rf_dagffwr.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_dagff.c
 *
 * code for creating fault-free DAGs
 *
 * $Locker:  $
 * $Log: rf_dagffwr.c,v $
 * Revision 1.19  1996/07/31  15:35:24  jimz
 * evenodd changes; bugfixes for double-degraded archs, generalize
 * some formerly PQ-only functions
 *
 * Revision 1.18  1996/07/28  20:31:39  jimz
 * i386netbsd port
 * true/false fixup
 *
 * Revision 1.17  1996/07/27  18:40:24  jimz
 * cleanup sweep
 *
 * Revision 1.16  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 1.15  1996/06/11  01:27:50  jimz
 * Fixed bug where diskthread shutdown would crash or hang. This
 * turned out to be two distinct bugs:
 * (1) [crash] The thread shutdown code wasn't properly waiting for
 * all the diskthreads to complete. This caused diskthreads that were
 * exiting+cleaning up to unlock a destroyed mutex.
 * (2) [hang] TerminateDiskQueues wasn't locking, and DiskIODequeue
 * only checked for termination _after_ a wakeup if the queues were
 * empty. This was a race where the termination wakeup could be lost
 * by the dequeueing thread, and the system would hang waiting for the
 * thread to exit, while the thread waited for an I/O or a signal to
 * check the termination flag.
 *
 * Revision 1.14  1996/06/10  22:24:01  wvcii
 * added write dags which do not have a commit node and are
 * used in forward and backward error recovery experiments.
 *
 * Revision 1.13  1996/06/07  22:26:27  jimz
 * type-ify which_ru (RF_ReconUnitNum_t)
 *
 * Revision 1.12  1996/06/07  21:33:04  jimz
 * begin using consistent types for sector numbers,
 * stripe numbers, row+col numbers, recon unit numbers
 *
 * Revision 1.11  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 1.10  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 1.9  1996/05/27  18:56:37  jimz
 * more code cleanup
 * better typing
 * compiles in all 3 environments
 *
 * Revision 1.8  1996/05/24  22:17:04  jimz
 * continue code + namespace cleanup
 * typed a bunch of flags
 *
 * Revision 1.7  1996/05/24  04:28:55  jimz
 * release cleanup ckpt
 *
 * Revision 1.6  1996/05/23  21:46:35  jimz
 * checkpoint in code cleanup (release prep)
 * lots of types, function names have been fixed
 *
 * Revision 1.5  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 1.4  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 1.3  1996/05/15  23:23:12  wvcii
 * fixed bug in small write read old q node succedent initialization
 *
 * 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  19:20:45  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_dagffrd.h"
#include "rf_memchunk.h"
#include "rf_general.h"
#include "rf_dagffwr.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_CreateNonRedundantWriteDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList,
  RF_IoType_t            type)
{
  rf_CreateNonredundantDAG(raidPtr, asmap, dag_h, bp, flags, allocList,
    RF_IO_TYPE_WRITE);
}

void rf_CreateRAID0WriteDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList,
  RF_IoType_t            type)
{
  rf_CreateNonredundantDAG(raidPtr, asmap, dag_h, bp, flags, allocList,
    RF_IO_TYPE_WRITE);
}

void rf_CreateSmallWriteDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList)
{
#if RF_FORWARD > 0
  rf_CommonCreateSmallWriteDAGFwd(raidPtr, asmap, dag_h, bp, flags, allocList,
    &rf_xorFuncs, NULL);
#else /* RF_FORWARD > 0 */
#if RF_BACKWARD > 0
  rf_CommonCreateSmallWriteDAGFwd(raidPtr, asmap, dag_h, bp, flags, allocList,
    &rf_xorFuncs, NULL);
#else /* RF_BACKWARD > 0 */
  /* "normal" rollaway */
  rf_CommonCreateSmallWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList,
    &rf_xorFuncs, NULL);
#endif /* RF_BACKWARD > 0 */
#endif /* RF_FORWARD > 0 */
}

void rf_CreateLargeWriteDAG(
  RF_Raid_t             *raidPtr,
  RF_AccessStripeMap_t  *asmap,
  RF_DagHeader_t        *dag_h,
  void                  *bp,
  RF_RaidAccessFlags_t   flags,
  RF_AllocListElem_t    *allocList)
{
#if RF_FORWARD > 0
  rf_CommonCreateLargeWriteDAGFwd(raidPtr, asmap, dag_h, bp, flags, allocList,
    1, rf_RegularXorFunc, RF_TRUE);
#else /* RF_FORWARD > 0 */
#if RF_BACKWARD > 0
  rf_CommonCreateLargeWriteDAGFwd(raidPtr, asmap, dag_h, bp, flags, allocList,
    1, rf_RegularXorFunc, RF_TRUE);
#else /* RF_BACKWARD > 0 */
  /* "normal" rollaway */
  rf_CommonCreateLargeWriteDAG(raidPtr, asmap, dag_h, bp, flags, allocList,
    1, rf_RegularXorFunc, RF_TRUE);
#endif /* RF_BACKWARD > 0 */
#endif /* RF_FORWARD > 0 */
}


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


/******************************************************************************
 *
 * creates a DAG to perform a large-write operation:
 *
 *           / Rod \           / Wnd \
 * H -- block- Rod - Xor - Cmt - Wnd --- T
 *           \ Rod /          \  Wnp /
 *                             \[Wnq]/
 *
 * The XOR node also does the Q calculation in the P+Q architecture.
 * All nodes are before the commit node (Cmt) are assumed to be atomic and
 * undoable - or - they make no changes to permanent state.
 *
 * Rod = read old data
 * Cmt = commit node
 * 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
 *              nfaults   - number of faults array can tolerate
 *                          (equal to # redundancy units in stripe)
 *              redfuncs  - list of redundancy generating functions
 *
 *****************************************************************************/

void rf_CommonCreateLargeWriteDAG(
  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,
  int                  (*redFunc)(),
  int                    allowBufferRecycle)
{
  RF_DagNode_t *nodes, *wndNodes, *rodNodes, *xorNode, *wnpNode;
  RF_DagNode_t *wnqNode, *blockNode, *commitNode, *termNode;
  int nWndNodes, nRodNodes, i, nodeNum, asmNum;
  RF_AccessStripeMapHeader_t *new_asm_h[2];
  RF_StripeNum_t parityStripeID;
  char *sosBuffer, *eosBuffer;
  RF_ReconUnitNum_t which_ru;
  RF_RaidLayout_t *layoutPtr;
  RF_PhysDiskAddr_t *pda;

  layoutPtr = &(raidPtr->Layout);
  parityStripeID = rf_RaidAddressToParityStripeID(layoutPtr, asmap->raidAddress,
    &which_ru);

  if (rf_dagDebug) {
    printf("[Creating large-write DAG]\n");
  }
  dag_h->creator = "LargeWriteDAG";

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

  /* alloc the nodes: Wnd, xor, commit, block, term, and  Wnp */
  nWndNodes = asmap->numStripeUnitsAccessed;
  RF_CallocAndAdd(nodes, nWndNodes + 4 + nfaults, sizeof(RF_DagNode_t),
    (RF_DagNode_t *), allocList);
  i = 0;
  wndNodes    = &nodes[i]; i += nWndNodes;
  xorNode     = &nodes[i]; i += 1;
  wnpNode     = &nodes[i]; i += 1;
  blockNode   = &nodes[i]; i += 1;
  commitNode  = &nodes[i]; i += 1;
  termNode    = &nodes[i]; i += 1;
  if (nfaults == 2) {
    wnqNode   = &nodes[i]; i += 1;
  }
  else {
    wnqNode = NULL;
  }
  rf_MapUnaccessedPortionOfStripe(raidPtr, layoutPtr, asmap, dag_h, new_asm_h,
    &nRodNodes, &sosBuffer, &eosBuffer, allocList);
  if (nRodNodes > 0) {
    RF_CallocAndAdd(rodNodes, nRodNodes, sizeof(RF_DagNode_t),
      (RF_DagNode_t *), allocList);
  }
  else {
    rodNodes = NULL;
  }

  /* begin node initialization */
  if (nRodNodes > 0) {
    rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,
      NULL, nRodNodes, 0, 0, 0, dag_h, "Nil", allocList);
  }
  else {
    rf_InitNode(blockNode, rf_wait, RF_FALSE, rf_NullNodeFunc, rf_NullNodeUndoFunc,