rf_reconbuffer.c

RAIDFrame是个非常好的磁盘阵列RAID仿真工具

    if (use_committed) {      /* if a buffer has been committed to us, use it */
      t = reconCtrlPtr->committedRbufs;
      RF_ASSERT(t);
      reconCtrlPtr->committedRbufs = t->next;
      t->next = NULL;
    } else if (reconCtrlPtr->floatingRbufs) {
      t = reconCtrlPtr->floatingRbufs;
      reconCtrlPtr->floatingRbufs = t->next;
      t->next = NULL;
    }
  }

  /* If we weren't able to acquire a buffer, 
   * append to the end of the buf list in the recon ctrl struct.
   */
  if (!t) {
    RF_ASSERT(!keep_it && !use_committed);
    Dprintf2("RECON: row %d col %d failed to acquire floating rbuf\n",rbuf->row, rbuf->col);

    raidPtr->procsInBufWait++;
    if ( (raidPtr->procsInBufWait == raidPtr->numCol -1) && (raidPtr->numFullReconBuffers == 0)) {
      printf("Buffer wait deadlock detected.  Exiting.\n");
      rf_PrintPSStatusTable(raidPtr, rbuf->row);
      RF_PANIC();
    }
    pssPtr->flags |= RF_PSS_BUFFERWAIT;
    cb = rf_AllocCallbackDesc();                      /* append to buf wait list in recon ctrl structure */
    cb->row = rbuf->row; cb->col = rbuf->col;
    cb->callbackArg.v  = rbuf->parityStripeID;
    cb->callbackArg2.v = rbuf->which_ru;
    cb->next = NULL;
    if (!reconCtrlPtr->bufferWaitList) reconCtrlPtr->bufferWaitList = cb;
    else {       /* might want to maintain head/tail pointers here rather than search for end of list */
      for (p = reconCtrlPtr->bufferWaitList; p->next; p=p->next);
      p->next = cb;
    }
    retcode = 1;
    goto out;
  }
  Dprintf2("RECON: row %d col %d acquired rbuf\n",rbuf->row, rbuf->col);
  RF_ETIMER_STOP(raidPtr->recon_tracerecs[rbuf->col].recon_timer);
  RF_ETIMER_EVAL(raidPtr->recon_tracerecs[rbuf->col].recon_timer);
  raidPtr->recon_tracerecs[rbuf->col].specific.recon.recon_return_to_submit_us +=
    RF_ETIMER_VAL_US(raidPtr->recon_tracerecs[rbuf->col].recon_timer);
  RF_ETIMER_START(raidPtr->recon_tracerecs[rbuf->col].recon_timer);
  
  rf_LogTraceRec(raidPtr, &raidPtr->recon_tracerecs[rbuf->col]);

  /* initialize the buffer */
  if (t!=rbuf) {
    t->row = rbuf->row; t->col = reconCtrlPtr->fcol;
    t->parityStripeID = rbuf->parityStripeID;
    t->which_ru = rbuf->which_ru;
    t->failedDiskSectorOffset = rbuf->failedDiskSectorOffset;
    t->spRow=rbuf->spRow;
    t->spCol=rbuf->spCol;
    t->spOffset=rbuf->spOffset;
    
    ta = t->buffer; t->buffer = rbuf->buffer; rbuf->buffer = ta;      /* swap buffers */
  }

  /* the first installation always gets installed as the destination buffer.
   * subsequent installations get stacked up to allow for multi-way XOR
   */
  if (!pssPtr->rbuf) {pssPtr->rbuf = t; t->count = 1;}
  else pssPtr->rbufsForXor[ pssPtr->xorBufCount++ ] = t;          /* install this buffer */

  rf_CheckForFullRbuf(raidPtr, reconCtrlPtr, pssPtr, layoutPtr->numDataCol);      /* the buffer is full if G=2 */

out:  
  RF_UNLOCK_PSS_MUTEX( raidPtr,rbuf->row,rbuf->parityStripeID);
  RF_UNLOCK_MUTEX( reconCtrlPtr->rb_mutex );
  return(retcode);
}

int rf_MultiWayReconXor(raidPtr, pssPtr)
  RF_Raid_t                     *raidPtr;
  RF_ReconParityStripeStatus_t  *pssPtr;   /* the pss descriptor for this parity stripe */
{
  int i, numBufs = pssPtr->xorBufCount;
  int numBytes = rf_RaidAddressToByte(raidPtr, raidPtr->Layout.sectorsPerStripeUnit * raidPtr->Layout.SUsPerRU);
  RF_ReconBuffer_t **rbufs = (RF_ReconBuffer_t **) pssPtr->rbufsForXor;
  RF_ReconBuffer_t *targetRbuf = (RF_ReconBuffer_t *) pssPtr->rbuf;
  
  RF_ASSERT(pssPtr->rbuf != NULL);
  RF_ASSERT(numBufs > 0 && numBufs < RF_PS_MAX_BUFS);
#ifdef KERNEL
  thread_block(); /* yield the processor before doing a big XOR */
#endif /* KERNEL */
  /*
   * XXX
   *
   * What if more than 9 bufs?
   */
  nWayXorFuncs[numBufs](pssPtr->rbufsForXor, targetRbuf, numBytes/sizeof(long));

  /* release all the reconstruction buffers except the last one, which belongs to the
   * the disk who's submission caused this XOR to take place
   */
  for (i=0; i < numBufs-1; i++) {
    if (rbufs[i]->type == RF_RBUF_TYPE_FLOATING) rf_ReleaseFloatingReconBuffer(raidPtr, rbufs[i]->row, rbufs[i]);
    else if (rbufs[i]->type == RF_RBUF_TYPE_FORCED) rf_FreeReconBuffer(rbufs[i]);
    else RF_ASSERT(0);
  }
  targetRbuf->count += pssPtr->xorBufCount;
  pssPtr->xorBufCount = 0;
  return(0);
}

/* removes one full buffer from one of the full-buffer lists and returns it.
 *
 * ASSUMES THE RB_MUTEX IS UNLOCKED AT ENTRY.
 */
RF_ReconBuffer_t *rf_GetFullReconBuffer(reconCtrlPtr)
  RF_ReconCtrl_t  *reconCtrlPtr;
{
  RF_ReconBuffer_t *p;

  RF_LOCK_MUTEX(reconCtrlPtr->rb_mutex);

  if ( (p=reconCtrlPtr->priorityList) != NULL) {
    reconCtrlPtr->priorityList = p->next;
    p->next = NULL;
    goto out;
  }
  if ( (p=reconCtrlPtr->fullBufferList) != NULL) {
    reconCtrlPtr->fullBufferList = p->next;
    p->next = NULL;
    goto out;
  }

out:
  RF_UNLOCK_MUTEX(reconCtrlPtr->rb_mutex);
  return(p);
}


/* if the reconstruction buffer is full, move it to the full list, which is maintained
 * sorted by failed disk sector offset
 *
 * ASSUMES THE RB_MUTEX IS LOCKED AT ENTRY.
 */
int rf_CheckForFullRbuf(raidPtr, reconCtrl, pssPtr, numDataCol)
  RF_Raid_t                     *raidPtr;
  RF_ReconCtrl_t                *reconCtrl;
  RF_ReconParityStripeStatus_t  *pssPtr;
  int                            numDataCol;
{
  RF_ReconBuffer_t *p, *pt, *rbuf = (RF_ReconBuffer_t *) pssPtr->rbuf;

  if (rbuf->count == numDataCol) {
    raidPtr->numFullReconBuffers++;
    Dprintf2("RECON: rbuf for psid %ld ru %d has filled\n",rbuf->parityStripeID, rbuf->which_ru);
    if (!reconCtrl->fullBufferList || (rbuf->failedDiskSectorOffset < reconCtrl->fullBufferList->failedDiskSectorOffset)) {
      Dprintf2("RECON: rbuf for psid %ld ru %d is head of list\n", rbuf->parityStripeID, rbuf->which_ru);
      rbuf->next = reconCtrl->fullBufferList;
      reconCtrl->fullBufferList = rbuf;
    }
    else {
      for (pt = reconCtrl->fullBufferList, p = pt->next; p && p->failedDiskSectorOffset < rbuf->failedDiskSectorOffset; pt=p, p=p->next);
      rbuf->next = p;
      pt->next = rbuf;
      Dprintf2("RECON: rbuf for psid %ld ru %d is in list\n", rbuf->parityStripeID, rbuf->which_ru);
    }
#if 0
    pssPtr->writeRbuf = pssPtr->rbuf;        /* DEBUG ONLY:  we like to be able to find this rbuf while it's awaiting write */
#else
    rbuf->pssPtr = pssPtr;
#endif
    pssPtr->rbuf = NULL;
    rf_CauseReconEvent(raidPtr, rbuf->row, rbuf->col, NULL, RF_REVENT_BUFREADY);
  }
  return(0);
}


/* release a floating recon buffer for someone else to use.
 * assumes the rb_mutex is LOCKED at entry
 */
void rf_ReleaseFloatingReconBuffer(raidPtr, row, rbuf)
  RF_Raid_t         *raidPtr;
  RF_RowCol_t        row;
  RF_ReconBuffer_t  *rbuf;
{
  RF_ReconCtrl_t *rcPtr = raidPtr->reconControl[row];
  RF_CallbackDesc_t *cb;

  Dprintf2("RECON: releasing rbuf for psid %ld ru %d\n",rbuf->parityStripeID, rbuf->which_ru);
  
  /* if anyone is waiting on buffers, wake one of them up.  They will subsequently wake up anyone
   * else waiting on their RU
   */
  if (rcPtr->bufferWaitList) {
    rbuf->next = rcPtr->committedRbufs;
    rcPtr->committedRbufs = rbuf;
    cb = rcPtr->bufferWaitList;
    rcPtr->bufferWaitList = cb->next;
    rf_CauseReconEvent(raidPtr, cb->row, cb->col, (void *) 1, RF_REVENT_BUFCLEAR);  /* arg==1 => we've committed a buffer */
    rf_FreeCallbackDesc(cb);
    raidPtr->procsInBufWait--;
  } else {
    rbuf->next = rcPtr->floatingRbufs;
    rcPtr->floatingRbufs = rbuf;
  }
}

/* release any disk that is waiting on a buffer for the indicated RU.
 * assumes the rb_mutex is LOCKED at entry
 */
void rf_ReleaseBufferWaiters(raidPtr, pssPtr)
  RF_Raid_t                     *raidPtr;
  RF_ReconParityStripeStatus_t  *pssPtr;
{
  RF_CallbackDesc_t *cb1, *cb = pssPtr->bufWaitList;

  Dprintf2("RECON: releasing buf waiters for psid %ld ru %d\n",pssPtr->parityStripeID, pssPtr->which_ru);
  pssPtr->flags &= ~RF_PSS_BUFFERWAIT;
  while (cb) {
    cb1 = cb->next;
    cb->next = NULL;
    rf_CauseReconEvent(raidPtr, cb->row, cb->col, (void *) 0, RF_REVENT_BUFCLEAR);  /* arg==0 => we haven't committed a buffer */
    rf_FreeCallbackDesc(cb);
    cb = cb1;
  }
  pssPtr->bufWaitList = NULL;
}

/* when reconstruction is forced on an RU, there may be some disks waiting to
 * acquire a buffer for that RU.  Since we allocate a new buffer as part of
 * the forced-reconstruction process, we no longer have to wait for any
 * buffers, so we wakeup any waiter that we find in the bufferWaitList
 *
 * assumes the rb_mutex is LOCKED at entry
 */
void rf_ReleaseBufferWaiter(rcPtr, rbuf)
  RF_ReconCtrl_t    *rcPtr;
  RF_ReconBuffer_t  *rbuf;
{
  RF_CallbackDesc_t *cb, *cbt;

  for (cbt = NULL, cb = rcPtr->bufferWaitList; cb; cbt = cb, cb=cb->next) {
    if ( (cb->callbackArg.v == rbuf->parityStripeID) && ( cb->callbackArg2.v == rbuf->which_ru)) {
      Dprintf2("RECON: Dropping row %d col %d from buffer wait list\n", cb->row, cb->col);
      if (cbt) cbt->next = cb->next;
      else rcPtr->bufferWaitList = cb->next;
      rf_CauseReconEvent((RF_Raid_t *) rbuf->raidPtr, cb->row, cb->col, (void *) 0, RF_REVENT_BUFREADY);  /* arg==0 => no committed buffer */
      rf_FreeCallbackDesc(cb);
      return;
    }
  }
}