rf_dagffwr.c

RAIDFrame是个非常好的磁盘阵列RAID仿真工具

      NULL, 1, 0, 0, 0, dag_h, "Nil", allocList);
  }

  rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc, NULL,
    nWndNodes + nfaults, 1, 0, 0, dag_h, "Cmt", allocList);
  rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc, NULL,
    0, nWndNodes + nfaults, 0, 0, dag_h, "Trm", allocList);

  /* initialize the Rod nodes */
  for (nodeNum = asmNum = 0; asmNum < 2; asmNum++) {
    if (new_asm_h[asmNum]) {
      pda = new_asm_h[asmNum]->stripeMap->physInfo;
      while (pda) {
        rf_InitNode(&rodNodes[nodeNum], rf_wait, RF_FALSE, rf_DiskReadFunc,
          rf_DiskReadUndoFunc,rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h,
          "Rod", allocList);
        rodNodes[nodeNum].params[0].p = pda;
        rodNodes[nodeNum].params[1].p = pda->bufPtr;
        rodNodes[nodeNum].params[2].v = parityStripeID;
        rodNodes[nodeNum].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY,
          0, 0, which_ru);
        nodeNum++;
        pda = pda->next;
      }
    }
  }
  RF_ASSERT(nodeNum == nRodNodes);

  /* initialize the wnd nodes */
  pda = asmap->physInfo;
  for (i=0; i < nWndNodes; i++) {
    rf_InitNode(&wndNodes[i], rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
      rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnd", allocList);
    RF_ASSERT(pda != NULL);
    wndNodes[i].params[0].p = pda;
    wndNodes[i].params[1].p = pda->bufPtr;
    wndNodes[i].params[2].v = parityStripeID;
    wndNodes[i].params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
    pda = pda->next;
  }

  /* initialize the redundancy node */
  if (nRodNodes > 0) {
    rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1,
      nRodNodes, 2 * (nWndNodes+nRodNodes) + 1, nfaults, dag_h,
      "Xr ", allocList);
  }
  else {
    rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc, NULL, 1,
      1, 2 * (nWndNodes+nRodNodes) + 1, nfaults, dag_h, "Xr ", allocList);
  }
  xorNode->flags |= RF_DAGNODE_FLAG_YIELD;
  for (i=0; i < nWndNodes; i++) {
    xorNode->params[2*i+0] = wndNodes[i].params[0];         /* pda */
    xorNode->params[2*i+1] = wndNodes[i].params[1];         /* buf ptr */
  }
  for (i=0; i < nRodNodes; i++) {
    xorNode->params[2*(nWndNodes+i)+0] = rodNodes[i].params[0];  /* pda */
    xorNode->params[2*(nWndNodes+i)+1] = rodNodes[i].params[1];  /* buf ptr */
  }
  /* xor node needs to get at RAID information */
  xorNode->params[2*(nWndNodes+nRodNodes)].p = raidPtr;
  
  /*
   * Look for an Rod node that reads a complete SU. If none, alloc a buffer
   * to receive the parity info. Note that we can't use a new data buffer
   * because it will not have gotten written when the xor occurs.
   */
  if (allowBufferRecycle) {
    for (i = 0; i < nRodNodes; i++) {
      if (((RF_PhysDiskAddr_t *)rodNodes[i].params[0].p)->numSector == raidPtr->Layout.sectorsPerStripeUnit)
        break;
    }
  }
  if ((!allowBufferRecycle) || (i == nRodNodes)) {
    RF_CallocAndAdd(xorNode->results[0], 1,
      rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit),
      (void *), allocList);
  }
  else {
    xorNode->results[0] = rodNodes[i].params[1].p;
  }

  /* initialize the Wnp node */
  rf_InitNode(wnpNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
    rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnp", allocList);
  wnpNode->params[0].p = asmap->parityInfo;
  wnpNode->params[1].p = xorNode->results[0];
  wnpNode->params[2].v = parityStripeID;
  wnpNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
  /* parityInfo must describe entire parity unit */
  RF_ASSERT(asmap->parityInfo->next == NULL);

  if (nfaults == 2) {
      /*
       * We never try to recycle a buffer for the Q calcuation
       * in addition to the parity. This would cause two buffers
       * to get smashed during the P and Q calculation, guaranteeing
       * one would be wrong.
       */
      RF_CallocAndAdd(xorNode->results[1], 1,
        rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit),
        (void *),allocList); 
      rf_InitNode(wnqNode, rf_wait, RF_FALSE, rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
        rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Wnq", allocList);
      wnqNode->params[0].p = asmap->qInfo;
      wnqNode->params[1].p = xorNode->results[1];
      wnqNode->params[2].v = parityStripeID;
      wnqNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
      /* parityInfo must describe entire parity unit */
      RF_ASSERT(asmap->parityInfo->next == NULL);
  }

  /*
   * Connect nodes to form graph.
   */

  /* connect dag header to block node */
  RF_ASSERT(blockNode->numAntecedents == 0);
  dag_h->succedents[0] = blockNode;

  if (nRodNodes > 0) {
    /* connect the block node to the Rod nodes */
    RF_ASSERT(blockNode->numSuccedents == nRodNodes);
    RF_ASSERT(xorNode->numAntecedents == nRodNodes);
    for (i = 0; i < nRodNodes; i++) {
      RF_ASSERT(rodNodes[i].numAntecedents == 1);
      blockNode->succedents[i] = &rodNodes[i];
      rodNodes[i].antecedents[0] = blockNode;
      rodNodes[i].antType[0] = rf_control;

      /* connect the Rod nodes to the Xor node */
      RF_ASSERT(rodNodes[i].numSuccedents == 1);
      rodNodes[i].succedents[0] = xorNode;
      xorNode->antecedents[i] = &rodNodes[i];
      xorNode->antType[i] = rf_trueData;
    }
  }
  else {
    /* connect the block node to the Xor node */
    RF_ASSERT(blockNode->numSuccedents == 1);
    RF_ASSERT(xorNode->numAntecedents == 1);
    blockNode->succedents[0] = xorNode;
    xorNode->antecedents[0] = blockNode;
    xorNode->antType[0] = rf_control;
  }

  /* connect the xor node to the commit node */
  RF_ASSERT(xorNode->numSuccedents == 1);
  RF_ASSERT(commitNode->numAntecedents == 1);
  xorNode->succedents[0] = commitNode;
  commitNode->antecedents[0] = xorNode;
  commitNode->antType[0] = rf_control;

  /* connect the commit node to the write nodes */
  RF_ASSERT(commitNode->numSuccedents == nWndNodes + nfaults);
  for (i = 0; i < nWndNodes; i++) {
    RF_ASSERT(wndNodes->numAntecedents == 1);
    commitNode->succedents[i] = &wndNodes[i];
    wndNodes[i].antecedents[0] = commitNode;
    wndNodes[i].antType[0] = rf_control;
  }
  RF_ASSERT(wnpNode->numAntecedents == 1);
  commitNode->succedents[nWndNodes] = wnpNode;
  wnpNode->antecedents[0]= commitNode;
  wnpNode->antType[0] = rf_trueData;
  if (nfaults == 2) {
    RF_ASSERT(wnqNode->numAntecedents == 1);
    commitNode->succedents[nWndNodes + 1] = wnqNode;
    wnqNode->antecedents[0] = commitNode;
    wnqNode->antType[0] = rf_trueData;
  }

  /* connect the write nodes to the term node */
  RF_ASSERT(termNode->numAntecedents == nWndNodes + nfaults);
  RF_ASSERT(termNode->numSuccedents == 0);
  for (i = 0; i < nWndNodes; i++) {
    RF_ASSERT(wndNodes->numSuccedents == 1);
    wndNodes[i].succedents[0] = termNode;
    termNode->antecedents[i] = &wndNodes[i];
    termNode->antType[i] = rf_control;
  }
  RF_ASSERT(wnpNode->numSuccedents == 1);
  wnpNode->succedents[0] = termNode;
  termNode->antecedents[nWndNodes] = wnpNode;
  termNode->antType[nWndNodes] = rf_control;
  if (nfaults == 2) {
    RF_ASSERT(wnqNode->numSuccedents == 1);
    wnqNode->succedents[0] = termNode;
    termNode->antecedents[nWndNodes + 1] = wnqNode;
    termNode->antType[nWndNodes + 1] = rf_control;
  }
}

/******************************************************************************
 *
 * creates a DAG to perform a small-write operation (either raid 5 or pq),
 * which is as follows:
 *
 * Hdr -> Nil -> Rop -> Xor -> Cmt ----> Wnp [Unp] --> Trm
 *            \- Rod X      /     \----> Wnd [Und]-/
 *           [\- Rod X     /       \---> Wnd [Und]-/]
 *           [\- Roq -> Q /         \--> Wnq [Unq]-/]
 *
 * Rop = read old parity
 * Rod = read old data
 * Roq = read old "q"
 * Cmt = commit node
 * Und = unlock data disk
 * Unp = unlock parity disk
 * Unq = unlock q disk
 * Wnp = write new parity
 * Wnd = write new data
 * Wnq = write new "q"
 * [ ] denotes optional segments in the graph
 *
 * Parameters:  raidPtr   - description of the physical array
 *              asmap     - logical & physical addresses for this access
 *              bp        - buffer ptr (holds write data)
 *              flags     - general flags (e.g. disk locking) 
 *              allocList - list of memory allocated in DAG creation
 *              pfuncs    - list of parity generating functions
 *              qfuncs    - list of q generating functions
 *
 * A null qfuncs indicates single fault tolerant
 *****************************************************************************/

void rf_CommonCreateSmallWriteDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList,
  RF_RedFuncs_t         *pfuncs,
  RF_RedFuncs_t         *qfuncs)
{
  RF_DagNode_t *readDataNodes, *readParityNodes, *readQNodes, *termNode;
  RF_DagNode_t *unlockDataNodes, *unlockParityNodes, *unlockQNodes;
  RF_DagNode_t *xorNodes, *qNodes, *blockNode, *commitNode, *nodes;
  RF_DagNode_t *writeDataNodes, *writeParityNodes, *writeQNodes;
  int i, j, nNodes, totalNumNodes, lu_flag;
  RF_ReconUnitNum_t which_ru;
  int (*func)(), (*undoFunc)(), (*qfunc)();
  int numDataNodes, numParityNodes;
  RF_StripeNum_t parityStripeID;
  RF_PhysDiskAddr_t *pda;
  char *name, *qname;
  long nfaults;

  nfaults = qfuncs ? 2 : 1;
  lu_flag = (rf_enableAtomicRMW) ? 1 : 0; /* lock/unlock flag */

  parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout),
    asmap->raidAddress, &which_ru);
  pda = asmap->physInfo;
  numDataNodes = asmap->numStripeUnitsAccessed;
  numParityNodes = (asmap->parityInfo->next) ? 2 : 1;

  if (rf_dagDebug) {
    printf("[Creating small-write DAG]\n");
  }
  RF_ASSERT(numDataNodes > 0);
  dag_h->creator = "SmallWriteDAG";

  dag_h->numCommitNodes = 1;
  dag_h->numCommits = 0;
  dag_h->numSuccedents = 1;

  /*
   * DAG creation occurs in four steps:
   * 1. count the number of nodes in the DAG
   * 2. create the nodes
   * 3. initialize the nodes
   * 4. connect the nodes
   */

  /*
   * Step 1. compute number of nodes in the graph
   */

  /* number of nodes:
   *  a read and write for each data unit
   *  a redundancy computation node for each parity node (nfaults * nparity)
   *  a read and write for each parity unit
   *  a block and commit node (2)
   *  a terminate node
   *  if atomic RMW
   *    an unlock node for each data unit, redundancy unit
   */
  totalNumNodes = (2 * numDataNodes) + (nfaults * numParityNodes)
    + (nfaults * 2 * numParityNodes) + 3;
  if (lu_flag) {
    totalNumNodes += (numDataNodes + (nfaults * numParityNodes));
  }

  /*
   * Step 2. create the nodes
   */
  RF_CallocAndAdd(nodes, totalNumNodes, sizeof(RF_DagNode_t),
    (RF_DagNode_t *), allocList);
  i = 0;
  blockNode        = &nodes[i]; i += 1;
  commitNode       = &nodes[i]; i += 1;
  readDataNodes    = &nodes[i]; i += numDataNodes;
  readParityNodes  = &nodes[i]; i += numParityNodes;
  writeDataNodes   = &nodes[i]; i += numDataNodes;
  writeParityNodes = &nodes[i]; i += numParityNodes;
  xorNodes         = &nodes[i]; i += numParityNodes;
  termNode         = &nodes[i]; i += 1;
  if (lu_flag) {
    unlockDataNodes   = &nodes[i]; i += numDataNodes;
    unlockParityNodes = &nodes[i]; i += numParityNodes;
  }
  else {
    unlockDataNodes = unlockParityNodes = NULL;
  }
  if (nfaults == 2) {
    readQNodes     = &nodes[i]; i += numParityNodes;
    writeQNodes    = &nodes[i]; i += numParityNodes;
    qNodes         = &nodes[i]; i += numParityNodes;
    if (lu_flag) {
      unlockQNodes    = &nodes[i]; i += numParityNodes;
    }
    else {
      unlockQNodes = NULL;
    }
  }
  else {
    readQNodes = writeQNodes = qNodes = unlockQNodes = NULL;
  }
  RF_ASSERT(i == totalNumNodes);
  
  /*
   * Step 3. initialize the nodes
   */
  /* initialize block node (Nil) */
  nNodes     = numDataNodes + (nfaults * numParityNodes);
  rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
    NULL, nNodes, 0, 0, 0, dag_h, "Nil", allocList);

  /* initialize commit node (Cmt) */
  rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
    NULL, nNodes, (nfaults * numParityNodes), 0, 0, dag_h, "Cmt", allocList);

  /* initialize terminate node (Trm) */