SQLite4  Check-in [acf38e32c8]

#define array_size(x) (sizeof(x)/sizeof(x[0]))


/* The size of each shared-memory chunk */
#define LSM_SHM_CHUNK_SIZE (32*1024)

/* The number of bytes reserved at the start of each shm chunk for MM. */
#define LSM_SHM_CHUNK_HDR  (sizeof(ShmChunk))

/* The number of available read locks. */
#define LSM_LOCK_NREADER   6

/* Lock definitions */
#define LSM_LOCK_DMS1         1
#define LSM_LOCK_DMS2         2

  LogRegion aRegion[3];           /* Log file regions (see docs in lsm_log.c) */
};

/*
** Tree header structure. 
*/
struct TreeHeader {
  u32 iUsedShmid;                 /* Id of first shm chunk used by this tree */
  u32 iNextShmid;                 /* Shm-id of next chunk allocated */
  u32 iFirst;                     /* Chunk number of smallest shm-id */
  u32 iTransId;                   /* Current transaction id */
  u32 nChunk;                     /* Number of chunks in shared-memory file */
  u32 iRoot;                      /* Offset of root node in shm file */
  u32 nHeight;                    /* Current height of tree structure */
  u32 iWrite;                     /* Write offset in shm file */


  u32 nByte;                      /* Size of current tree structure in bytes */
  DbLog log;                      /* Current layout of log file */ 
  u32 aCksum[2];                  /* Checksums 1 and 2. */
};

/*
** Database handle structure.

*/
#define segmentHasSeparators(pSegment) ((pSegment)->sep.iFirst>0)

/*
** The values that accompany the lock held by a database reader.
*/
struct ShmReader {
  u32 iTreeId;
  i64 iLsmId;
};

/*
** An instance of this structure is stored in the first shared-memory
** page. The shared-memory header.
**

};

/*
** An instance of this structure is stored at the start of each shared-memory
** chunk except the first (which is the header chunk - see above).
*/
struct ShmChunk {

  u32 iShmid;
  u32 iNext;
};

/* Return true if shm-sequence "a" is larger than or equal to "b" */
#define shm_sequence_ge(a, b) (((u32)a-(u32)b) < (1<<30))

#define LSM_APPLIST_SZ 4

typedef struct Freelist Freelist;
typedef struct FreelistEntry FreelistEntry;

/*

/* 
** Functions from file "lsm_tree.c".
*/
int lsmTreeNew(lsm_env *, int (*)(void *, int, void *, int), Tree **ppTree);
void lsmTreeRelease(lsm_env *, Tree *);
void lsmTreeClear(lsm_db *);
int lsmTreeInit(lsm_db *);

int lsmTreeSize(lsm_db *);
int lsmTreeEndTransaction(lsm_db *pDb, int bCommit);
int lsmTreeBeginTransaction(lsm_db *pDb);
int lsmTreeLoadHeader(lsm_db *pDb);

int lsmTreeInsert(lsm_db *pDb, void *pKey, int nKey, void *pVal, int nVal);

#ifdef LSM_DEBUG
void lsmShmHasLock(lsm_db *db, int iLock, int eOp);
#else
# define lsmShmHasLock(x,y,z)
#endif

int lsmReadlock(lsm_db *, i64 iLsm, u32 iTree);
int lsmReleaseReadlock(lsm_db *);

int lsmLsmInUse(lsm_db *db, i64 iLsmId, int *pbInUse);
int lsmTreeInUse(lsm_db *db, u32 iLsmId, int *pbInUse);
int lsmFreelistAppend(lsm_env *pEnv, Freelist *p, int iBlk, i64 iId);

int lsmDbMultiProc(lsm_db *);

  int iPass;
  int nJump = 0;                  /* Number of LSM_LOG_JUMP records in pass 0 */
  DbLog *pLog;

  rc = lsmFsOpenLog(pDb->pFS);
  if( rc!=LSM_OK ) return rc;

  rc = lsmTreeInit(pDb);
  if( rc!=LSM_OK ) return rc;

  pLog = &pDb->treehdr.log;
  lsmCheckpointLogoffset(pDb->pShmhdr->aWorker, pLog);

  logReaderInit(pDb, pLog, 1, &reader);
  lsmStringInit(&buf1, pDb->pEnv);
  lsmStringInit(&buf2, pDb->pEnv);

  }
}

LsmFile *lsmDbRecycleFd(lsm_db *db){
  LsmFile *pRet;
  Database *p = db->pDatabase;
  lsmMutexEnter(db->pEnv, p->pClientMutex);
  if( (pRet = p->pLsmFile)!=0 ){
    p->pLsmFile = pRet->pNext;
  }
  lsmMutexLeave(db->pEnv, p->pClientMutex);
  return pRet;
}

/*

    /* Take a read-lock on the tree and snapshot just loaded. Then check
    ** that the shared-memory still contains the same values. If so, proceed.
    ** Otherwise, relinquish the read-lock and retry the whole procedure
    ** (starting with loading the in-memory tree header).  */
    if( rc==LSM_OK ){
      ShmHeader *pShm = pDb->pShmhdr;
      u32 iShmchunk = pDb->treehdr.iUsedShmid;
      i64 iSnap = lsmCheckpointId(pDb->aSnapshot, 0);
      rc = lsmReadlock(pDb, iSnap, iShmchunk);
      if( rc==LSM_OK ){

        if( 0==memcmp(pShm->hdr1.aCksum, pDb->treehdr.aCksum, sizeof(u32)*2)
         && iSnap==lsmCheckpointId(pShm->aClient, 0)
        ){
          /* Read lock has been successfully obtained. Deserialize the 
          ** checkpoint just loaded. TODO: This will be removed after 
          ** lsm_sorted.c is changed to work directly from the serialized
          ** version of the snapshot.  */
          rc = lsmCheckpointDeserialize(pDb, 0, pDb->aSnapshot, &pDb->pClient);
          assert( (rc==LSM_OK)==(pDb->pClient!=0) );

#endif

/*
** Obtain a read-lock on database version identified by the combination
** of snapshot iLsm and tree iTree. Return LSM_OK if successful, or
** an LSM error code otherwise.
*/
int lsmReadlock(lsm_db *db, i64 iLsm, u32 iTree){
  ShmHeader *pShm = db->pShmhdr;
  int i;
  int rc = LSM_OK;

  assert( db->iReader<0 );

  /* Search for an exact match. */

      if( rc==LSM_OK ) db->iReader = i;
    }
  }

  /* Search for any usable slot */
  for(i=0; db->iReader<0 && rc==LSM_OK && i<LSM_LOCK_NREADER; i++){
    ShmReader *p = &pShm->aReader[i];
    if( p->iLsmId && p->iLsmId<=iLsm && shm_sequence_ge(iTree, p->iTreeId) ){
      rc = lsmShmLock(db, LSM_LOCK_READER(i), LSM_LOCK_SHARED, 0);
      if( rc==LSM_OK ){
        if( p->iLsmId && p->iLsmId<=iLsm && shm_sequence_ge(iTree,p->iTreeId) ){
          db->iReader = i;
        }
      }else if( rc==LSM_BUSY ){
        rc = LSM_OK;
      }
    }
  }

  return rc;
}

/*
** This is used to check if there exists a read-lock locking a particular
** version of either the in-memory tree or database file. 
**
** If iLsmId is non-zero, then it is a snapshot id. If there exists a 
** read-lock using this snapshot or newer, set *pbInUse to true. Or,
** if there is no such read-lock, set it to false.
**
** Or, if iLsmId is zero, then iShmid is a shared-memory sequence id.
** Search for a read-lock using this sequence id or newer. etc.
*/
static int isInUse(lsm_db *db, i64 iLsmId, u32 iShmid, int *pbInUse){
  ShmHeader *pShm = db->pShmhdr;
  int i;
  int rc = LSM_OK;

  for(i=0; rc==LSM_OK && i<LSM_LOCK_NREADER; i++){
    ShmReader *p = &pShm->aReader[i];
    if( p->iLsmId ){
      if( (iLsmId!=0 && iLsmId>=p->iLsmId) 
       || (iLsmId==0 && shm_sequence_ge(p->iTreeId, iShmid))
      ){
        rc = lsmShmLock(db, LSM_LOCK_READER(i), LSM_LOCK_EXCL, 0);
        if( rc==LSM_OK ){
          p->iLsmId = 0;
          lsmShmLock(db, LSM_LOCK_READER(i), LSM_LOCK_UNLOCK, 0);
        }
      }
    }
  }

  if( rc==LSM_BUSY ){
    *pbInUse = 1;
    return LSM_OK;
  }
  *pbInUse = 0;
  return rc;
}

int lsmTreeInUse(lsm_db *db, u32 iShmid, int *pbInUse){
  if( db->treehdr.iUsedShmid==iShmid ){
    *pbInUse = 1;
    return LSM_OK;
  }
  return isInUse(db, 0, iShmid, pbInUse);
}

int lsmLsmInUse(lsm_db *db, i64 iLsmId, int *pbInUse){
  if( db->pClient && db->pClient->iId<=iLsmId ){
    *pbInUse = 1;
    return LSM_OK;
  }

  return 0;
}

static ShmChunk * treeShmChunk(lsm_db *pDb, int iChunk){
  int rcdummy = LSM_OK;
  return (ShmChunk *)treeShmptr(pDb, iChunk*LSM_SHM_CHUNK_SIZE, &rcdummy);
}

static ShmChunk * treeShmChunkRc(lsm_db *pDb, int iChunk, int *pRc){
  return (ShmChunk *)treeShmptr(pDb, iChunk*LSM_SHM_CHUNK_SIZE, pRc);
}

/* Values for the third argument to treeShmkey(). */
#define TK_LOADKEY  1
#define TK_LOADVAL  2

static TreeKey *treeShmkey(
  lsm_db *pDb,                    /* Database handle */

    assert( iWrite );
    iChunk = treeOffsetToChunk(iWrite-1);
    iEof = (iChunk+1) * CHUNK_SIZE;
    assert( iEof>=iWrite && (iEof-iWrite)<CHUNK_SIZE );
    if( (iWrite+nByte)>iEof ){
      ShmChunk *pHdr;           /* Header of chunk just finished (iChunk) */
      ShmChunk *pFirst;         /* Header of chunk treehdr.iFirst */
      ShmChunk *pNext;          /* Header of new chunk */
      int iNext = 0;            /* Next chunk */
      int rc = LSM_OK;

      pFirst = treeShmChunk(pDb, pDb->treehdr.iFirst);

      /* Check if the chunk at the start of the linked list is still in
      ** use. If not, reuse it. If so, allocate a new chunk by appending
      ** to the *-shm file.  */
      assert( shm_sequence_ge(pDb->treehdr.iUsedShmid, pFirst->iShmid) );
      if( pDb->treehdr.iUsedShmid!=pFirst->iShmid ){
        int bInUse;

        rc = lsmTreeInUse(pDb, pFirst->iShmid, &bInUse);
        if( rc!=LSM_OK ){
          *pRc = rc;
          return 0;
        }
        if( bInUse==0 ){
          iNext = pDb->treehdr.iFirst;
          pDb->treehdr.iFirst = pFirst->iNext;


          assert( pDb->treehdr.iFirst );

        }
      }
      if( iNext==0 ) iNext = pDb->treehdr.nChunk++;

      /* Set the header values for the new chunk */
      pNext = treeShmChunkRc(pDb, iNext, &rc);
      if( pNext ){
        pNext->iNext = 0;
        pNext->iShmid = (pDb->treehdr.iNextShmid++);
      }else{
        *pRc = rc;
        return 0;
      }

      /* Set the header values for the chunk just finished */
      pHdr = (ShmChunk *)treeShmptr(pDb, iChunk*CHUNK_SIZE, pRc);

      pHdr->iNext = iNext;

      /* Advance to the next chunk */
      iWrite = iNext * CHUNK_SIZE + CHUNK_HDR;
    }

    /* Allocate space at iWrite. */

  return rc;
}

/*
** Empty the contents of the in-memory tree.
*/
void lsmTreeClear(lsm_db *pDb){

  pDb->treehdr.iTransId = 1;
  pDb->treehdr.iRoot = 0;
  pDb->treehdr.nHeight = 0;
  pDb->treehdr.nByte = 0;
}

/*
** This function is called during recovery to initialize the 
** tree header. Only the database connections private copy of the tree-header
** is initialized here - it will be copied into shared memory if log file
** recovery is successful.
*/
int lsmTreeInit(lsm_db *pDb){
  ShmChunk *pOne;
  int rc = LSM_OK;

  pDb->treehdr.iTransId = 1;
  pDb->treehdr.iFirst = 1;
  pDb->treehdr.nChunk = 2;
  pDb->treehdr.iWrite = LSM_SHM_CHUNK_SIZE + LSM_SHM_CHUNK_HDR;
  pDb->treehdr.iNextShmid = 2;
  pDb->treehdr.iUsedShmid = 1;

  pOne = treeShmChunkRc(pDb, 1, &rc);
  if( pOne ){
    pOne->iNext = 0;
    pOne->iShmid = 1;
  }
  return rc;
}

/*
** Insert a new entry into the in-memory tree.
**
** If the value of the 5th parameter, nVal, is negative, then a delete-marker
** is inserted into the tree. In this case the value pointer, pVal, must be
** NULL.
*/
int lsmTreeInsert(
  lsm_db *pDb,                    /* Database handle */
  void *pKey,                     /* Pointer to key data */
  int nKey,                       /* Size of key data in bytes */
  void *pVal,                     /* Pointer to value data (or NULL) */
  int nVal                        /* Bytes in value data (or -ve for delete) */
){
  int rc = LSM_OK;                /* Return Code */
  TreeKey *pTreeKey;              /* New key-value being inserted */

  u32 iTreeKey;

  TreeHeader *pHdr = &pDb->treehdr;

  assert( nVal>=0 || pVal==0 );
  assert_tree_looks_ok(LSM_OK, pTree);
#if 0
  dump_tree_contents(pDb, "before");
#endif

** contents back to their current state.
*/
void lsmTreeMark(lsm_db *pDb, TreeMark *pMark){
  pMark->iRoot = pDb->treehdr.iRoot;
  pMark->nHeight = pDb->treehdr.nHeight;
  pMark->iWrite = pDb->treehdr.iWrite;
  pMark->nChunk = pDb->treehdr.nChunk;

  pMark->iRollback = intArraySize(&pDb->rollback);
}

/*
** Roll back to mark pMark. Structure *pMark should have been previously
** populated by a call to lsmTreeMark().
*/
void lsmTreeRollback(lsm_db *pDb, TreeMark *pMark){
  int rcdummy = LSM_OK;
  int iIdx;
  int nIdx;
  u32 iNext;
  ShmChunk *pChunk;
  u32 iChunk;
  u32 iShmid;

  /* Revert all required v2 pointers. */
  nIdx = intArraySize(&pDb->rollback);
  for(iIdx = pMark->iRollback; iIdx<nIdx; iIdx++){
    TreeNode *pNode;
    pNode = treeShmptr(pDb, intArrayEntry(&pDb->rollback, iIdx), &rcdummy);
    assert( pNode && rcdummy==LSM_OK );
    pNode->iV2 = 0;
    pNode->iV2Child = 0;
    pNode->iV2Ptr = 0;
  }
  intArrayTruncate(&pDb->rollback, pMark->iRollback);

  /* Restore the free-chunk list. */
  assert( pMark->iWrite!=0 );
  iChunk = treeOffsetToChunk(pMark->iWrite-1);
  pChunk = treeShmChunk(pDb, iChunk);
  iNext = pChunk->iNext;
  pChunk->iNext = 0;

  pChunk = treeShmChunk(pDb, pDb->treehdr.iFirst);
  iShmid = pChunk->iShmid-1;


  while( iNext ){
    u32 iFree = iNext;            /* Current chunk being rollback-freed */
    ShmChunk *pFree;              /* Pointer to chunk iFree */

    pFree = treeShmChunk(pDb, iFree);
    iNext = pFree->iNext;

    if( iFree<pMark->nChunk ){
      pFree->iNext = pDb->treehdr.iFirst;
      pFree->iShmid = iShmid--;
      pDb->treehdr.iFirst = iFree;
    }
  }

  /* Restore the tree-header fields */
  pDb->treehdr.iRoot = pMark->iRoot;
  pDb->treehdr.nHeight = pMark->nHeight;
  pDb->treehdr.iWrite = pMark->iWrite;