->epoch % 2 == 1); // election

  assert(m->epoch == epoch);

  if (electing_me) {
    // thanks
    acked_me[from].cluster_features = m->get_connection()->get_features();
    acked_me[from].mon_features = m->mon_features;
    acked_me[from].metadata = m->metadata;
    dout(5) << " so far i have {";
    for (map<int, elector_info_t>::const_iterator p = acked_me.begin();
         p != acked_me.end();
         ++p) {
      if (p != acked_me.begin())
        *_dout << ",";
      *_dout << " mon." << p->first << ":"
             << " features " << p->second.cluster_features
             << " " << p->second.mon_features;
    }
    *_dout << " }" << dendl;

    // is that _everyone_?  /*如果所有成员都承认了自己的leader地位,那么宣布获胜,调用victory*/
    /*
    *注意,如果所有的人都承认自己leader地位,那么可以宣布获胜。但是有些情况下,无法等到所有的回应。
    *比如某个ceph-mon进程已经不在了,是不可能得到其承认的。为了防止出现这种情况下,在通知其他节点
    *选自己的start函数设置了定时器“
    */
    if (acked_me.size() == mon->monmap->size()) {//如果所有的mon端都进行了ack,说明都defer了自己,自己成为了leader
      // if yes, shortcut to election finish
      victory();
    }
  } else {
    // ignore, i‘m deferring already.
    assert(leader_acked >= 0);
  }
}

当发起端接收到所有mon的ack信息后,开始调用victory()函数,说明本端成为了leader。继续分析

void Elector::victory()
{
  leader_acked = -1;//已经选举成功了,那么把相关标记信息复位
  electing_me = false;//同上

  cancel_timer();

  assert(epoch % 2 == 1);  // election
  bump_epoch(epoch+1);     // is over! 这里epoch变为偶数,稳定状态。

  // tell everyone!
  for (set<int>::iterator p = quorum.begin();
       p != quorum.end();
       ++p) {
    if (*p == mon->rank) continue;
    MMonElection *m = new MMonElection(MMonElection::OP_VICTORY, epoch,/*注意这里会发送OP_VICTORY 消息来帮自己设置为leader节点*/
                       mon->monmap);
    m->quorum = quorum;
    m->quorum_features = cluster_features;
    m->mon_features = mon_features;
    m->sharing_bl = mon->get_local_commands_bl(mon_features);
    mon->messenger->send_message(m, mon->monmap->get_inst(*p));
  }

    // tell monitor
  mon->win_election(epoch, quorum,
                    cluster_features, mon_features, metadata);
}

void Monitor::win_election(epoch_t epoch, set<int>& active, uint64_t features,
                           const mon_feature_t& mon_features,
               const map<int,Metadata>& metadata)
{
  assert(is_electing());
  state = STATE_LEADER;// 修改选举状态
  clog->info() << "mon." << name << " is new leader, mons " << get_quorum_names()
      << " in quorum (ranks " << quorum << ")";  
  }
  paxos->leader_init();//函数里paoxs状态变为, state=STATE_RECOVERING,开始进行数据恢复操作
  monmon()->election_finished();
  _finish_svc_election();
  health_monitor->start(epoch);  

阶段5:选举失败端(peon)处理OP_VICTORY消息

//OP_VICTORY消息对应的处理函数
void Elector::handle_victory(MonOpRequestRef op)
{
  op->mark_event("elector:handle_victory");
  MMonElection *m = static_cast<MMonElection*>(op->get_req());
  dout(5) << "handle_victory from " << m->get_source()
          << " quorum_features " << m->quorum_features
          << " " << m->mon_features
          << dendl;
  int from = m->get_source().num();

  assert(from < mon->rank);// 已经选举成功了,远端的rank自然是小的
  assert(m->epoch % 2 == 0);  //victory 函数中已经对epoch+1 了

  leader_acked = -1;

  bump_epoch(m->epoch);//设置本端的epoch以及重置选举状态。此时非leader端mon也进入稳定状态

  // they win 输掉选举的话,把自己设置为peon节点
  mon->lose_election(epoch, m->quorum, from,
                     m->quorum_features, m->mon_features);

  // cancel my timer
  cancel_timer();

  // stash leader‘s commands
  assert(m->sharing_bl.length());
  vector<MonCommand> new_cmds;
  bufferlist::iterator bi = m->sharing_bl.begin();
  MonCommand::decode_vector(new_cmds, bi);
  mon->set_leader_commands(new_cmds);
}

void Monitor::lose_election(epoch_t epoch, set<int> &q, int l,
                            uint64_t features,
                            const mon_feature_t& mon_features)
{
  dout(10) << "lose_election, epoch " << epoch << " leader is mon" << leader
       << " quorum is " << quorum << " features are " << quorum_con_features
           << " mon_features are " << quorum_mon_features
           << dendl;
  state = STATE_PEON;//设置peon状态
  paxos->peon_init();
  _finish_svc_election();
  health_monitor->start(epoch);
 }

 void Paxos::peon_init()
{
  cancel_events();
  new_value.clear();

  state = STATE_RECOVERING;// peon 端的paxos也进入STATE_RECOVERING状态
}

ceph mon选举流程

标签:active   option   环境   数据   form   orm   默认   paxos   his   

原文地址:https://blog.51cto.com/wendashuai/2519254

版权声明:完美者 发表于 2020-08-12 15:55:33。
转载请注明:ceph mon选举流程 | 完美导航

暂无评论

暂无评论...