dlmdomain.c

ocfs1.4.1 oracle分布式文件系统

	memset(&cancel_msg, 0, sizeof(cancel_msg));
	cancel_msg.node_idx = dlm->node_num;
	cancel_msg.name_len = strlen(dlm->name);
	memcpy(cancel_msg.domain, dlm->name, cancel_msg.name_len);

	status = o2net_send_message(DLM_CANCEL_JOIN_MSG, DLM_MOD_KEY,
				    &cancel_msg, sizeof(cancel_msg), node,
				    NULL);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

bail:
	return status;
}

/* map_size should be in bytes. */
static int dlm_send_join_cancels(struct dlm_ctxt *dlm,
				 unsigned long *node_map,
				 unsigned int map_size)
{
	int status, tmpstat;
	unsigned int node;

	if (map_size != (BITS_TO_LONGS(O2NM_MAX_NODES) *
			 sizeof(unsigned long))) {
		mlog(ML_ERROR,
		     "map_size %u != BITS_TO_LONGS(O2NM_MAX_NODES) %u\n",
		     map_size, (unsigned)BITS_TO_LONGS(O2NM_MAX_NODES));
		return -EINVAL;
	}

	status = 0;
	node = -1;
	while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
				     node + 1)) < O2NM_MAX_NODES) {
		if (node == dlm->node_num)
			continue;

		tmpstat = dlm_send_one_join_cancel(dlm, node);
		if (tmpstat) {
			mlog(ML_ERROR, "Error return %d cancelling join on "
			     "node %d\n", tmpstat, node);
			if (!status)
				status = tmpstat;
		}
	}

	if (status)
		mlog_errno(status);
	return status;
}

static int dlm_request_join(struct dlm_ctxt *dlm,
			    int node,
			    enum dlm_query_join_response_code *response)
{
	int status;
	struct dlm_query_join_request join_msg;
	struct dlm_query_join_packet packet;
	u32 join_resp;

	mlog(0, "querying node %d\n", node);

	memset(&join_msg, 0, sizeof(join_msg));
	join_msg.node_idx = dlm->node_num;
	join_msg.name_len = strlen(dlm->name);
	memcpy(join_msg.domain, dlm->name, join_msg.name_len);
	join_msg.dlm_proto = dlm->dlm_locking_proto;
	join_msg.fs_proto = dlm->fs_locking_proto;

	/* copy live node map to join message */
	byte_copymap(join_msg.node_map, dlm->live_nodes_map, O2NM_MAX_NODES);

	status = o2net_send_message(DLM_QUERY_JOIN_MSG, DLM_MOD_KEY, &join_msg,
				    sizeof(join_msg), node,
				    &join_resp);
	if (status < 0 && status != -ENOPROTOOPT) {
		mlog_errno(status);
		goto bail;
	}
	dlm_query_join_wire_to_packet(join_resp, &packet);

	/* -ENOPROTOOPT from the net code means the other side isn't
	    listening for our message type -- that's fine, it means
	    his dlm isn't up, so we can consider him a 'yes' but not
	    joined into the domain.  */
	if (status == -ENOPROTOOPT) {
		status = 0;
		*response = JOIN_OK_NO_MAP;
	} else if (packet.code == JOIN_DISALLOW ||
		   packet.code == JOIN_OK_NO_MAP) {
		*response = packet.code;
	} else if (packet.code == JOIN_PROTOCOL_MISMATCH) {
		mlog(ML_NOTICE,
		     "This node requested DLM locking protocol %u.%u and "
		     "filesystem locking protocol %u.%u.  At least one of "
		     "the protocol versions on node %d is not compatible, "
		     "disconnecting\n",
		     dlm->dlm_locking_proto.pv_major,
		     dlm->dlm_locking_proto.pv_minor,
		     dlm->fs_locking_proto.pv_major,
		     dlm->fs_locking_proto.pv_minor,
		     node);
		status = -EPROTO;
		*response = packet.code;
	} else if (packet.code == JOIN_OK) {
		*response = packet.code;
		/* Use the same locking protocol as the remote node */
		dlm->dlm_locking_proto.pv_minor = packet.dlm_minor;
		dlm->fs_locking_proto.pv_minor = packet.fs_minor;
		mlog(0,
		     "Node %d responds JOIN_OK with DLM locking protocol "
		     "%u.%u and fs locking protocol %u.%u\n",
		     node,
		     dlm->dlm_locking_proto.pv_major,
		     dlm->dlm_locking_proto.pv_minor,
		     dlm->fs_locking_proto.pv_major,
		     dlm->fs_locking_proto.pv_minor);
	} else {
		status = -EINVAL;
		mlog(ML_ERROR, "invalid response %d from node %u\n",
		     packet.code, node);
	}

	mlog(0, "status %d, node %d response is %d\n", status, node,
	     *response);

bail:
	return status;
}

static int dlm_send_one_join_assert(struct dlm_ctxt *dlm,
				    unsigned int node)
{
	int status;
	struct dlm_assert_joined assert_msg;

	mlog(0, "Sending join assert to node %u\n", node);

	memset(&assert_msg, 0, sizeof(assert_msg));
	assert_msg.node_idx = dlm->node_num;
	assert_msg.name_len = strlen(dlm->name);
	memcpy(assert_msg.domain, dlm->name, assert_msg.name_len);

	status = o2net_send_message(DLM_ASSERT_JOINED_MSG, DLM_MOD_KEY,
				    &assert_msg, sizeof(assert_msg), node,
				    NULL);
	if (status < 0)
		mlog_errno(status);

	return status;
}

static void dlm_send_join_asserts(struct dlm_ctxt *dlm,
				  unsigned long *node_map)
{
	int status, node, live;

	status = 0;
	node = -1;
	while ((node = find_next_bit(node_map, O2NM_MAX_NODES,
				     node + 1)) < O2NM_MAX_NODES) {
		if (node == dlm->node_num)
			continue;

		do {
			/* It is very important that this message be
			 * received so we spin until either the node
			 * has died or it gets the message. */
			status = dlm_send_one_join_assert(dlm, node);

			spin_lock(&dlm->spinlock);
			live = test_bit(node, dlm->live_nodes_map);
			spin_unlock(&dlm->spinlock);

			if (status) {
				mlog(ML_ERROR, "Error return %d asserting "
				     "join on node %d\n", status, node);

				/* give us some time between errors... */
				if (live)
					msleep(DLM_DOMAIN_BACKOFF_MS);
			}
		} while (status && live);
	}
}

struct domain_join_ctxt {
	unsigned long live_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
	unsigned long yes_resp_map[BITS_TO_LONGS(O2NM_MAX_NODES)];
};

static int dlm_should_restart_join(struct dlm_ctxt *dlm,
				   struct domain_join_ctxt *ctxt,
				   enum dlm_query_join_response_code response)
{
	int ret;

	if (response == JOIN_DISALLOW) {
		mlog(0, "Latest response of disallow -- should restart\n");
		return 1;
	}

	spin_lock(&dlm->spinlock);
	/* For now, we restart the process if the node maps have
	 * changed at all */
	ret = memcmp(ctxt->live_map, dlm->live_nodes_map,
		     sizeof(dlm->live_nodes_map));
	spin_unlock(&dlm->spinlock);

	if (ret)
		mlog(0, "Node maps changed -- should restart\n");

	return ret;
}

static int dlm_try_to_join_domain(struct dlm_ctxt *dlm)
{
	int status = 0, tmpstat, node;
	struct domain_join_ctxt *ctxt;
	enum dlm_query_join_response_code response = JOIN_DISALLOW;

	mlog_entry("%p", dlm);

	ctxt = kzalloc(sizeof(*ctxt), GFP_KERNEL);
	if (!ctxt) {
		status = -ENOMEM;
		mlog_errno(status);
		goto bail;
	}

	/* group sem locking should work for us here -- we're already
	 * registered for heartbeat events so filling this should be
	 * atomic wrt getting those handlers called. */
	o2hb_fill_node_map(dlm->live_nodes_map, sizeof(dlm->live_nodes_map));

	spin_lock(&dlm->spinlock);
	memcpy(ctxt->live_map, dlm->live_nodes_map, sizeof(ctxt->live_map));

	__dlm_set_joining_node(dlm, dlm->node_num);

	spin_unlock(&dlm->spinlock);

	node = -1;
	while ((node = find_next_bit(ctxt->live_map, O2NM_MAX_NODES,
				     node + 1)) < O2NM_MAX_NODES) {
		if (node == dlm->node_num)
			continue;

		status = dlm_request_join(dlm, node, &response);
		if (status < 0) {
			mlog_errno(status);
			goto bail;
		}

		/* Ok, either we got a response or the node doesn't have a
		 * dlm up. */
		if (response == JOIN_OK)
			set_bit(node, ctxt->yes_resp_map);

		if (dlm_should_restart_join(dlm, ctxt, response)) {
			status = -EAGAIN;
			goto bail;
		}
	}

	mlog(0, "Yay, done querying nodes!\n");

	/* Yay, everyone agree's we can join the domain. My domain is
	 * comprised of all nodes who were put in the
	 * yes_resp_map. Copy that into our domain map and send a join
	 * assert message to clean up everyone elses state. */
	spin_lock(&dlm->spinlock);
	memcpy(dlm->domain_map, ctxt->yes_resp_map,
	       sizeof(ctxt->yes_resp_map));
	set_bit(dlm->node_num, dlm->domain_map);
	spin_unlock(&dlm->spinlock);

	dlm_send_join_asserts(dlm, ctxt->yes_resp_map);

	/* Joined state *must* be set before the joining node
	 * information, otherwise the query_join handler may read no
	 * current joiner but a state of NEW and tell joining nodes
	 * we're not in the domain. */
	spin_lock(&dlm_domain_lock);
	dlm->dlm_state = DLM_CTXT_JOINED;
	dlm->num_joins++;
	spin_unlock(&dlm_domain_lock);

bail:
	spin_lock(&dlm->spinlock);
	__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);
	if (!status)
		__dlm_print_nodes(dlm);
	spin_unlock(&dlm->spinlock);

	if (ctxt) {
		/* Do we need to send a cancel message to any nodes? */
		if (status < 0) {
			tmpstat = dlm_send_join_cancels(dlm,
							ctxt->yes_resp_map,
							sizeof(ctxt->yes_resp_map));
			if (tmpstat < 0)
				mlog_errno(tmpstat);
		}
		kfree(ctxt);
	}

	mlog(0, "returning %d\n", status);
	return status;
}

static void dlm_unregister_domain_handlers(struct dlm_ctxt *dlm)
{
	o2hb_unregister_callback(NULL, &dlm->dlm_hb_up);
	o2hb_unregister_callback(NULL, &dlm->dlm_hb_down);
	o2net_unregister_handler_list(&dlm->dlm_domain_handlers);
}

static int dlm_register_domain_handlers(struct dlm_ctxt *dlm)
{
	int status;

	mlog(0, "registering handlers.\n");

	o2hb_setup_callback(&dlm->dlm_hb_down, O2HB_NODE_DOWN_CB,
			    dlm_hb_node_down_cb, dlm, DLM_HB_NODE_DOWN_PRI);
	status = o2hb_register_callback(NULL, &dlm->dlm_hb_down);
	if (status)
		goto bail;

	o2hb_setup_callback(&dlm->dlm_hb_up, O2HB_NODE_UP_CB,
			    dlm_hb_node_up_cb, dlm, DLM_HB_NODE_UP_PRI);
	status = o2hb_register_callback(NULL, &dlm->dlm_hb_up);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_MASTER_REQUEST_MSG, dlm->key,
					sizeof(struct dlm_master_request),
					dlm_master_request_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_ASSERT_MASTER_MSG, dlm->key,
					sizeof(struct dlm_assert_master),
					dlm_assert_master_handler,
					dlm, dlm_assert_master_post_handler,
					&dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_CREATE_LOCK_MSG, dlm->key,
					sizeof(struct dlm_create_lock),
					dlm_create_lock_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_CONVERT_LOCK_MSG, dlm->key,
					DLM_CONVERT_LOCK_MAX_LEN,
					dlm_convert_lock_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_UNLOCK_LOCK_MSG, dlm->key,
					DLM_UNLOCK_LOCK_MAX_LEN,
					dlm_unlock_lock_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_PROXY_AST_MSG, dlm->key,
					DLM_PROXY_AST_MAX_LEN,
					dlm_proxy_ast_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_EXIT_DOMAIN_MSG, dlm->key,
					sizeof(struct dlm_exit_domain),
					dlm_exit_domain_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_DEREF_LOCKRES_MSG, dlm->key,
					sizeof(struct dlm_deref_lockres),
					dlm_deref_lockres_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_MIGRATE_REQUEST_MSG, dlm->key,
					sizeof(struct dlm_migrate_request),
					dlm_migrate_request_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_MIG_LOCKRES_MSG, dlm->key,
					DLM_MIG_LOCKRES_MAX_LEN,
					dlm_mig_lockres_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_MASTER_REQUERY_MSG, dlm->key,
					sizeof(struct dlm_master_requery),
					dlm_master_requery_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_LOCK_REQUEST_MSG, dlm->key,
					sizeof(struct dlm_lock_request),
					dlm_request_all_locks_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_RECO_DATA_DONE_MSG, dlm->key,
					sizeof(struct dlm_reco_data_done),
					dlm_reco_data_done_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_BEGIN_RECO_MSG, dlm->key,
					sizeof(struct dlm_begin_reco),
					dlm_begin_reco_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

	status = o2net_register_handler(DLM_FINALIZE_RECO_MSG, dlm->key,
					sizeof(struct dlm_finalize_reco),
					dlm_finalize_reco_handler,
					dlm, NULL, &dlm->dlm_domain_handlers);
	if (status)
		goto bail;

bail:
	if (status)
		dlm_unregister_domain_handlers(dlm);

	return status;
}

static int dlm_join_domain(struct dlm_ctxt *dlm)
{
	int status;
	unsigned int backoff;
	unsigned int total_backoff = 0;

	BUG_ON(!dlm);

	mlog(0, "Join domain %s\n", dlm->name);

	status = dlm_register_domain_handlers(dlm);
	if (status) {
		mlog_errno(status);
		goto bail;
	}

	status = dlm_debug_init(dlm);
	if (status < 0) {
		mlog_errno(status);
		goto bail;
	}

	status = dlm_launch_thread(dlm);