pager.c

最新的sqlite3.6.2源代码

  u32 *pDbSize
){
  int rc;
  unsigned char aMagic[8]; /* A buffer to hold the magic header */
  i64 jrnlOff;
  int iPageSize;

  seekJournalHdr(pPager);
  if( pPager->journalOff+JOURNAL_HDR_SZ(pPager) > journalSize ){
    return SQLITE_DONE;
  }
  jrnlOff = pPager->journalOff;

  rc = sqlite3OsRead(pPager->jfd, aMagic, sizeof(aMagic), jrnlOff);
  if( rc ) return rc;
  jrnlOff += sizeof(aMagic);

  if( memcmp(aMagic, aJournalMagic, sizeof(aMagic))!=0 ){
    return SQLITE_DONE;
  }

  rc = read32bits(pPager->jfd, jrnlOff, pNRec);
  if( rc ) return rc;

  rc = read32bits(pPager->jfd, jrnlOff+4, &pPager->cksumInit);
  if( rc ) return rc;

  rc = read32bits(pPager->jfd, jrnlOff+8, pDbSize);
  if( rc ) return rc;

  rc = read32bits(pPager->jfd, jrnlOff+16, (u32 *)&iPageSize);
  if( rc==SQLITE_OK 
   && iPageSize>=512 
   && iPageSize<=SQLITE_MAX_PAGE_SIZE 
   && ((iPageSize-1)&iPageSize)==0 
  ){
    u16 pagesize = iPageSize;
    rc = sqlite3PagerSetPagesize(pPager, &pagesize);
  }
  if( rc ) return rc;

  /* Update the assumed sector-size to match the value used by 
  ** the process that created this journal. If this journal was
  ** created by a process other than this one, then this routine
  ** is being called from within pager_playback(). The local value
  ** of Pager.sectorSize is restored at the end of that routine.
  */
  rc = read32bits(pPager->jfd, jrnlOff+12, (u32 *)&pPager->sectorSize);
  if( rc ) return rc;

  pPager->journalOff += JOURNAL_HDR_SZ(pPager);
  return SQLITE_OK;
}


/*
** Write the supplied master journal name into the journal file for pager
** pPager at the current location. The master journal name must be the last
** thing written to a journal file. If the pager is in full-sync mode, the
** journal file descriptor is advanced to the next sector boundary before
** anything is written. The format is:
**
** + 4 bytes: PAGER_MJ_PGNO.
** + N bytes: length of master journal name.
** + 4 bytes: N
** + 4 bytes: Master journal name checksum.
** + 8 bytes: aJournalMagic[].
**
** The master journal page checksum is the sum of the bytes in the master
** journal name.
**
** If zMaster is a NULL pointer (occurs for a single database transaction), 
** this call is a no-op.
*/
static int writeMasterJournal(Pager *pPager, const char *zMaster){
  int rc;
  int len; 
  int i; 
  i64 jrnlOff;
  i64 jrnlSize;
  u32 cksum = 0;
  char zBuf[sizeof(aJournalMagic)+2*4];

  if( !zMaster || pPager->setMaster) return SQLITE_OK;
  pPager->setMaster = 1;

  len = strlen(zMaster);
  for(i=0; i<len; i++){
    cksum += zMaster[i];
  }

  /* If in full-sync mode, advance to the next disk sector before writing
  ** the master journal name. This is in case the previous page written to
  ** the journal has already been synced.
  */
  if( pPager->fullSync ){
    seekJournalHdr(pPager);
  }
  jrnlOff = pPager->journalOff;
  pPager->journalOff += (len+20);

  rc = write32bits(pPager->jfd, jrnlOff, PAGER_MJ_PGNO(pPager));
  if( rc!=SQLITE_OK ) return rc;
  jrnlOff += 4;

  rc = sqlite3OsWrite(pPager->jfd, zMaster, len, jrnlOff);
  if( rc!=SQLITE_OK ) return rc;
  jrnlOff += len;

  put32bits(zBuf, len);
  put32bits(&zBuf[4], cksum);
  memcpy(&zBuf[8], aJournalMagic, sizeof(aJournalMagic));
  rc = sqlite3OsWrite(pPager->jfd, zBuf, 8+sizeof(aJournalMagic), jrnlOff);
  jrnlOff += 8+sizeof(aJournalMagic);
  pPager->needSync = !pPager->noSync;

  /* If the pager is in peristent-journal mode, then the physical 
  ** journal-file may extend past the end of the master-journal name
  ** and 8 bytes of magic data just written to the file. This is 
  ** dangerous because the code to rollback a hot-journal file
  ** will not be able to find the master-journal name to determine 
  ** whether or not the journal is hot. 
  **
  ** Easiest thing to do in this scenario is to truncate the journal 
  ** file to the required size.
  */ 
  if( (rc==SQLITE_OK)
   && (rc = sqlite3OsFileSize(pPager->jfd, &jrnlSize))==SQLITE_OK
   && jrnlSize>jrnlOff
  ){
    rc = sqlite3OsTruncate(pPager->jfd, jrnlOff);
  }
  return rc;
}

/*
** 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;
  sqlite3PcacheFetch(pPager->pPCache, pgno, 0, &p);
  return p;
}

/*
** 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){
  if( pPager->errCode ) return;
  sqlite3PcacheClear(pPager->pPCache);
}

/*
** Unlock the database file. 
**
** If the pager is currently in error state, discard the contents of 
** the cache and reset the Pager structure internal state. If there is
** an open journal-file, then the next time a shared-lock is obtained
** on the pager file (by this or any other process), it will be
** treated as a hot-journal and rolled back.
*/
static void pager_unlock(Pager *pPager){
  if( !pPager->exclusiveMode ){
    if( !MEMDB ){
      int rc = osUnlock(pPager->fd, NO_LOCK);
      if( rc ) pPager->errCode = rc;
      pPager->dbSize = -1;
      IOTRACE(("UNLOCK %p\n", pPager))

      /* Always close the journal file when dropping the database lock.
      ** Otherwise, another connection with journal_mode=delete might
      ** delete the file out from under us.
      */
      if( pPager->journalOpen ){
        sqlite3OsClose(pPager->jfd);
        pPager->journalOpen = 0;
        sqlite3BitvecDestroy(pPager->pInJournal);
        pPager->pInJournal = 0;
        sqlite3BitvecDestroy(pPager->pAlwaysRollback);
        pPager->pAlwaysRollback = 0;
      }

      /* If Pager.errCode is set, the contents of the pager cache cannot be
      ** trusted. Now that the pager file is unlocked, the contents of the
      ** cache can be discarded and the error code safely cleared.
      */
      if( pPager->errCode ){
        if( rc==SQLITE_OK ) pPager->errCode = SQLITE_OK;
        pager_reset(pPager);
        if( pPager->stmtOpen ){
          sqlite3OsClose(pPager->stfd);
          sqlite3BitvecDestroy(pPager->pInStmt);
          pPager->pInStmt = 0;
        }
        pPager->stmtOpen = 0;
        pPager->stmtInUse = 0;
        pPager->journalOff = 0;
        pPager->journalStarted = 0;
        pPager->stmtAutoopen = 0;
        pPager->origDbSize = 0;
      }
    }

    if( !MEMDB || pPager->errCode==SQLITE_OK ){
      pPager->state = PAGER_UNLOCK;
      pPager->changeCountDone = 0;
    }
  }
}

/*
** Execute a rollback if a transaction is active and unlock the 
** database file. If the pager has already entered the error state, 
** do not attempt the rollback.
*/
static void pagerUnlockAndRollback(Pager *p){
  if( p->errCode==SQLITE_OK && p->state>=PAGER_RESERVED ){
    sqlite3BeginBenignMalloc();
    sqlite3PagerRollback(p);
    sqlite3EndBenignMalloc();
  }
  pager_unlock(p);
}

/*
** This routine ends a transaction.  A transaction is ended by either
** a COMMIT or a ROLLBACK.
**
** When this routine is called, the pager has the journal file open and
** a RESERVED or EXCLUSIVE lock on the database.  This routine will release
** the database lock and acquires a SHARED lock in its place if that is
** the appropriate thing to do.  Release locks usually is appropriate,
** unless we are in exclusive access mode or unless this is a 
** COMMIT AND BEGIN or ROLLBACK AND BEGIN operation.
**
** The journal file is either deleted or truncated.
**
** 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_end_transaction(Pager *pPager, int hasMaster){
  int rc = SQLITE_OK;
  int rc2 = SQLITE_OK;
  assert( !MEMDB );
  if( pPager->state<PAGER_RESERVED ){
    return SQLITE_OK;
  }
  sqlite3PagerStmtCommit(pPager);
  if( pPager->stmtOpen && !pPager->exclusiveMode ){
    sqlite3OsClose(pPager->stfd);
    pPager->stmtOpen = 0;
  }
  if( pPager->journalOpen ){
    if( pPager->exclusiveMode 
     || pPager->journalMode==PAGER_JOURNALMODE_PERSIST
    ){
      rc = zeroJournalHdr(pPager, hasMaster);
      pager_error(pPager, rc);
      pPager->journalOff = 0;
      pPager->journalStarted = 0;
    }else{
      sqlite3OsClose(pPager->jfd);
      pPager->journalOpen = 0;
      if( rc==SQLITE_OK && !pPager->tempFile ){
        rc = sqlite3OsDelete(pPager->pVfs, pPager->zJournal, 0);
      }
    }
    sqlite3BitvecDestroy(pPager->pInJournal);
    pPager->pInJournal = 0;
    sqlite3BitvecDestroy(pPager->pAlwaysRollback);
    pPager->pAlwaysRollback = 0;
    sqlite3PcacheCleanAll(pPager->pPCache);
#ifdef SQLITE_CHECK_PAGES
    sqlite3PcacheIterate(pPager->pPCache, pager_set_pagehash);
#endif
    sqlite3PcacheSetFlags(pPager->pPCache,
       ~(PGHDR_IN_JOURNAL | PGHDR_NEED_SYNC), 0
    );
    pPager->dirtyCache = 0;
    pPager->nRec = 0;
  }else{
    assert( pPager->pInJournal==0 );
  }

  if( !pPager->exclusiveMode ){
    rc2 = osUnlock(pPager->fd, SHARED_LOCK);
    pPager->state = PAGER_SHARED;
  }else if( pPager->state==PAGER_SYNCED ){
    pPager->state = PAGER_EXCLUSIVE;
  }
  pPager->origDbSize = 0;
  pPager->setMaster = 0;
  pPager->needSync = 0;
  /* lruListSetFirstSynced(pPager); */
  pPager->dbSize = -1;
  pPager->dbModified = 0;

  return (rc==SQLITE_OK?rc2: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 experimented with
** a checksum of the entire data, but that was found to be too slow.
**
** Note that the page number is stored at the beginning of data and
** the checksum is stored at the end.  This is important.  If journal
** corruption occurs due to a power failure, the most likely scenario
** is that one end or the other of the record will be changed.  It is
** much less likely that the two ends of the journal record will be
** correct and the middle be corrupt.  Thus, this "checksum" scheme,
** though fast and simple, catches the mostly likely kind of corruption.
**
** FIX ME:  Consider adding every 200th (or so) byte of the data to the
** checksum.  That way if a single page spans 3 or more disk sectors and
** only the middle sector is corrupt, we will still have a reasonable
** chance of failing the checksum and thus detecting the problem.
*/
static u32 pager_cksum(Pager *pPager, const u8 *aData){
  u32 cksum = pPager->cksumInit;
  int i = pPager->pageSize-200;
  while( i>0 ){
    cksum += aData[i];
    i -= 200;
  }
  return cksum;
}

/* Forward declaration */
static void makeClean(PgHdr*);

/*
** Read a single page from the journal file opened on file descriptor
** jfd.  Playback this one page.
**
** The isMainJrnl flag is true if this is the main rollback journal and
** false for the statement journal.  The main rollback journal uses
** checksums - the statement journal does not.
*/
static int pager_playback_one_page(
  Pager *pPager,       /* The pager being played back */
  sqlite3_file *jfd,   /* The file that is the journal being rolled back */
  i64 offset,          /* Offset of the page within the journal */
  int isMainJrnl       /* True for main rollback journal. False for Stmt jrnl */
){
  int rc;
  PgHdr *pPg;                   /* An existing page in the cache */
  Pgno pgno;                    /* The page number of a page in journal */
  u32 cksum;                    /* Checksum used for sanity checking */
  u8 *aData = (u8 *)pPager->pTmpSpace;   /* Temp storage for a page */