藍色情懷

(This article was written with the Winsock programmer in mind, but the information in it can be used by Unix programmers, as well as administrators and technicians.)

TCP is a simple protocol in a certain sense: you send data, it delivers it. Because it was engineered for reliability in networks of uncertain quality, it works around a lot of problems without bothering the end user. But partially because of this reliability, TCP exhibits behaviors that surprise those that don't truly understand the protocol. This tutorial will introduce you to the most important of these issues, but it's really the tip of the iceberg. For the submerged part, see TCP/IP Illustrated; the state/transition diagram below comes from volume 2 of that series. It happens to be printed in Volume 1 of the series as well, and in Stevens' Unix Network Programming, volume 1. You can also get that diagram in Postscript format off the web; see the Miscellaneous Resources section of the FAQ for a pointer.

In this tutorial, the term "packet" refers to data wrapped in a TCP frame. As a TCP packet goes through the network from one peer to another, the packet may be split into separate hardware frames, or the data within multiple TCP frames may be coalesced. This is as opposed to the "datagram" sense of the term "packet," for an inviolable block of data that is unchanged from sender to receiver.

TCP Control Bits

When a TCP implementation decides to send a packet of data to the remote peer, it first wraps the data with 20-plus bytes of data called the "header". Headers are an essential part of network protocols, because they enable the participants in the network make decisions regarding the data flowing over it. Every protocol adds headers (and sometimes trailers) to your data. We won't discuss the TCP and IP headers in detail here, as that's better left to books like W. Richard Stevens'.

Within the header is a field that I will call the "control bits," for lack of a better term. The bits that interest us here are called SYN, ACK, FIN and RST, for "synchronize," "acknowledge," "finish," and "reset," respectively. These bits are set in TCP packets for the sole benefit of the remote peer's network stack — that is, they are the machinery under the hood that most people never have occasion to examine.

The State/Transition Diagram

Below is the state/transition diagram for the TCP protocol. The states are in round-ended boxes, and the transitions are the labelled arrows. The transitions show how how your program can make TCP move from one state to another. It also shows how the remote peer can make your stack change TCP states, and how you can recognize these changes at the application level. Note that transition labels come from the names of BSD sockets functions; although there are differences in the Winsock API, the effects are the same at this level. (I apologize for the so-so readability of the text in this image, but it's already too big at 20K, so I'm unwilling to make it any bigger — if you want a pretty, readable diagram, get the Postscript file and print your own copy.)

Understanding this diagram is really one of the keys to understanding TCP, so let's go through a few exercises. But first, you need to know about the netstat tool. This tool comes with all Microsoft TCP/IP stacks, and probably others as well. It is modelled after a Unix tool of the same name, with virtually the same output. (The differences between each version of this tool are slight enough that once you learn to use one, the rest are trivial to pick up.)

The netstat tool is usually run from the command line, often with the -n flag to make it faster. (-n suppresses the DNS name lookups, displaying the raw address and port numbers instead.) Another useful flag is the -a flag, which shows "all" entries, including listeners. (This only works correctly in Windows NT derivatives.) It is also very helpful to use this tool in combination with a "grep" tool — I recommend installing Cygwin, which includes GNU grep.

Microsoft netstats output four columns: the protocol (e.g. TCP or UDP), the local address/port combination, the remote address/port combination, and the current state of that connection. The first three columns are self-explanatory, and are often collectively called the "connection 5-tuple," which uniquely describes a given TCP or UDP connection. The last column corresponds directly to the states in the diagram above.

A Micro-FAQ

Now for those exercises I promised:

1. Problem: From the default CLOSED state, how does a client program normally get to the ESTABLISHED state?
   
   Solution: The client calls the connect() function (or similar), which causes TCP to send an empty packet with the SYN control bit set (SYN_SENT). The remote peer's stack sees this "synchronize" request, and sends back an empty packet with the SYN and ACK bits set (i.e. "I acknowledge your synchronize request"). When the client receives the SYN/ACK packet, it sends back an ACK packet, and reports a successful connection to the client program.
   
   
2. Problem: What is the normal TCP shutdown sequence?
   
   Solution: The important thing to understand is that TCP is a bidirectional protocol. So, the connection is shut down in two identical stages, one for each "direction". One peer sends a packet with the FIN bit set, which the other end ACKnowledges; when the other end is also finished sending data, it sends out a FIN packet, which the other end ACKs, closing the connection.
   
   
3. Problem: What is the significance of the RST bit?
   
   Solution: This is an abnormal close, also called "slamming the connection shut." It happens under several circumstances, but none of the common ones are documented in the Stevens diagram. Two of these cases you can cause from Winsock: the first method is to set SO_LINGER to 0 with setsockopt() and then call closesocket(). The second method is to call shutdown() with how equal to 2, optionally followed by a closesocket() call.
   
   From the Winsock client level, the two other common RST occurrences are "connection refused" and "remote peer terminated connection." The first happens when you try to connect to a port that isn't open on a remote machine. The second happens as a result of the remote peer using one of the two RST-forcing methods above; alternately, the application could have crashed, and the peer's stack sent out a RST for its connection. Another way this can happen is the remote peer catastrophically crashed, and then after the remote machine came back up, your program sent it a packet which the stack rightfully had no way of delivering, so it replied with a RST packet, because the connection's 5-tuple is now invalid.
   
   Generally speaking, RST signals a problem of some kind: either something bad happened to the connection, or there's a bug somewhere. For example, some firewalls improperly use the RST bit to signal a closed connection. The solution, of course, is to replace the firewall product. :)
   
   
4. Problem: Netstat shows lots of sockets in the TIME_WAIT state. What's wrong?
   
   Solution: Nothing's wrong. TIME_WAIT is absolutely normal. If you go through the state-transition diagram above, you see that every socket that gets closed normally goes through this state after it's closed. The TIME_WAIT state is a safety mechanism, to catch stray packets for that connection after the connection is "officially" closed. Since the maximum time that such stray packets can exist is 2 times the "maximum segment lifetime" (the time for a packet to go from your machine to the remote peer and for the response to come back), the TIME_WAIT state lasts for 2 * MSL. The only tricky bit is, there is no easy way to estimate MSL on the fly, so stacks traditionally hard-code a value for it, from 15 to 60 seconds. Thus, TIME_WAIT usually lasts 30-120 seconds.
   
   
5. Problem: My sockets keep getting into a FIN_WAIT_x state. What's wrong?
   
   Solution: Either your program or the remote peer is not closing the socket properly. If you walk through the state-transition diagram above, you can see that FIN_WAIT_1 usually happens when the local program calls shutdown() with the "how" parameter set to 1 or SD_SEND, but the remote peer doesn't respond. Likewise FIN_WAIT_2 happens when the remote peer shuts down its sending half of the connection, and your program doesn't respond. Since FIN_WAIT states often last up to 10 minutes, it's well worth the effort to fix the problem that's causing these FIN_WAIT states. (The exact length of the state depends on the stack and the circumstances that got it into that state.)
   
   
6. Problem: Often my calls to bind() fail when I try to re-bind to a port that I was just using. What's wrong?
   
   Solution: The socket is probably in one of the FIN_WAIT states or in the TIME_WAIT state. If it's a FIN_WAIT problem and you can't fix it, or if it's the normal TIME_WAIT state, the best thing to do is to redesign your program so that it doesn't need to keep re-binding. For example, a server program generally keeps its listener socket alive so that it doesn't have to keep re-binding it to the port; if you closesocket() the listener for some reason after each successful connection, your listener socket will go into the TIME_WAIT state for somewhere between 30 and 120 seconds, during which you won't be able to re-bind to that port. However, if you find that re-binding is absolutely necessary, setting the SOREUSEADDR option with setsockopt() will usually get you around the problem.
   
   

Tools for TCPers

Below is a small batch file I find helpful in dealing with TCP state issues, which I call "showwait." Basically, it shows you the current WAIT states every second until you hit Ctrl-C. I have a similar script on my Unix machines as well.

		@echo off
		:loop
		netstat -na |grep WAIT
		delay 1
		goto loop

This script depends on a 4DOS feature called "delay". If you don't use this shell, get an implementation of the "sleep" command, which does the same thing. The Cygwin toolset, mentioned above, includes one. (Are you maybe getting the impression that I'm a closet Unixhead? Oh, noooo.... :) )

There's one problem with this tool: it only catches problems with "WAIT" in their name. Less common states like LAST_ACK and SYN_RCVD won't be seen by this script. SYN_RCVD in particular signals serious problems if it stays around for a prolonged amount of time, because it indicates that a remote machine sent your machine a SYN packet, your machine ACK'd it, and the remote machine has failed to ACK your SYN/ACK. Since this exchange typically only takes from a few tens to a few hundreds of milliseconds, a persistent SYN_RCVD indicates a badly-written network stack, or a very "crashy" computer. If you see many of these states at once, it may mean you're under a "SYN attack", one of several Denial of Service attacks that are going around these days. At that point, it's time to break out the network sniffer and start some detective work.

Conclusion

The techniques and information in this article reflect the basic mental tools that your organization needs to develop, even if it's just appointing a single "networking guru" who will master this material, and become a resource for the other developers in the company. This knowledge is very widely useful; for example, it can make reading sniffer dumps less painful and more productive. Also, these techniques can reasonably be applied by technicians working with knowledgeable users over the phone to gather information about failures in your program that otherwise would get logged as random failures.

Copyright © 1998-2000 by Warren Young. All rights reserved.