TIME_WAIT实验解析精讲

时间 : 17-11-12 栏目 : linux精讲 作者 : 老薛 评论 : 1 点击 : 455 次

TIME_WAIT是排查TCP连接问题经常遇到的一种TCP状态。首先我们来看TIME_WAIT状态是如何产生的。TCP关闭连接的状态变化图如下:

可以看到,主动关闭连接的一端在发送最后一个ACK后进入TIME_WAIT状态。TIME_WAIT状态会持续2倍的MSL(Maximum Segment Lifetime),MSL是指一个TCP分段在网络上存在的最大时间,因而TIME_WAIT状态也被称为2MSL状态。

TIME_WAIT状态的作用主要有两个:

  • 可靠的关闭连接。假设主动关闭连接的一端发送的最后一个ACK分段在网络中丢包或被延迟了,被动关闭的一端因收不到ACK,会重新发送FIN包。这时如果没有TIME_WAIT状态直接处于CLOSED状态,就会直接响应RST而不会响应ACK。极端情况下,ACK到达对端需要一个MSL,对端重发的FIN到达需要一个MSL,当2个MSL之后仍未收到重传的FIN,则认为对端已经收到了ACK。
  • 防止上一次连接中的分段延迟到达后影响新连接。TCP连接由五元组(协议,源IP,源端口,目的IP,目的端口, 因而协议固定是TCP,也可以说是四元组)唯一标识。假设没有TIME_WAIT状态,一个连接关闭后,可能使用相同的五元组的新连接被建立,这时若原连接上的TCP分段因为网络延时刚刚到达,且它的序列号刚好在新连接的接收窗口,则会令新连接接收的数据混乱。尽管每次建立连接使用的序列号都是随机产生的,但是序列号的长度只有32位,在高速网络上可能很快出现序列号循环。TIME_WAIT状态持续2MSL后,原连接的数据包都已经在网络上消失,不会再干扰新连接。

现在我们知道TIME_WAIT是TCP的正常状态。然而,当机器上存在大量TIME_WAIT状态的socket时, 不主动连接(客户端)和被动连接(服务器端)的场景下对服务器有不同的影响。

首先来看主动连接场景。

主动连接端会占用本地端口,TIME_WAIT状态socket占用的本地端口过多的话,会导致本地端口不足,TCP连接不能成功建立。

可以调整参数net.ipv4.ip_local_port_range来增加本地端口的选择范围,但这样效果有限。还可以启用net.ipv4.tcp_tw_reuse参数来重用TIME_WAIT状态的socket。

内核文档对于该参数的描述如下:

tcp_tw_reuse - BOOLEAN
    Allow to reuse TIME-WAIT sockets for new connections when it is safe from protocol viewpoint. Default value is 0.
    It should not be changed without advice/request of technical experts.

下面来看该参数具体如何实现。

本地端口的选取过程如图:


net/ipv4/inet_hashtables.c中的函数__inet_check_established用来检查已经在使用的端口是否能够重用,对于TIME_WAIT状态的socket会调用tcp_twsk_unique来判断。

tcp_twsk_unique的源码如下:

int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp)
{
    const struct tcp_timewait_sock *tcptw = tcp_twsk(sktw);
    struct tcp_sock *tp = tcp_sk(sk);
    /* With PAWS, it is safe from the viewpoint
       of data integrity. Even without PAWS it is safe provided sequence
       spaces do not overlap i.e. at data rates <= 80Mbit/sec.
       Actually, the idea is close to VJ's one, only timestamp cache is
       held not per host, but per port pair and TW bucket is used as state
       holder.
       If TW bucket has been already destroyed we fall back to VJ's scheme
       and use initial timestamp retrieved from peer table.
     */
    if (tcptw->tw_ts_recent_stamp &&
        (twp == NULL || (sysctl_tcp_tw_reuse &&
                 get_seconds() - tcptw->tw_ts_recent_stamp > 1))) {
        tp->write_seq = tcptw->tw_snd_nxt + 65535 + 2;
        if (tp->write_seq == 0)
            tp->write_seq = 1;
        tp->rx_opt.ts_recent       = tcptw->tw_ts_recent;
        tp->rx_opt.ts_recent_stamp = tcptw->tw_ts_recent_stamp;
        sock_hold(sktw);
        return 1;
    }
    return 0;
}

只有当net.ipv4.tcp_timestamps选项启用时,tcptw->tw_ts_recent_stamp才会为真, net.ipv4.tcp_tw_reuse启用时,sysctl_tcp_tw_reuse才会为真。从代码可以看到,当tcp_timestamps和tcp_tw_reuse两个选项开启时,若该socket距离上次收到数据包已经超过1秒,则可以重用该端口。

Linux的TCP实现将TIME_WAIT状态持续时间硬编码在了源码中, 设置为60秒,在include/net/tcp.h:

#define TCP_TIMEWAIT_LEN (60*HZ) /* how long to wait to destroy TIME-WAIT   * state, about 60 seconds */

这个值并不能通过sysctl来修改,如果需要修改该时间值,则需要修改内核代码重新编译。

下面来验证tcp_tw_reuse的效果: 我们准备一个服务器,运行NGINX并且不主动关闭连接,客户端使用curl来访问服务器并主动关闭连接来产生TIME_WAIT状态的socket。curl连接成功时返回值为0,失败时返回值为7。我们以此判断连接是否成功。

为了使实验简单,将本地端口范围修改为只有三个端口可用:

sysctl -w net.ipv4.ip_local_port_range="32768 32770"
确保tcp_tw_reuse选项为关闭状态:

[root@localhost ~]# sysctl -a |grep reuse
net.ipv4.tcp_tw_reuse = 0
编写实验脚本,脚本逻辑为先执行3次curl访问服务器,占满全部本地端口,2秒之后再执行一次curl访问。脚本内容如下:

for i in `seq 1 3`
do
  echo $i
  date
  curl http://10.95.48.11/ -o /dev/null -s
  echo "RETURN: " $?
  ss -atn |grep TIME
done

sleep 2

echo "4"
date
curl http://10.95.48.11/ -o /dev/null -s
echo "RETURN: " $?
在客户端执行脚本, 可以看到第4次请求会失败:

[root@localhost ~]# sh t.sh

Thu Nov  9 15:51:57 CST 2017
RETURN:  0
TIME-WAIT  0      0            192.168.10.131:32769          10.95.48.11:80

Thu Nov  9 15:51:57 CST 2017
RETURN:  0
TIME-WAIT  0      0            192.168.10.131:32770          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32769          10.95.48.11:80

Thu Nov  9 15:51:57 CST 2017
RETURN:  0
TIME-WAIT  0      0            192.168.10.131:32768          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32770          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32769          10.95.48.11:80

Thu Nov  9 15:51:59 CST 2017
RETURN:  7
接着我们将tcp_tw_reuse和tcp_timestamps选项启用:

[root@localhost ~]# sysctl -w net.ipv4.tcp_tw_reuse=1
net.ipv4.tcp_tw_reuse = 1
[root@localhost ~]# sysctl -w net.ipv4.tcp_timestamps=1
net.ipv4.tcp_timestamps = 1
再次执行shell脚本, 第4次请求会成功返回0:

[root@localhost ~]# sh t.sh

Thu Nov  9 15:50:10 CST 2017
RETURN:  0
TIME-WAIT  0      0            192.168.10.131:32768          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32770          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32769          10.95.48.11:80

Thu Nov  9 15:50:10 CST 2017
RETURN:  0
TIME-WAIT  0      0            192.168.10.131:32768          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32770          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32769          10.95.48.11:80

Thu Nov  9 15:50:10 CST 2017
RETURN:  0
TIME-WAIT  0      0            192.168.10.131:32768          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32770          10.95.48.11:80
TIME-WAIT  0      0            192.168.10.131:32769          10.95.48.11:80

Thu Nov  9 15:50:12 CST 2017
RETURN:  0
开启tcp_tw_reuse选项可以使主动连接的一端在主动关闭的场景下每秒能建立6万多个连接。如果每秒要建立更多的连接呢?可以使用tcp_tw_recycle选项。

内核文档对于该参数的描述如下:

tcp_tw_recycle - BOOLEAN
    Enable fast recycling TIME-WAIT sockets. Default value is 0.
    It should not be changed without advice/request of technical
    experts.
相较于tcp_tw_reuse只在需要时重用TIME_WAIT状态socket, tcp_tw_recycle处理更激进,它会快速回收TIME_WAIT状态的socket。

内核代码中有定时器来调用tcp_time_wait函数来处理TIME_WAIT状态的socket,函数源码如下:

/*
 * Move a socket to time-wait or dead fin-wait-2 state.
 */
void tcp_time_wait(struct sock *sk, int state, int timeo)
{
    struct inet_timewait_sock *tw = NULL;
    const struct inet_connection_sock *icsk = inet_csk(sk);
    const struct tcp_sock *tp = tcp_sk(sk);
    int recycle_ok = 0;
    if (tcp_death_row.sysctl_tw_recycle && tp->rx_opt.ts_recent_stamp)
        recycle_ok = icsk->icsk_af_ops->remember_stamp(sk);
    if (tcp_death_row.tw_count < tcp_death_row.sysctl_max_tw_buckets)
        tw = inet_twsk_alloc(sk, state);
    if (tw != NULL) {
        struct tcp_timewait_sock *tcptw = tcp_twsk((struct sock *)tw);
        const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
        tw->tw_rcv_wscale   = tp->rx_opt.rcv_wscale;
        tcptw->tw_rcv_nxt   = tp->rcv_nxt;
        tcptw->tw_snd_nxt   = tp->snd_nxt;
        tcptw->tw_rcv_wnd   = tcp_receive_window(tp);
        tcptw->tw_ts_recent = tp->rx_opt.ts_recent;
        tcptw->tw_ts_recent_stamp = tp->rx_opt.ts_recent_stamp;
#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
        if (tw->tw_family == PF_INET6) {
            struct ipv6_pinfo *np = inet6_sk(sk);
            struct inet6_timewait_sock *tw6;
            tw->tw_ipv6_offset = inet6_tw_offset(sk->sk_prot);
            tw6 = inet6_twsk((struct sock *)tw);
            ipv6_addr_copy(&tw6->tw_v6_daddr, &np->daddr);
            ipv6_addr_copy(&tw6->tw_v6_rcv_saddr, &np->rcv_saddr);
            tw->tw_ipv6only = np->ipv6only;
        }
#endif
#ifdef CONFIG_TCP_MD5SIG
        /*
         * The timewait bucket does not have the key DB from the
         * sock structure. We just make a quick copy of the
         * md5 key being used (if indeed we are using one)
         * so the timewait ack generating code has the key.
         */
        do {
            struct tcp_md5sig_key *key;
            memset(tcptw->tw_md5_key, 0, sizeof(tcptw->tw_md5_key));
            tcptw->tw_md5_keylen = 0;
            key = tp->af_specific->md5_lookup(sk, sk);
            if (key != NULL) {
                memcpy(&tcptw->tw_md5_key, key->key, key->keylen);
                tcptw->tw_md5_keylen = key->keylen;
                if (tcp_alloc_md5sig_pool(sk) == NULL)
                    BUG();
            }
        } while (0);
#endif
        /* Linkage updates. */
        __inet_twsk_hashdance(tw, sk, &tcp_hashinfo);
        /* Get the TIME_WAIT timeout firing. */
        if (timeo < rto)
            timeo = rto;
        if (recycle_ok) {
            tw->tw_timeout = rto;
        } else {
            tw->tw_timeout = TCP_TIMEWAIT_LEN;
            if (state == TCP_TIME_WAIT)
                timeo = TCP_TIMEWAIT_LEN;
        }
        inet_twsk_schedule(tw, &tcp_death_row, timeo,
                   TCP_TIMEWAIT_LEN);
        inet_twsk_put(tw);
    } else {
        /* Sorry, if we're out of memory, just CLOSE this
         * socket up.  We've got bigger problems than
         * non-graceful socket closings.
         */
        LIMIT_NETDEBUG(KERN_INFO "TCP: time wait bucket table overflow\n");
    }
    tcp_update_metrics(sk);
    tcp_done(sk);
}
从代码上可以看到只有当tcp_timestamps和tcp_tw_recycle都开启时,才会快速回收。而根据代码:

const int rto = (icsk->icsk_rto << 2) - (icsk->icsk_rto >> 1);
可以看到回收的超时时间为3.5 * RTO, RTO是由TCP分段中timestamp选项计算得到的,一般场景下这个时间在几百毫秒左右。

从上面的tcp_time_wait源码也可以看出, 当TIME_WAIT状态的socket数量超过tcp_max_tw_buckets选项指定的数量值时,会直接关闭socket,进入CLOSED状态,内核日志中会报错:

TCP: time wait bucket table overflow
的错误。若把tcp_max_tw_buckets选项设置为0,则可以直接跳过TIME_WAIT状态。

然而,tcp_tw_recycle选项在NAT环境使用有一些隐患,下面来分析一下。

协议栈收到syn包时会调用到函数tcp_v4_conn_request, 该函数部分源码如下:

    /* VJ's idea. We save last timestamp seen
     * from the destination in peer table, when entering
     * state TIME-WAIT, and check against it before
     * accepting new connection request.
     *
     * If "isn" is not zero, this request hit alive
     * timewait bucket, so that all the necessary checks
     * are made in the function processing timewait state.
     */
    if (tmp_opt.saw_tstamp &&
        tcp_death_row.sysctl_tw_recycle &&
        (dst = inet_csk_route_req(sk, req)) != NULL &&
        (peer = rt_get_peer((struct rtable *)dst)) != NULL &&
        peer->v4daddr == saddr) {
        if (get_seconds() < peer->tcp_ts_stamp + TCP_PAWS_MSL &&
            (s32)(peer->tcp_ts - req->ts_recent) >
                        TCP_PAWS_WINDOW) {
            NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSPASSIVEREJECTED);
            goto drop_and_release;
        }
    }
其中TCP_PAWS_MSL和TCP_PAWS_WINDOW分别为60和1:

#define TCP_PAWS_MSL    60      /* Per-host timestamps are invalidated
                     * after this time. It should be equal
                     * (or greater than) TCP_TIMEWAIT_LEN
                     * to provide reliability equal to one
                     * provided by timewait state.
                     */
#define TCP_PAWS_WINDOW 1       /* Replay window for per-host
                     * timestamps. It must be less than
                     * minimal timewait lifetime.
                     */
从代码上我们可以看到,当开启tcp_timestamps和tcp_tw_recycle选项时,60秒内来自同一源IP主机的TCP分段的时间戳必须递增,否则该分段会被直接丢弃。

假如多个客户端从NAT环境访问服务器,服务器端看到的对端IP是一样的,但是TCP分段的时间戳会不一样。当时间戳较大的客户端连接成功后的60秒内,时间戳较小的客户端再次连接服务器,syn包会被服务器直接丢弃,导致连接失败。

我们来通过实验验证。首先基于libvirt构造NAT环境。 编辑实验网络描述文件n1.xml:

<network>
  <name>n1</name>
  <bridge name="n1"/>
  <forward mode="nat" />
  <ip address="192.168.10.1" netmask="255.255.255.0">
    <dhcp>
      <range start="192.168.10.2" end="192.168.10.254"/>
    </dhcp>
  </ip>
</network>
启用实验网络:

virsh net-define n1.xml
virsh net-start n1
再启动两台虚拟机做为客户端,将其接口设置在实验网络中:

<interface type="network">
  <source network="n1" />
</interface>
虚拟机启动后,在两台机器上抓包确认出两台机器TCP时间戳的大小。

将服务器端的NGINX配置keepalive_timeout设为0,令服务器端主动关闭连接,重新加载NGINX配置生效。

将服务器的tcp_timestamps和tcp_tw_recycle选项启动:

[root@bogon conf]# sysctl -w net.ipv4.tcp_tw_recycle=1
net.ipv4.tcp_tw_recycle = 1
[root@bogon conf]# sysctl -w net.ipv4.tcp_timestamps=1
net.ipv4.tcp_timestamps = 1
首先在时间戳较小的客户端上执行curl, 连接成功:

[root@localhost ~]# curl http://10.95.48.11 -o /dev/null -s -v
* About to connect() to 10.95.48.11 port 80 (#0)
*   Trying 10.95.48.11... connected
* Connected to 10.95.48.11 (10.95.48.11) port 80 (#0)
> GET / HTTP/1.1
> User-Agent: curl/7.19.7 (x86_64-redhat-linux-gnu) libcurl/7.19.7 NSS/3.21 Basic ECC zlib/1.2.3 libidn/1.18 libssh2/1.4.2
> Host: 10.95.48.11
> Accept: */*
>
< HTTP/1.1 200 OK
< Server: openresty/1.11.2.5
< Date: Thu, 09 Nov 2017 11:07:55 GMT
< Content-Type: text/html
< Content-Length: 562
< Last-Modified: Thu, 21 Sep 2017 20:02:30 GMT
< Connection: close
< ETag: "59c41ad6-232"
< Accept-Ranges: bytes
<
{ [data not shown]
* Closing connection #0
接着在时间戳较大的客户端上执行curl,连接也成功:

[root@localhost ~]# curl http://10.95.48.11 -o /dev/null -s -v
* About to connect() to 10.95.48.11 port 80 (#0)
*   Trying 10.95.48.11... connected
* Connected to 10.95.48.11 (10.95.48.11) port 80 (#0)
> GET / HTTP/1.1
> User-Agent: curl/7.19.7 (x86_64-redhat-linux-gnu) libcurl/7.19.7 NSS/3.21 Basic ECC zlib/1.2.3 libidn/1.18 libssh2/1.4.2
> Host: 10.95.48.11
> Accept: */*
>
< HTTP/1.1 200 OK
< Server: openresty/1.11.2.5
< Date: Thu, 09 Nov 2017 11:08:07 GMT
< Content-Type: text/html
< Content-Length: 562
< Last-Modified: Thu, 21 Sep 2017 20:02:30 GMT
< Connection: close
< ETag: "59c41ad6-232"
< Accept-Ranges: bytes
<
{ [data not shown]
* Closing connection #0
而这时立即再次从时间戳较小机器上执行,可以发现连接失败了:

[root@localhost ~]# curl http://10.95.48.11 -o /dev/null -s -v
* About to connect() to 10.95.48.11 port 80 (#0)
*   Trying 10.95.48.11... Connection timed out
* couldn't connect to host
* Closing connection #0
一分钟之后重试, 则连接成功:

[root@localhost ~]# curl http://10.95.48.11 -o /dev/null -s -v
* About to connect() to 10.95.48.11 port 80 (#0)
*   Trying 10.95.48.11... connected
* Connected to 10.95.48.11 (10.95.48.11) port 80 (#0)
> GET / HTTP/1.1
> User-Agent: curl/7.19.7 (x86_64-redhat-linux-gnu) libcurl/7.19.7 NSS/3.21 Basic ECC zlib/1.2.3 libidn/1.18 libssh2/1.4.2
> Host: 10.95.48.11
> Accept: */*
>
< HTTP/1.1 200 OK
< Server: openresty/1.11.2.5
< Date: Thu, 09 Nov 2017 11:11:51 GMT
< Content-Type: text/html
< Content-Length: 562
< Last-Modified: Thu, 21 Sep 2017 20:02:30 GMT
< Connection: close
< ETag: "59c41ad6-232"
< Accept-Ranges: bytes
<
{ [data not shown]
* Closing connection #0
我们从服务器端的抓包结果可以看到失败请求的TCP时间戳小于之前的时间戳:

在被动连接的情况下,若产生了大量的TIME_WAIT状态,并没有特别大的性能影响,只是会多消耗一些内存与端口链表查找的CPU资源。

了解了这些关于TIME_WAIT相关的知识背景就可以根据应用场景来灵活调整相应的参数设置了。

本文标签 ,

除非注明,文章均为( 老薛 )原创,转载请保留链接: http://www.bdkyr.com/operation/2459.html

TIME_WAIT实验解析精讲:目前有1 条留言

  1. 沙发
    天天
    Post: 2017-11-15 上午10:52

    专业

发表评论

2 + 5 = ?


博主微信号,很高兴为您提供帮助

随便看看

1