bufmgr.c

关系型数据库 Postgresql 6.5.2

					bufHdr->flags |= BM_IO_ERROR;
					elog(ERROR, "BufferSync: cannot write %u for %s",
						 bufHdr->tag.blockNum, bufHdr->sb_relname);
				}
				BufferFlushCount++;

				/*
				 * If this buffer was marked by someone as DIRTY while we
				 * were flushing it out we must not clear DIRTY flag -
				 * vadim 01/17/97
				 */
				if (!(bufHdr->flags & BM_JUST_DIRTIED))
					bufHdr->flags &= ~BM_DIRTY;
				if (reln != (Relation) NULL)
					RelationDecrementReferenceCount(reln);
			}
		}
	}
	SpinRelease(BufMgrLock);

	LocalBufferSync();
}


/*
 * WaitIO -- Block until the IO_IN_PROGRESS flag on 'buf'
 *		is cleared.  Because IO_IN_PROGRESS conflicts are
 *		expected to be rare, there is only one BufferIO
 *		lock in the entire system.	All processes block
 *		on this semaphore when they try to use a buffer
 *		that someone else is faulting in.  Whenever a
 *		process finishes an IO and someone is waiting for
 *		the buffer, BufferIO is signaled (SignalIO).  All
 *		waiting processes then wake up and check to see
 *		if their buffer is now ready.  This implementation
 *		is simple, but efficient enough if WaitIO is
 *		rarely called by multiple processes simultaneously.
 *
 *	ProcSleep atomically releases the spinlock and goes to
 *		sleep.
 *
 *	Note: there is an easy fix if the queue becomes long.
 *		save the id of the buffer we are waiting for in
 *		the queue structure.  That way signal can figure
 *		out which proc to wake up.
 */
#ifdef HAS_TEST_AND_SET
static void
WaitIO(BufferDesc *buf, SPINLOCK spinlock)
{
	SpinRelease(spinlock);
	S_LOCK(&(buf->io_in_progress_lock));
	S_UNLOCK(&(buf->io_in_progress_lock));
	SpinAcquire(spinlock);
}

#else							/* HAS_TEST_AND_SET */
IpcSemaphoreId WaitIOSemId;
IpcSemaphoreId WaitCLSemId;

static void
WaitIO(BufferDesc *buf, SPINLOCK spinlock)
{
	bool		inProgress;

	for (;;)
	{

		/* wait until someone releases IO lock */
		(*NWaitIOBackendP)++;
		SpinRelease(spinlock);
		IpcSemaphoreLock(WaitIOSemId, 0, 1);
		SpinAcquire(spinlock);
		inProgress = (buf->flags & BM_IO_IN_PROGRESS);
		if (!inProgress)
			break;
	}
}

/*
 * SignalIO
 */
static void
SignalIO(BufferDesc *buf)
{
	/* somebody better be waiting. */
	Assert(buf->refcount > 1);
	IpcSemaphoreUnlock(WaitIOSemId, 0, *NWaitIOBackendP);
	*NWaitIOBackendP = 0;
}

#endif	 /* HAS_TEST_AND_SET */

long		NDirectFileRead;	/* some I/O's are direct file access.
								 * bypass bufmgr */
long		NDirectFileWrite;	/* e.g., I/O in psort and hashjoin.					*/

void
PrintBufferUsage(FILE *statfp)
{
	float		hitrate;
	float		localhitrate;

	if (ReadBufferCount == 0)
		hitrate = 0.0;
	else
		hitrate = (float) BufferHitCount *100.0 / ReadBufferCount;

	if (ReadLocalBufferCount == 0)
		localhitrate = 0.0;
	else
		localhitrate = (float) LocalBufferHitCount *100.0 / ReadLocalBufferCount;

	fprintf(statfp, "!\tShared blocks: %10ld read, %10ld written, buffer hit rate = %.2f%%\n",
			ReadBufferCount - BufferHitCount, BufferFlushCount, hitrate);
	fprintf(statfp, "!\tLocal  blocks: %10ld read, %10ld written, buffer hit rate = %.2f%%\n",
			ReadLocalBufferCount - LocalBufferHitCount, LocalBufferFlushCount, localhitrate);
	fprintf(statfp, "!\tDirect blocks: %10ld read, %10ld written\n",
			NDirectFileRead, NDirectFileWrite);
}

void
ResetBufferUsage()
{
	BufferHitCount = 0;
	ReadBufferCount = 0;
	BufferFlushCount = 0;
	LocalBufferHitCount = 0;
	ReadLocalBufferCount = 0;
	LocalBufferFlushCount = 0;
	NDirectFileRead = 0;
	NDirectFileWrite = 0;
}

/* ----------------------------------------------
 *		ResetBufferPool
 *
 *		this routine is supposed to be called when a transaction aborts.
 *		it will release all the buffer pins held by the transaciton.
 *
 * ----------------------------------------------
 */
void
ResetBufferPool()
{
	int			i;

	for (i = 1; i <= NBuffers; i++)
	{
		CommitInfoNeedsSave[i - 1] = 0;
		if (BufferIsValid(i))
		{
			while (PrivateRefCount[i - 1] > 0)
				ReleaseBuffer(i);
		}
		LastRefCount[i - 1] = 0;
	}

	ResetLocalBufferPool();
}

/* -----------------------------------------------
 *		BufferPoolCheckLeak
 *
 *		check if there is buffer leak
 *
 * -----------------------------------------------
 */
int
BufferPoolCheckLeak()
{
	int			i;
	int			result = 0;

	for (i = 1; i <= NBuffers; i++)
	{
		if (BufferIsValid(i))
		{
			BufferDesc *buf = &(BufferDescriptors[i - 1]);

			elog(NOTICE,
				 "Buffer Leak: [%03d] (freeNext=%d, freePrev=%d, \
relname=%s, blockNum=%d, flags=0x%x, refcount=%d %d)",
				 i - 1, buf->freeNext, buf->freePrev,
				 buf->sb_relname, buf->tag.blockNum, buf->flags,
				 buf->refcount, PrivateRefCount[i - 1]);
			result = 1;
		}
	}
	return (result);
}

/* ------------------------------------------------
 *		FlushBufferPool
 *
 *		flush all dirty blocks in buffer pool to disk
 *
 * ------------------------------------------------
 */
void
FlushBufferPool(int StableMainMemoryFlag)
{
	if (!StableMainMemoryFlag)
	{
		BufferSync();
		smgrcommit();
	}
}

/*
 * BufferGetBlockNumber
 *		Returns the block number associated with a buffer.
 *
 * Note:
 *		Assumes that the buffer is valid.
 */
BlockNumber
BufferGetBlockNumber(Buffer buffer)
{
	Assert(BufferIsValid(buffer));

	/* XXX should be a critical section */
	if (BufferIsLocal(buffer))
		return LocalBufferDescriptors[-buffer - 1].tag.blockNum;
	else
		return BufferDescriptors[buffer - 1].tag.blockNum;
}

#ifdef NOT_USED
/*
 * BufferGetRelation
 *		Returns the relation desciptor associated with a buffer.
 *
 * Note:
 *		Assumes buffer is valid.
 */
Relation
BufferGetRelation(Buffer buffer)
{
	Relation	relation;
	Oid			relid;

	Assert(BufferIsValid(buffer));
	Assert(!BufferIsLocal(buffer));		/* not supported for local buffers */

	/* XXX should be a critical section */
	relid = BufferDescriptors[buffer - 1].tag.relId.relId;
	relation = RelationIdGetRelation(relid);

	RelationDecrementReferenceCount(relation);

	if (RelationHasReferenceCountZero(relation))
	{

		/*
		 * elog(NOTICE, "BufferGetRelation: 0->1");
		 */

		RelationIncrementReferenceCount(relation);
	}

	return relation;
}

#endif

/*
 * BufferReplace
 *
 * Flush the buffer corresponding to 'bufHdr'
 *
 */
static int
BufferReplace(BufferDesc *bufHdr, bool bufferLockHeld)
{
	Relation	reln;
	Oid			bufdb,
				bufrel;
	int			status;

	if (!bufferLockHeld)
		SpinAcquire(BufMgrLock);

	/*
	 * first try to find the reldesc in the cache, if no luck, don't
	 * bother to build the reldesc from scratch, just do a blind write.
	 */

	bufdb = bufHdr->tag.relId.dbId;
	bufrel = bufHdr->tag.relId.relId;

	if (bufdb == MyDatabaseId || bufdb == (Oid) NULL)
		reln = RelationIdCacheGetRelation(bufrel);
	else
		reln = (Relation) NULL;

	/* To check if block content changed while flushing. - vadim 01/17/97 */
	bufHdr->flags &= ~BM_JUST_DIRTIED;

	SpinRelease(BufMgrLock);

	if (reln != (Relation) NULL)
	{
		status = smgrflush(DEFAULT_SMGR, reln, bufHdr->tag.blockNum,
						   (char *) MAKE_PTR(bufHdr->data));
	}
	else
	{

		/* blind write always flushes */
		status = smgrblindwrt(DEFAULT_SMGR, bufHdr->sb_dbname,
							  bufHdr->sb_relname, bufdb, bufrel,
							  bufHdr->tag.blockNum,
							  (char *) MAKE_PTR(bufHdr->data));
	}

	if (reln != (Relation) NULL)
		RelationDecrementReferenceCount(reln);

	if (status == SM_FAIL)
		return FALSE;

	BufferFlushCount++;

	return TRUE;
}

/*
 * RelationGetNumberOfBlocks
 *		Returns the buffer descriptor associated with a page in a relation.
 *
 * Note:
 *		XXX may fail for huge relations.
 *		XXX should be elsewhere.
 *		XXX maybe should be hidden
 */
BlockNumber
RelationGetNumberOfBlocks(Relation relation)
{
	return ((relation->rd_myxactonly) ? relation->rd_nblocks :
			smgrnblocks(DEFAULT_SMGR, relation));
}

/* ---------------------------------------------------------------------
 *		ReleaseRelationBuffers
 *
 *		this function unmarks all the dirty pages of a relation
 *		in the buffer pool so that at the end of transaction
 *		these pages will not be flushed.
 *		XXX currently it sequentially searches the buffer pool, should be
 *		changed to more clever ways of searching.
 * --------------------------------------------------------------------
 */
void
ReleaseRelationBuffers(Relation rel)
{
	int			i;
	int			holding = 0;
	BufferDesc *buf;

	if (rel->rd_myxactonly)
	{
		for (i = 0; i < NLocBuffer; i++)
		{
			buf = &LocalBufferDescriptors[i];
			if ((buf->flags & BM_DIRTY) &&
				(buf->tag.relId.relId == RelationGetRelid(rel)))
				buf->flags &= ~BM_DIRTY;
		}
		return;
	}

	for (i = 1; i <= NBuffers; i++)
	{
		buf = &BufferDescriptors[i - 1];
		if (!holding)
		{
			SpinAcquire(BufMgrLock);
			holding = 1;
		}
		if ((buf->flags & BM_DIRTY) &&
			(buf->tag.relId.dbId == MyDatabaseId) &&
			(buf->tag.relId.relId == RelationGetRelid(rel)))
		{
			buf->flags &= ~BM_DIRTY;
			if (!(buf->flags & BM_FREE))
			{
				SpinRelease(BufMgrLock);
				holding = 0;
				ReleaseBuffer(i);
			}
		}
	}
	if (holding)
		SpinRelease(BufMgrLock);
}

/* ---------------------------------------------------------------------
 *		DropBuffers
 *
 *		This function marks all the buffers in the buffer cache for a
 *		particular database as clean.  This is used when we destroy a
 *		database, to avoid trying to flush data to disk when the directory
 *		tree no longer exists.
 *
 *		This is an exceedingly non-public interface.
 * --------------------------------------------------------------------
 */
void
DropBuffers(Oid dbid)
{
	int			i;
	BufferDesc *buf;

	SpinAcquire(BufMgrLock);
	for (i = 1; i <= NBuffers; i++)
	{
		buf = &BufferDescriptors[i - 1];
		if ((buf->tag.relId.dbId == dbid) && (buf->flags & BM_DIRTY))
			buf->flags &= ~BM_DIRTY;
	}
	SpinRelease(BufMgrLock);
}

/* -----------------------------------------------------------------
 *		PrintBufferDescs
 *
 *		this function prints all the buffer descriptors, for debugging
 *		use only.
 * -----------------------------------------------------------------
 */
void
PrintBufferDescs()
{
	int			i;
	BufferDesc *buf = BufferDescriptors;

	if (IsUnderPostmaster)
	{
		SpinAcquire(BufMgrLock);
		for (i = 0; i < NBuffers; ++i, ++buf)
		{
			elog(DEBUG, "[%02d] (freeNext=%d, freePrev=%d, relname=%s, \
blockNum=%d, flags=0x%x, refcount=%d %d)",
				 i, buf->freeNext, buf->freePrev,
				 buf->sb_relname, buf->tag.blockNum, buf->flags,
				 buf->refcount, PrivateRefCount[i]);
		}
		SpinRelease(BufMgrLock);
	}
	else
	{
		/* interactive backend */
		for (i = 0; i < NBuffers; ++i, ++buf)
		{
			printf("[%-2d] (%s, %d) flags=0x%x, refcnt=%d %ld)\n",
				   i, buf->sb_relname, buf->tag.blockNum,
				   buf->flags, buf->refcount, PrivateRefCount[i]);
		}
	}
}

void
PrintPinnedBufs()
{
	int			i;
	BufferDesc *buf = BufferDescriptors;

	SpinAcquire(BufMgrLock);
	for (i = 0; i < NBuffers; ++i, ++buf)
	{
		if (PrivateRefCount[i] > 0)
			elog(NOTICE, "[%02d] (freeNext=%d, freePrev=%d, relname=%s, \
blockNum=%d, flags=0x%x, refcount=%d %d)\n",
				 i, buf->freeNext, buf->freePrev, buf->sb_relname,
				 buf->tag.blockNum, buf->flags,
				 buf->refcount, PrivateRefCount[i]);
	}
	SpinRelease(BufMgrLock);
}

/*
 * BufferPoolBlowaway
 *
 * this routine is solely for the purpose of experiments -- sometimes
 * you may want to blowaway whatever is left from the past in buffer
 * pool and start measuring some performance with a clean empty buffer
 * pool.
 */
#ifdef NOT_USED
void
BufferPoolBlowaway()
{
	int			i;

	BufferSync();
	for (i = 1; i <= NBuffers; i++)
	{
		if (BufferIsValid(i))
		{
			while (BufferIsValid(i))
				ReleaseBuffer(i);
		}
		BufTableDelete(&BufferDescriptors[i - 1]);
	}
}

#endif

/* ---------------------------------------------------------------------
 *		BlowawayRelationBuffers
 *
 *		This function blowaway all the pages with blocknumber >= passed
 *		of a relation in the buffer pool. Used by vacuum before truncation...
 *
 *		Returns: 0 - Ok, -1 - DIRTY, -2 - PINNED
 *
 *		XXX currently it sequentially searches the buffer pool, should be
 *		changed to more clever ways of searching.
 * --------------------------------------------------------------------
 */
int
BlowawayRelationBuffers(Relation rel, BlockNumber block)
{
	int			i;
	BufferDesc *buf;