ZeroTierOne/controller/src/controller.rs

// (c) 2020-2022 ZeroTier, Inc. -- currently proprietary pending actual release and licensing. See LICENSE.md.

use std::collections::HashMap;
use std::error::Error;
use std::sync::{Arc, Mutex, RwLock, Weak};

use tokio::time::{Duration, Instant};

use zerotier_network_hypervisor::protocol;
use zerotier_network_hypervisor::protocol::{PacketBuffer, DEFAULT_MULTICAST_LIMIT, ZEROTIER_VIRTUAL_NETWORK_DEFAULT_MTU};
use zerotier_network_hypervisor::vl1::*;
use zerotier_network_hypervisor::vl2;
use zerotier_network_hypervisor::vl2::multicastauthority::MulticastAuthority;
use zerotier_network_hypervisor::vl2::networkconfig::*;
use zerotier_network_hypervisor::vl2::{NetworkId, Revocation};
use zerotier_utils::blob::Blob;
use zerotier_utils::buffer::OutOfBoundsError;
use zerotier_utils::error::InvalidParameterError;
use zerotier_utils::reaper::Reaper;
use zerotier_utils::tokio;
use zerotier_utils::{ms_monotonic, ms_since_epoch};
use zerotier_vl1_service::VL1Service;

use crate::database::*;
use crate::model::{AuthenticationResult, Member, RequestLogItem, CREDENTIAL_WINDOW_SIZE_DEFAULT};

// A netconf per-query task timeout, just a sanity limit.
const REQUEST_TIMEOUT: Duration = Duration::from_secs(10);

/// ZeroTier VL2 network controller packet handler, answers VL2 netconf queries.
pub struct Controller {
    self_ref: Weak<Self>,
    service: RwLock<Weak<VL1Service<dyn Database, Self, Self>>>,
    reaper: Reaper,
    runtime: tokio::runtime::Handle,
    database: Arc<dyn Database>,
    local_identity: Verified<Identity>,

    /// Handler for MULTICAST_LIKE and MULTICAST_GATHER messages.
    multicast_authority: MulticastAuthority,

    /// Async tasks that should be killed when the controller is dropped.
    daemons: Mutex<Vec<tokio::task::JoinHandle<()>>>, // drop() aborts these

    /// Recently authorized network members and when that authorization expires (in monotonic ticks).
    /// Note that this is not and should not be used for real authentication, just for locking up multicast info.
    recently_authorized: RwLock<HashMap<[u8; Identity::FINGERPRINT_SIZE], HashMap<NetworkId, i64>>>,
}

impl Controller {
    /// Start an inner protocol handler answer ZeroTier VL2 network controller queries.
    ///
    /// The start() method must be called once the service this will run within is also created.
    pub async fn new(database: Arc<dyn Database>, runtime: tokio::runtime::Handle) -> Result<Arc<Self>, Box<dyn Error>> {
        if let Some(local_identity) = database.load_node_identity() {
            assert!(local_identity.secret.is_some());
            Ok(Arc::new_cyclic(|r| Self {
                self_ref: r.clone(),
                service: RwLock::new(Weak::default()),
                reaper: Reaper::new(&runtime),
                runtime,
                database: database.clone(),
                local_identity: local_identity,
                multicast_authority: MulticastAuthority::new(),
                daemons: Mutex::new(Vec::with_capacity(2)),
                recently_authorized: RwLock::new(HashMap::new()),
            }))
        } else {
            Err(Box::new(InvalidParameterError(
                "local controller's identity not readable by database",
            )))
        }
    }

    /// Set the service and HostSystem implementation for this controller and start daemons.
    ///
    /// This must be called once the service that uses this handler is up or the controller
    /// won't actually do anything. The controller holds a weak reference to VL1Service so
    /// be sure it's not dropped.
    pub async fn start(&self, service: &Arc<VL1Service<dyn Database, Self, Self>>) {
        *self.service.write().unwrap() = Arc::downgrade(service);

        // Create database change listener.
        if let Some(cw) = self.database.changes().await.map(|mut ch| {
            let self2 = self.self_ref.clone();
            self.runtime.spawn(async move {
                loop {
                    if let Ok(change) = ch.recv().await {
                        if let Some(self2) = self2.upgrade() {
                            self2.reaper.add(
                                self2.runtime.spawn(self2.clone().handle_change_notification(change)),
                                Instant::now().checked_add(REQUEST_TIMEOUT).unwrap(),
                            );
                        } else {
                            break;
                        }
                    }
                }
            })
        }) {
            self.daemons.lock().unwrap().push(cw);
        }

        // Create background task to expire multicast subscriptions and recent authorizations.
        let self2 = self.self_ref.clone();
        self.daemons.lock().unwrap().push(self.runtime.spawn(async move {
            let sleep_duration = Duration::from_millis((protocol::VL2_DEFAULT_MULTICAST_LIKE_EXPIRE / 2).min(2500) as u64);
            loop {
                tokio::time::sleep(sleep_duration).await;

                if let Some(self2) = self2.upgrade() {
                    let time_ticks = ms_monotonic();
                    self2.multicast_authority.clean(time_ticks);
                    self2.recently_authorized.write().unwrap().retain(|_, by_network| {
                        by_network.retain(|_, timeout| *timeout > time_ticks);
                        !by_network.is_empty()
                    });
                } else {
                    break;
                }
            }
        }));
    }

    /// Launched as a task when the DB informs us of a change.
    async fn handle_change_notification(self: Arc<Self>, change: Change) {
        match change {
            Change::NetworkCreated(_) => {}
            Change::NetworkChanged(_, _) => {}
            Change::NetworkDeleted(_, _) => {} // TODO: somehow poison whole network
            Change::MemberCreated(_) => {}
            Change::MemberChanged(old_member, new_member) => {
                if !new_member.authorized() && old_member.authorized() {
                    self.deauthorize_member(&new_member).await;
                }
            }
            Change::MemberDeleted(member) => self.deauthorize_member(&member).await,
        }
    }

    /// Compose and send network configuration packet (either V1 or V2)
    fn send_network_config(
        &self,
        peer: &Peer,
        config: &NetworkConfig,
        in_re_message_id: Option<u64>, // None for unsolicited push
    ) {
        if let Some(host_system) = self.service.read().unwrap().upgrade() {
            peer.send(
                host_system.as_ref(),
                host_system.node(),
                None,
                ms_monotonic(),
                |packet| -> Result<(), OutOfBoundsError> {
                    let payload_start = packet.len();

                    if let Some(in_re_message_id) = in_re_message_id {
                        let ok_header = packet.append_struct_get_mut::<protocol::OkHeader>()?;
                        ok_header.verb = protocol::message_type::VL1_OK;
                        ok_header.in_re_verb = protocol::message_type::VL2_NETWORK_CONFIG_REQUEST;
                        ok_header.in_re_message_id = in_re_message_id.to_be_bytes();
                    } else {
                        packet.append_u8(protocol::message_type::VL2_NETWORK_CONFIG)?;
                    }

                    if peer.is_v2() {
                        todo!()
                    } else {
                        let config_data = if let Some(config_dict) = config.v1_proto_to_dictionary(&self.local_identity) {
                            config_dict.to_bytes()
                        } else {
                            eprintln!("WARNING: unexpected error serializing network config into V1 format dictionary");
                            return Err(OutOfBoundsError); // abort
                        };
                        if config_data.len() > (u16::MAX as usize) {
                            eprintln!("WARNING: network config is larger than 65536 bytes!");
                            return Err(OutOfBoundsError); // abort
                        }

                        packet.append_u64(config.network_id.into())?;
                        packet.append_u16(config_data.len() as u16)?;
                        packet.append_bytes(config_data.as_slice())?;

                        // TODO: for V1 we may need to introduce use of the chunking mechanism for large configs.
                    }

                    let new_payload_len = protocol::compress(&mut packet.as_bytes_mut()[payload_start..]);
                    packet.set_size(payload_start + new_payload_len);

                    Ok(())
                },
            );
        }
    }

    /// Send one or more revocation object(s) to a peer. The provided vector is drained.
    fn send_revocations(&self, peer: &Peer, revocations: &mut Vec<Revocation>) {
        if let Some(host_system) = self.service.read().unwrap().upgrade() {
            let time_ticks = ms_monotonic();
            while !revocations.is_empty() {
                let send_count = revocations.len().min(protocol::UDP_DEFAULT_MTU / 256);
                debug_assert!(send_count <= (u16::MAX as usize));
                peer.send(
                    host_system.as_ref(),
                    host_system.node(),
                    None,
                    time_ticks,
                    |packet| -> Result<(), OutOfBoundsError> {
                        let payload_start = packet.len();

                        packet.append_u8(protocol::message_type::VL2_NETWORK_CREDENTIALS)?;
                        packet.append_u8(0)?;
                        packet.append_u16(0)?;
                        packet.append_u16(0)?;
                        packet.append_u16(send_count as u16)?;
                        for _ in 0..send_count {
                            let r = revocations.pop().unwrap();
                            packet.append_bytes(r.v1_proto_to_bytes(self.local_identity.address).as_bytes())?;
                        }
                        packet.append_u16(0)?;

                        let new_payload_len = protocol::compress(&mut packet.as_bytes_mut()[payload_start..]);
                        packet.set_size(payload_start + new_payload_len);

                        Ok(())
                    },
                );
            }
        }
    }

    async fn deauthorize_member(&self, member: &Member) {
        let time_clock = ms_since_epoch();
        let mut revocations = Vec::with_capacity(1);
        if let Ok(all_network_members) = self.database.list_members(member.network_id).await {
            for m in all_network_members.iter() {
                if member.node_id != *m {
                    if let Some(peer) = self.service.read().unwrap().upgrade().and_then(|s| s.node().peer(*m)) {
                        revocations.clear();
                        Revocation::new(member.network_id, time_clock, member.node_id, *m, &self.local_identity, false)
                            .map(|r| revocations.push(r));
                        self.send_revocations(&peer, &mut revocations);
                    }
                }
            }
        }
    }

    /// Attempt to create a network configuration and return the result.
    ///
    /// This is the central function of the controller that looks up members, checks their
    /// permissions, and generates a network config and other credentials (or not).
    ///
    /// An error is only returned if a database or other unusual error occurs. Otherwise a rejection
    /// reason is returned with None or an acceptance reason with a network configuration is returned.
    async fn authorize(
        self: &Arc<Self>,
        source_identity: &Verified<Identity>,
        network_id: NetworkId,
        time_clock: i64,
    ) -> Result<(AuthenticationResult, Option<NetworkConfig>), Box<dyn Error + Send + Sync>> {
        let network = self.database.get_network(network_id).await?;
        if network.is_none() {
            return Ok((AuthenticationResult::Rejected, None));
        }
        let network = network.unwrap();

        let mut member = self.database.get_member(network_id, source_identity.address).await?;
        let mut member_changed = false;

        // SECURITY WARNING: this is a critical code path where members of networks are authorized.
        // Read and modify with extreme care.

        // If we have a member object and a pinned identity, check to make sure it matches. Also accept
        // upgraded identities to replace old versions if they are properly formed and their signatures
        // all check out (see Identity::is_upgraded_from()). Note that we do not pin the identity here
        // if it is unspecified. That's not done until we fully authorize this member, since we don't
        // want to have a way to somehow pin the wrong person's identity (if someone manages to somehow
        // create a colliding identity and get it to us).
        if let Some(member) = member.as_mut() {
            if let Some(pinned_identity) = member.identity.as_ref() {
                if !pinned_identity.eq(&source_identity) {
                    if source_identity.is_upgraded_from(pinned_identity) {
                        // Upgrade identity types if we have a V2 identity upgraded from a V1 identity.
                        let _ = member.identity.replace(source_identity.clone_without_secret());
                        let _ = member.identity_fingerprint.replace(Blob::from(source_identity.fingerprint));
                        member_changed = true;
                    } else {
                        return Ok((AuthenticationResult::RejectedIdentityMismatch, None));
                    }
                }
            }

            if let Some(pinned_fingerprint) = member.identity_fingerprint.as_ref() {
                if pinned_fingerprint.as_bytes().eq(&source_identity.fingerprint) {
                    if member.identity.is_none() {
                        // Learn the FULL identity if the fingerprint is pinned and they match. This
                        // lets us add members by address/fingerprint with full SHA384 identity
                        // verification instead of just by short address.
                        let _ = member.identity.replace(source_identity.clone_without_secret());
                        member_changed = true;
                    }
                } else {
                    return Ok((AuthenticationResult::RejectedIdentityMismatch, None));
                }
            }
        }

        let mut authentication_result = AuthenticationResult::Rejected;

        // This is the main "authorized" state of the member record. If it is true then the member is allowed,
        // but with the caveat that SSO must be checked if it's enabled on the network. If this is false then
        // the member is rejected unless auto-authorized via a mechanism like public networks below.
        let mut member_authorized = member.as_ref().map_or(false, |m| m.authorized());

        // If the member isn't authorized, check to see if it should be auto-authorized.
        if !member_authorized {
            if member.is_none() {
                if network.learn_members.unwrap_or(true) {
                    let _ = member.insert(Member::new_with_identity(source_identity.as_ref().clone(), network_id));
                    member_changed = true;
                } else {
                    return Ok((AuthenticationResult::Rejected, None));
                }
            }

            if network.private {
                // TODO: check token authorization
            } else {
                authentication_result = AuthenticationResult::ApprovedIsPublicNetwork;
                member.as_mut().unwrap().last_authorized_time = Some(time_clock);
                member_authorized = true;
                member_changed = true;
            }
        }

        let mut member = member.unwrap();

        // If the member is authorized set the final verdict to reflect this unless SSO (third party auth)
        // is enabled on the network and disagrees. Skip if the verdict is already one of the approved
        // values, which would indicate auth-authorization above.
        if member_authorized {
            if !authentication_result.approved() {
                // TODO: check SSO if enabled on network!
                authentication_result = AuthenticationResult::Approved;
            }
        } else {
            // This should not be able to be in approved state if member_authorized is still false.
            assert!(!authentication_result.approved());
        }

        // drop 'mut' from these since they should no longer change
        let member_authorized = member_authorized;
        let authentication_result = authentication_result;

        // Generate network configuration if the member is authorized.
        let network_config = if authentication_result.approved() {
            // We should not be able to make it here if this is still false.
            assert!(member_authorized);

            // Pin member identity if not pinned already. This is analogous to SSH "trust on first use" except
            // that the ZeroTier address is akin to the host name. Once we've seen the full identity once then
            // it becomes truly "impossible" to collide the address. (Unless you can break ECC and SHA384.)
            if member.identity.is_none() {
                let _ = member.identity.replace(source_identity.clone_without_secret());
                debug_assert!(member.identity_fingerprint.is_none());
                let _ = member.identity_fingerprint.replace(Blob::from(source_identity.fingerprint));
                member_changed = true;
            }

            // Make sure these agree. It should be impossible to end up with a member that's authorized and
            // whose identity and identity fingerprint don't match.
            if !member
                .identity
                .as_ref()
                .unwrap()
                .fingerprint
                .eq(member.identity_fingerprint.as_ref().unwrap().as_bytes())
            {
                debug_assert!(false);
                return Ok((AuthenticationResult::RejectedDueToError, None));
            }

            // Figure out TTL for credentials (time window in V1).
            let credential_ttl = network.credential_ttl.unwrap_or(CREDENTIAL_WINDOW_SIZE_DEFAULT);

            // Check and if necessary auto-assign static IPs for this member.
            member_changed |= network.assign_ip_addresses(self.database.as_ref(), &mut member).await;

            let mut nc = NetworkConfig::new(network_id, source_identity.address);

            nc.name = network.name.clone();
            nc.private = network.private;
            nc.timestamp = time_clock;
            nc.multicast_limit = network.multicast_limit.unwrap_or(DEFAULT_MULTICAST_LIMIT as u32);
            nc.multicast_like_expire = Some(protocol::VL2_DEFAULT_MULTICAST_LIKE_EXPIRE as u32);
            nc.mtu = network.mtu.unwrap_or(ZEROTIER_VIRTUAL_NETWORK_DEFAULT_MTU as u16);
            nc.routes = network.ip_routes.iter().cloned().collect();
            nc.static_ips = member.ip_assignments.iter().cloned().collect();

            // For any members that have been deauthorized but may still be in the cert agreement window,
            // insert rules to drop packets to/from those members. This lets us ban them without
            // adjusting the window, which is a simpler approach and has less risk of interrupting
            // connectivity between valid members.
            if let Ok(mut deauthed_members_still_in_window) = self
                .database
                .list_members_deauthorized_after(network.id, time_clock - (credential_ttl as i64))
                .await
            {
                if !deauthed_members_still_in_window.is_empty() {
                    deauthed_members_still_in_window.sort_unstable(); // may improve packet compression
                    nc.rules.reserve(deauthed_members_still_in_window.len() + 1);
                    let mut or = false;
                    for dead in deauthed_members_still_in_window.iter() {
                        nc.rules.push(vl2::rule::Rule::match_source_zerotier_address(false, or, *dead));
                        or = true;
                    }
                    nc.rules.push(vl2::rule::Rule::action_drop());
                }
            }

            // Then add the rest of the user-defined rules, or a blanket accept if there are none.
            if let Some(rules) = network.rules.as_ref() {
                nc.rules.reserve(rules.len());
                for r in rules.iter() {
                    nc.rules.push(r.clone());
                }
            } else {
                nc.rules.push(vl2::rule::Rule::action_accept());
            }

            nc.dns = network.dns.iter().map(|(k, v)| (k.clone(), v.iter().cloned().collect())).collect();

            if network.min_supported_version.unwrap_or(0) < (protocol::PROTOCOL_VERSION_V2 as u32) {
                // If this network supports V1 nodes we have to include V1 credentials. Otherwise we can skip
                // the overhead (bandwidth and CPU) of generating these.

                if let Some(com) =
                    vl2::v1::CertificateOfMembership::new(&self.local_identity, network_id, &source_identity, time_clock, credential_ttl)
                {
                    let mut v1cred = V1Credentials {
                        revision: time_clock as u64,
                        max_delta: credential_ttl,
                        certificate_of_membership: com,
                        certificates_of_ownership: Vec::new(),
                        tags: HashMap::new(),
                    };

                    if !nc.static_ips.is_empty() {
                        let mut coo = vl2::v1::CertificateOfOwnership::new(network_id, time_clock, source_identity.address);
                        for ip in nc.static_ips.iter() {
                            coo.add_ip(ip);
                        }
                        if !coo.sign(&self.local_identity, &source_identity) {
                            return Ok((AuthenticationResult::RejectedDueToError, None));
                        }
                        v1cred.certificates_of_ownership.push(coo);
                    }

                    for (id, value) in member.tags.iter() {
                        let tag = vl2::v1::Tag::new(*id, *value, &self.local_identity, network_id, &source_identity, time_clock);
                        if tag.is_none() {
                            return Ok((AuthenticationResult::RejectedDueToError, None));
                        }
                        let _ = v1cred.tags.insert(*id, tag.unwrap());
                    }

                    nc.v1_credentials = Some(v1cred);
                } else {
                    return Ok((AuthenticationResult::RejectedDueToError, None));
                }
            }

            if source_identity.p384.is_some() {
                // TODO: create V2 type credential for V2 nodes
            }

            // Log this member in the recently authorized cache, which is currently just used to filter whether we should
            // handle multicast subscription traffic.
            let _ = self
                .recently_authorized
                .write()
                .unwrap()
                .entry(source_identity.fingerprint)
                .or_default()
                .insert(network_id, ms_monotonic() + (credential_ttl as i64));

            Some(nc)
        } else {
            None
        };

        // Save any changes to member record.
        if member_changed {
            self.database.save_member(member, false).await?;
        }

        Ok((authentication_result, network_config))
    }
}

impl InnerProtocol for Controller {
    fn handle_packet<HostSystemImpl: HostSystem + ?Sized>(
        &self,
        host_system: &HostSystemImpl,
        _: &Node,
        source: &Arc<Peer>,
        source_path: &Arc<Path>,
        source_hops: u8,
        message_id: u64,
        verb: u8,
        payload: &PacketBuffer,
        mut cursor: usize,
    ) -> PacketHandlerResult {
        match verb {
            protocol::message_type::VL2_NETWORK_CONFIG_REQUEST => {
                if !host_system.should_respond_to(&source.identity) {
                    return PacketHandlerResult::Ok; // handled and ignored
                }

                let network_id = payload.read_u64(&mut cursor);
                if network_id.is_err() {
                    return PacketHandlerResult::Error;
                }
                let network_id = NetworkId::from_u64(network_id.unwrap());
                if network_id.is_none() {
                    return PacketHandlerResult::Error;
                }
                let network_id = network_id.unwrap();

                debug_event!(
                    host_system,
                    "[vl2] NETWORK_CONFIG_REQUEST from {}({}) for {:0>16x}",
                    source.identity.address.to_string(),
                    source_path.endpoint.to_string(),
                    u64::from(network_id)
                );

                let metadata = if (cursor + 2) < payload.len() {
                    let meta_data_len = payload.read_u16(&mut cursor);
                    if meta_data_len.is_err() {
                        return PacketHandlerResult::Error;
                    }
                    if let Ok(d) = payload.read_bytes(meta_data_len.unwrap() as usize, &mut cursor) {
                        d.to_vec()
                    } else {
                        return PacketHandlerResult::Error;
                    }
                } else {
                    Vec::new()
                };

                // Launch handler as an async background task.
                let (self2, source, source_remote_endpoint) =
                    (self.self_ref.upgrade().unwrap(), source.clone(), source_path.endpoint.clone());
                self.reaper.add(
                    self.runtime.spawn(async move {
                        let node_id = source.identity.address;
                        let node_fingerprint = Blob::from(source.identity.fingerprint);
                        let now = ms_since_epoch();

                        let (result, config) = match self2.authorize(&source.identity, network_id, now).await {
                            Result::Ok((result, Some(config))) => {
                                //println!("{}", serde_yaml::to_string(&config).unwrap());
                                self2.send_network_config(source.as_ref(), &config, Some(message_id));
                                (result, Some(config))
                            }
                            Result::Ok((result, None)) => (result, None),
                            Result::Err(e) => {
                                #[cfg(debug_assertions)]
                                let host = self2.service.read().unwrap().clone().upgrade().unwrap();
                                debug_event!(host, "[vl2] ERROR getting network config: {}", e.to_string());
                                return;
                            }
                        };

                        let _ = self2
                            .database
                            .log_request(RequestLogItem {
                                network_id,
                                node_id,
                                node_fingerprint,
                                controller_node_id: self2.local_identity.address,
                                metadata,
                                peer_version: source.version(),
                                peer_protocol_version: source.protocol_version(),
                                timestamp: now,
                                source_remote_endpoint,
                                source_hops,
                                result,
                                config,
                            })
                            .await;
                    }),
                    Instant::now().checked_add(REQUEST_TIMEOUT).unwrap(),
                );

                PacketHandlerResult::Ok
            }

            protocol::message_type::VL2_MULTICAST_LIKE => {
                let auth = self.recently_authorized.read().unwrap();
                let time_ticks = ms_monotonic();
                self.multicast_authority.handle_vl2_multicast_like(
                    |network_id, identity| {
                        auth.get(&identity.fingerprint)
                            .map_or(false, |t| t.get(&network_id).map_or(false, |t| *t > time_ticks))
                    },
                    time_ticks,
                    source,
                    payload,
                    cursor,
                );
                PacketHandlerResult::Ok
            }

            protocol::message_type::VL2_MULTICAST_GATHER => {
                if let Some(service) = self.service.read().unwrap().upgrade() {
                    let auth = self.recently_authorized.read().unwrap();
                    let time_ticks = ms_monotonic();
                    self.multicast_authority.handle_vl2_multicast_gather(
                        |network_id, identity| {
                            auth.get(&identity.fingerprint)
                                .map_or(false, |t| t.get(&network_id).map_or(false, |t| *t > time_ticks))
                        },
                        time_ticks,
                        service.as_ref(),
                        service.node(),
                        source,
                        message_id,
                        payload,
                        cursor,
                    );
                }
                PacketHandlerResult::Ok
            }

            _ => PacketHandlerResult::NotHandled,
        }
    }
}

impl VL1AuthProvider for Controller {
    #[inline(always)]
    fn should_respond_to(&self, _: &Verified<Identity>) -> bool {
        // Controllers always have to establish sessions to process requests. We don't really know if
        // a member is relevant until we have looked up both the network and the member, since whether
        // or not to "learn" unknown members is a network level option.
        true
    }

    fn has_trust_relationship(&self, id: &Verified<Identity>) -> bool {
        self.recently_authorized
            .read()
            .unwrap()
            .get(&id.fingerprint)
            .map_or(false, |by_network| by_network.values().any(|t| *t > ms_monotonic()))
    }
}

impl Drop for Controller {
    fn drop(&mut self) {
        for h in self.daemons.lock().unwrap().drain(..) {
            h.abort();
        }
    }
}