root/dev/raidframe/rf_dagdegwr.c

DEFINITIONS

1. RF_CREATE_DAG_FUNC_DECL
2. rf_CreateDegradedWriteDAG
3. rf_CommonCreateSimpleDegradedWriteDAG
4. rf_WriteGenerateFailedAccessASMs
5. rf_DoubleDegSmallWrite

    1 /*      $OpenBSD: rf_dagdegwr.c,v 1.6 2006/07/09 22:10:05 mk Exp $      */
    2 /*      $NetBSD: rf_dagdegwr.c,v 1.5 2000/01/07 03:40:57 oster Exp $    */
    4 /*
    5  * Copyright (c) 1995 Carnegie-Mellon University.
    6  * All rights reserved.
    7  *
    8  * Author: Mark Holland, Daniel Stodolsky, William V. Courtright II
    9  *
   10  * Permission to use, copy, modify and distribute this software and
   11  * its documentation is hereby granted, provided that both the copyright
   12  * notice and this permission notice appear in all copies of the
   13  * software, derivative works or modified versions, and any portions
   14  * thereof, and that both notices appear in supporting documentation.
   15  *
   16  * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
   17  * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
   18  * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
   19  *
   20  * Carnegie Mellon requests users of this software to return to
   21  *
   22  *  Software Distribution Coordinator  or  Software.Distribution@CS.CMU.EDU
   23  *  School of Computer Science
   24  *  Carnegie Mellon University
   25  *  Pittsburgh PA 15213-3890
   26  *
   27  * any improvements or extensions that they make and grant Carnegie the
   28  * rights to redistribute these changes.
   29  */
   31 /*
   32  * rf_dagdegwr.c
   33  *
   34  * Code for creating degraded write DAGs.
   35  *
   36  */
   38 #include "rf_types.h"
   39 #include "rf_raid.h"
   40 #include "rf_dag.h"
   41 #include "rf_dagutils.h"
   42 #include "rf_dagfuncs.h"
   43 #include "rf_debugMem.h"
   44 #include "rf_memchunk.h"
   45 #include "rf_general.h"
   46 #include "rf_dagdegwr.h"
   49 /*****************************************************************************
   50  *
   51  * General comments on DAG creation:
   52  *
   53  * All DAGs in this file use roll-away error recovery. Each DAG has a single
   54  * commit node, usually called "Cmt". If an error occurs before the Cmt node
   55  * is reached, the execution engine will halt forward execution and work
   56  * backward through the graph, executing the undo functions. Assuming that
   57  * each node in the graph prior to the Cmt node are undoable and atomic - or -
   58  * does not make changes to permanent state, the graph will fail atomically.
   59  * If an error occurs after the Cmt node executes, the engine will roll-forward
   60  * through the graph, blindly executing nodes until it reaches the end.
   61  * If a graph reaches the end, it is assumed to have completed successfully.
   62  *
   63  * A graph has only 1 Cmt node.
   64  *
   65  *****************************************************************************/
   68 /*****************************************************************************
   69  *
   70  * The following wrappers map the standard DAG creation interface to the
   71  * DAG creation routines. Additionally, these wrappers enable experimentation
   72  * with new DAG structures by providing an extra level of indirection, allowing
   73  * the DAG creation routines to be replaced at this single point.
   74  *
   75  *****************************************************************************/
RF_CREATE_DAG_FUNC_DECL(rf_CreateSimpleDegradedWriteDAG)
   78 {
rf_CommonCreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp,
flags, allocList, 1, rf_RecoveryXorFunc, RF_TRUE);
   81 }
   83 void
rf_CreateDegradedWriteDAG(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap,
RF_DagHeader_t *dag_h, void *bp, RF_RaidAccessFlags_t flags,
RF_AllocListElem_t *allocList)
   87 {
RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
RF_PhysDiskAddr_t *failedPDA = asmap->failedPDAs[0];
RF_ASSERT(asmap->numDataFailed == 1);
dag_h->creator = "DegradedWriteDAG";
   94         /*
   95          * If the access writes only a portion of the failed unit, and also
   96          * writes some portion of at least one surviving unit, we create two
   97          * DAGs, one for the failed component and one for the non-failed
   98          * component, and do them sequentially. Note that the fact that we're
   99          * accessing only a portion of the failed unit indicates that the
  100          * access either starts or ends in the failed unit, and hence we need
  101          * create only two dags. This is inefficient in that the same data or
  102          * parity can get read and written twice using this structure. I need
  103          * to fix this to do the access all at once.
  104          */
RF_ASSERT(!(asmap->numStripeUnitsAccessed != 1 &&
failedPDA->numSector != layoutPtr->sectorsPerStripeUnit));
rf_CreateSimpleDegradedWriteDAG(raidPtr, asmap, dag_h, bp,
flags, allocList);
  109 }
  113 /*****************************************************************************
  114  *
  115  * DAG creation code begins here.
  116  *
  117  *****************************************************************************/
  120 /*****************************************************************************
  121  *
  122  * CommonCreateSimpleDegradedWriteDAG -- creates a DAG to do a degraded-mode
  123  * write, which is as follows
  124  *
  125  *                                        / {Wnq} --\
  126  * hdr -> blockNode ->  Rod -> Xor -> Cmt -> Wnp ----> unblock -> term
  127  *                  \  {Rod} /            |  Wnd ---/
  128  *                                        \ {Wnd} -/
  129  *
  130  * Commit nodes: Xor, Wnd
  131  *
  132  * IMPORTANT:
  133  * This DAG generator does not work for double-degraded archs since it does not
  134  * generate Q.
  135  *
  136  * This dag is essentially identical to the large-write dag, except that the
  137  * write to the failed data unit is suppressed.
  138  *
  139  * IMPORTANT:  this dag does not work in the case where the access writes only
  140  * a portion of the failed unit, and also writes some portion of at least one
  141  * surviving SU. this case is handled in CreateDegradedWriteDAG above.
  142  *
  143  * The block & unblock nodes are leftovers from a previous version. They
  144  * do nothing, but I haven't deleted them because it would be a tremendous
  145  * effort to put them back in.
  146  *
  147  * This dag is used whenever one of the data units in a write has failed.
  148  * If it is the parity unit that failed, the nonredundant write dag (below)
  149  * is used.
  150  *
  151  *****************************************************************************/
  153 void
rf_CommonCreateSimpleDegradedWriteDAG(RF_Raid_t *raidPtr,
RF_AccessStripeMap_t *asmap, RF_DagHeader_t *dag_h, void *bp,
RF_RaidAccessFlags_t flags, RF_AllocListElem_t *allocList, int nfaults,
  157     int (*redFunc) (RF_DagNode_t *), int allowBufferRecycle)
  158 {
  159         int nNodes, nRrdNodes, nWndNodes, nXorBufs, i, j, paramNum,
rdnodesFaked;
RF_DagNode_t *blockNode, *unblockNode, *wnpNode, *wnqNode, *termNode;
RF_DagNode_t *nodes, *wndNodes, *rrdNodes, *xorNode, *commitNode;
RF_SectorCount_t sectorsPerSU;
RF_ReconUnitNum_t which_ru;
  165         char *xorTargetBuf = NULL;      /*
  166                                          * The target buffer for the XOR
  167                                          * operation.
  168                                          */
  169         char *overlappingPDAs;          /* A temporary array of flags. */
RF_AccessStripeMapHeader_t *new_asm_h[2];
RF_PhysDiskAddr_t *pda, *parityPDA;
RF_StripeNum_t parityStripeID;
RF_PhysDiskAddr_t *failedPDA;
RF_RaidLayout_t *layoutPtr;
layoutPtr = &(raidPtr->Layout);
parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr,
asmap->raidAddress, &which_ru);
sectorsPerSU = layoutPtr->sectorsPerStripeUnit;
  180         /*
  181          * failedPDA points to the pda within the asm that targets
  182          * the failed disk.
  183          */
failedPDA = asmap->failedPDAs[0];
  186         if (rf_dagDebug)
printf("[Creating degraded-write DAG]\n");
RF_ASSERT(asmap->numDataFailed == 1);
dag_h->creator = "SimpleDegradedWriteDAG";
  192         /*
  193          * Generate two ASMs identifying the surviving data
  194          * we need in order to recover the lost data.
  195          */
  196         /* overlappingPDAs array must be zero'd */
RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed,
  198             sizeof(char), (char *));
rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h,
new_asm_h, &nXorBufs, NULL, overlappingPDAs, allocList);
  202         /* Create all the nodes at once. */
nWndNodes = asmap->numStripeUnitsAccessed - 1;  /*
  204                                                          * No access is
  205                                                          * generated for the
  206                                                          * failed pda.
  207                                                          */
nRrdNodes = ((new_asm_h[0]) ?
new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) +
  211             ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed
  212                             : 0);
  213         /*
  214          * XXX
  215          *
  216          * There's a bug with a complete stripe overwrite- that means 0 reads
  217          * of old data, and the rest of the DAG generation code doesn't like
  218          * that. A release is coming, and I don't wanna risk breaking a
  219          * critical DAG generator, so here's what I'm gonna do- if there's
  220          * no read nodes, I'm gonna fake there being a read node, and I'm
  221          * gonna swap in a no-op node in its place (to make all the link-up
  222          * code happy).
  223          * This should be fixed at some point. --jimz
  224          */
  225         if (nRrdNodes == 0) {
nRrdNodes = 1;
rdnodesFaked = 1;
  228         } else {
rdnodesFaked = 0;
  230         }
  231         /* Lock, unlock, xor, Wnd, Rrd, W(nfaults). */
nNodes = 5 + nfaults + nWndNodes + nRrdNodes;
RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t),
  234             (RF_DagNode_t *), allocList);
i = 0;
blockNode = &nodes[i];
i += 1;
commitNode = &nodes[i];
i += 1;
unblockNode = &nodes[i];
i += 1;
termNode = &nodes[i];
i += 1;
xorNode = &nodes[i];
i += 1;
wnpNode = &nodes[i];
i += 1;
wndNodes = &nodes[i];
i += nWndNodes;
rrdNodes = &nodes[i];
i += nRrdNodes;
  252         if (nfaults == 2) {
wnqNode = &nodes[i];
i += 1;
  255         } else {
wnqNode = NULL;
  257         }
RF_ASSERT(i == nNodes);
  260         /*
  261          * This dag can not commit until all rrd and xor Nodes have
  262          * completed.
  263          */
dag_h->numCommitNodes = 1;
dag_h->numCommits = 0;
dag_h->numSuccedents = 1;
RF_ASSERT(nRrdNodes > 0);
rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, nRrdNodes, 0, 0, 0, dag_h,
  271             "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(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, 1, nWndNodes + nfaults, 0, 0,
dag_h, "Nil", allocList);
rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc,
rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);
rf_InitNode(xorNode, rf_wait, RF_FALSE, redFunc, rf_NullNodeUndoFunc,
NULL, 1, nRrdNodes, 2 * nXorBufs + 2, nfaults, dag_h, "Xrc",
allocList);
  284         /*
  285          * Fill in the Rrd nodes. If any of the rrd buffers are the same size
  286          * as the failed buffer, save a pointer to it so we can use it as the
  287          * target of the XOR. The pdas in the rrd nodes have been range-
  288          * restricted, so if a buffer is the same size as the failed buffer,
  289          * it must also be at the same alignment within the SU.
  290          */
i = 0;
  292         if (new_asm_h[0]) {
  293                 for (i = 0, pda = new_asm_h[0]->stripeMap->physInfo;
i < new_asm_h[0]->stripeMap->numStripeUnitsAccessed;
i++, pda = pda->next) {
rf_InitNode(&rrdNodes[i], rf_wait, RF_FALSE,
rf_DiskReadFunc, rf_DiskReadUndoFunc,
rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h,
  299                             "Rrd", allocList);
RF_ASSERT(pda);
rrdNodes[i].params[0].p = pda;
rrdNodes[i].params[1].p = pda->bufPtr;
rrdNodes[i].params[2].v = parityStripeID;
rrdNodes[i].params[3].v = RF_CREATE_PARAM3(
RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
  306                 }
  307         }
  308         /* i now equals the number of stripe units accessed in new_asm_h[0]. */
  309         if (new_asm_h[1]) {
  310                 for (j = 0, pda = new_asm_h[1]->stripeMap->physInfo;
j < new_asm_h[1]->stripeMap->numStripeUnitsAccessed;
j++, pda = pda->next) {
rf_InitNode(&rrdNodes[i + j], rf_wait, RF_FALSE,
rf_DiskReadFunc, rf_DiskReadUndoFunc,
rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h,
  316                             "Rrd", allocList);
RF_ASSERT(pda);
rrdNodes[i + j].params[0].p = pda;
rrdNodes[i + j].params[1].p = pda->bufPtr;
rrdNodes[i + j].params[2].v = parityStripeID;
rrdNodes[i + j].params[3].v = RF_CREATE_PARAM3(
RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
  323                         if (allowBufferRecycle &&
  324                             (pda->numSector == failedPDA->numSector))
xorTargetBuf = pda->bufPtr;
  326                 }
  327         }
  328         if (rdnodesFaked) {
  329                 /*
  330                  * This is where we'll init that fake noop read node.
  331                  * (XXX should the wakeup func be different ?)
  332                  */
rf_InitNode(&rrdNodes[0], rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, 1, 1, 0, 0, dag_h, "RrN",
allocList);
  336         }
  337         /*
  338          * Make a PDA for the parity unit. The parity PDA should start at
  339          * the same offset into the SU as the failed PDA.
  340          */
  341         /*
  342          * Danner comment: I don't think this copy is really necessary. We are
  343          * in one of two cases here.
  344          * (1) The entire failed unit is written. Then asmap->parityInfo will
  345          *     describe the entire parity.
  346          * (2) We are only writing a subset of the failed unit and nothing else.
  347          *     Then the asmap->parityInfo describes the failed unit and the copy
  348          *     can also be avoided.
  349          */
RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t),
  352             (RF_PhysDiskAddr_t *), allocList);
parityPDA->row = asmap->parityInfo->row;
parityPDA->col = asmap->parityInfo->col;
parityPDA->startSector = ((asmap->parityInfo->startSector /
sectorsPerSU) * sectorsPerSU) + (failedPDA->startSector %
sectorsPerSU);
parityPDA->numSector = failedPDA->numSector;
  360         if (!xorTargetBuf) {
RF_CallocAndAdd(xorTargetBuf, 1, rf_RaidAddressToByte(raidPtr,
failedPDA->numSector), (char *), allocList);
  363         }
  364         /* Init 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 = parityPDA;
wnpNode->params[1].p = xorTargetBuf;
wnpNode->params[2].v = parityStripeID;
wnpNode->params[3].v = RF_CREATE_PARAM3(
RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
  374         /* Fill in the Wnq Node. */
  375         if (nfaults == 2) {
  376                 {
RF_MallocAndAdd(parityPDA, sizeof(RF_PhysDiskAddr_t),
  378                             (RF_PhysDiskAddr_t *), allocList);
parityPDA->row = asmap->qInfo->row;
parityPDA->col = asmap->qInfo->col;
parityPDA->startSector = ((asmap->qInfo->startSector /
sectorsPerSU) * sectorsPerSU) +
  383                             (failedPDA->startSector % sectorsPerSU);
parityPDA->numSector = failedPDA->numSector;
rf_InitNode(wnqNode, rf_wait, RF_FALSE,
rf_DiskWriteFunc, rf_DiskWriteUndoFunc,
rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h,
  389                             "Wnq", allocList);
wnqNode->params[0].p = parityPDA;
RF_CallocAndAdd(xorNode->results[1], 1,
rf_RaidAddressToByte(raidPtr, failedPDA->numSector),
  393                             (char *), allocList);
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);
  398                 }
  399         }
  400         /* Fill in the Wnd nodes. */
  401         for (pda = asmap->physInfo, i = 0; i < nWndNodes;
i++, pda = pda->next) {
  403                 if (pda == failedPDA) {
i--;
  405                         continue;
  406                 }
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);
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);
  416         }
  418         /* Fill in the results of the xor node. */
xorNode->results[0] = xorTargetBuf;
  421         /* Fill in the params of the xor node. */
paramNum = 0;
  424         if (rdnodesFaked == 0) {
  425                 for (i = 0; i < nRrdNodes; i++) {
  426                         /* All the Rrd nodes need to be xored together. */
xorNode->params[paramNum++] = rrdNodes[i].params[0];
xorNode->params[paramNum++] = rrdNodes[i].params[1];
  429                 }
  430         }
  431         for (i = 0; i < nWndNodes; i++) {
  432                 /*
  433                  * Any Wnd nodes that overlap the failed access need to be
  434                  * xored in.
  435                  */
  436                 if (overlappingPDAs[i]) {
RF_MallocAndAdd(pda, sizeof(RF_PhysDiskAddr_t),
  438                             (RF_PhysDiskAddr_t *), allocList);
bcopy((char *) wndNodes[i].params[0].p, (char *) pda,
  440                             sizeof(RF_PhysDiskAddr_t));
rf_RangeRestrictPDA(raidPtr, failedPDA, pda,
RF_RESTRICT_DOBUFFER, 0);
xorNode->params[paramNum++].p = pda;
xorNode->params[paramNum++].p = pda->bufPtr;
  445                 }
  446         }
RF_Free(overlappingPDAs, asmap->numStripeUnitsAccessed * sizeof(char));
  449         /*
  450          * Install the failed PDA into the xor param list so that the
  451          * new data gets xor'd in.
  452          */
xorNode->params[paramNum++].p = failedPDA;
xorNode->params[paramNum++].p = failedPDA->bufPtr;
  456         /*
  457          * The last 2 params to the recovery xor node are always the failed
  458          * PDA and the raidPtr. Install the failedPDA even though we have just
  459          * done so above. This allows us to use the same XOR function for both
  460          * degraded reads and degraded writes.
  461          */
xorNode->params[paramNum++].p = failedPDA;
xorNode->params[paramNum++].p = raidPtr;
RF_ASSERT(paramNum == 2 * nXorBufs + 2);
  466         /*
  467          * Code to link nodes begins here.
  468          */
  470         /* Link header to block node. */
RF_ASSERT(blockNode->numAntecedents == 0);
dag_h->succedents[0] = blockNode;
  474         /* Link block node to rd nodes. */
RF_ASSERT(blockNode->numSuccedents == nRrdNodes);
  476         for (i = 0; i < nRrdNodes; i++) {
RF_ASSERT(rrdNodes[i].numAntecedents == 1);
blockNode->succedents[i] = &rrdNodes[i];
rrdNodes[i].antecedents[0] = blockNode;
rrdNodes[i].antType[0] = rf_control;
  481         }
  483         /* Link read nodes to xor node. */
RF_ASSERT(xorNode->numAntecedents == nRrdNodes);
  485         for (i = 0; i < nRrdNodes; i++) {
RF_ASSERT(rrdNodes[i].numSuccedents == 1);
rrdNodes[i].succedents[0] = xorNode;
xorNode->antecedents[i] = &rrdNodes[i];
xorNode->antType[i] = rf_trueData;
  490         }
  492         /* Link xor node to 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;
  499         /* Link commit node to wnd nodes. */
RF_ASSERT(commitNode->numSuccedents == nfaults + nWndNodes);
  501         for (i = 0; i < nWndNodes; i++) {
RF_ASSERT(wndNodes[i].numAntecedents == 1);
commitNode->succedents[i] = &wndNodes[i];
wndNodes[i].antecedents[0] = commitNode;
wndNodes[i].antType[0] = rf_control;
  506         }
  508         /* Link the commit node to wnp, wnq nodes. */
RF_ASSERT(wnpNode->numAntecedents == 1);
commitNode->succedents[nWndNodes] = wnpNode;
wnpNode->antecedents[0] = commitNode;
wnpNode->antType[0] = rf_control;
  513         if (nfaults == 2) {
RF_ASSERT(wnqNode->numAntecedents == 1);
commitNode->succedents[nWndNodes + 1] = wnqNode;
wnqNode->antecedents[0] = commitNode;
wnqNode->antType[0] = rf_control;
  518         }
  519         /* Link write new data nodes to unblock node. */
RF_ASSERT(unblockNode->numAntecedents == (nWndNodes + nfaults));
  521         for (i = 0; i < nWndNodes; i++) {
RF_ASSERT(wndNodes[i].numSuccedents == 1);
wndNodes[i].succedents[0] = unblockNode;
unblockNode->antecedents[i] = &wndNodes[i];
unblockNode->antType[i] = rf_control;
  526         }
  528         /* Link write new parity node to unblock node. */
RF_ASSERT(wnpNode->numSuccedents == 1);
wnpNode->succedents[0] = unblockNode;
unblockNode->antecedents[nWndNodes] = wnpNode;
unblockNode->antType[nWndNodes] = rf_control;
  534         /* Link write new q node to unblock node. */
  535         if (nfaults == 2) {
RF_ASSERT(wnqNode->numSuccedents == 1);
wnqNode->succedents[0] = unblockNode;
unblockNode->antecedents[nWndNodes + 1] = wnqNode;
unblockNode->antType[nWndNodes + 1] = rf_control;
  540         }
  541         /* Link unblock node to term node. */
RF_ASSERT(unblockNode->numSuccedents == 1);
RF_ASSERT(termNode->numAntecedents == 1);
RF_ASSERT(termNode->numSuccedents == 0);
unblockNode->succedents[0] = termNode;
termNode->antecedents[0] = unblockNode;
termNode->antType[0] = rf_control;
  548 }
  550 #define CONS_PDA(if,start,num)  do {                                    \
pda_p->row = asmap->if->row;                                    \
pda_p->col = asmap->if->col;                                    \
pda_p->startSector = ((asmap->if->startSector / secPerSU) *     \
secPerSU) + start;                                          \
pda_p->numSector = num;                                         \
pda_p->next = NULL;                                             \
RF_MallocAndAdd(pda_p->bufPtr,                                  \
rf_RaidAddressToByte(raidPtr,num),(char *), allocList);     \
  559 } while (0)
  561 void
rf_WriteGenerateFailedAccessASMs(RF_Raid_t *raidPtr,
RF_AccessStripeMap_t *asmap, RF_PhysDiskAddr_t **pdap, int *nNodep,
RF_PhysDiskAddr_t **pqpdap, int *nPQNodep, RF_AllocListElem_t *allocList)
  565 {
RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
  567         int PDAPerDisk, i;
RF_SectorCount_t secPerSU = layoutPtr->sectorsPerStripeUnit;
  569         int numDataCol = layoutPtr->numDataCol;
  570         int state;
  571         unsigned napdas;
RF_SectorNum_t fone_start, fone_end, ftwo_start = 0, ftwo_end;
RF_PhysDiskAddr_t *fone = asmap->failedPDAs[0];
RF_PhysDiskAddr_t *ftwo = asmap->failedPDAs[1];
RF_PhysDiskAddr_t *pda_p;
RF_RaidAddr_t sosAddr;
  578         /*
  579          * Determine how many pda's we will have to generate per unaccessed
  580          * stripe. If there is only one failed data unit, it is one; if two,
  581          * possibly two, depending whether they overlap.
  582          */
fone_start = rf_StripeUnitOffset(layoutPtr, fone->startSector);
fone_end = fone_start + fone->numSector;
  587         if (asmap->numDataFailed == 1) {
PDAPerDisk = 1;
state = 1;
RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t),
  591                     (RF_PhysDiskAddr_t *), allocList);
pda_p = *pqpdap;
  593                 /* Build p. */
CONS_PDA(parityInfo, fone_start, fone->numSector);
pda_p->type = RF_PDA_TYPE_PARITY;
pda_p++;
  597                 /* Build q. */
CONS_PDA(qInfo, fone_start, fone->numSector);
pda_p->type = RF_PDA_TYPE_Q;
  600         } else {
ftwo_start = rf_StripeUnitOffset(layoutPtr, ftwo->startSector);
ftwo_end = ftwo_start + ftwo->numSector;
  603                 if (fone->numSector + ftwo->numSector > secPerSU) {
PDAPerDisk = 1;
state = 2;
RF_MallocAndAdd(*pqpdap, 2 * sizeof(RF_PhysDiskAddr_t),
  607                             (RF_PhysDiskAddr_t *), allocList);
pda_p = *pqpdap;
CONS_PDA(parityInfo, 0, secPerSU);
pda_p->type = RF_PDA_TYPE_PARITY;
pda_p++;
CONS_PDA(qInfo, 0, secPerSU);
pda_p->type = RF_PDA_TYPE_Q;
  614                 } else {
PDAPerDisk = 2;
state = 3;
  617                         /* Four of them, fone, then ftwo. */
RF_MallocAndAdd(*pqpdap, 4 * sizeof(RF_PhysDiskAddr_t),
  619                             (RF_PhysDiskAddr_t *), allocList);
pda_p = *pqpdap;
CONS_PDA(parityInfo, fone_start, fone->numSector);
pda_p->type = RF_PDA_TYPE_PARITY;
pda_p++;
CONS_PDA(qInfo, fone_start, fone->numSector);
pda_p->type = RF_PDA_TYPE_Q;
pda_p++;
CONS_PDA(parityInfo, ftwo_start, ftwo->numSector);
pda_p->type = RF_PDA_TYPE_PARITY;
pda_p++;
CONS_PDA(qInfo, ftwo_start, ftwo->numSector);
pda_p->type = RF_PDA_TYPE_Q;
  632                 }
  633         }
  634         /* Figure out number of nonaccessed pda. */
napdas = PDAPerDisk * (numDataCol - 2);
  636         *nPQNodep = PDAPerDisk;
  638         *nNodep = napdas;
  639         if (napdas == 0)
  640                 return;         /* Short circuit. */
  642         /* Allocate up our list of pda's. */
RF_CallocAndAdd(pda_p, napdas, sizeof(RF_PhysDiskAddr_t),
  645             (RF_PhysDiskAddr_t *), allocList);
  646         *pdap = pda_p;
  648         /* Link them together. */
  649         for (i = 0; i < (napdas - 1); i++)
pda_p[i].next = pda_p + (i + 1);
sosAddr = rf_RaidAddressOfPrevStripeBoundary(layoutPtr,
asmap->raidAddress);
  654         for (i = 0; i < numDataCol; i++) {
  655                 if ((pda_p - (*pdap)) == napdas)
  656                         continue;
pda_p->type = RF_PDA_TYPE_DATA;
pda_p->raidAddress = sosAddr + (i * secPerSU);
  659                 (raidPtr->Layout.map->MapSector) (raidPtr, pda_p->raidAddress,
  660                     &(pda_p->row), &(pda_p->col), &(pda_p->startSector), 0);
  661                 /* Skip over dead disks. */
  662                 if (RF_DEAD_DISK(raidPtr->Disks[pda_p->row][pda_p->col].status))
  663                         continue;
  664                 switch (state) {
  665                 case 1: /* Fone. */
pda_p->numSector = fone->numSector;
pda_p->raidAddress += fone_start;
pda_p->startSector += fone_start;
RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(
raidPtr, pda_p->numSector), (char *), allocList);
  671                         break;
  672                 case 2: /* Full stripe. */
pda_p->numSector = secPerSU;
RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(
raidPtr, secPerSU), (char *), allocList);
  676                         break;
  677                 case 3: /* Two slabs. */
pda_p->numSector = fone->numSector;
pda_p->raidAddress += fone_start;
pda_p->startSector += fone_start;
RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(
raidPtr, pda_p->numSector), (char *), allocList);
pda_p++;
pda_p->type = RF_PDA_TYPE_DATA;
pda_p->raidAddress = sosAddr + (i * secPerSU);
  686                         (raidPtr->Layout.map->MapSector) (raidPtr,
pda_p->raidAddress, &(pda_p->row), &(pda_p->col),
  688                             &(pda_p->startSector), 0);
pda_p->numSector = ftwo->numSector;
pda_p->raidAddress += ftwo_start;
pda_p->startSector += ftwo_start;
RF_MallocAndAdd(pda_p->bufPtr, rf_RaidAddressToByte(
raidPtr, pda_p->numSector), (char *), allocList);
  694                         break;
  695                 default:
RF_PANIC();
  697                 }
pda_p++;
  699         }
RF_ASSERT(pda_p - *pdap == napdas);
  702         return;
  703 }
  705 #define DISK_NODE_PDA(node)     ((node)->params[0].p)
  707 #define DISK_NODE_PARAMS(_node_,_p_)    do {                            \
  708         (_node_).params[0].p = _p_ ;                                    \
  709         (_node_).params[1].p = (_p_)->bufPtr;                           \
  710         (_node_).params[2].v = parityStripeID;                          \
  711         (_node_).params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY,  \
  712             0, 0, which_ru);                                            \
  713 } while (0)
  715 void
rf_DoubleDegSmallWrite(RF_Raid_t *raidPtr, RF_AccessStripeMap_t *asmap,
RF_DagHeader_t *dag_h, void *bp, RF_RaidAccessFlags_t flags,
RF_AllocListElem_t *allocList, char *redundantReadNodeName,
  719     char *redundantWriteNodeName, char *recoveryNodeName,
  720     int (*recovFunc) (RF_DagNode_t *))
  721 {
RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
RF_DagNode_t *nodes, *wudNodes, *rrdNodes, *recoveryNode, *blockNode,
  724             *unblockNode, *rpNodes, *rqNodes, *wpNodes, *wqNodes, *termNode;
RF_PhysDiskAddr_t *pda, *pqPDAs;
RF_PhysDiskAddr_t *npdas;
  727         int nWriteNodes, nNodes, nReadNodes, nRrdNodes, nWudNodes, i;
RF_ReconUnitNum_t which_ru;
  729         int nPQNodes;
RF_StripeNum_t parityStripeID = rf_RaidAddressToParityStripeID(
layoutPtr, asmap->raidAddress, &which_ru);
  733         /*
  734          * Simple small write case - First part looks like a reconstruct-read
  735          * of the failed data units. Then a write of all data units not
  736          * failed.
  737          */
  740         /*
  741          * Hdr | ------Block- /  /         \   Rrd  Rrd ...  Rrd  Rp Rq \  \
  742          * /  -------PQ----- /   \   \ Wud   Wp  WQ          \    |   /
  743          * --Unblock- | T
  744          *
  745          * Rrd = read recovery data (potentially none)
  746          * Wud = write user data (not incl. failed disks)
  747          * Wp = Write P (could be two)
  748          * Wq = Write Q (could be two)
  749          *
  750          */
rf_WriteGenerateFailedAccessASMs(raidPtr, asmap, &npdas, &nRrdNodes,
  753             &pqPDAs, &nPQNodes, allocList);
RF_ASSERT(asmap->numDataFailed == 1);
nWudNodes = asmap->numStripeUnitsAccessed - (asmap->numDataFailed);
nReadNodes = nRrdNodes + 2 * nPQNodes;
nWriteNodes = nWudNodes + 2 * nPQNodes;
nNodes = 4 + nReadNodes + nWriteNodes;
RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *),
allocList);
blockNode = nodes;
unblockNode = blockNode + 1;
termNode = unblockNode + 1;
recoveryNode = termNode + 1;
rrdNodes = recoveryNode + 1;
rpNodes = rrdNodes + nRrdNodes;
rqNodes = rpNodes + nPQNodes;
wudNodes = rqNodes + nPQNodes;
wpNodes = wudNodes + nWudNodes;
wqNodes = wpNodes + nPQNodes;
dag_h->creator = "PQ_DDSimpleSmallWrite";
dag_h->numSuccedents = 1;
dag_h->succedents[0] = blockNode;
rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc,
rf_TerminateUndoFunc, NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);
termNode->antecedents[0] = unblockNode;
termNode->antType[0] = rf_control;
  783         /* Init the block and unblock nodes. */
  784         /* The block node has all the read nodes as successors. */
rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, nReadNodes, 0, 0, 0, dag_h,
  787             "Nil", allocList);
  788         for (i = 0; i < nReadNodes; i++)
blockNode->succedents[i] = rrdNodes + i;
  791         /* The unblock node has all the writes as successors. */
rf_InitNode(unblockNode, rf_wait, RF_FALSE, rf_NullNodeFunc,
rf_NullNodeUndoFunc, NULL, 1, nWriteNodes, 0, 0, dag_h,
  794             "Nil", allocList);
  795         for (i = 0; i < nWriteNodes; i++) {
unblockNode->antecedents[i] = wudNodes + i;
unblockNode->antType[i] = rf_control;
  798         }
unblockNode->succedents[0] = termNode;
  801 #define INIT_READ_NODE(node,name)       do {                            \
rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskReadFunc,           \
rf_DiskReadUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0,      \
dag_h, name, allocList);                                    \
  805         (node)->succedents[0] = recoveryNode;                           \
  806         (node)->antecedents[0] = blockNode;                             \
  807         (node)->antType[0] = rf_control;                                \
  808 } while (0)
  810         /* Build the read nodes. */
pda = npdas;
  812         for (i = 0; i < nRrdNodes; i++, pda = pda->next) {
INIT_READ_NODE(rrdNodes + i, "rrd");
DISK_NODE_PARAMS(rrdNodes[i], pda);
  815         }
  817         /* Read redundancy pdas. */
pda = pqPDAs;
INIT_READ_NODE(rpNodes, "Rp");
RF_ASSERT(pda);
DISK_NODE_PARAMS(rpNodes[0], pda);
pda++;
INIT_READ_NODE(rqNodes, redundantReadNodeName);
RF_ASSERT(pda);
DISK_NODE_PARAMS(rqNodes[0], pda);
  826         if (nPQNodes == 2) {
pda++;
INIT_READ_NODE(rpNodes + 1, "Rp");
RF_ASSERT(pda);
DISK_NODE_PARAMS(rpNodes[1], pda);
pda++;
INIT_READ_NODE(rqNodes + 1, redundantReadNodeName);
RF_ASSERT(pda);
DISK_NODE_PARAMS(rqNodes[1], pda);
  835         }
  836         /*
  837          * The recovery node has all reads as precedessors and all writes as
  838          * successors. It generates a result for every write P or write Q
  839          * node. As parameters, it takes a pda per read and a pda per stripe
  840          * of user data written. It also takes as the last params the raidPtr
  841          * and asm. For results, it takes PDA for P & Q.
  842          */
rf_InitNode(recoveryNode, rf_wait, RF_FALSE, recovFunc,
rf_NullNodeUndoFunc, NULL,
nWriteNodes,                /* succesors */
nReadNodes,                 /* preds */
nReadNodes + nWudNodes + 3, /* params */
  849             2 * nPQNodes,               /* results */
dag_h, recoveryNodeName, allocList);
  854         for (i = 0; i < nReadNodes; i++) {
recoveryNode->antecedents[i] = rrdNodes + i;
recoveryNode->antType[i] = rf_control;
recoveryNode->params[i].p = DISK_NODE_PDA(rrdNodes + i);
  858         }
  859         for (i = 0; i < nWudNodes; i++) {
recoveryNode->succedents[i] = wudNodes + i;
  861         }
recoveryNode->params[nReadNodes + nWudNodes].p = asmap->failedPDAs[0];
recoveryNode->params[nReadNodes + nWudNodes + 1].p = raidPtr;
recoveryNode->params[nReadNodes + nWudNodes + 2].p = asmap;
  866         for (; i < nWriteNodes; i++)
recoveryNode->succedents[i] = wudNodes + i;
pda = pqPDAs;
recoveryNode->results[0] = pda;
pda++;
recoveryNode->results[1] = pda;
  873         if (nPQNodes == 2) {
pda++;
recoveryNode->results[2] = pda;
pda++;
recoveryNode->results[3] = pda;
  878         }
  879         /* Fill writes. */
  880 #define INIT_WRITE_NODE(node,name)      do {                            \
rf_InitNode(node, rf_wait, RF_FALSE, rf_DiskWriteFunc,          \
rf_DiskWriteUndoFunc, rf_GenericWakeupFunc, 1, 1, 4, 0,     \
dag_h, name, allocList);                                    \
  884         (node)->succedents[0] = unblockNode;                            \
  885         (node)->antecedents[0] = recoveryNode;                          \
  886         (node)->antType[0] = rf_control;                                \
  887 } while (0)
pda = asmap->physInfo;
  890         for (i = 0; i < nWudNodes; i++) {
INIT_WRITE_NODE(wudNodes + i, "Wd");
DISK_NODE_PARAMS(wudNodes[i], pda);
recoveryNode->params[nReadNodes + i].p =
DISK_NODE_PDA(wudNodes + i);
pda = pda->next;
  896         }
  897         /* Write redundancy pdas. */
pda = pqPDAs;
INIT_WRITE_NODE(wpNodes, "Wp");
RF_ASSERT(pda);
DISK_NODE_PARAMS(wpNodes[0], pda);
pda++;
INIT_WRITE_NODE(wqNodes, "Wq");
RF_ASSERT(pda);
DISK_NODE_PARAMS(wqNodes[0], pda);
  906         if (nPQNodes == 2) {
pda++;
INIT_WRITE_NODE(wpNodes + 1, "Wp");
RF_ASSERT(pda);
DISK_NODE_PARAMS(wpNodes[1], pda);
pda++;
INIT_WRITE_NODE(wqNodes + 1, "Wq");
RF_ASSERT(pda);
DISK_NODE_PARAMS(wqNodes[1], pda);
  915         }
  916 }