rf_paritylogging.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: 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.
 */

/* $Locker:  $
 * $Log: rf_paritylogging.c,v $
 * Revision 1.41  1996/07/31  16:56:18  jimz
 * dataBytesPerStripe, sectorsPerDisk init arch-indep.
 *
 * Revision 1.40  1996/07/28  20:31:39  jimz
 * i386netbsd port
 * true/false fixup
 *
 * Revision 1.39  1996/07/18  22:57:14  jimz
 * port simulator to AIX
 *
 * Revision 1.38  1996/07/13  00:00:59  jimz
 * sanitized generalized reconstruction architecture
 * cleaned up head sep, rbuf problems
 *
 * Revision 1.37  1996/06/17  03:24:14  jimz
 * switch to new shutdown function typing
 *
 * Revision 1.36  1996/06/14  23:15:38  jimz
 * attempt to deal with thread GC problem
 *
 * Revision 1.35  1996/06/11  13:48:30  jimz
 * get it to compile in-kernel
 *
 * Revision 1.34  1996/06/11  10:16:35  jimz
 * Check return values on array configuration- back out if failed.
 * Reorder shutdown to avoid using deallocated resources.
 * Get rid of bogus join op in shutdown.
 *
 * Revision 1.33  1996/06/10  18:29:17  wvcii
 * fixed bug in rf_IdentifyStripeParityLogging
 * - added array initialization
 *
 * Revision 1.32  1996/06/10  11:55:47  jimz
 * Straightened out some per-array/not-per-array distinctions, fixed
 * a couple bugs related to confusion. Added shutdown lists. Removed
 * layout shutdown function (now subsumed by shutdown lists).
 *
 * Revision 1.31  1996/06/07  22:26:27  jimz
 * type-ify which_ru (RF_ReconUnitNum_t)
 *
 * Revision 1.30  1996/06/07  21:33:04  jimz
 * begin using consistent types for sector numbers,
 * stripe numbers, row+col numbers, recon unit numbers
 *
 * Revision 1.29  1996/06/05  18:06:02  jimz
 * Major code cleanup. The Great Renaming is now done.
 * Better modularity. Better typing. Fixed a bunch of
 * synchronization bugs. Made a lot of global stuff
 * per-desc or per-array. Removed dead code.
 *
 * Revision 1.28  1996/06/03  23:28:26  jimz
 * more bugfixes
 * check in tree to sync for IPDS runs with current bugfixes
 * there still may be a problem with threads in the script test
 * getting I/Os stuck- not trivially reproducible (runs ~50 times
 * in a row without getting stuck)
 *
 * Revision 1.27  1996/06/02  17:31:48  jimz
 * Moved a lot of global stuff into array structure, where it belongs.
 * Fixed up paritylogging, pss modules in this manner. Some general
 * code cleanup. Removed lots of dead code, some dead files.
 *
 * Revision 1.26  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 1.25  1996/05/30  23:22:16  jimz
 * bugfixes of serialization, timing problems
 * more cleanup
 *
 * Revision 1.24  1996/05/27  18:56:37  jimz
 * more code cleanup
 * better typing
 * compiles in all 3 environments
 *
 * Revision 1.23  1996/05/24  22:17:04  jimz
 * continue code + namespace cleanup
 * typed a bunch of flags
 *
 * Revision 1.22  1996/05/24  01:59:45  jimz
 * another checkpoint in code cleanup for release
 * time to sync kernel tree
 *
 * Revision 1.21  1996/05/23  21:46:35  jimz
 * checkpoint in code cleanup (release prep)
 * lots of types, function names have been fixed
 *
 * Revision 1.20  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 1.19  1996/05/20  16:16:30  jimz
 * switch to rf_{mutex,cond}_{init,destroy}
 *
 * Revision 1.18  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 1.17  1996/05/03  19:47:11  wvcii
 * added includes of new dag library
 *
 * Revision 1.16  1995/12/12  18:10:06  jimz
 * MIN -> RF_MIN, MAX -> RF_MAX, ASSERT -> RF_ASSERT
 * fix 80-column brain damage in comments
 *
 * Revision 1.15  1995/12/06  20:57:43  wvcii
 * added prototypes
 * reintegration of logs on shutdown now conditional on forceParityLogReint
 *
 * Revision 1.14  1995/11/30  16:06:42  wvcii
 * added copyright info
 *
 * Revision 1.13  1995/11/17  19:01:29  wvcii
 * added prototyping to MapParity
 *
 * Revision 1.12  1995/11/07  15:36:03  wvcii
 * changed ParityLoggingDagSelect prototype
 * function no longer returns numHdrSucc, numTermAnt
 *
 * Revision 1.11  1995/10/08  20:42:54  wvcii
 * lots of random debugging - debugging incomplete
 *
 * Revision 1.10  1995/09/07  01:26:55  jimz
 * Achive basic compilation in kernel. Kernel functionality
 * is not guaranteed at all, but it'll compile. Mostly. I hope.
 *
 * Revision 1.9  1995/09/06  19:21:17  wvcii
 * explicit shutdown (forced reintegration) for simulator version
 *
 * Revision 1.8  1995/07/08  18:19:16  rachad
 * Parity verifies can not be done in the simulator.
 *
 * Revision 1.7  1995/07/07  00:17:20  wvcii
 * this version free from deadlock, fails parity verification
 *
 * Revision 1.6  1995/06/23  13:39:59  robby
 * updeated to prototypes in rf_layout.h
 *
 * Revision 1.5  1995/06/09  13:14:56  wvcii
 * code is now nonblocking
 *
 * Revision 1.4  95/06/01  17:02:23  wvcii
 * code debug
 * 
 * Revision 1.3  95/05/31  13:08:57  wvcii
 * code debug
 * 
 * Revision 1.2  95/05/21  15:35:00  wvcii
 * code debug
 * 
 *
 *
 */

/*
  parity logging configuration, dag selection, and mapping is implemented here
 */

#include "rf_archs.h"

#if RF_INCLUDE_PARITYLOGGING > 0

#include "rf_types.h"
#include "rf_raid.h"
#include "rf_dag.h"
#include "rf_dagutils.h"
#include "rf_dagfuncs.h"
#include "rf_dagffrd.h"
#include "rf_dagffwr.h"
#include "rf_dagdegrd.h"
#include "rf_dagdegwr.h"
#include "rf_threadid.h"
#include "rf_paritylog.h"
#include "rf_paritylogDiskMgr.h"
#include "rf_paritylogging.h"
#include "rf_parityloggingdags.h"
#include "rf_general.h"
#include "rf_map.h"
#include "rf_utils.h"
#include "rf_shutdown.h"

typedef struct RF_ParityLoggingConfigInfo_s {
  RF_RowCol_t  **stripeIdentifier;  /* filled in at config time & used by IdentifyStripe */
} RF_ParityLoggingConfigInfo_t;

static void FreeRegionInfo(RF_Raid_t *raidPtr, RF_RegionId_t regionID);
static void rf_ShutdownParityLogging(RF_ThreadArg_t arg);
static void rf_ShutdownParityLoggingRegionInfo(RF_ThreadArg_t arg);
static void rf_ShutdownParityLoggingPool(RF_ThreadArg_t arg);
static void rf_ShutdownParityLoggingRegionBufferPool(RF_ThreadArg_t arg);
static void rf_ShutdownParityLoggingParityBufferPool(RF_ThreadArg_t arg);
static void rf_ShutdownParityLoggingDiskQueue(RF_ThreadArg_t arg);

int rf_ConfigureParityLogging(
  RF_ShutdownList_t  **listp,
  RF_Raid_t           *raidPtr,
  RF_Config_t         *cfgPtr)
{
  int i, j, startdisk, rc;
  RF_SectorCount_t totalLogCapacity, fragmentation, lastRegionCapacity;
  RF_SectorCount_t parityBufferCapacity, maxRegionParityRange;
  RF_RaidLayout_t *layoutPtr = &raidPtr->Layout;
  RF_ParityLoggingConfigInfo_t *info;
  RF_ParityLog_t *l, *next;
  caddr_t lHeapPtr;

  /*
   * We create multiple entries on the shutdown list here, since
   * this configuration routine is fairly complicated in and of
   * itself, and this makes backing out of a failed configuration
   * much simpler.
   */

  raidPtr->numSectorsPerLog = RF_DEFAULT_NUM_SECTORS_PER_LOG;

  /* create a parity logging configuration structure */
  RF_MallocAndAdd(info, sizeof(RF_ParityLoggingConfigInfo_t), (RF_ParityLoggingConfigInfo_t *), raidPtr->cleanupList);
  if (info == NULL)
    return(ENOMEM);
  layoutPtr->layoutSpecificInfo = (void *) info;

  RF_ASSERT(raidPtr->numRow == 1);
  
  /* the stripe identifier must identify the disks in each stripe,
   * IN THE ORDER THAT THEY APPEAR IN THE STRIPE.
   */
  info->stripeIdentifier = rf_make_2d_array((raidPtr->numCol), (raidPtr->numCol), raidPtr->cleanupList);
  if (info->stripeIdentifier == NULL)
    return(ENOMEM);

  startdisk = 0;
  for (i=0; i<(raidPtr->numCol); i++)
    {
      for (j=0; j<(raidPtr->numCol); j++)
	{
	  info->stripeIdentifier[i][j] = (startdisk + j) % (raidPtr->numCol - 1);
	}
      if ((--startdisk) < 0)
	startdisk = raidPtr->numCol-1-1;
    }

  /* fill in the remaining layout parameters */
  layoutPtr->numStripe = layoutPtr->stripeUnitsPerDisk;
  layoutPtr->bytesPerStripeUnit = layoutPtr->sectorsPerStripeUnit << raidPtr->logBytesPerSector;
  layoutPtr->numParityCol = 1;
  layoutPtr->numParityLogCol = 1;
  layoutPtr->numDataCol = raidPtr->numCol - layoutPtr->numParityCol - layoutPtr->numParityLogCol;
  layoutPtr->dataSectorsPerStripe = layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;
  layoutPtr->dataStripeUnitsPerDisk = layoutPtr->stripeUnitsPerDisk;
  raidPtr->sectorsPerDisk = layoutPtr->stripeUnitsPerDisk * layoutPtr->sectorsPerStripeUnit;

  raidPtr->totalSectors = layoutPtr->stripeUnitsPerDisk * layoutPtr->numDataCol * layoutPtr->sectorsPerStripeUnit;

  /* configure parity log parameters

     parameter               comment/constraints
     ----------------        -------------------
   * numParityRegions        all regions (except possibly last) of equal size
   * totalInCoreLogCapacity  amount of memory in bytes available for in-core logs (default 1 MB)
   # numSectorsPerLog        capacity of an in-core log in sectors (1 disk track)
     numParityLogs           total number of in-core logs, should be at least numParityRegions
     regionLogCapacity       size of a region log (except possibly last one) in sectors
     totalLogCapacity        total amount of log space in sectors

   * denotes a user settable parameter.
   # logs are fixed to be the size of a disk track, value #defined in rf_paritylog.h

  */

  totalLogCapacity = layoutPtr->stripeUnitsPerDisk * layoutPtr->sectorsPerStripeUnit * layoutPtr->numParityLogCol;
  raidPtr->regionLogCapacity = totalLogCapacity / rf_numParityRegions;
  if (rf_parityLogDebug)
    printf("bytes per sector %d\n", raidPtr->bytesPerSector);

  /* reduce fragmentation within a disk region by adjusting the number of regions
     in an attempt to allow an integral number of logs to fit into a disk region */
  fragmentation = raidPtr->regionLogCapacity % raidPtr->numSectorsPerLog;
  if (fragmentation > 0)
    for (i = 1; i < (raidPtr->numSectorsPerLog / 2); i++)
      {
	if (((totalLogCapacity / (rf_numParityRegions + i)) % raidPtr->numSectorsPerLog) < fragmentation)
	  {
	    rf_numParityRegions++;
	    raidPtr->regionLogCapacity = totalLogCapacity / rf_numParityRegions;
	    fragmentation = raidPtr->regionLogCapacity % raidPtr->numSectorsPerLog;
	  }
	if (((totalLogCapacity / (rf_numParityRegions - i)) % raidPtr->numSectorsPerLog) < fragmentation)
	  {
	    rf_numParityRegions--;
	    raidPtr->regionLogCapacity = totalLogCapacity / rf_numParityRegions;
	    fragmentation = raidPtr->regionLogCapacity % raidPtr->numSectorsPerLog;
	  }
      }
  /* ensure integral number of regions per log */
  raidPtr->regionLogCapacity = (raidPtr->regionLogCapacity / raidPtr->numSectorsPerLog) * raidPtr->numSectorsPerLog;

  raidPtr->numParityLogs = rf_totalInCoreLogCapacity / (raidPtr->bytesPerSector * raidPtr->numSectorsPerLog);
  /* to avoid deadlock, must ensure that enough logs exist for each region to have one simultaneously */
  if (raidPtr->numParityLogs < rf_numParityRegions)
    raidPtr->numParityLogs = rf_numParityRegions;

  /* create region information structs */
  RF_Malloc(raidPtr->regionInfo, (rf_numParityRegions * sizeof(RF_RegionInfo_t)), (RF_RegionInfo_t *));
  if (raidPtr->regionInfo == NULL)
    return(ENOMEM);

  /* last region may not be full capacity */
  lastRegionCapacity = raidPtr->regionLogCapacity;
  while ((rf_numParityRegions - 1) * raidPtr->regionLogCapacity + lastRegionCapacity > totalLogCapacity)
    lastRegionCapacity = lastRegionCapacity - raidPtr->numSectorsPerLog;

  raidPtr->regionParityRange = raidPtr->sectorsPerDisk / rf_numParityRegions;
  maxRegionParityRange = raidPtr->regionParityRange;

/* i can't remember why this line is in the code -wvcii 6/30/95 */
/*  if (raidPtr->sectorsPerDisk % rf_numParityRegions > 0)
    regionParityRange++; */

  /* build pool of unused parity logs */
  RF_Malloc(raidPtr->parityLogBufferHeap, raidPtr->numParityLogs * raidPtr->numSectorsPerLog * raidPtr->bytesPerSector, (caddr_t));
  if (raidPtr->parityLogBufferHeap == NULL)
    return(ENOMEM);
  lHeapPtr = raidPtr->parityLogBufferHeap;
  rc = rf_mutex_init(&raidPtr->parityLogPool.mutex);
  if (rc) {
    RF_ERRORMSG3("Unable to init mutex file %s line %d rc=%d\n", __FILE__,
      __LINE__, rc);
    RF_Free(raidPtr->parityLogBufferHeap, raidPtr->numParityLogs * raidPtr->numSectorsPerLog * raidPtr->bytesPerSector);
    return(ENOMEM);
  }
  for (i = 0; i < raidPtr->numParityLogs; i++)