sqlite 提供了一种 redo log 型事务实现,支持读写的并发,见 write-ahead log(
https://sqlite.org/wal.html)。本文将介绍 wal 原理,并源码剖析 checkpoint 过程,同时讨论下 wal 使用中的一些注意点。由于 sqlite 的复杂性,会省略掉一些细节,重点放在核心流程和 wal 并发的实现。
1.1 redo log
sqlite wal 是一种简单的 redo log 事务实现,redo log 概念这里简述下。数据库事务需要满足满足 acid,其中原子性(a),即一次事务内的多个修改,要么全部提交成功要么全部提交失败,不存在部分提交到 db 的情况。 redo log 的解决思路是将修改后的日志按序先写入 log 文件(wal 文件),每个完成的事务会添加 checksum,可鉴别事务的完整性。事务写入日志文件后,即代表提交成功,读取时日志和 db 文件合并的结果构成了 db 的完整内容。同时定期 checkpoint,同步 wal 中的事务到 db 文件,使 wal 文件保持在合理的大小。日志文件持久化到磁盘后,已提交成功的事务按序 checkpoint 执行的结果都是一样的,不受 crash 和掉电的影响。
sqlite 的 wal 也是这种思路的实现,只是 sqlite 提供的是一种简化实现,同时只允许一个写者操作日志文件,日志也是 page 这种物理日志。redo log 还能将 undo log 的随机写转化为顺序写,具有更高的写入性能,这里不赘述。
想对 redo log 进一步了解,可以参考以下资料:
https://zhuanlan.zhihu.com/p/35574452
https://developer.aliyun.com/article/1009683
1.2 sqlite wal
sqlite wal 写操作不直接写入 db 主文件,而是写到“db-wal”文件(以下简称'wal'文件)的末尾。读操作时,将结合 db 主文件以及 wal 的内容返回结果。wal 模式同时具有简单的 mvvc 实现,支持文件级别的读写并发,提供了相对 delete(rollback) 模式 (undo log 事务) 更高的并发性。 具体可看图加深理解。
下图中:
关于写
关于读
关于 checkpoint:
wal 的实现大部分代码集中在 wal.c 中,从 sqlite 的架构划分应该主要算是 pager 层的实现。
https://www.sqlite.org/arch.html。wal 实现从逻辑上由 3 部分组成:
2.1 wal 和 wal-index 文件格式
文件格式定义,官方文档见:
https://www.sqlite.org/walformat.html
https://www.sqlite.org/fileformat2.html
这一层细节比较多,主要是些二进制定义。核心是 wal 格式提供了一种 page 格式的 redo log 组织格式,保证 crash 后 recover 过程满足一致性。
wal-index 文件(db-shm)只是一种对 wal 文件的快速索引,后文为了省事,也统称 wal 文件。
2.2 文件多副本抽象
即 wal 和 db 文件对外表现为一个统一的文件抽象,并提供文件级别的 mvcc,对 pager 层屏蔽 wal 细节。
由于 wal 和 db 一样都是以 pgno 的方式索引 page,按 pgno 替换就可以构造出不同版本的 b 树,比较简单。mvcc 主要通过 read lock 的 read mark 实现,前面有介绍过, 后面并发控制部分会详细举例介绍。
具体实现可看:
写入:https://github.com/sqlite/sqlite/blob/version-3.15.2/src/pager.c#L3077
读取:https://github.com/sqlite/sqlite/blob/version-3.15.2/src/wal.c#L2593
2.3 并发控制
通过文件锁保证并发操作不会损坏数据库文件,下一节详细讲解。
wal 支持读读、读写的并发,相比最初的 rollback journal 模式提供了更大的并发力度。但 wal 实现的是文件级别的并发,没有 mysql 表锁行锁的概念,一个 db 文件同时的并发写事务同时只能存在一个,不支持写的同时并发。checkpoint 也可能会 block 读写。
wal 并发实现上主要通过文件锁,和文件级别 mvcc 来实现文件级别的读写并发。 锁即下文源码中的 WAL_CKPT_LOCK,WAL_WRITE_LOCK 和WAL_READ_LOCK,出于简化问题考虑省略了 WAL_RECOVER_LOCK 等相关性不大的其他锁的讨论。mvcc 即通过文件多副本和 read mark 实现,后文也会详细介绍。
3.1 锁的分类和作用
官方介绍:https://www.sqlite.org/walformat.html
可看 2.3.1节 How the various locks are used
也可看下面简化分析:
3.2 锁的持有情况
数据库的访问,可以分为 3 类:读、写和checkpoint。事务对锁的持有不总是在事务一开始就持有,后文为了简化分析,会假设读写事务对锁的持有在事务开始时是已知的,并且与事务同生命周期。实际在读事务某些执行路径上也可能会持有 write lock,这里专注主线逻辑。
3.3 锁的应用
这部分可以和源码分析部分参照起来看,是整个 wal 里面相对复杂的部分,重点,需要来回反复看。
commit transaction:表示已经提交但没有 checkpoint 的事务,蓝框中表示事务修改的页面。
ongoing transition : 表示正在进行中的事务,同时也表示一个活跃的数据库连接,蓝线表示 read mark 的位置。
pgx.y: 表示 page 的页号和版本。
3.3.1 读写
如图可知:
3.3.2 checkpoint
这部分要和源码分析结合,如果此时发起 checkpoint。
源码对应 sqlite 3.15.2,通过直接调用 checkpoint 观察整个过程。
https://github.com/sqlite/sqlite/tree/version-3.15.2/src
4.1 调用链路
4.2 sqlite3_wal_checkpoint_v2
https://github.com/sqlite/sqlite/blob/version-3.15.2/src/main.c#L2065
主要是加锁和一些参数校验。
4.3 sqlite3Checkpoint
https://github.com/sqlite/sqlite/blob/version-3.15.2/src/main.c#L2146
ndb 上循环 checkpoint,大多数时候只有一个 db 文件。
4.4 sqlite3BtreeCheckpoint
https://github.com/sqlite/sqlite/blob/version-3.15.2/src/btree.c#L9472
检查 btree 是否 locked,也是前置检查逻辑。
4.5 sqlite3PagerCheckpoint
https://github.com/sqlite/sqlite/blob/version-3.15.2/src/pager.c#L7198
也是前置的处理逻辑。不过有个和 checkpoint 逻辑有关的。
/* 只在非SQLITE_CHECKPOINT_PASSIVE模式时设置xBusyHandler * 即SQLITE_CHECKPOINT_PASSIVE时如果获取不到锁,立即返回,不进行等待并retry */ if( pPager->pWal ){ rc = sqlite3WalCheckpoint(pPager->pWal, db, eMode, (eMode==SQLITE_CHECKPOINT_PASSIVE ? 0 : pPager->xBusyHandler), pPager->pBusyHandlerArg, pPager->walSyncFlags, pPager->pageSize, (u8 *)pPager->pTmpSpace, pnLog, pnCkpt ); }
4.6 sqlite3WalCheckpoint
https://github.com/sqlite/sqlite/blob/version-3.15.2/src/wal.c#L3192
int sqlite3WalCheckpoint( Wal *pWal, /* Wal connection */ int eMode, /* PASSIVE, FULL, RESTART, or TRUNCATE */ int (*xBusy)(void*), /* Function to call when busy */ void *pBusyArg, /* Context argument for xBusyHandler */ int sync_flags, /* Flags to sync db file with (or 0) */ int nBuf, /* Size of temporary buffer */ u8 *zBuf, /* Temporary buffer to use */ int *pnLog, /* OUT: Number of frames in WAL */ int *pnCkpt /* OUT: Number of backfilled frames in WAL */){ int rc; /* Return code */ int isChanged = 0; /* True if a new wal-index header is loaded */ int eMode2 = eMode; /* Mode to pass to walCheckpoint() */ int (*xBusy2)(void*) = xBusy; /* Busy handler for eMode2 */ assert( pWal->ckptLock==0 ); assert( pWal->writeLock==0 ); /* EVIDENCE-OF: R-62920-47450 The busy-handler callback is never invoked ** in the SQLITE_CHECKPOINT_PASSIVE mode. */ assert( eMode!=SQLITE_CHECKPOINT_PASSIVE || xBusy==0 ); if( pWal->readOnly ) return SQLITE_READONLY; WALTRACE(("WAL%p: checkpoint begins\n", pWal)); /* IMPLEMENTATION-OF: R-62028-47212 All calls obtain an exclusive ** "checkpoint" lock on the database file. */ // 独占获取WAL_CKPT_LOCK锁 rc = walLockExclusive(pWal, WAL_CKPT_LOCK, 1); if( rc ){ /* EVIDENCE-OF: R-10421-19736 If any other process is running a ** checkpoint operation at the same time, the lock cannot be obtained and ** SQLITE_BUSY is returned. ** EVIDENCE-OF: R-53820-33897 Even if there is a busy-handler configured, ** it will not be invoked in this case. */ testcase( rc==SQLITE_BUSY ); testcase( xBusy!=0 ); return rc; } pWal->ckptLock = 1; /* IMPLEMENTATION-OF: R-59782-36818 The SQLITE_CHECKPOINT_FULL, RESTART and ** TRUNCATE modes also obtain the exclusive "writer" lock on the database ** file. ** ** EVIDENCE-OF: R-60642-04082 If the writer lock cannot be obtained ** immediately, and a busy-handler is configured, it is invoked and the ** writer lock retried until either the busy-handler returns 0 or the ** lock is successfully obtained. */ // 非SQLITE_CHECKPOINT_PASSIVE时,独占获取WAL_WRITE_LOCK锁,并进行busy retry if( eMode!=SQLITE_CHECKPOINT_PASSIVE ){ rc = walBusyLock(pWal, xBusy, pBusyArg, WAL_WRITE_LOCK, 1); if( rc==SQLITE_OK ){ pWal->writeLock = 1; }else if( rc==SQLITE_BUSY ){ eMode2 = SQLITE_CHECKPOINT_PASSIVE; xBusy2 = 0; rc = SQLITE_OK; } } //如果wal-index显示db有变化,unfetch db文件,和主线逻辑关系不大 /* Read the wal-index header. */ if( rc==SQLITE_OK ){ rc = walIndexReadHdr(pWal, &isChanged); if( isChanged && pWal->pDbFd->pMethods->iVersion>=3 ){ sqlite3OsUnfetch(pWal->pDbFd, 0, 0); } } /* Copy data from the log to the database file. */ if( rc==SQLITE_OK ){ if( pWal->hdr.mxFrame && walPagesize(pWal)!=nBuf ){ rc = SQLITE_CORRUPT_BKPT; }else{ // checkpoint rc = walCheckpoint(pWal, eMode2, xBusy2, pBusyArg, sync_flags, zBuf); } /* If no error occurred, set the output variables. */ if( rc==SQLITE_OK || rc==SQLITE_BUSY ){ if( pnLog ) *pnLog = (int)pWal->hdr.mxFrame; if( pnCkpt ) *pnCkpt = (int)(walCkptInfo(pWal)->nBackfill); } } // release wal index,非主线逻辑 if( isChanged ){ /* If a new wal-index header was loaded before the checkpoint was ** performed, then the pager-cache associated with pWal is now ** out of date. So zero the cached wal-index header to ensure that ** next time the pager opens a snapshot on this database it knows that ** the cache needs to be reset. */ memset(&pWal->hdr, 0, sizeof(WalIndexHdr)); } // 释放锁,返回 /* Release the locks. */ sqlite3WalEndWriteTransaction(pWal); walUnlockExclusive(pWal, WAL_CKPT_LOCK, 1); pWal->ckptLock = 0; WALTRACE(("WAL%p: checkpoint %s\n", pWal, rc ? "failed" : "ok")); return (rc==SQLITE_OK && eMode!=eMode2 ? SQLITE_BUSY : rc);}
4.7 walCheckpoint
https://github.com/sqlite/sqlite/blob/version-3.15.2/src/wal.c#L1724
static int walCheckpoint( Wal *pWal, /* Wal connection */ int eMode, /* One of PASSIVE, FULL or RESTART */ int (*xBusy)(void*), /* Function to call when busy */ void *pBusyArg, /* Context argument for xBusyHandler */ int sync_flags, /* Flags for OsSync() (or 0) */ u8 *zBuf /* Temporary buffer to use */){ int rc = SQLITE_OK; /* Return code */ int szPage; /* Database page-size */ WalIterator *pIter = 0; /* Wal iterator context */ u32 iDbpage = 0; /* Next database page to write */ u32 iFrame = 0; /* Wal frame containing data for iDbpage */ u32 mxSafeFrame; /* Max frame that can be backfilled */ u32 mxPage; /* Max database page to write */ int i; /* Loop counter */ volatile WalCkptInfo *pInfo; /* The checkpoint status information */ szPage = walPagesize(pWal); testcase( szPage<=32768 ); testcase( szPage>=65536 ); pInfo = walCkptInfo(pWal); if( pInfo->nBackfill<pWal->hdr.mxFrame ){ /* Allocate the iterator */ rc = walIteratorInit(pWal, &pIter); if( rc!=SQLITE_OK ){ return rc; } assert( pIter ); /* EVIDENCE-OF: R-62920-47450 The busy-handler callback is never invoked ** in the SQLITE_CHECKPOINT_PASSIVE mode. */ assert( eMode!=SQLITE_CHECKPOINT_PASSIVE || xBusy==0 ); /* Compute in mxSafeFrame the index of the last frame of the WAL that is ** safe to write into the database. Frames beyond mxSafeFrame might ** overwrite database pages that are in use by active readers and thus ** cannot be backfilled from the WAL. */ mxSafeFrame = pWal->hdr.mxFrame; mxPage = pWal->hdr.nPage; /* 计算mxSafeFrame * 会尝试独占的获取aReadMark锁,如果获取到,则代表原先持有对应aReadMark锁的事务已经结束。 * 会不断的用busy rerty逻辑等待对应的读锁释放。 * 如果对应事物一直没有释放aReadMark锁,最终的 mxSafeFrame = MIN(unfinished_aReadMarks) */ for(i=1; i<WAL_NREADER; i++){ /* Thread-sanitizer reports that the following is an unsafe read, ** as some other thread may be in the process of updating the value ** of the aReadMark[] slot. The assumption here is that if that is ** happening, the other client may only be increasing the value, ** not decreasing it. So assuming either that either the "old" or ** "new" version of the value is read, and not some arbitrary value ** that would never be written by a real client, things are still ** safe. */ u32 y = pInfo->aReadMark[i]; if( mxSafeFrame>y ){ assert( y<=pWal->hdr.mxFrame ); // 尝试获取 WAL_READ_LOCK(i)锁,并进行忙等待 rc = walBusyLock(pWal, xBusy, pBusyArg, WAL_READ_LOCK(i), 1); if( rc==SQLITE_OK ){ // 成功获取 WAL_READ_LOCK(i)锁,设置为READMARK_NOT_USED;i==1,是个treak,不影响主流程 pInfo->aReadMark[i] = (i==1 ? mxSafeFrame : READMARK_NOT_USED); walUnlockExclusive(pWal, WAL_READ_LOCK(i), 1); }else if( rc==SQLITE_BUSY ){ // 一直没有获取对应WAL_READ_LOCK(i)锁,设置mxSafeFrame为y mxSafeFrame = y; xBusy = 0; }else{ goto walcheckpoint_out; } } } // 开始从wal文件写回db文件,此时独占的持有WAL_READ_LOCK(0) if( pInfo->nBackfill<mxSafeFrame && (rc = walBusyLock(pWal, xBusy, pBusyArg, WAL_READ_LOCK(0),1))==SQLITE_OK ){ i64 nSize; /* Current size of database file */ u32 nBackfill = pInfo->nBackfill; pInfo->nBackfillAttempted = mxSafeFrame; /* Sync the WAL to disk */ if( sync_flags ){ rc = sqlite3OsSync(pWal->pWalFd, sync_flags); } /* If the database may grow as a result of this checkpoint, hint ** about the eventual size of the db file to the VFS layer. */ if( rc==SQLITE_OK ){ i64 nReq = ((i64)mxPage * szPage); rc = sqlite3OsFileSize(pWal->pDbFd, &nSize); if( rc==SQLITE_OK && nSize<nReq ){ sqlite3OsFileControlHint(pWal->pDbFd, SQLITE_FCNTL_SIZE_HINT, &nReq); } } // 逻辑比较简单,遍历并回写 /* Iterate through the contents of the WAL, copying data to the db file */ while( rc==SQLITE_OK && 0==walIteratorNext(pIter, &iDbpage, &iFrame) ){ i64 iOffset; assert( walFramePgno(pWal, iFrame)==iDbpage ); if( iFrame<=nBackfill || iFrame>mxSafeFrame || iDbpage>mxPage ){ continue; } iOffset = walFrameOffset(iFrame, szPage) + WAL_FRAME_HDRSIZE; /* testcase( IS_BIG_INT(iOffset) ); // requires a 4GiB WAL file */ rc = sqlite3OsRead(pWal->pWalFd, zBuf, szPage, iOffset); if( rc!=SQLITE_OK ) break; iOffset = (iDbpage-1)*(i64)szPage; testcase( IS_BIG_INT(iOffset) ); rc = sqlite3OsWrite(pWal->pDbFd, zBuf, szPage, iOffset); if( rc!=SQLITE_OK ) break; } /* If work was actually accomplished... */ if( rc==SQLITE_OK ){ if( mxSafeFrame==walIndexHdr(pWal)->mxFrame ){ i64 szDb = pWal->hdr.nPage*(i64)szPage; testcase( IS_BIG_INT(szDb) ); rc = sqlite3OsTruncate(pWal->pDbFd, szDb); if( rc==SQLITE_OK && sync_flags ){ rc = sqlite3OsSync(pWal->pDbFd, sync_flags); } } if( rc==SQLITE_OK ){ /* 更新nBackfill为已经checkpoint的部分 * nBackfill记录当前已经checkpoint的部分 */ pInfo->nBackfill = mxSafeFrame; } } /* Release the reader lock held while backfilling */ // 释放 WAL_READ_LOCK(0) walUnlockExclusive(pWal, WAL_READ_LOCK(0), 1); } if( rc==SQLITE_BUSY ){ /* Reset the return code so as not to report a checkpoint failure ** just because there are active readers. */ rc = SQLITE_OK; } } /* If this is an SQLITE_CHECKPOINT_RESTART or TRUNCATE operation, and the ** entire wal file has been copied into the database file, then block ** until all readers have finished using the wal file. This ensures that ** the next process to write to the database restarts the wal file. */ // 非passive的checkpoint的区别都在这里 if( rc==SQLITE_OK && eMode!=SQLITE_CHECKPOINT_PASSIVE ){ assert( pWal->writeLock ); if( pInfo->nBackfill<pWal->hdr.mxFrame ){ // 没有全部checkpoint rc = SQLITE_BUSY; }else if( eMode>=SQLITE_CHECKPOINT_RESTART ){ // RESTART or TRUNCATE u32 salt1; sqlite3_randomness(4, &salt1); assert( pInfo->nBackfill==pWal->hdr.mxFrame ); // 获取所有读锁, 保证下一个事物能够重新开始restart,即循环利用wal文件 rc = walBusyLock(pWal, xBusy, pBusyArg, WAL_READ_LOCK(1), WAL_NREADER-1); if( rc==SQLITE_OK ){ if( eMode==SQLITE_CHECKPOINT_TRUNCATE ){ /* IMPLEMENTATION-OF: R-44699-57140 This mode works the same way as ** SQLITE_CHECKPOINT_RESTART with the addition that it also ** truncates the log file to zero bytes just prior to a ** successful return. ** ** In theory, it might be safe to do this without updating the ** wal-index header in shared memory, as all subsequent reader or ** writer clients should see that the entire log file has been ** checkpointed and behave accordingly. This seems unsafe though, ** as it would leave the system in a state where the contents of ** the wal-index header do not match the contents of the ** file-system. To avoid this, update the wal-index header to ** indicate that the log file contains zero valid frames. */ walRestartHdr(pWal, salt1); // Truncate wal文件 rc = sqlite3OsTruncate(pWal->pWalFd, 0); } walUnlockExclusive(pWal, WAL_READ_LOCK(1), WAL_NREADER-1); } } } walcheckpoint_out: walIteratorFree(pIter); return rc;}
5.1 checkpoint 何时触发
5.2 checkpoint 四种 mode 的区别
5.3 wal 下读写和 checkpoint 的并发性
可看看上面不同操作对锁的持有情况:
5.4 wal 文件巨大的原因 & 如何解决
5.4.1 原因
wal 文件提供的操作模型非常简单,只有在一次完整的 checkpoint 后才会重头开始循环利用 wal 文件,如果 checkpoint 一直没有提交当前的 wal 文件中所有更新,会导致 wal 文件无限增大。同时只有在 truncate 模式 checkpoint 才会缩减 wal 文件。
大概有以下原因会导致 wal 不能完全提交,核心都是 checkpoint 竞争不到锁。
在 checkpoint 中不能如预料中的获得锁,主要有两种可能:
5.4.2 解决方案
综上要解决 wal 无限增大主要有: