What's Mean "TCP: Treason uncloaked! Peer..shrinks window..Repaired."

« The Promise-无极 | Blog首页 | MT Scode 插件错误解决 »

最近在系统的dmesg信息中,经常看到这样的信息:

TCP: Treason uncloaked! Peer 203.168.193.2:2682/80 shrinks window 
                                     2217805154:2217892754 Repaired.
TCP: Treason uncloaked! Peer 203.168.193.2:10721/80 shrinks window ... Repaired.
TCP: Treason uncloaked! Peer 203.168.193.2:10721/80 shrinks window ... Repaired.
TCP: Treason uncloaked! Peer 203.168.193.2:44016/80 shrinks window ... Repaired.
TCP: Treason uncloaked! Peer 203.168.193.2:44016/80 shrinks window ... Repaired.
TCP: Treason uncloaked! Peer 203.168.193.2:55074/80 shrinks window ... Repaired.

操作系统版本为Redhat Linux AD3:

[root@eygle /]# cat /etc/redhat-release 
Red Hat Enterprise Linux AS release 3 (Taroon)
[root@eygle /]# uname -a
Linux eygle 2.4.21-4.EL #1 Fri Oct 3 18:13:58 EDT 2003 i686 i686 i386 GNU/Linux

以前一直没时间关注,今天抽点时间来研究一下.

首先在网上看到这样的一段解释:

The remote host decided to shrink the TCP window size without negotiating such with your Linux box. The message is of the informational level, meaning Linux doesn't like what it is seeing but will cope with it and carry on.

大致意思是说:这是一段一般性提示信息,说明远程主机在未经Linux主机"同意"就收缩了TCP window size,虽然Linux主机不喜欢这类举动,但是仍将继续处理这类请求.

这是一种解释,也就是这类信息可能并非危险.

另外一段解释来自Debian的邮件列表,首先回复者引用了一段源代码说明错误的来源:

>>From /usr/src/linux/net/ipv4/tcp_timer.c:

        if (tp->snd_wnd == 0 && !sk->dead &&
            !((1<<sk->state)&(TCPF_SYN_SENT|TCPF_SYN_RECV))) {
                /* Receiver dastardly shrinks window. Our retransmits
                 * become zero probes, but we should not timeout this
                 * connection. If the socket is an orphan, time it out,
                 * we cannot allow such beasts to hang infinitely.
                 */
#ifdef TCP_DEBUG
                if (net_ratelimit())
                        printk(KERN_DEBUG "TCP: Treason uncloaked! Peer
%u.%u.%u.%u:%u/%u shrinks window %u:%u. Repaired.\n",
                               NIPQUAD(sk->daddr), htons(sk->dport), sk->num,
                               tp->snd_una, tp->snd_nxt);
#endif

具体的解释是:

So it appears that someone is running some sort of "tar-pit" system that is
designed to keep sockets in a bad state and run you out of kernel memory.

I suspect that this ties in with the spam blocking things we recently
discussed. Maybe you should tell your ISP that they are to blame for such
actions being done to you and that they should "give you face" (I think that
was the term you used) by closing their open relays.

作者认为这可能和tar-pit攻击相关.并且建议联系ISP提供解决方案.

还有如下的解释:

The reason Linux is printing such messages is because your client guy is shrinking the TCP Window to 0, and the server has something to retransmit. There is something seriously wrong with your client's stack. Which Stack/OS are you using on he client side, and which browser?

That could explain your browser showing some html tags as the server fails to send the whole page across and based on what browser you are using it is failing to parse it out.

意思是,这类错误也可能是客户端Stack错误引起的.

另一段同样引用了源代码的解释似乎直指了问题的本质:

> Treason uncloaked! Peer [IP address]:515/1022 shrinks window
> 3957222360:3957222379. Repaired.

> Our researches so far indicate the problem may be a buggy TCP stack
> in the client, that is in the DP301P+. But we still do not know
> exactly what caused the problem, nor how to prevent it happening
> again.

That comes from the kernel tcp code below. Looks like the DLink has
returned information yielding a transmit window smaller than it
previously did; specifically it returned a window of zero plus an ack
of up to byte 3957222360, thus indicating that it can accept nothing
after that byte. Previously it had sent some ack+wnd values
indicating that it would accept up to byte 3957222379.

The Linux side is now supposed to send a packet every now and then
forever until the returned window is nonzero. It does.

However, the dlink is apparently not responding in a timely manner.
Any response would either open the window or update the rcv timestamp
such that the thing will retransmit forever. It may be responding
very slowly, or just not responding at all.

The kernel prints the message after it expected but did not see a
response to the probe packet it sent to check for a nonzero window.
The kernel implements exponential backoff retransmissions until it
hasn't seen any response in 2m, then it will bail and close the
connection. This is reasonable. It's unclear from your report if the
connections are failing outright or just sometimes having to
retransmit a probe against a peer that shrank the window.