Ceph 学习——OSD读写流程与源码分析（一）

消息从客户端发送而来，之前几节介绍了 客户端下 对象存储、块存储库的实现以及他们在客户端下API请求的发送过程（Ceph学习——Librados与Osdc实现源码解析 、 Ceph学习——客户端读写操作分析 、 Ceph学习——Librbd块存储库与RBD读写流程源码分析）。当请求被封装后，通过消息发送模块（Ceph学习——Ceph网络通信机制与源码分析）将请求及其相关信息发送到服务端实现真正的对数据的操作。服务端的操作模块便是由OSD、OS模块完成的，这节先介绍OSD模块。

• OSD 模块主要的类
  • OSD类
  • PrimaryLogPG类
  • PGBackend类
• OSD读写函数调用流程

直接上图：

同样当前最新的版本，和之前的版本有所不同，有一些模块简化了，类的名字也改了。先介绍图中涉及的相关的类，然后在对类中具体函数主要调用流程进行分析。

OSD 模块主要的类

盗图：其中ReplicatedPG 在最新的版本中去掉了，更改为PrimaryLogPG类

OSD类

OSD和OSDService是核心类，他们直接在顶层负责一个OSD节点的工作，从客户端的得到的消息，就是先到达OSD类中，通过OSD类的处理，在调用PrimaryLogPG（之前为ReplicatedPG 类）类进行处理。该类中，在读写流程中的主要工作是消息（Message）封装为 RequestOp，检查epoch （版本）是否需要更新，并获取PG句柄，并做PG相关的检查，最后将请求加入队列。

PrimaryLogPG类

该类继承自PG类，PGBackend::Listener（该类是一个抽象类）类PG类处理相关状态的维护，以及实现PG层面的功能，核心功能是用boost库的statechart状态机来做PG状态转换。它实现了PG内的数据读写等功能。

PGBackend类

该类主要功能是将请求数据通过事务的形式同步到一个PG的其它从OSD上（注意：主OSD的操作PrimaryLogPG来完成）。
他有两个子类，分别是 ReplicatedBackend和ECBackend，对应着PG的的两种类型的实现。

OSD读写函数调用流程

1）OSD::ms_fast_dispatch 函数是接收消息Message的入口函数，他被网络模块的接收线程调用。主要工作是 检查service服务 、把Message封装为OpRequest类型、获取session、获取最新的OSdMap，最后dispatch_session_waiting，进入下一步。

void OSD::ms_fast_dispatch(Message *m)
{
  FUNCTRACE();
  if (service.is_stopping()) {//检查service，如果停止了直接返回
    m->put();
    return;
  }
  OpRequestRef op = op_tracker.create_request<OpRequest, Message*>(m);//把Message封装为OpRequest类型
...
...

  if (m->get_connection()->has_features(CEPH_FEATUREMASK_RESEND_ON_SPLIT) ||
      m->get_type() != CEPH_MSG_OSD_OP) {
    // queue it directly直接调用enqueue_op处理
    enqueue_op(
      static_cast<MOSDFastDispatchOp*>(m)->get_spg(),
      op,
      static_cast<MOSDFastDispatchOp*>(m)->get_map_epoch());
  } else {
    Session *session = static_cast<Session*>(m->get_connection()->get_priv());//获取 session 其中包含了一个Connection的相关信息
    if (session) {
      {
    Mutex::Locker l(session->session_dispatch_lock);
    op->get();
    session->waiting_on_map.push_back(*op);//将请求加如waiting_on_map的列表里
    OSDMapRef nextmap = service.get_nextmap_reserved();//获取最新的OSDMAP
    dispatch_session_waiting(session, nextmap);//该函数中 循环处理请求
    service.release_map(nextmap);
      }
      session->put();
    }
  }
  OID_EVENT_TRACE_WITH_MSG(m, "MS_FAST_DISPATCH_END", false); 
}

2）OSD::dispatch_session_waiting 主要工作是循环处理队列waiting_on_map中的元素，对比OSDmap，以及获取他们的pgid，最后调用enqueue_op处理。

void OSD::dispatch_session_waiting(Session *session, OSDMapRef osdmap)
{
  assert(session->session_dispatch_lock.is_locked());

  auto i = session->waiting_on_map.begin();
  while (i != session->waiting_on_map.end()) {//循环处理waiting_on_map中的元素
    OpRequestRef op = &(*i);
    assert(ms_can_fast_dispatch(op->get_req()));
    const MOSDFastDispatchOp *m = static_cast<const MOSDFastDispatchOp*>(
      op->get_req());
    if (m->get_min_epoch() > osdmap->get_epoch()) {//osdmap版本不对应
      break;
    }
    session->waiting_on_map.erase(i++);
    op->put();

    spg_t pgid;
    if (m->get_type() == CEPH_MSG_OSD_OP) {
      pg_t actual_pgid = osdmap->raw_pg_to_pg(
    static_cast<const MOSDOp*>(m)->get_pg());
      //osdmap->get_primary_shard(actual_pgid, &pgid)获取 pgid 该PG的主OSD
      if (!osdmap->get_primary_shard(actual_pgid, &pgid)) {
    continue;
      }
    } else {
      pgid = m->get_spg();
    }
    enqueue_op(pgid, op, m->get_map_epoch());//获取成功则调用enqueue_op处理
  }

  if (session->waiting_on_map.empty()) {
    clear_session_waiting_on_map(session);
  } else {
    register_session_waiting_on_map(session);
  }
}

3）OSD::enqueue_op 的主要工作是将求情加入到op_shardedwq队列中

void OSD::enqueue_op(spg_t pg, OpRequestRef& op, epoch_t epoch)
{
...
  op->osd_trace.event("enqueue op");
  op->osd_trace.keyval("priority", op->get_req()->get_priority());
  op->osd_trace.keyval("cost", op->get_req()->get_cost());
  op->mark_queued_for_pg();
  logger->tinc(l_osd_op_before_queue_op_lat, latency);
  //加入op_shardedwq队列中
  op_shardedwq.queue(
    OpQueueItem(
      unique_ptr<OpQueueItem::OpQueueable>(new PGOpItem(pg, op)),
      op->get_req()->get_cost(),
      op->get_req()->get_priority(),
      op->get_req()->get_recv_stamp(),
      op->get_req()->get_source().num(),
      epoch));
}

4）OSD::dequeue_op 调用函数进行osdmap的更新，调用do_request进入PG处理流程

void OSD::dequeue_op(
  PGRef pg, OpRequestRef op,
  ThreadPool::TPHandle &handle)
{
...
...
  logger->tinc(l_osd_op_before_dequeue_op_lat, latency);

  Session *session = static_cast<Session *>(
    op->get_req()->get_connection()->get_priv());
  if (session) {
      //调用该函数进行 osdmap的更新
    maybe_share_map(session, op, pg->get_osdmap());
    session->put();
  }
  //正在是删除、直接返回
  if (pg->is_deleting())
    return;

  op->mark_reached_pg();
  op->osd_trace.event("dequeue_op");
  //调用pg的do_request处理
  pg->do_request(op, handle);

  // finish
  dout(10) << "dequeue_op " << op << " finish" << dendl;
  OID_EVENT_TRACE_WITH_MSG(op->get_req(), "DEQUEUE_OP_END", false);
}

5）PrimaryLogPG::do_request该函数 主要你检查PG的状态，以及根据消息类型进行不同处理

void PrimaryLogPG::do_request(
  OpRequestRef& op,
  ThreadPool::TPHandle &handle)
{
...
  // make sure we have a new enough map
  //检查 osdmap
  auto p = waiting_for_map.find(op->get_source());
...

  //是否可以丢弃
  if (can_discard_request(op)) {
    return;
  }
...
...
  //PG还没有peered
  if (!is_peered()) {
    // Delay unless PGBackend says it's ok
      //检查pgbackend是否可以处理这个请求
    if (pgbackend->can_handle_while_inactive(op)) {
      bool handled = pgbackend->handle_message(op);//可以处理，则调用该函数处理
      assert(handled);
      return;
    } else {
      waiting_for_peered.push_back(op);//不可以则加入waiting_for_peered队列
      op->mark_delayed("waiting for peered");
      return;
    }
  }
  ...
  ...
  //PG处于Peered 并且flushes_in_progress为0的状态下
  assert(is_peered() && flushes_in_progress == 0);
  if (pgbackend->handle_message(op))
    return;

// 根据不同的消息请求类型，进行相应的处理
  switch (op->get_req()->get_type()) {
  case CEPH_MSG_OSD_OP:
  case CEPH_MSG_OSD_BACKOFF:
    if (!is_active()) {//该PG状态 为非active状态
      dout(20) << " peered, not active, waiting for active on " << op << dendl;
      waiting_for_active.push_back(op);//加入队列
      op->mark_delayed("waiting for active");
      return;
    }
    switch (op->get_req()->get_type()) {
    case CEPH_MSG_OSD_OP:
      // verify client features 如果是cache pool ，操作没有带CEPH_FEATURE_OSD_CACHEPOOL的feature标志，返回错误信息
      if ((pool.info.has_tiers() || pool.info.is_tier()) &&
      !op->has_feature(CEPH_FEATURE_OSD_CACHEPOOL)) {
    osd->reply_op_error(op, -EOPNOTSUPP);
    return;
      }
      do_op(op);//调用do_op 处理
      break;
    case CEPH_MSG_OSD_BACKOFF:
      // object-level backoff acks handled in osdop context
      handle_backoff(op);
      break;
    }
    break;

...
//各种消息类型
...

  default:
    assert(0 == "bad message type in do_request");
  }
}

6）PrimaryLogPG::do_op 函数很长很负责，这里着看相关调用流程好了，主要功能是解析出操作来，然后对操作的个中参数进行检查，检查相关对象的状态，以及该对象的head、snap、clone对象的状态等，并调用函数获取对象的上下文、操作的上下文（ObjectContext、OPContext）

void PrimaryLogPG::do_op(OpRequestRef& op)
{
  FUNCTRACE();
  // NOTE: take a non-const pointer here; we must be careful not to
  // change anything that will break other reads on m (operator<<).
  MOSDOp *m = static_cast<MOSDOp*>(op->get_nonconst_req());
  assert(m->get_type() == CEPH_MSG_OSD_OP);
  //解析字段，从bufferlist解析数据
  if (m->finish_decode()) {
    op->reset_desc();   // for TrackedOp
    m->clear_payload();
  }
...
...
  if ((m->get_flags() & (CEPH_OSD_FLAG_BALANCE_READS |
             CEPH_OSD_FLAG_LOCALIZE_READS)) &&
      op->may_read() &&
      !(op->may_write() || op->may_cache())) {
    // balanced reads; any replica will do 平衡读，则主从OSD都可以读取
    if (!(is_primary() || is_replica())) {
      osd->handle_misdirected_op(this, op);
      return;
    }
  } else {
    // normal case; must be primary 否则只能读取主OSD
    if (!is_primary()) {
      osd->handle_misdirected_op(this, op);
      return;
    }
  }

  if (!op_has_sufficient_caps(op)) {
    osd->reply_op_error(op, -EPERM);
    return;
  }
  //op中包含includes_pg_op该操作，则调用 do_pg_op(op)处理
  if (op->includes_pg_op()) {
    return do_pg_op(op);
  }

  // object name too long?
  //检查名字是否太长
  if (m->get_oid().name.size() > cct->_conf->osd_max_object_name_len) {
    dout(4) << "do_op name is longer than "
        << cct->_conf->osd_max_object_name_len
        << " bytes" << dendl;
    osd->reply_op_error(op, -ENAMETOOLONG);
    return;
  }
...
...
  // blacklisted?
  //发送请求的客户端是黑名单中的一个
  if (get_osdmap()->is_blacklisted(m->get_source_addr())) {
    dout(10) << "do_op " << m->get_source_addr() << " is blacklisted" << dendl;
    osd->reply_op_error(op, -EBLACKLISTED);
    return;
  }
...
...
  // missing object?
  //head对象是否处于缺失状态
  if (is_unreadable_object(head)) {
    if (!is_primary()) {
      osd->reply_op_error(op, -EAGAIN);
      return;
    }
    if (can_backoff &&
    (g_conf->osd_backoff_on_degraded ||
     (g_conf->osd_backoff_on_unfound && missing_loc.is_unfound(head)))) {
      add_backoff(session, head, head);
      maybe_kick_recovery(head);
    } else {
      wait_for_unreadable_object(head, op);//加入队列，等待恢复完成
    }
    return;
  }

  // degraded object?
  //顺序写 且head对象正在恢复状态
  if (write_ordered && is_degraded_or_backfilling_object(head)) {
    if (can_backoff && g_conf->osd_backoff_on_degraded) {
      add_backoff(session, head, head);
      maybe_kick_recovery(head);
    } else {
      wait_for_degraded_object(head, op);//加入队列，等待
    }
    return;
  }
  //顺序写，切处于数据一致性检查 scrub时期
  if (write_ordered && scrubber.is_chunky_scrub_active() &&
      scrubber.write_blocked_by_scrub(head)) {
    dout(20) << __func__ << ": waiting for scrub" << dendl;
    waiting_for_scrub.push_back(op);
    op->mark_delayed("waiting for scrub");
    return;
  }
 ...
 ...
  //若果是顺序写，并且该对象在该队列中
  if (write_ordered && objects_blocked_on_cache_full.count(head)) {
    block_write_on_full_cache(head, op);
    return;
  }
  ...
  ...
  // io blocked on obc?
  //检查对象是否被blocked
  if (!m->has_flag(CEPH_OSD_FLAG_FLUSH) &&
      maybe_await_blocked_head(oid, op)) {
    return;
  }

  //调用find_object_context获取object_context
  int r = find_object_context(
    oid, &obc, can_create,
    m->has_flag(CEPH_OSD_FLAG_MAP_SNAP_CLONE),
    &missing_oid);


// hit.set 不为空 则设置
  bool in_hit_set = false;
  if (hit_set) {
    if (obc.get()) {
      if (obc->obs.oi.soid != hobject_t() && hit_set->contains(obc->obs.oi.soid))
    in_hit_set = true;
    } else {
      if (missing_oid != hobject_t() && hit_set->contains(missing_oid))
        in_hit_set = true;
    }
    if (!op->hitset_inserted) {
      hit_set->insert(oid);
      op->hitset_inserted = true;
      if (hit_set->is_full() ||
          hit_set_start_stamp + pool.info.hit_set_period <= m->get_recv_stamp()) {
        hit_set_persist();
      }
    }
  }
  //agent_state 不为空
  if (agent_state) {
    if (agent_choose_mode(false, op))// 调用该函数进行选择agent的状态
      return;
  }
...
...
...
  op->mark_started();

  execute_ctx(ctx);//调用该函数，执行相关操作
  utime_t prepare_latency = ceph_clock_now();
  prepare_latency -= op->get_dequeued_time();
  osd->logger->tinc(l_osd_op_prepare_lat, prepare_latency);
  if (op->may_read() && op->may_write()) {
    osd->logger->tinc(l_osd_op_rw_prepare_lat, prepare_latency);
  } else if (op->may_read()) {
    osd->logger->tinc(l_osd_op_r_prepare_lat, prepare_latency);
  } else if (op->may_write() || op->may_cache()) {
    osd->logger->tinc(l_osd_op_w_prepare_lat, prepare_latency);
  }

  // force recovery of the oldest missing object if too many logs
  maybe_force_recovery();
}

7） PrimaryLogPG::find_object_context 函数主要根据 不同发情况 通过调用 PrimaryLogPG::get_object_context函数获取 对象上下文。

/* * If we return an error, and set *pmissing, then promoting that * object may help. * * If we return -EAGAIN, we will always set *pmissing to the missing * object to wait for. * * If we return an error but do not set *pmissing, then we know the * object does not exist. */
//获取一个对象的ObjectContext
int PrimaryLogPG::find_object_context(const hobject_t& oid,
                      ObjectContextRef *pobc,
                      bool can_create,
                      bool map_snapid_to_clone,
                      hobject_t *pmissing)
{
  FUNCTRACE();
  assert(oid.pool == static_cast<int64_t>(info.pgid.pool()));
  // want the head?
  if (oid.snap == CEPH_NOSNAP) {
    ObjectContextRef obc = get_object_context(oid, can_create);//如果是想要原始对象（head）直接调用
    if (!obc) {
      if (pmissing)
        *pmissing = oid;
      return -ENOENT;
    }
    dout(10) << "find_object_context " << oid
       << " @" << oid.snap
       << " oi=" << obc->obs.oi
       << dendl;
    *pobc = obc;

    return 0;
  }

  hobject_t head = oid.get_head();

  // we want a snap
  //不是map_snapid_to_clone对象且，该snap快照已经被删除，直接返回-ENOENT
  if (!map_snapid_to_clone && pool.info.is_removed_snap(oid.snap)) {
    dout(10) << __func__ << " snap " << oid.snap << " is removed" << dendl;
    return -ENOENT;
  }

  SnapSetContext *ssc = get_snapset_context(oid, can_create);//调用get_snapset_context对象来获取SnapSetContext对象。
  if (!ssc || !(ssc->exists || can_create)) {
    dout(20) << __func__ << " " << oid << " no snapset" << dendl;
    if (pmissing)
      *pmissing = head;  // start by getting the head
    if (ssc)
      put_snapset_context(ssc);
    return -ENOENT;
  }
//如果是map_snapid_to_clone
  if (map_snapid_to_clone) {
    dout(10) << "find_object_context " << oid << " @" << oid.snap
         << " snapset " << ssc->snapset
         << " map_snapid_to_clone=true" << dendl;
    if (oid.snap > ssc->snapset.seq) {//大于说明 该快照最新，osd还没完成相关信息的更新，直接返回head对象的上下文
      // already must be readable
      ObjectContextRef obc = get_object_context(head, false);//直接返回head对象的上下文
      dout(10) << "find_object_context " << oid << " @" << oid.snap
           << " snapset " << ssc->snapset
           << " maps to head" << dendl;
      *pobc = obc;
      put_snapset_context(ssc);
      return (obc && obc->obs.exists) ? 0 : -ENOENT;
    } else {
      vector<snapid_t>::const_iterator citer = std::find(//否则检查snapset的克隆列表
    ssc->snapset.clones.begin(),
    ssc->snapset.clones.end(),
    oid.snap);
      if (citer == ssc->snapset.clones.end()) {
    dout(10) << "find_object_context " << oid << " @" << oid.snap
         << " snapset " << ssc->snapset
         << " maps to nothing" << dendl;
    put_snapset_context(ssc);
    return -ENOENT;
      }
      ...
      ...
      //找到，但处于缺失状态
      if (pg_log.get_missing().is_missing(oid)) {
    dout(10) << "find_object_context " << oid << " @" << oid.snap
         << " snapset " << ssc->snapset
         << " " << oid << " is missing" << dendl;
    if (pmissing)
      *pmissing = oid;
    put_snapset_context(ssc);
    return -EAGAIN;
      }
    ...
    ...//各种情况下的find_object_context
}

8）get_object_context 实际去获取上下文，先在缓存里面找，如果没有在调用函数去获取。另外在调用get_snapset_context获取SnapSetContext。

ObjectContextRef PrimaryLogPG::get_object_context(
  const hobject_t& soid,
  bool can_create,
  const map<string, bufferlist> *attrs)
{
...
//先在缓存里面找
  ObjectContextRef obc = object_contexts.lookup(soid);
  osd->logger->inc(l_osd_object_ctx_cache_total);
  if (obc) {
    osd->logger->inc(l_osd_object_ctx_cache_hit);
    dout(10) << __func__ << ": found obc in cache: " << obc
         << dendl;
  } else {
    dout(10) << __func__ << ": obc NOT found in cache: " << soid << dendl;
    // check disk
    bufferlist bv;
    if (attrs) {
      assert(attrs->count(OI_ATTR));
      bv = attrs->find(OI_ATTR)->second;
    } else {
      int r = pgbackend->objects_get_attr(soid, OI_ATTR, &bv);//缓存没有就调用函数去获取
      if (r < 0) {
    if (!can_create) {
      dout(10) << __func__ << ": no obc for soid "
           << soid << " and !can_create"
           << dendl;
      return ObjectContextRef();   // -ENOENT!
    }

    dout(10) << __func__ << ": no obc for soid "
         << soid << " but can_create"
         << dendl;
    // new object.
    object_info_t oi(soid);
    //调用get_snapset_context获取 SnapSetContext
    SnapSetContext *ssc = get_snapset_context(
      soid, true, 0, false);
        assert(ssc);
    obc = create_object_context(oi, ssc);
    dout(10) << __func__ << ": " << obc << " " << soid
         << " " << obc->rwstate
         << " oi: " << obc->obs.oi
         << " ssc: " << obc->ssc
         << " snapset: " << obc->ssc->snapset << dendl;
    return obc;
      }
    }
 ...
 ...

  }
}

9）

SnapSetContext *PrimaryLogPG::get_snapset_context(
  const hobject_t& oid,
  bool can_create,
  const map<string, bufferlist> *attrs,
  bool oid_existed)
{
  Mutex::Locker l(snapset_contexts_lock);
  SnapSetContext *ssc;
  map<hobject_t, SnapSetContext*>::iterator p = snapset_contexts.find(
    oid.get_snapdir());
  if (p != snapset_contexts.end()) {
    if (can_create || p->second->exists) {
      ssc = p->second;
    } else {
      return NULL;
    }
  } else {
    bufferlist bv;
    if (!attrs) {
      int r = -ENOENT;
      if (!(oid.is_head() && !oid_existed)) {
    r = pgbackend->objects_get_attr(oid.get_head(), SS_ATTR, &bv);
      }
      if (r < 0 && !can_create)
    return NULL;
    } else {
      assert(attrs->count(SS_ATTR));
      bv = attrs->find(SS_ATTR)->second;
    }
    ssc = new SnapSetContext(oid.get_snapdir());
    _register_snapset_context(ssc);
    if (bv.length()) {
      bufferlist::iterator bvp = bv.begin();
      try {
    ssc->snapset.decode(bvp);
      } catch (buffer::error& e) {
        dout(0) << __func__ << " Can't decode snapset: " << e << dendl;
    return NULL;
      }
      ssc->exists = true;
    } else {
      ssc->exists = false;
    }
  }
  assert(ssc);
  ssc->ref++;
  return ssc;
}

10）该函数是由do_op调用的， 主要工作是检查对象状态和上下文相关信息的获取，并调用函数prepare _transactions 把操作封装到事务中。如果是读取操作，则调用相关读取函数（同步、异步）。如果是写操作，则 调用calc_trim_to计算是否将旧的PG log日志进行trim操作、 issue_repop(repop, ctx)向各个副本发送同步操作请求、eval_repop(repop)检查发向各个副本的同步操作请求是否已经reply成功

void PrimaryLogPG::execute_ctx(OpContext *ctx)
{
  FUNCTRACE();
  dout(10) << __func__ << " " << ctx << dendl;
  ctx->reset_obs(ctx->obc);
  ctx->update_log_only = false; // reset in case finish_copyfrom() is re-running execute_ctx
  OpRequestRef op = ctx->op;
  const MOSDOp *m = static_cast<const MOSDOp*>(op->get_req());
  ObjectContextRef obc = ctx->obc;
  const hobject_t& soid = obc->obs.oi.soid;

  // this method must be idempotent since we may call it several times
  // before we finally apply the resulting transaction.
  ctx->op_t.reset(new PGTransaction);

  if (op->may_write() || op->may_cache()) {
    // snap
    if (!(m->has_flag(CEPH_OSD_FLAG_ENFORCE_SNAPC)) &&//如果是对整个pool的快照操作
    pool.info.is_pool_snaps_mode()) {
      // use pool's snapc
      ctx->snapc = pool.snapc;//设置为该值 pool的信息
    } else {//如果是用户特定的快照 如RBD
      // client specified snapc
      ctx->snapc.seq = m->get_snap_seq();//设置为信息带的相关信息
      ctx->snapc.snaps = m->get_snaps();
      filter_snapc(ctx->snapc.snaps);
    }
    if ((m->has_flag(CEPH_OSD_FLAG_ORDERSNAP)) &&
    ctx->snapc.seq < obc->ssc->snapset.seq) {//客户端的 snap序号小于服务端的 返回错误
      dout(10) << " ORDERSNAP flag set and snapc seq " << ctx->snapc.seq
           << " < snapset seq " << obc->ssc->snapset.seq
           << " on " << obc->obs.oi.soid << dendl;
      reply_ctx(ctx, -EOLDSNAPC);
      return;
    }
...

  if (!ctx->user_at_version)
    ctx->user_at_version = obc->obs.oi.user_version;
  dout(30) << __func__ << " user_at_version " << ctx->user_at_version << dendl;
//若是读操作，给objectContext加上ondisk_read_lock锁
  if (op->may_read()) {
    dout(10) << " taking ondisk_read_lock" << dendl;
    obc->ondisk_read_lock();
  }

  {
#ifdef WITH_LTTNG
    osd_reqid_t reqid = ctx->op->get_reqid();
#endif
    tracepoint(osd, prepare_tx_enter, reqid.name._type,
        reqid.name._num, reqid.tid, reqid.inc);
  }

  int result = prepare_transaction(ctx);//将相关的操作封装到 ctx->op_t中 封装成事务
  {
#ifdef WITH_LTTNG
    osd_reqid_t reqid = ctx->op->get_reqid();
#endif
    tracepoint(osd, prepare_tx_exit, reqid.name._type,
        reqid.name._num, reqid.tid, reqid.inc);
  }

  if (op->may_read()) {
    dout(10) << " dropping ondisk_read_lock" << dendl;
    obc->ondisk_read_unlock();
  }

  bool pending_async_reads = !ctx->pending_async_reobc->ondisk_read_lock();ads.empty();
  if (result == -EINPROGRESS || pending_async_reads) {
    // come back later.
    if (pending_async_reads) {
      assert(pool.info.is_erasure());
      in_progress_async_reads.push_back(make_pair(op, ctx));
      ctx->start_async_reads(this);//如果是，则调用该函数 异步读取
    }
    return;
  }

  if (result == -EAGAIN) {
    // clean up after the ctx
    close_op_ctx(ctx);
    return;
  }

  bool successful_write = !ctx->op_t->empty() && op->may_write() && result >= 0;
  // prepare the reply
  ctx->reply = new MOSDOpReply(m, 0, get_osdmap()->get_epoch(), 0,
                   successful_write, op->qos_resp);

  // read or error?
  if ((ctx->op_t->empty() || result < 0) && !ctx->update_log_only) {
    // finish side-effects
    if (result >= 0)
      do_osd_op_effects(ctx, m->get_connection());

    complete_read_ctx(result, ctx);//同步读取，
    return;
  }

  ctx->reply->set_reply_versions(ctx->at_version, ctx->user_at_version);

  assert(op->may_write() || op->may_cache());

  // trim log?
  calc_trim_to();//调用函数 计算是否将旧的PG log日志进行trim操作
  ...
  ...
  issue_repop(repop, ctx);//向各个副本发送同步操作请求
  eval_repop(repop);//检查发向各个副本的同步操作请求是否已经reply成功
  repop->put();
}

11）PrimaryLogPG::issue_repop函数 主要是把讲求发送到 副本OSD上进行处理

void PrimaryLogPG::issue_repop(RepGather *repop, OpContext *ctx)
{
  FUNCTRACE();
  const hobject_t& soid = ctx->obs->oi.soid;
  dout(7) << "issue_repop rep_tid " << repop->rep_tid
          << " o " << soid
          << dendl;

  repop->v = ctx->at_version;
  if (ctx->at_version > eversion_t()) {
    for (set<pg_shard_t>::iterator i = actingbackfill.begin();
     i != actingbackfill.end();
     ++i) {
      if (*i == get_primary()) continue;
      pg_info_t &pinfo = peer_info[*i];
      // keep peer_info up to date
      if (pinfo.last_complete == pinfo.last_update)
    pinfo.last_complete = ctx->at_version;
      pinfo.last_update = ctx->at_version;
    }
  }
 //为写做准备 给相关对象加ondisk_write_lock锁
  ctx->obc->ondisk_write_lock();

  ctx->op_t->add_obc(ctx->obc);
  if (ctx->clone_obc) {
    ctx->clone_obc->ondisk_write_lock();
    ctx->op_t->add_obc(ctx->clone_obc);
  }
  if (ctx->head_obc) {
    ctx->head_obc->ondisk_write_lock();
    ctx->op_t->add_obc(ctx->head_obc);
  }

  Context *on_all_commit = new C_OSD_RepopCommit(this, repop);
  Context *on_all_applied = new C_OSD_RepopApplied(this, repop);
  Context *onapplied_sync = new C_OSD_OndiskWriteUnlock(
    ctx->obc,
    ctx->clone_obc,
    ctx->head_obc);
  if (!(ctx->log.empty())) {
    assert(ctx->at_version >= projected_last_update);
    projected_last_update = ctx->at_version;
  }
  for (auto &&entry: ctx->log) {
    projected_log.add(entry);
  }
  //将事务发送到OSD处理，对于不同的PG实现，调用不同的类，PGBackend有两个子类，ReplicatedBackend 和 ECBackend 两个类对应不同的实现
  pgbackend->submit_transaction(
    soid,
    ctx->delta_stats,
    ctx->at_version,
    std::move(ctx->op_t),
    pg_trim_to,
    min_last_complete_ondisk,
    ctx->log,
    ctx->updated_hset_history,
    onapplied_sync,
    on_all_applied,
    on_all_commit,
    repop->rep_tid,
    ctx->reqid,
    ctx->op);
}

12）该函数用于最终调用网络接口，把更新请求发送给从OSD，并调用queue_transactions 函数对该PG的主OSD上的实现更改。

void ReplicatedBackend::submit_transaction(
  const hobject_t &soid,
  const object_stat_sum_t &delta_stats,
  const eversion_t &at_version,
  PGTransactionUPtr &&_t,
  const eversion_t &trim_to,
  const eversion_t &roll_forward_to,
  const vector<pg_log_entry_t> &_log_entries,
  boost::optional<pg_hit_set_history_t> &hset_history,
  Context *on_local_applied_sync,
  Context *on_all_acked,
  Context *on_all_commit,
  ceph_tid_t tid,
  osd_reqid_t reqid,
  OpRequestRef orig_op)
{
  parent->apply_stats(
    soid,
    delta_stats);

  vector<pg_log_entry_t> log_entries(_log_entries);
  ObjectStore::Transaction op_t;
  PGTransactionUPtr t(std::move(_t));
  set<hobject_t> added, removed;
  generate_transaction(
    t,
    coll,
    log_entries,
    &op_t,
    &added,
    &removed);
  assert(added.size() <= 1);
  assert(removed.size() <= 1);

  auto insert_res = in_progress_ops.insert(
    make_pair(
      tid,
      InProgressOp(
    tid, on_all_commit, on_all_acked,
    orig_op, at_version)
      )
    );
  assert(insert_res.second);
  //构件InProgressOp请求记录
  InProgressOp &op = insert_res.first->second;

  op.waiting_for_applied.insert(
    parent->get_actingbackfill_shards().begin(),
    parent->get_actingbackfill_shards().end());
  op.waiting_for_commit.insert(
    parent->get_actingbackfill_shards().begin(),
    parent->get_actingbackfill_shards().end());

  //调用该函数，把请求发送出去，发送到从OSD
  issue_op(
    soid,
    at_version,
    tid,
    reqid,
    trim_to,
    at_version,
    added.size() ? *(added.begin()) : hobject_t(),
    removed.size() ? *(removed.begin()) : hobject_t(),
    log_entries,
    hset_history,
    &op,
    op_t);

  add_temp_objs(added);
  clear_temp_objs(removed);

  parent->log_operation(
    log_entries,
    hset_history,
    trim_to,
    at_version,
    true,
    op_t);

  op_t.register_on_applied_sync(on_local_applied_sync);
  op_t.register_on_applied(
    parent->bless_context(
      new C_OSD_OnOpApplied(this, &op)));
  op_t.register_on_commit(
    parent->bless_context(
      new C_OSD_OnOpCommit(this, &op)));

  vector<ObjectStore::Transaction> tls;
  tls.push_back(std::move(op_t));

  parent->queue_transactions(tls, op.op);//调用该函数完成最后的操作，对该PG的主OSD上的本地对象完成操作
}

13） 调用的queue_transactions函数，会调用到os层。
调用的函数位于 PrinaryLogPG.h

  void queue_transactions(vector<ObjectStore::Transaction>& tls,
              OpRequestRef op) override {
    osd->store->queue_transactions(osr.get(), tls, 0, 0, 0, op, NULL);//最终调用到os层
  }

其中 osd->store 定义为
ObjectStore *store;