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。

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OSDMAPdispatch_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OSDif (!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);// finishdout(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 osdmapauto p = waiting_for_map.find(op->get_source());
...//是否可以丢弃if (can_discard_request(op)) {return;}
...
...//PG还没有peeredif (!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 contexthandle_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 TrackedOpm->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OSDif (!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blockedif (!m->has_flag(CEPH_OSD_FLAG_FLUSH) &&maybe_await_blocked_head(oid, op)) {return;}//调用find_object_context获取object_contextint 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 logsmaybe_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-ENOENTif (!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 headif (ssc)put_snapset_context(ssc);return -ENOENT;}
//如果是map_snapid_to_cloneif (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 readableObjectContextRef 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 diskbufferlist 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 SnapSetContextSnapSetContext *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_ctxOpRequestRef 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()) {// snapif (!(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 snapcctx->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 snapcctx->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_LTTNGosd_reqid_t reqid = ctx->op->get_reqid();
#endiftracepoint(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_LTTNGosd_reqid_t reqid = ctx->op->get_reqid();
#endiftracepoint(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 ctxclose_op_ctx(ctx);return;}bool successful_write = !ctx->op_t->empty() && op->may_write() && result >= 0;// prepare the replyctx->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-effectsif (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 dateif (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OSDissue_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;