rf_paritylogging.c

RAIDFrame是个非常好的磁盘阵列RAID仿真工具

}

static void rf_ShutdownParityLoggingRegionBufferPool(RF_ThreadArg_t arg)
{
  RF_Raid_t *raidPtr;

  raidPtr = (RF_Raid_t *)arg;
  if (rf_parityLogDebug) {
    int tid;
    rf_get_threadid(tid);
    printf("[%d] ShutdownParityLoggingRegionBufferPool\n", tid);
  }
  FreeRegionBufferQueue(&raidPtr->regionBufferPool);
}

static void rf_ShutdownParityLoggingParityBufferPool(RF_ThreadArg_t arg)
{
  RF_Raid_t *raidPtr;

  raidPtr = (RF_Raid_t *)arg;
  if (rf_parityLogDebug) {
    int tid;
    rf_get_threadid(tid);
    printf("[%d] ShutdownParityLoggingParityBufferPool\n", tid);
  }
  FreeRegionBufferQueue(&raidPtr->parityBufferPool);
}

static void rf_ShutdownParityLoggingDiskQueue(RF_ThreadArg_t arg)
{
  RF_ParityLogData_t *d;
  RF_CommonLogData_t *c;
  RF_Raid_t *raidPtr;

  raidPtr = (RF_Raid_t *)arg;
  if (rf_parityLogDebug) {
    int tid;
    rf_get_threadid(tid);
    printf("[%d] ShutdownParityLoggingDiskQueue\n", tid);
  }
  /* free disk manager stuff */
  RF_ASSERT(raidPtr->parityLogDiskQueue.bufHead == NULL);
  RF_ASSERT(raidPtr->parityLogDiskQueue.bufTail == NULL);
  RF_ASSERT(raidPtr->parityLogDiskQueue.reintHead == NULL);
  RF_ASSERT(raidPtr->parityLogDiskQueue.reintTail == NULL);
  while (raidPtr->parityLogDiskQueue.freeDataList)
    {
      d = raidPtr->parityLogDiskQueue.freeDataList;
      raidPtr->parityLogDiskQueue.freeDataList = raidPtr->parityLogDiskQueue.freeDataList->next;
      RF_Free(d, sizeof(RF_ParityLogData_t));
    }
  while (raidPtr->parityLogDiskQueue.freeCommonList)
    {
      c = raidPtr->parityLogDiskQueue.freeCommonList;
      rf_mutex_destroy(&c->mutex);
      raidPtr->parityLogDiskQueue.freeCommonList = raidPtr->parityLogDiskQueue.freeCommonList->next;
      RF_Free(c, sizeof(RF_CommonLogData_t));
    }
}

static void rf_ShutdownParityLogging(RF_ThreadArg_t arg)
{
  RF_Raid_t *raidPtr;
  int status;

  raidPtr = (RF_Raid_t *)arg;
  if (rf_parityLogDebug) {
    int tid;
    rf_get_threadid(tid);
    printf("[%d] ShutdownParityLogging\n", tid);
  }
#ifndef SIMULATE
  /* shutdown disk thread */
  /* This has the desirable side-effect of forcing all regions to be
     reintegrated.  This is necessary since all parity log maps are
     currently held in volatile memory. */

  RF_LOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex);
  raidPtr->parityLogDiskQueue.threadState |= RF_PLOG_TERMINATE;
  RF_UNLOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex);
  RF_SIGNAL_COND(raidPtr->parityLogDiskQueue.cond);
  /*
   * pLogDiskThread will now terminate when queues are cleared
   * now wait for it to be done
   */
  RF_LOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex);
  while(!(raidPtr->parityLogDiskQueue.threadState&RF_PLOG_SHUTDOWN)) {
    RF_WAIT_COND(raidPtr->parityLogDiskQueue.cond, raidPtr->parityLogDiskQueue.mutex);
  }
  RF_UNLOCK_MUTEX(raidPtr->parityLogDiskQueue.mutex);
#else /* !SIMULATE */
  /* explicitly call shutdown routines which force reintegration */
  rf_ShutdownLogging(raidPtr);
#endif /* !SIMULATE */
  if (rf_parityLogDebug) {
    int tid;
    rf_get_threadid(tid);
    printf("[%d] ShutdownParityLogging done (thread completed)\n", tid);
  }
}

int rf_GetDefaultNumFloatingReconBuffersParityLogging(RF_Raid_t *raidPtr)
{
  return(20);
}

RF_HeadSepLimit_t rf_GetDefaultHeadSepLimitParityLogging(RF_Raid_t *raidPtr)
{
  return(10);
}

/* return the region ID for a given RAID address */
RF_RegionId_t rf_MapRegionIDParityLogging(
  RF_Raid_t       *raidPtr,
  RF_SectorNum_t   address)
{
  RF_RegionId_t regionID;

/*  regionID = address / (raidPtr->regionParityRange * raidPtr->Layout.numDataCol); */
  regionID = address / raidPtr->regionParityRange;
  if (regionID == rf_numParityRegions)
    {
      /* last region may be larger than other regions */
      regionID--;
    }
  RF_ASSERT(address >= raidPtr->regionInfo[regionID].parityStartAddr);
  RF_ASSERT(address < raidPtr->regionInfo[regionID].parityStartAddr + raidPtr->regionInfo[regionID].numSectorsParity);
  RF_ASSERT(regionID < rf_numParityRegions);
  return(regionID);
}


/* given a logical RAID sector, determine physical disk address of data */
void rf_MapSectorParityLogging(
  RF_Raid_t         *raidPtr,
  RF_RaidAddr_t      raidSector,
  RF_RowCol_t       *row,
  RF_RowCol_t       *col,
  RF_SectorNum_t    *diskSector,
  int                remap)
{
  RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;
  *row = 0;
  /* *col = (SUID % (raidPtr->numCol - raidPtr->Layout.numParityLogCol)); */
  *col = SUID % raidPtr->Layout.numDataCol;
  *diskSector = (SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
    (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
}


/* given a logical RAID sector, determine physical disk address of parity  */
void rf_MapParityParityLogging(
  RF_Raid_t       *raidPtr,
  RF_RaidAddr_t    raidSector,
  RF_RowCol_t     *row,
  RF_RowCol_t     *col,
  RF_SectorNum_t  *diskSector,
  int              remap)
{
  RF_StripeNum_t SUID = raidSector / raidPtr->Layout.sectorsPerStripeUnit;

  *row = 0;
  /* *col = raidPtr->Layout.numDataCol-(SUID/raidPtr->Layout.numDataCol)%(raidPtr->numCol - raidPtr->Layout.numParityLogCol); */
  *col = raidPtr->Layout.numDataCol;
  *diskSector =(SUID / (raidPtr->Layout.numDataCol)) * raidPtr->Layout.sectorsPerStripeUnit +
    (raidSector % raidPtr->Layout.sectorsPerStripeUnit);
}


/* given a regionID and sector offset, determine the physical disk address of the parity log */
void rf_MapLogParityLogging(
  RF_Raid_t       *raidPtr,
  RF_RegionId_t    regionID,
  RF_SectorNum_t   regionOffset,
  RF_RowCol_t     *row,
  RF_RowCol_t     *col,
  RF_SectorNum_t  *startSector)
{
  *row = 0;
  *col = raidPtr->numCol - 1;
  *startSector = raidPtr->regionInfo[regionID].regionStartAddr + regionOffset;
}


/* given a regionID, determine the physical disk address of the logged parity for that region */
void rf_MapRegionParity(
  RF_Raid_t         *raidPtr,
  RF_RegionId_t      regionID,
  RF_RowCol_t       *row,
  RF_RowCol_t       *col,
  RF_SectorNum_t    *startSector,
  RF_SectorCount_t  *numSector)
{
  *row = 0;
  *col = raidPtr->numCol - 2;
  *startSector = raidPtr->regionInfo[regionID].parityStartAddr;
  *numSector = raidPtr->regionInfo[regionID].numSectorsParity;
}


/* given a logical RAID address, determine the participating disks in the stripe */
void rf_IdentifyStripeParityLogging(
  RF_Raid_t        *raidPtr,
  RF_RaidAddr_t     addr,
  RF_RowCol_t     **diskids,
  RF_RowCol_t      *outRow)
{
  RF_StripeNum_t stripeID = rf_RaidAddressToStripeID(&raidPtr->Layout, addr);
  RF_ParityLoggingConfigInfo_t *info = (RF_ParityLoggingConfigInfo_t *) raidPtr->Layout.layoutSpecificInfo;
  *outRow = 0;
  *diskids = info->stripeIdentifier[ stripeID % raidPtr->numCol ];
}


void rf_MapSIDToPSIDParityLogging(
  RF_RaidLayout_t    *layoutPtr,
  RF_StripeNum_t      stripeID,
  RF_StripeNum_t     *psID,
  RF_ReconUnitNum_t  *which_ru)
{
  *which_ru = 0;
  *psID = stripeID;
}


/* select an algorithm for performing an access.  Returns two pointers,
 * one to a function that will return information about the DAG, and
 * another to a function that will create the dag.
 */
void rf_ParityLoggingDagSelect(
  RF_Raid_t             *raidPtr,
  RF_IoType_t            type,
  RF_AccessStripeMap_t  *asmp,
  RF_VoidFuncPtr        *createFunc)
{
  RF_RaidLayout_t *layoutPtr = &(raidPtr->Layout);
  RF_PhysDiskAddr_t *failedPDA;
  RF_RowCol_t frow, fcol;
  RF_RowStatus_t rstat;
  int prior_recon;
  int tid;
  
  RF_ASSERT(RF_IO_IS_R_OR_W(type));

  if (asmp->numDataFailed + asmp->numParityFailed > 1) {
    RF_ERRORMSG("Multiple disks failed in a single group!  Aborting I/O operation.\n");
    /* *infoFunc = */ *createFunc = NULL;
    return;
  } else if (asmp->numDataFailed + asmp->numParityFailed == 1) {
    
    /* if under recon & already reconstructed, redirect the access to the spare drive 
     * and eliminate the failure indication 
     */
    failedPDA = asmp->failedPDA;
    frow = failedPDA->row; fcol = failedPDA->col;
    rstat = raidPtr->status[failedPDA->row];
    prior_recon = (rstat == rf_rs_reconfigured) || (
      (rstat == rf_rs_reconstructing) ?
      rf_CheckRUReconstructed(raidPtr->reconControl[frow]->reconMap, failedPDA->startSector) : 0
      );
    if (prior_recon) {
      RF_RowCol_t or = failedPDA->row,oc=failedPDA->col;
      RF_SectorNum_t oo=failedPDA->startSector;
      if (layoutPtr->map->flags & RF_DISTRIBUTE_SPARE) {         /* redirect to dist spare space */

        if (failedPDA == asmp->parityInfo) {

          /* parity has failed */
          (layoutPtr->map->MapParity)(raidPtr, failedPDA->raidAddress, &failedPDA->row, 
                                      &failedPDA->col, &failedPDA->startSector, RF_REMAP);

          if (asmp->parityInfo->next) {                          /* redir 2nd component, if any */
            RF_PhysDiskAddr_t *p = asmp->parityInfo->next;
            RF_SectorNum_t SUoffs = p->startSector % layoutPtr->sectorsPerStripeUnit;
            p->row = failedPDA->row;
            p->col = failedPDA->col;
            p->startSector = rf_RaidAddressOfPrevStripeUnitBoundary(layoutPtr, failedPDA->startSector) +
                             SUoffs;    /* cheating:  startSector is not really a RAID address */
          }

        } else if (asmp->parityInfo->next && failedPDA == asmp->parityInfo->next) {
          RF_ASSERT(0);            /* should not ever happen */
        } else {

          /* data has failed */
          (layoutPtr->map->MapSector)(raidPtr, failedPDA->raidAddress, &failedPDA->row, 
                                      &failedPDA->col, &failedPDA->startSector, RF_REMAP);

        }

      } else {                                                 /* redirect to dedicated spare space */

        failedPDA->row = raidPtr->Disks[frow][fcol].spareRow;
        failedPDA->col = raidPtr->Disks[frow][fcol].spareCol;

        /* the parity may have two distinct components, both of which may need to be redirected */
        if (asmp->parityInfo->next) {
          if (failedPDA == asmp->parityInfo) {
            failedPDA->next->row = failedPDA->row;
            failedPDA->next->col = failedPDA->col;
          } else if (failedPDA == asmp->parityInfo->next) {    /* paranoid:  should never occur */
            asmp->parityInfo->row = failedPDA->row;
            asmp->parityInfo->col = failedPDA->col;
          }
        }
      }

      RF_ASSERT(failedPDA->col != -1);
       
      if (rf_dagDebug || rf_mapDebug) {
        rf_get_threadid(tid);
        printf("[%d] Redirected type '%c' r %d c %d o %ld -> r %d c %d o %ld\n",
             tid,type,or,oc,oo,failedPDA->row,failedPDA->col,failedPDA->startSector);
      }
      
      asmp->numDataFailed = asmp->numParityFailed = 0;
    }

  }

  
  if (type == RF_IO_TYPE_READ) {

    if (asmp->numDataFailed == 0)
      *createFunc = rf_CreateFaultFreeReadDAG;
    else
      *createFunc = rf_CreateRaidFiveDegradedReadDAG;
    
  }
  else {

    
    /* if mirroring, always use large writes.  If the access requires two distinct parity updates,
     * always do a small write.  If the stripe contains a failure but the access does not, do a
     * small write.
     * The first conditional (numStripeUnitsAccessed <= numDataCol/2) uses a less-than-or-equal
     * rather than just a less-than because when G is 3 or 4, numDataCol/2 is 1, and I want
     * single-stripe-unit updates to use just one disk.
     */
    if ( (asmp->numDataFailed + asmp->numParityFailed) == 0) {
      if (((asmp->numStripeUnitsAccessed <= (layoutPtr->numDataCol / 2)) && (layoutPtr->numDataCol!=1)) ||
          (asmp->parityInfo->next!=NULL) || rf_CheckStripeForFailures(raidPtr, asmp)) {
	*createFunc = rf_CreateParityLoggingSmallWriteDAG;
      }
      else
        *createFunc = rf_CreateParityLoggingLargeWriteDAG;
    } 
    else
      if (asmp->numParityFailed == 1)
        *createFunc = (void (*)())rf_CreateNonRedundantWriteDAG;
      else
        if (asmp->numStripeUnitsAccessed != 1 && failedPDA->numSector != layoutPtr->sectorsPerStripeUnit)
          *createFunc = NULL;
        else
          *createFunc = rf_CreateDegradedWriteDAG;
  }
}

#endif /* RF_INCLUDE_PARITYLOGGING > 0 */