dlmdomain.c

linux 内核源代码

	int node, clear_node, status;

	/* At this point we've migrated away all our locks and won't
	 * accept mastership of new ones. The dlm is responsible for
	 * almost nothing now. We make sure not to confuse any joining
	 * nodes and then commence shutdown procedure. */

	spin_lock(&dlm->spinlock);
	/* Clear ourselves from the domain map */
	clear_bit(dlm->node_num, dlm->domain_map);
	while ((node = find_next_bit(dlm->domain_map, O2NM_MAX_NODES,
				     0)) < O2NM_MAX_NODES) {
		/* Drop the dlm spinlock. This is safe wrt the domain_map.
		 * -nodes cannot be added now as the
		 *   query_join_handlers knows to respond with OK_NO_MAP
		 * -we catch the right network errors if a node is
		 *   removed from the map while we're sending him the
		 *   exit message. */
		spin_unlock(&dlm->spinlock);

		clear_node = 1;

		status = dlm_send_one_domain_exit(dlm, node);
		if (status < 0 &&
		    status != -ENOPROTOOPT &&
		    status != -ENOTCONN) {
			mlog(ML_NOTICE, "Error %d sending domain exit message "
			     "to node %d\n", status, node);

			/* Not sure what to do here but lets sleep for
			 * a bit in case this was a transient
			 * error... */
			msleep(DLM_DOMAIN_BACKOFF_MS);
			clear_node = 0;
		}

		spin_lock(&dlm->spinlock);
		/* If we're not clearing the node bit then we intend
		 * to loop back around to try again. */
		if (clear_node)
			clear_bit(node, dlm->domain_map);
	}
	spin_unlock(&dlm->spinlock);
}

int dlm_joined(struct dlm_ctxt *dlm)
{
	int ret = 0;

	spin_lock(&dlm_domain_lock);

	if (dlm->dlm_state == DLM_CTXT_JOINED)
		ret = 1;

	spin_unlock(&dlm_domain_lock);

	return ret;
}

int dlm_shutting_down(struct dlm_ctxt *dlm)
{
	int ret = 0;

	spin_lock(&dlm_domain_lock);

	if (dlm->dlm_state == DLM_CTXT_IN_SHUTDOWN)
		ret = 1;

	spin_unlock(&dlm_domain_lock);

	return ret;
}

void dlm_unregister_domain(struct dlm_ctxt *dlm)
{
	int leave = 0;

	spin_lock(&dlm_domain_lock);
	BUG_ON(dlm->dlm_state != DLM_CTXT_JOINED);
	BUG_ON(!dlm->num_joins);

	dlm->num_joins--;
	if (!dlm->num_joins) {
		/* We mark it "in shutdown" now so new register
		 * requests wait until we've completely left the
		 * domain. Don't use DLM_CTXT_LEAVING yet as we still
		 * want new domain joins to communicate with us at
		 * least until we've completed migration of our
		 * resources. */
		dlm->dlm_state = DLM_CTXT_IN_SHUTDOWN;
		leave = 1;
	}
	spin_unlock(&dlm_domain_lock);

	if (leave) {
		mlog(0, "shutting down domain %s\n", dlm->name);

		/* We changed dlm state, notify the thread */
		dlm_kick_thread(dlm, NULL);

		while (dlm_migrate_all_locks(dlm)) {
			/* Give dlm_thread time to purge the lockres' */
			msleep(500);
			mlog(0, "%s: more migration to do\n", dlm->name);
		}
		dlm_mark_domain_leaving(dlm);
		dlm_leave_domain(dlm);
		dlm_complete_dlm_shutdown(dlm);
	}
	dlm_put(dlm);
}
EXPORT_SYMBOL_GPL(dlm_unregister_domain);

static int dlm_query_join_handler(struct o2net_msg *msg, u32 len, void *data,
				  void **ret_data)
{
	struct dlm_query_join_request *query;
	enum dlm_query_join_response response;
	struct dlm_ctxt *dlm = NULL;
	u8 nodenum;

	query = (struct dlm_query_join_request *) msg->buf;

	mlog(0, "node %u wants to join domain %s\n", query->node_idx,
		  query->domain);

	/*
	 * If heartbeat doesn't consider the node live, tell it
	 * to back off and try again.  This gives heartbeat a chance
	 * to catch up.
	 */
	if (!o2hb_check_node_heartbeating(query->node_idx)) {
		mlog(0, "node %u is not in our live map yet\n",
		     query->node_idx);

		response = JOIN_DISALLOW;
		goto respond;
	}

	response = JOIN_OK_NO_MAP;

	spin_lock(&dlm_domain_lock);
	dlm = __dlm_lookup_domain_full(query->domain, query->name_len);
	if (!dlm)
		goto unlock_respond;

	/*
	 * There is a small window where the joining node may not see the
	 * node(s) that just left but still part of the cluster. DISALLOW
	 * join request if joining node has different node map.
	 */
	nodenum=0;
	while (nodenum < O2NM_MAX_NODES) {
		if (test_bit(nodenum, dlm->domain_map)) {
			if (!byte_test_bit(nodenum, query->node_map)) {
				mlog(0, "disallow join as node %u does not "
				     "have node %u in its nodemap\n",
				     query->node_idx, nodenum);
				response = JOIN_DISALLOW;
				goto unlock_respond;
			}
		}
		nodenum++;
	}

	/* Once the dlm ctxt is marked as leaving then we don't want
	 * to be put in someone's domain map. 
	 * Also, explicitly disallow joining at certain troublesome
	 * times (ie. during recovery). */
	if (dlm && dlm->dlm_state != DLM_CTXT_LEAVING) {
		int bit = query->node_idx;
		spin_lock(&dlm->spinlock);

		if (dlm->dlm_state == DLM_CTXT_NEW &&
		    dlm->joining_node == DLM_LOCK_RES_OWNER_UNKNOWN) {
			/*If this is a brand new context and we
			 * haven't started our join process yet, then
			 * the other node won the race. */
			response = JOIN_OK_NO_MAP;
		} else if (dlm->joining_node != DLM_LOCK_RES_OWNER_UNKNOWN) {
			/* Disallow parallel joins. */
			response = JOIN_DISALLOW;
		} else if (dlm->reco.state & DLM_RECO_STATE_ACTIVE) {
			mlog(0, "node %u trying to join, but recovery "
			     "is ongoing.\n", bit);
			response = JOIN_DISALLOW;
		} else if (test_bit(bit, dlm->recovery_map)) {
			mlog(0, "node %u trying to join, but it "
			     "still needs recovery.\n", bit);
			response = JOIN_DISALLOW;
		} else if (test_bit(bit, dlm->domain_map)) {
			mlog(0, "node %u trying to join, but it "
			     "is still in the domain! needs recovery?\n",
			     bit);
			response = JOIN_DISALLOW;
		} else {
			/* Alright we're fully a part of this domain
			 * so we keep some state as to who's joining
			 * and indicate to him that needs to be fixed
			 * up. */
			response = JOIN_OK;
			__dlm_set_joining_node(dlm, query->node_idx);
		}

		spin_unlock(&dlm->spinlock);
	}
unlock_respond:
	spin_unlock(&dlm_domain_lock);

respond:
	mlog(0, "We respond with %u\n", response);

	return response;
}

static int dlm_assert_joined_handler(struct o2net_msg *msg, u32 len, void *data,
				     void **ret_data)
{
	struct dlm_assert_joined *assert;
	struct dlm_ctxt *dlm = NULL;

	assert = (struct dlm_assert_joined *) msg->buf;

	mlog(0, "node %u asserts join on domain %s\n", assert->node_idx,
		  assert->domain);

	spin_lock(&dlm_domain_lock);
	dlm = __dlm_lookup_domain_full(assert->domain, assert->name_len);
	/* XXX should we consider no dlm ctxt an error? */
	if (dlm) {
		spin_lock(&dlm->spinlock);

		/* Alright, this node has officially joined our
		 * domain. Set him in the map and clean up our
		 * leftover join state. */
		BUG_ON(dlm->joining_node != assert->node_idx);
		set_bit(assert->node_idx, dlm->domain_map);
		__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);

		printk(KERN_INFO "ocfs2_dlm: Node %u joins domain %s\n",
		       assert->node_idx, dlm->name);
		__dlm_print_nodes(dlm);

		/* notify anything attached to the heartbeat events */
		dlm_hb_event_notify_attached(dlm, assert->node_idx, 1);

		spin_unlock(&dlm->spinlock);
	}
	spin_unlock(&dlm_domain_lock);

	return 0;
}

static int dlm_cancel_join_handler(struct o2net_msg *msg, u32 len, void *data,
				   void **ret_data)
{
	struct dlm_cancel_join *cancel;
	struct dlm_ctxt *dlm = NULL;

	cancel = (struct dlm_cancel_join *) msg->buf;

	mlog(0, "node %u cancels join on domain %s\n", cancel->node_idx,
		  cancel->domain);

	spin_lock(&dlm_domain_lock);
	dlm = __dlm_lookup_domain_full(cancel->domain, cancel->name_len);

	if (dlm) {
		spin_lock(&dlm->spinlock);

		/* Yikes, this guy wants to cancel his join. No
		 * problem, we simply cleanup our join state. */
		BUG_ON(dlm->joining_node != cancel->node_idx);
		__dlm_set_joining_node(dlm, DLM_LOCK_RES_OWNER_UNKNOWN);

		spin_unlock(&dlm->spinlock);
	}
	spin_unlock(&dlm_domain_lock);

	return 0;
}

static int dlm_send_one_join_cancel(struct dlm_ctxt *dlm,
				    unsigned int node)
{
	int status;
	struct dlm_cancel_join cancel_msg;

	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, 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 *response)
{
	int status, retval;
	struct dlm_query_join_request join_msg;

	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);

	/* 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, &retval);
	if (status < 0 && status != -ENOPROTOOPT) {
		mlog_errno(status);
		goto bail;
	}

	/* -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 (retval == JOIN_DISALLOW ||
		   retval == JOIN_OK ||
		   retval == JOIN_OK_NO_MAP) {
		*response = retval;
	} else {
		status = -EINVAL;
		mlog(ML_ERROR, "invalid response %d from node %u\n", retval,
		     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 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 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? */