bufmgr.c

关系型数据库 Postgresql 6.5.2

			else
			{

				/*
				 * BM_JUST_DIRTIED cleared by BufferReplace and shouldn't
				 * be setted by anyone.		- vadim 01/17/97
				 */
				if (buf->flags & BM_JUST_DIRTIED)
				{
					elog(FATAL, "BufferAlloc: content of block %u (%s) changed while flushing",
						 buf->tag.blockNum, buf->sb_relname);
				}
				else
					buf->flags &= ~BM_DIRTY;
			}

			/*
			 * Somebody could have pinned the buffer while we were doing
			 * the I/O and had given up the BufMgrLock (though they would
			 * be waiting for us to clear the BM_IO_IN_PROGRESS flag).
			 * That's why this is a loop -- if so, we need to clear the
			 * I/O flags, remove our pin and start all over again.
			 *
			 * People may be making buffers free at any time, so there's no
			 * reason to think that we have an immediate disaster on our
			 * hands.
			 */
			if (buf && buf->refcount > 1)
			{
				inProgress = FALSE;
				buf->flags &= ~BM_IO_IN_PROGRESS;
#ifdef HAS_TEST_AND_SET
				S_UNLOCK(&(buf->io_in_progress_lock));
#else							/* !HAS_TEST_AND_SET */
				if (buf->refcount > 1)
					SignalIO(buf);
#endif	 /* !HAS_TEST_AND_SET */
				PrivateRefCount[BufferDescriptorGetBuffer(buf) - 1] = 0;
				buf->refcount--;
				buf = (BufferDesc *) NULL;
			}

			/*
			 * Somebody could have allocated another buffer for the same
			 * block we are about to read in. (While we flush out the
			 * dirty buffer, we don't hold the lock and someone could have
			 * allocated another buffer for the same block. The problem is
			 * we haven't gotten around to insert the new tag into the
			 * buffer table. So we need to check here.		-ay 3/95
			 */
			buf2 = BufTableLookup(&newTag);
			if (buf2 != NULL)
			{

				/*
				 * Found it. Someone has already done what we're about to
				 * do. We'll just handle this as if it were found in the
				 * buffer pool in the first place.
				 */
				if (buf != NULL)
				{
#ifdef HAS_TEST_AND_SET
					S_UNLOCK(&(buf->io_in_progress_lock));
#else							/* !HAS_TEST_AND_SET */
					if (buf->refcount > 1)
						SignalIO(buf);
#endif	 /* !HAS_TEST_AND_SET */
					/* give up the buffer since we don't need it any more */
					buf->refcount--;
					PrivateRefCount[BufferDescriptorGetBuffer(buf) - 1] = 0;
					AddBufferToFreelist(buf);
					buf->flags |= BM_FREE;
					buf->flags &= ~BM_IO_IN_PROGRESS;
				}

				PinBuffer(buf2);
				inProgress = (buf2->flags & BM_IO_IN_PROGRESS);

				*foundPtr = TRUE;
				if (inProgress)
				{
					WaitIO(buf2, BufMgrLock);
					if (buf2->flags & BM_IO_ERROR)
						*foundPtr = FALSE;
				}

				SpinRelease(BufMgrLock);

				return buf2;
			}
		}
	}

	/*
	 * At this point we should have the sole pin on a non-dirty buffer and
	 * we may or may not already have the BM_IO_IN_PROGRESS flag set.
	 */

	/*
	 * Change the name of the buffer in the lookup table:
	 *
	 * Need to update the lookup table before the read starts. If someone
	 * comes along looking for the buffer while we are reading it in, we
	 * don't want them to allocate a new buffer.  For the same reason, we
	 * didn't want to erase the buf table entry for the buffer we were
	 * writing back until now, either.
	 */

	if (!BufTableDelete(buf))
	{
		SpinRelease(BufMgrLock);
		elog(FATAL, "buffer wasn't in the buffer table\n");

	}

	/* record the database name and relation name for this buffer */
	strcpy(buf->sb_relname, reln->rd_rel->relname.data);
	strcpy(buf->sb_dbname, DatabaseName);

	INIT_BUFFERTAG(&(buf->tag), reln, blockNum);
	if (!BufTableInsert(buf))
	{
		SpinRelease(BufMgrLock);
		elog(FATAL, "Buffer in lookup table twice \n");
	}

	/*
	 * Buffer contents are currently invalid.  Have to mark IO IN PROGRESS
	 * so no one fiddles with them until the read completes.  If this
	 * routine has been called simply to allocate a buffer, no io will be
	 * attempted, so the flag isnt set.
	 */
	if (!inProgress)
	{
		buf->flags |= BM_IO_IN_PROGRESS;
#ifdef HAS_TEST_AND_SET
		Assert(S_LOCK_FREE(&(buf->io_in_progress_lock)));
		S_LOCK(&(buf->io_in_progress_lock));
#endif	 /* HAS_TEST_AND_SET */
	}

#ifdef BMTRACE
	_bm_trace((reln->rd_rel->relisshared ? 0 : MyDatabaseId), RelationGetRelid(reln), blockNum, BufferDescriptorGetBuffer(buf), BMT_ALLOCNOTFND);
#endif	 /* BMTRACE */

	SpinRelease(BufMgrLock);

	return buf;
}

/*
 * WriteBuffer
 *
 *		Pushes buffer contents to disk if WriteMode is BUFFER_FLUSH_WRITE.
 *		Otherwise, marks contents as dirty.
 *
 * Assume that buffer is pinned.  Assume that reln is
 *		valid.
 *
 * Side Effects:
 *		Pin count is decremented.
 */

#undef WriteBuffer

int
WriteBuffer(Buffer buffer)
{
	BufferDesc *bufHdr;

	if (WriteMode == BUFFER_FLUSH_WRITE)
		return FlushBuffer(buffer, TRUE);
	else
	{

		if (BufferIsLocal(buffer))
			return WriteLocalBuffer(buffer, TRUE);

		if (BAD_BUFFER_ID(buffer))
			return FALSE;

		bufHdr = &BufferDescriptors[buffer - 1];

		SharedBufferChanged = true;

		SpinAcquire(BufMgrLock);
		Assert(bufHdr->refcount > 0);
		bufHdr->flags |= (BM_DIRTY | BM_JUST_DIRTIED);
		UnpinBuffer(bufHdr);
		SpinRelease(BufMgrLock);
		CommitInfoNeedsSave[buffer - 1] = 0;
	}
	return TRUE;
}

#ifdef NOT_USED
void
WriteBuffer_Debug(char *file, int line, Buffer buffer)
{
	WriteBuffer(buffer);
	if (ShowPinTrace && BufferIsLocal(buffer) && is_userbuffer(buffer))
	{
		BufferDesc *buf;

		buf = &BufferDescriptors[buffer - 1];
		fprintf(stderr, "UNPIN(WR) %ld relname = %s, blockNum = %d, \
refcount = %ld, file: %s, line: %d\n",
				buffer, buf->sb_relname, buf->tag.blockNum,
				PrivateRefCount[buffer - 1], file, line);
	}
}

#endif

/*
 * DirtyBufferCopy() -- For a given dbid/relid/blockno, if the buffer is
 *						in the cache and is dirty, mark it clean and copy
 *						it to the requested location.  This is a logical
 *						write, and has been installed to support the cache
 *						management code for write-once storage managers.
 *
 *	DirtyBufferCopy() -- Copy a given dirty buffer to the requested
 *						 destination.
 *
 *		We treat this as a write.  If the requested buffer is in the pool
 *		and is dirty, we copy it to the location requested and mark it
 *		clean.	This routine supports the Sony jukebox storage manager,
 *		which agrees to take responsibility for the data once we mark
 *		it clean.
 *
 *	NOTE: used by sony jukebox code in postgres 4.2   - ay 2/95
 */
#ifdef NOT_USED
void
DirtyBufferCopy(Oid dbid, Oid relid, BlockNumber blkno, char *dest)
{
	BufferDesc *buf;
	BufferTag	btag;

	btag.relId.relId = relid;
	btag.relId.dbId = dbid;
	btag.blockNum = blkno;

	SpinAcquire(BufMgrLock);
	buf = BufTableLookup(&btag);

	if (buf == (BufferDesc *) NULL
		|| !(buf->flags & BM_DIRTY)
		|| !(buf->flags & BM_VALID))
	{
		SpinRelease(BufMgrLock);
		return;
	}

	/*
	 * hate to do this holding the lock, but release and reacquire is
	 * slower
	 */
	memmove(dest, (char *) MAKE_PTR(buf->data), BLCKSZ);

	buf->flags &= ~BM_DIRTY;

	SpinRelease(BufMgrLock);
}

#endif

/*
 * FlushBuffer -- like WriteBuffer, but force the page to disk.
 *
 * 'buffer' is known to be dirty/pinned, so there should not be a
 * problem reading the BufferDesc members without the BufMgrLock
 * (nobody should be able to change tags, flags, etc. out from under
 * us).
 */
static int
FlushBuffer(Buffer buffer, bool release)
{
	BufferDesc *bufHdr;
	Oid			bufdb;
	Relation	bufrel;
	int			status;

	if (BufferIsLocal(buffer))
		return FlushLocalBuffer(buffer, release);

	if (BAD_BUFFER_ID(buffer))
		return STATUS_ERROR;

	bufHdr = &BufferDescriptors[buffer - 1];
	bufdb = bufHdr->tag.relId.dbId;

	Assert(bufdb == MyDatabaseId || bufdb == (Oid) NULL);
	bufrel = RelationIdCacheGetRelation(bufHdr->tag.relId.relId);
	Assert(bufrel != (Relation) NULL);

	SharedBufferChanged = true;

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

	status = smgrflush(DEFAULT_SMGR, bufrel, bufHdr->tag.blockNum,
					   (char *) MAKE_PTR(bufHdr->data));

	RelationDecrementReferenceCount(bufrel);

	if (status == SM_FAIL)
	{
		elog(ERROR, "FlushBuffer: cannot flush block %u of the relation %s",
			 bufHdr->tag.blockNum, bufHdr->sb_relname);
		return STATUS_ERROR;
	}
	BufferFlushCount++;

	SpinAcquire(BufMgrLock);

	/*
	 * 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)
	{
		elog(NOTICE, "FlusfBuffer: content of block %u (%s) changed while flushing",
			 bufHdr->tag.blockNum, bufHdr->sb_relname);
	}
	else
		bufHdr->flags &= ~BM_DIRTY;
	if (release)
		UnpinBuffer(bufHdr);
	SpinRelease(BufMgrLock);
	CommitInfoNeedsSave[buffer - 1] = 0;

	return STATUS_OK;
}

/*
 * WriteNoReleaseBuffer -- like WriteBuffer, but do not unpin the buffer
 *						   when the operation is complete.
 *
 *		We know that the buffer is for a relation in our private cache,
 *		because this routine is called only to write out buffers that
 *		were changed by the executing backend.
 */
int
WriteNoReleaseBuffer(Buffer buffer)
{
	BufferDesc *bufHdr;

	if (WriteMode == BUFFER_FLUSH_WRITE)
		return FlushBuffer(buffer, FALSE);
	else
	{

		if (BufferIsLocal(buffer))
			return WriteLocalBuffer(buffer, FALSE);

		if (BAD_BUFFER_ID(buffer))
			return STATUS_ERROR;

		bufHdr = &BufferDescriptors[buffer - 1];

		SharedBufferChanged = true;

		SpinAcquire(BufMgrLock);
		bufHdr->flags |= (BM_DIRTY | BM_JUST_DIRTIED);
		SpinRelease(BufMgrLock);
		CommitInfoNeedsSave[buffer - 1] = 0;
	}
	return STATUS_OK;
}


#undef ReleaseAndReadBuffer
/*
 * ReleaseAndReadBuffer -- combine ReleaseBuffer() and ReadBuffer()
 *		so that only one semop needs to be called.
 *
 */
Buffer
ReleaseAndReadBuffer(Buffer buffer,
					 Relation relation,
					 BlockNumber blockNum)
{
	BufferDesc *bufHdr;
	Buffer		retbuf;

	if (BufferIsLocal(buffer))
	{
		Assert(LocalRefCount[-buffer - 1] > 0);
		LocalRefCount[-buffer - 1]--;
	}
	else
	{
		if (BufferIsValid(buffer))
		{
			bufHdr = &BufferDescriptors[buffer - 1];
			Assert(PrivateRefCount[buffer - 1] > 0);
			PrivateRefCount[buffer - 1]--;
			if (PrivateRefCount[buffer - 1] == 0 &&
				LastRefCount[buffer - 1] == 0)
			{

				/*
				 * only release buffer if it is not pinned in previous
				 * ExecMain level
				 */
				SpinAcquire(BufMgrLock);
				bufHdr->refcount--;
				if (bufHdr->refcount == 0)
				{
					AddBufferToFreelist(bufHdr);
					bufHdr->flags |= BM_FREE;
				}
				if (CommitInfoNeedsSave[buffer - 1])
				{
					bufHdr->flags |= (BM_DIRTY | BM_JUST_DIRTIED);
					CommitInfoNeedsSave[buffer - 1] = 0;
				}
				retbuf = ReadBufferWithBufferLock(relation, blockNum, true);
				return retbuf;
			}
		}
	}

	return ReadBuffer(relation, blockNum);
}

/*
 * BufferSync -- Flush all dirty buffers in the pool.
 *
 *		This is called at transaction commit time.	It does the wrong thing,
 *		right now.	We should flush only our own changes to stable storage,
 *		and we should obey the lock protocol on the buffer manager metadata
 *		as we do it.  Also, we need to be sure that no other transaction is
 *		modifying the page as we flush it.	This is only a problem for objects
 *		that use a non-two-phase locking protocol, like btree indices.	For
 *		those objects, we would like to set a write lock for the duration of
 *		our IO.  Another possibility is to code updates to btree pages
 *		carefully, so that writing them out out of order cannot cause
 *		any unrecoverable errors.
 *
 *		I don't want to think hard about this right now, so I will try
 *		to come back to it later.
 */
static void
BufferSync()
{
	int			i;
	Oid			bufdb;
	Oid			bufrel;
	Relation	reln;
	BufferDesc *bufHdr;
	int			status;

	SpinAcquire(BufMgrLock);
	for (i = 0, bufHdr = BufferDescriptors; i < NBuffers; i++, bufHdr++)
	{
		if ((bufHdr->flags & BM_VALID) && (bufHdr->flags & BM_DIRTY))
		{
			bufdb = bufHdr->tag.relId.dbId;
			bufrel = bufHdr->tag.relId.relId;
			if (bufdb == MyDatabaseId || bufdb == (Oid) 0)
			{
				reln = RelationIdCacheGetRelation(bufrel);

				/*
				 * We have to pin buffer to keep anyone from stealing it
				 * from the buffer pool while we are flushing it or
				 * waiting in WaitIO. It's bad for GetFreeBuffer in
				 * BufferAlloc, but there is no other way to prevent
				 * writing into disk block data from some other buffer,
				 * getting smgr status of some other block and clearing
				 * BM_DIRTY of ...			  - VAdim 09/16/96
				 */
				PinBuffer(bufHdr);
				if (bufHdr->flags & BM_IO_IN_PROGRESS)
				{
					WaitIO(bufHdr, BufMgrLock);
					UnpinBuffer(bufHdr);
					if (bufHdr->flags & BM_IO_ERROR)
					{
						elog(ERROR, "BufferSync: write error %u for %s",
							 bufHdr->tag.blockNum, bufHdr->sb_relname);
					}
					if (reln != (Relation) NULL)
						RelationDecrementReferenceCount(reln);
					continue;
				}

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

				/*
				 * If we didn't have the reldesc in our local cache, flush
				 * this page out using the 'blind write' storage manager
				 * routine.  If we did find it, use the standard
				 * interface.
				 */

#ifndef OPTIMIZE_SINGLE
				SpinRelease(BufMgrLock);
#endif	 /* OPTIMIZE_SINGLE */
				if (reln == (Relation) NULL)
				{
					status = smgrblindwrt(DEFAULT_SMGR, bufHdr->sb_dbname,
									   bufHdr->sb_relname, bufdb, bufrel,
										  bufHdr->tag.blockNum,
										(char *) MAKE_PTR(bufHdr->data));
				}
				else
				{
					status = smgrwrite(DEFAULT_SMGR, reln,
									   bufHdr->tag.blockNum,
									   (char *) MAKE_PTR(bufHdr->data));
				}
#ifndef OPTIMIZE_SINGLE
				SpinAcquire(BufMgrLock);
#endif	 /* OPTIMIZE_SINGLE */

				UnpinBuffer(bufHdr);
				if (status == SM_FAIL)
				{