linkerd2 proxy destination 学习笔记

在本文章中,能粗略了解到 linker2 的代理服务 proxy 组件中关于 destination 的交互原理。

作者 哗啦啦mesh团队 发表于 2019年10月29日

作者: 哗啦啦 mesh团队,热衷于kubernetes、devops、apollo、istio、linkerd、openstack、calico 等领域技术。

linkerd2介绍

Linkerd由控制平面数据平面组成:

  • 控制平面是在所属的Kubernetes命名空间(linkerd默认情况下)中运行的一组服务,这些服务可以完成汇聚遥测数据,提供面向用户的API,并向数据平面代理提供控制数据等,它们共同驱动数据平面。

  • 数据平面用Rust编写的轻量级代理,该代理安装在服务的每个pod中,并成为数据平面的一部分,它接收Pod的所有接入流量,并通过initContainer配置iptables正确转发流量的拦截所有传出流量,因为它是附加工具,并且拦截服务的所有传入和传出流量,所以不需要更改代码,甚至可以将其添加到正在运行的服务中。

借用官方的图:

proxy-destination

proxy由rust开发完成,其内部的异步运行时采用了Tokio框架,服务组件用到了tower

本文主要关注proxy与destination组件交互相关的整体逻辑,分析proxy内部的运行逻辑。

流程分析

初始化

proxy启动后:

  1. app::init初始化配置
  2. app::Main::new创建主逻辑main
  3. main.run_until内新加一任务 ProxyParts::build_proxy_task

ProxyParts::build_proxy_task中会进行一系列的初始化工作,此处只关注dst_svc,其创建代码为:

    dst_svc = svc::stack(connect::svc(keepalive))
                .push(tls::client::layer(local_identity.clone()))
                .push_timeout(config.control_connect_timeout)
                .push(control::client::layer())
                .push(control::resolve::layer(dns_resolver.clone()))
                .push(reconnect::layer({
                    let backoff = config.control_backoff.clone();
                    move |_| Ok(backoff.stream())
                }))
                .push(http_metrics::layer::<_, classify::Response>(
                    ctl_http_metrics.clone(),
                ))
                .push(proxy::grpc::req_body_as_payload::layer().per_make())
                .push(control::add_origin::layer())
                .push_buffer_pending(
                    config.destination_buffer_capacity,
                    config.control_dispatch_timeout,
                )
                .into_inner()
                .make(config.destination_addr.clone())

dst_svc一共有2处引用,一是crate::resolve::Resolver的创建会涉及;另一个就是ProfilesClient的创建。

Resolver

  1. api_resolve::Resolve::new(dst_svc.clone())创建resolver对象
  2. 调用outbound::resolve创建 map_endpoint::Resolve类型对象,并当做参数resolve传入outbound::spawn函数开启出口线程

outbound::spawn中,resolve被用于创建负载均衡控制层,并用于后续路由控制:

let balancer_layer = svc::layers()
        .push_spawn_ready()
        .push(discover::Layer::new(
            DISCOVER_UPDATE_BUFFER_CAPACITY,
            resolve,
        ))
        .push(balance::layer(EWMA_DEFAULT_RTT, EWMA_DECAY));

discover::Layer::layer中:

let from_resolve = FromResolve::new(self.resolve.clone());
let make_discover = MakeEndpoint::new(make_endpoint, from_resolve);
Buffer::new(self.capacity, make_discover)

Profiles

  1. ProfilesClient::new中调用api::client::Destination::new(dst_svc)创建grpc的client端并存于成员变量service
  2. 接着profiles_client对象会被用于inboundoutbound的创建(省略无关代码):
    let dst_stack = svc::stack(...)...
        .push(profiles::router::layer(
            profile_suffixes,
            profiles_client,
            dst_route_stack,
        ))
        ...

其中profiles::router::layer会创建一个Layer对象,并将profiles_client赋予get_routes成员。然后在service方法中,会调到Layer::layer方法,里面会创建一个MakeSvc对象,其get_routes成员的值即为profiles_client

运行

新的连接过来时,从listen拿到连接对象后,会交给linkerd_proxy::transport::tls::accept::AcceptTlscall,然后是linkerd2_proxy::proxy::server::Servercall,并最终分别调用linkerd2_proxy_http::balance::MakeSvc::calllinkerd2_proxy_http::profiles::router::MakeSvc::call方法。

balance

linkerd2_proxy_http::balance::MakeSvc::call中:

  1. 调用inner.call(target),此处的inner即是前面Buffer::new的结果。
  2. 生成一个新的linkerd2_proxy_http::balance::MakeSvc对象,当做Future返回

先看inner.call。它内部经过层层调用,依次触发BufferMakeEndpointFromResolve等结构的call方法,最终会触发最开始创建的resolve.resolve(target),其内部调用api_resolve::Resolve::call

api_resolve::Resolve::call中:

    fn call(&mut self, target: T) -> Self::Future {
        let path = target.to_string();
        trace!("resolve {:?}", path);
        self.service
            // GRPC请求,获取k8s的endpoint
            .get(grpc::Request::new(api::GetDestination {
                path,
                scheme: self.scheme.clone(),
                context_token: self.context_token.clone(),
            }))
            .map(|rsp| {
                debug!(metadata = ?rsp.metadata());
                // 拿到结果stream
                Resolution {
                    inner: rsp.into_inner(),
                }
            })
    }

将返回的Resolution再次放入MakeSvc中,然后看其poll:

    fn poll(&mut self) -> Poll<Self::Item, Self::Error> {
        // 这个poll会依次调用:
        //    linkerd2_proxy_api_resolve::resolve::Resolution::poll
        //    linkerd2_proxy_discover::from_resolve::DiscoverFuture::poll
        //    linkerd2_proxy_discover::make_endpoint::DiscoverFuture::poll
        // 最终获得Poll<Change<SocketAddr, Endpoint>> 
        let discover = try_ready!(self.inner.poll());
        let instrument = PendingUntilFirstData::default();
        let loaded = PeakEwmaDiscover::new(discover, self.default_rtt, self.decay, instrument);
        let balance = Balance::new(loaded, self.rng.clone());
        Ok(Async::Ready(balance))
    }

最终返回service Balance

当具体请求过来后,先会判断Balance::poll_ready

    fn poll_ready(&mut self) -> Poll<(), Self::Error> {
        // 获取Update<Endpoint>
        // 将Remove的从self.ready_services中删掉
        // 将Insert的构造UnreadyService结构加到self.unready_services
        self.poll_discover()?;
        // 对UnreadyService,调用其poll,内部会调用到svc的poll_ready判断endpoint是否可用
        // 可用时,将其加入self.ready_services
        self.poll_unready();
        
        loop {
            if let Some(index) = self.next_ready_index {
                // 找到对应的endpoint,可用则返回
                if let Ok(Async::Ready(())) = self.poll_ready_index_or_evict(index) {
                    return Ok(Async::Ready(()));
                }
            }
            // 选择负载比较低的endpoint
            self.next_ready_index = self.p2c_next_ready_index();
            if self.next_ready_index.is_none() {
                // 
                return Ok(Async::NotReady);
            }
        }
    }

就绪后,对请求req调用call

    fn call(&mut self, request: Req) -> Self::Future {
        // 找到下一个可用的svc,并将其从ready_services中删除
        let index = self.next_ready_index.take().expect("not ready");
        let (key, mut svc) = self
            .ready_services
            .swap_remove_index(index)
            .expect("invalid ready index");

        // 将请求转过去
        let fut = svc.call(request);
        // 加到unready
        self.push_unready(key, svc);

        fut.map_err(Into::into)
    }

profiles

linkerd2_proxy_http::profiles::router::MakeSvc::call中:

        // Initiate a stream to get route and dst_override updates for this
        // destination.
        let route_stream = match target.get_destination() {
            Some(ref dst) => {
                if self.suffixes.iter().any(|s| s.contains(dst.name())) {
                    debug!("fetching routes for {:?}", dst);
                    self.get_routes.get_routes(&dst)
                } else {
                    debug!("skipping route discovery for dst={:?}", dst);
                    None
                }
            }
            None => {
                debug!("no destination for routes");
                None
            }
        };

经过若干判断后,会调用ProfilesClient::get_routes并将结果存于route_stream

进入get_routes

    fn get_routes(&self, dst: &NameAddr) -> Option<Self::Stream> {
        // 创建通道
        let (tx, rx) = mpsc::channel(1);
        // This oneshot allows the daemon to be notified when the Self::Stream
        // is dropped.
        let (hangup_tx, hangup_rx) = oneshot::channel();
        // 创建Daemon对象(Future任务)
        let daemon = Daemon {
            tx,
            hangup: hangup_rx,
            dst: format!("{}", dst),
            state: State::Disconnected,
            service: self.service.clone(),
            backoff: self.backoff,
            context_token: self.context_token.clone(),
        };
        // 调用Daemon::poll
        let spawn = DefaultExecutor::current().spawn(Box::new(daemon.map_err(|_| ())));
        // 将通道接收端传出
        spawn.ok().map(|_| Rx {
            rx,
            _hangup: hangup_tx,
        })
    }

接着看Daemon::poll

    fn poll(&mut self) -> Poll<Self::Item, Self::Error> {
        loop {
            // 遍历state成员状态
            self.state = match self.state {
                // 未连接时
                State::Disconnected => {
                    match self.service.poll_ready() {
                        Ok(Async::NotReady) => return Ok(Async::NotReady),
                        Ok(Async::Ready(())) => {}
                        Err(err) => {
                            error!(
                                "profile service unexpected error (dst = {}): {:?}",
                                self.dst, err,
                            );
                            return Ok(Async::Ready(()));
                        }
                    };
                    // 构造grpc请求
                    let req = api::GetDestination {
                        scheme: "k8s".to_owned(),
                        path: self.dst.clone(),
                        context_token: self.context_token.clone(),
                    };
                    debug!("getting profile: {:?}", req);
                    // 获取请求任务
                    let rspf = self.service.get_profile(grpc::Request::new(req));
                    State::Waiting(rspf)
                }
                // 正在请求时,从请求中获取回复
                State::Waiting(ref mut f) => match f.poll() {
                    Ok(Async::NotReady) => return Ok(Async::NotReady),
                    // 正常回复
                    Ok(Async::Ready(rsp)) => {
                        trace!("response received");
                        // 流式回复
                        State::Streaming(rsp.into_inner())
                    }
                    Err(e) => {
                        warn!("error fetching profile for {}: {:?}", self.dst, e);
                        State::Backoff(Delay::new(clock::now() + self.backoff))
                    }
                },
                // 接收回复
                State::Streaming(ref mut s) => {
                    // 处理回复流
                    // 注意此处,参数1是get_profile请求的回复流,
                    //   参数2是之前创建的通道发送端
                    match Self::proxy_stream(s, &mut self.tx, &mut self.hangup) {
                        Async::NotReady => return Ok(Async::NotReady),
                        Async::Ready(StreamState::SendLost) => return Ok(().into()),
                        Async::Ready(StreamState::RecvDone) => {
                            State::Backoff(Delay::new(clock::now() + self.backoff))
                        }
                    }
                }
                // 异常,结束请求
                State::Backoff(ref mut f) => match f.poll() {
                    Ok(Async::NotReady) => return Ok(Async::NotReady),
                    Err(_) | Ok(Async::Ready(())) => State::Disconnected,
                },
            };
        }
    }

接着 proxy_stream

    fn proxy_stream(
        rx: &mut grpc::Streaming<api::DestinationProfile, T::ResponseBody>,
        tx: &mut mpsc::Sender<profiles::Routes>,
        hangup: &mut oneshot::Receiver<Never>,
    ) -> Async<StreamState> {
        loop {
            // 发送端是否就绪
            match tx.poll_ready() {
                Ok(Async::NotReady) => return Async::NotReady,
                Ok(Async::Ready(())) => {}
                Err(_) => return StreamState::SendLost.into(),
            }

            // 从grpc stream中取得一条数据
            match rx.poll() {
                Ok(Async::NotReady) => match hangup.poll() {
                    Ok(Async::Ready(never)) => match never {}, // unreachable!
                    Ok(Async::NotReady) => {
                        // We are now scheduled to be notified if the hangup tx
                        // is dropped.
                        return Async::NotReady;
                    }
                    Err(_) => {
                        // Hangup tx has been dropped.
                        debug!("profile stream cancelled");
                        return StreamState::SendLost.into();
                    }
                },
                Ok(Async::Ready(None)) => return StreamState::RecvDone.into(),
                // 正确取得profile结构
                Ok(Async::Ready(Some(profile))) => {
                    debug!("profile received: {:?}", profile);
                    // 解析数据
                    let retry_budget = profile.retry_budget.and_then(convert_retry_budget);
                    let routes = profile
                        .routes
                        .into_iter()
                        .filter_map(move |orig| convert_route(orig, retry_budget.as_ref()))
                        .collect();
                    let dst_overrides = profile
                        .dst_overrides
                        .into_iter()
                        .filter_map(convert_dst_override)
                        .collect();
                    // 构造profiles::Routes结构并推到发送端
                    match tx.start_send(profiles::Routes {
                        routes,
                        dst_overrides,
                    }) {
                        Ok(AsyncSink::Ready) => {} // continue
                        Ok(AsyncSink::NotReady(_)) => {
                            info!("dropping profile update due to a full buffer");
                            // This must have been because another task stole
                            // our tx slot? It seems pretty unlikely, but possible?
                            return Async::NotReady;
                        }
                        Err(_) => {
                            return StreamState::SendLost.into();
                        }
                    }
                }
                Err(e) => {
                    warn!("profile stream failed: {:?}", e);
                    return StreamState::RecvDone.into();
                }
            }
        }
    }

回到MakeSvc::call方法,前面创建的route_stream会被用于创建一个linkerd2_proxy::proxy::http::profiles::router::Service任务对象,并在其poll_ready方法中通过poll_route_streamroute_steam获取profiles::Routes并调用update_routes创建具体可用的路由规则linkerd2_router::Router,至此,路由规则已建好,就等具体的请求过来然后在call中调用linkerd2_router::call进行对请求的路由判断。

图示

profile

proxy-destination

总结

proxy采用的tower框架,每个处理逻辑都是其中的一个layer,开发时只需层层堆叠即可。不过,也正因如此,各层之间的接口都极其相似,须得小心不可调错。 对于destination这部分逻辑,linkerd2的destination组件收到来自proxy的grpc请求后,每当endpoint或service profile有任何变动,都会立即通过stream发送过去,proxy收到后根据endpoint调整负载均衡策略,根据service profile调整路由,然后通过它们来处理用户服务的实际请求。

Tags: