db.c

这是国外的resip协议栈

	    (ret = __dbreg_new_id(dbp, txn)) != 0)
		return (ret);

	/*
	 * Insert ourselves into the DB_ENV's dblist.  We allocate a
	 * unique ID to each {fileid, meta page number} pair, and to
	 * each temporary file (since they all have a zero fileid).
	 * This ID gives us something to use to tell which DB handles
	 * go with which databases in all the cursor adjustment
	 * routines, where we don't want to do a lot of ugly and
	 * expensive memcmps.
	 */
	MUTEX_THREAD_LOCK(dbenv, dbenv->dblist_mutexp);
	for (maxid = 0, ldbp = LIST_FIRST(&dbenv->dblist);
	    ldbp != NULL; ldbp = LIST_NEXT(ldbp, dblistlinks)) {
		if (fname != NULL &&
		    memcmp(ldbp->fileid, dbp->fileid, DB_FILE_ID_LEN) == 0 &&
		    ldbp->meta_pgno == dbp->meta_pgno)
			break;
		if (ldbp->adj_fileid > maxid)
			maxid = ldbp->adj_fileid;
	}

	/*
	 * If ldbp is NULL, we didn't find a match, or we weren't
	 * really looking because fname is NULL.  Assign the dbp an
	 * adj_fileid one higher than the largest we found, and
	 * insert it at the head of the master dbp list.
	 *
	 * If ldbp is not NULL, it is a match for our dbp.  Give dbp
	 * the same ID that ldbp has, and add it after ldbp so they're
	 * together in the list.
	 */
	if (ldbp == NULL) {
		dbp->adj_fileid = maxid + 1;
		LIST_INSERT_HEAD(&dbenv->dblist, dbp, dblistlinks);
	} else {
		dbp->adj_fileid = ldbp->adj_fileid;
		LIST_INSERT_AFTER(ldbp, dbp, dblistlinks);
	}
	MUTEX_THREAD_UNLOCK(dbenv, dbenv->dblist_mutexp);

	return (0);
}

/*
 * __db_dbenv_mpool --
 *	Set up the underlying environment cache during a db_open.
 *
 * PUBLIC: int __db_dbenv_mpool __P((DB *, const char *, u_int32_t));
 */
int
__db_dbenv_mpool(dbp, fname, flags)
	DB *dbp;
	const char *fname;
	u_int32_t flags;
{
	DB_ENV *dbenv;
	DBT pgcookie;
	DB_MPOOLFILE *mpf;
	DB_PGINFO pginfo;
	u_int32_t clear_len;
	int ftype, ret;

	COMPQUIET(mpf, NULL);

	dbenv = dbp->dbenv;

	/*
	 * If we need to pre- or post-process a file's pages on I/O, set the
	 * file type.  If it's a hash file, always call the pgin and pgout
	 * routines.  This means that hash files can never be mapped into
	 * process memory.  If it's a btree file and requires swapping, we
	 * need to page the file in and out.  This has to be right -- we can't
	 * mmap files that are being paged in and out.
	 */
	switch (dbp->type) {
	case DB_BTREE:
	case DB_RECNO:
		ftype = F_ISSET(dbp, DB_AM_SWAP | DB_AM_ENCRYPT | DB_AM_CHKSUM)
		    ? DB_FTYPE_SET : DB_FTYPE_NOTSET;
		clear_len = CRYPTO_ON(dbenv) ? dbp->pgsize : DB_PAGE_DB_LEN;
		break;
	case DB_HASH:
		ftype = DB_FTYPE_SET;
		clear_len = CRYPTO_ON(dbenv) ? dbp->pgsize : DB_PAGE_DB_LEN;
		break;
	case DB_QUEUE:
		ftype = F_ISSET(dbp,
		    DB_AM_SWAP | DB_AM_ENCRYPT | DB_AM_CHKSUM) ?
		    DB_FTYPE_SET : DB_FTYPE_NOTSET;
		clear_len = CRYPTO_ON(dbenv) ? dbp->pgsize : DB_PAGE_QUEUE_LEN;
		break;
	case DB_UNKNOWN:
		/*
		 * If we're running in the verifier, our database might
		 * be corrupt and we might not know its type--but we may
		 * still want to be able to verify and salvage.
		 *
		 * If we can't identify the type, it's not going to be safe
		 * to call __db_pgin--we pretty much have to give up all
		 * hope of salvaging cross-endianness.  Proceed anyway;
		 * at worst, the database will just appear more corrupt
		 * than it actually is, but at best, we may be able
		 * to salvage some data even with no metadata page.
		 */
		if (F_ISSET(dbp, DB_AM_VERIFYING)) {
			ftype = DB_FTYPE_NOTSET;
			clear_len = DB_PAGE_DB_LEN;
			break;
		}
		/* FALLTHROUGH */
	default:
		return (__db_unknown_type(dbenv, "DB->open", dbp->type));
	}

	mpf = dbp->mpf;

	(void)__memp_set_clear_len(mpf, clear_len);
	(void)__memp_set_fileid(mpf, dbp->fileid);
	(void)__memp_set_ftype(mpf, ftype);
	(void)__memp_set_lsn_offset(mpf, 0);

	pginfo.db_pagesize = dbp->pgsize;
	pginfo.flags =
	    F_ISSET(dbp, (DB_AM_CHKSUM | DB_AM_ENCRYPT | DB_AM_SWAP));
	pginfo.type = dbp->type;
	pgcookie.data = &pginfo;
	pgcookie.size = sizeof(DB_PGINFO);
	(void)__memp_set_pgcookie(mpf, &pgcookie);

	if ((ret = __memp_fopen(mpf, NULL, fname,
	    LF_ISSET(DB_RDONLY | DB_NOMMAP |
	    DB_ODDFILESIZE | DB_TRUNCATE) |
	    (F_ISSET(dbenv, DB_ENV_DIRECT_DB) ? DB_DIRECT : 0) |
	    (F_ISSET(dbp, DB_AM_NOT_DURABLE) ? DB_TXN_NOT_DURABLE : 0),
	    0, dbp->pgsize)) != 0)
		return (ret);

	return (0);
}

/*
 * __db_close --
 *	DB->close method.
 *
 * PUBLIC: int __db_close __P((DB *, DB_TXN *, u_int32_t));
 */
int
__db_close(dbp, txn, flags)
	DB *dbp;
	DB_TXN *txn;
	u_int32_t flags;
{
	DB_ENV *dbenv;
	int db_ref, deferred_close, ret, t_ret;

	dbenv = dbp->dbenv;
	deferred_close = ret = 0;

	/*
	 * Validate arguments, but as a DB handle destructor, we can't fail.
	 *
	 * Check for consistent transaction usage -- ignore errors.  Only
	 * internal callers specify transactions, so it's a serious problem
	 * if we get error messages.
	 */
	if (txn != NULL)
		(void)__db_check_txn(dbp, txn, DB_LOCK_INVALIDID, 0);

	/* Refresh the structure and close any underlying resources. */
	ret = __db_refresh(dbp, txn, flags, &deferred_close);

	/*
	 * If we've deferred the close because the logging of the close failed,
	 * return our failure right away without destroying the handle.
	 */
	if (deferred_close)
		return (ret);

	/* !!!
	 * This code has an apparent race between the moment we read and
	 * decrement dbenv->db_ref and the moment we check whether it's 0.
	 * However, if the environment is DBLOCAL, the user shouldn't have a
	 * reference to the dbenv handle anyway;  the only way we can get
	 * multiple dbps sharing a local dbenv is if we open them internally
	 * during something like a subdatabase open.  If any such thing is
	 * going on while the user is closing the original dbp with a local
	 * dbenv, someone's already badly screwed up, so there's no reason
	 * to bother engineering around this possibility.
	 */
	MUTEX_THREAD_LOCK(dbenv, dbenv->dblist_mutexp);
	db_ref = --dbenv->db_ref;
	MUTEX_THREAD_UNLOCK(dbenv, dbenv->dblist_mutexp);
	if (F_ISSET(dbenv, DB_ENV_DBLOCAL) && db_ref == 0 &&
	    (t_ret = __dbenv_close(dbenv, 0)) != 0 && ret == 0)
		ret = t_ret;

	/* Free the database handle. */
	memset(dbp, CLEAR_BYTE, sizeof(*dbp));
	__os_free(dbenv, dbp);

	return (ret);
}

/*
 * __db_refresh --
 *	Refresh the DB structure, releasing any allocated resources.
 * This does most of the work of closing files now because refresh
 * is what is used during abort processing (since we can't destroy
 * the actual handle) and during abort processing, we may have a
 * fully opened handle.
 *
 * PUBLIC: int __db_refresh __P((DB *, DB_TXN *, u_int32_t, int *));
 */
int
__db_refresh(dbp, txn, flags, deferred_closep)
	DB *dbp;
	DB_TXN *txn;
	u_int32_t flags;
	int *deferred_closep;
{
	DB *sdbp;
	DBC *dbc;
	DB_ENV *dbenv;
	DB_LOCKREQ lreq;
	DB_MPOOL *dbmp;
	int resync, ret, t_ret;

	ret = 0;

	dbenv = dbp->dbenv;

	/* If never opened, or not currently open, it's easy. */
	if (!F_ISSET(dbp, DB_AM_OPEN_CALLED))
		goto never_opened;

	/*
	 * If we have any secondary indices, disassociate them from us.
	 * We don't bother with the mutex here;  it only protects some
	 * of the ops that will make us core-dump mid-close anyway, and
	 * if you're trying to do something with a secondary *while* you're
	 * closing the primary, you deserve what you get.  The disassociation
	 * is mostly done just so we can close primaries and secondaries in
	 * any order--but within one thread of control.
	 */
	for (sdbp = LIST_FIRST(&dbp->s_secondaries);
	    sdbp != NULL; sdbp = LIST_NEXT(sdbp, s_links)) {
		LIST_REMOVE(sdbp, s_links);
		if ((t_ret = __db_disassociate(sdbp)) != 0 && ret == 0)
			ret = t_ret;
	}

	/*
	 * Sync the underlying access method.  Do before closing the cursors
	 * because DB->sync allocates cursors in order to write Recno backing
	 * source text files.
	 *
	 * Sync is slow on some systems, notably Solaris filesystems where the
	 * entire buffer cache is searched.  If we're in recovery, don't flush
	 * the file, it's not necessary.
	 */
	if (!LF_ISSET(DB_NOSYNC) &&
	    !F_ISSET(dbp, DB_AM_DISCARD | DB_AM_RECOVER) &&
	    (t_ret = __db_sync(dbp)) != 0 && ret == 0)
		ret = t_ret;

	/*
	 * Go through the active cursors and call the cursor recycle routine,
	 * which resolves pending operations and moves the cursors onto the
	 * free list.  Then, walk the free list and call the cursor destroy
	 * routine.  Note that any failure on a close is considered "really
	 * bad" and we just break out of the loop and force forward.
	 */
	resync = TAILQ_FIRST(&dbp->active_queue) == NULL ? 0 : 1;
	while ((dbc = TAILQ_FIRST(&dbp->active_queue)) != NULL)
		if ((t_ret = __db_c_close(dbc)) != 0) {
			if (ret == 0)
				ret = t_ret;
			break;
		}

	while ((dbc = TAILQ_FIRST(&dbp->free_queue)) != NULL)
		if ((t_ret = __db_c_destroy(dbc)) != 0) {
			if (ret == 0)
				ret = t_ret;
			break;
		}

	/*
	 * Close any outstanding join cursors.  Join cursors destroy themselves
	 * on close and have no separate destroy routine.  We don't have to set
	 * the resync flag here, because join cursors aren't write cursors.
	 */
	while ((dbc = TAILQ_FIRST(&dbp->join_queue)) != NULL)
		if ((t_ret = __db_join_close(dbc)) != 0) {
			if (ret == 0)
				ret = t_ret;
			break;
		}

	/*
	 * Sync the memory pool, even though we've already called DB->sync,
	 * because closing cursors can dirty pages by deleting items they
	 * referenced.
	 *
	 * Sync is slow on some systems, notably Solaris filesystems where the
	 * entire buffer cache is searched.  If we're in recovery, don't flush
	 * the file, it's not necessary.
	 */
	if (resync && !LF_ISSET(DB_NOSYNC) &&
	    !F_ISSET(dbp, DB_AM_DISCARD | DB_AM_RECOVER) &&
	    (t_ret = __memp_fsync(dbp->mpf)) != 0 && ret == 0)
		ret = t_ret;

	/*
	 * At this point, we haven't done anything to render the DB
	 * handle unusable, at least by a transaction abort.  Take the
	 * opportunity now to log the file close.  If this log fails
	 * and we're in a transaction, we have to bail out of the attempted
	 * close; we'll need a dbp in order to successfully abort the
	 * transaction, and we can't conjure a new one up because we haven't
	 * gotten out the dbreg_register record that represents the close.
	 * In this case, we put off actually closing the dbp until we've
	 * performed the abort.
	 */
	if (LOGGING_ON(dbp->dbenv)) {
		/*
		 * Discard the log file id, if any.  We want to log the close
		 * if and only if this is not a recovery dbp.
		 */
		if (F_ISSET(dbp, DB_AM_RECOVER))
			t_ret = __dbreg_revoke_id(dbp, 0, DB_LOGFILEID_INVALID);
		else {
			if ((t_ret = __dbreg_close_id(dbp,
			    txn, DBREG_CLOSE)) != 0 && txn != NULL) {
				/*
				 * We're in a txn and the attempt to log the
				 * close failed;  let the txn subsystem know
				 * that we need to destroy this dbp once we're
				 * done with the abort, then bail from the
				 * close.
				 *
				 * Note that if the attempt to put off the
				 * close -also- fails--which it won't unless
				 * we're out of heap memory--we're really
				 * screwed.  Panic.
				 */
				if ((ret =
				    __txn_closeevent(dbenv, txn, dbp)) != 0)
					return (__db_panic(dbenv, ret));
				if (deferred_closep != NULL)
					*deferred_closep = 1;
				return (t_ret);
			}
		}

		if (ret == 0)
			ret = t_ret;

		/* Discard the log FNAME. */
		if ((t_ret = __dbreg_teardown(dbp)) != 0 && ret == 0)
			ret = t_ret;
	}

	/* Close any handle we've been holding since the open.  */
	if (dbp->saved_open_fhp != NULL &&
	    (t_ret = __os_closehandle(dbenv, dbp->saved_open_fhp)) != 0 &&
	    ret == 0)
		ret = t_ret;

never_opened:
	/*
	 * Remove this DB handle from the DB_ENV's dblist, if it's been added.
	 *
	 * Close our reference to the underlying cache while locked, we don't
	 * want to race with a thread searching for our underlying cache link
	 * while opening a DB handle.
	 */
	MUTEX_THREAD_LOCK(dbenv, dbenv->dblist_mutexp);
	if (dbp->dblistlinks.le_prev != NULL) {
		LIST_REMOVE(dbp, dblistlinks);
		dbp->dblistlinks.le_prev = NULL;
	}

	/* Close the memory pool file handle. */
	if (dbp->mpf != NULL) {
		if ((t_ret = __memp_fclose(dbp->mpf,
		    F_ISSET(dbp, DB_AM_DISCARD) ? DB_MPOOL_DISCARD : 0)) != 0 &&
		    ret == 0)
			ret = t_ret;
		dbp->mpf = NULL;
	}

	MUTEX_THREAD_UNLOCK(dbenv, dbenv->dblist_mutexp);

	/*
	 * Call the access specific close function.
	 *
	 * We do this here rather than in __db_close as we need to do this when
	 * aborting an open so that file descriptors are closed and abort of
	 * renames can succeed on platforms that lock open files (such as
	 * Windows).  In particular, we need to ensure that all the extents
	 * associated with a queue are closed so that queue renames can be
	 * aborted.
	 *
	 * It is also important that we do this before releasing the handle
	 * lock, because dbremove and dbrename assume that once they have the
	 * handle lock, it is safe to modify the underlying file(s).
	 *
	 * !!!
	 * Because of where these functions are called in the DB handle close
	 * process, these routines can't do anything that would dirty pages or
	 * otherwise affect closing down the database.  Specifically, we can't
	 * abort and recover any of the information they control.
	 */
	if ((t_ret = __bam_db_close(dbp)) != 0 && ret == 0)
		ret = t_ret;
	if ((t_ret = __ham_db_close(dbp)) != 0 && ret == 0)
		ret = t_ret;
	if ((t_ret = __qam_db_close(dbp, dbp->flags)) != 0 && ret == 0)
		ret = t_ret;

	/*
	 * !!!
	 * At this point, the access-method specific information has been
	 * freed.  From now on, we can use the dbp, but not touch any
	 * access-method specific data.
	 */

	if (dbp->lid != DB_LOCK_INVALIDID) {
		/* We may have pending trade operations on this dbp. */
		if (txn != NULL)
			__txn_remlock(dbenv, txn, &dbp->handle_lock, dbp->lid);

		/* We may be holding the handle lock; release it. */
		lreq.op = DB_LOCK_PUT_ALL;
		lreq.obj = NULL;
		if ((t_ret = __lock_vec(dbenv,
		    dbp->lid, 0, &lreq, 1, NULL)) != 0 && ret == 0)
			ret = t_ret;

		if ((t_ret = __lock_id_free(dbenv, dbp->lid)) != 0 && ret == 0)
			ret = t_ret;
		dbp->lid = DB_LOCK_INVALIDID;
		LOCK_INIT(dbp->handle_lock);
	}

	/* Discard the locker ID allocated as the fileid. */
	if (F_ISSET(dbp, DB_AM_INMEM) && LOCKING_ON(dbenv) &&
	    (t_ret = __lock_id_free(dbenv, *(u_int32_t *)dbp->fileid)) != 0 &&
	    ret == 0)
		ret = t_ret;

	dbp->type = DB_UNKNOWN;