rf_dagdegrd.c

RAIDFrame是个非常好的磁盘阵列RAID仿真工具

/*
 * Copyright (c) 1995 Carnegie-Mellon University.
 * All rights reserved.
 *
 * Author: Mark Holland, Daniel Stodolsky, William V. Courtright II
 *
 * Permission to use, copy, modify and distribute this software and
 * its documentation is hereby granted, provided that both the copyright
 * notice and this permission notice appear in all copies of the
 * software, derivative works or modified versions, and any portions
 * thereof, and that both notices appear in supporting documentation.
 *
 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
 * CONDITION.  CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
 *
 * Carnegie Mellon requests users of this software to return to
 *
 *  Software Distribution Coordinator  or  [e-mail]
 *  School of Computer Science
 *  Carnegie Mellon University
 *  Pittsburgh PA 15213-3890
 *
 * any improvements or extensions that they make and grant Carnegie the
 * rights to redistribute these changes.
 */

/*
 * rf_dagdegrd.c
 *
 * code for creating degraded read DAGs
 *
 * $Locker:  $
 * $Log: rf_dagdegrd.c,v $
 * Revision 1.19  1996/08/19  23:30:36  jimz
 * fix chained declustered accesses in degraded mode when mirror copy is failed
 * (workload shifting not allowed when there are no duplicate copies extant)
 *
 * Revision 1.18  1996/07/31  16:29:01  jimz
 * asm/asmap re-fix (EO merge)
 *
 * Revision 1.17  1996/07/31  15:34:34  jimz
 * evenodd changes; bugfixes for double-degraded archs, generalize
 * some formerly PQ-only functions
 *
 * Revision 1.16  1996/07/28  20:31:39  jimz
 * i386netbsd port
 * true/false fixup
 *
 * Revision 1.15  1996/07/27  23:36:08  jimz
 * Solaris port of simulator
 *
 * Revision 1.14  1996/07/22  19:52:16  jimz
 * switched node params to RF_DagParam_t, a union of
 * a 64-bit int and a void *, for better portability
 * attempted hpux port, but failed partway through for
 * lack of a single C compiler capable of compiling all
 * source files
 *
 * Revision 1.13  1996/06/09  02:36:46  jimz
 * lots of little crufty cleanup- fixup whitespace
 * issues, comment #ifdefs, improve typing in some
 * places (esp size-related)
 *
 * Revision 1.12  1996/06/07  22:26:27  jimz
 * type-ify which_ru (RF_ReconUnitNum_t)
 *
 * Revision 1.11  1996/06/07  21:33:04  jimz
 * begin using consistent types for sector numbers,
 * stripe numbers, row+col numbers, recon unit numbers
 *
 * Revision 1.10  1996/05/31  22:26:54  jimz
 * fix a lot of mapping problems, memory allocation problems
 * found some weird lock issues, fixed 'em
 * more code cleanup
 *
 * Revision 1.9  1996/05/30  11:29:41  jimz
 * Numerous bug fixes. Stripe lock release code disagreed with the taking code
 * about when stripes should be locked (I made it consistent: no parity, no lock)
 * There was a lot of extra serialization of I/Os which I've removed- a lot of
 * it was to calculate values for the cache code, which is no longer with us.
 * More types, function, macro cleanup. Added code to properly quiesce the array
 * on shutdown. Made a lot of stuff array-specific which was (bogusly) general
 * before. Fixed memory allocation, freeing bugs.
 *
 * Revision 1.8  1996/05/27  18:56:37  jimz
 * more code cleanup
 * better typing
 * compiles in all 3 environments
 *
 * Revision 1.7  1996/05/24  22:17:04  jimz
 * continue code + namespace cleanup
 * typed a bunch of flags
 *
 * Revision 1.6  1996/05/24  04:28:55  jimz
 * release cleanup ckpt
 *
 * Revision 1.5  1996/05/23  21:46:35  jimz
 * checkpoint in code cleanup (release prep)
 * lots of types, function names have been fixed
 *
 * Revision 1.4  1996/05/23  00:33:23  jimz
 * code cleanup: move all debug decls to rf_options.c, all extern
 * debug decls to rf_options.h, all debug vars preceded by rf_
 *
 * Revision 1.3  1996/05/18  19:51:34  jimz
 * major code cleanup- fix syntax, make some types consistent,
 * add prototypes, clean out dead code, et cetera
 *
 * Revision 1.2  1996/05/08  21:01:24  jimz
 * fixed up enum type names that were conflicting with other
 * enums and function names (ie, "panic")
 * future naming trends will be towards RF_ and rf_ for
 * everything raidframe-related
 *
 * Revision 1.1  1996/05/03  19:22:23  wvcii
 * Initial revision
 *
 */

#include "rf_types.h"
#include "rf_raid.h"
#include "rf_dag.h"
#include "rf_dagutils.h"
#include "rf_dagfuncs.h"
#include "rf_threadid.h"
#include "rf_debugMem.h"
#include "rf_memchunk.h"
#include "rf_general.h"
#include "rf_dagdegrd.h"
#include "rf_sys.h"


/******************************************************************************
 *
 * General comments on DAG creation:
 * 
 * All DAGs in this file use roll-away error recovery.  Each DAG has a single
 * commit node, usually called "Cmt."  If an error occurs before the Cmt node
 * is reached, the execution engine will halt forward execution and work
 * backward through the graph, executing the undo functions.  Assuming that
 * each node in the graph prior to the Cmt node are undoable and atomic - or -
 * does not make changes to permanent state, the graph will fail atomically.
 * If an error occurs after the Cmt node executes, the engine will roll-forward
 * through the graph, blindly executing nodes until it reaches the end.
 * If a graph reaches the end, it is assumed to have completed successfully.
 *
 * A graph has only 1 Cmt node.
 *
 */


/******************************************************************************
 *
 * The following wrappers map the standard DAG creation interface to the
 * DAG creation routines.  Additionally, these wrappers enable experimentation
 * with new DAG structures by providing an extra level of indirection, allowing
 * the DAG creation routines to be replaced at this single point.
 */

void rf_CreateRaidFiveDegradedReadDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList)
{
  rf_CreateDegradedReadDAG(raidPtr, asmap, dag_h, bp, flags, allocList,
    &rf_xorRecoveryFuncs);
}


/******************************************************************************
 *
 * DAG creation code begins here
 */


/******************************************************************************
 * Create a degraded read DAG for RAID level 1
 *
 * Hdr -> Nil -> R(p/s)d -> Commit -> Trm
 *
 * The "Rd" node reads data from the surviving disk in the mirror pair
 *   Rpd - read of primary copy
 *   Rsd - read of secondary copy
 *
 * Parameters:  raidPtr   - description of the physical array
 *              asmap     - logical & physical addresses for this access
 *              bp        - buffer ptr (for holding write data)
 *              flags     - general flags (e.g. disk locking) 
 *              allocList - list of memory allocated in DAG creation
 *****************************************************************************/

void rf_CreateRaidOneDegradedReadDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList)
{
  RF_DagNode_t *nodes, *rdNode, *blockNode, *commitNode, *termNode;
  RF_StripeNum_t parityStripeID;
  RF_ReconUnitNum_t which_ru;
  RF_PhysDiskAddr_t *pda;
  int useMirror, i;

  useMirror = 0;
  parityStripeID = rf_RaidAddressToParityStripeID(&(raidPtr->Layout),
    asmap->raidAddress, &which_ru);
  if (rf_dagDebug) {
    printf("[Creating RAID level 1 degraded read DAG]\n");
  }
  dag_h->creator = "RaidOneDegradedReadDAG";
  /* alloc the Wnd nodes and the Wmir node */
  if (asmap->numDataFailed == 0)
    useMirror = RF_FALSE;
  else
    useMirror = RF_TRUE;

  /* total number of nodes = 1 + (block + commit + terminator) */
  RF_CallocAndAdd(nodes, 4, sizeof(RF_DagNode_t), (RF_DagNode_t *), allocList);
  i = 0;
  rdNode      = &nodes[i]; i++;
  blockNode   = &nodes[i]; i++;
  commitNode = &nodes[i]; i++;
  termNode    = &nodes[i]; i++;

  /* this dag can not commit until the commit node is reached.   errors prior
   * to the commit point imply the dag has failed and must be retried
   */
  dag_h->numCommitNodes = 1;
  dag_h->numCommits = 0;
  dag_h->numSuccedents = 1;

  /* initialize the block, commit, and terminator nodes */
  rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
    NULL, 1, 0, 0, 0, dag_h, "Nil", allocList);
  rf_InitNode(commitNode, rf_wait, RF_TRUE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
    NULL, 1, 1, 0, 0, dag_h, "Cmt", allocList);
  rf_InitNode(termNode, rf_wait, RF_FALSE, rf_TerminateFunc, rf_TerminateUndoFunc,
    NULL, 0, 1, 0, 0, dag_h, "Trm", allocList);

  pda = asmap->physInfo;
  RF_ASSERT(pda != NULL);
  /* parityInfo must describe entire parity unit */
  RF_ASSERT(asmap->parityInfo->next == NULL);

  /* initialize the data node */
  if (!useMirror) {
    /* read primary copy of data */
    rf_InitNode(rdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc,
      rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rpd", allocList);
    rdNode->params[0].p = pda;
    rdNode->params[1].p = pda->bufPtr;
    rdNode->params[2].v = parityStripeID;
    rdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
  }
  else {
    /* read secondary copy of data */
    rf_InitNode(rdNode, rf_wait, RF_FALSE, rf_DiskReadFunc, rf_DiskReadUndoFunc,
      rf_GenericWakeupFunc, 1, 1, 4, 0, dag_h, "Rsd", allocList);
    rdNode->params[0].p = asmap->parityInfo;
    rdNode->params[1].p = pda->bufPtr;
    rdNode->params[2].v = parityStripeID;
    rdNode->params[3].v = RF_CREATE_PARAM3(RF_IO_NORMAL_PRIORITY, 0, 0, which_ru);
  }

  /* connect header to block node */
  RF_ASSERT(dag_h->numSuccedents == 1);
  RF_ASSERT(blockNode->numAntecedents == 0);
  dag_h->succedents[0] = blockNode;

  /* connect block node to rdnode */
  RF_ASSERT(blockNode->numSuccedents == 1);
  RF_ASSERT(rdNode->numAntecedents == 1);
  blockNode->succedents[0] = rdNode;
  rdNode->antecedents[0] = blockNode;
  rdNode->antType[0] = rf_control;

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

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



/******************************************************************************
 *
 * creates a DAG to perform a degraded-mode read of data within one stripe.
 * This DAG is as follows:
 *
 * Hdr -> Block -> Rud -> Xor -> Cmt -> T
 *              -> Rrd ->
 *              -> Rp -->
 *
 * Each R node is a successor of the L node
 * One successor arc from each R node goes to C, and the other to X
 * There is one Rud for each chunk of surviving user data requested by the
 * user, and one Rrd for each chunk of surviving user data _not_ being read by
 * the user
 * R = read, ud = user data, rd = recovery (surviving) data, p = parity
 * X = XOR, C = Commit, T = terminate
 *
 * The block node guarantees a single source node.
 *
 * Note:  The target buffer for the XOR node is set to the actual user buffer
 * where the failed data is supposed to end up.  This buffer is zero'd by the
 * code here.  Thus, if you create a degraded read dag, use it, and then
 * re-use, you have to be sure to zero the target buffer prior to the re-use.
 *
 * The recfunc argument at the end specifies the name and function used for
 * the redundancy
 * recovery function. 
 *
 *****************************************************************************/

void rf_CreateDegradedReadDAG(
  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         *recFunc)
{
  RF_DagNode_t *nodes, *rudNodes, *rrdNodes, *xorNode, *blockNode;
  RF_DagNode_t *commitNode, *rpNode, *termNode;
  int nNodes, nRrdNodes, nRudNodes, nXorBufs, i;
  int j, paramNum;
  RF_SectorCount_t sectorsPerSU;
  RF_ReconUnitNum_t which_ru;
  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;
  char *rpBuf;

  layoutPtr = &(raidPtr->Layout);
  /* failedPDA points to the pda within the asm that targets the failed disk */
  failedPDA = asmap->failedPDA;
  parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr,
    asmap->raidAddress, &which_ru);
  sectorsPerSU = layoutPtr->sectorsPerStripeUnit;

  if (rf_dagDebug) {
    printf("[Creating degraded read DAG]\n");
  }

  RF_ASSERT( asmap->numDataFailed == 1 );
  dag_h->creator = "DegradedReadDAG";

  /*
   * generate two ASMs identifying the surviving data we need
   * in order to recover the lost data
   */

  /* overlappingPDAs array must be zero'd */
  RF_Calloc(overlappingPDAs, asmap->numStripeUnitsAccessed, sizeof(char), (char *));
  rf_GenerateFailedAccessASMs(raidPtr, asmap, failedPDA, dag_h, new_asm_h, &nXorBufs,
    &rpBuf, overlappingPDAs, allocList);

  /*
   * create all the nodes at once
   *
   * -1 because no access is generated for the failed pda
   */
  nRudNodes = asmap->numStripeUnitsAccessed-1;
  nRrdNodes = ((new_asm_h[0]) ? new_asm_h[0]->stripeMap->numStripeUnitsAccessed : 0) +
              ((new_asm_h[1]) ? new_asm_h[1]->stripeMap->numStripeUnitsAccessed : 0);
  nNodes = 5 + nRudNodes + nRrdNodes; /* lock, unlock, xor, Rp, Rud, Rrd */
  RF_CallocAndAdd(nodes, nNodes, sizeof(RF_DagNode_t), (RF_DagNode_t *),
    allocList);
  i = 0;
  blockNode   = &nodes[i]; i++;
  commitNode  = &nodes[i]; i++;
  xorNode     = &nodes[i]; i++;
  rpNode      = &nodes[i]; i++;
  termNode    = &nodes[i]; i++;
  rudNodes    = &nodes[i]; i += nRudNodes;
  rrdNodes    = &nodes[i]; i += nRrdNodes;
  RF_ASSERT(i == nNodes);

  /* initialize nodes */
  dag_h->numCommitNodes = 1;
  dag_h->numCommits = 0;
  /* this dag can not commit until the commit node is reached
   * errors prior to the commit point imply the dag has failed