md.c

关系型数据库 Postgresql 6.5.2

/*-------------------------------------------------------------------------
 *
 * md.c
 *	  This code manages relations that reside on magnetic disk.
 *
 * Copyright (c) 1994, Regents of the University of California
 *
 *
 * IDENTIFICATION
 *	  $Header: /usr/local/cvsroot/pgsql/src/backend/storage/smgr/md.c,v 1.46.2.2 1999/09/06 20:00:15 tgl Exp $
 *
 *-------------------------------------------------------------------------
 */
#include <unistd.h>
#include <fcntl.h>
#include <sys/file.h>

#include "postgres.h"

#include "catalog/catalog.h"
#include "miscadmin.h"
#include "storage/smgr.h"

#undef DIAGNOSTIC

/*
 *	The magnetic disk storage manager keeps track of open file descriptors
 *	in its own descriptor pool.  This happens for two reasons.	First, at
 *	transaction boundaries, we walk the list of descriptors and flush
 *	anything that we've dirtied in the current transaction.  Second, we want
 *	to support relations larger than the OS' file size limit (often 2GBytes).
 *	In order to do that, we break relations up into chunks of < 2GBytes
 *	and store one chunk in each of several files that represent the relation.
 *	See the BLCKSZ and RELSEG_SIZE configuration constants in include/config.h.
 *
 *	The file descriptor stored in the relation cache (see RelationGetFile())
 *	is actually an index into the Md_fdvec array.  -1 indicates not open.
 *
 *	When a relation is broken into multiple chunks, only the first chunk
 *	has its own entry in the Md_fdvec array; the remaining chunks have
 *	palloc'd MdfdVec objects that are chained onto the first chunk via the
 *	mdfd_chain links.  All chunks except the last MUST have size exactly
 *	equal to RELSEG_SIZE blocks --- see mdnblocks() and mdtruncate().
 */

typedef struct _MdfdVec
{
	int			mdfd_vfd;		/* fd number in vfd pool */
	uint16		mdfd_flags;		/* clean, dirty, free */
	int			mdfd_lstbcnt;	/* most recent block count */
	int			mdfd_nextFree;	/* next free vector */
#ifndef LET_OS_MANAGE_FILESIZE
	struct _MdfdVec *mdfd_chain;/* for large relations */
#endif
} MdfdVec;

static int	Nfds = 100;			/* initial/current size of Md_fdvec array */
static MdfdVec *Md_fdvec = (MdfdVec *) NULL;
static int	Md_Free = -1;		/* head of freelist of unused fdvec entries */
static int	CurFd = 0;			/* first never-used fdvec index */
static MemoryContext MdCxt;		/* context for all my allocations */

#define MDFD_DIRTY		(uint16) 0x01
#define MDFD_FREE		(uint16) 0x02

/* routines declared here */
static int _mdfd_getrelnfd(Relation reln);
static MdfdVec *_mdfd_openseg(Relation reln, int segno, int oflags);
static MdfdVec *_mdfd_getseg(Relation reln, int blkno);
static int	_fdvec_alloc(void);
static void _fdvec_free(int);
static BlockNumber _mdnblocks(File file, Size blcksz);

/*
 *	mdinit() -- Initialize private state for magnetic disk storage manager.
 *
 *		We keep a private table of all file descriptors.  Whenever we do
 *		a write to one, we mark it dirty in our table.	Whenever we force
 *		changes to disk, we mark the file descriptor clean.  At transaction
 *		commit, we force changes to disk for all dirty file descriptors.
 *		This routine allocates and initializes the table.
 *
 *		Returns SM_SUCCESS or SM_FAIL with errno set as appropriate.
 */
int
mdinit()
{
	MemoryContext oldcxt;
	int			i;

	MdCxt = (MemoryContext) CreateGlobalMemory("MdSmgr");
	if (MdCxt == (MemoryContext) NULL)
		return SM_FAIL;

	oldcxt = MemoryContextSwitchTo(MdCxt);
	Md_fdvec = (MdfdVec *) palloc(Nfds * sizeof(MdfdVec));
	MemoryContextSwitchTo(oldcxt);

	if (Md_fdvec == (MdfdVec *) NULL)
		return SM_FAIL;

	MemSet(Md_fdvec, 0, Nfds * sizeof(MdfdVec));

	/* Set free list */
	for (i = 0; i < Nfds; i++)
	{
		Md_fdvec[i].mdfd_nextFree = i + 1;
		Md_fdvec[i].mdfd_flags = MDFD_FREE;
	}
	Md_Free = 0;
	Md_fdvec[Nfds - 1].mdfd_nextFree = -1;

	return SM_SUCCESS;
}

int
mdcreate(Relation reln)
{
	int			fd,
				vfd;
	char	   *path;

	path = relpath(reln->rd_rel->relname.data);
#ifndef __CYGWIN32__
	fd = FileNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL, 0600);
#else
	fd = FileNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | O_BINARY, 0600);
#endif

	/*
	 * If the file already exists and is empty, we pretend that the create
	 * succeeded.  During bootstrap processing, we skip that check,
	 * because pg_time, pg_variable, and pg_log get created before their
	 * .bki file entries are processed.
	 *
	 * As the result of this pretence it was possible to have in pg_class > 1
	 * records with the same relname. Actually, it should be fixed in
	 * upper levels, too, but... -	vadim 05/06/97
	 */

	if (fd < 0)
	{
		if (!IsBootstrapProcessingMode())
			return -1;
#ifndef __CYGWIN32__
		fd = FileNameOpenFile(path, O_RDWR, 0600);		/* Bootstrap */
#else
		fd = FileNameOpenFile(path, O_RDWR | O_BINARY, 0600);	/* Bootstrap */
#endif
		if (fd < 0)
			return -1;
	}

	vfd = _fdvec_alloc();
	if (vfd < 0)
		return -1;

	Md_fdvec[vfd].mdfd_vfd = fd;
	Md_fdvec[vfd].mdfd_flags = (uint16) 0;
#ifndef LET_OS_MANAGE_FILESIZE
	Md_fdvec[vfd].mdfd_chain = (MdfdVec *) NULL;
#endif
	Md_fdvec[vfd].mdfd_lstbcnt = 0;

	return vfd;
}

/*
 *	mdunlink() -- Unlink a relation.
 */
int
mdunlink(Relation reln)
{
	int			nblocks;
	int			fd;
	MdfdVec    *v;
	MemoryContext oldcxt;

	/*
	 * Force all segments of the relation to be opened, so that we
	 * won't miss deleting any of them.
	 */
	nblocks = mdnblocks(reln);

	/*
	 * Clean out the mdfd vector, letting fd.c unlink the physical files.
	 *
	 * NOTE: We truncate the file(s) before deleting 'em, because if other
	 * backends are holding the files open, the unlink will fail on some
	 * platforms (think Microsoft).  Better a zero-size file gets left around
	 * than a big file.  Those other backends will be forced to close the
	 * relation by cache invalidation, but that probably hasn't happened yet.
	 */
	fd = RelationGetFile(reln);
	if (fd < 0)					/* should not happen */
		elog(ERROR, "mdunlink: mdnblocks didn't open relation");

	Md_fdvec[fd].mdfd_flags = (uint16) 0;

	oldcxt = MemoryContextSwitchTo(MdCxt);
#ifndef LET_OS_MANAGE_FILESIZE
	for (v = &Md_fdvec[fd]; v != (MdfdVec *) NULL;)
	{
		MdfdVec    *ov = v;
		FileTruncate(v->mdfd_vfd, 0);
		FileUnlink(v->mdfd_vfd);
		v = v->mdfd_chain;
		if (ov != &Md_fdvec[fd])
			pfree(ov);
	}
	Md_fdvec[fd].mdfd_chain = (MdfdVec *) NULL;
#else
	v = &Md_fdvec[fd];
	FileTruncate(v->mdfd_vfd, 0);
	FileUnlink(v->mdfd_vfd);
#endif
	MemoryContextSwitchTo(oldcxt);

	_fdvec_free(fd);

	/* be sure to mark relation closed */
	reln->rd_fd = -1;

	return SM_SUCCESS;
}

/*
 *	mdextend() -- Add a block to the specified relation.
 *
 *		This routine returns SM_FAIL or SM_SUCCESS, with errno set as
 *		appropriate.
 */
int
mdextend(Relation reln, char *buffer)
{
	long		pos;
	int			nblocks;
	MdfdVec    *v;

	nblocks = mdnblocks(reln);
	v = _mdfd_getseg(reln, nblocks);

	if ((pos = FileSeek(v->mdfd_vfd, 0L, SEEK_END)) < 0)
		return SM_FAIL;

	if (FileWrite(v->mdfd_vfd, buffer, BLCKSZ) != BLCKSZ)
		return SM_FAIL;

	/* remember that we did a write, so we can sync at xact commit */
	v->mdfd_flags |= MDFD_DIRTY;

	/* try to keep the last block count current, though it's just a hint */
#ifndef LET_OS_MANAGE_FILESIZE
	if ((v->mdfd_lstbcnt = (++nblocks % RELSEG_SIZE)) == 0)
		v->mdfd_lstbcnt = RELSEG_SIZE;

#ifdef DIAGNOSTIC
	if (_mdnblocks(v->mdfd_vfd, BLCKSZ) > RELSEG_SIZE
		|| v->mdfd_lstbcnt > RELSEG_SIZE)
		elog(FATAL, "segment too big!");
#endif
#else
	v->mdfd_lstbcnt = ++nblocks;
#endif

	return SM_SUCCESS;
}

/*
 *	mdopen() -- Open the specified relation.
 */
int
mdopen(Relation reln)
{
	char	   *path;
	int			fd;
	int			vfd;

	path = relpath(reln->rd_rel->relname.data);

#ifndef __CYGWIN32__
	fd = FileNameOpenFile(path, O_RDWR, 0600);
#else
	fd = FileNameOpenFile(path, O_RDWR | O_BINARY, 0600);
#endif

	if (fd < 0)
	{
		/* in bootstrap mode, accept mdopen as substitute for mdcreate */
		if (IsBootstrapProcessingMode())
		{
#ifndef __CYGWIN32__
			fd = FileNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL, 0600);
#else
			fd = FileNameOpenFile(path, O_RDWR | O_CREAT | O_EXCL | O_BINARY, 0600);
#endif
		}
		if (fd < 0)
		{
			elog(ERROR, "mdopen: couldn't open %s: %m", path);
			return -1;
		}
	}

	vfd = _fdvec_alloc();
	if (vfd < 0)
		return -1;

	Md_fdvec[vfd].mdfd_vfd = fd;
	Md_fdvec[vfd].mdfd_flags = (uint16) 0;
	Md_fdvec[vfd].mdfd_lstbcnt = _mdnblocks(fd, BLCKSZ);
#ifndef LET_OS_MANAGE_FILESIZE
	Md_fdvec[vfd].mdfd_chain = (MdfdVec *) NULL;

#ifdef DIAGNOSTIC
	if (Md_fdvec[vfd].mdfd_lstbcnt > RELSEG_SIZE)
		elog(FATAL, "segment too big on relopen!");
#endif
#endif

	return vfd;
}

/*
 *	mdclose() -- Close the specified relation, if it isn't closed already.
 *
 *		AND FREE fd vector! It may be re-used for other relation!
 *		reln should be flushed from cache after closing !..
 *
 *		Returns SM_SUCCESS or SM_FAIL with errno set as appropriate.
 */
int
mdclose(Relation reln)
{
	int			fd;
	MdfdVec    *v;
	MemoryContext oldcxt;

	fd = RelationGetFile(reln);
	if (fd < 0)
		return SM_SUCCESS;		/* already closed, so no work */

	oldcxt = MemoryContextSwitchTo(MdCxt);
#ifndef LET_OS_MANAGE_FILESIZE
	for (v = &Md_fdvec[fd]; v != (MdfdVec *) NULL;)
	{
		MdfdVec    *ov = v;

		/* if not closed already */
		if (v->mdfd_vfd >= 0)
		{

			/*
			 * We sync the file descriptor so that we don't need to reopen
			 * it at transaction commit to force changes to disk.
			 */

			FileSync(v->mdfd_vfd);
			FileClose(v->mdfd_vfd);

			/* mark this file descriptor as clean in our private table */
			v->mdfd_flags &= ~MDFD_DIRTY;
		}
		/* Now free vector */
		v = v->mdfd_chain;
		if (ov != &Md_fdvec[fd])
			pfree(ov);
	}

	Md_fdvec[fd].mdfd_chain = (MdfdVec *) NULL;
#else
	v = &Md_fdvec[fd];
	if (v != (MdfdVec *) NULL)
	{
		if (v->mdfd_vfd >= 0)
		{

			/*
			 * We sync the file descriptor so that we don't need to reopen
			 * it at transaction commit to force changes to disk.
			 */

			FileSync(v->mdfd_vfd);
			FileClose(v->mdfd_vfd);

			/* mark this file descriptor as clean in our private table */
			v->mdfd_flags &= ~MDFD_DIRTY;
		}
	}
#endif
	MemoryContextSwitchTo(oldcxt);

	_fdvec_free(fd);

	/* be sure to mark relation closed */
	reln->rd_fd = -1;

	return SM_SUCCESS;
}

/*
 *	mdread() -- Read the specified block from a relation.
 *
 *		Returns SM_SUCCESS or SM_FAIL.
 */
int
mdread(Relation reln, BlockNumber blocknum, char *buffer)
{
	int			status;
	long		seekpos;
	int			nbytes;
	MdfdVec    *v;

	v = _mdfd_getseg(reln, blocknum);

#ifndef LET_OS_MANAGE_FILESIZE
	seekpos = (long) (BLCKSZ * (blocknum % RELSEG_SIZE));

#ifdef DIAGNOSTIC
	if (seekpos >= BLCKSZ * RELSEG_SIZE)
		elog(FATAL, "seekpos too big!");
#endif
#else
	seekpos = (long) (BLCKSZ * (blocknum));
#endif

	if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
		return SM_FAIL;

	status = SM_SUCCESS;
	if ((nbytes = FileRead(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
	{
		if (nbytes == 0)
			MemSet(buffer, 0, BLCKSZ);
		else
			status = SM_FAIL;
	}

	return status;
}

/*
 *	mdwrite() -- Write the supplied block at the appropriate location.
 *
 *		Returns SM_SUCCESS or SM_FAIL.
 */
int
mdwrite(Relation reln, BlockNumber blocknum, char *buffer)
{
	int			status;
	long		seekpos;
	MdfdVec    *v;

	v = _mdfd_getseg(reln, blocknum);

#ifndef LET_OS_MANAGE_FILESIZE
	seekpos = (long) (BLCKSZ * (blocknum % RELSEG_SIZE));
#ifdef DIAGNOSTIC
	if (seekpos >= BLCKSZ * RELSEG_SIZE)
		elog(FATAL, "seekpos too big!");
#endif
#else
	seekpos = (long) (BLCKSZ * (blocknum));
#endif

	if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
		return SM_FAIL;

	status = SM_SUCCESS;
	if (FileWrite(v->mdfd_vfd, buffer, BLCKSZ) != BLCKSZ)
		status = SM_FAIL;

	v->mdfd_flags |= MDFD_DIRTY;

	return status;
}

/*
 *	mdflush() -- Synchronously write a block to disk.
 *
 *		This is exactly like mdwrite(), but doesn't return until the file
 *		system buffer cache has been flushed.
 */
int
mdflush(Relation reln, BlockNumber blocknum, char *buffer)
{
	int			status;
	long		seekpos;
	MdfdVec    *v;

	v = _mdfd_getseg(reln, blocknum);

#ifndef LET_OS_MANAGE_FILESIZE
	seekpos = (long) (BLCKSZ * (blocknum % RELSEG_SIZE));
#ifdef DIAGNOSTIC
	if (seekpos >= BLCKSZ * RELSEG_SIZE)
		elog(FATAL, "seekpos too big!");
#endif
#else
	seekpos = (long) (BLCKSZ * (blocknum));
#endif

	if (FileSeek(v->mdfd_vfd, seekpos, SEEK_SET) != seekpos)
		return SM_FAIL;

	/* write and sync the block */
	status = SM_SUCCESS;
	if (FileWrite(v->mdfd_vfd, buffer, BLCKSZ) != BLCKSZ
		|| FileSync(v->mdfd_vfd) < 0)
		status = SM_FAIL;

	/*
	 * By here, the block is written and changes have been forced to
	 * stable storage.	Mark the descriptor as clean until the next write,
	 * so we don't sync it again unnecessarily at transaction commit.
	 */

	v->mdfd_flags &= ~MDFD_DIRTY;

	return status;
}

/*
 *	mdblindwrt() -- Write a block to disk blind.
 *
 *		We have to be able to do this using only the name and OID of
 *		the database and relation in which the block belongs.  This
 *		is a synchronous write.
 */
int
mdblindwrt(char *dbstr,
		   char *relstr,
		   Oid dbid,
		   Oid relid,
		   BlockNumber blkno,
		   char *buffer)
{
	int			fd;
	int			segno;
	long		seekpos;