bufmgr.c

PostgreSQL7.4.6 for Linux

	/*
	 * relation->rd_nblocks should be accurate already if the relation is
	 * new or temp, because no one else should be modifying it.  Otherwise
	 * we need to ask the smgr for the current physical file length.
	 *
	 * Don't call smgr on a view, either.
	 */
	if (relation->rd_rel->relkind == RELKIND_VIEW)
		relation->rd_nblocks = 0;
	else if (relation->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
		relation->rd_nblocks = 0;
	else if (!relation->rd_isnew && !relation->rd_istemp)
		relation->rd_nblocks = smgrnblocks(DEFAULT_SMGR, relation);
	return relation->rd_nblocks;
}

/*
 * RelationUpdateNumberOfBlocks
 *		Forcibly update relation->rd_nblocks.
 *
 * If the relcache drops an entry for a temp relation, it must call this
 * routine after recreating the relcache entry, so that rd_nblocks is
 * re-sync'd with reality.  See RelationGetNumberOfBlocks.
 */
void
RelationUpdateNumberOfBlocks(Relation relation)
{
	if (relation->rd_rel->relkind == RELKIND_VIEW)
		relation->rd_nblocks = 0;
	else if (relation->rd_rel->relkind == RELKIND_COMPOSITE_TYPE)
		relation->rd_nblocks = 0;
	else
		relation->rd_nblocks = smgrnblocks(DEFAULT_SMGR, relation);
}

/* ---------------------------------------------------------------------
 *		DropRelationBuffers
 *
 *		This function removes all the buffered pages for a relation
 *		from the buffer pool.  Dirty pages are simply dropped, without
 *		bothering to write them out first.	This is NOT rollback-able,
 *		and so should be used only with extreme caution!
 *
 *		We assume that the caller holds an exclusive lock on the relation,
 *		which should assure that no new buffers will be acquired for the rel
 *		meanwhile.
 * --------------------------------------------------------------------
 */
void
DropRelationBuffers(Relation rel)
{
	DropRelFileNodeBuffers(rel->rd_node, rel->rd_istemp);
}

/* ---------------------------------------------------------------------
 *		DropRelFileNodeBuffers
 *
 *		This is the same as DropRelationBuffers, except that the target
 *		relation is specified by RelFileNode and temp status.
 *
 *		This is NOT rollback-able.	One legitimate use is to clear the
 *		buffer cache of buffers for a relation that is being deleted
 *		during transaction abort.
 * --------------------------------------------------------------------
 */
void
DropRelFileNodeBuffers(RelFileNode rnode, bool istemp)
{
	int			i;
	BufferDesc *bufHdr;

	if (istemp)
	{
		for (i = 0; i < NLocBuffer; i++)
		{
			bufHdr = &LocalBufferDescriptors[i];
			if (RelFileNodeEquals(bufHdr->tag.rnode, rnode))
			{
				bufHdr->flags &= ~(BM_DIRTY | BM_JUST_DIRTIED);
				bufHdr->cntxDirty = false;
				LocalRefCount[i] = 0;
				bufHdr->tag.rnode.relNode = InvalidOid;
			}
		}
		return;
	}

	LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);

	for (i = 1; i <= NBuffers; i++)
	{
		bufHdr = &BufferDescriptors[i - 1];
recheck:
		if (RelFileNodeEquals(bufHdr->tag.rnode, rnode))
		{
			/*
			 * If there is I/O in progress, better wait till it's done;
			 * don't want to delete the relation out from under someone
			 * who's just trying to flush the buffer!
			 */
			if (bufHdr->flags & BM_IO_IN_PROGRESS)
			{
				WaitIO(bufHdr);

				/*
				 * By now, the buffer very possibly belongs to some other
				 * rel, so check again before proceeding.
				 */
				goto recheck;
			}
			/* Now we can do what we came for */
			bufHdr->flags &= ~(BM_DIRTY | BM_JUST_DIRTIED);
			bufHdr->cntxDirty = false;

			/*
			 * Release any refcount we may have.  If someone else has a
			 * pin on the buffer, we got trouble.
			 */
			if (!(bufHdr->flags & BM_FREE))
			{
				/* the sole pin should be ours */
				if (bufHdr->refcount != 1 || PrivateRefCount[i - 1] == 0)
					elog(FATAL, "block %u of %u/%u is still referenced (private %ld, global %d)",
						 bufHdr->tag.blockNum,
						 bufHdr->tag.rnode.tblNode,
						 bufHdr->tag.rnode.relNode,
						 PrivateRefCount[i - 1], bufHdr->refcount);
				/* Make sure it will be released */
				PrivateRefCount[i - 1] = 1;
				UnpinBuffer(bufHdr);
			}

			/*
			 * And mark the buffer as no longer occupied by this rel.
			 */
			BufTableDelete(bufHdr);
		}
	}

	LWLockRelease(BufMgrLock);
}

/* ---------------------------------------------------------------------
 *		DropBuffers
 *
 *		This function removes all the buffers in the buffer cache for a
 *		particular database.  Dirty pages are simply dropped, without
 *		bothering to write them out first.	This is used when we destroy a
 *		database, to avoid trying to flush data to disk when the directory
 *		tree no longer exists.	Implementation is pretty similar to
 *		DropRelationBuffers() which is for destroying just one relation.
 * --------------------------------------------------------------------
 */
void
DropBuffers(Oid dbid)
{
	int			i;
	BufferDesc *bufHdr;

	LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);

	for (i = 1; i <= NBuffers; i++)
	{
		bufHdr = &BufferDescriptors[i - 1];
recheck:

		/*
		 * We know that currently database OID is tblNode but this
		 * probably will be changed in future and this func will be used
		 * to drop tablespace buffers.
		 */
		if (bufHdr->tag.rnode.tblNode == dbid)
		{
			/*
			 * If there is I/O in progress, better wait till it's done;
			 * don't want to delete the database out from under someone
			 * who's just trying to flush the buffer!
			 */
			if (bufHdr->flags & BM_IO_IN_PROGRESS)
			{
				WaitIO(bufHdr);

				/*
				 * By now, the buffer very possibly belongs to some other
				 * DB, so check again before proceeding.
				 */
				goto recheck;
			}
			/* Now we can do what we came for */
			bufHdr->flags &= ~(BM_DIRTY | BM_JUST_DIRTIED);
			bufHdr->cntxDirty = false;

			/*
			 * The thing should be free, if caller has checked that no
			 * backends are running in that database.
			 */
			Assert(bufHdr->flags & BM_FREE);

			/*
			 * And mark the buffer as no longer occupied by this page.
			 */
			BufTableDelete(bufHdr);
		}
	}

	LWLockRelease(BufMgrLock);
}

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

	if (IsUnderPostmaster)
	{
		LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);
		for (i = 0; i < NBuffers; ++i, ++buf)
		{
			elog(LOG, "[%02d] (freeNext=%d, freePrev=%d, rel=%u/%u, \
blockNum=%u, flags=0x%x, refcount=%d %ld)",
				 i, buf->freeNext, buf->freePrev,
				 buf->tag.rnode.tblNode, buf->tag.rnode.relNode,
				 buf->tag.blockNum, buf->flags,
				 buf->refcount, PrivateRefCount[i]);
		}
		LWLockRelease(BufMgrLock);
	}
	else
	{
		/* interactive backend */
		for (i = 0; i < NBuffers; ++i, ++buf)
		{
			printf("[%-2d] (%u/%u, %u) flags=0x%x, refcnt=%d %ld)\n",
				   i, buf->tag.rnode.tblNode, buf->tag.rnode.relNode,
				   buf->tag.blockNum,
				   buf->flags, buf->refcount, PrivateRefCount[i]);
		}
	}
}
#endif

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

	LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);
	for (i = 0; i < NBuffers; ++i, ++buf)
	{
		if (PrivateRefCount[i] > 0)
			elog(WARNING, "[%02d] (freeNext=%d, freePrev=%d, rel=%u/%u, \
blockNum=%u, flags=0x%x, refcount=%d %ld)",
				 i, buf->freeNext, buf->freePrev,
				 buf->tag.rnode.tblNode, buf->tag.rnode.relNode,
				 buf->tag.blockNum, buf->flags,
				 buf->refcount, PrivateRefCount[i]);
	}
	LWLockRelease(BufMgrLock);
}
#endif

/* ---------------------------------------------------------------------
 *		FlushRelationBuffers
 *
 *		This function writes all dirty pages of a relation out to disk.
 *		Furthermore, pages that have blocknumber >= firstDelBlock are
 *		actually removed from the buffer pool.	An error code is returned
 *		if we fail to dump a dirty buffer or if we find one of
 *		the target pages is pinned into the cache.
 *
 *		This is called by DROP TABLE to clear buffers for the relation
 *		from the buffer pool.  Note that we must write dirty buffers,
 *		rather than just dropping the changes, because our transaction
 *		might abort later on; we want to roll back safely in that case.
 *
 *		This is also called by VACUUM before truncating the relation to the
 *		given number of blocks.  It might seem unnecessary for VACUUM to
 *		write dirty pages before firstDelBlock, since VACUUM should already
 *		have committed its changes.  However, it is possible for there still
 *		to be dirty pages: if some page had unwritten on-row tuple status
 *		updates from a prior transaction, and VACUUM had no additional
 *		changes to make to that page, then VACUUM won't have written it.
 *		This is harmless in most cases but will break pg_upgrade, which
 *		relies on VACUUM to ensure that *all* tuples have correct on-row
 *		status.  So, we check and flush all dirty pages of the rel
 *		regardless of block number.
 *
 *		In all cases, the caller should be holding AccessExclusiveLock on
 *		the target relation to ensure that no other backend is busy reading
 *		more blocks of the relation (or might do so before we commit).
 *
 *		Formerly, we considered it an error condition if we found dirty
 *		buffers here.	However, since BufferSync no longer forces out all
 *		dirty buffers at every xact commit, it's possible for dirty buffers
 *		to still be present in the cache due to failure of an earlier
 *		transaction.  So, must flush dirty buffers without complaint.
 *
 *		Returns: 0 - Ok, -1 - FAILED TO WRITE DIRTY BUFFER, -2 - PINNED
 *
 *		XXX currently it sequentially searches the buffer pool, should be
 *		changed to more clever ways of searching.
 * --------------------------------------------------------------------
 */
int
FlushRelationBuffers(Relation rel, BlockNumber firstDelBlock)
{
	int			i;
	BufferDesc *bufHdr;
	XLogRecPtr	recptr;
	int			status;
	ErrorContextCallback errcontext;

	/* Setup error traceback support for ereport() */
	errcontext.callback = buffer_write_error_callback;
	errcontext.arg = NULL;
	errcontext.previous = error_context_stack;
	error_context_stack = &errcontext;

	if (rel->rd_istemp)
	{
		for (i = 0; i < NLocBuffer; i++)
		{
			bufHdr = &LocalBufferDescriptors[i];
			errcontext.arg = bufHdr;
			if (RelFileNodeEquals(bufHdr->tag.rnode, rel->rd_node))
			{
				if (bufHdr->flags & BM_DIRTY || bufHdr->cntxDirty)
				{
					status = smgrwrite(DEFAULT_SMGR, rel,
									   bufHdr->tag.blockNum,
									   (char *) MAKE_PTR(bufHdr->data));
					if (status == SM_FAIL)
					{
						error_context_stack = errcontext.previous;
						elog(WARNING, "FlushRelationBuffers(\"%s\" (local), %u): block %u is dirty, could not flush it",
							 RelationGetRelationName(rel), firstDelBlock,
							 bufHdr->tag.blockNum);
						return (-1);
					}
					bufHdr->flags &= ~(BM_DIRTY | BM_JUST_DIRTIED);
					bufHdr->cntxDirty = false;
				}
				if (LocalRefCount[i] > 0)
				{
					error_context_stack = errcontext.previous;
					elog(WARNING, "FlushRelationBuffers(\"%s\" (local), %u): block %u is referenced (%ld)",
						 RelationGetRelationName(rel), firstDelBlock,
						 bufHdr->tag.blockNum, LocalRefCount[i]);
					return (-2);
				}
				if (bufHdr->tag.blockNum >= firstDelBlock)
					bufHdr->tag.rnode.relNode = InvalidOid;
			}
		}

		/* Pop the error context stack */
		error_context_stack = errcontext.previous;

		return 0;
	}

	LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);

	for (i = 0; i < NBuffers; i++)
	{
		bufHdr = &BufferDescriptors[i];
		errcontext.arg = bufHdr;
		if (RelFileNodeEquals(bufHdr->tag.rnode, rel->rd_node))
		{
			if (bufHdr->flags & BM_DIRTY || bufHdr->cntxDirty)
			{
				PinBuffer(bufHdr);
				if (bufHdr->flags & BM_IO_IN_PROGRESS)
					WaitIO(bufHdr);
				LWLockRelease(BufMgrLock);

				/*
				 * Force XLOG flush for buffer' LSN
				 */
				recptr = BufferGetLSN(bufHdr);
				XLogFlush(recptr);

				/*
				 * Now it's safe to write buffer to disk
				 */

				LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);
				if (bufHdr->flags & BM_IO_IN_PROGRESS)
					WaitIO(bufHdr);

				if (bufHdr->flags & BM_DIRTY || bufHdr->cntxDirty)
				{
					bufHdr->flags &= ~BM_JUST_DIRTIED;
					StartBufferIO(bufHdr, false);		/* output IO start */

					LWLockRelease(BufMgrLock);

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

					if (status == SM_FAIL)		/* disk failure ?! */
						ereport(PANIC,
								(errcode(ERRCODE_IO_ERROR),
							  errmsg("could not write block %u of %u/%u",
									 bufHdr->tag.blockNum,
									 bufHdr->tag.rnode.tblNode,
									 bufHdr->tag.rnode.relNode)));

					BufferFlushCount++;

					LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);
					bufHdr->flags &= ~BM_IO_IN_PROGRESS;
					TerminateBufferIO(bufHdr);
					Assert(!(bufHdr->flags & BM_JUST_DIRTIED));
					bufHdr->flags &= ~BM_DIRTY;

					/*
					 * Note that it's safe to change cntxDirty here
					 * because of we protect it from upper writers by
					 * AccessExclusiveLock and from other bufmgr routines
					 * by BM_IO_IN_PROGRESS
					 */
					bufHdr->cntxDirty = false;
				}
				UnpinBuffer(bufHdr);
			}
			if (!(bufHdr->flags & BM_FREE))
			{
				LWLockRelease(BufMgrLock);
				error_context_stack = errcontext.previous;
				elog(WARNING, "FlushRelationBuffers(\"%s\", %u): block %u is referenced (private %ld, global %d)",
					 RelationGetRelationName(rel), firstDelBlock,
					 bufHdr->tag.blockNum,
					 PrivateRefCount[i], bufHdr->refcount);
				return -2;
			}
			if (bufHdr->tag.blockNum >= firstDelBlock)
				BufTableDelete(bufHdr);
		}
	}

	LWLockRelease(BufMgrLock);

	/* Pop the error context stack */
	error_context_stack = errcontext.previous;

	return 0;
}

#undef ReleaseBuffer

/*
 * ReleaseBuffer -- remove the pin on a buffer without
 *		marking it dirty.
 */
int
ReleaseBuffer(Buffer buffer)
{
	BufferDesc *bufHdr;

	if (BufferIsLocal(buffer))
	{
		Assert(LocalRefCount[-buffer - 1] > 0);
		LocalRefCount[-buffer - 1]--;
		return STATUS_OK;
	}

	if (BAD_BUFFER_ID(buffer))
		return STATUS_ERROR;

	bufHdr = &BufferDescriptors[buffer - 1];

	Assert(PrivateRefCount[buffer - 1] > 0);
	if (PrivateRefCount[buffer - 1] > 1)
		PrivateRefCount[buffer - 1]--;
	else
	{
		LWLockAcquire(BufMgrLock, LW_EXCLUSIVE);
		UnpinBuffer(bufHdr);
		LWLockRelease(BufMgrLock);
	}

	return STATUS_OK;
}

#ifdef NOT_USED
void
IncrBufferRefCount_Debug(char *file, int line, Buffer buffer)
{
	IncrBufferRefCount(buffer);
	if (ShowPinTrace && !BufferIsLocal(buffer) && is_userbuffer(buffer))
	{
		BufferDesc *buf = &BufferDescriptors[buffer - 1];

		fprintf(stderr, "PIN(Incr) %d rel = %u/%u, blockNum = %u, \
refcount = %ld, file: %s, line: %d\n",
				buffer,
				buf->tag.rnode.tblNode, buf->tag.rnode.relNode,
				buf->tag.blockNum,
				PrivateRefCount[buffer - 1], file, line);
	}
}
#endif

#ifdef NOT_USED
void
ReleaseBuffer_Debug(char *file, int line, Buffer buffer)
{
	ReleaseBuffer(buffer);
	if (ShowPinTrace && !BufferIsLocal(buffer) && is_userbuffer(buffer))
	{
		BufferDesc *buf = &BufferDescriptors[buffer - 1];

		fprintf(stderr, "UNPIN(Rel) %d rel = %u/%u, blockNum = %u, \
refcount = %ld, file: %s, line: %d\n",
				buffer,
				buf->tag.rnode.tblNode, buf->tag.rnode.relNode,
				buf->tag.blockNum,
				PrivateRefCount[buffer - 1], file, line);
	}
}
#endif

#ifdef NOT_USED
Buffer
ReleaseAndReadBuffer_Debug(char *file,
						   int line,
						   Buffer buffer,
						   Relation relation,
						   BlockNumber blockNum)
{
	bool		bufferValid;
	Buffer		b;

	bufferValid = BufferIsValid(buffer);
	b = ReleaseAndReadBuffer(buffer, relation, blockNum);
	if (ShowPinTrace && bufferValid && BufferIsLocal(buffer)
		&& is_userbuffer(buffer))
	{