rf_stripelocks.c

RAIDFrame是个非常好的磁盘阵列RAID仿真工具

  RF_ASSERT(RF_IO_IS_R_OR_W(lockReqDesc->type));

  if (rf_stripeLockDebug) {
    rf_get_threadid(tid);
    if (stripeID == -1) Dprintf1("[%d] Lock acquisition supressed (stripeID == -1)\n",tid);
    else {
      Dprintf8("[%d] Trying to acquire stripe lock table 0x%lx SID %ld type %c range %ld-%ld, range2 %ld-%ld hashval %d\n",
        tid, (unsigned long) lockTable, stripeID, lockReqDesc->type, lockReqDesc->start,
        lockReqDesc->stop, lockReqDesc->start2, lockReqDesc->stop2);
      Dprintf3("[%d] lock %ld hashval %d\n", tid, stripeID, hashval);
      FLUSH;
    }
  }
  if (stripeID == -1) return(0);
  lockReqDesc->next = NULL;                       /* just to be sure */
  
  RF_LOCK_MUTEX(lockTable[hashval].mutex);
  for (lockDesc = lockTable[hashval].descList; lockDesc; lockDesc=lockDesc->next) {
    if (lockDesc->stripeID == stripeID) break;
  }
  
  if (!lockDesc) {            /* no entry in table => no one reading or writing */
    lockDesc = AllocStripeLockDesc(stripeID);
    lockDesc->next = lockTable[hashval].descList;
    lockTable[hashval].descList = lockDesc;
    if (lockReqDesc->type == RF_IO_TYPE_WRITE) lockDesc->nWriters++;
    lockDesc->granted = lockReqDesc;
    if (rf_stripeLockDebug) {Dprintf7("[%d] no one waiting: lock %ld %c %ld-%ld %ld-%ld granted\n",
				tid,stripeID,lockReqDesc->type,lockReqDesc->start,lockReqDesc->stop,lockReqDesc->start2,lockReqDesc->stop2); FLUSH;}
  } else {

    if (lockReqDesc->type == RF_IO_TYPE_WRITE) lockDesc->nWriters++;
    
    if (lockDesc->nWriters == 0) {                        /* no need to search any lists if there are no writers anywhere */
      lockReqDesc->next = lockDesc->granted;
      lockDesc->granted = lockReqDesc;
      if (rf_stripeLockDebug) {Dprintf7("[%d] no writers: lock %ld %c %ld-%ld %ld-%ld granted\n",
				  tid,stripeID,lockReqDesc->type,lockReqDesc->start,lockReqDesc->stop,lockReqDesc->start2,lockReqDesc->stop2); FLUSH;}
    } else {

      /* search the granted & waiting lists for a conflict.  stop searching as soon as we find one */
      retcode = 0;
      for (p = lockDesc->granted; p; p=p->next) if (STRIPELOCK_CONFLICT(lockReqDesc, p)) {retcode = 1; break;}
      if (!retcode) for (p = lockDesc->waitersH; p; p=p->next) if (STRIPELOCK_CONFLICT(lockReqDesc, p)) {retcode = 2; break;}

      if (!retcode) {
	lockReqDesc->next = lockDesc->granted;                                   /* no conflicts found => grant lock */
	lockDesc->granted = lockReqDesc;
	if (rf_stripeLockDebug) {
		Dprintf7("[%d] no conflicts: lock %ld %c %ld-%ld %ld-%ld granted\n",
			tid,stripeID,lockReqDesc->type,lockReqDesc->start,lockReqDesc->stop,
			lockReqDesc->start2,lockReqDesc->stop2);
		FLUSH;
	}
      } else {
	if (rf_stripeLockDebug) {
		Dprintf6("[%d] conflict: lock %ld %c %ld-%ld hashval=%d not granted\n",
				tid,stripeID,lockReqDesc->type,lockReqDesc->start,lockReqDesc->stop,
				hashval);
		Dprintf3("[%d] lock %ld retcode=%d\n", tid, stripeID, retcode);
		FLUSH;
	}
	AddToWaitersQueue(lockTable, lockDesc, lockReqDesc);                     /* conflict => the current access must wait */
      }
    }
  }

  RF_UNLOCK_MUTEX(lockTable[hashval].mutex);
  return(retcode);
}

void rf_ReleaseStripeLock(
  RF_LockTableEntry_t  *lockTable,
  RF_StripeNum_t        stripeID,
  RF_LockReqDesc_t     *lockReqDesc)
{
  RF_StripeLockDesc_t *lockDesc, *ld_t;
  RF_LockReqDesc_t    *lr, *lr_t, *callbacklist, *t;
  RF_IoType_t type = lockReqDesc->type;
  int tid, hashval = HASH_STRIPEID(stripeID);
  int release_it, consider_it;
  RF_LockReqDesc_t *candidate, *candidate_t, *predecessor;

  RF_ASSERT(RF_IO_IS_R_OR_W(type));

  if (rf_stripeLockDebug) {
    rf_get_threadid(tid);
    if (stripeID == -1) Dprintf1("[%d] Lock release supressed (stripeID == -1)\n",tid);
    else {Dprintf8("[%d] Releasing stripe lock on stripe ID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
		 tid,stripeID,lockReqDesc->type,lockReqDesc->start,lockReqDesc->stop,lockReqDesc->start2,lockReqDesc->stop2, lockTable); FLUSH;}
  }

  if (stripeID == -1) return;
  
  RF_LOCK_MUTEX(lockTable[hashval].mutex);

  /* find the stripe lock descriptor */
  for (ld_t = NULL, lockDesc = lockTable[hashval].descList; lockDesc; ld_t = lockDesc, lockDesc=lockDesc->next) {
    if (lockDesc->stripeID == stripeID) break;
  }
  RF_ASSERT(lockDesc);                                    /* major error to release a lock that doesn't exist */

  /* find the stripe lock request descriptor & delete it from the list */
  for (lr_t = NULL, lr = lockDesc->granted; lr; lr_t = lr, lr=lr->next) if (lr == lockReqDesc) break;
  
  RF_ASSERT(lr && (lr == lockReqDesc));                            /* major error to release a lock that hasn't been granted */
  if (lr_t) lr_t->next = lr->next; else {
    RF_ASSERT(lr == lockDesc->granted);
    lockDesc->granted = lr->next;
  }
  lr->next = NULL;
  
  if (lockReqDesc->type == RF_IO_TYPE_WRITE) lockDesc->nWriters--;

  /* search through the waiters list to see if anyone needs to be woken up.
   * for each such descriptor in the wait list, we check it against everything granted and against
   * everything _in front_ of it in the waiters queue.  If it conflicts with none of these, we release it.
   *
   * DON'T TOUCH THE TEMPLINK POINTER OF ANYTHING IN THE GRANTED LIST HERE.  This will roach the case where
   * the callback tries to acquire a new lock in the same stripe.  There are some asserts to try and detect this.
   *
   * We apply 2 performance optimizations:
   * (1) if releasing this lock results in no more writers to this stripe, we just release everybody waiting,
   * since we place no restrictions on the number of concurrent reads.
   * (2) we consider as candidates for wakeup only those waiters that have a range overlap with either
   * the descriptor being woken up or with something in the callbacklist (i.e. something we've just now woken up).
   * This allows us to avoid the long evaluation for some descriptors.
   */
  
  callbacklist = NULL;
  if (lockDesc->nWriters == 0) {                          /* performance tweak (1) */
    while (lockDesc->waitersH) {
      
      lr = lockDesc->waitersH;                            /* delete from waiters list */
      lockDesc->waitersH = lr->next;
      
      RF_ASSERT(lr->type == RF_IO_TYPE_READ);
      
      lr->next = lockDesc->granted;                       /* add to granted list */
      lockDesc->granted = lr;

      RF_ASSERT(!lr->templink);
      lr->templink = callbacklist;                        /* put on callback list so that we'll invoke callback below */
      callbacklist = lr;
      if (rf_stripeLockDebug) {Dprintf8("[%d] No writers: granting lock stripe ID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
				     tid,stripeID,lr->type,lr->start,lr->stop,lr->start2,lr->stop2,(unsigned long) lockTable); FLUSH;}
    }
    lockDesc->waitersT = NULL;                            /* we've purged the whole waiters list */
      
  } else for (candidate_t = NULL, candidate = lockDesc->waitersH; candidate; ) {
  
    /* performance tweak (2) */
    consider_it = 0;
    if (RANGE_OVERLAP(lockReqDesc, candidate)) consider_it = 1;
    else for (t = callbacklist; t; t=t->templink) if (RANGE_OVERLAP(t, candidate)) {
      consider_it = 1;
      break;
    }
    if (!consider_it) {
      if (rf_stripeLockDebug) {Dprintf8("[%d] No overlap: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
				     tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
				     (unsigned long) lockTable); FLUSH;}
      candidate_t = candidate; candidate = candidate->next;
      continue;
    }
    
    
    /* we have a candidate for release.  check to make sure it is not blocked by any granted locks */
    release_it = 1;
    for (predecessor = lockDesc->granted; predecessor; predecessor = predecessor->next) {
      if (STRIPELOCK_CONFLICT(candidate, predecessor)) {
	if (rf_stripeLockDebug) {
	  Dprintf8("[%d] Conflicts with granted lock: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
		   tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
		   (unsigned long) lockTable); FLUSH;
	}
	release_it = 0; break;
      }
    }
    
    /* now check to see if the candidate is blocked by any waiters that occur before it it the wait queue */
    if (release_it) for (predecessor = lockDesc->waitersH; predecessor != candidate; predecessor = predecessor->next) {
      if (STRIPELOCK_CONFLICT(candidate, predecessor)) {
	if (rf_stripeLockDebug) {
	  Dprintf8("[%d] Conflicts with waiting lock: rejecting candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
		   tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
		   (unsigned long) lockTable); FLUSH;
	}
	release_it = 0; break;
      }
    }
    
    /* release it if indicated */
    if (release_it) {
      if (rf_stripeLockDebug) {Dprintf8("[%d] Granting lock to candidate stripeID %ld, type %c range %ld-%ld %ld-%ld table 0x%lx\n",
				     tid, stripeID, candidate->type, candidate->start, candidate->stop, candidate->start2, candidate->stop2,
				     (unsigned long) lockTable); FLUSH;}
      if (candidate_t) {
	candidate_t->next = candidate->next;
	if (lockDesc->waitersT == candidate) lockDesc->waitersT = candidate_t;       /* cannot be waitersH since candidate_t is not NULL */
      } else {
	RF_ASSERT(candidate == lockDesc->waitersH);
	lockDesc->waitersH = lockDesc->waitersH->next;
	if (!lockDesc->waitersH) lockDesc->waitersT = NULL;
      }
      candidate->next = lockDesc->granted;                    /* move it to the granted list */
      lockDesc->granted = candidate;
      
      RF_ASSERT(!candidate->templink);
      candidate->templink = callbacklist;                     /* put it on the list of things to be called after we release the mutex */
      callbacklist = candidate;
      
      if (!candidate_t) candidate = lockDesc->waitersH; else candidate = candidate_t->next;     /* continue with the rest of the list */
    } else {
      candidate_t = candidate; candidate = candidate->next;                                     /* continue with the rest of the list */
    }
  }

  /* delete the descriptor if no one is waiting or active */
  if (!lockDesc->granted && !lockDesc->waitersH) {
    RF_ASSERT(lockDesc->nWriters == 0);
    if (rf_stripeLockDebug) {
      Dprintf3("[%d] Last lock released (table 0x%lx): deleting desc for stripeID %ld\n",tid,(unsigned long) lockTable, stripeID); FLUSH;
    }
    if (ld_t) ld_t->next = lockDesc->next; else {
      RF_ASSERT(lockDesc == lockTable[hashval].descList);
      lockTable[hashval].descList = lockDesc->next;
    }
    FreeStripeLockDesc(lockDesc);
    lockDesc = NULL;                                                               /* only for the ASSERT below */
  }
	
  RF_UNLOCK_MUTEX(lockTable[hashval].mutex);

  /* now that we've unlocked the mutex, invoke the callback on all the descriptors in the list */
  RF_ASSERT(!( (callbacklist) && (!lockDesc) ));                   /* if we deleted the descriptor, we should have no callbacks to do */
  for (candidate = callbacklist; candidate; ) {
    t = candidate;
    candidate = candidate->templink;
    t->templink = NULL;
    (t->cbFunc)(t->cbArg);
  }
}

/* must have the indicated lock table mutex upon entry */
static void AddToWaitersQueue(
  RF_LockTableEntry_t  *lockTable,
  RF_StripeLockDesc_t  *lockDesc,
  RF_LockReqDesc_t     *lockReqDesc)
{
  int tid;

  if (rf_stripeLockDebug) {
    rf_get_threadid(tid);
    Dprintf3("[%d] Waiting on lock for stripe %ld table 0x%lx\n", tid, lockDesc->stripeID, (unsigned long) lockTable); FLUSH;
  }
  if (!lockDesc->waitersH) {
    lockDesc->waitersH = lockDesc->waitersT = lockReqDesc;
  } else {
    lockDesc->waitersT->next = lockReqDesc;
    lockDesc->waitersT = lockReqDesc;
  }
}

static RF_StripeLockDesc_t *AllocStripeLockDesc(RF_StripeNum_t stripeID)
{
	RF_StripeLockDesc_t *p;

	RF_FREELIST_GET(rf_stripelock_freelist,p,next,(RF_StripeLockDesc_t *));
	if (p) {
		p->stripeID = stripeID;
	}
	return(p);
}

static void FreeStripeLockDesc(RF_StripeLockDesc_t *p)
{
	RF_FREELIST_FREE(rf_stripelock_freelist,p,next);
}

static void PrintLockedStripes(lockTable)
  RF_LockTableEntry_t  *lockTable;
{
  int i, j, foundone = 0, did;
  RF_StripeLockDesc_t *p;
  RF_LockReqDesc_t *q;

  RF_LOCK_MUTEX(rf_printf_mutex);
  printf("Locked stripes:\n");
  for (i=0; i<rf_lockTableSize; i++) if (lockTable[i].descList) {
    foundone = 1;
    for (p = lockTable[i].descList; p; p=p->next) {
      printf("Stripe ID 0x%lx (%d) nWriters %d\n",p->stripeID, p->stripeID, p->nWriters);
      
      if (! (p->granted) ) printf("Granted: (none)\n"); else printf("Granted:\n");
      for (did=1,j=0,q = p->granted; q; j++,q=q->next) {
	printf("  %c(%ld-%ld",q->type,q->start,q->stop);
	if (q->start2 != -1) printf(",%ld-%ld) ",q->start2,q->stop2); else printf(") ");
	if (j && !(j%4)) {printf("\n"); did=1;} else did=0;
      }
      if (!did) printf("\n");
      
      if (! (p->waitersH) ) printf("Waiting: (none)\n"); else printf("Waiting:\n");
      for (did=1,j=0,q = p->waitersH; q; j++,q=q->next) {
	printf("%c(%ld-%ld",q->type,q->start,q->stop);
	if (q->start2 != -1) printf(",%ld-%ld) ",q->start2,q->stop2); else printf(") ");
	if (j && !(j%4)) {printf("\n         "); did=1;} else did=0;
      }
      if (!did) printf("\n");
    }
  }
  if (!foundone) printf("(none)\n"); else printf("\n");
  RF_UNLOCK_MUTEX(rf_printf_mutex);
}