pager.c

来自「SQLite 2.8.6 源代码,用来在Linux/Unix/Windows上编」· C语言 代码 · 共 2,008 行 · 第 1/5 页

  PgHdr *pPg, *pNext;
  for(pPg=pPager->pAll; pPg; pPg=pNext){
    pNext = pPg->pNextAll;
    sqliteFree(pPg);
  }
  pPager->pFirst = 0;
  pPager->pFirstSynced = 0;
  pPager->pLast = 0;
  pPager->pAll = 0;
  memset(pPager->aHash, 0, sizeof(pPager->aHash));
  pPager->nPage = 0;
  if( pPager->state>=SQLITE_WRITELOCK ){
    sqlitepager_rollback(pPager);
  }
  sqliteOsUnlock(&pPager->fd);
  pPager->state = SQLITE_UNLOCK;
  pPager->dbSize = -1;
  pPager->nRef = 0;
  assert( pPager->journalOpen==0 );
}

/*
** When this routine is called, the pager has the journal file open and
** a write lock on the database.  This routine releases the database
** write lock and acquires a read lock in its place.  The journal file
** is deleted and closed.
*/
static int pager_unwritelock(Pager *pPager){
  int rc;
  PgHdr *pPg;
  if( pPager->state<SQLITE_WRITELOCK ) return SQLITE_OK;
  sqlitepager_ckpt_commit(pPager);
  if( pPager->ckptOpen ){
    sqliteOsClose(&pPager->cpfd);
    pPager->ckptOpen = 0;
  }
  if( pPager->journalOpen ){
    sqliteOsClose(&pPager->jfd);
    pPager->journalOpen = 0;
    sqliteOsDelete(pPager->zJournal);
    sqliteFree( pPager->aInJournal );
    pPager->aInJournal = 0;
    for(pPg=pPager->pAll; pPg; pPg=pPg->pNextAll){
      pPg->inJournal = 0;
      pPg->dirty = 0;
      pPg->needSync = 0;
    }
  }else{
    assert( pPager->dirtyFile==0 || pPager->useJournal==0 );
  }
  rc = sqliteOsReadLock(&pPager->fd);
  if( rc==SQLITE_OK ){
    pPager->state = SQLITE_READLOCK;
  }else{
    /* This can only happen if a process does a BEGIN, then forks and the
    ** child process does the COMMIT.  Because of the semantics of unix
    ** file locking, the unlock will fail.
    */
    pPager->state = SQLITE_UNLOCK;
  }
  return rc;
}

/*
** Compute and return a checksum for the page of data.
*/
static u32 pager_cksum(Pager *pPager, Pgno pgno, const char *aData){
  u32 cksum = pPager->cksumInit + pgno;
  return cksum;
}

/*
** Read a single page from the journal file opened on file descriptor
** jfd.  Playback this one page.
**
** There are three different journal formats.  The format parameter determines
** which format is used by the journal that is played back.
*/
static int pager_playback_one_page(Pager *pPager, OsFile *jfd, int format){
  int rc;
  PgHdr *pPg;              /* An existing page in the cache */
  PageRecord pgRec;
  u32 cksum;

  rc = read32bits(format, jfd, &pgRec.pgno);
  if( rc!=SQLITE_OK ) return rc;
  rc = sqliteOsRead(jfd, &pgRec.aData, sizeof(pgRec.aData));
  if( rc!=SQLITE_OK ) return rc;

  /* Sanity checking on the page.  This is more important that I originally
  ** thought.  If a power failure occurs while the journal is being written,
  ** it could cause invalid data to be written into the journal.  We need to
  ** detect this invalid data (with high probability) and ignore it.
  */
  if( pgRec.pgno==0 ){
    return SQLITE_DONE;
  }
  if( pgRec.pgno>(unsigned)pPager->dbSize ){
    return SQLITE_OK;
  }
  if( format>=JOURNAL_FORMAT_3 ){
    rc = read32bits(format, jfd, &cksum);
    if( rc ) return rc;
    if( pager_cksum(pPager, pgRec.pgno, pgRec.aData)!=cksum ){
      return SQLITE_DONE;
    }
  }

  /* Playback the page.  Update the in-memory copy of the page
  ** at the same time, if there is one.
  */
  pPg = pager_lookup(pPager, pgRec.pgno);
  TRACE2("PLAYBACK %d\n", pgRec.pgno);
  sqliteOsSeek(&pPager->fd, (pgRec.pgno-1)*(off_t)SQLITE_PAGE_SIZE);
  rc = sqliteOsWrite(&pPager->fd, pgRec.aData, SQLITE_PAGE_SIZE);
  if( pPg ){
    if( pPg->nRef==0 ||
        memcmp(PGHDR_TO_DATA(pPg), pgRec.aData, SQLITE_PAGE_SIZE)==0
    ){
      /* Do not update the data on this page if the page is in use
      ** and the page has never been modified.  This avoids resetting
      ** the "extra" data.  That in turn avoids invalidating BTree cursors
      ** in trees that have never been modified.  The end result is that
      ** you can have a SELECT going on in one table and ROLLBACK changes
      ** to a different table and the SELECT is unaffected by the ROLLBACK.
      */
      memcpy(PGHDR_TO_DATA(pPg), pgRec.aData, SQLITE_PAGE_SIZE);
      memset(PGHDR_TO_EXTRA(pPg), 0, pPager->nExtra);
    }
    pPg->dirty = 0;
    pPg->needSync = 0;
  }
  return rc;
}

/*
** Playback the journal and thus restore the database file to
** the state it was in before we started making changes.  
**
** The journal file format is as follows:  There is an initial
** file-type string for sanity checking.  Then there is a single
** Pgno number which is the number of pages in the database before
** changes were made.  The database is truncated to this size.
** Next come zero or more page records where each page record
** consists of a Pgno and SQLITE_PAGE_SIZE bytes of data.  See
** the PageRecord structure for details.
**
** If the file opened as the journal file is not a well-formed
** journal file (as determined by looking at the magic number
** at the beginning) then this routine returns SQLITE_PROTOCOL.
** If any other errors occur during playback, the database will
** likely be corrupted, so the PAGER_ERR_CORRUPT bit is set in
** pPager->errMask and SQLITE_CORRUPT is returned.  If it all
** works, then this routine returns SQLITE_OK.
*/
static int pager_playback(Pager *pPager, int useJournalSize){
  off_t szJ;               /* Size of the journal file in bytes */
  int nRec;                /* Number of Records in the journal */
  int i;                   /* Loop counter */
  Pgno mxPg = 0;           /* Size of the original file in pages */
  int format;              /* Format of the journal file. */
  unsigned char aMagic[sizeof(aJournalMagic1)];
  int rc;

  /* Figure out how many records are in the journal.  Abort early if
  ** the journal is empty.
  */
  assert( pPager->journalOpen );
  sqliteOsSeek(&pPager->jfd, 0);
  rc = sqliteOsFileSize(&pPager->jfd, &szJ);
  if( rc!=SQLITE_OK ){
    goto end_playback;
  }
  if( szJ < sizeof(aMagic)+sizeof(Pgno) ){
    goto end_playback;
  }

  /* Read the beginning of the journal and truncate the
  ** database file back to its original size.
  */
  rc = sqliteOsRead(&pPager->jfd, aMagic, sizeof(aMagic));
  if( rc!=SQLITE_OK ){
    rc = SQLITE_PROTOCOL;
    goto end_playback;
  }
  if( memcmp(aMagic, aJournalMagic3, sizeof(aMagic))==0 ){
    format = JOURNAL_FORMAT_3;
  }else if( memcmp(aMagic, aJournalMagic2, sizeof(aMagic))==0 ){
    format = JOURNAL_FORMAT_2;
  }else if( memcmp(aMagic, aJournalMagic1, sizeof(aMagic))==0 ){
    format = JOURNAL_FORMAT_1;
  }else{
    rc = SQLITE_PROTOCOL;
    goto end_playback;
  }
  if( format>=JOURNAL_FORMAT_3 ){
    rc = read32bits(format, &pPager->jfd, &nRec);
    if( rc ) goto end_playback;
    rc = read32bits(format, &pPager->jfd, &pPager->cksumInit);
    if( rc ) goto end_playback;
    if( nRec==0xffffffff || useJournalSize ){
      nRec = (szJ - JOURNAL_HDR_SZ(3))/JOURNAL_PG_SZ(3);
    }
  }else{
    nRec = (szJ - JOURNAL_HDR_SZ(2))/JOURNAL_PG_SZ(2);
    assert( nRec*JOURNAL_PG_SZ(2)+JOURNAL_HDR_SZ(2)==szJ );
  }
  rc = read32bits(format, &pPager->jfd, &mxPg);
  if( rc!=SQLITE_OK ){
    goto end_playback;
  }
  assert( pPager->origDbSize==0 || pPager->origDbSize==mxPg );
  rc = sqliteOsTruncate(&pPager->fd, SQLITE_PAGE_SIZE*(off_t)mxPg);
  if( rc!=SQLITE_OK ){
    goto end_playback;
  }
  pPager->dbSize = mxPg;
  
  /* Copy original pages out of the journal and back into the database file.
  */
  for(i=0; i<nRec; i++){
    rc = pager_playback_one_page(pPager, &pPager->jfd, format);
    if( rc!=SQLITE_OK ){
      if( rc==SQLITE_DONE ){
        rc = SQLITE_OK;
      }
      break;
    }
  }

  /* Pages that have been written to the journal but never synced
  ** where not restored by the loop above.  We have to restore those
  ** pages by reading the back from the original database.
  */
  if( rc==SQLITE_OK ){
    PgHdr *pPg;
    for(pPg=pPager->pAll; pPg; pPg=pPg->pNextAll){
      char zBuf[SQLITE_PAGE_SIZE];
      if( !pPg->dirty ) continue;
      if( (int)pPg->pgno <= pPager->origDbSize ){
        sqliteOsSeek(&pPager->fd, SQLITE_PAGE_SIZE*(off_t)(pPg->pgno-1));
        rc = sqliteOsRead(&pPager->fd, zBuf, SQLITE_PAGE_SIZE);
        if( rc ) break;
      }else{
        memset(zBuf, 0, SQLITE_PAGE_SIZE);
      }
      if( pPg->nRef==0 || memcmp(zBuf, PGHDR_TO_DATA(pPg), SQLITE_PAGE_SIZE) ){
        memcpy(PGHDR_TO_DATA(pPg), zBuf, SQLITE_PAGE_SIZE);
        memset(PGHDR_TO_EXTRA(pPg), 0, pPager->nExtra);
      }
      pPg->needSync = 0;
      pPg->dirty = 0;
    }
  }

end_playback:
  if( rc!=SQLITE_OK ){
    pager_unwritelock(pPager);
    pPager->errMask |= PAGER_ERR_CORRUPT;
    rc = SQLITE_CORRUPT;
  }else{
    rc = pager_unwritelock(pPager);
  }
  return rc;
}

/*
** Playback the checkpoint journal.
**
** This is similar to playing back the transaction journal but with
** a few extra twists.
**
**    (1)  The number of pages in the database file at the start of
**         the checkpoint is stored in pPager->ckptSize, not in the
**         journal file itself.
**
**    (2)  In addition to playing back the checkpoint journal, also
**         playback all pages of the transaction journal beginning
**         at offset pPager->ckptJSize.
*/
static int pager_ckpt_playback(Pager *pPager){
  off_t szJ;               /* Size of the full journal */
  int nRec;                /* Number of Records */
  int i;                   /* Loop counter */
  int rc;

  /* Truncate the database back to its original size.
  */
  rc = sqliteOsTruncate(&pPager->fd, SQLITE_PAGE_SIZE*(off_t)pPager->ckptSize);
  pPager->dbSize = pPager->ckptSize;

  /* Figure out how many records are in the checkpoint journal.
  */
  assert( pPager->ckptInUse && pPager->journalOpen );
  sqliteOsSeek(&pPager->cpfd, 0);
  nRec = pPager->ckptNRec;
  
  /* Copy original pages out of the checkpoint journal and back into the
  ** database file.  Note that the checkpoint journal always uses format
  ** 2 instead of format 3 since it does not need to be concerned with
  ** power failures corrupting the journal and can thus omit the checksums.
  */
  for(i=nRec-1; i>=0; i--){
    rc = pager_playback_one_page(pPager, &pPager->cpfd, 2);
    assert( rc!=SQLITE_DONE );
    if( rc!=SQLITE_OK ) goto end_ckpt_playback;
  }

  /* Figure out how many pages need to be copied out of the transaction
  ** journal.
  */
  rc = sqliteOsSeek(&pPager->jfd, pPager->ckptJSize);
  if( rc!=SQLITE_OK ){
    goto end_ckpt_playback;
  }
  rc = sqliteOsFileSize(&pPager->jfd, &szJ);
  if( rc!=SQLITE_OK ){
    goto end_ckpt_playback;
  }
  nRec = (szJ - pPager->ckptJSize)/JOURNAL_PG_SZ(journal_format);
  for(i=nRec-1; i>=0; i--){
    rc = pager_playback_one_page(pPager, &pPager->jfd, journal_format);
    if( rc!=SQLITE_OK ){
      assert( rc!=SQLITE_DONE );
      goto end_ckpt_playback;
    }
  }
  
end_ckpt_playback:
  if( rc!=SQLITE_OK ){
    pPager->errMask |= PAGER_ERR_CORRUPT;
    rc = SQLITE_CORRUPT;
  }
  return rc;
}

/*
** Change the maximum number of in-memory pages that are allowed.
**
** The maximum number is the absolute value of the mxPage parameter.
** If mxPage is negative, the noSync flag is also set.  noSync bypasses
** calls to sqliteOsSync().  The pager runs much faster with noSync on,
** but if the operating system crashes or there is an abrupt power 
** failure, the database file might be left in an inconsistent and
** unrepairable state.  
*/
void sqlitepager_set_cachesize(Pager *pPager, int mxPage){
  if( mxPage>=0 ){
    pPager->noSync = pPager->tempFile;
  }else{
    pPager->noSync = 1;
    mxPage = -mxPage;
  }
  if( mxPage>10 ){
    pPager->mxPage = mxPage;
  }
}

/*
** Adjust the robustness of the database to damage due to OS crashes
** or power failures by changing the number of syncs()s when writing
** the rollback journal.  There are three levels:
**
**    OFF       sqliteOsSync() is never called.  This is the default
**              for temporary and transient files.
**
**    NORMAL    The journal is synced once before writes begin on the
**              database.  This is normally adequate protection, but
**              it is theoretically possible, though very unlikely,
**              that an inopertune power failure could leave the journal
**              in a state which would cause damage to the database
**              when it is rolled back.
**
**    FULL      The journal is synced twice before writes begin on the
**              database (with some additional information being written
**              in between the two syncs.  If we assume that writing a
**              single disk sector is atomic, then this mode provides
**              assurance that the journal will not be corrupted to the
**              point of causing damage to the database during rollback.
**
** Numeric values associated with these states are OFF==1, NORMAL=2,
** and FULL=3.
*/
void sqlitepager_set_safety_level(Pager *pPager, int level){
  pPager->noSync =  level==1 || pPager->tempFile;
  pPager->fullSync = level==3 && !pPager->tempFile;
}

/*
** Open a temporary file.  Write the name of the file into zName
** (zName must be at least SQLITE_TEMPNAME_SIZE bytes long.)  Write
** the file descriptor into *fd.  Return SQLITE_OK on success or some
** other error code if we fail.
**
** The OS will automatically delete the temporary file when it is
** closed.
*/
static int sqlitepager_opentemp(char *zFile, OsFile *fd){
  int cnt = 8;
  int rc;
  do{
    cnt--;