freespace.c

PostgreSQL7.4.6 for Linux

	/*
	 * Note we don't record info about a relation unless there's already
	 * an FSM entry for it, implying someone has done GetPageWithFreeSpace
	 * for it.	Inactive rels thus will not clutter the map simply by
	 * being vacuumed.
	 */
	fsmrel = lookup_fsm_rel(rel);
	if (fsmrel)
	{
		int			curAlloc;
		int			curAllocPages;
		FSMPageData *newLocation;

		curAlloc = realloc_fsm_rel(fsmrel, nPages, false);
		curAllocPages = curAlloc * CHUNKPAGES;

		/*
		 * If the data fits in our current allocation, just copy it;
		 * otherwise must compress.
		 */
		newLocation = (FSMPageData *)
			(FreeSpaceMap->arena + fsmrel->firstChunk * CHUNKBYTES);
		if (nPages <= curAllocPages)
		{
			int			i;

			for (i = 0; i < nPages; i++)
			{
				BlockNumber page = pageSpaces[i].blkno;
				Size		avail = pageSpaces[i].avail;

				/* Check caller provides sorted data */
				if (i > 0 && page <= pageSpaces[i - 1].blkno)
					elog(ERROR, "free-space data is not in page order");
				FSMPageSetPageNum(newLocation, page);
				FSMPageSetSpace(newLocation, avail);
				newLocation++;
			}
			fsmrel->storedPages = nPages;
		}
		else
		{
			pack_incoming_pages(newLocation, curAllocPages,
								pageSpaces, nPages);
			fsmrel->storedPages = curAllocPages;
		}
	}
	LWLockRelease(FreeSpaceLock);
}

/*
 * GetFreeIndexPage - like GetPageWithFreeSpace, but for indexes
 */
BlockNumber
GetFreeIndexPage(RelFileNode *rel)
{
	FSMRelation *fsmrel;
	BlockNumber freepage;

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);

	/*
	 * We always add a rel to the hashtable when it is inquired about.
	 */
	fsmrel = create_fsm_rel(rel);

	freepage = find_index_free_space(fsmrel);
	LWLockRelease(FreeSpaceLock);
	return freepage;
}

/*
 * RecordIndexFreeSpace - like RecordRelationFreeSpace, but for indexes
 */
void
RecordIndexFreeSpace(RelFileNode *rel,
					 int nPages,
					 BlockNumber *pages)
{
	FSMRelation *fsmrel;

	/* Limit nPages to something sane */
	if (nPages < 0)
		nPages = 0;
	else if (nPages > MaxFSMPages)
		nPages = MaxFSMPages;

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);

	/*
	 * Note we don't record info about a relation unless there's already
	 * an FSM entry for it, implying someone has done GetFreeIndexPage for
	 * it.	Inactive rels thus will not clutter the map simply by being
	 * vacuumed.
	 */
	fsmrel = lookup_fsm_rel(rel);
	if (fsmrel)
	{
		int			curAlloc;
		int			curAllocPages;
		int			i;
		IndexFSMPageData *newLocation;

		curAlloc = realloc_fsm_rel(fsmrel, nPages, true);
		curAllocPages = curAlloc * INDEXCHUNKPAGES;

		/*
		 * If the data fits in our current allocation, just copy it;
		 * otherwise must compress.  But compression is easy: we merely
		 * forget extra pages.
		 */
		newLocation = (IndexFSMPageData *)
			(FreeSpaceMap->arena + fsmrel->firstChunk * CHUNKBYTES);
		if (nPages > curAllocPages)
			nPages = curAllocPages;

		for (i = 0; i < nPages; i++)
		{
			BlockNumber page = pages[i];

			/* Check caller provides sorted data */
			if (i > 0 && page <= pages[i - 1])
				elog(ERROR, "free-space data is not in page order");
			IndexFSMPageSetPageNum(newLocation, page);
			newLocation++;
		}
		fsmrel->storedPages = nPages;
	}
	LWLockRelease(FreeSpaceLock);
}

/*
 * FreeSpaceMapTruncateRel - adjust for truncation of a relation.
 *
 * We need to delete any stored data past the new relation length, so that
 * we don't bogusly return removed block numbers.
 */
void
FreeSpaceMapTruncateRel(RelFileNode *rel, BlockNumber nblocks)
{
	FSMRelation *fsmrel;

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);
	fsmrel = lookup_fsm_rel(rel);
	if (fsmrel)
	{
		int			pageIndex;

		/* Use lookup to locate first entry >= nblocks */
		(void) lookup_fsm_page_entry(fsmrel, nblocks, &pageIndex);
		/* Delete all such entries */
		fsmrel->storedPages = pageIndex;
		/* XXX should we adjust rel's lastPageCount and sumRequests? */
	}
	LWLockRelease(FreeSpaceLock);
}

/*
 * FreeSpaceMapForgetRel - forget all about a relation.
 *
 * This is called when a relation is deleted.  Although we could just let
 * the rel age out of the map, it's better to reclaim and reuse the space
 * sooner.
 */
void
FreeSpaceMapForgetRel(RelFileNode *rel)
{
	FSMRelation *fsmrel;

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);
	fsmrel = lookup_fsm_rel(rel);
	if (fsmrel)
		delete_fsm_rel(fsmrel);
	LWLockRelease(FreeSpaceLock);
}

/*
 * FreeSpaceMapForgetDatabase - forget all relations of a database.
 *
 * This is called during DROP DATABASE.  As above, might as well reclaim
 * map space sooner instead of later.
 *
 * XXX when we implement tablespaces, target Oid will need to be tablespace
 * ID not database ID.
 */
void
FreeSpaceMapForgetDatabase(Oid dbid)
{
	FSMRelation *fsmrel,
			   *nextrel;

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);
	for (fsmrel = FreeSpaceMap->usageList; fsmrel; fsmrel = nextrel)
	{
		nextrel = fsmrel->nextUsage;	/* in case we delete it */
		if (fsmrel->key.tblNode == dbid)
			delete_fsm_rel(fsmrel);
	}
	LWLockRelease(FreeSpaceLock);
}

/*
 * PrintFreeSpaceMapStatistics - print statistics about FSM contents
 *
 * The info is sent to ereport() with the specified message level.	This is
 * intended for use during VACUUM.
 */
void
PrintFreeSpaceMapStatistics(int elevel)
{
	FSMRelation *fsmrel;
	int			storedPages = 0;
	int			numRels;
	double		sumRequests;
	double		needed;

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);
	/* Count total space used --- tedious, but seems useful */
	for (fsmrel = FreeSpaceMap->firstRel;
		 fsmrel != NULL;
		 fsmrel = fsmrel->nextPhysical)
		storedPages += fsmrel->storedPages;
	/* Copy other stats before dropping lock */
	numRels = FreeSpaceMap->numRels;
	sumRequests = FreeSpaceMap->sumRequests;
	LWLockRelease(FreeSpaceLock);

	/* Convert stats to actual number of page slots needed */
	needed = (sumRequests + numRels) * CHUNKPAGES;

	ereport(elevel,
			(errmsg("free space map: %d relations, %d pages stored; %.0f total pages needed",
					numRels, storedPages, needed),
			 errdetail("Allocated FSM size: %d relations + %d pages = %.0f kB shared memory.",
					   MaxFSMRelations, MaxFSMPages,
					   (double) FreeSpaceShmemSize() / 1024.0)));
}

/*
 * DumpFreeSpaceMap - dump contents of FSM into a disk file for later reload
 *
 * This is expected to be called during database shutdown, after updates to
 * the FSM have stopped.  We lock the FreeSpaceLock but that's purely pro
 * forma --- if anyone else is still accessing FSM, there's a problem.
 */
void
DumpFreeSpaceMap(void)
{
	FILE	   *fp;
	char		cachefilename[MAXPGPATH];
	FsmCacheFileHeader header;
	FSMRelation *fsmrel;

	/* Try to create file */
	snprintf(cachefilename, sizeof(cachefilename), "%s/%s",
			 DataDir, FSM_CACHE_FILENAME);

	unlink(cachefilename);		/* in case it exists w/wrong permissions */

	fp = AllocateFile(cachefilename, PG_BINARY_W);
	if (fp == NULL)
	{
		elog(LOG, "could not write \"%s\": %m", cachefilename);
		return;
	}

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);

	/* Write file header */
	MemSet(&header, 0, sizeof(header));
	strcpy(header.label, FSM_CACHE_LABEL);
	header.endian = FSM_CACHE_ENDIAN;
	header.version = FSM_CACHE_VERSION;
	header.numRels = FreeSpaceMap->numRels;
	if (fwrite(&header, 1, sizeof(header), fp) != sizeof(header))
		goto write_failed;

	/* For each relation, in order from least to most recently used... */
	for (fsmrel = FreeSpaceMap->usageListTail;
		 fsmrel != NULL;
		 fsmrel = fsmrel->priorUsage)
	{
		FsmCacheRelHeader relheader;
		int			nPages;

		/* Write relation header */
		MemSet(&relheader, 0, sizeof(relheader));
		relheader.key = fsmrel->key;
		relheader.isIndex = fsmrel->isIndex;
		relheader.avgRequest = fsmrel->avgRequest;
		relheader.lastPageCount = fsmrel->lastPageCount;
		relheader.storedPages = fsmrel->storedPages;
		if (fwrite(&relheader, 1, sizeof(relheader), fp) != sizeof(relheader))
			goto write_failed;

		/* Write the per-page data directly from the arena */
		nPages = fsmrel->storedPages;
		if (nPages > 0)
		{
			Size		len;
			char	   *data;

			if (fsmrel->isIndex)
				len = nPages * sizeof(IndexFSMPageData);
			else
				len = nPages * sizeof(FSMPageData);
			data = (char *)
				(FreeSpaceMap->arena + fsmrel->firstChunk * CHUNKBYTES);
			if (fwrite(data, 1, len, fp) != len)
				goto write_failed;
		}
	}

	/* Clean up */
	LWLockRelease(FreeSpaceLock);

	FreeFile(fp);

	return;

write_failed:
	elog(LOG, "could not write \"%s\": %m", cachefilename);

	/* Clean up */
	LWLockRelease(FreeSpaceLock);

	FreeFile(fp);

	/* Remove busted cache file */
	unlink(cachefilename);
}

/*
 * LoadFreeSpaceMap - load contents of FSM from a disk file
 *
 * This is expected to be called during database startup, before any FSM
 * updates begin.  We lock the FreeSpaceLock but that's purely pro
 * forma --- if anyone else is accessing FSM yet, there's a problem.
 *
 * Notes: no complaint is issued if no cache file is found.  If the file is
 * found, it is deleted after reading.	Thus, if we crash without a clean
 * shutdown, the next cycle of life starts with no FSM data.  To do otherwise,
 * we'd need to do significantly more validation in this routine, because of
 * the likelihood that what is in the dump file would be out-of-date, eg
 * there might be entries for deleted or truncated rels.
 */
void
LoadFreeSpaceMap(void)
{
	FILE	   *fp;
	char		cachefilename[MAXPGPATH];
	FsmCacheFileHeader header;
	int			relno;

	/* Try to open file */
	snprintf(cachefilename, sizeof(cachefilename), "%s/%s",
			 DataDir, FSM_CACHE_FILENAME);

	fp = AllocateFile(cachefilename, PG_BINARY_R);
	if (fp == NULL)
	{
		if (errno != ENOENT)
			elog(LOG, "could not read \"%s\": %m", cachefilename);
		return;
	}

	LWLockAcquire(FreeSpaceLock, LW_EXCLUSIVE);

	/* Read and verify file header */
	if (fread(&header, 1, sizeof(header), fp) != sizeof(header) ||
		strcmp(header.label, FSM_CACHE_LABEL) != 0 ||
		header.endian != FSM_CACHE_ENDIAN ||
		header.version != FSM_CACHE_VERSION ||
		header.numRels < 0)
	{
		elog(LOG, "bogus file header in \"%s\"", cachefilename);
		goto read_failed;
	}

	/* For each relation, in order from least to most recently used... */
	for (relno = 0; relno < header.numRels; relno++)
	{
		FsmCacheRelHeader relheader;
		Size		len;
		char	   *data;
		FSMRelation *fsmrel;
		int			nPages;
		int			curAlloc;
		int			curAllocPages;

		/* Read and verify relation header, as best we can */
		if (fread(&relheader, 1, sizeof(relheader), fp) != sizeof(relheader) ||
			(relheader.isIndex != false && relheader.isIndex != true) ||
			relheader.avgRequest >= BLCKSZ ||
			relheader.lastPageCount < 0 ||
			relheader.storedPages < 0)
		{
			elog(LOG, "bogus rel header in \"%s\"", cachefilename);
			goto read_failed;
		}

		/* Make sure lastPageCount doesn't exceed current MaxFSMPages */
		if (relheader.lastPageCount > MaxFSMPages)
			relheader.lastPageCount = MaxFSMPages;

		/* Read the per-page data */
		nPages = relheader.storedPages;
		if (relheader.isIndex)
			len = nPages * sizeof(IndexFSMPageData);
		else
			len = nPages * sizeof(FSMPageData);
		data = (char *) palloc(len + 1);		/* +1 to avoid palloc(0) */
		if (fread(data, 1, len, fp) != len)
		{
			elog(LOG, "premature EOF in \"%s\"", cachefilename);
			pfree(data);
			goto read_failed;
		}

		/*
		 * Okay, create the FSM entry and insert data into it.	Since the
		 * rels were stored in reverse usage order, at the end of the loop
		 * they will be correctly usage-ordered in memory; and if
		 * MaxFSMRelations is less than it used to be, we will correctly
		 * drop the least recently used ones.
		 */
		fsmrel = create_fsm_rel(&relheader.key);
		fsmrel->avgRequest = relheader.avgRequest;

		curAlloc = realloc_fsm_rel(fsmrel, relheader.lastPageCount,
								   relheader.isIndex);
		if (relheader.isIndex)
		{
			IndexFSMPageData *newLocation;

			curAllocPages = curAlloc * INDEXCHUNKPAGES;

			/*
			 * If the data fits in our current allocation, just copy it;
			 * otherwise must compress.  But compression is easy: we
			 * merely forget extra pages.
			 */
			newLocation = (IndexFSMPageData *)
				(FreeSpaceMap->arena + fsmrel->firstChunk * CHUNKBYTES);
			if (nPages > curAllocPages)
				nPages = curAllocPages;
			memcpy(newLocation, data, nPages * sizeof(IndexFSMPageData));
			fsmrel->storedPages = nPages;
		}
		else
		{
			FSMPageData *newLocation;

			curAllocPages = curAlloc * CHUNKPAGES;

			/*
			 * If the data fits in our current allocation, just copy it;
			 * otherwise must compress.
			 */
			newLocation = (FSMPageData *)
				(FreeSpaceMap->arena + fsmrel->firstChunk * CHUNKBYTES);
			if (nPages <= curAllocPages)
			{
				memcpy(newLocation, data, nPages * sizeof(FSMPageData));
				fsmrel->storedPages = nPages;
			}
			else
			{
				pack_existing_pages(newLocation, curAllocPages,
									(FSMPageData *) data, nPages);
				fsmrel->storedPages = curAllocPages;