dpvs源码分析(续二)

原创

榴莲其实还可以

发布于 2018-08-10 15:31:49

2.1K00

代码可运行

文章被收录于专栏：后台开发+音视频+ffmpeg

运行总次数：0

代码可运行

前言

在上一篇<dpvs源码分析(续)>中，我们以tcp为例，讲到了连接的建立，同时也提到了full-nat，snat这些术语。在该篇中，我们再来讲讲连接建立的过程。

建立连接

static int tcp_conn_sched(struct dp_vs_proto *proto, 
                          const struct dp_vs_iphdr *iph,
                          struct rte_mbuf *mbuf, 
                          struct dp_vs_conn **conn,
                          int *verdict)
{
    struct tcphdr *th, _tcph;
    struct dp_vs_service *svc;
    assert(proto && iph && mbuf && conn && verdict);

    th = mbuf_header_pointer(mbuf, iph->len, sizeof(_tcph), &_tcph);
    if (unlikely(!th)) {
        *verdict = INET_DROP;
        return EDPVS_INVPKT;
    }

    /* Syn-proxy step 2 logic: receive client's 3-handshacke ack packet */
    /* When synproxy disabled, only SYN packets can arrive here.
     * So don't judge SYNPROXY flag here! If SYNPROXY flag judged, and syn_proxy
     * got disbled and keepalived reloaded, SYN packets for RS may never be sent. */
    if (dp_vs_synproxy_ack_rcv(iph->af, mbuf, th, proto, conn, iph, verdict) == 0) {
        /* Attention: First ACK packet is also stored in conn->ack_mbuf */
        return EDPVS_PKTSTOLEN;
    }

   ..................
    svc = dp_vs_service_lookup(iph->af, iph->proto, 
                               &iph->daddr, th->dest, 0, mbuf, NULL);
    if (!svc) {
        *verdict = INET_ACCEPT;
        return EDPVS_NOSERV;
    }
 
    ..............................

    *conn = dp_vs_schedule(svc, iph, mbuf, false);
    if (!*conn) {
        dp_vs_service_put(svc);
        *verdict = INET_DROP;
        return EDPVS_RESOURCE;
    }

    dp_vs_service_put(svc);

    return EDPVS_OK;
}

代码依然是有删减的，删除的部分用"..................."替代了。但是主体逻辑依然存在。该函数中dp_vs_synproxy_ack_rcv

是tcp握手代理，也就是说client先和部署了dpvs的机器进行3次握手，完了之后，dpvs机器才会和真正的服务器RS去握手，此处细节就不深究了。然后第二个调用是dp_vs_service_lookup，这个函数干嘛呢？看名字大概就能猜出来了——通过传递的参数，协议，ip, port查找service。service是dest的集合，dest就是真正的服务器。找了了svc之后，就可以准备建立连接了。

2，dp_vs_schedule

/* select an RS by service's scheduler and create a connection */
struct dp_vs_conn *dp_vs_schedule(struct dp_vs_service *svc, 
                                  const struct dp_vs_iphdr *iph,
                                  struct rte_mbuf *mbuf,
                                  bool is_synproxy_on)
{
    uint16_t _ports[2], *ports; /* sport, dport */
    struct dp_vs_dest *dest;
    struct dp_vs_conn *conn;
    struct dp_vs_conn_param param;
    struct sockaddr_in daddr, saddr;
    int err;

    assert(svc && iph && mbuf);

    ports = mbuf_header_pointer(mbuf, iph->len, sizeof(_ports), _ports);
    if (!ports)
        return NULL;
        
    /* persistent service */
    if (svc->flags & DP_VS_SVC_F_PERSISTENT)
        return dp_vs_sched_persist(svc, iph,  mbuf, is_synproxy_on);

    dest = svc->scheduler->schedule(svc, mbuf);
    if (!dest) {
        RTE_LOG(WARNING, IPVS, "%s: no dest found.\n", __func__);
        return NULL;
    }
        
    if (dest->fwdmode == DPVS_FWD_MODE_SNAT) {
        if (unlikely(iph->proto == IPPROTO_ICMP)) {
            struct icmphdr *ich, _icmph;
            ich = mbuf_header_pointer(mbuf, iph->len, sizeof(_icmph), &_icmph);
            if (!ich)
                return NULL;

            ports = _ports;
            _ports[0] = icmp4_id(ich);
            _ports[1] = ich->type << 8 | ich->code;

            /* ID may confict for diff host,
             * need we use ID pool ? */
            dp_vs_conn_fill_param(iph->af, iph->proto,
                                  &iph->daddr, &dest->addr,
                                  ports[1], ports[0],
                                  0, &param);
        } else {
            /* we cannot inherit dest (host's src port),
             * that may confict for diff hosts,
             * and using dest->port is worse choice. */
            memset(&daddr, 0, sizeof(daddr));
            daddr.sin_family = AF_INET;
            daddr.sin_addr = iph->daddr.in;
            daddr.sin_port = ports[1];
            memset(&saddr, 0, sizeof(saddr));
            saddr.sin_family = AF_INET;
            saddr.sin_addr = dest->addr.in;
            saddr.sin_port = 0;

            err = sa_fetch(NULL, &daddr, &saddr);
            if (err != 0)
                return NULL;

            dp_vs_conn_fill_param(iph->af, iph->proto,
                                  &iph->daddr, &dest->addr,
                                  ports[1], saddr.sin_port,
                                  0, &param);
        }
    } else {
        if (unlikely(iph->proto == IPPROTO_ICMP)) {
            struct icmphdr *ich, _icmph;
            ich = mbuf_header_pointer(mbuf, iph->len, sizeof(_icmph), &_icmph);
            if (!ich)
                return NULL;

            ports = _ports;
            _ports[0] = icmp4_id(ich);
            _ports[1] = ich->type << 8 | ich->code;

            dp_vs_conn_fill_param(iph->af, iph->proto,
                                  &iph->saddr, &iph->daddr,
                                  ports[0], ports[1], 0, &param);
        } else {
            dp_vs_conn_fill_param(iph->af, iph->proto,
                                  &iph->saddr, &iph->daddr,
                                  ports[0], ports[1], 0, &param);
        }
    }

    conn = dp_vs_conn_new(mbuf, &param, dest,
            is_synproxy_on ? DPVS_CONN_F_SYNPROXY : 0);
    if (!conn) {
        if (dest->fwdmode == DPVS_FWD_MODE_SNAT && iph->proto != IPPROTO_ICMP)
            sa_release(NULL, &daddr, &saddr);
        return NULL;
    }

    dp_vs_stats_conn(conn);
    return conn;
}

恩，继续看代码，其实注释的也挺清楚的——通过service的scheduler函数，选择一个RS(real server), 然后新建一个连接。该函数首先调用了 dest = svc->scheduler->schedule(svc, mbuf); 这句话的意思也很简单，按照service的调度规则，从中选择一个dest(也就是一台RS)。svc的结构体叫struct dp_vs_service，其中有个成员struct dp_vs_scheduler *scheduler，看一下这个结构体：

struct dp_vs_scheduler {
    struct list_head    n_list;
    char                *name;
//    rte_atomic32_t      refcnt;

    struct dp_vs_dest *
        (*schedule)(struct dp_vs_service *svc, 
                    const struct rte_mbuf *mbuf);

    int (*init_service)(struct dp_vs_service *svc);
    int (*exit_service)(struct dp_vs_service *svc);
    int (*update_service)(struct dp_vs_service *svc);
} __rte_cache_aligned;

也就是说不同的dp_vs_scheduler 对象，是有不同的调度方式的。这里看起来是不是和前面两篇中的job，pkt_type有些相似呢？在dpvs中这个scheduler有3中：dp_vs_rr_scheduler,

static struct dp_vs_scheduler dp_vs_rr_scheduler = {
    .name = "rr",       /* name */
//    .refcnt = ATOMIC_INIT(0),
    .n_list = LIST_HEAD_INIT(dp_vs_rr_scheduler.n_list),
    .init_service = dp_vs_rr_init_svc,
    .update_service = dp_vs_rr_update_svc,
    .schedule = dp_vs_rr_schedule,
};

static struct dp_vs_scheduler dp_vs_wlc_scheduler = {
    .name = "wlc",
    .n_list = LIST_HEAD_INIT(dp_vs_wlc_scheduler.n_list),
    .schedule = dp_vs_wlc_schedule,
};

static struct dp_vs_scheduler dp_vs_wrr_scheduler = {
    .name = "wrr",
    .n_list = LIST_HEAD_INIT(dp_vs_wrr_scheduler.n_list),
    .init_service = dp_vs_wrr_init_svc,
    .exit_service = dp_vs_wrr_done_svc,
    .update_service = dp_vs_wrr_update_svc,
    .schedule = dp_vs_wrr_schedule,
};

这个注册调用是dpvs_init->dp_vs_sched_init，在dp_vs_sched_init中分别调用了dp_vs_rr_init, dp_vs_wrr_init以及dp_vs_wlc_init分别对这三个scheduler进行了注册。具体的调度策略我们就不详述了。之所以追溯这些代码，其实是想将dpvs的一些代码串联起来。一般在我们刚看dpvs的代码的时候，main函数就是一堆init，然后就是数据平面与控制平面线程启动。刚看起来可能会一脸懵逼，但是从数据平面线程看，往后一直追溯，就会发现，前面的初始化都会在后面有用到。

再回到dp_vs_schedule这个函数，当我们找到一个dest后，然后会有很多dp_vs_conn_fill_param这个函数的调用，dp_vs_conn_fill_param主要就是填充最后一个参数param。这个参数保存了源ip端口和目的ip端口信息。这个参数在dp_vs_conn_new会用到。

struct dp_vs_conn * dp_vs_conn_new(struct rte_mbuf *mbuf,
                                   struct dp_vs_conn_param *param,
                                   struct dp_vs_dest *dest, uint32_t flags)

dp_vs_conn_new这个函数比较长，就不都贴出来了，大体讲下其中的逻辑就行了。从这个函数的定义就可以看出，函数最后会构建了一个struct dp_vs_conn对象，这个结构体里面保存有连接的源和目的ip,port。此外这个函数中还会保存inboud和outbound对应的地址端口信息。以snat为例，client发包给dpvs，dpvs转发给RS，处理完成之后RS会发包到dpvs,再由dpvs转发到client，那么在dpvs这个层面是需要保存这个对应关系的.还需说明dp_vs_conn这个结构体中，有两个成员变量packet_xmit，packet_out_xmit函数指针，一个是转发到RS，一个是转发给client。这两个函数指针真正的赋值是在dp_vs_conn_new中调用的conn_bind_dest中完成的。不同的转发模式赋值不同。

结语

到此整个dpvs连接的就算完了，回到第二篇<dpvs源码分析(续)>最后的dp_vs_in函数，连接建立完了之后，如果是client发包过来，会调用xmit_inbound，xmit_inbound函数又会调用到dp_vs_conn::packet_xmit进行真正的转发。

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

nat

原创声明：本文系作者授权腾讯云开发者社区发表，未经许可，不得转载。

如有侵权，请联系 cloudcommunity@tencent.com 删除。

nat

登录后参与评论

0 条评论

热度