SQLite4  Check-in [65642c32ba]

typedef sqlite4_int64 i64;
typedef sqlite4_uint64 u64;
typedef unsigned int u32;
typedef unsigned short u16;
typedef unsigned char u8;

/* 
** Special error codes used internally. These must be distinct from SQLite4
** error codes. Which is easy - SQLite4 error codes are all greater than or
** equal to zero.
*/
#define BT_BLOCKFULL -1


/* Number of elements in an array object. */
#define array_size(x) (sizeof(x)/sizeof(x[0]))

/* Number of read-lock slots in shared memory */
#define BT_NREADER 4

/* By default blocks are 512K bytes in size. */
#define BT_DEFAULT_BLKSZ (512*1024)

typedef struct BtDbHdr BtDbHdr;
struct BtDbHdr {
  u32 pgsz;                       /* Page size in bytes */
  u32 blksz;                      /* Block size in bytes */
  u32 nPg;                        /* Size of database file in pages */

  u32 iRoot;                      /* B-tree root page */
  u32 iMRoot;                     /* Root page of meta-tree */
  u32 iSRoot;                     /* Root page of schedule-tree */

  u32 iSubRoot;                   /* Root of current sub-tree */

    memcpy(&hdr, aData, sizeof(BtDbHdrCksum));
    btLogChecksum32(1, (u8*)&hdr, offsetof(BtDbHdrCksum, aCksum), 0, aCksum);
  }

  if( aData==0 || aCksum[0]!=hdr.aCksum[0] || aCksum[1]!=hdr.aCksum[1] ){
    memset(&hdr, 0, sizeof(BtDbHdrCksum));
    hdr.hdr.pgsz = BT_DEFAULT_PGSZ;
    hdr.hdr.blksz = BT_DEFAULT_BLKSZ;
    hdr.hdr.nPg = 2;
    hdr.hdr.iRoot = 2;
  }

  memcpy(pHdr, &hdr, sizeof(BtDbHdr));
}

      btPutU16(&aData[pgsz-6], iWrite);
    }

  }else{
    /* The new entry will not fit on the leaf page. Entries will have
    ** to be shuffled between existing leaves and new leaves may need
    ** to be added to make space for it. */
    bt_db *db = pCsr->base.pDb;
    if( db->bFastInsertOp ){
      BtDbHdr *pHdr = sqlite4BtPagerDbhdr(db->pPager);
      if( pHdr->nSubPg >= (pHdr->blksz / pHdr->pgsz) ){
        rc = BT_BLOCKFULL;
        pHdr->iSubRoot = 0;
      }
    }
    if( rc==SQLITE4_OK && pCsr->nPg==1 ){
      rc = btExtendTree(pCsr);
    }
    if( rc==SQLITE4_OK ){
      rc = btBalance(pCsr, bLeaf, nKV, apKV);
    }
  }

      }
    }
  }

  return rc;
}

static int btFastInsertRoot(bt_db *db, BtDbHdr *pHdr, u32 *piRoot);

static int btReplaceEntry(
  bt_db *db,                      /* Database handle */
  u32 iRoot,                      /* Root page of b-tree to update */
  const void *pK, int nK,         /* Key to insert */
  const void *pV, int nV          /* Value to insert. (nV<0) -> delete */
){
  BtDbHdr *pHdr = sqlite4BtPagerDbhdr(db->pPager);
  int rc;                         /* Return code */
  BtCursor csr;                  /* Cursor object to seek to insert point */
  u32 iRootPg = iRoot;

  if( rc==SQLITE4_OK && iRoot==0 ){
    rc = btFastInsertRoot(db, pHdr, &iRootPg);
  }

  /* Seek stack cursor csr to the b-tree page that key pK/nK is/would be
  ** stored on.  */
  btCsrSetup(db, iRootPg, &csr);
  rc = btCsrSeek(&csr, pK, nK, BT_SEEK_GE, BT_CSRSEEK_UPDATE);

  if( rc==SQLITE4_OK ){
    /* The cursor currently points to an entry with key pK/nK. This call
    ** should therefore replace that entry. So delete it and then re-seek
    ** the cursor.  */
    rc = sqlite4BtDelete(&csr.base);

    if( rc==SQLITE4_OK && nV>=0 ){
      rc = btCsrSeek(&csr, pK, nK, BT_SEEK_GE, BT_CSRSEEK_UPDATE);
      if( rc==SQLITE4_OK ) rc = btErrorBkpt(SQLITE4_CORRUPT);
    }
  }

  if( nV>=0 && (rc==SQLITE4_NOTFOUND || rc==SQLITE4_INEXACT ) ){

    KeyValue kv;
    kv.pgno = 0;
    kv.eType = KV_VALUE;
    kv.pK = pK; kv.nK = nK;
    kv.pV = pV; kv.nV = nV;

    rc = btOverflowAssign(db, &kv);
    if( rc==SQLITE4_OK ){
      do{
        /* Insert the new KV pair into the current leaf. */
        rc = btInsertAndBalance(&csr, 1, &kv);

        /* Unless this is a block-full error, break out of the loop */
        if( rc!=BT_BLOCKFULL ) break;
        assert( iRoot==0 );

        rc = btFastInsertRoot(db, pHdr, &iRootPg);
        if( rc==SQLITE4_OK ){
          btCsrReset(&csr, 1);
          btCsrSetup(db, iRootPg, &csr);
          rc = btCsrSeek(&csr, pK, nK, BT_SEEK_GE, BT_CSRSEEK_UPDATE);
        }
      }while( rc==SQLITE4_NOTFOUND || rc==SQLITE4_INEXACT );
    }

    if( kv.eType==KV_CELL ){
      sqlite4_free(db->pEnv, (void*)kv.pV);
    }
  }

  btCsrReset(&csr, 1);

  return rc;
}

static int btAllocateNewRoot(bt_db *db, u32 *piNew){
  u32 iNew = 0;
  BtPage *pPg;
  int rc;

  rc = btAllocateNonOverflow(db, &pPg);
  if( rc==SQLITE4_OK ){
    iNew = sqlite4BtPagePgno(pPg);
    sqlite4BtPageRelease(pPg);
  }

  *piNew = iNew;
  return rc;
}

static int btDecodeMetatreeKey(
  BtCursor *pCsr,
  u32 *piAge,
  u32 *piLevel,
  u8 **paKey,
  int *pnKey
){
  u8 *aK; int nK;
  int rc = sqlite4BtCsrKey((bt_cursor*)pCsr, (const void**)&aK, &nK);
  if( rc==SQLITE4_OK ){
    *piAge = btGetU32(&aK[0]);
    *piLevel = ~btGetU32(&aK[4]);
    if( paKey ){
      *paKey = &aK[8];
      *pnKey = nK-8;
    }
  }
  return rc;
}

static int btFindNextLevel(
  bt_db *db, 
  BtDbHdr *pHdr, 
  u32 *piNext
){
  int rc;
  BtCursor csr;

  btCsrSetup(db, pHdr->iMRoot, &csr);
  rc = btCsrEnd(&csr, 0);
  assert( rc!=SQLITE4_INEXACT );

  *piNext = 0;
  if( rc==SQLITE4_OK ){
    u32 iAge;
    u32 iLevel;
    rc = btDecodeMetatreeKey(&csr, &iAge, &iLevel, 0, 0);
    if( rc==SQLITE4_OK && iAge==0 ){
      *piNext = iLevel+1;
    }
  }else if( rc==SQLITE4_NOTFOUND ){
    rc = SQLITE4_OK;

  }
  btCsrReset(&csr, 1);

  return rc;
}

static int btFastInsertRoot(
  bt_db *db, 
  BtDbHdr *pHdr, 
  u32 *piRoot
){
  int rc = SQLITE4_OK;
  u32 iSubRoot = 0;

  if( pHdr->iMRoot==0 ){
    rc = btAllocateNewRoot(db, &pHdr->iMRoot);
  }
  iSubRoot = pHdr->iSubRoot;






  /* If no writable sub-tree has been discovered, create one now. */
  if( iSubRoot==0 ){
    u32 iLevel;





    rc = btFindNextLevel(db, pHdr, &iLevel);

    if( rc==SQLITE4_OK ){
      rc = btAllocateNewRoot(db, &iSubRoot);
    }

    if( rc==SQLITE4_OK ){
      u8 aKey[8];
      u8 aVal[8];
      pHdr->iSubRoot = iSubRoot;
      pHdr->nSubPg = 0;

      /* The key for the new entry consists of the concatentation of two 
      ** 32-bit big-endian integers - the <age> and <level-no>. The age
      ** of the new segment is 0. The level number is one greater than the
      ** level number of the previous segment.  */
      btPutU32(&aKey[0], 0);
      btPutU32(&aKey[4], ~iLevel);

      btPutU32(&aVal[0], iSubRoot);
      btPutU32(&aVal[4], 1);
      rc = btReplaceEntry(db, pHdr->iMRoot, aKey, 8, aVal, 8);
    }
  }

  *piRoot = iSubRoot;
  return rc;
}

  assert( rc!=SQLITE4_NOTFOUND && rc!=SQLITE4_INEXACT );
  btCheckPageRefs(db);

  /* Seek stack cursor csr to the b-tree page that key pK/nK is/would be
  ** stored on.  */
  if( rc==SQLITE4_OK ){
    BtDbHdr *pHdr = sqlite4BtPagerDbhdr(db->pPager);
    u32 iRoot = 0;
    if( !db->bFastInsertOp ){


      iRoot = pHdr->iRoot;
    }
    if( rc==SQLITE4_OK ){
      rc = btReplaceEntry(db, iRoot, pK, nK, pV, nV);
    }
  }

  btCheckPageRefs(db);
  db->bFastInsertOp = 0;
  return rc;
}

** or SQLITE4_OK otherwise.
*/
int sqlite4BtPagerHdrdump(BtPager *pPager, sqlite4_buffer *pBuf){
  BtDbHdr *pHdr = pPager->pHdr;
  int rc = SQLITE4_OK;

  sqlite4BtBufAppendf(pBuf, 
      "pgsz=%d blksz=%d nPg=%d"
      " iRoot=%d iMRoot=%d iSRoot=%d"
      " iCookie=%d iFreePg=%d iFreeBlk=%d",
      pHdr->pgsz, pHdr->blksz, pHdr->nPg, 
      pHdr->iRoot, pHdr->iMRoot, pHdr->iSRoot,
      pHdr->iCookie, pHdr->iFreePg, pHdr->iFreeBlk
  );

  return rc;
}

#ifndef NDEBUG

  <li><p>Each segment is assigned to a level. Each level contains a 
  set of non-overlapping segments. All segments in level N are newer
  than those in segment (N+1).

  <li><p>Segments are linked together using a tree-like data structure
  stored on regular database pages (mixed in with b-tree data pages).
</ul>

<h3> Meta-Tree Format</h3>

<p> The meta-tree is a b-tree stored in the same format as the main user
b-tree. It serves as an index for all segments in the database. The keys
in the meta-tree are formated as follows:

<ul>
  <li> Segment age (big-endian 32-bit integer).
  <li> Level number (ones-compliment of value as a big-endian 32-bit integer).
  <li> Separator key. All keys within the segment are guaranteed to be as
       large or larger than the separator key. The separator key may be
       zero bytes in size.
</ul>


<h2 id=in-memory_tree>3.2. In-Memory Tree</h2>

<p>This design uses the same multi-version in-memory tree that LSM does. 
The tree structure may be located in heap or shared memory. If shared memory
is used then a separate file (*-shm2 or some such) is used for the data. The 
tree-header used by LSM to identify the latest tree version is combined with