线程模型

srs使用了state-threads协程库，是单线程多协程模型。
这个协程的概念类似于lua的协程，都是单线程中可以创建多个协程。而golang中的goroutine协程是多线程并发的，goroutine有可能运行在同一个线程也可能在不同线程，这样就有了线程安全问题，所以需要chan通信或者mutex加锁共享资源。
而srs因为是单线程多协程所以不用考虑线程安全，数据不用加锁。

主流程分析

撇掉程序启动的一些初始化和设置，直接进入：

int SrsServer::listen()
{
    int ret = ERROR_SUCCESS;
    
    if ((ret = listen_rtmp()) != ERROR_SUCCESS) {
        return ret;
    }
    
    if ((ret = listen_http_api()) != ERROR_SUCCESS) {
        return ret;
    }
    
    if ((ret = listen_http_stream()) != ERROR_SUCCESS) {
        return ret;
    }
    
    if ((ret = listen_stream_caster()) != ERROR_SUCCESS) {
        return ret;
    }
    
    return ret;
}

先看看listen_rtmp():

int SrsServer::listen_rtmp()
{
    int ret = ERROR_SUCCESS;
    
    // stream service port.
    std::vector<std::string> ip_ports = _srs_config->get_listens();
    srs_assert((int)ip_ports.size() > 0);
    
    close_listeners(SrsListenerRtmpStream);
    
    for (int i = 0; i < (int)ip_ports.size(); i++) {
        SrsListener* listener = new SrsStreamListener(this, SrsListenerRtmpStream);
        listeners.push_back(listener);
        
        std::string ip;
        int port;
        srs_parse_endpoint(ip_ports[i], ip, port);
        
        if ((ret = listener->listen(ip, port)) != ERROR_SUCCESS) {
            srs_error("RTMP stream listen at %s:%d failed. ret=%d", ip.c_str(), port, ret);
            return ret;
        }
    }
    
    return ret;
}

创建了SrsStreamListener，在SrsStreamListener::listen中又创建了SrsTcpListener进行listen。

SrsTcpListener::SrsTcpListener(ISrsTcpHandler* h, string i, int p)
{
    handler = h;
    ip = i;
    port = p;

    _fd = -1;
    _stfd = NULL;

    pthread = new SrsReusableThread("tcp", this);
}

在SrsTcpListener中创建了pthread: SrsReusableThread。
在int SrsTcpListener::listen()中调用了pthread->start()，协程会回调到int SrsTcpListener::cycle()。

int SrsTcpListener::cycle()
{
    int ret = ERROR_SUCCESS;
    
    st_netfd_t client_stfd = st_accept(_stfd, NULL, NULL, ST_UTIME_NO_TIMEOUT);
    
    if(client_stfd == NULL){
        // ignore error.
        if (errno != EINTR) {
            srs_error("ignore accept thread stoppped for accept client error");
        }
        return ret;
    }
    srs_verbose("get a client. fd=%d", st_netfd_fileno(client_stfd));
    
    if ((ret = handler->on_tcp_client(client_stfd)) != ERROR_SUCCESS) {
        srs_warn("accept client error. ret=%d", ret);
        return ret;
    }
    
    return ret;
}

accept连接后，回调到on_tcp_client。
也就是SrsStreamListener::on_tcp_client：

int SrsStreamListener::on_tcp_client(st_netfd_t stfd)
{
    int ret = ERROR_SUCCESS;
    
    if ((ret = server->accept_client(type, stfd)) != ERROR_SUCCESS) {
        srs_warn("accept client error. ret=%d", ret);
        return ret;
    }

    return ret;
}

int SrsServer::accept_client(SrsListenerType type, st_netfd_t client_stfd)
{
...
    SrsConnection* conn = NULL;
    if (type == SrsListenerRtmpStream) {
        conn = new SrsRtmpConn(this, client_stfd);
    } else if (type == SrsListenerHttpApi) {
#ifdef SRS_AUTO_HTTP_API
        conn = new SrsHttpApi(this, client_stfd, http_api_mux);
#else
        srs_warn("close http client for server not support http-api");
        srs_close_stfd(client_stfd);
        return ret;
#endif
    } else if (type == SrsListenerHttpStream) {
#ifdef SRS_AUTO_HTTP_SERVER
        conn = new SrsResponseOnlyHttpConn(this, client_stfd, http_server);
#else
        srs_warn("close http client for server not support http-server");
        srs_close_stfd(client_stfd);
        return ret;
#endif
    } else {
        // TODO: FIXME: handler others
    }
    srs_assert(conn);
    
    // directly enqueue, the cycle thread will remove the client.
    conns.push_back(conn);
    srs_verbose("add conn to vector.");
    
    // cycle will start process thread and when finished remove the client.
    // @remark never use the conn, for it maybe destroyed.
    if ((ret = conn->start()) != ERROR_SUCCESS) {
        return ret;
    }
    srs_verbose("conn started success.");

    srs_verbose("accept client finished. conns=%d, ret=%d", (int)conns.size(), ret);
    
    return ret;
}

在上面根据type创建不同的SrsConnection，Rtmp创建了SrsRtmpConn，并且加入到std::vector<SrsConnection*> conns;中，然后执行conn->start()。

SrsConnection基类创建了一个协程pthread: SrsOneCycleThread，上面的conn->start()，实际上是pthread->start()：

SrsConnection::SrsConnection(IConnectionManager* cm, st_netfd_t c)
{
    id = 0;
    manager = cm;
    stfd = c;
    disposed = false;
    expired = false;
    
    // the client thread should reap itself, 
    // so we never use joinable.
    // TODO: FIXME: maybe other thread need to stop it.
    // @see: https://github.com/ossrs/srs/issues/78
    pthread = new SrsOneCycleThread("conn", this);
}

int SrsConnection::start()
{
    return pthread->start();
}

在int SrsConnection::cycle()调用了do_cycle()，派生类实现了这个方法。

int SrsRtmpConn::do_cycle()
{
    int ret = ERROR_SUCCESS;
    
    srs_trace("RTMP client ip=%s", ip.c_str());

    rtmp->set_recv_timeout(SRS_CONSTS_RTMP_RECV_TIMEOUT_US);
    rtmp->set_send_timeout(SRS_CONSTS_RTMP_SEND_TIMEOUT_US);
    
    //正式进入rtmp握手。
    if ((ret = rtmp->handshake()) != ERROR_SUCCESS) {
        srs_error("rtmp handshake failed. ret=%d", ret);
        return ret;
    }
    srs_verbose("rtmp handshake success");
    
    if ((ret = rtmp->connect_app(req)) != ERROR_SUCCESS) {
        srs_error("rtmp connect vhost/app failed. ret=%d", ret);
        return ret;
    }
    srs_verbose("rtmp connect app success");
    
    // set client ip to request.
    req->ip = ip;
    
    srs_trace("connect app, "
        "tcUrl=%s, pageUrl=%s, swfUrl=%s, schema=%s, vhost=%s, port=%s, app=%s, args=%s", 
        req->tcUrl.c_str(), req->pageUrl.c_str(), req->swfUrl.c_str(), 
        req->schema.c_str(), req->vhost.c_str(), req->port.c_str(),
        req->app.c_str(), (req->args? "(obj)":"null"));
    
    // show client identity
    if(req->args) {
        std::string srs_version;
        std::string srs_server_ip;
        int srs_pid = 0;
        int srs_id = 0;
        
        SrsAmf0Any* prop = NULL;
        if ((prop = req->args->ensure_property_string("srs_version")) != NULL) {
            srs_version = prop->to_str();
        }
        if ((prop = req->args->ensure_property_string("srs_server_ip")) != NULL) {
            srs_server_ip = prop->to_str();
        }
        if ((prop = req->args->ensure_property_number("srs_pid")) != NULL) {
            srs_pid = (int)prop->to_number();
        }
        if ((prop = req->args->ensure_property_number("srs_id")) != NULL) {
            srs_id = (int)prop->to_number();
        }
        
        srs_info("edge-srs ip=%s, version=%s, pid=%d, id=%d", 
            srs_server_ip.c_str(), srs_version.c_str(), srs_pid, srs_id);
        if (srs_pid > 0) {
            srs_trace("edge-srs ip=%s, version=%s, pid=%d, id=%d", 
                srs_server_ip.c_str(), srs_version.c_str(), srs_pid, srs_id);
        }
    }
    
    ret = service_cycle();
    
    http_hooks_on_close();

    return ret;
}

在这儿正式进入rtmp协议处理阶段。先进行握手：rtmp->handshake()等操作，然后进入service_cycle();。

int SrsRtmpConn::service_cycle()
{    
  ...
    while (!disposed) {
        ret = stream_service_cycle();
        
        // stream service must terminated with error, never success.
        // when terminated with success, it's user required to stop.
        if (ret == ERROR_SUCCESS) {
            continue;
        }
        
        // when not system control error, fatal error, return.
        if (!srs_is_system_control_error(ret)) {
            if (ret != ERROR_SOCKET_TIMEOUT && !srs_is_client_gracefully_close(ret)) {
                srs_error("stream service cycle failed. ret=%d", ret);
            }
            return ret;
        }
        
        // for republish, continue service
        if (ret == ERROR_CONTROL_REPUBLISH) {
            // set timeout to a larger value, wait for encoder to republish.
            rtmp->set_send_timeout(SRS_REPUBLISH_RECV_TIMEOUT_US);
            rtmp->set_recv_timeout(SRS_REPUBLISH_SEND_TIMEOUT_US);
            
            srs_trace("control message(unpublish) accept, retry stream service.");
            continue;
        }
        
        // for "some" system control error, 
        // logical accept and retry stream service.
        if (ret == ERROR_CONTROL_RTMP_CLOSE) {
            // TODO: FIXME: use ping message to anti-death of socket.
            // @see: https://github.com/ossrs/srs/issues/39
            // set timeout to a larger value, for user paused.
            rtmp->set_recv_timeout(SRS_PAUSED_RECV_TIMEOUT_US);
            rtmp->set_send_timeout(SRS_PAUSED_SEND_TIMEOUT_US);
            
            srs_trace("control message(close) accept, retry stream service.");
            continue;
        }
        
        // for other system control message, fatal error.
        srs_error("control message(%d) reject as error. ret=%d", ret, ret);
        return ret;
    }
    
    return ret;
}

stream_service_cycle:

int SrsRtmpConn::stream_service_cycle()
{
    int ret = ERROR_SUCCESS;
        
    SrsRtmpConnType type;
    if ((ret = rtmp->identify_client(res->stream_id, type, req->stream, req->duration)) != ERROR_SUCCESS) {
        if (!srs_is_client_gracefully_close(ret)) {
            srs_error("identify client failed. ret=%d", ret);
        }
        return ret;
    }
    
    srs_discovery_tc_url(req->tcUrl, req->schema, req->host, req->vhost, req->app, req->stream, req->port, req->param);
    req->strip();
    srs_trace("client identified, type=%s, stream_name=%s, duration=%.2f, param=%s",
        srs_client_type_string(type).c_str(), req->stream.c_str(), req->duration, req->param.c_str());
    
    // discovery vhost, resolve the vhost from config
    SrsConfDirective* parsed_vhost = _srs_config->get_vhost(req->vhost);
    if (parsed_vhost) {
        req->vhost = parsed_vhost->arg0();
    }
    
    if (req->schema.empty() || req->vhost.empty() || req->port.empty() || req->app.empty()) {
        ret = ERROR_RTMP_REQ_TCURL;
        srs_error("discovery tcUrl failed. "
                  "tcUrl=%s, schema=%s, vhost=%s, port=%s, app=%s, ret=%d",
                  req->tcUrl.c_str(), req->schema.c_str(), req->vhost.c_str(), req->port.c_str(), req->app.c_str(), ret);
        return ret;
    }
    
    if ((ret = check_vhost()) != ERROR_SUCCESS) {
        srs_error("check vhost failed. ret=%d", ret);
        return ret;
    }
    
    srs_trace("connected stream, tcUrl=%s, pageUrl=%s, swfUrl=%s, schema=%s, vhost=%s, port=%s, app=%s, stream=%s, param=%s, args=%s",
        req->tcUrl.c_str(), req->pageUrl.c_str(), req->swfUrl.c_str(),
        req->schema.c_str(), req->vhost.c_str(), req->port.c_str(),
        req->app.c_str(), req->stream.c_str(), req->param.c_str(), (req->args? "(obj)":"null"));
    
    // do token traverse before serve it.
    // @see https://github.com/ossrs/srs/pull/239
    if (true) {
        bool vhost_is_edge = _srs_config->get_vhost_is_edge(req->vhost);
        bool edge_traverse = _srs_config->get_vhost_edge_token_traverse(req->vhost);
        if (vhost_is_edge && edge_traverse) {
            if ((ret = check_edge_token_traverse_auth()) != ERROR_SUCCESS) {
                srs_warn("token auth failed, ret=%d", ret);
                return ret;
            }
        }
    }
    
    // security check
    if ((ret = security->check(type, ip, req)) != ERROR_SUCCESS) {
        srs_error("security check failed. ret=%d", ret);
        return ret;
    }
    srs_info("security check ok");
    
    // Never allow the empty stream name, for HLS may write to a file with empty name.
    // @see https://github.com/ossrs/srs/issues/834
    if (req->stream.empty()) {
        ret = ERROR_RTMP_STREAM_NAME_EMPTY;
        srs_error("RTMP: Empty stream name not allowed, ret=%d", ret);
        return ret;
    }

    // client is identified, set the timeout to service timeout.
    rtmp->set_recv_timeout(SRS_CONSTS_RTMP_RECV_TIMEOUT_US);
    rtmp->set_send_timeout(SRS_CONSTS_RTMP_SEND_TIMEOUT_US);
    
    // find a source to serve.
    SrsSource* source = NULL;
    if ((ret = SrsSource::fetch_or_create(req, server, &source)) != ERROR_SUCCESS) {
        return ret;
    }
    srs_assert(source != NULL);
    
    // update the statistic when source disconveried.
    SrsStatistic* stat = SrsStatistic::instance();
    if ((ret = stat->on_client(_srs_context->get_id(), req, this, type)) != ERROR_SUCCESS) {
        srs_error("stat client failed. ret=%d", ret);
        return ret;
    }

    bool vhost_is_edge = _srs_config->get_vhost_is_edge(req->vhost);
    bool enabled_cache = _srs_config->get_gop_cache(req->vhost);
    srs_trace("source url=%s, ip=%s, cache=%d, is_edge=%d, source_id=%d[%d]",
        req->get_stream_url().c_str(), ip.c_str(), enabled_cache, vhost_is_edge, 
        source->source_id(), source->source_id());
    source->set_cache(enabled_cache);
    
    client_type = type;
    //根据客户端类型进入不同分支
    switch (type) {
        case SrsRtmpConnPlay: {
            srs_verbose("start to play stream %s.", req->stream.c_str());
            
            // response connection start play
            if ((ret = rtmp->start_play(res->stream_id)) != ERROR_SUCCESS) {
                srs_error("start to play stream failed. ret=%d", ret);
                return ret;
            }
            if ((ret = http_hooks_on_play()) != ERROR_SUCCESS) {
                srs_error("http hook on_play failed. ret=%d", ret);
                return ret;
            }
            
            srs_info("start to play stream %s success", req->stream.c_str());
            ret = playing(source);
            http_hooks_on_stop();
            
            return ret;
        }
        case SrsRtmpConnFMLEPublish: {
            srs_verbose("FMLE start to publish stream %s.", req->stream.c_str());
            
            if ((ret = rtmp->start_fmle_publish(res->stream_id)) != ERROR_SUCCESS) {
                srs_error("start to publish stream failed. ret=%d", ret);
                return ret;
            }
            
            return publishing(source);
        }
        case SrsRtmpConnHaivisionPublish: {
            srs_verbose("Haivision start to publish stream %s.", req->stream.c_str());
            
            if ((ret = rtmp->start_haivision_publish(res->stream_id)) != ERROR_SUCCESS) {
                srs_error("start to publish stream failed. ret=%d", ret);
                return ret;
            }
            
            return publishing(source);
        }
        case SrsRtmpConnFlashPublish: {
            srs_verbose("flash start to publish stream %s.", req->stream.c_str());
            
            if ((ret = rtmp->start_flash_publish(res->stream_id)) != ERROR_SUCCESS) {
                srs_error("flash start to publish stream failed. ret=%d", ret);
                return ret;
            }
            
            return publishing(source);
        }
        default: {
            ret = ERROR_SYSTEM_CLIENT_INVALID;
            srs_info("invalid client type=%d. ret=%d", type, ret);
            return ret;
        }
    }

    return ret;
}

先进行tmp->identify_client客户端身份识别。
然后根据根据客户端类型(type)进入不同分支。
SrsRtmpConnPlay 客户端播流。
SrsRtmpConnFMLEPublish Rtmp推流到服务器。
SrsRtmpConnHaivisionPublish 应该是海康威视推流到服务器？
SrsRtmpConnFlashPublish Flash推流到服务器。
这儿只看SrsRtmpConnFMLEPublish：
进入int SrsRtmpConn::publishing(SrsSource* source)，然后int SrsRtmpConn::do_publishing(SrsSource* source, SrsPublishRecvThread* trd)，

int SrsRtmpConn::do_publishing(SrsSource* source, SrsPublishRecvThread* trd)
{
...
    // start isolate recv thread.
    if ((ret = trd->start()) != ERROR_SUCCESS) {
        srs_error("start isolate recv thread failed. ret=%d", ret);
        return ret;
    }
    ...
}

trd协程运行，协程循环：执行rtmp->recv_message(&msg)后调用int SrsPublishRecvThread::handle(SrsCommonMessage* msg)。
再回调到int SrsRtmpConn::handle_publish_message(SrsSource* source, SrsCommonMessage* msg, bool is_fmle, bool vhost_is_edge)。
之后处理收到的数据：

int SrsRtmpConn::process_publish_message(SrsSource* source, SrsCommonMessage* msg, bool vhost_is_edge)
{
    int ret = ERROR_SUCCESS;
    
    // for edge, directly proxy message to origin.
    if (vhost_is_edge) {
        if ((ret = source->on_edge_proxy_publish(msg)) != ERROR_SUCCESS) {
            srs_error("edge publish proxy msg failed. ret=%d", ret);
            return ret;
        }
        return ret;
    }
    
    // process audio packet
    if (msg->header.is_audio()) {
        if ((ret = source->on_audio(msg)) != ERROR_SUCCESS) {
            srs_error("source process audio message failed. ret=%d", ret);
            return ret;
        }
        return ret;
    }
    // process video packet
    if (msg->header.is_video()) {
        if ((ret = source->on_video(msg)) != ERROR_SUCCESS) {
            srs_error("source process video message failed. ret=%d", ret);
            return ret;
        }
        return ret;
    }
    
    // process aggregate packet
    if (msg->header.is_aggregate()) {
        if ((ret = source->on_aggregate(msg)) != ERROR_SUCCESS) {
            srs_error("source process aggregate message failed. ret=%d", ret);
            return ret;
        }
        return ret;
    }
    
    // process onMetaData
    if (msg->header.is_amf0_data() || msg->header.is_amf3_data()) {
        SrsPacket* pkt = NULL;
        if ((ret = rtmp->decode_message(msg, &pkt)) != ERROR_SUCCESS) {
            srs_error("decode onMetaData message failed. ret=%d", ret);
            return ret;
        }
        SrsAutoFree(SrsPacket, pkt);
    
        if (dynamic_cast<SrsOnMetaDataPacket*>(pkt)) {
            SrsOnMetaDataPacket* metadata = dynamic_cast<SrsOnMetaDataPacket*>(pkt);
            if ((ret = source->on_meta_data(msg, metadata)) != ERROR_SUCCESS) {
                srs_error("source process onMetaData message failed. ret=%d", ret);
                return ret;
            }
            srs_info("process onMetaData message success.");
            return ret;
        }
        
        srs_info("ignore AMF0/AMF3 data message.");
        return ret;
    }
    
    return ret;
}

如果本服务器是edge边缘服务器（vhost_is_edge）直接推流回源到源服务器。
audio和video分开处理。
这儿只看一下video的处理：

int SrsSource::on_video(SrsCommonMessage* shared_video)
{
    int ret = ERROR_SUCCESS;
    
    // monotically increase detect.
    if (!mix_correct && is_monotonically_increase) {
        if (last_packet_time > 0 && shared_video->header.timestamp < last_packet_time) {
            is_monotonically_increase = false;
            srs_warn("VIDEO: stream not monotonically increase, please open mix_correct.");
        }
    }
    last_packet_time = shared_video->header.timestamp;
    
    // drop any unknown header video.
    // @see https://github.com/ossrs/srs/issues/421
    if (!SrsFlvCodec::video_is_acceptable(shared_video->payload, shared_video->size)) {
        char b0 = 0x00;
        if (shared_video->size > 0) {
            b0 = shared_video->payload[0];
        }
        
        srs_warn("drop unknown header video, size=%d, bytes[0]=%#x", shared_video->size, b0);
        return ret;
    }
    
    // convert shared_video to msg, user should not use shared_video again.
    // the payload is transfer to msg, and set to NULL in shared_video.
    SrsSharedPtrMessage msg;
    if ((ret = msg.create(shared_video)) != ERROR_SUCCESS) {
        srs_error("initialize the video failed. ret=%d", ret);
        return ret;
    }
    srs_info("Video dts=%"PRId64", size=%d", msg.timestamp, msg.size);
    
    // directly process the audio message.
    if (!mix_correct) {
        return on_video_imp(&msg);
    }
    
    // insert msg to the queue.
    mix_queue->push(msg.copy());
    
    // fetch someone from mix queue.
    SrsSharedPtrMessage* m = mix_queue->pop();
    if (!m) {
        return ret;
    }
    
    // consume the monotonically increase message.
    if (m->is_audio()) {
        ret = on_audio_imp(m);
    } else {
        ret = on_video_imp(m);
    }
    srs_freep(m);
    
    return ret;
}

把shared_video转换为SrsSharedPtrMessage。
调用on_video_imp。

int SrsSource::on_video_imp(SrsSharedPtrMessage* msg)
{
    int ret = ERROR_SUCCESS;
    
    srs_info("Video dts=%"PRId64", size=%d", msg->timestamp, msg->size);
    
    bool is_sequence_header = SrsFlvCodec::video_is_sequence_header(msg->payload, msg->size);
    
    // whether consumer should drop for the duplicated sequence header.
    bool drop_for_reduce = false;
    if (is_sequence_header && cache_sh_video && _srs_config->get_reduce_sequence_header(_req->vhost)) {
        if (cache_sh_video->size == msg->size) {
            drop_for_reduce = srs_bytes_equals(cache_sh_video->payload, msg->payload, msg->size);
            srs_warn("drop for reduce sh video, size=%d", msg->size);
        }
    }
    
    // cache the sequence header if h264
    // donot cache the sequence header to gop_cache, return here.
    if (is_sequence_header) {
        srs_freep(cache_sh_video);
        cache_sh_video = msg->copy();
        
        // parse detail audio codec
        SrsAvcAacCodec codec;
        
        // user can disable the sps parse to workaround when parse sps failed.
        // @see https://github.com/ossrs/srs/issues/474
        codec.avc_parse_sps = _srs_config->get_parse_sps(_req->vhost);
        
        SrsCodecSample sample;
        if ((ret = codec.video_avc_demux(msg->payload, msg->size, &sample)) != ERROR_SUCCESS) {
            srs_error("source codec demux video failed. ret=%d", ret);
            return ret;
        }
        
        // when got video stream info.
        SrsStatistic* stat = SrsStatistic::instance();
        if ((ret = stat->on_video_info(_req, SrsCodecVideoAVC, codec.avc_profile, codec.avc_level)) != ERROR_SUCCESS) {
            return ret;
        }
        
        srs_trace("%dB video sh,  codec(%d, profile=%s, level=%s, %dx%d, %dkbps, %dfps, %ds)",
            msg->size, codec.video_codec_id,
            srs_codec_avc_profile2str(codec.avc_profile).c_str(),
            srs_codec_avc_level2str(codec.avc_level).c_str(), codec.width, codec.height,
            codec.video_data_rate / 1000, codec.frame_rate, codec.duration);
    }
    
#ifdef SRS_AUTO_HLS
    if ((ret = hls->on_video(msg, is_sequence_header)) != ERROR_SUCCESS) {
        // apply the error strategy for hls.
        // @see https://github.com/ossrs/srs/issues/264
        std::string hls_error_strategy = _srs_config->get_hls_on_error(_req->vhost);
        if (srs_config_hls_is_on_error_ignore(hls_error_strategy)) {
            srs_warn("hls process video message failed, ignore and disable hls. ret=%d", ret);
            
            // unpublish, ignore ret.
            hls->on_unpublish();
            
            // ignore.
            ret = ERROR_SUCCESS;
        } else if (srs_config_hls_is_on_error_continue(hls_error_strategy)) {
            if (srs_hls_can_continue(ret, cache_sh_video, msg)) {
                ret = ERROR_SUCCESS;
            } else {
                srs_warn("hls continue video failed. ret=%d", ret);
                return ret;
            }
        } else {
            srs_warn("hls disconnect publisher for video error. ret=%d", ret);
            return ret;
        }
    }
#endif
    
#ifdef SRS_AUTO_DVR
    if ((ret = dvr->on_video(msg)) != ERROR_SUCCESS) {
        srs_warn("dvr process video message failed, ignore and disable dvr. ret=%d", ret);
        
        // unpublish, ignore ret.
        dvr->on_unpublish();
        
        // ignore.
        ret = ERROR_SUCCESS;
    }
#endif

#ifdef SRS_AUTO_HDS
    if ((ret = hds->on_video(msg)) != ERROR_SUCCESS) {
        srs_warn("hds process video message failed, ignore and disable dvr. ret=%d", ret);
        
        // unpublish, ignore ret.
        hds->on_unpublish();
        // ignore.
        ret = ERROR_SUCCESS;
    }
#endif
    
    // copy to all consumer
    if (!drop_for_reduce) {
        for (int i = 0; i < (int)consumers.size(); i++) {
            SrsConsumer* consumer = consumers.at(i);
            if ((ret = consumer->enqueue(msg, atc, jitter_algorithm)) != ERROR_SUCCESS) {
                srs_error("dispatch the video failed. ret=%d", ret);
                return ret;
            }
        }
        srs_info("dispatch video success.");
    }

    // copy to all forwarders.
    if (!forwarders.empty()) {
        std::vector<SrsForwarder*>::iterator it;
        for (it = forwarders.begin(); it != forwarders.end(); ++it) {
            SrsForwarder* forwarder = *it;
            if ((ret = forwarder->on_video(msg)) != ERROR_SUCCESS) {
                srs_error("forwarder process video message failed. ret=%d", ret);
                return ret;
            }
        }
    }
    
    // when sequence header, donot push to gop cache and adjust the timestamp.
    if (is_sequence_header) {
        return ret;
    }

    // cache the last gop packets
    if ((ret = gop_cache->cache(msg)) != ERROR_SUCCESS) {
        srs_error("gop cache msg failed. ret=%d", ret);
        return ret;
    }
    srs_verbose("cache gop success.");
    
    // if atc, update the sequence header to abs time.
    if (atc) {
        if (cache_sh_video) {
            cache_sh_video->timestamp = msg->timestamp;
        }
        if (cache_metadata) {
            cache_metadata->timestamp = msg->timestamp;
        }
    }
    
    return ret;
}

以上进行了缓存h264 sequence header，hls分发，客户端消费者分发，forwarders推流等等。
这里主要看一下消费者分发：

// copy to all consumer
    if (!drop_for_reduce) {
        for (int i = 0; i < (int)consumers.size(); i++) {
            SrsConsumer* consumer = consumers.at(i);
            if ((ret = consumer->enqueue(msg, atc, jitter_algorithm)) != ERROR_SUCCESS) {
                srs_error("dispatch the video failed. ret=%d", ret);
                return ret;
            }
        }
        srs_info("dispatch video success.");
    }

int SrsConsumer::enqueue(SrsSharedPtrMessage* shared_msg, bool atc, SrsRtmpJitterAlgorithm ag)
{
    int ret = ERROR_SUCCESS;
    
  //这儿的copy操作只是增加引用计数，没有实际的内存拷贝。
    SrsSharedPtrMessage* msg = shared_msg->copy();

    if (!atc) {
        if ((ret = jitter->correct(msg, ag)) != ERROR_SUCCESS) {
            srs_freep(msg);
            return ret;
        }
    }
    
    if ((ret = queue->enqueue(msg, NULL)) != ERROR_SUCCESS) {
        return ret;
    }
    
#ifdef SRS_PERF_QUEUE_COND_WAIT
    srs_verbose("enqueue msg, time=%"PRId64", size=%d, duration=%d, waiting=%d, min_msg=%d", 
        msg->timestamp, msg->size, queue->duration(), mw_waiting, mw_min_msgs);
        
    // fire the mw when msgs is enough.
    if (mw_waiting) {
        int duration_ms = queue->duration();
        bool match_min_msgs = queue->size() > mw_min_msgs;
        
        // For ATC, maybe the SH timestamp bigger than A/V packet,
        // when encoder republish or overflow.
        // @see https://github.com/ossrs/srs/pull/749
        if (atc && duration_ms < 0) {
            st_cond_signal(mw_wait);
            mw_waiting = false;
            return ret;
        }
        
        // when duration ok, signal to flush.
        if (match_min_msgs && duration_ms > mw_duration) {
            st_cond_signal(mw_wait);
            mw_waiting = false;
            return ret;
        }
    }
#endif
    
    return ret;
}

每个SrsConsumer消费者拥有独立的SrsMessageQueue* queue队列。内部队列实现实际上是std::multimap<int64_t, SrsSharedPtrMessage*> msgs。
SrsMessageQueue有数量大小限制，当队列满的时候删除丢弃旧的messages：

队列大小限制queue_size设置为配置文件中的"queue_length"。如果没设置则默认#define SRS_PERF_PLAY_QUEUE 30。
queue_size_ms = (int)(queue_size * 1000);

int SrsMessageQueue::enqueue(SrsSharedPtrMessage* msg, bool* is_overflow)
{
    int ret = ERROR_SUCCESS;
    
    if (msg->is_av()) {
        if (av_start_time == -1) {
            av_start_time = msg->timestamp;
        }
        
        av_end_time = msg->timestamp;
    }
    
    msgs.push_back(msg);

    while (av_end_time - av_start_time > queue_size_ms) {
        // notice the caller queue already overflow and shrinked.
        if (is_overflow) {
            *is_overflow = true;
        }
        
        shrink();
    }
    
    return ret;
}

void SrsMessageQueue::shrink()
{
    SrsSharedPtrMessage* video_sh = NULL;
    SrsSharedPtrMessage* audio_sh = NULL;
    int msgs_size = (int)msgs.size();
    
    // remove all msg
    // igone the sequence header
    for (int i = 0; i < (int)msgs.size(); i++) {
        SrsSharedPtrMessage* msg = msgs.at(i);

        if (msg->is_video() && SrsFlvCodec::video_is_sequence_header(msg->payload, msg->size)) {
            srs_freep(video_sh);
            video_sh = msg;
            continue;
        }
        else if (msg->is_audio() && SrsFlvCodec::audio_is_sequence_header(msg->payload, msg->size)) {
            srs_freep(audio_sh);
            audio_sh = msg;
            continue;
        }

        srs_freep(msg);
    }
    msgs.clear();  

    // update av_start_time
    av_start_time = av_end_time;
    //push_back secquence header and update timestamp
    if (video_sh) {
        video_sh->timestamp = av_end_time;
        msgs.push_back(video_sh);
    }
    if (audio_sh) {
        audio_sh->timestamp = av_end_time;
        msgs.push_back(audio_sh);
    }
    
    if (_ignore_shrink) {
        srs_info("shrink the cache queue, size=%d, removed=%d, max=%.2f", 
            (int)msgs.size(), msgs_size - (int)msgs.size(), queue_size_ms / 1000.0);
    } else {
        srs_trace("shrink the cache queue, size=%d, removed=%d, max=%.2f", 
            (int)msgs.size(), msgs_size - (int)msgs.size(), queue_size_ms / 1000.0);
    }
}

保存最近的sequence_header，然后清除其他messages。

AVC sequence header
AAC sequence header
这两个header非常重要，是客户端解码的必需部分，所以不能删除。

这个丢包策略没有根据整个GOP进行丢包，而是直接丢掉除sequence_header的包，有可能会造成客户端花屏。

总结

客户端Rtmp推流到服务器，服务器将消息缓存到各个客户端消费者自己的队列中，数据使用引用计数没有内存拷贝操作。过期数据将被清除。
客户端消费者是SrsRtmpConnPlay类型，消费者播放流的流程在下一篇文章中介绍。