pager.c

sqlite-source-2_8_16.zip 轻量级嵌入式数据库源码

** The Pager keeps a separate list of pages that are currently in
** the checkpoint journal.  This helps the sqlitepager_ckpt_commit()
** routine run MUCH faster for the common case where there are many
** pages in memory but only a few are in the checkpoint journal.
*/
static void page_add_to_ckpt_list(PgHdr *pPg){
  Pager *pPager = pPg->pPager;
  if( pPg->inCkpt ) return;
  assert( pPg->pPrevCkpt==0 && pPg->pNextCkpt==0 );
  pPg->pPrevCkpt = 0;
  if( pPager->pCkpt ){
    pPager->pCkpt->pPrevCkpt = pPg;
  }
  pPg->pNextCkpt = pPager->pCkpt;
  pPager->pCkpt = pPg;
  pPg->inCkpt = 1;
}
static void page_remove_from_ckpt_list(PgHdr *pPg){
  if( !pPg->inCkpt ) return;
  if( pPg->pPrevCkpt ){
    assert( pPg->pPrevCkpt->pNextCkpt==pPg );
    pPg->pPrevCkpt->pNextCkpt = pPg->pNextCkpt;
  }else{
    assert( pPg->pPager->pCkpt==pPg );
    pPg->pPager->pCkpt = pPg->pNextCkpt;
  }
  if( pPg->pNextCkpt ){
    assert( pPg->pNextCkpt->pPrevCkpt==pPg );
    pPg->pNextCkpt->pPrevCkpt = pPg->pPrevCkpt;
  }
  pPg->pNextCkpt = 0;
  pPg->pPrevCkpt = 0;
  pPg->inCkpt = 0;
}

/*
** Find a page in the hash table given its page number.  Return
** a pointer to the page or NULL if not found.
*/
static PgHdr *pager_lookup(Pager *pPager, Pgno pgno){
  PgHdr *p = pPager->aHash[pager_hash(pgno)];
  while( p && p->pgno!=pgno ){
    p = p->pNextHash;
  }
  return p;
}

/*
** Unlock the database and clear the in-memory cache.  This routine
** sets the state of the pager back to what it was when it was first
** opened.  Any outstanding pages are invalidated and subsequent attempts
** to access those pages will likely result in a coredump.
*/
static void pager_reset(Pager *pPager){
  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.
**
** TODO: Consider keeping the journal file open for temporary databases.
** This might give a performance improvement on windows where opening
** a file is an expensive operation.
*/
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.
**
** This is not a real checksum.  It is really just the sum of the 
** random initial value and the page number.  We considered do a checksum
** of the database, but that was found to be too slow.
*/
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 ){
    /* No page should ever be rolled back that is in use, except for page
    ** 1 which is held in use in order to keep the lock on the database
    ** active.
    */
    assert( pPg->nRef==0 || pPg->pgno==1 );
    memcpy(PGHDR_TO_DATA(pPg), pgRec.aData, SQLITE_PAGE_SIZE);
    memset(PGHDR_TO_EXTRA(pPg), 0, pPager->nExtra);
    pPg->dirty = 0;
    pPg->needSync = 0;
    CODEC(pPager, PGHDR_TO_DATA(pPg), pPg->pgno, 3);
  }
  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: 
**
**    *  8 byte prefix.  One of the aJournalMagic123 vectors defined
**       above.  The format of the journal file is determined by which
**       of the three prefix vectors is seen.
**    *  4 byte big-endian integer which is the number of valid page records
**       in the journal.  If this value is 0xffffffff, then compute the
**       number of page records from the journal size.  This field appears
**       in format 3 only.
**    *  4 byte big-endian integer which is the initial value for the 
**       sanity checksum.  This field appears in format 3 only.
**    *  4 byte integer which is the number of pages to truncate the
**       database to during a rollback.
**    *  Zero or more pages instances, each as follows:
**        +  4 byte page number.
**        +  SQLITE_PAGE_SIZE bytes of data.
**        +  4 byte checksum (format 3 only)
**
** When we speak of the journal header, we mean the first 4 bullets above.
** Each entry in the journal is an instance of the 5th bullet.  Note that
** bullets 2 and 3 only appear in format-3 journals.
**
** Call the value from the second bullet "nRec".  nRec is the number of
** valid page entries in the journal.  In most cases, you can compute the
** value of nRec from the size of the journal file.  But if a power
** failure occurred while the journal was being written, it could be the
** case that the size of the journal file had already been increased but
** the extra entries had not yet made it safely to disk.  In such a case,
** the value of nRec computed from the file size would be too large.  For
** that reason, we always use the nRec value in the header.
**
** If the nRec value is 0xffffffff it means that nRec should be computed
** from the file size.  This value is used when the user selects the
** no-sync option for the journal.  A power failure could lead to corruption
** in this case.  But for things like temporary table (which will be
** deleted when the power is restored) we don't care.  
**
** Journal formats 1 and 2 do not have an nRec value in the header so we
** have to compute nRec from the file size.  This has risks (as described
** above) which is why all persistent tables have been changed to use
** format 3.
**
** If the file opened as the journal file is not a well-formed
** journal file then the database will likely already 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 the journal file is too small to contain a complete header,
  ** it must mean that the process that created the journal was just
  ** beginning to write the journal file when it died.  In that case,
  ** the database file should have still been completely unchanged.
  ** Nothing needs to be rolled back.  We can safely ignore this journal.
  */
  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 ){
    if( szJ < sizeof(aMagic) + 3*sizeof(u32) ){
      /* Ignore the journal if it is too small to contain a complete
      ** header.  We already did this test once above, but at the prior
      ** test, we did not know the journal format and so we had to assume
      ** the smallest possible header.  Now we know the header is bigger
      ** than the minimum so we test again.
      */
      goto end_playback;
    }
    rc = read32bits(format, &pPager->jfd, (u32*)&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 them 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);
        TRACE2("REFETCH %d\n", pPg->pgno);
        CODEC(pPager, zBuf, pPg->pgno, 2);
        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.
*/