百度文件系统bfs源码分析系列(三)
2019-02-16 本文已影响8人
fooboo
写流程(下)
综合前两篇博文的分析，在正式开始写文件前，会先根据文件路径信息，在nameserver服务上会作些预处理，每一步都可能出现问题及不一致性，怎么解决会在后面统一回答。
这里先总结下之前的，put一个文件时，所发生的事情，此时并没有把文件中的内容上传至bfs：
- 对文件绝对路径所在每一层目录进行block编码，写入leveldb，除了该文件；
- 检查文件是否存在；
- 可能会删除原有文件已存在的block元数据;(nameserver/chunkserver)；
- 存该文件的block编码信息入leveldb；
- 可能要删除4)中的block元数据；
- 如果有ha功能，则保存log，后期通过raft同步log到follower后，会执行log_callback_，即1)和4)，follower节点收到日志后会在某个时间点进行apply：
602 void RaftNodeImpl::Init(std::function<void (const std::string& log)> callback,
603 std::function<void (int32_t, std::string*)> /*snapshot_callback*/) {
604 log_callback_ = callback; //NameSpace::TailLog
605 ApplyLog();
606 }
494 void RaftNodeImpl::ApplyLog() {
495 MutexLock lock(&mu_);
496 if (applying_ || !log_callback_) {
497 return;
498 }
499 applying_ = true;
500 for (int64_t i = last_applied_ + 1; i <= commit_index_; i++) {
501 std::string log;
502 StatusCode s = log_db_->Read(i, &log);
503 if (s != kOK) {
505 }
506 LogEntry entry;
507 bool ret = entry.ParseFromString(log);
508 assert(ret);
509 if (entry.type() == kUserLog) {
510 mu_.Unlock();
513 log_callback_(entry.log_data());///NameSpace::TailLog
514 mu_.Lock();
515 }
516 last_applied_ = entry.index();
517 }
519 StoreContext("last_applied", last_applied_);
520 applying_ = false;
521 }
756 void NameSpace::TailLog(const std::string& logstr) {
757 NameServerLog log;
758 if(!log.ParseFromString(logstr)) {
760 }
761 for (int i = 0; i < log.entries_size(); i++) {
762 const NsLogEntry& entry = log.entries(i);
763 int type = entry.type();
764 leveldb::Status s;
765 if (type == kSyncWrite) {
766 s = db_->Put(leveldb::WriteOptions(), entry.key(), entry.value());
767 } else if (type == kSyncDelete) {
768 s = db_->Delete(leveldb::WriteOptions(), entry.key());
769 }
770 if (!s.ok()) {
772 }
773 }
774 }
593 if (leader_commit > commit_index_) {//apply log
594 commit_index_ = leader_commit;
595 }
596 if (commit_index_ > last_applied_) {
597 thread_pool_->AddTask(std::bind(&RaftNodeImpl::ApplyLog, this));
598 }
以上说的block编码只是由parent_id和文件名(目录名)组成的key，由文件(目录)基本信息组成的value，且和chunkserver上的block不一样的，引用“可以通过Namespace中的内容以及ChunkServer的block report重建出来。每次NameServer重启时都会重建Blockmapping中的内容。耗时主要是顺序扫描leveldb的过程。”：
328 EncodingStoreKey(parent_id, fname, &file_key);
115 void NameSpace::EncodingStoreKey(int64_t entry_id,
116 const std::string& path,
117 std::string* key_str) {
118 key_str->resize(8);
119 common::util::EncodeBigEndian(&(*key_str)[0], (uint64_t)entry_id);
120 key_str->append(path);
121 }
当nameserver_client_->SendRequest返回成功后，会实例化FileImplWrapper类对象*file = new FileImplWrapper(this, rpc_client_, path, flags, write_option)：
62 class File {
63 public:
64 File() {}
65 virtual ~File() {}
66 virtual int32_t Pread(char* buf, int32_t read_size, int64_t offset, bool reada = false) = 0;
67 //for files opened with O_WRONLY, only support Seek(0, SEEK_CUR)
68 virtual int64_t Seek(int64_t offset, int32_t whence) = 0;
69 virtual int32_t Read(char* buf, int32_t read_size) = 0;
70 virtual int32_t Write(const char* buf, int32_t write_size) = 0;
71 virtual int32_t Flush() = 0;
72 virtual int32_t Sync() = 0;
73 virtual int32_t Close() = 0;
74 private:
75 // No copying allowed
76 File(const File&);
77 void operator=(const File&);
78 };
20 class FileImplWrapper : public File {
21 public:
22 FileImplWrapper(FSImpl* fs, RpcClient* rpc_client,
23 const std::string& name, int32_t flags, const WriteOptions& options);
26 FileImplWrapper(FileImpl* file_impl);
27 virtual ~FileImplWrapper();
28 //接口声明原型同File一样
35 private:
36 std::shared_ptr<FileImpl> impl_;
37 // No copying allowed
38 FileImplWrapper(const FileImplWrapper&);
39 void operator=(const FileImplWrapper&);
40 };
12 FileImplWrapper::FileImplWrapper(FSImpl* fs, RpcClient* rpc_client,
13 const std::string& name, int32_t flags, const WriteOptions& options) :
14 impl_(new FileImpl(fs, rpc_client, name, flags, options)) {}
上面只是封装的操作文件的类，最后还是调用FileImpl的接口，FileImpl太复杂，这里只列举因为写bfs而进行打开，写，关闭操作来分析，后面分析因写本地文件而读bfs的时候再分析：
69 class FileImpl : public File, public std::enable_shared_from_this<FileImpl> {
70 public:
71 FileImpl(FSImpl* fs, RpcClient* rpc_client, const std::string& name,
72 int32_t flags, const WriteOptions& options);
75 ~FileImpl ();
79 int32_t Write(const char* buf, int32_t write_size);
99 int32_t Flush();
100 int32_t Sync();
101 int32_t Close();
150 std::map<std::string, ChunkServer_Stub*> chunkservers_; ///< located chunkservers
126 private:
以下是部分BfsPut的实现：
253 char buf[10240];
255 int32_t bytes = 0;
256 while ( (bytes = fread(buf, 1, sizeof(buf), fp)) > 0) {//从源文件读
257 int32_t write_bytes = file->Write(buf, bytes);//写file对象
258 if (write_bytes < bytes) { ///判断是否写失败
260 ret = 2;
261 break;
262 }
264 }
265 fclose(fp);//关闭文件
266 if (file->Close() != 0) { //关闭file
268 ret = 1;
269 }
下面开始真正的写file对象：
388 int32_t FileImpl::Write(const char* buf, int32_t len) {
391 //more code...
402 if (open_flags_ & O_WRONLY) {
403 // Add block
404 MutexLock lock(&mu_, "Write AddBlock", 1000);
405 if (chunkservers_.empty()) {
406 int ret = 0;
407 for (int i = 0; i < FLAGS_sdk_createblock_retry; i++) {
408 ret = AddBlock();
409 if (ret == kOK) break;
410 sleep(10);
411 }
412 if (ret != kOK) {
414 common::atomic_dec(&back_writing_);
415 return ret;
416 }
417 }
418 }
419 //more code...
chunkservers_类似与chunkserver服务的连接，具体把内容写到哪台服务上，一开始为空，后面会创建；这里的实现是一个文件一个block，并没有按照一定的大小切割成多个block，分别存在不同的chunkserver中；下面是请求创建block的协议和实现：
9 message ChunkServerInfo {//
10 optional int32 id = 1;
11 optional string address = 2;
13 optional string start_time = 17;
14 optional int32 last_heartbeat = 3;
15 optional int64 data_size = 4;
17 optional int32 block_num = 7;
18 optional bool is_dead = 8;
19 optional ChunkServerStatus status = 9;
24 //more
38 }
54 message LocatedBlock {
55 optional int64 block_id = 1;
56 optional int64 block_size = 2;
57 repeated ChunkServerInfo chains = 3;
58 optional int32 status = 4;
59 }
105 message AddBlockRequest {
106 optional int64 sequence_id = 1;
107 optional string file_name = 2;
108 optional string client_address = 3;
109 }
110 message AddBlockResponse {
111 optional int64 sequence_id = 1;
112 optional StatusCode status = 2;
113 optional LocatedBlock block = 3;
114 }
314 int32_t FileImpl::AddBlock() {
315 AddBlockRequest request;
316 AddBlockResponse response;
317 request.set_sequence_id(0);
318 request.set_file_name(name_);
319 const std::string& local_host_name = fs_->local_host_name_;
320 request.set_client_address(local_host_name);
321 bool ret = fs_->nameserver_client_->SendRequest(&NameServer_Stub::AddBlock,
322 &request, &response, 15, 1);
323 //假设addblock成功(后面分析)
332 block_for_write_ = new LocatedBlock(response.block());
333 bool chains_write = IsChainsWrite();
334 int cs_size = chains_write ? 1 : block_for_write_->chains_size();
335 for (int i = 0; i < cs_size; i++) {
336 const std::string& addr = block_for_write_->chains(i).address();
337 rpc_client_->GetStub(addr, &chunkservers_[addr]);
338 write_windows_[addr] = new common::SlidingWindow<int>(100,
339 std::bind(&FileImpl::OnWriteCommit,
340 std::placeholders::_1, std::placeholders::_2));
341 cs_write_queue_[addr] = new WriteBufferQueue;
342 cs_errors_[addr] = false;
343 WriteBlockRequest create_request;
344 int64_t seq = common::timer::get_micros();
345 create_request.set_sequence_id(seq);
346 create_request.set_block_id(block_for_write_->block_id());
347 create_request.set_databuf("", 0);//第一次写
348 create_request.set_offset(0);//第一次写
349 create_request.set_is_last(false); //第一次写
350 create_request.set_packet_seq(0);//第一次写
351 WriteBlockResponse create_response;
352 if (chains_write) {
353 for (int i = 0; i < block_for_write_->chains_size(); i++) {
354 const std::string& cs_addr = block_for_write_->chains(i).address();
355 create_request.add_chunkservers(cs_addr);
356 }
357 }
358 bool ret = rpc_client_->SendRequest(chunkservers_[addr],
359 &ChunkServer_Stub::WriteBlock,
360 &create_request, &create_response,
361 25, 1);
362 if (!ret || create_response.status() != 0) {
365 //失败的处理
382 }
383 write_windows_[addr]->Add(0, 0);
384 }
385 last_seq_ = 0;
386 return OK;
387 }
以上是AddBlock实现，大概是sdk请求nameserver服务创建block管理数据，这部分后面分析，姑且认为这里创建成功并返回给sdk；然后IsChainsWrite()判断是否是链式写还是扇区写；循环要写的备份数，分别与chunkserver通信(protobuf rpc)，并建立类似窗口机制SlidingWindow，回调函数为OnWriteCommit，这里的实现后面分析，并为每个chunkserver维护写队列，其中这里的WriteBlockRequest对象为初始化数据，还没写：
138 std::map<std::string, common::SlidingWindow<int>* > write_windows_;
139 struct WriteBufferQueue {
140 Mutex mu;
141 std::queue<WriteBuffer*> buffers;
142 };
143 std::map<std::string, WriteBufferQueue*> cs_write_queue_;
然后sdk带着nameserver分配的block_id，请求chunkserver写block，即WriteBlockRequest，这里会判断是否是链式写，是的话由chunkserver写完交给下一个chunkserver写，最后返回到sdk，否则异步写chunkserver，后面会进行成功和失败的处理，跳过；此时还未真正的写文件内容至bfs，目前做的是sdk跟nameserver通信，分配blockid等这些，然后sdk跟chunkserver通信，写block请求，这里可能有点模糊，这里具体解释下先。
一）sdk与nameserver的通信，请求和响应协议为AddBlockRequest/AddBlockResponse，nameserver收到AddBlock请求后：
449 void NameServerImpl::AddBlock(...) {
453 if (!is_leader_) {
454 //不是leader
456 return;
457 }
458 response->set_sequence_id(request->sequence_id());
459 if (readonly_) {//默认为true，在LeaveReadOnly为false
460 //只能读不能写
463 return;
464 }
466 std::string path = NameSpace::NormalizePath(request->file_name());
467 FileLockGuard file_lock_guard(new WriteLock(path));
468 FileInfo file_info;
469 if (!namespace_->GetFileInfo(path, &file_info)) {
470 //无法获取路径信息，由之前的createfile编码
473 return;
474 }
476 if (file_info.blocks_size() > 0) {//移除之前的block信息，新文件一般不存在这样的情况
477 std::map<int64_t, std::set<int32_t> > block_cs;
478 block_mapping_manager_->RemoveBlocksForFile(file_info, &block_cs);
479 for (std::map<int64_t, std::set<int32_t> >::iterator it = block_cs.begin();
480 it != block_cs.end(); ++it) {
481 const std::set<int32_t>& cs = it->second;
482 for (std::set<int32_t>::iterator cs_it = cs.begin(); cs_it != cs.end(); ++cs_it) {
483 chunkserver_manager_->RemoveBlock(*cs_it, it->first);
484 }
485 }
486 file_info.clear_blocks();//这一步在createfile中没有
487 }
代码行476至487处理的事情是接着createfile处理的，在createfile中，如果文件存在但有截断选项会进行block_mapping_manager_->RemoveBlock(block_id)，这里可以认为是删除block_id列表(动态构建出来的)，并没有chunkserver_manager_->RemoveBlock，后面多进行了一次RemoveBlocksForFile(具体原因可能要参考后面代码)，且真正的file_info.clear_blocks()；
然后根据副本数进行一定规则选出chunkserver服务地址ChunkServerManager::GetChunkServerChains，后续由sdk与之通信；
489 int replica_num = file_info.replicas();
490 /// check lease for write
491 std::vector<std::pair<int32_t, std::string> > chains;
492 common::timer::TimeChecker add_block_timer;
493 if (chunkserver_manager_->GetChunkServerChains(replica_num, &chains, request->client_addres s())) {
495 int64_t new_block_id = namespace_->GetNewBlockId();
498 file_info.add_blocks(new_block_id);
499 file_info.set_version(-1);
500 ///TODO: Lost update? Get&Update not atomic.
501 for (int i = 0; i < replica_num; i++) {
502 file_info.add_cs_addrs(chunkserver_manager_->GetChunkServerAddr(chains[i].first));
503 }
504 NameServerLog log;
505 if (!namespace_->UpdateFileInfo(file_info, &log)) {
507 response->set_status(kUpdateError);
508 }
509 LocatedBlock* block = response->mutable_block();
510 std::vector<int32_t> replicas;
511 for (int i = 0; i < replica_num; i++) {
512 ChunkServerInfo* info = block->add_chains();
513 int32_t cs_id = chains[i].first;
514 info->set_address(chains[i].second);
517 replicas.push_back(cs_id);
518 // update cs -> block
519 add_block_timer.Reset();
520 chunkserver_manager_->AddBlock(cs_id, new_block_id);
521 add_block_timer.Check(50 * 1000, "AddBlock");
522 }
523 block_mapping_manager_->AddBlock(new_block_id, replica_num, replicas);
525 block->set_block_id(new_block_id);
526 response->set_status(kOK);
527 LogRemote(log, std::bind(&NameServerImpl::SyncLogCallback, this,
528 controller, request, response, done,
529 (std::vector<FileInfo>*)NULL,
530 file_lock_guard, std::placeholders::_1));
531 }//失败处理
分配下一个new_block_id，设置file_info数据，根据副本数格式化LocatedBlock给sdk，并设置chunkserver_manager_->AddBlock(cs_id, new_block_id)，即哪个chunkserver管理着哪些block：
22 class Blocks {
23 public:
24 Blocks(int32_t cs_id) : report_id_(-1), cs_id_(cs_id) {}
25 int64_t GetReportId();
26 void Insert(int64_t block_id);
27 void Remove(int64_t block_id);
28 void CleanUp(std::set<int64_t>* blocks);
29 void MoveNew();
30 int64_t CheckLost(int64_t report_id, const std::set<int64_t>& blocks,
31 int64_t start, int64_t end, std::vector<int64_t>* lost);
32 private:
33 Mutex block_mu_;
34 std::set<int64_t> blocks_; ////后面分析为啥会有new_blocks_
35 Mutex new_blocks_mu_;
36 std::set<int64_t> new_blocks_;
37 int64_t report_id_;
38 int32_t cs_id_; // for debug msg
39 };
62 void Blocks::MoveNew() {
63 MutexLock blocks_lock(&block_mu_);
64 MutexLock new_block_lock(&new_blocks_mu_);
65 std::set<int64_t> tmp;
66 blocks_.insert(new_blocks_.begin(), new_blocks_.end());
67 std::swap(tmp, new_blocks_);
68 }
接着上面，然后block_mapping_manager_->AddBlock，接着raft做同步，逻辑同createfile中一样：
86 void BlockMapping::AddBlock(int64_t block_id, int32_t replica,
87 const std::vector<int32_t>& init_replicas) {
88 NSBlock* nsblock = NULL;
89 nsblock = new NSBlock(block_id, replica, -1, 0);
90 if (nsblock->recover_stat == kNotInRecover) {
91 nsblock->replica.insert(init_replicas.begin(), init_replicas.end());
92 } else {
93 nsblock->incomplete_replica.insert(init_replicas.begin(), init_replicas.end());
94 }
95 //more code...
二）sdk与chunkserver的通信，请求和响应协议为WriteBlockRequest/WriteBlockResponse，当chunkserver服务收到请求后：
360 void ChunkServerImpl::WriteBlock(...) {
364 int64_t block_id = request->block_id();
365 const std::string& databuf = request->databuf();
366 int64_t offset = request->offset();
367 int32_t packet_seq = request->packet_seq();
368
369 if (!response->has_sequence_id() &&
370 g_unfinished_bytes.Add(databuf.size()) > FLAGS_chunkserver_max_unfinished_bytes) {
371 response->set_sequence_id(request->sequence_id());
372 //有太多的写没完成
375 response->set_status(kCsTooMuchUnfinishedWrite);
376 g_unfinished_bytes.Sub(databuf.size());
377 done->Run();
378 return;
379 }
380 if (!response->has_sequence_id()) {
381 response->set_sequence_id(request->sequence_id());
382 /// Flow control
383 if (g_block_buffers.Get() > FLAGS_chunkserver_max_pending_buffers) {
384 response->set_status(kCsTooMuchPendingBuffer);
388 g_unfinished_bytes.Sub(databuf.size());
389 done->Run();
390 g_refuse_ops.Inc();
391 return;
392 }
395 response->add_timestamp(common::timer::get_micros());
396 std::function<void ()> task =
397 std::bind(&ChunkServerImpl::WriteBlock, this, controller, request, response, done);
398 work_thread_pool_->AddTask(task);
399 return;
400 }
sdk过来的请求，response的sequence_id是没有值的，默认为false，所以最后都进381行代码，然后交给工作线程继续执行这个函数ChunkServerImpl::WriteBlock，后续respone就有sequence_id：
1615 inline void WriteBlockResponse::set_has_sequence_id() {
1616 _has_bits_[0] |= 0x00000002u;
1617 }
360 void ChunkServerImpl::WriteBlock(...) {
361 //more code...
406 int next_cs_offset = -1;
407 for (int i = 0; i < request->chunkservers_size(); i++) {
408 if (request->chunkservers(i) == data_server_addr_) {
409 next_cs_offset = i + 1;
410 break;
411 }
412 }
413 if (next_cs_offset >= 0 && next_cs_offset < request->chunkservers_size()) {//链式写
414 // share same write request
415 const WriteBlockRequest* next_request = request;
416 WriteBlockResponse* next_response = new WriteBlockResponse();
417 ChunkServer_Stub* stub = NULL;
418 const std::string& next_server = request->chunkservers(next_cs_offset); //下一个chunkserver的地址
419 rpc_client_->GetStub(next_server, &stub);
420 WriteNext(next_server, stub, next_request, next_response, request, response, done);
421 } else {//扇区写
422 std::function<void ()> callback =
423 std::bind(&ChunkServerImpl::LocalWriteBlock, this, request, response, done);
424 work_thread_pool_->AddTask(callback);
425 }
426 }
以上判断是链式写还是扇区写，即chunkservers个数大于1的时候是链式写；
如果是链式写，那么会把写请求转发给下一个chunkserver，代码行413〜420，即WriteNext(next_server, stub, next_request, next_response, request, response, done)：
428 void ChunkServerImpl::WriteNext(...) {
439 std::function<void (const WriteBlockRequest*, WriteBlockResponse*, bool, int)> callback =
440 std::bind(&ChunkServerImpl::WriteNextCallback,
441 this, std::placeholders::_1, std::placeholders::_2, std::placeholders::_3,
442 std::placeholders::_4, next_server, std::make_pair(request, response), done, stub);
443 rpc_client_->AsyncRequest(stub, &ChunkServer_Stub::WriteBlock,
444 next_request, next_response, callback, 30, 3);
445 }
75 template <class Stub, class Request, class Response, class Callback>
76 void AsyncRequest(Stub* stub, void(Stub::*func)(
77 google::protobuf::RpcController*,
78 const Request*, Response*, Callback*),
79 const Request* request, Response* response,
80 std::function<void (const Request*, Response*, bool, int)> callback,
81 int32_t rpc_timeout, int retry_times) {
82 sofa::pbrpc::RpcController* controller = new sofa::pbrpc::RpcController();
83 controller->SetTimeout(rpc_timeout * 1000L);
84 google::protobuf::Closure* done =
85 sofa::pbrpc::NewClosure(&RpcClient::template RpcCallback<Request, Response, Callback >,
86 controller, request, response, callback);//todo
87 (stub->*func)(controller, request, response, done);
88 }
即在下一个chunkserver下执行WriteBlock，回调为WriteNextCallback，下一个chunkserver执行完后会回调它，再判断是否要重试：
447 void ChunkServerImpl::WriteNextCallback(const WriteBlockRequest* next_request,
448 WriteBlockResponse* next_response,
449 bool failed, int error,
450 const std::string& next_server,
451 std::pair<const WriteBlockRequest*, WriteBlockResponse*> origin, ...) {
454 const WriteBlockRequest* request = origin.first;
455 WriteBlockResponse* response = origin.second;
456 /// If RPC_ERROR_SEND_BUFFER_FULL retry send.
457 if (failed && error == sofa::pbrpc::RPC_ERROR_SEND_BUFFER_FULL) { //重试
458 std::function<void ()> callback =
459 std::bind(&ChunkServerImpl::WriteNext, this, next_server,
460 stub, next_request, next_response, request, response, done);
461 work_thread_pool_->DelayTask(10, callback);
462 return;
463 }
464 delete stub;
467 const std::string& databuf = request->databuf();
470 if (failed || next_response->status() != kOK) {
471 //failed
480 delete next_response;
481 g_unfinished_bytes.Sub(databuf.size());
482 done->Run();
483 return;
484 } else {
485 //success
486 delete next_response;
487 }
488
489 std::function<void ()> callback =
490 std::bind(&ChunkServerImpl::LocalWriteBlock, this, request, response, done);
491 work_thread_pool_->AddTask(callback);
492 }
最终调用LocalWriteBlock进行本地写，扇区写也是如此：
494 void ChunkServerImpl::LocalWriteBlock(const WriteBlockRequest* request,
495 WriteBlockResponse* response,
496 ::google::protobuf::Closure* done) {
497 int64_t block_id = request->block_id();
498 const std::string& databuf = request->databuf();
499 int64_t offset = request->offset();
500 int32_t packet_seq = request->packet_seq();
503 //set response->set_status(kOK);
506 int64_t find_start = common::timer::get_micros();
507 /// search;
508 Block* block = NULL;
509
510 if (packet_seq == 0) {//FileImpl::AddBlock时packet_seq为0
511 StatusCode s;
512 block = block_manager_->CreateBlock(block_id, &s);
513 if (s != kOK) {
516 if (s == kBlockExist) {
517 int64_t expected_size = block->GetExpectedSize();
518 if (expected_size != request->total_size()) {
522 s = kWriteError;
523 } else {
524 response->set_current_size(block->Size());
525 response->set_current_seq(block->GetLastSeq());
528 }
529 }
530 response->set_status(s);
531 g_unfinished_bytes.Sub(databuf.size());
532 done->Run();
533 return;
这里姑且认为不存在处理，则会进行block_manager_->CreateBlock：
4 message BlockMeta {
5 optional int64 block_id = 1;
6 optional int64 block_size = 2;
7 optional int64 checksum = 3;
8 optional int64 version = 4 [default = -1];
9 optional string store_path = 5;
10 }
236 Block* BlockManager::CreateBlock(int64_t block_id, StatusCode* status) {
237 BlockMeta meta;
238 meta.set_block_id(block_id);
239 Disk* disk = PickDisk(block_id);//选择负载较小的disk
240 if (!disk) {
241 *status = kNotEnoughQuota;
243 return NULL;
244 }
245 meta.set_store_path(disk->Path());
246 Block* block = new Block(meta, disk, file_cache_); //disk_file_ = meta.store_path() + BuildFilePath(meta_.block_id());
247 //more code...
276 return block;
277 }
152 bool Disk::SyncBlockMeta(const BlockMeta& meta) {
153 std::string idstr = BlockId2Str(meta.block_id());
154 leveldb::WriteOptions options;
155 // options.sync = true;
156 std::string meta_buf;
157 meta.SerializeToString(&meta_buf);
158 leveldb::Status s = metadb_->Put(options, idstr, meta_buf);
159 if (!s.ok()) {
161 return false;
162 }
163 return true;
164 }
千呼万唤始出来，真正存储数据的block，后面在写具体内容时再分析：
39 class Block {
40 public:
41 Block(const BlockMeta& meta, Disk* disk, FileCache* file_cache);
42 ~Block();
83 private:
84 enum Type {
85 InDisk,
86 InMem,
87 };
101 std::string disk_file_;
114 };
packet_seq不为0则进行查找，接着写真正的内容：
494 void ChunkServerImpl::LocalWriteBlock(...) {
565 if (!block->Write(packet_seq, offset, databuf.data(),
566 databuf.size(), &add_used)) {
567 block->DecRef();
568 response->set_status(kWriteError);
569 g_unfinished_bytes.Sub(databuf.size());
570 done->Run();
571 return;
572 }
574 if (request->is_last()) {
575 if (request->has_recover_version()) {
576 block->SetVersion(request->recover_version());
579 }
580 block->SetSliceNum(packet_seq + 1);
581 }
583 // If complete, close block, and report only once(close block return true).
584 int64_t report_start = write_end;
585 if (block->IsComplete() &&
586 block_manager_->CloseBlock(block, request->sync_on_close())) {
590 ReportFinish(block);
591 }
340 bool ChunkServerImpl::ReportFinish(Block* block) {
341 BlockReceivedRequest request;
342 request.set_chunkserver_id(chunkserver_id_);
343 request.set_chunkserver_addr(data_server_addr_);
344
345 ReportBlockInfo* info = request.add_blocks();
346 info->set_block_id(block->Id());
347 info->set_block_size(block->Size());
348 info->set_version(block->GetVersion());
349 info->set_is_recover(block->IsRecover());
350 BlockReceivedResponse response;
351 if (!nameserver_->SendRequest(&NameServer_Stub::BlockReceived, &request, &response, 20)) {
353 return false;
354 }
357 return true;
358 }
具体的写后面再分析，这里put一个文件时，会写一个的0的block(实际实现过程中，Sdk在AddBlock时，会首先发送一个长度为0的写请求，促使ChunkServer提前创建好写block所需的文件，但是这时该block的大小为0)；并向nameserver报告写完，nameserver更新状态：
202 void NameServerImpl::BlockReceived(...) {
206 if (!is_leader_) {
207 response->set_status(kIsFollower);
208 done->Run();
209 return;
210 }
211 g_block_report.Inc();
212 response->set_sequence_id(request->sequence_id());
213 int32_t cs_id = request->chunkserver_id();
216 const ::google::protobuf::RepeatedPtrField<ReportBlockInfo>& blocks = request->blocks();
217
219 int old_id = chunkserver_manager_->GetChunkServerId(request->chunkserver_addr());
221 if (cs_id != old_id) {
224 response->set_status(kUnknownCs);
225 done->Run();
226 return;
227 }
228 for (int i = 0; i < blocks.size(); i++) {
230 const ReportBlockInfo& block = blocks.Get(i);
231 int64_t block_id = block.block_id();
232 int64_t block_size = block.block_size();
233 int64_t block_version = block.version();
236 // update block -> cs;
238 if (block_mapping_manager_->UpdateBlockInfo(block_id, cs_id, block_size, block_version) ) {
240 // update cs -> block
241 chunkserver_manager_->AddBlock(cs_id, block_id);
243 } else {
246 }
247 }
248 response->set_status(kOK);
249 done->Run();
250 }
以上差不多是nameserver和chunkserver整个写的大概过程，有些细节没有具体分析，会在下一篇中分析。
像nameserver上有些数据是动态维护的，并不会保存到db中，如果在某个删除过程中造成不一致，比如sdk与nameserver通信，nameserver处理到一半，然后sdk那边返回错误，则会在后面由chunkserver通过心跳或者report等消息，重新恢复不一致的数据，后期分析的差不多会总结下整个系统的设计。
但是有些不一致性的数据没多大关系，比如put文件时，先对路径编码进leveldb，后面如果不传文件或者sdk返回错误，通信中断啥的，虽然在nameserver上占用一定的空间，可以在后面定时的处理那些空的目录啥的情况。
以上建立初始化后，真正向bfs写文件内容，即向chunkserver传输字节内容，放在下一篇分析，这部分内容也多。
AsyncRequest和SendRequest从代码实现和接口名上来看，应该是个异步和同步调用，可能要深入析下sofa::pbrpc代码实现；
这里raft使用的log是自己实现的，有时间分析下。
后面分析下每个server启动时做了哪些工作，初始化哪些数据结构，和通信。
一些bfs qa：
Q：BlockMapping内的东西不落地么？重启需重新加载么？耗时多少？
A：BlockMapping内的东西确实是不落地的。可以通过Namespace中的内容以及ChunkServer的block report重建出来。每次NameServer重启时都会重建Blockmapping中的内容。耗时主要是顺序扫描leveldb的过程。
Q：AddBlock在什么时候会被调用？
A：AddBlock在SDK首次向 一个新创建的文件调用Write时被调用，然后sdk会与NameServer通信，Namserver为其分配block id，以及挑选出足够的ChunkServer供sdk写入使用。
Q：为什么不在创建文件的时候就直接AddBlock？
A：因为有可能为一些空文件，这里使用lazy的模式，到最后一步再去创建。
Q：什么时候chunkserver上会有一个大小为零的block？
A：Sdk创建了一个文件后，根本没进行过Write，这时，只有在NameServer中有该文件的元信息，但是ChunkServer上是没有的。
Sdk向一个新建的文件发起了写请求，但是该写请求在反映到ChunkServer中时，ChunkServer刚创建完文件就挂了，此里该block的大小为0。
实际实现过程中，Sdk在AddBlock时，会首先发送一个长度为0的写请求，促使ChunkServer提前创建好写block所需的文件，但是这时该block的大小为0.（@世光和@丽媛昨天说的应该是这种情况？）
