Chinaunix首页 | 论坛 | 博客
  • 博客访问: 3929124
  • 博文数量: 93
  • 博客积分: 3189
  • 博客等级: 中校
  • 技术积分: 4229
  • 用 户 组: 普通用户
  • 注册时间: 2009-02-02 13:29
个人简介

出没于杭州和青岛的程序猿一枚,对内核略懂一二

文章分类

全部博文(93)

文章存档

2016年(2)

2015年(3)

2014年(11)

2013年(29)

2012年(16)

2011年(5)

2010年(5)

2009年(22)

分类: LINUX

2012-10-12 08:18:16

关于TCP connect 返回错误99,可以能大家都会遇到,这里就分析一下这个错误的真正含义:
基于内核2.6.32
应用层调用connect,对应的系统调用的套接口实现是inet_stream_connect,对应tcp协议对应的传输层接口是tcp_v4_connect, 这里就从这个函数作为入口,该函数定义在net/ipv4/tcp_ipv4.c文件中

点击(此处)折叠或打开

  1. int tcp_v4_connect(struct sock *sk, struct sockaddr *uaddr, int addr_len)
  2. {
  3.     struct inet_sock *inet = inet_sk(sk);
  4.     struct tcp_sock *tp = tcp_sk(sk);
  5.     struct sockaddr_in *usin = (struct sockaddr_in *)uaddr;
  6.     struct rtable *rt;
  7.     __be32 daddr, nexthop;
  8.     int tmp;
  9.     int err;

  10.     if (addr_len < sizeof(struct sockaddr_in))
  11.         return -EINVAL;

  12.     if (usin->sin_family != AF_INET)
  13.         return -EAFNOSUPPORT;

  14.     nexthop = daddr = usin->sin_addr.s_addr;
  15.     if (inet->opt && inet->opt->srr) {
  16.         if (!daddr)
  17.             return -EINVAL;
  18.         nexthop = inet->opt->faddr;
  19.     }

  20.     tmp = ip_route_connect(&rt, nexthop, inet->saddr,
  21.              RT_CONN_FLAGS(sk), sk->sk_bound_dev_if,
  22.              IPPROTO_TCP,
  23.              inet->sport, usin->sin_port, sk, 1);
  24.     if (tmp < 0) {
  25.         if (tmp == -ENETUNREACH)
  26.             IP_INC_STATS_BH(sock_net(sk), IPSTATS_MIB_OUTNOROUTES);
  27.         return tmp;
  28.     }

  29.     if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
  30.         ip_rt_put(rt);
  31.         return -ENETUNREACH;
  32.     }

  33.     if (!inet->opt || !inet->opt->srr)
  34.         daddr = rt->rt_dst;

  35.     if (!inet->saddr)
  36.         inet->saddr = rt->rt_src;
  37.     inet->rcv_saddr = inet->saddr;

  38.     if (tp->rx_opt.ts_recent_stamp && inet->daddr != daddr) {
  39.         /* Reset inherited state */
  40.         tp->rx_opt.ts_recent     = 0;
  41.         tp->rx_opt.ts_recent_stamp = 0;
  42.         tp->write_seq         = 0;
  43.     }

  44.     if (tcp_death_row.sysctl_tw_recycle &&
  45.      !tp->rx_opt.ts_recent_stamp && rt->rt_dst == daddr) {
  46.         struct inet_peer *peer = rt_get_peer(rt);
  47.         /*
  48.          * VJ's idea. We save last timestamp seen from
  49.          * the destination in peer table, when entering state
  50.          * TIME-WAIT * and initialize rx_opt.ts_recent from it,
  51.          * when trying new connection.
  52.          */
  53.         if (peer != NULL &&
  54.          peer->tcp_ts_stamp + TCP_PAWS_MSL >= get_seconds()) {
  55.             tp->rx_opt.ts_recent_stamp = peer->tcp_ts_stamp;
  56.             tp->rx_opt.ts_recent = peer->tcp_ts;
  57.         }
  58.     }

  59.     inet->dport = usin->sin_port;
  60.     inet->daddr = daddr;

  61.     inet_csk(sk)->icsk_ext_hdr_len = 0;
  62.     if (inet->opt)
  63.         inet_csk(sk)->icsk_ext_hdr_len = inet->opt->optlen;

  64.     tp->rx_opt.mss_clamp = 536;

  65.     /* Socket identity is still unknown (sport may be zero).
  66.      * However we set state to SYN-SENT and not releasing socket
  67.      * lock select source port, enter ourselves into the hash tables and
  68.      * complete initialization after this.
  69.      */
  70.     tcp_set_state(sk, TCP_SYN_SENT);
  71.     err = inet_hash_connect(&tcp_death_row, sk);
  72.     if (err)
  73.         goto failure;

  74.     err = ip_route_newports(&rt, IPPROTO_TCP,
  75.                 inet->sport, inet->dport, sk);
  76.     if (err)
  77.         goto failure;

  78.     /* OK, now commit destination to socket. */
  79.     sk->sk_gso_type = SKB_GSO_TCPV4;
  80.     sk_setup_caps(sk, &rt->u.dst);

  81.     if (!tp->write_seq)
  82.         tp->write_seq = secure_tcp_sequence_number(inet->saddr,
  83.                              inet->daddr,
  84.                              inet->sport,
  85.                              usin->sin_port);

  86.     inet->id = tp->write_seq ^ jiffies;

  87.     err = tcp_connect(sk);
  88.     rt = NULL;
  89.     if (err)
  90.         goto failure;

  91.     return 0;

  92. failure:
  93.     /*
  94.      * This unhashes the socket and releases the local port,
  95.      * if necessary.
  96.      */
  97.     tcp_set_state(sk, TCP_CLOSE);
  98.     ip_rt_put(rt);
  99.     sk->sk_route_caps = 0;
  100.     inet->dport = 0;
  101.     return err;
  102. }
这里进入第84行代码的函数inet_hash_connect 查找一个本地可用端口与服务器建立连接

点击(此处)折叠或打开

  1. int inet_hash_connect(struct inet_timewait_death_row *death_row,
  2.          struct sock *sk)
  3. {
  4.     return __inet_hash_connect(death_row, sk, inet_sk_port_offset(sk),
  5.             __inet_check_established, __inet_hash_nolisten);
  6. }
该函数实际上调用_inet_hash_connect

点击(此处)折叠或打开

  1. int __inet_hash_connect(struct inet_timewait_death_row *death_row,
  2.         struct sock *sk, u32 port_offset,
  3.         int (*check_established)(struct inet_timewait_death_row *,
  4.             struct sock *, __u16, struct inet_timewait_sock **),
  5.         void (*hash)(struct sock *sk))
  6. {
  7.     struct inet_hashinfo *hinfo = death_row->hashinfo;
  8.     const unsigned short snum = inet_sk(sk)->num;
  9.     struct inet_bind_hashbucket *head;
  10.     struct inet_bind_bucket *tb;
  11.     int ret;
  12.     struct net *net = sock_net(sk);

  13.     if (!snum) {
  14.         int i, remaining, low, high, port;
  15.         static u32 hint;
  16.         u32 offset = hint + port_offset;
  17.         struct hlist_node *node;
  18.         struct inet_timewait_sock *tw = NULL;

  19.         inet_get_local_port_range(&low, &high);
  20.         remaining = (high - low) + 1;

  21.         local_bh_disable();
  22.         for (i = 1; i <= remaining; i++) {
  23.             port = low + (i + offset) % remaining;
  24.             head = &hinfo->bhash[inet_bhashfn(net, port,
  25.                     hinfo->bhash_size)];
  26.             spin_lock(&head->lock);

  27.             /* Does not bother with rcv_saddr checks,
  28.              * because the established check is already
  29.              * unique enough.
  30.              */
  31.             inet_bind_bucket_for_each(tb, node, &head->chain) {
  32.                 if (ib_net(tb) == net && tb->port == port) {
  33.                     if (tb->fastreuse >= 0)
  34.                         goto next_port;
  35.                     WARN_ON(hlist_empty(&tb->owners));
  36.                     if (!check_established(death_row, sk,
  37.                                 port, &tw))
  38.                         goto ok;
  39.                     goto next_port;
  40.                 }
  41.             }

  42.             tb = inet_bind_bucket_create(hinfo->bind_bucket_cachep,
  43.                     net, head, port);
  44.             if (!tb) {
  45.                 spin_unlock(&head->lock);
  46.                 break;
  47.             }
  48.             tb->fastreuse = -1;
  49.             goto ok;

  50.         next_port:
  51.             spin_unlock(&head->lock);
  52.         }
  53.         local_bh_enable();

  54.         return -EADDRNOTAVAIL;

  55. ok:
  56.         hint += i;

  57.         /* Head lock still held and bh's disabled */
  58.         inet_bind_hash(sk, tb, port);
  59.         if (sk_unhashed(sk)) {
  60.             inet_sk(sk)->sport = htons(port);
  61.             hash(sk);
  62.         }
  63.         spin_unlock(&head->lock);

  64.         if (tw) {
  65.             inet_twsk_deschedule(tw, death_row);
  66.             inet_twsk_put(tw);
  67.         }

  68.         ret = 0;
  69.         goto out;
  70.     }

  71.     head = &hinfo->bhash[inet_bhashfn(net, snum, hinfo->bhash_size)];
  72.     tb = inet_csk(sk)->icsk_bind_hash;
  73.     spin_lock_bh(&head->lock);
  74.     if (sk_head(&tb->owners) == sk && !sk->sk_bind_node.next) {
  75.         hash(sk);
  76.         spin_unlock_bh(&head->lock);
  77.         return 0;
  78.     } else {
  79.         spin_unlock(&head->lock);
  80.         /* No definite answer... Walk to established hash table */
  81.         ret = check_established(death_row, sk, snum, NULL);
  82. out:
  83.         local_bh_enable();
  84.         return ret;
  85.     }
  86. }
注意第61行代码,这里返回了EADDRNOTAVAIL错误。
这里分析一下这个函数实现:
调用inet_get_local_port_range(&low, &high) 获取可用的端口列表,这个值就是/proc/sys/net/ipv4/ip_local_port_range 中的值。

点击(此处)折叠或打开

  1. void inet_get_local_port_range(int *low, int *high)
  2. {
  3.     unsigned seq;
  4.     do {
  5.         seq = read_seqbegin(&sysctl_local_ports.lock);

  6.         *low = sysctl_local_ports.range[0];
  7.         *high = sysctl_local_ports.range[1];
  8.     } while (read_seqretry(&sysctl_local_ports.lock, seq));
  9. }
然后内核在这个范围内选择一个可用的端口作为本地端口去connect服务器,如果没有可用的端口可用,比如这个范围内的端口都处于如下状态中的一种:
1. bind使用的端口
2. 端口处于非TIME_WAIT状态
3. 端口处于TIME_WAIT状态,但是没有启用tcp_tw_reuse
那么就会返回EADDRNOTAVAIL错误。

一般情况下,出现这个错误应该是代码设计的问题,如果确定代码没有问题,那么根据上面的原则,可用使用如下方法解决问题:
1. 增大可选端口的范围,修改/proc/sys/net/ipv4/ip_local_port_range的值。
2. 开启tcp_tw_reuse,允许使用TIME_WAIT状态的端口。
阅读(18803) | 评论(2) | 转发(0) |
给主人留下些什么吧!~~

firocu2013-04-02 17:32:32

firocu:你好,请问tcp_v4_connect()的if这句怎么理解:
 if (tcp_death_row.sysctl_tw_recycle &&  !tp->rx_opt.ts_recent_stamp && rt->rt_dst == daddr) 

拜谢了。

今日再次求教:如可以请帮我指明:
 !tp->rx_opt.ts_recent_stamp  和  rt->rt_dst == daddr 这两个表达式所表达的含义,分别是什么?理解不能。

回复 | 举报

firocu2013-03-17 12:32:12

你好,请问tcp_v4_connect()的if这句怎么理解:
 if (tcp_death_row.sysctl_tw_recycle &&  !tp->rx_opt.ts_recent_stamp && rt->rt_dst == daddr) 

拜谢了。