SQLite---Page Cache之事务处理(2)

写在前面：个人认为pager层是SQLite实现最为核心的模块，它具有四大功能：I/O，页面缓存，并发控制和日志恢复。而这些功能不仅是上层Btree的基础，而且对系统的性能和健壮性有关至关重要的影响。其中并发控制和日志恢复是事务处理实现的基础。SQLite并发控制的机制非常简单——封锁机制；别外，它的查询优化机制也非常简单——基于索引。这一切使得整个SQLite的实现变得简单，SQLite变得很小，运行速度也非常快，所以，特别适合嵌入式设备。好了，接下来讨论事务的剩余部分。

6、修改位于用户进程空间的页面(Changing Database Pages In User Space)

页面的原始数据写入日志之后，就可以修改页面了——位于用户进程空间。每个数据库连接都有自己私有的空间，所以页面的变化只对该连接可见，而对其它连接的数据仍然是磁盘缓存中的数据。从这里可以明白一件事：一个进程在修改页面数据的同时，其它进程可以继续进行读操作。图中的红色表示修改的页面。

7、日志文件刷入磁盘(Flushing The Rollback Journal File To Mass Storage)

接下来把日志文件的内容刷入磁盘，这对于数据库从意外中恢复来说是至关重要的一步。而且这通常也是一个耗时的操作，因为磁盘I/O速度很慢。 这个步骤不只把日志文件刷入磁盘那么简单，它的实现实际上分成两步：首先把日志文件的内容刷入磁盘（即页面数据）；然后把日志文件中页面的数目写入日志文件头，再把header刷入磁盘（这一过程在代码中清晰可见）。 代码如下：

/*
**Sync日志文件,保证所有的脏页面写入磁盘日志文件
*/
static int syncJournal(Pager *pPager){
  PgHdr *pPg;
  int rc = SQLITE_OK;

  /* Sync the journal before modifying the main database
  ** (assuming there is a journal and it needs to be synced.)
  */
  if( pPager->needSync ){
    if( !pPager->tempFile ){
      assert( pPager->journalOpen );
      /* assert( !pPager->noSync ); // noSync might be set if synchronous
      ** was turned off after the transaction was started.  Ticket #615 */
#ifndef NDEBUG
      {
        /* Make sure the pPager->nRec counter we are keeping agrees
        ** with the nRec computed from the size of the journal file.
        */
        i64 jSz;
        rc = sqlite3OsFileSize(pPager->jfd, &jSz);
        if( rc!=0 ) return rc;
        assert( pPager->journalOff==jSz );
      }
#endif
      {
        /* Write the nRec value into the journal file header. If in
        ** full-synchronous mode, sync the journal first. This ensures that
        ** all data has really hit the disk before nRec is updated to mark
        ** it as a candidate for rollback. 
        */
        if( pPager->fullSync ){
          TRACE2("SYNC journal of %d\n", PAGERID(pPager));
    //首先保证脏页面中所有的数据都已经写入日志文件
          rc = sqlite3OsSync(pPager->jfd, 0);
          if( rc!=0 ) return rc;
        }
        rc = sqlite3OsSeek(pPager->jfd,
                           pPager->journalHdr + sizeof(aJournalMagic));
        if( rc ) return rc;
    //页面的数目写入日志文件
        rc = write32bits(pPager->jfd, pPager->nRec);
        if( rc ) return rc;

        rc = sqlite3OsSeek(pPager->jfd, pPager->journalOff);
        if( rc ) return rc;
      }
      TRACE2("SYNC journal of %d\n", PAGERID(pPager));
      rc = sqlite3OsSync(pPager->jfd, pPager->full_fsync);
      if( rc!=0 ) return rc;
      pPager->journalStarted = 1;
    }
    pPager->needSync = 0;

    /* Erase the needSync flag from every page.
    */
    //清除needSync标志位
    for(pPg=pPager->pAll; pPg; pPg=pPg->pNextAll){
      pPg->needSync = 0;
    }
    pPager->pFirstSynced = pPager->pFirst;
  }

#ifndef NDEBUG
  /* If the Pager.needSync flag is clear then the PgHdr.needSync
  ** flag must also be clear for all pages.  Verify that this
  ** invariant is true.
  */
  else{
    for(pPg=pPager->pAll; pPg; pPg=pPg->pNextAll){
      assert( pPg->needSync==0 );
    }
    assert( pPager->pFirstSynced==pPager->pFirst );
  }
#endif

  return rc;
}

8、获取排斥锁（Obtaining An Exclusive Lock）

在对数据库文件进行修改之前(注：这里不是内存中的页面),我们必须得到数据库文件的排斥锁(Exclusive Lock)。得到排斥锁的过程可分为两步：首先得到Pending lock；然后Pending lock升级到exclusive lock。 Pending lock允许其它已经存在的Shared lock继续读数据库文件，但是不允许产生新的shared lock，这样做目的是为了防止写操作发生饿死情况。一旦所有的shared lock完成操作，则pending lock升级到exclusive lock。

9、修改的页面写入文件(Writing Changes To The Database File)

一旦得到exclusive lock，其它的进程就不能进行读操作，此时就可以把修改的页面写回数据库文件，但是通常OS都把结果暂时保存到磁盘缓存中，直到某个时刻才会真正把结果写入磁盘。

以上两步的实现代码：

//把所有的脏页面写入数据库
//到这里开始获取EXCLUSIVEQ锁，并将页面写回操作系统文件
static int pager_write_pagelist(PgHdr *pList){
  Pager *pPager;
  int rc;

  if( pList==0 ) return SQLITE_OK;
  pPager = pList->pPager;

  /* At this point there may be either a RESERVED or EXCLUSIVE lock on the
  ** database file. If there is already an EXCLUSIVE lock, the following
  ** calls to sqlite3OsLock() are no-ops.
  **
  ** Moving the lock from RESERVED to EXCLUSIVE actually involves going
  ** through an intermediate state PENDING.   A PENDING lock prevents new
  ** readers from attaching to the database but is unsufficient for us to
  ** write.  The idea of a PENDING lock is to prevent new readers from
  ** coming in while we wait for existing readers to clear.
  **
  ** While the pager is in the RESERVED state, the original database file
  ** is unchanged and we can rollback without having to playback the
  ** journal into the original database file.  Once we transition to
  ** EXCLUSIVE, it means the database file has been changed and any rollback
  ** will require a journal playback.
  */
  //加EXCLUSIVE_LOCK锁
  rc = pager_wait_on_lock(pPager, EXCLUSIVE_LOCK);
  if( rc!=SQLITE_OK ){
    return rc;
  }

  while( pList ){
    assert( pList->dirty );
    rc = sqlite3OsSeek(pPager->fd, (pList->pgno-1)*(i64)pPager->pageSize);
    if( rc ) return rc;
    /* If there are dirty pages in the page cache with page numbers greater
    ** than Pager.dbSize, this means sqlite3pager_truncate() was called to
    ** make the file smaller (presumably by auto-vacuum code). Do not write
    ** any such pages to the file.
    */
    if( pList->pgno<=pPager->dbSize ){
      char *pData = CODEC2(pPager, PGHDR_TO_DATA(pList), pList->pgno, 6);
      TRACE3("STORE %d page %d\n", PAGERID(pPager), pList->pgno);
      //写入文件
      rc = sqlite3OsWrite(pPager->fd, pData, pPager->pageSize);
      TEST_INCR(pPager->nWrite);
    }
#ifndef NDEBUG
    else{
      TRACE3("NOSTORE %d page %d\n", PAGERID(pPager), pList->pgno);
    }
#endif
    if( rc ) return rc;
    //设置dirty
    pList->dirty = 0;
#ifdef SQLITE_CHECK_PAGES
    pList->pageHash = pager_pagehash(pList);
#endif
//指向下一个脏页面
    pList = pList->pDirty;
  }
  return SQLITE_OK;
}

10、修改结果刷入存储设备(Flushing Changes To Mass Storage)

为了保证修改结果真正写入磁盘，这一步必不要少。对于数据库存的完整性，这一步也是关键的一步。由于要进行实际的I/O操作，所以和第7步一样，将花费较多的时间。

最后来看看这几步是如何实现的：

其实以上以上几步是在函数sqlite3BtreeSync()---btree.c中调用的(而关于该函数的调用后面再讲)。

代码如下：

//同步btree对应的数据库文件
//该函数返回之后，只需要提交写事务,删除日志文件
int sqlite3BtreeSync(Btree *p, const char *zMaster){
  int rc = SQLITE_OK;
  if( p->inTrans==TRANS_WRITE ){
    BtShared *pBt = p->pBt;
    Pgno nTrunc = 0;
#ifndef SQLITE_OMIT_AUTOVACUUM
    if( pBt->autoVacuum ){
      rc = autoVacuumCommit(pBt, &nTrunc); 
      if( rc!=SQLITE_OK ){
        return rc;
      }
    }
#endif

 //调用pager进行sync
    rc = sqlite3pager_sync(pBt->pPager, zMaster, nTrunc);
  }
  return rc;
}

//把pager所有脏页面写回文件
int sqlite3pager_sync(Pager *pPager, const char *zMaster, Pgno nTrunc){
  int rc = SQLITE_OK;

  TRACE4("DATABASE SYNC: File=%s zMaster=%s nTrunc=%d\n", 
      pPager->zFilename, zMaster, nTrunc);

  /* If this is an in-memory db, or no pages have been written to, or this
  ** function has already been called, it is a no-op.
  */
  //pager不处于PAGER_SYNCED状态,dirtyCache为1,
  //则进行sync操作
  if( pPager->state!=PAGER_SYNCED && !MEMDB && pPager->dirtyCache ){
    PgHdr *pPg;
    assert( pPager->journalOpen );

    /* If a master journal file name has already been written to the
    ** journal file, then no sync is required. This happens when it is
    ** written, then the process fails to upgrade from a RESERVED to an
    ** EXCLUSIVE lock. The next time the process tries to commit the
    ** transaction the m-j name will have already been written.
    */
    if( !pPager->setMaster ){
        //pager修改计数
      rc = pager_incr_changecounter(pPager);
      if( rc!=SQLITE_OK ) goto sync_exit;
#ifndef SQLITE_OMIT_AUTOVACUUM
      if( nTrunc!=0 ){
        /* If this transaction has made the database smaller, then all pages
        ** being discarded by the truncation must be written to the journal
        ** file.
        */
        Pgno i;
        void *pPage;
        int iSkip = PAGER_MJ_PGNO(pPager);
        for( i=nTrunc+1; i<=pPager->origDbSize; i++ ){
          if( !(pPager->aInJournal[i/8] & (1<<(i&7))) && i!=iSkip ){
            rc = sqlite3pager_get(pPager, i, &pPage);
            if( rc!=SQLITE_OK ) goto sync_exit;
            rc = sqlite3pager_write(pPage);
            sqlite3pager_unref(pPage);
            if( rc!=SQLITE_OK ) goto sync_exit;
          }
        } 
      }
#endif
      rc = writeMasterJournal(pPager, zMaster);
      if( rc!=SQLITE_OK ) goto sync_exit;

        //sync日志文件
      rc = syncJournal(pPager);
      if( rc!=SQLITE_OK ) goto sync_exit;
    }

#ifndef SQLITE_OMIT_AUTOVACUUM
    if( nTrunc!=0 ){
      rc = sqlite3pager_truncate(pPager, nTrunc);
      if( rc!=SQLITE_OK ) goto sync_exit;
    }
#endif

    /* Write all dirty pages to the database file */
    pPg = pager_get_all_dirty_pages(pPager);


   //把所有脏页面写回操作系统文件
    rc = pager_write_pagelist(pPg);
    if( rc!=SQLITE_OK ) goto sync_exit;

    /* Sync the database file. */
       //sync数据库文件
    if( !pPager->noSync ){
      rc = sqlite3OsSync(pPager->fd, 0);
    }

    pPager->state = PAGER_SYNCED;
  }else if( MEMDB && nTrunc!=0 ){
    rc = sqlite3pager_truncate(pPager, nTrunc);
  }

sync_exit:
  return rc;
}

下图可以进一步解释该过程：