SQLite4  Check-in [ae3c8da44d]

** This function is a no-op if *pRc is non-zero when it is called.
**
** Open the LSM database identified by zFile and compute its checksum
** (a string, as returned by testCksumDatabase()). If the checksum is
** identical to zExpect1 or, if it is not NULL, zExpect2, the test passes.
** Otherwise, print an error message and set *pRc to 1.
*/
static void testCompareCksumLsmdb(
  const char *zFile,              /* Path to LSM database */
  int bCompress,                  /* True if db is compressed */
  const char *zExpect1,           /* Expected checksum 1 */
  const char *zExpect2,           /* Expected checksum 2 (or NULL) */
  int *pRc                        /* IN/OUT: Test case error code */
){
  if( *pRc==0 ){
    char zCksum[TEST_CKSUM_BYTES];
    TestDb *pDb;

    *pRc = tdb_lsm_open((bCompress?"compression=1 mmap=0":""), zFile, 0, &pDb);
    testCksumDatabase(pDb, zCksum);
    testClose(&pDb);

    if( *pRc==0 ){
      int r1 = 0;
      int r2 = -1;

** should really be in this file.
*************************************************************************/

/*
** This test verifies that if a system crash occurs while doing merge work
** on the db, no data is lost.
*/
static void crash_test1(int bCompress, int *pRc){
  const char *DBNAME = "testdb.lsm";

  const DatasourceDefn defn = {TEST_DATASOURCE_RANDOM, 12, 16, 200, 200};

  const int nRow = 5000;          /* Database size */
  const int nIter = 200;          /* Number of test iterations */
  const int nWork = 20;           /* Maximum lsm_work() calls per iteration */
  const int nPage = 15;           /* Pages per lsm_work call */

  int i;
  int iDot = 0;
  Datasource *pData;
  CksumDb *pCksumDb;
  TestDb *pDb;
  char *zCfg;

  const char *azConfig[2] = {
    "page_size=1024 block_size=65536 write_buffer=16384 safety=2 mmap=0", 
    "page_size=1024 block_size=65536 write_buffer=16384 safety=2 "
    " compression=1 mmap=0"
  };
  assert( bCompress==0 || bCompress==1 );

  /* Allocate datasource. And calculate the expected checksums. */
  pData = testDatasourceNew(&defn);
  pCksumDb = testCksumArrayNew(pData, nRow, nRow, 1);

  /* Setup and save the initial database. */

  zCfg = testMallocPrintf("%s nmerge=7", azConfig[bCompress]);
  testSetupSavedLsmdb(zCfg, DBNAME, pData, 5000, pRc);
  testFree(zCfg);

  for(i=0; i<nIter && *pRc==0; i++){
    int iWork;
    int testrc = 0;

    testCaseProgress(i, nIter, testCaseNDot(), &iDot);

    /* Restore and open the database. */
    testRestoreLsmdb(DBNAME);
    testrc = tdb_lsm_open(azConfig[bCompress], DBNAME, 0, &pDb);
    assert( testrc==0 );

    /* Call lsm_work() on the db */
    tdb_lsm_prepare_sync_crash(pDb, 1 + (i%(nWork*2)));
    for(iWork=0; testrc==0 && iWork<nWork; iWork++){
      int nWrite = 0;
      lsm_db *db = tdb_lsm(pDb);
      testrc = lsm_work(db, 0, nPage, &nWrite);
      assert( testrc!=0 || nWrite>0 );
      if( testrc==0 ) testrc = lsm_checkpoint(db, 0);
    }
    tdb_close(pDb);

    /* Check that the database content is still correct */
    testCompareCksumLsmdb(DBNAME, 
        bCompress, testCksumArrayGet(pCksumDb, nRow), 0, pRc);
  }

  testCksumArrayFree(pCksumDb);
  testDatasourceFree(pData);
}

/*
** This test verifies that if a system crash occurs while committing a
** transaction to the log file, no earlier transactions are lost or damaged.
*/
static void crash_test2(int bCompress, int *pRc){
  const char *DBNAME = "testdb.lsm";
  const DatasourceDefn defn = {TEST_DATASOURCE_RANDOM, 12, 16, 1000, 1000};

  const int nIter = 200;
  const int nInsert = 20;

  int i;

      testDatasourceEntry(pData, 100+iIns, &pKey, &nKey, &pVal, &nVal);
      testrc = tdb_write(pDb, pKey, nKey, pVal, nVal);
      if( testrc ) break;
    }
    tdb_close(pDb);

    /* Check that no data was lost when the system crashed. */
    testCompareCksumLsmdb(DBNAME, bCompress,
      testCksumArrayGet(pCksumDb, 100 + iIns),
      testCksumArrayGet(pCksumDb, 100 + iIns + 1),
      pRc
    );
  }

  testDatasourceFree(pData);
  testCksumArrayFree(pCksumDb);
}

/*
** This test verifies that if a system crash occurs when checkpointing
** the database, data is not lost (assuming that any writes not synced
** to the db have been synced into the log file).
*/
static void crash_test3(int bCompress, int *pRc){
  const char *DBNAME = "testdb.lsm";
  const int nIter = 100;
  const DatasourceDefn defn = {TEST_DATASOURCE_RANDOM, 12, 16, 1000, 1000};

  int i;
  int iDot = 0;
  Datasource *pData;

      /* Schedule a crash simulation then close the db. */
      tdb_lsm_prepare_sync_crash(pDb, 1 + (i%2));
      tdb_close(pDb);

      /* Open the database and check that the crash did not cause any
      ** data loss.  */
      testCompareCksumLsmdb(DBNAME, bCompress,
        testCksumArrayGet(pCksumDb, 110 + iOpen*10), 0,
        pRc
      );
    }
  }

  testDatasourceFree(pData);
  testCksumArrayFree(pCksumDb);
}

void do_crash_test(const char *zPattern, int *pRc){
  struct Test {
    const char *zTest;
    void (*x)(int, int *);
    int bCompress;
  } aTest [] = {
    { "crash.lsm.1",     crash_test1, 0 },
    { "crash.lsm_zip.1", crash_test1, 1 },
    { "crash.lsm.2",     crash_test2, 0 },
    { "crash.lsm.3",     crash_test3, 0 },
  };
  int i;

  for(i=0; *pRc==LSM_OK && i<ArraySize(aTest); i++){
    struct Test *p = &aTest[i];
    if( testCaseBegin(pRc, zPattern, "%s", p->zTest) ){
      p->x(p->bCompress, pRc);
      testCaseFinish(*pRc);
    }
  }
}

        int iOpt = testPrngValue(iSeed++) % 3;
        switch( iOpt ){
          case 0:
            break;

          case 1:
            testPrngArray(iSeed++, (u32 *)aOld, pDb->szSector/4);
            /* Fall-through */

          case 2:
            pEnv->xWrite(
                pFile, (lsm_i64)i * pDb->szSector, aOld, pDb->szSector
            );
            break;
        }

int lsmFsFileid(lsm_db *pDb, void **ppId, int *pnId);

/* Creating, populating, gobbling and deleting sorted runs. */
void lsmFsGobble(lsm_db *, Segment *, Pgno *, int);
int lsmFsSortedDelete(FileSystem *, Snapshot *, int, Segment *);
int lsmFsSortedFinish(FileSystem *, Segment *);
int lsmFsSortedAppend(FileSystem *, Snapshot *, Segment *, Page **);
int lsmFsSortedPadding(FileSystem *, Snapshot *, Segment *);

/* Functions to retrieve the lsm_env pointer from a FileSystem or Page object */
lsm_env *lsmFsEnv(FileSystem *);
lsm_env *lsmPageEnv(Page *);
FileSystem *lsmPageFS(Page *);

int lsmFsSectorSize(FileSystem *);

  int nPagesize;                  /* Database page-size in bytes */
  int nBlocksize;                 /* Database block-size in bytes */

  /* r/w file descriptors for both files. */
  LsmFile *pLsmFile;
  lsm_file *fdDb;                 /* Database file */
  lsm_file *fdLog;                /* Log file */
  int szSector;                   /* Database file sector size */

  /* If this is a compressed database, a pointer to the compression methods.
  ** For an uncompressed database, a NULL pointer.  */
  lsm_compress *pCompress;
  u8 *aBuffer;                    /* Buffer to compress into */
  int nBuffer;                    /* Allocated size of aBuffer[] in bytes */

        pFS->fdDb = fsOpenFile(pFS, 0, &rc);
      }
    }

    if( rc!=LSM_OK ){
      lsmFsClose(pFS);
      pFS = 0;
    }else{
      pFS->szSector = lsmEnvSectorSize(pFS->pEnv, pFS->fdDb);
    }
  }

  pDb->pFS = pFS;
  return rc;
}

      }
      lsmSortedRemap(pFS->pDb);
    }
    *pRc = rc;
  }
}

static int fsPageGet(FileSystem *, Pgno, int, Page **, int *);

/*
** Parameter iBlock is a database file block. This function reads the value 
** stored in the blocks "next block" pointer and stores it in *piNext.
** LSM_OK is returned if everything is successful, or an LSM error code
** otherwise.
*/

    rc = lsmEnvRead(pFS->pEnv, pFS->fdDb, iOff, aNext, sizeof(aNext));
    if( rc==LSM_OK ){
      *piNext = (int)lsmGetU32(aNext);
    }
  }else{
    const int nPagePerBlock = (pFS->nBlocksize / pFS->nPagesize);
    Page *pLast;
    rc = fsPageGet(pFS, iBlock*nPagePerBlock, 0, &pLast, 0);
    if( rc==LSM_OK ){
      *piNext = fsPageToBlock(pFS, lsmGetU32(&pLast->aData[pFS->nPagesize-4]));
      lsmFsPageRelease(pLast);
    }
  }
  return rc;
}

  aBuf[2] = (u8)nByte | 0x80;
}
static int getRecordSize(u8 *aBuf, int *pbFree){
  int nByte;
  nByte  = (aBuf[0] & 0x7F) << 14;
  nByte += (aBuf[1] & 0x7F) << 7;
  nByte += (aBuf[2] & 0x7F);
  *pbFree = !(aBuf[1] & 0x80);
  return nByte;
}

static int fsSubtractOffset(FileSystem *pFS, i64 iOff, int iSub, i64 *piRes){
  i64 iStart;
  int iBlk;
  int rc;

  return rc;
}

static int fsAllocateBuffer(FileSystem *pFS){
  assert( pFS->pCompress );
  if( pFS->aBuffer==0 ){
    pFS->nBuffer = pFS->pCompress->xBound(pFS->pCompress->pCtx, pFS->nPagesize);
    if( pFS->nBuffer<(pFS->szSector+6) ){
      pFS->nBuffer = pFS->szSector+6;
    }
    pFS->aBuffer = lsmMalloc(pFS->pEnv, LSM_MAX(pFS->nBuffer, pFS->nPagesize));
    if( pFS->aBuffer==0 ) return LSM_NOMEM_BKPT;
  }
  return LSM_OK;
}

/*
** This function is only called in compressed database mode. It reads and
** uncompresses the compressed data for page pPg from the database and
** populates the pPg->aData[] buffer and pPg->nCompress field.
**
** LSM_OK is returned if successful, or an LSM error code otherwise.
*/
static int fsReadPagedata(
  FileSystem *pFS,                /* File-system handle */
  Page *pPg,                      /* Page to read and uncompress data for */
  int *pnSpace                    /* OUT: Total bytes of free space */
){
  lsm_compress *p = pFS->pCompress;
  i64 iOff = pPg->iPg;
  u8 aSz[3];
  int rc;

  assert( p && pPg->nCompress==0 );

  if( fsAllocateBuffer(pFS) ) return LSM_NOMEM;

  rc = fsReadData(pFS, iOff, aSz, sizeof(aSz));

  if( rc==LSM_OK ){
    int bFree;
    pPg->nCompress = (int)getRecordSize(aSz, &bFree);
    if( bFree ){
      if( pnSpace ){
        *pnSpace = pPg->nCompress + sizeof(aSz)*2;
      }else{
        rc = LSM_CORRUPT_BKPT;
      }
    }else{
      rc = fsAddOffset(pFS, iOff, 3, &iOff);
      if( rc==LSM_OK ){
        if( pPg->nCompress>pFS->nBuffer ){
          rc = LSM_CORRUPT_BKPT;
        }else{
          rc = fsReadData(pFS, iOff, pFS->aBuffer, pPg->nCompress);
        }
        if( rc==LSM_OK ){
          int n = pFS->nPagesize;
          rc = p->xUncompress(p->pCtx, 
              (char *)pPg->aData, &n, 
              (const char *)pFS->aBuffer, pPg->nCompress
              );
          if( rc==LSM_OK && n!=pPg->nData ){
            rc = LSM_CORRUPT_BKPT;
          }
        }
      }
    }
  }
  return rc;
}

/*
** Return a handle for a database page.
*/
static int fsPageGet(
  FileSystem *pFS,                /* File-system handle */
  Pgno iPg,                       /* Page id */
  int noContent,                  /* True to not load content from disk */
  Page **ppPg,                    /* OUT: New page handle */
  int *pnSpace                    /* OUT: Bytes of free space */
){
  Page *p;
  int iHash;
  int rc = LSM_OK;

  assert( iPg>=fsFirstPageOnBlock(pFS, 1) );
  *ppPg = 0;

    for(p=pFS->apHash[iHash]; p; p=p->pHashNext){
      if( p->iPg==iPg) break;
    }

    if( p==0 ){
      rc = fsPageBuffer(pFS, 1, &p);
      if( rc==LSM_OK ){
        int nSpace = 0;
        p->iPg = iPg;
        p->nRef = 0;
        p->pFS = pFS;
        assert( p->flags==0 || p->flags==PAGE_FREE );
        if( pFS->pCompress==0 && (fsIsLast(pFS, iPg) || fsIsFirst(pFS, iPg)) ){
          p->flags |= PAGE_SHORT;
        }
        p->nData =  pFS->nPagesize - (p->flags & PAGE_SHORT);

#ifdef LSM_DEBUG
        memset(p->aData, 0x56, pFS->nPagesize);
#endif
        assert( p->pLruNext==0 && p->pLruPrev==0 );
        if( noContent==0 ){
          if( pFS->pCompress ){
            rc = fsReadPagedata(pFS, p, &nSpace);
          }else{
            int nByte = pFS->nPagesize;
            i64 iOff = (i64)(iPg-1) * pFS->nPagesize;
            rc = lsmEnvRead(pFS->pEnv, pFS->fdDb, iOff, p->aData, nByte);
          }
          pFS->nRead++;
        }

        /* If the xRead() call was successful (or not attempted), link the
         ** page into the page-cache hash-table. Otherwise, if it failed,
         ** free the buffer. */
        if( rc==LSM_OK && nSpace==0 ){
          p->pHashNext = pFS->apHash[iHash];
          pFS->apHash[iHash] = p;
        }else{
          fsPageBufferFree(p);
          p = 0;
          if( pnSpace ) *pnSpace = nSpace;
        }
      }
    }else if( p->nRef==0 ){
      fsPageRemoveFromLru(pFS, p);
    }

    assert( (rc==LSM_OK && (p || (pnSpace && *pnSpace)))
         || (rc!=LSM_OK && p==0) 
    );
  }
  if( rc==LSM_OK && p ){
    pFS->nOut += (p->nRef==0);
    p->nRef++;
  }
  *ppPg = p;
  return rc;
}

    iBlk = iNext;
  }

  pRun->nSize -= (pRun->iFirst - fsFirstPageOnBlock(pFS, iBlk));
  assert( pRun->nSize>0 );
}

static int fsNextPageOffset(
  FileSystem *pFS,                /* File system object */
  Segment *pSeg,                  /* Segment to move within */
  Pgno iPg,                       /* Offset of current page */
  int nByte,                      /* Size of current page including headers */
  Pgno *piNext                    /* OUT: Offset of next page. Or zero (EOF) */
){
  Pgno iNext;
  int rc;

  assert( pFS->pCompress );

  rc = fsAddOffset(pFS, iPg, nByte-1, &iNext);
  if( pSeg && iNext==pSeg->iLastPg ){
    iNext = 0;
  }else if( rc==LSM_OK ){
    rc = fsAddOffset(pFS, iNext, 1, &iNext);
  }

  *piNext = iNext;
  return rc;
}

static int fsGetPageBefore(FileSystem *pFS, i64 iOff, Pgno *piPrev){

** is set to point to it and LSM_OK is returned. Otherwise, if an error 
** occurs, *ppNext is set to NULL and and lsm error code returned.
**
** Page references returned by this function should be released by the 
** caller using lsmFsPageRelease().
*/
int lsmFsDbPageNext(Segment *pRun, Page *pPg, int eDir, Page **ppNext){
  int rc = LSM_OK;
  FileSystem *pFS = pPg->pFS;
  Pgno iPg = pPg->iPg;

  if( pFS->pCompress ){
    int nSpace = pPg->nCompress + 2*3;

    do {
      if( eDir>0 ){
        rc = fsNextPageOffset(pFS, pRun, iPg, nSpace, &iPg);
      }else{
        if( iPg==pRun->iFirst ){
          iPg = 0;
        }else{
          rc = fsGetPageBefore(pFS, iPg, &iPg);
        }
      }

      nSpace = 0;
      if( iPg!=0 ){
        rc = fsPageGet(pFS, iPg, 0, ppNext, &nSpace);
        assert( (*ppNext==0)==(rc!=LSM_OK || nSpace>0) );
      }else{


        *ppNext = 0;

      }
    }while( nSpace>0 && rc==LSM_OK );

  }else{
    assert( eDir==1 || eDir==-1 );
    if( eDir<0 ){
      if( pRun && iPg==pRun->iFirst ){
        *ppNext = 0;
        return LSM_OK;
      }else if( fsIsFirst(pFS, iPg) ){

        return LSM_OK;
      }else if( fsIsLast(pFS, iPg) ){
        iPg = lsmGetU32(&pPg->aData[pFS->nPagesize-4]);
      }else{
        iPg++;
      }
    }
    rc = fsPageGet(pFS, iPg, 0, ppNext, 0);
  }

  return rc;
}

static Pgno findAppendPoint(FileSystem *pFS){
  int i;
  Pgno *aiAppend = pFS->pDb->pWorker->aiAppend;
  u32 iRet = 0;

      }else{
        iNext = fsFirstPageOnBlock(pFS, iNew);
      }
    }

    /* Grab the new page. */
    pPg = 0;
    rc = fsPageGet(pFS, iApp, 1, &pPg, 0);
    assert( rc==LSM_OK || pPg==0 );

    /* If this is the first or last page of a block, fill in the pointer 
     ** value at the end of the new page. */
    if( rc==LSM_OK ){
      p->nSize++;
      p->iLastPg = iApp;

/*
** Mark the sorted run passed as the second argument as finished. 
*/
int lsmFsSortedFinish(FileSystem *pFS, Segment *p){
  int rc = LSM_OK;
  if( p && p->iLastPg ){

    int iBlk;

    /* Check if the last page of this run happens to be the last of a block.
    ** If it is, then an extra block has already been allocated for this run.
    ** Shift this extra block back to the free-block list. 
    **
    ** Otherwise, add the first free page in the last block used by the run

        if( aiAppend[i]==0 ){
          aiAppend[i] = p->iLastPg+1;
          break;
        }
      }
    }else if( pFS->pCompress==0 ){
      Page *pLast;
      rc = fsPageGet(pFS, p->iLastPg, 0, &pLast, 0);
      if( rc==LSM_OK ){
        int iPg = (int)lsmGetU32(&pLast->aData[pFS->nPagesize-4]);
        lsmBlockRefree(pFS->pDb, fsPageToBlock(pFS, iPg));
        lsmFsPageRelease(pLast);
      }
    }else{
      int iBlk = 0;
      rc = fsBlockNext(pFS, fsPageToBlock(pFS, p->iLastPg), &iBlk);
      if( rc==LSM_OK ){
        lsmBlockRefree(pFS->pDb, iBlk);
      }
    }
  }
  return rc;
}

/*
** Obtain a reference to page number iPg.
*/
int lsmFsDbPageGet(FileSystem *pFS, Pgno iPg, Page **ppPg){
  assert( pFS );
  return fsPageGet(pFS, iPg, 0, ppPg, 0);
}

/*
** Obtain a reference to the last page in the segment passed as the 
** second argument.
*/
int lsmFsDbPageLast(FileSystem *pFS, Segment *pSeg, Page **ppPg){
  int rc;
  Pgno iPg = pSeg->iLastPg;
  if( pFS->pCompress ){
    int nSpace;
    iPg++;
    do {
      nSpace = 0;
      rc = fsGetPageBefore(pFS, iPg, &iPg);
      if( rc==LSM_OK ){
        rc = fsPageGet(pFS, iPg, 0, ppPg, &nSpace);
      }
    }while( rc==LSM_OK && nSpace>0 );

  }else{
    rc = fsPageGet(pFS, iPg, 0, ppPg, 0);
  }
  return rc;
}

/*
** Return a reference to meta-page iPg. If successful, LSM_OK is returned
** and *ppPg populated with the new page reference. The reference should
** be released by the caller using lsmFsPageRelease().
**

){
  Pgno iRet = 0;
  int rc = *pRc;
  assert( pFS->pCompress );
  if( rc==LSM_OK ){
    int nRem;
    int nWrite;
    Pgno iLastOnBlock;
    Pgno iApp = pSeg->iLastPg+1;

    /* If this is the first data written into the segment, find an append-point
    ** or allocate a new block.  */
    if( iApp==1 ){
      pSeg->iFirst = iApp = findAppendPoint(pFS);
      if( iApp==0 ){
        int iBlk;
        rc = lsmBlockAllocate(pFS->pDb, &iBlk);
        pSeg->iFirst = iApp = fsFirstPageOnBlock(pFS, iBlk);
      }
    }

    iRet = iApp;

    /* Write as much data as is possible at iApp (usually all of it). */
    iLastOnBlock = fsLastPageOnBlock(pFS, fsPageToBlock(pFS, iApp));
    if( rc==LSM_OK ){
      int nSpace = iLastOnBlock - iApp + 1;
      nWrite = LSM_MIN(nData, nSpace);
      nRem = nData - nWrite;
      assert( nWrite>=0 );
      if( nWrite!=0 ){
        rc = lsmEnvWrite(pFS->pEnv, pFS->fdDb, iApp, aData, nWrite);
      }
      iApp += nWrite;
    }

    /* If required, allocate a new block and write the rest of the data
    ** into it. Set the next and previous block pointers to link the new
    ** block to the old.  */
    assert( nRem<=0 || (iApp-1)==iLastOnBlock );
    if( rc==LSM_OK && (iApp-1)==iLastOnBlock ){
      u8 aPtr[4];                 /* Space to serialize a u32 */
      int iBlk;                   /* New block number */

      if( nWrite>0 ){
        /* Allocate a new block. */
        rc = lsmBlockAllocate(pFS->pDb, &iBlk);

        /* Set the "next" pointer on the old block */
        if( rc==LSM_OK ){
          assert( iApp==(fsPageToBlock(pFS, iApp)*pFS->nBlocksize)-4 );
          lsmPutU32(aPtr, iBlk);
          rc = lsmEnvWrite(pFS->pEnv, pFS->fdDb, iApp, aPtr, sizeof(aPtr));
        }

        /* Set the "prev" pointer on the new block */
        if( rc==LSM_OK ){
          Pgno iWrite;
          lsmPutU32(aPtr, fsPageToBlock(pFS, iApp));
          iWrite = fsFirstPageOnBlock(pFS, iBlk);
          rc = lsmEnvWrite(pFS->pEnv, pFS->fdDb, iWrite-4, aPtr, sizeof(aPtr));
          if( nRem>0 ) iApp = iWrite;
        }
      }else{
        /* The next block is already allocated. */
        assert( nRem>0 );
        rc = fsBlockNext(pFS, fsPageToBlock(pFS, iApp), &iBlk);
        iApp = fsFirstPageOnBlock(pFS, iBlk);
      }

      /* Write the remaining data into the new block */
      if( rc==LSM_OK && nRem>0 ){
        rc = lsmEnvWrite(pFS->pEnv, pFS->fdDb, iApp, &aData[nWrite], nRem);
        iApp += nRem;
      }
    }

    pSeg->iLastPg = iApp-1;
    *pRc = rc;

    }
    pPg->flags &= ~PAGE_DIRTY;
    pFS->nWrite++;
  }

  return rc;
}

/*
** Add a padding record to the segment passed as the third argument.
*/
int lsmFsSortedPadding(
  FileSystem *pFS, 
  Snapshot *pSnapshot,
  Segment *pSeg
){
  int rc = LSM_OK;
  if( pFS->pCompress ){
    Pgno iLast2;
    Pgno iLast = pSeg->iLastPg;     /* Current last page of segment */
    int nPad;                       /* Bytes of padding required */
    u8 aSz[3];

    nPad = pFS->szSector - 1 - (iLast % pFS->szSector);
    if( nPad==0 
     || (nPad==4 && iLast==fsLastPageOnBlock(pFS, fsPageToBlock(pFS, iLast)) )
    ){
      return LSM_OK;
    }

    iLast2 = (1 + (iLast + 6)/pFS->szSector) * pFS->szSector - 1;
    assert( fsPageToBlock(pFS, iLast)==fsPageToBlock(pFS, iLast2) );
    nPad = iLast2 - iLast;

    if( iLast2>fsLastPageOnBlock(pFS, fsPageToBlock(pFS, iLast)) ){
      nPad -= 4;
      if( nPad<6 ){
        nPad += (pFS->szSector - 4);
      }
    }
    assert( nPad>=6 );

#if 0
    printf("padding segment with %d bytes at %d...\n", nPad, (int)iLast);
#endif

    pSeg->nSize += nPad;
    nPad -= 6;
    putRecordSize(aSz, nPad, 1);

    fsAppendData(pFS, pSeg, aSz, sizeof(aSz), &rc);
    memset(pFS->aBuffer, 0, nPad);
    fsAppendData(pFS, pSeg, pFS->aBuffer, nPad, &rc);
    fsAppendData(pFS, pSeg, aSz, sizeof(aSz), &rc);


    assert( rc!=LSM_OK 
     || pSeg->iLastPg==fsLastPageOnBlock(pFS, fsPageToBlock(pFS, pSeg->iLastPg))
     || ((pSeg->iLastPg + 1) % pFS->szSector)==0
    );
  }

  return rc;
}


/*
** Increment the reference count on the page object passed as the first
** argument.
*/
void lsmFsPageRef(Page *pPg){
  if( pPg ){

  Segment *pSeg,
  int bExtra,                     /* If true, count the "next" block if any */
  int nUsed,
  u8 *aUsed
){
  if( pSeg ){
    if( pSeg && pSeg->nSize>0 ){

      Pgno iLast = pSeg->iLastPg;
      int iBlk;
      int iLastBlk;
      iBlk = fsPageToBlock(pFS, pSeg->iFirst);
      iLastBlk = fsPageToBlock(pFS, pSeg->iLastPg);

      while( iBlk ){
        assert( iBlk<=nUsed );
        /* assert( aUsed[iBlk-1]==0 ); */
        aUsed[iBlk-1] = 1;
        if( iBlk!=iLastBlk ){
          fsBlockNext(pFS, iBlk, &iBlk);
        }else{
          iBlk = 0;
        }
      }

      if( bExtra && iLast==fsLastPageOnBlock(pFS, fsPageToBlock(pFS, iLast)) ){
        fsBlockNext(pFS, iLastBlk, &iBlk);
        aUsed[iBlk-1] = 1;
      }
    }
  }
}

/*
** Return true if pPg happens to be the last page in segment pSeg. Or false
** otherwise. This function is only invoked as part of assert() conditions.
*/
int lsmFsDbPageIsLast(Segment *pSeg, Page *pPg){
  if( pPg->pFS->pCompress ){
    Pgno iNext = 0;
    int rc;
    rc = fsNextPageOffset(pPg->pFS, pSeg, pPg->iPg, pPg->nCompress+6, &iNext);
    return (rc!=LSM_OK || iNext==0);
  }
  return (pPg->iPg==pSeg->iLastPg);
}
#endif

    lsmFree(pMW->pDb->pEnv, pMW->hier.apHier);
    pMW->hier.apHier = 0;
    pMW->hier.nHier = 0;
  }

  return rc;
}

static int mergeWorkerAddPadding(
  MergeWorker *pMW                /* Merge worker object */
){
  FileSystem *pFS = pMW->pDb->pFS;
  return lsmFsSortedPadding(pFS, pMW->pDb->pWorker, &pMW->pLevel->lhs);
}

static int keyszToSkip(FileSystem *pFS, int nKey){
  int nPgsz;                /* Nominal database page size */
  nPgsz = lsmFsPageSize(pFS);
  return LSM_MIN(((nKey * 4) / nPgsz), 3);
}

  Merge *pMerge;                  /* Persistent part of level merge state */
  int nHdr;                       /* Space required for this record header */
  Page *pPg;                      /* Page to write to */
  u8 *aData;                      /* Data buffer for page pWriter->pPage */
  int nData;                      /* Size of buffer aData[] in bytes */
  int nRec;                       /* Number of records on page pPg */
  int iFPtr;                      /* Value of pointer in footer of pPg */
  int iRPtr = 0;                  /* Value of pointer written into record */
  int iOff;                       /* Current write offset within page pPg */
  Segment *pSeg;                  /* Segment being written */
  int flags = 0;                  /* If != 0, flags value for page footer */
  int bFirst = 0;                 /* True for first key of output run */
  void *pVal;
  int nVal;

  lsmMCursorClose(pCsr);

  /* Persist and release the output page. */
  if( rc==LSM_OK ) rc = mergeWorkerPersistAndRelease(pMW);
  if( rc==LSM_OK ) rc = mergeWorkerBtreeIndirect(pMW);
  if( rc==LSM_OK ) rc = mergeWorkerFinishHierarchy(pMW);
  if( rc==LSM_OK ) rc = mergeWorkerAddPadding(pMW);

  lsmFree(pMW->pDb->pEnv, pMW->aGobble);
  pMW->aGobble = 0;
  pMW->pCsr = 0;
  pMW->pPage = 0;
  pMW->pPage = 0;