leveldb的write代码初看瞎搞一堆，细看则实为短小精悍。1 Status DBImpl::Write(const WriteOptions& options, WriteBatch* my_batch) {
2 // -----A begin-------
3 Writer w(&mutex_);
4 w.batch = my_batch;
5 w.sync = options.sync;
6 w.done = false;
7 // -----A end --------
8
9
10 // -----B begin-------
11 MutexLock l(&mutex_);
12 writers_.push_back(&w);
13 while (!w.done && &w != writers_.front()) {
14 w.cv.Wait();
15 }
16 if (w.done) {
17 return w.status;
18 }
19 // -----B end -------
20
21 // May temporarily unlock and wait.
22 Status status = MakeRoomForWrite(my_batch == NULL);
23 uint64_t last_sequence = versions_->LastSequence();
24 Writer* last_writer = &w;
25 if (status.ok() && my_batch != NULL) { // NULL batch is for compactions
26 WriteBatch* updates = BuildBatchGroup(&last_writer);
27 WriteBatchInternal::SetSequence(updates, last_sequence + 1);
28 last_sequence += WriteBatchInternal::Count(updates);
29
30 // Add to log and apply to memtable. We can release the lock
31 // during this phase since &w is currently responsible for logging
32 // and protects against concurrent loggers and concurrent writes
33 // into mem_.
34 {
35 // -----C begin-------
36 mutex_.Unlock();
37 // -----C end -------
// add batch data to log files
38 status = log_->AddRecord(WriteBatchInternal::Contents(updates));
39 if (status.ok() && options.sync) {
40 status = logfile_->Sync();
41 }
42 if (status.ok()) {
43 status = WriteBatchInternal::InsertInto(updates, mem_);
44 }
45 // -----D begin-------
46 mutex_.Lock();
47 // -----D end -------
48 }
49 if (updates == tmp_batch_) tmp_batch_->Clear();
50
51 versions_->SetLastSequence(last_sequence);
52 }
53
54 // -----E begin-------
55 while (true) {
56 Writer* ready = writers_.front();
57 writers_.pop_front();
58 if (ready != &w) {
59 ready->status = status;
60 ready->done = true;
61 ready->cv.Signal();
62 }
63 if (ready == last_writer) break;
64 }
65 // -----E end -------
66
67
68 // -----F begin-------
69 // Notify new head of write queue
70 if (!writers_.empty()) {
71 writers_.front()->cv.Signal();
72 }
73 // -----F end-------
74
75 return status;
76 }
如上，A段代码定义一个Writer w, w的成员包括了batch信息，同时初始化了一个条件变量成员(port::CondVar)
假设同时有w1, w2, w3, w4, w5, w6 并发请求写入。
B段代码让竞争到mutex资源的w1获取了锁。添加到writers队列里去，此时队列只有一个w1, 从而其顺利的进行BuildBatchGroup。当运行到c段代码时，mutex互斥锁释放，这时(w2, w3, w4, w5, w6)会竞争锁，由于B段代码中不满足队首条件，均等待并释放锁了。从而队列可能会如(w3, w5, w2, w4).
继而w1进行log写入和memtable写入，之所以这里在无锁状况下时安全的，因为其它的写操作都不满足队首条件，进而不会进入log和memtable写入阶段。 当w1完成log和memtable写入后，进入d段代码，则mutex又锁住，这时B段代码中队列因为获取不到锁则队列不会修改。
进入E段代码后，w1被pop出来，由于reader==w, 并且ready==last_writer,所以直接到F段代码，唤醒了此时处于队首的w3.
w3唤醒时，发现自己是队首，可以顺利的进行进入BuildBatchGroup，在该函数中，遍历了目前所有的队列元素，形成一个update的batch，即将w3, w5, w2, w4合并为一个batch. 并将last_writer置为此时处于队尾的最后一个元素w4，c段代码运行后，因为释放了锁资源，队列可能随着DBImpl::Write的调用而更改，如队列状况可能为(w3, w5, w2, w4, w6, w9, w8).
C段和D段间的代码将w3, w5, w2, w4整个的batch写入log和memtable. 到E段时，分别对w5, w2, w4进行了一次cond signal.当判断到完w4 == lastwriter时，则退出E段代码。F段则对队首的w6唤醒，从而按上述步骤依次进行下去。
这样就形成了多个并发write 合并为一个batch写入log和memtable的机制。
