mp_fopen.c

这是国外的resip协议栈

/*-
 * See the file LICENSE for redistribution information.
 *
 * Copyright (c) 1996-2004
 *	Sleepycat Software.  All rights reserved.
 *
 * $Id: mp_fopen.c,v 11.143 2004/10/15 16:59:43 bostic Exp $
 */

#include "db_config.h"

#ifndef NO_SYSTEM_INCLUDES
#include <sys/types.h>

#include <string.h>
#endif

#include "db_int.h"
#include "dbinc/db_shash.h"
#include "dbinc/log.h"
#include "dbinc/mp.h"

/*
 * __memp_fopen_pp --
 *	DB_MPOOLFILE->open pre/post processing.
 *
 * PUBLIC: int __memp_fopen_pp
 * PUBLIC:     __P((DB_MPOOLFILE *, const char *, u_int32_t, int, size_t));
 */
int
__memp_fopen_pp(dbmfp, path, flags, mode, pagesize)
	DB_MPOOLFILE *dbmfp;
	const char *path;
	u_int32_t flags;
	int mode;
	size_t pagesize;
{
	DB_ENV *dbenv;
	int rep_check, ret;

	dbenv = dbmfp->dbenv;

	PANIC_CHECK(dbenv);

	/* Validate arguments. */
	if ((ret = __db_fchk(dbenv, "DB_MPOOLFILE->open", flags,
	    DB_CREATE | DB_DIRECT | DB_EXTENT |
	    DB_NOMMAP | DB_ODDFILESIZE | DB_RDONLY | DB_TRUNCATE)) != 0)
		return (ret);

	/*
	 * Require a non-zero, power-of-two pagesize, smaller than the
	 * clear length.
	 */
	if (pagesize == 0 || !POWER_OF_TWO(pagesize)) {
		__db_err(dbenv,
		    "DB_MPOOLFILE->open: page sizes must be a power-of-2");
		return (EINVAL);
	}
	if (dbmfp->clear_len > pagesize) {
		__db_err(dbenv,
		    "DB_MPOOLFILE->open: clear length larger than page size");
		return (EINVAL);
	}

	/* Read-only checks, and local flag. */
	if (LF_ISSET(DB_RDONLY) && path == NULL) {
		__db_err(dbenv,
		    "DB_MPOOLFILE->open: temporary files can't be readonly");
		return (EINVAL);
	}

	rep_check = IS_ENV_REPLICATED(dbenv) ? 1 : 0;
	if (rep_check)
		__env_rep_enter(dbenv);
	ret = __memp_fopen(dbmfp, NULL, path, flags, mode, pagesize);
	if (rep_check)
		__env_db_rep_exit(dbenv);
	return (ret);
}

/*
 * __memp_fopen --
 *	DB_MPOOLFILE->open.
 *
 * PUBLIC: int __memp_fopen __P((DB_MPOOLFILE *,
 * PUBLIC:     MPOOLFILE *, const char *, u_int32_t, int, size_t));
 */
int
__memp_fopen(dbmfp, mfp, path, flags, mode, pgsize)
	DB_MPOOLFILE *dbmfp;
	MPOOLFILE *mfp;
	const char *path;
	u_int32_t flags;
	int mode;
	size_t pgsize;
{
	DB_ENV *dbenv;
	DB_MPOOL *dbmp;
	DB_MPOOLFILE *tmp_dbmfp;
	MPOOL *mp;
	db_pgno_t last_pgno;
	size_t maxmap;
	u_int32_t mbytes, bytes, oflags, pagesize;
	int refinc, ret;
	char *rpath;
	void *p;

	dbenv = dbmfp->dbenv;
	dbmp = dbenv->mp_handle;
	mp = dbmp->reginfo[0].primary;
	refinc = ret = 0;
	rpath = NULL;

	/*
	 * We're keeping the page size as a size_t in the public API, but
	 * it's a u_int32_t everywhere internally.
	 */
	pagesize = (u_int32_t)pgsize;

	/*
	 * We're called internally with a specified mfp, in which case the
	 * path is NULL, but we'll get the path from the underlying region
	 * information.  Otherwise, if the path is NULL, it's a temporary
	 * file -- we know we can't join any existing files, and we'll delay
	 * the open until we actually need to write the file.
	 */
	DB_ASSERT(mfp == NULL || path == NULL);

	if (mfp == NULL && path == NULL)
		goto alloc;

	/*
	 * Our caller may be able to tell us which underlying MPOOLFILE we
	 * need a handle for.
	 */
	if (mfp != NULL) {
		/*
		 * Deadfile can only be set if mpf_cnt goes to zero (or if we
		 * failed creating the file DB_AM_DISCARD).  Increment the ref
		 * count so the file cannot become dead and be unlinked.
		 */
		MUTEX_LOCK(dbenv, &mfp->mutex);
		if (!mfp->deadfile) {
			++mfp->mpf_cnt;
			refinc = 1;
		}
		MUTEX_UNLOCK(dbenv, &mfp->mutex);

		/*
		 * Test one last time to see if the file is dead -- it may have
		 * been removed.  This happens when a checkpoint trying to open
		 * the file to flush a buffer races with the Db::remove method.
		 * The error will be ignored, so don't output an error message.
		 */
		if (mfp->deadfile)
			return (EINVAL);
	}

	/* Convert MP open flags to DB OS-layer open flags. */
	oflags = 0;
	if (LF_ISSET(DB_CREATE))
		oflags |= DB_OSO_CREATE;
	if (LF_ISSET(DB_DIRECT))
		oflags |= DB_OSO_DIRECT;
	if (LF_ISSET(DB_RDONLY)) {
		F_SET(dbmfp, MP_READONLY);
		oflags |= DB_OSO_RDONLY;
	}

	/*
	 * Get the real name for this file and open it.
	 *
	 * Supply a page size so os_open can decide whether to turn buffering
	 * off if the DB_DIRECT_DB flag is set.
	 *
	 * Acquire the region lock if we're using a path from an underlying
	 * MPOOLFILE -- there's a race in accessing the path name stored in
	 * the region, __memp_nameop may be simultaneously renaming the file.
	 */
	if (mfp != NULL) {
		R_LOCK(dbenv, dbmp->reginfo);
		path = R_ADDR(dbmp->reginfo, mfp->path_off);
	}
	if ((ret =
	    __db_appname(dbenv, DB_APP_DATA, path, 0, NULL, &rpath)) == 0)
		ret = __os_open_extend(dbenv,
		    rpath, (u_int32_t)pagesize, oflags, mode, &dbmfp->fhp);
	if (mfp != NULL)
		R_UNLOCK(dbenv, dbmp->reginfo);
	if (ret != 0) {
		/* If it's a Queue extent file, it may not exist, that's OK. */
		if (!LF_ISSET(DB_EXTENT))
			__db_err(dbenv, "%s: %s", rpath, db_strerror(ret));
		goto err;
	}

	/*
	 * Cache file handles are shared, and have mutexes to protect the
	 * underlying file handle across seek and read/write calls.
	 */
	dbmfp->fhp->ref = 1;
	if (F_ISSET(dbenv, DB_ENV_THREAD) &&
	    (ret = __db_mutex_setup(dbenv, dbmp->reginfo,
	    &dbmfp->fhp->mutexp, MUTEX_ALLOC | MUTEX_THREAD)) != 0)
		goto err;

	/*
	 * Figure out the file's size.
	 *
	 * !!!
	 * We can't use off_t's here, or in any code in the mainline library
	 * for that matter.  (We have to use them in the os stubs, of course,
	 * as there are system calls that take them as arguments.)  The reason
	 * is some customers build in environments where an off_t is 32-bits,
	 * but still run where offsets are 64-bits, and they pay us a lot of
	 * money.
	 */
	if ((ret = __os_ioinfo(
	    dbenv, rpath, dbmfp->fhp, &mbytes, &bytes, NULL)) != 0) {
		__db_err(dbenv, "%s: %s", rpath, db_strerror(ret));
		goto err;
	}

	/*
	 * Get the file id if we weren't given one.  Generated file id's
	 * don't use timestamps, otherwise there'd be no chance of any
	 * other process joining the party.
	 */
	if (!F_ISSET(dbmfp, MP_FILEID_SET) &&
	    (ret = __os_fileid(dbenv, rpath, 0, dbmfp->fileid)) != 0)
		goto err;

	if (mfp != NULL)
		goto have_mfp;

	/*
	 * If not creating a temporary file, walk the list of MPOOLFILE's,
	 * looking for a matching file.  Files backed by temporary files
	 * or previously removed files can't match.
	 *
	 * DB_TRUNCATE support.
	 *
	 * The fileID is a filesystem unique number (e.g., a UNIX dev/inode
	 * pair) plus a timestamp.  If files are removed and created in less
	 * than a second, the fileID can be repeated.  The problem with
	 * repetition happens when the file that previously had the fileID
	 * value still has pages in the pool, since we don't want to use them
	 * to satisfy requests for the new file.
	 *
	 * Because the DB_TRUNCATE flag reuses the dev/inode pair, repeated
	 * opens with that flag set guarantees matching fileIDs when the
	 * machine can open a file and then re-open with truncate within a
	 * second.  For this reason, we pass that flag down, and, if we find
	 * a matching entry, we ensure that it's never found again, and we
	 * create a new entry for the current request.
	 */
	R_LOCK(dbenv, dbmp->reginfo);
	for (mfp = SH_TAILQ_FIRST(&mp->mpfq, __mpoolfile);
	    mfp != NULL; mfp = SH_TAILQ_NEXT(mfp, q, __mpoolfile)) {
		/* Skip dead files and temporary files. */
		if (mfp->deadfile || F_ISSET(mfp, MP_TEMP))
			continue;

		/* Skip non-matching files. */
		if (memcmp(dbmfp->fileid, R_ADDR(dbmp->reginfo,
		    mfp->fileid_off), DB_FILE_ID_LEN) != 0)
			continue;

		/*
		 * If the file is being truncated, remove it from the system
		 * and create a new entry.
		 *
		 * !!!
		 * We should be able to set mfp to NULL and break out of the
		 * loop, but I like the idea of checking all the entries.
		 */
		if (LF_ISSET(DB_TRUNCATE)) {
			MUTEX_LOCK(dbenv, &mfp->mutex);
			mfp->deadfile = 1;
			MUTEX_UNLOCK(dbenv, &mfp->mutex);
			continue;
		}

		/*
		 * Some things about a file cannot be changed: the clear length,
		 * page size, or lSN location.
		 *
		 * The file type can change if the application's pre- and post-
		 * processing needs change.  For example, an application that
		 * created a hash subdatabase in a database that was previously
		 * all btree.
		 *
		 * !!!
		 * We do not check to see if the pgcookie information changed,
		 * or update it if it is.
		 */
		if (dbmfp->clear_len != mfp->clear_len ||
		    pagesize != mfp->stat.st_pagesize ||
		    dbmfp->lsn_offset != mfp->lsn_off) {
			__db_err(dbenv,
		    "%s: clear length, page size or LSN location changed",
			    path);
			R_UNLOCK(dbenv, dbmp->reginfo);
			ret = EINVAL;
			goto err;
		}

		/*
		 * Check to see if this file has died while we waited.
		 *
		 * We normally don't lock the deadfile field when we read it as
		 * we only care if the field is zero or non-zero.  We do lock
		 * on read when searching for a matching MPOOLFILE so that two
		 * threads of control don't race between setting the deadfile
		 * bit and incrementing the reference count, that is, a thread
		 * of control decrementing the reference count and then setting
		 * deadfile because the reference count is 0 blocks us finding
		 * the file without knowing it's about to be marked dead.
		 */
		MUTEX_LOCK(dbenv, &mfp->mutex);
		if (mfp->deadfile) {
			MUTEX_UNLOCK(dbenv, &mfp->mutex);
			continue;
		}
		++mfp->mpf_cnt;
		refinc = 1;
		MUTEX_UNLOCK(dbenv, &mfp->mutex);

		if (dbmfp->ftype != 0)
			mfp->ftype = dbmfp->ftype;

		break;
	}
	R_UNLOCK(dbenv, dbmp->reginfo);

	if (mfp != NULL)
		goto have_mfp;

alloc:	/* Allocate and initialize a new MPOOLFILE. */
	if ((ret = __memp_alloc(
	    dbmp, dbmp->reginfo, NULL, sizeof(MPOOLFILE), NULL, &mfp)) != 0)
		goto err;
	memset(mfp, 0, sizeof(MPOOLFILE));
	mfp->mpf_cnt = 1;
	mfp->ftype = dbmfp->ftype;
	mfp->stat.st_pagesize = pagesize;
	mfp->lsn_off = dbmfp->lsn_offset;
	mfp->clear_len = dbmfp->clear_len;
	mfp->priority = dbmfp->priority;
	if (dbmfp->gbytes != 0 || dbmfp->bytes != 0) {
		mfp->maxpgno = (db_pgno_t)
		    (dbmfp->gbytes * (GIGABYTE / mfp->stat.st_pagesize));
		mfp->maxpgno += (db_pgno_t)
		    ((dbmfp->bytes + mfp->stat.st_pagesize - 1) /
		    mfp->stat.st_pagesize);
	}
	if (FLD_ISSET(dbmfp->config_flags, DB_MPOOL_NOFILE))
		mfp->no_backing_file = 1;
	if (FLD_ISSET(dbmfp->config_flags, DB_MPOOL_UNLINK))
		mfp->unlink_on_close = 1;

	if (LF_ISSET(DB_TXN_NOT_DURABLE))
		F_SET(mfp, MP_NOT_DURABLE);
	if (LF_ISSET(DB_DURABLE_UNKNOWN | DB_RDONLY))
		F_SET(mfp, MP_DURABLE_UNKNOWN);
	if (LF_ISSET(DB_DIRECT))
		F_SET(mfp, MP_DIRECT);
	if (LF_ISSET(DB_EXTENT))
		F_SET(mfp, MP_EXTENT);
	F_SET(mfp, MP_CAN_MMAP);

	if (path == NULL)
		F_SET(mfp, MP_TEMP);
	else {
		/*
		 * Don't permit files that aren't a multiple of the pagesize,
		 * and find the number of the last page in the file, all the
		 * time being careful not to overflow 32 bits.
		 *
		 * During verify or recovery, we might have to cope with a
		 * truncated file; if the file size is not a multiple of the
		 * page size, round down to a page, we'll take care of the
		 * partial page outside the mpool system.
		 */
		if (bytes % pagesize != 0) {
			if (LF_ISSET(DB_ODDFILESIZE))
				bytes -= (u_int32_t)(bytes % pagesize);
			else {
				__db_err(dbenv,
		    "%s: file size not a multiple of the pagesize", rpath);
				ret = EINVAL;
				goto err;
			}
		}

		/*
		 * If the user specifies DB_MPOOL_LAST or DB_MPOOL_NEW on a
		 * page get, we have to increment the last page in the file.
		 * Figure it out and save it away.
		 *
		 * Note correction: page numbers are zero-based, not 1-based.
		 */
		last_pgno = (db_pgno_t)(mbytes * (MEGABYTE / pagesize));
		last_pgno += (db_pgno_t)(bytes / pagesize);
		if (last_pgno != 0)
			--last_pgno;
		mfp->orig_last_pgno = mfp->last_pgno = last_pgno;

		/* Copy the file path into shared memory. */