db_am.c

File system using stacked.

	DB_FH *fhp;
	int ret;

	PANIC_CHECK(dbp->dbenv);
	DB_ILLEGAL_BEFORE_OPEN(dbp, "DB->fd");

	/*
	 * XXX
	 * Truly spectacular layering violation.
	 */
	if ((ret = __mp_xxx_fh(dbp->mpf, &fhp)) != 0)
		return (ret);

	if (F_ISSET(fhp, DB_FH_VALID)) {
		*fdp = fhp->fd;
		return (0);
	} else {
#ifdef KBDB
		*fdp = NULL;
#else
		*fdp = -1;
#endif
		__db_err(dbp->dbenv, "DB does not have a valid file handle");
		return (ENOENT);
	}
}

/*
 * __db_get --
 *	Return a key/data pair.
 *
 * PUBLIC: int __db_get __P((DB *, DB_TXN *, DBT *, DBT *, u_int32_t));
 */
int
__db_get(dbp, txn, key, data, flags)
	DB *dbp;
	DB_TXN *txn;
	DBT *key, *data;
	u_int32_t flags;
{
	DBC *dbc;
	int mode, ret, t_ret;

	PANIC_CHECK(dbp->dbenv);
	DB_ILLEGAL_BEFORE_OPEN(dbp, "DB->get");

	if ((ret = __db_getchk(dbp, key, data, flags)) != 0)
		return (ret);

	/* Check for consistent transaction usage. */
	if ((ret = __db_check_txn(dbp, txn, DB_LOCK_INVALIDID, 1)) != 0)
		return (ret);

	mode = 0;
	if (LF_ISSET(DB_DIRTY_READ)) {
		mode = DB_DIRTY_READ;
		LF_CLR(DB_DIRTY_READ);
	}
	else if (flags == DB_CONSUME || flags == DB_CONSUME_WAIT)
		mode = DB_WRITELOCK;
	if ((ret = dbp->cursor(dbp, txn, &dbc, mode)) != 0)
		return (ret);

	DEBUG_LREAD(dbc, txn, "__db_get", key, NULL, flags);

	/*
	 * The DBC_TRANSIENT flag indicates that we're just doing a
	 * single operation with this cursor, and that in case of
	 * error we don't need to restore it to its old position--we're
	 * going to close it right away.  Thus, we can perform the get
	 * without duplicating the cursor, saving some cycles in this
	 * common case.
	 *
	 * SET_RET_MEM indicates that if key and/or data have no DBT
	 * flags set and DB manages the returned-data memory, that memory
	 * will belong to this handle, not to the underlying cursor.
	 */
	F_SET(dbc, DBC_TRANSIENT);
	SET_RET_MEM(dbc, dbp);

	if (LF_ISSET(~(DB_RMW | DB_MULTIPLE)) == 0)
		LF_SET(DB_SET);
	ret = dbc->c_get(dbc, key, data, flags);

	if ((t_ret = __db_c_close(dbc)) != 0 && ret == 0)
		ret = t_ret;

	return (ret);
}

/*
 * __db_put --
 *	Store a key/data pair.
 *
 * PUBLIC: int __db_put __P((DB *, DB_TXN *, DBT *, DBT *, u_int32_t));
 */
int
__db_put(dbp, txn, key, data, flags)
	DB *dbp;
	DB_TXN *txn;
	DBT *key, *data;
	u_int32_t flags;
{
	DBC *dbc;
	DBT tdata;
	DB_ENV *dbenv;
	int ret, t_ret, txn_local;

	dbc = NULL;
	dbenv = dbp->dbenv;
	txn_local = 0;

	PANIC_CHECK(dbenv);
	DB_ILLEGAL_BEFORE_OPEN(dbp, "DB->put");

	/* Validate arguments. */
	if ((ret = __db_putchk(dbp, key, data,
	    flags, F_ISSET(dbp, DB_AM_DUP) || F_ISSET(key, DB_DBT_DUPOK))) != 0)
		return (ret);

	/* Create local transaction as necessary. */
	if (IS_AUTO_COMMIT(dbenv, txn, flags)) {
		if ((ret = __db_txn_auto(dbp, &txn)) != 0)
			return (ret);
		txn_local = 1;
		LF_CLR(DB_AUTO_COMMIT);
	}

	/* Check for consistent transaction usage. */
	if ((ret = __db_check_txn(dbp, txn, DB_LOCK_INVALIDID, 0)) != 0)
		goto err;

	if ((ret = dbp->cursor(dbp, txn, &dbc, DB_WRITELOCK)) != 0)
		goto err;

	DEBUG_LWRITE(dbc, txn, "db_put", key, data, flags);

	SET_RET_MEM(dbc, dbp);

	/*
	 * See the comment in __db_get().
	 *
	 * Note that the c_get in the DB_NOOVERWRITE case is safe to
	 * do with this flag set;  if it errors in any way other than
	 * DB_NOTFOUND, we're going to close the cursor without doing
	 * anything else, and if it returns DB_NOTFOUND then it's safe
	 * to do a c_put(DB_KEYLAST) even if an access method moved the
	 * cursor, since that's not position-dependent.
	 */
	F_SET(dbc, DBC_TRANSIENT);

	switch (flags) {
	case DB_APPEND:
		/*
		 * If there is an append callback, the value stored in
		 * data->data may be replaced and then freed.  To avoid
		 * passing a freed pointer back to the user, just operate
		 * on a copy of the data DBT.
		 */
		tdata = *data;

		/*
		 * Append isn't a normal put operation;  call the appropriate
		 * access method's append function.
		 */
		switch (dbp->type) {
		case DB_QUEUE:
			if ((ret = __qam_append(dbc, key, &tdata)) != 0)
				goto err;
			break;
		case DB_RECNO:
			if ((ret = __ram_append(dbc, key, &tdata)) != 0)
				goto err;
			break;
		default:
			/* The interface should prevent this. */
			DB_ASSERT(0);
			ret = __db_ferr(dbenv, "__db_put", flags);
			goto err;
		}

		/*
		 * Secondary indices:  since we've returned zero from
		 * an append function, we've just put a record, and done
		 * so outside __db_c_put.  We know we're not a secondary--
		 * the interface prevents puts on them--but we may be a
		 * primary.  If so, update our secondary indices
		 * appropriately.
		 */
		DB_ASSERT(!F_ISSET(dbp, DB_AM_SECONDARY));

		if (LIST_FIRST(&dbp->s_secondaries) != NULL)
			ret = __db_append_primary(dbc, key, &tdata);

		/*
		 * The append callback, if one exists, may have allocated
		 * a new tdata.data buffer.  If so, free it.
		 */
		FREE_IF_NEEDED(dbp, &tdata);

		/* No need for a cursor put;  we're done. */
		goto err;
	case DB_NOOVERWRITE:
		flags = 0;
		/*
		 * Set DB_DBT_USERMEM, this might be a threaded application and
		 * the flags checking will catch us.  We don't want the actual
		 * data, so request a partial of length 0.
		 */
		memset(&tdata, 0, sizeof(tdata));
		F_SET(&tdata, DB_DBT_USERMEM | DB_DBT_PARTIAL);

		/*
		 * If we're doing page-level locking, set the read-modify-write
		 * flag, we're going to overwrite immediately.
		 */
		if ((ret = dbc->c_get(dbc, key, &tdata,
		    DB_SET | (STD_LOCKING(dbc) ? DB_RMW : 0))) == 0)
			ret = DB_KEYEXIST;
		else if (ret == DB_NOTFOUND || ret == DB_KEYEMPTY)
			ret = 0;
		break;
	default:
		/* Fall through to normal cursor put. */
		break;
	}
	if (ret == 0)
		ret = dbc->c_put(dbc,
		    key, data, flags == 0 ? DB_KEYLAST : flags);

err:	/* Close the cursor. */
	if (dbc != NULL && (t_ret = __db_c_close(dbc)) != 0 && ret == 0)
		ret = t_ret;

	/* Commit for DB_AUTO_COMMIT. */
	if (txn_local) {
		if (ret == 0)
			ret = txn->commit(txn, 0);
		else
			if ((t_ret = txn->abort(txn)) != 0)
				ret = __db_panic(dbenv, t_ret);
	}

	return (ret);
}

/*
 * __db_delete --
 *	Delete the items referenced by a key.
 *
 * PUBLIC: int __db_delete __P((DB *, DB_TXN *, DBT *, u_int32_t));
 */
int
__db_delete(dbp, txn, key, flags)
	DB *dbp;
	DB_TXN *txn;
	DBT *key;
	u_int32_t flags;
{
	DBC *dbc;
	DBT data, lkey;
	DB_ENV *dbenv;
	u_int32_t f_init, f_next;
	int ret, t_ret, txn_local;

	dbc = NULL;
	dbenv = dbp->dbenv;
	txn_local = 0;

	PANIC_CHECK(dbenv);
	DB_ILLEGAL_BEFORE_OPEN(dbp, "DB->del");

	/* Check for invalid flags. */
	if ((ret = __db_delchk(dbp, key, flags)) != 0)
		return (ret);

	/* Create local transaction as necessary. */
	if (IS_AUTO_COMMIT(dbenv, txn, flags)) {
		if ((ret = __db_txn_auto(dbp, &txn)) != 0)
			return (ret);
		txn_local = 1;
		LF_CLR(DB_AUTO_COMMIT);
	}

	/* Check for consistent transaction usage. */
	if ((ret = __db_check_txn(dbp, txn, DB_LOCK_INVALIDID, 0)) != 0)
		goto err;

	/* Allocate a cursor. */
	if ((ret = dbp->cursor(dbp, txn, &dbc, DB_WRITELOCK)) != 0)
		goto err;

	DEBUG_LWRITE(dbc, txn, "db_delete", key, NULL, flags);

	/*
	 * Walk a cursor through the key/data pairs, deleting as we go.  Set
	 * the DB_DBT_USERMEM flag, as this might be a threaded application
	 * and the flags checking will catch us.  We don't actually want the
	 * keys or data, so request a partial of length 0.
	 */
	memset(&lkey, 0, sizeof(lkey));
	F_SET(&lkey, DB_DBT_USERMEM | DB_DBT_PARTIAL);
	memset(&data, 0, sizeof(data));
	F_SET(&data, DB_DBT_USERMEM | DB_DBT_PARTIAL);

	/*
	 * If locking (and we haven't already acquired CDB locks), set the
	 * read-modify-write flag.
	 */
	f_init = DB_SET;
	f_next = DB_NEXT_DUP;
	if (STD_LOCKING(dbc)) {
		f_init |= DB_RMW;
		f_next |= DB_RMW;
	}

	/* Walk through the set of key/data pairs, deleting as we go. */
	if ((ret = dbc->c_get(dbc, key, &data, f_init)) != 0)
		goto err;

	/*
	 * Hash permits an optimization in DB->del:  since on-page
	 * duplicates are stored in a single HKEYDATA structure, it's
	 * possible to delete an entire set of them at once, and as
	 * the HKEYDATA has to be rebuilt and re-put each time it
	 * changes, this is much faster than deleting the duplicates
	 * one by one.  Thus, if we're not pointing at an off-page
	 * duplicate set, and we're not using secondary indices (in
	 * which case we'd have to examine the items one by one anyway),
	 * let hash do this "quick delete".
	 *
	 * !!!
	 * Note that this is the only application-executed delete call in
	 * Berkeley DB that does not go through the __db_c_del function.
	 * If anything other than the delete itself (like a secondary index
	 * update) has to happen there in a particular situation, the
	 * conditions here should be modified not to call __ham_quick_delete.
	 * The ordinary AM-independent alternative will work just fine with
	 * a hash;  it'll just be slower.
	 */
	if (dbp->type == DB_HASH) {
		if (LIST_FIRST(&dbp->s_secondaries) == NULL &&
		    !F_ISSET(dbp, DB_AM_SECONDARY) &&
		    dbc->internal->opd == NULL) {
			ret = __ham_quick_delete(dbc);
			goto err;
		}
	}

	for (;;) {
		if ((ret = dbc->c_del(dbc, 0)) != 0)
			goto err;
		if ((ret = dbc->c_get(dbc, &lkey, &data, f_next)) != 0) {
			if (ret == DB_NOTFOUND) {
				ret = 0;
				break;
			}
			goto err;
		}
	}

err:	/* Discard the cursor. */
	if (dbc != NULL && (t_ret = dbc->c_close(dbc)) != 0 && ret == 0)
		ret = t_ret;

	/* Commit for DB_AUTO_COMMIT. */
	if (txn_local) {
		if (ret == 0)
			ret = txn->commit(txn, 0);
		else
			if ((t_ret = txn->abort(txn)) != 0)
				ret = __db_panic(dbenv, t_ret);
	}

	return (ret);
}

/*
 * __db_sync --
 *	Flush the database cache.
 *
 * PUBLIC: int __db_sync __P((DB *, u_int32_t));
 */
int
__db_sync(dbp, flags)
	DB *dbp;
	u_int32_t flags;
{
	int ret, t_ret;

	PANIC_CHECK(dbp->dbenv);
	DB_ILLEGAL_BEFORE_OPEN(dbp, "DB->sync");

	if ((ret = __db_syncchk(dbp, flags)) != 0)
		return (ret);

	/* Read-only trees never need to be sync'd. */
	if (F_ISSET(dbp, DB_AM_RDONLY))
		return (0);

	/* If it's a Recno tree, write the backing source text file. */
	if (dbp->type == DB_RECNO)
		ret = __ram_writeback(dbp);

	/* If the tree was never backed by a database file, we're done. */
	if (F_ISSET(dbp, DB_AM_INMEM))
		return (0);

	/* Flush any dirty pages from the cache to the backing file. */
	if ((t_ret = dbp->mpf->sync(dbp->mpf)) != 0 && ret == 0)
		ret = t_ret;
	return (ret);
}

/*
 * __db_associate --
 *	Associate another database as a secondary index to this one.
 *
 * PUBLIC: int __db_associate __P((DB *, DB_TXN *, DB *,
 * PUBLIC:     int (*)(DB *, const DBT *, const DBT *, DBT *), u_int32_t));
 */
int
__db_associate(dbp, txn, sdbp, callback, flags)
	DB *dbp, *sdbp;
	DB_TXN *txn;
	int (*callback) __P((DB *, const DBT *, const DBT *, DBT *));