WiredTiger存储引擎分析

初始化

mongodb中wiredtiger初始化

MONGO_INITIALIZER_WITH_PREREQUISITES(WiredTigerEngineInit, ("SetGlobalEnvironment"))
(InitializerContext* context) {
    getGlobalServiceContext()->registerStorageEngine(kWiredTigerEngineName,
                                                     new WiredTigerFactory());

    return Status::OK();
}

主要的初始化相关处理：

1、storeMongodOptions()：初始化存储引擎相关配置参数。
2、KVStorageEngine类实例初始化时会调用_engine->createRecordStore、_catalog->init(&opCtx);，然后初始化所有已有的collection及对应db。
3、WiredTigerKVEngine中调用如下函数，建立wiredtiger存储引擎的_conn。

wiredtiger_open(path.c_str(), &_eventHandler, config.c_str(), &_conn);

4、WiredTigerKVEngine构造函数中，实例化了wiredtiger存储引擎的_sessionCache。

_sessionCache.reset(new WiredTigerSessionCache(this));

class WiredTigerSessionCache {
   ... ...
   typedef std::vector<WiredTigerSession*> SessionCache;
   SessionCache _sessions;
   ...
}
class WiredTigerSession {
   ... ...
   WiredTigerSession(WT_CONNECTION* conn, uint64_t epoch = 0, uint64_t cursorEpoch = 0);
   ... ...
   typedef std::list<WiredTigerCachedCursor> CursorCache;
   WT_SESSION* _session;  // owned
   CursorCache _cursors;  // owned
   ... ...
}

collection初始化流程

1、以mongodb接收到insert命令后的处理为例：

CmdInsert::CmdInsert() : WriteCmd("insert", BatchedCommandRequest::BatchType_Insert) {}

bool WriteCmd::run(...){
    ... ...
    writeBatchExecutor.executeBatch(request, &response);
    ... ...
}

WriteBatchExecutor::executeBatch() -> WriteBatchExecutor::bulkExecute() -> WriteBatchExecutor::execInserts() 
-> WriteBatchExecutor::insertMany() -> insertOne()

static void insertOne(WriteBatchExecutor::ExecInsertsState* state, WriteOpResult* result){
    ... ...
    if (state->lockAndCheck(result)) {
    ... ...
}

调用lockAndCheck函数，最终会调用Database::createCollection创建并初始化collection。
2、Database::createCollection会调用如下代码创建collection：

... ...
Status status = _dbEntry->createCollection(txn, ns, options, true /*allocateDefaultSpace*/);
Collection* collection = _getOrCreateCollectionInstance(txn, ns);
_collections[ns] = collection;
... ...

其中，KVDatabaseCatalogEntry中createCollection函数会初始化collection相关名字和对应的WT_SESSION。

Status KVDatabaseCatalogEntry::createCollection(OperationContext* txn, StringData ns, 
                    const CollectionOptions& options, bool allocateDefaultSpace)
{
    ... ...
    Status status = _engine->getCatalog()->newCollection(txn, ns, options);
    string ident = _engine->getCatalog()->getCollectionIdent(ns);
    status = _engine->getEngine()->createRecordStore(txn, ns, ident, options);
    RecordStore* rs = _engine->getEngine()->getRecordStore(txn, ns, ident, options);
    _collections[ns.toString()] = new KVCollectionCatalogEntry(_engine->getEngine(), _engine->getCatalog(), ns, ident, rs);
    ... ...
}

1）KVCatalog::newCollection()函数会调用_newUniqueIdent()函数获得collection对应的文件名字（ident），并且，会将collection的ns、ident存储到元数据文件_mdb_catalog中。

元数据文件_mdb_catalog对应的io结构体和btree会在KVStorageEngined的构造函数中初始化完成。

Status status = _engine->createRecordStore(&opCtx, catalogInfo, catalogInfo, CollectionOptions());
_catalogRecordStore.reset( _engine->getRecordStore(&opCtx, catalogInfo, catalogInfo, CollectionOptions()));
_catalog.reset(new KVCatalog(_catalogRecordStore.get(),_supportsDocLocking,
                      _options.directoryPerDB,_options.directoryForIndexes));
_catalog->init(&opCtx);
//由于每个collection创建时都会存储到元数据文件_mdb_catalog中，
//因此，可以直接从这个文件中得到所有已创建的collection。
_catalog->getAllCollections(&collections);  

2）KVDatabaseCatalogEntry::createCollection()函数会调用WiredTigerKVEngine::createRecordStore()函数创建和初始化WT_SESSION结构体。同时会调用__session_create函数创建文件和初始化对应的btree。

3、 WiredTigerKVEngine::createRecordStore()函数会调用WT_SESSION结构体的__session_create函数创建collection对应ident文件和初始化对应的btree。

Status WiredTigerKVEngine::createRecordStore(OperationContext* opCtx, StringData ns,
                                 StringData ident, const CollectionOptions& options) {
    ... ...
    //ident是类似这种形式：collection-0-rand.wt，_uri(ident)返回的是string("table:") + ident.toString();
    string uri = _uri(ident);  
    WT_SESSION* s = session.getSession();
    return wtRCToStatus(s->create(s, uri.c_str(), config.c_str()));
}

4、WT_SESSION结构体的__session_create函数会创建collection对应ident文件和初始化对应的btree。其中，uri的格式为table:collection-x-xxxxx。

__session_create(WT_SESSION *wt_session, const char *uri, const char *config) -> __wt_session_create() ->
__wt_schema_create() -> __create_table() 

static int __create_table(WT_SESSION_IMPL *session, const char *name, bool exclusive, const char *config){
    ... ...
    if (!exists) {
        WT_ERR(__wt_metadata_insert(session, name, tableconf));
        WT_ERR(__wt_schema_get_table(session, tablename, strlen(tablename), true, &table));

        if (ncolgroups == 0) {
            cgsize = strlen("colgroup:") + strlen(tablename) + 1;
            WT_ERR(__wt_calloc_def(session, cgsize, &cgname));
            snprintf(cgname, cgsize, "colgroup:%s", tablename);
            WT_ERR(__create_colgroup(session, cgname, exclusive, config));
        }
    }
    ... ...
}

5、如果ncolgroups为0，即配置文件中没有配置colgroup，则会调用__create_colgroup()函数，在该函数中又会调用__wt_schema_create()函数会创建该table对应的source file和初始化对应的btree。

static int __create_colgroup(WT_SESSION_IMPL *session,const char *name, bool exclusive, const char *config) {
    ... ...
    //__wt_schema_colgroup_source会得到file:collection-x-xxxxx.wt的文件名，后面设置到配置文件中source参数中。
    WT_ERR(__wt_schema_colgroup_source(session, table, cgname, config, &namebuf));
    source = namebuf.data;
    WT_ERR(__wt_buf_fmt(session, &confbuf, "source=\"%s\"", source));
    ... ...
    WT_ERR(__wt_schema_create(session, source, sourceconf));  //创建之前得到的source文件名对应文件的io结构体和btree。
    ... ...
    if (!exists) {
        WT_ERR(__wt_metadata_insert(session, name, cgconf));
        WT_ERR(__wt_schema_open_colgroups(session, table));
    }
}

6、__wt_schema_create()函数会调用__create_file()函数实际完成创建table对应的source file文件和初始化对应的btree的功能。

static int __create_file(WT_SESSION_IMPL *session, const char *uri, bool exclusive, const char *config){
    /* create、open对应的source文件*/
    WT_ERR(__wt_block_manager_create(session, filename, allocsize));
    ... ...
    //__wt_session_get_btree()函数中会调用__wt_conn_btree_open()函数创建对应的btree。
    WT_ERR(__wt_session_get_btree(session, uri, NULL, NULL, WT_DHANDLE_EXCLUSIVE));
}

int __wt_conn_btree_open(WT_SESSION_IMPL *session, const char *cfg[], uint32_t flags){
    ... ...
    WT_ERR(__wt_btree_open(session, cfg));
    ... ...
}

int __wt_btree_open(WT_SESSION_IMPL *session, const char *op_cfg[]){
    ... ...
    // Initialize and configure the WT_BTREE structure
    WT_ERR(__btree_conf(session, &ckpt));
    //__wt_block_manager_open()函数内部会调用__bm_method_set()函数设置btree->bm结构体(WT_BM)的相关方法。
    WT_ERR(__wt_block_manager_open(session, filename, dhandle->cfg, forced_salvage, readonly, btree->allocsize, &btree->bm));
    ... ...
    WT_ERR(bm->checkpoint_load(bm, session, ckpt.raw.data, ckpt.raw.size, root_addr, &root_addr_size, readonly));
    if (creation || root_addr_size == 0)
        WT_ERR(__btree_tree_open_empty(session, creation));
    else {
        WT_ERR(__wt_btree_tree_open(session, root_addr, root_addr_size));
        if (!F_ISSET(btree, WT_BTREE_REBALANCE)) {
            WT_WITH_PAGE_INDEX(session, ret = __btree_preload(session));
            if (btree->type != BTREE_ROW)  WT_ERR(__btree_get_last_recno(session));
        }
    }
}

__btree_tree_open_empty()或者__wt_btree_tree_open()函数会初始化一颗btree，每次分配都是按linux的一个page的大小分配的，__btree_conf()函数里会调用__btree_page_sizes()函数设置btree相关的。

collection的insert流程

1、以insert一个collection为例，当insertOne()函数中创建了collection后，会调用collection的insertDocument函数插入相关文档数据。

static void insertOne(WriteBatchExecutor::ExecInsertsState* state, WriteOpResult* result){
    ... ...
    Status status = state->getCollection()->insertDocument(txn, insertDoc, true);
    ... ...
}

Status Collection::insertDocument(OperationContext* txn, const DocWriter* doc, bool enforceQuota) {
    ... ...
    StatusWith<RecordId> loc = _recordStore->insertRecord(txn, doc, _enforceQuota(enforceQuota));
    ... ...
}

_recordStore->insertRecord(xxx)最终会调用如下函数：

Status WiredTigerRecordStore::insertRecords(OperationContext* txn, std::vector<Record>* records, bool enforceQuota) {
    ... ...
    //会先调用__session_open_cursor()函数打开一个cursor，uri是table:collection-x-xxxxx的格式
    WiredTigerCursor curwrap(_uri, _tableId, true, txn);
    WT_CURSOR* c = curwrap.get();
    ... ...
    for (auto& record : *records) {
        c->set_key(c, _makeKey(record.id));
        WiredTigerItem value(record.data.data(), record.data.size());
        c->set_value(c, value.Get());
        int ret = WT_OP_CHECK(c->insert(c));
    }
    ... ...
}

2、__session_open_cursor()函数会获取或打开btree相关结构及初始化WT_CURSOR结构体的相关函数指针成员。

static int __session_open_cursor(WT_SESSION *wt_session, const char *uri, WT_CURSOR *to_dup, 
                      const char *config, WT_CURSOR **cursorp){
    ... ...
    WT_ERR(__session_open_cursor_int(session, uri, NULL, statjoin ? to_dup : NULL, cfg, &cursor));
    ... ...
}

static int __session_open_cursor_int(WT_SESSION_IMPL *session, const char *uri,
        WT_CURSOR *owner, WT_CURSOR *other, const char *cfg[], WT_CURSOR **cursorp){
    ... ...
    WT_RET(__wt_curtable_open(session, uri, owner, cfg, cursorp));
    ... ...
}

int __wt_curtable_open(WT_SESSION_IMPL *session, const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp){
    ... ...
    if (table->is_simple) {
        ret = __wt_open_cursor(session, table->cgroups[0]->source, NULL, cfg, cursorp);
    }
    ... ...
}

int __wt_open_cursor(WT_SESSION_IMPL *session, const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp)
{
    return (__session_open_cursor_int(session, uri, owner, NULL, cfg, cursorp));
}

static int __session_open_cursor_int(WT_SESSION_IMPL *session, const char *uri,
            WT_CURSOR *owner, WT_CURSOR *other, const char *cfg[], WT_CURSOR **cursorp){
    ... ...
    WT_RET(__wt_curfile_open(session, uri, owner, cfg, cursorp));
    ... ...
}

int __wt_curfile_open(WT_SESSION_IMPL *session, const char *uri, WT_CURSOR *owner, const char *cfg[], WT_CURSOR **cursorp){
    ... ...
    ret = __wt_session_get_btree_ckpt(session, uri, cfg, flags);
    ... ...
    WT_ERR(__wt_curfile_create(session, owner, cfg, bulk, bitmap, cursorp));
    ... ...
}

__wt_curfile_create()函数会初始化WT_CURSOR结构体的相关函数指针成员，c->insert(c)调用时即会调用WT_CURSOR结构体中初始化的insert指针成员。

3、c->insert(c)最终会调用__curfile_insert()函数完成内存数据的修改和log文件的写入。

1、第一步会先调用__txn_logrec_init()函数初始化__wt_txn::logrec，然后再插入__wt_txn::logrec中。
__curfile_insert() -> __wt_btcur_insert() -> __cursor_row_modify() -> __wt_row_modify() -> __wt_txn_log_op() -> __txn_op_log()。
2、第一步的操作完成后，会调用WiredTigerRecoveryUnit::_commit()函数执行log日志的提交。
insertOne() -> WriteUnitOfWork::commit() -> WiredTigerRecoveryUnit::commitUnitOfWork() -> WiredTigerRecoveryUnit::_commit() -> WiredTigerRecoveryUnit::_txnClose() -> s->commit_transaction(…) -> __session_commit_transaction() -> __wt_txn_commit() -> __wt_txn_log_commit()函数完成日志写入。

config_def.c文件中wiredtiger_open_all的配置设置了transaction_sync.method会刷入硬盘。

3、wiredtiger_open() -> __wt_connection_workers() -> __wt_logmgr_open()，在函数中会创建线程，执行__log_server()函数，专门执行journal日志的落地。

#define WT_LOG_FILENAME "WiredTigerLog"     /* Log file name */
#define WT_LOG_PREPNAME "WiredTigerPreplog" /* Log pre-allocated name */
#define WT_LOG_TMPNAME  "WiredTigerTmplog"  /* Log temporary name */

cache淘汰

__wt_btcur_insert() -> __cursor_func_init() -> __curfile_enter() -> __cursor_enter() -> __wt_cache_eviction_check() -> __wt_eviction_needed()、__wt_cache_eviction_worker()

CheckPoint + WAL

MongoDB 3.2.9 请求 hang 分析及 wiredtiger 调优
MongoDB WiredTiger 存储引擎(1) cache_pool设计
MongoDB WiredTiger 存储引擎cache_pool设计 (下) – 实践篇
Mongodb如何使用wiredTiger
7-10倍写入性能提升:剖析WiredTiger数据页无锁及压缩黑科技
解析MongoDB存储引擎WiredTiger：事务实现