db_am.c

File system using stacked.

	u_int32_t flags;
{
	DB_ENV *dbenv;
	DBC *pdbc, *sdbc;
	DBT skey, key, data;
	int build, ret, t_ret, txn_local;

	dbenv = dbp->dbenv;

	PANIC_CHECK(dbenv);

	txn_local = 0;
	pdbc = NULL;

	memset(&key, 0, sizeof(DBT));
	memset(&data, 0, sizeof(DBT));
	memset(&skey, 0, sizeof(DBT));

	if ((ret = __db_associatechk(dbp, sdbp, callback, flags)) != 0)
		return (ret);

	/*
	 * Create a local transaction as necessary, check for consistent
	 * transaction usage, and, if we have no transaction but do have
	 * locking on, acquire a locker id for the handle lock acquisition.
	 */
	if (IS_AUTO_COMMIT(dbenv, txn, flags)) {
		if ((ret = __db_txn_auto(dbp, &txn)) != 0)
			return (ret);
		txn_local = 1;
	} else if (txn != NULL && !TXN_ON(dbenv))
		return (__db_not_txn_env(dbenv));

	/*
	 * Check that if an open transaction is in progress, we're in it,
	 * for other common transaction errors, and for concurrent associates.
	 */
	if ((ret = __db_check_txn(dbp, txn, DB_LOCK_INVALIDID, 0)) != 0)
		return (ret);

	sdbp->s_callback = callback;
	sdbp->s_primary = dbp;

	sdbp->stored_get = sdbp->get;
	sdbp->get = __db_secondary_get;

	sdbp->stored_close = sdbp->close;
	sdbp->close = __db_secondary_close;

	/*
	 * Secondary cursors may have the primary's lock file ID, so we
	 * need to make sure that no older cursors are lying around
	 * when we make the transition.
	 */
	if (TAILQ_FIRST(&sdbp->active_queue) != NULL ||
	    TAILQ_FIRST(&sdbp->join_queue) != NULL) {
		__db_err(dbenv,
    "Databases may not become secondary indices while cursors are open");
		ret = EINVAL;
		goto err;
	}
	while ((sdbc = TAILQ_FIRST(&sdbp->free_queue)) != NULL)
		if ((ret = __db_c_destroy(sdbc)) != 0)
			goto err;

	F_SET(sdbp, DB_AM_SECONDARY);

	/*
	 * Check to see if the secondary is empty--and thus if we should
	 * build it--before we link it in and risk making it show up in
	 * other threads.
	 */
	build = 0;
	if (LF_ISSET(DB_CREATE)) {
		if ((ret = sdbp->cursor(sdbp, txn, &sdbc, 0)) != 0)
			goto err;

		memset(&key, 0, sizeof(DBT));
		memset(&data, 0, sizeof(DBT));

		/*
		 * We don't care about key or data;  we're just doing
		 * an existence check.
		 */
		F_SET(&key, DB_DBT_PARTIAL | DB_DBT_USERMEM);
		F_SET(&data, DB_DBT_PARTIAL | DB_DBT_USERMEM);
		if ((ret = sdbc->c_real_get(sdbc, &key, &data,
		    (STD_LOCKING(sdbc) ? DB_RMW : 0) |
		    DB_FIRST)) == DB_NOTFOUND) {
			build = 1;
			ret = 0;
		}

		/*
		 * Secondary cursors have special refcounting close
		 * methods.  Be careful.
		 */
		if ((t_ret = __db_c_close(sdbc)) != 0)
			ret = t_ret;
		if (ret != 0)
			goto err;
	}

	/*
	 * Add the secondary to the list on the primary.  Do it here
	 * so that we see any updates that occur while we're walking
	 * the primary.
	 */
	MUTEX_THREAD_LOCK(dbenv, dbp->mutexp);

	/* See __db_s_next for an explanation of secondary refcounting. */
	DB_ASSERT(sdbp->s_refcnt == 0);
	sdbp->s_refcnt = 1;
	LIST_INSERT_HEAD(&dbp->s_secondaries, sdbp, s_links);
	MUTEX_THREAD_UNLOCK(dbenv, dbp->mutexp);

	if (build) {
		/*
		 * We loop through the primary, putting each item we
		 * find into the new secondary.
		 *
		 * If we're using CDB, opening these two cursors puts us
		 * in a bit of a locking tangle:  CDB locks are done on the
		 * primary, so that we stay deadlock-free, but that means
		 * that updating the secondary while we have a read cursor
		 * open on the primary will self-block.  To get around this,
		 * we force the primary cursor to use the same locker ID
		 * as the secondary, so they won't conflict.  This should
		 * be harmless even if we're not using CDB.
		 */
		if ((ret = sdbp->cursor(sdbp, txn, &sdbc,
		    CDB_LOCKING(sdbp->dbenv) ? DB_WRITECURSOR : 0)) != 0)
			goto err;
		if ((ret = __db_icursor(dbp,
		    txn, dbp->type, PGNO_INVALID, 0, sdbc->locker, &pdbc)) != 0)
			goto err;

		/* Lock out other threads, now that we have a locker ID. */
		dbp->associate_lid = sdbc->locker;

		memset(&key, 0, sizeof(DBT));
		memset(&data, 0, sizeof(DBT));
		while ((ret = pdbc->c_get(pdbc, &key, &data, DB_NEXT)) == 0) {
			memset(&skey, 0, sizeof(DBT));
			if ((ret = callback(sdbp, &key, &data, &skey)) != 0) {
				if (ret == DB_DONOTINDEX)
					continue;
				else
					goto err;
			}
			if ((ret = sdbc->c_put(sdbc,
			    &skey, &key, DB_UPDATE_SECONDARY)) != 0) {
				FREE_IF_NEEDED(sdbp, &skey);
				goto err;
			}

			FREE_IF_NEEDED(sdbp, &skey);
		}
		if (ret == DB_NOTFOUND)
			ret = 0;

		if ((ret = sdbc->c_close(sdbc)) != 0)
			goto err;
	}

err:	if (pdbc != NULL && (t_ret = pdbc->c_close(pdbc)) != 0 && ret == 0)
		ret = t_ret;

	dbp->associate_lid = DB_LOCK_INVALIDID;

	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_pget --
 *	Return a primary key/data pair given a secondary key.
 *
 * PUBLIC: int __db_pget __P((DB *, DB_TXN *, DBT *, DBT *, DBT *, u_int32_t));
 */
int
__db_pget(dbp, txn, skey, pkey, data, flags)
	DB *dbp;
	DB_TXN *txn;
	DBT *skey, *pkey, *data;
	u_int32_t flags;
{
	DBC *dbc;
	int ret, t_ret;

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

	if ((ret = __db_pgetchk(dbp, skey, pkey, data, flags)) != 0)
		return (ret);

	if ((ret = dbp->cursor(dbp, txn, &dbc, 0)) != 0)
		return (ret);
	SET_RET_MEM(dbc, dbp);

	/*
	 * The underlying cursor pget will fill in a default DBT for null
	 * pkeys, and use the cursor's returned-key memory internally to
	 * store any intermediate primary keys.  However, we've just set
	 * the returned-key memory to the DB handle's key memory, which
	 * is unsafe to use if the DB handle is threaded.  If the pkey
	 * argument is NULL, use the DBC-owned returned-key memory
	 * instead;  it'll go away when we close the cursor before we
	 * return, but in this case that's just fine, as we're not
	 * returning the primary key.
	 */
	if (pkey == NULL)
		dbc->rkey = &dbc->my_rkey;

	DEBUG_LREAD(dbc, txn, "__db_pget", skey, NULL, flags);

	/*
	 * The cursor is just a perfectly ordinary secondary database
	 * cursor.  Call its c_pget() method to do the dirty work.
	 */
	if (flags == 0 || flags == DB_RMW)
		flags |= DB_SET;
	ret = dbc->c_pget(dbc, skey, pkey, data, flags);

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

/*
 * __db_secondary_get --
 *	This wrapper function for DB->pget() is the DB->get() function
 *	on a database which has been made into a secondary index.
 */
static int
__db_secondary_get(sdbp, txn, skey, data, flags)
	DB *sdbp;
	DB_TXN *txn;
	DBT *skey, *data;
	u_int32_t flags;
{

	DB_ASSERT(F_ISSET(sdbp, DB_AM_SECONDARY));
	return (sdbp->pget(sdbp, txn, skey, NULL, data, flags));
}

/*
 * __db_secondary_close --
 *	Wrapper function for DB->close() which we use on secondaries to
 *	manage refcounting and make sure we don't close them underneath
 *	a primary that is updating.
 */
static int
__db_secondary_close(sdbp, flags)
	DB *sdbp;
	u_int32_t flags;
{
	DB *primary;
	int doclose;

	doclose = 0;
	primary = sdbp->s_primary;

	MUTEX_THREAD_LOCK(primary->dbenv, primary->mutexp);
	/*
	 * Check the refcount--if it was at 1 when we were called, no
	 * thread is currently updating this secondary through the primary,
	 * so it's safe to close it for real.
	 *
	 * If it's not safe to do the close now, we do nothing;  the
	 * database will actually be closed when the refcount is decremented,
	 * which can happen in either __db_s_next or __db_s_done.
	 */
	DB_ASSERT(sdbp->s_refcnt != 0);
	if (--sdbp->s_refcnt == 0) {
		LIST_REMOVE(sdbp, s_links);
		/* We don't want to call close while the mutex is held. */
		doclose = 1;
	}
	MUTEX_THREAD_UNLOCK(primary->dbenv, primary->mutexp);

	/*
	 * sdbp->close is this function;  call the real one explicitly if
	 * need be.
	 */
	return (doclose ? __db_close(sdbp, flags) : 0);
}

/*
 * __db_append_primary --
 *	Perform the secondary index updates necessary to put(DB_APPEND)
 *	a record to a primary database.
 */
static int
__db_append_primary(dbc, key, data)
	DBC *dbc;
	DBT *key, *data;
{
	DB *dbp, *sdbp;
	DBC *sdbc, *pdbc;
	DBT oldpkey, pkey, pdata, skey;
	int cmp, ret, t_ret;

	dbp = dbc->dbp;
	sdbp = NULL;
	ret = 0;

	/*
	 * Worrying about partial appends seems a little like worrying
	 * about Linear A character encodings.  But we support those
	 * too if your application understands them.
	 */
	pdbc = NULL;
	if (F_ISSET(data, DB_DBT_PARTIAL) || F_ISSET(key, DB_DBT_PARTIAL)) {
		/*
		 * The dbc we were passed is all set to pass things
		 * back to the user;  we can't safely do a call on it.
		 * Dup the cursor, grab the real data item (we don't
		 * care what the key is--we've been passed it directly),
		 * and use that instead of the data DBT we were passed.
		 *
		 * Note that we can get away with this simple get because
		 * an appended item is by definition new, and the
		 * correctly-constructed full data item from this partial
		 * put is on the page waiting for us.
		 */
		if ((ret = __db_c_idup(dbc, &pdbc, DB_POSITIONI)) != 0)
			return (ret);
		memset(&pkey, 0, sizeof(DBT));
		memset(&pdata, 0, sizeof(DBT));

		if ((ret = pdbc->c_get(pdbc, &pkey, &pdata, DB_CURRENT)) != 0)
			goto err;

		key = &pkey;
		data = &pdata;
	}

	/*
	 * Loop through the secondary indices, putting a new item in
	 * each that points to the appended item.
	 *
	 * This is much like the loop in "step 3" in __db_c_put, so
	 * I'm not commenting heavily here;  it was unclean to excerpt
	 * just that section into a common function, but the basic
	 * overview is the same here.
	 */
	for (sdbp = __db_s_first(dbp);
	    sdbp != NULL && ret == 0; ret = __db_s_next(&sdbp)) {
		memset(&skey, 0, sizeof(DBT));
		if ((ret = sdbp->s_callback(sdbp, key, data, &skey)) != 0) {
			if (ret == DB_DONOTINDEX)
				continue;
			else
				goto err;
		}

		if ((ret = __db_icursor(sdbp, dbc->txn, sdbp->type,
		    PGNO_INVALID, 0, dbc->locker, &sdbc)) != 0) {
			FREE_IF_NEEDED(sdbp, &skey);
			goto err;
		}
		if (CDB_LOCKING(sdbp->dbenv)) {
			DB_ASSERT(sdbc->mylock.off == LOCK_INVALID);
			F_SET(sdbc, DBC_WRITER);
		}

		/*
		 * Since we know we have a new primary key, it can't be a
		 * duplicate duplicate in the secondary.  It can be a
		 * duplicate in a secondary that doesn't support duplicates,
		 * however, so we need to be careful to avoid an overwrite
		 * (which would corrupt our index).
		 */
		if (!F_ISSET(sdbp, DB_AM_DUP)) {
			memset(&oldpkey, 0, sizeof(DBT));
			F_SET(&oldpkey, DB_DBT_MALLOC);
			ret = sdbc->c_real_get(sdbc, &skey, &oldpkey,
			    DB_SET | (STD_LOCKING(dbc) ? DB_RMW : 0));
			if (ret == 0) {
				cmp = __bam_defcmp(sdbp, &oldpkey, key);
				/*
				 * XXX
				 * This needs to use the right free function
				 * as soon as this is possible.
				 */
				__os_ufree(sdbp->dbenv,
				    oldpkey.data);
				if (cmp != 0) {
					__db_err(sdbp->dbenv, "%s%s",
			    "Append results in a non-unique secondary key in",
			    " an index not configured to support duplicates");
					ret = EINVAL;
					goto err1;
				}
			} else if (ret != DB_NOTFOUND && ret != DB_KEYEMPTY)
				goto err1;
		}

		ret = sdbc->c_put(sdbc, &skey, key, DB_UPDATE_SECONDARY);

err1:		FREE_IF_NEEDED(sdbp, &skey);

		if ((t_ret = sdbc->c_close(sdbc)) != 0 && ret == 0)
			ret = t_ret;

		if (ret != 0)
			goto err;
	}

err:	if (pdbc != NULL && (t_ret = pdbc->c_close(pdbc)) != 0 && ret == 0)
		ret = t_ret;
	if (sdbp != NULL && (t_ret = __db_s_done(sdbp)) != 0 && ret == 0)
		ret = t_ret;
	return (ret);
}