ip_route_output_key函数分析（1）

http://blog.chinaunix.net/uid-24673811-id-1754229.html

上面的文章读了net/ipv4/route.c 中的ip_route_input函数，是协议站对收到报文的路由查找函数。继续阅读一下协议栈发包的时候路由查找的调用函数ip_roue_output_key。

int ip_route_output_key(struct net *net, struct rtable **rp, struct flowi *flp)
{
    return ip_route_output_flow(net, rp, flp, NULL, 0);
}
只是一个函数封装，真正的处理函数是ip_route_output_flow.
int ip_route_output_flow(struct net *net, struct rtable **rp, struct flowi *flp,
             struct sock *sk, int flags)
{
    int err;

    /*路由查找*/
    if ((err = __ip_route_output_key(net, rp, flp)) != 0)
        return err;

    /*IPSec 的处理代码*/
    if (flp->proto) {
        if (!flp->fl4_src)
            flp->fl4_src = (*rp)->rt_src;
        if (!flp->fl4_dst)
            flp->fl4_dst = (*rp)->rt_dst;
        err = __xfrm_lookup(net, (struct dst_entry **)rp, flp, sk,
                    flags ? XFRM_LOOKUP_WAIT : 0);
        if (err == -EREMOTE)
            err = ipv4_dst_blackhole(net, rp, flp);

        return err;
    }

    return 0;
}

很眼熟的__ip_route_output_key函数。
int __ip_route_output_key(struct net *net, struct rtable **rp,
              const struct flowi *flp)
{
    unsigned hash;
    struct rtable *rth;

    if (!rt_caching(net))
        goto slow_output;

    /*类似于ip_route_input,先在cache中查找路由， 找到就返回*/
    hash = rt_hash(flp->fl4_dst, flp->fl4_src, flp->oif, rt_genid(net));

    rcu_read_lock_bh();
    for (rth = rcu_dereference(rt_hash_table[hash].chain); rth;
        rth = rcu_dereference(rth->u.dst.rt_next)) {
        if (rth->fl.fl4_dst == flp->fl4_dst &&
            rth->fl.fl4_src == flp->fl4_src &&
            rth->fl.iif == 0 &&
            rth->fl.oif == flp->oif &&
            rth->fl.mark == flp->mark &&
            !((rth->fl.fl4_tos ^ flp->fl4_tos) &
                (IPTOS_RT_MASK | RTO_ONLINK)) &&
            net_eq(dev_net(rth->u.dst.dev), net) &&
            !rt_is_expired(rth)) {
            dst_use(&rth->u.dst, jiffies);
            RT_CACHE_STAT_INC(out_hit);
            rcu_read_unlock_bh();
            *rp = rth;
            return 0;
        }
        RT_CACHE_STAT_INC(out_hlist_search);
    }
    rcu_read_unlock_bh();

/*不支持cache 或在cache中没找到相应的路由信息，在路由表中查找*/
slow_output:
    return ip_route_output_slow(net, rp, flp);
}

继续探险
/*
 * Major route resolver routine.
 */

static int ip_route_output_slow(struct net *net, struct rtable **rp,
                const struct flowi *oldflp)
{
    u32 tos    = RT_FL_TOS(oldflp); /*获取tos和当前的RTO_ONLINK（？）标志*/
    struct flowi fl = { .nl_u = { .ip4_u =
                      { .daddr = oldflp->fl4_dst,
                    .saddr = oldflp->fl4_src,
                    .tos = tos & IPTOS_RT_MASK,
                    .scope = ((tos & RTO_ONLINK) ? /*根据这个标志，得出路由的scope*/
                          RT_SCOPE_LINK :
                          RT_SCOPE_UNIVERSE),
                      } },
                .mark = oldflp->mark,
                .iif = net->loopback_dev->ifindex, /*设备号为lo的设备号?*/
                .oif = oldflp->oif };
    struct fib_result res;
    unsigned flags = 0;
    struct net_device *dev_out = NULL;
    int free_res = 0;
    int err;


    res.fi        = NULL;
#ifdef CONFIG_IP_MULTIPLE_TABLES

    res.r        = NULL;
#endif

    /*先是对源地址， 发包接口号和目的地址进行判断分类处理。下面的每一个红色跳转就是一种情况*/ 
    if (oldflp->fl4_src) { /*源*/
        err = -EINVAL;
        if (ipv4_is_multicast(oldflp->fl4_src) ||
            ipv4_is_lbcast(oldflp->fl4_src) ||
            ipv4_is_zeronet(oldflp->fl4_src))
            goto out;
       
        /*上面是对报文源地址的合理性检查，源地址是多播，广播或0地址时，返回错误*/

        /* I removed check for oif == dev_out->oif here.
           It was wrong for two reasons:
           我在这里删去检查oif == dev_out->oif是否成立，因为有两个原因说明这个检查时错误的：
           1. ip_dev_find(net, saddr) can return wrong iface, if saddr
              is assigned to multiple interfaces
              如果源地址是一个多播接口的地址，函数ip_dev_find(net, saddr)可能返回错误的设备接口。
           2. Moreover, we are allowed to send packets with saddr  
              of another iface. --ANK
              而且可以用另外设备接口的源地址发送报文
         */

        if (oldflp->oif == 0
            && (ipv4_is_multicast(oldflp->fl4_dst) ||
            oldflp->fl4_dst == htonl(0xFFFFFFFF))) { /*发包接口为lo，目的地址是广播或多播时查找发包设备，ip_dev_find返回与所给定的源地址相等的第一个设备*/
            /* It is equivalent to inet_addr_type(saddr) == RTN_LOCAL */
            dev_out = ip_dev_find(net, oldflp->fl4_src);
            if (dev_out == NULL)
                goto out;

            /* Special hack: user can direct multicasts
               and limited broadcast via necessary interface
               without fiddling with IP_MULTICAST_IF or IP_PKTINFO.
               This hack is not just for fun, it allows
               vic,vat and friends to work.
               They bind socket to loopback, set ttl to zero
               and expect that it will work.
               From the viewpoint of routing cache they are broken,
               because we are not allowed to build multicast path
               with loopback source addr (look, routing cache
               cannot know, that ttl is zero, so that packet
               will not leave this host and route is valid).
               Luckily, this hack is good workaround.
             */

            /*当报文初始化的出接口为lo接口源地址不为空目的地址是多播或广播地址时，找到源地址所对应的接口重新为出接口赋值， 然后创建cache路由项*/
            fl.oif = dev_out->ifindex;
            goto make_route;
        }

        /*?????*/
        if (!(oldflp->flags & FLOWI_FLAG_ANYSRC)) {
            /* It is equivalent to inet_addr_type(saddr) == RTN_LOCAL */
            dev_out = ip_dev_find(net, oldflp->fl4_src);
            if (dev_out == NULL)
                goto out;
            dev_put(dev_out);
            dev_out = NULL;
        }
    }


    if (oldflp->oif) {/*发包设备不为空*/
        /*检测出接口是否存在*/
        dev_out = dev_get_by_index(net, oldflp->oif);
        err = -ENODEV;
        if (dev_out == NULL)
            goto out;

        /* RACE: Check return value of inet_select_addr instead. */
        /*看设备是否是多地址*/
        if (__in_dev_get_rtnl(dev_out) == NULL) {
            dev_put(dev_out);
            goto out;    /* Wrong error code */
        }

        /*当目的地址是本地多播地址或广播地址，并且报文源地址为空时，找出出接口设备上IP地址scope小于RT_SCOPE_LINK的地址，并赋值，然后往cache中添加路由表项*/
        if (ipv4_is_local_multicast(oldflp->fl4_dst) ||
            oldflp->fl4_dst == htonl(0xFFFFFFFF)) {
            if (!fl.fl4_src)
                fl.fl4_src = inet_select_addr(dev_out, 0,
                                  RT_SCOPE_LINK);
            goto make_route;
        }
        /*目的地址是单播地址或空，源地址为空，那就选一个小于特定scope的IP地址*/
        if (!fl.fl4_src) {
            if (ipv4_is_multicast(oldflp->fl4_dst))
                fl.fl4_src = inet_select_addr(dev_out, 0,
                                  fl.fl4_scope);
            else if (!oldflp->fl4_dst)
                fl.fl4_src = inet_select_addr(dev_out, 0,
                                  RT_SCOPE_HOST);
        }
    }

    if (!fl.fl4_dst) {/*目的地址为空*/
        fl.fl4_dst = fl.fl4_src;
        if (!fl.fl4_dst)
            fl.fl4_dst = fl.fl4_src = htonl(INADDR_LOOPBACK);/*目的和源地址都是空，则赋值为lo接口地址*/
        if (dev_out)
            dev_put(dev_out);
        dev_out = net->loopback_dev;
        dev_hold(dev_out);
        fl.oif = net->loopback_dev->ifindex;
        res.type = RTN_LOCAL;
        flags |= RTCF_LOCAL;
        /*为发给本机的报文添加cache路由*/
        goto make_route;
    }

    /*一种情况是源地址目的地址不为空，目的地址为空，出接口为lo*/
    /*还有其他几种情况，就是目的地址和出接口必须对应*/
    if (fib_lookup(net, &fl, &res)) {
        res.fi = NULL;
        if (oldflp->oif) {
            /* Apparently, routing tables are wrong. Assume,
               that the destination is on link.

               WHY? DW.
               Because we are allowed to send to iface
               even if it has NO routes and NO assigned
               addresses. When oif is specified, routing
               tables are looked up with only one purpose:
               to catch if destination is gatewayed, rather than
               direct. Moreover, if MSG_DONTROUTE is set,
               we send packet, ignoring both routing tables
               and ifaddr state. --ANK


               We could make it even if oif is unknown,
               likely IPv6, but we do not.
             */

            if (fl.fl4_src == 0)
                fl.fl4_src = inet_select_addr(dev_out, 0,
                                  RT_SCOPE_LINK);
            res.type = RTN_UNICAST;
            /*没有查到路由，并且出接口不为lo*/
            goto make_route;
        }
        if (dev_out)
            dev_put(dev_out);
        err = -ENETUNREACH;
        goto out;
    }
    /*找到路由*/
    free_res = 1;

    /*路由指向本地*/
    if (res.type == RTN_LOCAL) {
        if (!fl.fl4_src)
            fl.fl4_src = fl.fl4_dst;
        if (dev_out)
            dev_put(dev_out);
        dev_out = net->loopback_dev;
        dev_hold(dev_out);
        fl.oif = dev_out->ifindex;
        if (res.fi)
            fib_info_put(res.fi);
        res.fi = NULL;
        flags |= RTCF_LOCAL;
        goto make_route;
    }

/*是否支持多路径路由*/
#ifdef CONFIG_IP_ROUTE_MULTIPATH
    if (res.fi->fib_nhs > 1 && fl.oif == 0)
        fib_select_multipath(&fl, &res);
    else
#endif
    if (!res.prefixlen && res.type == RTN_UNICAST && !fl.oif)
        fib_select_default(net, &fl, &res);

    if (!fl.fl4_src)
        fl.fl4_src = FIB_RES_PREFSRC(res);

    if (dev_out)
        dev_put(dev_out);
    dev_out = FIB_RES_DEV(res);
    dev_hold(dev_out);
    fl.oif = dev_out->ifindex;

/*往cache中添加相应的路由项*/
make_route:
    err = ip_mkroute_output(rp, &res, &fl, oldflp, dev_out, flags);


    if (free_res)
        fib_res_put(&res);
    if (dev_out)
        dev_put(dev_out);
out:    return err;
}