I have a C/C++ application set up as follows:
A non-blocking TCP server socket on a linux platform
A thread which writes a small packet (less than 20 bytes) to the socket at 1 Hz
The socket is configured with keepalive enabled and with: keepidle=5, keepintvl=5 and keepcnt=3
My intention is that the keepalive mechanism should detect a physical disconnection of the network link. However, when the link is cut, I do not see the zero-length packets which should be generated by the keepalive mechanism (I am using tcpdump to monitor traffic). My impression is that what happens is this: after the cable disconnection, the application keeps making send requests and the fact that there are pending send requests prevents the keepalive mechanism from being activated. Is this explanation valid?
In order to check my explanation, I have modified my test as follows:
A non-blocking TCP server socket on a linux platform
A cyclical thread which writes a small packet (about 100 bytes) to the socket with a period of 30 seconds
The socket is configured with keepalive enabled and with: keepidle=5, keepintvl=5 and keepcnt=2
In this case, if I cut the connection, the keepalive mechanism triggers within about 15-20 seconds (which is what I would expect).
On a related point, I would like to understand the exact semantics of tcp_keepidle. This is defined as: "The number of seconds a connection needs to be idle before TCP begins sending out keep-alive probes". What exactly does 'idle' means in this context? Does it simply mean that nothing is received and nothing is put on the network; or does it mean that nothing is received and no send requests are made to the socket?
Half-Established Connections
With a half-established connection I mean a connection for which the client's call to connect() returned successfully, but the servers call to accept() didn't. This can happen the following way: The client calls connect(), resulting in a SYN packet to the server. The server goes into state SYN-RECEIVED and sends a SYN-ACK packet to the client. This causes the client to reply with ACK, go into state ESTABLISHED and return from the connect() call. If the final ACK is lost (or ignored, due to a full accept queue at the server, which is probably the more likely scenario), the server is still in state SYN-RECEIVED and the accept() does not return. Due to timeouts associated with the SYN-RECEIVED state the SYN-ACK will be resend, allowing the client to resend the ACK. If the server is able to process the ACK eventually, it will go into state ESTABLISHED as well. Otherwise it will eventually reset the connection (i.e. send a RST to the client).
You can create this scenario by starting lots of connections on a single listen socket (if you do not adjust the backlog and tcp_max_syn_backlog). See this questions and this article for more details.
Experiments
I performed several experiments (with variations of this code) and observed some behaviour I cannot explain. All experiments where performed using Erlang's gen_tcp and a current Linux, but I strongly suspect that the answers are not specific to this setup, so I tried to keep it more general here.
connect() -> wait -> send() -> receive()
My starting point was to establish a connection from the client, wait between 1 to 5 seconds, send a "Ping" message to the server and wait for the reply. With this setup I observed that the receive() failed with the error closed when I had a half-established connection. There was never an error during the send() on a half-established connection. You can find a more detailed description of this setup here.
connect() -> long wait -> send()
To see, if I can get errors while sending data on a half-established connection I waited for 4 minutes before sending data. The 4 minutes should cover all timeouts and retries associated with the half-established connection. Sending data was still possible, i.e. send() returned without error.
connect() -> receive()
Next I tested what happens if I only call receive() with a very long timeout (5 minutes). My expectation was to get an closed error for the half-established connections, as in the original experiments. Alas, nothing happend, no error was thrown and the receive eventually timed out.
My questions
Is there a common name for what I call a half-established connection?
Why is the send() on a half-established connection successful?
Why does a receive() only fail if I send data first?
Any help, especially links to detailed explanations, are welcome.
From the client's point of view, the session is fully established, it sent SYN, got back SYN/ACK and sent ACK. It is only on the server side that you have a half-established state. (Even if it gets a repeated SYN/ACK from the server, it will just re-ACK because it's in the established state.)
The send on this session works fine because as far as the client is concerned, the session is established. The sent data does not have to be acknowledged by the far side in order to succeed (the send system call is finished when the data is copied into kernel buffers) but see below.
I believe here that the send actually is generating an error on the connection (probably a RST) because the receiving system cannot ACK data on a session it has not finished establishing. My guess is that any system call referencing the socket on the client side that happens after the send plus a short delay (i.e. when the RST has had a chance to come back) will result in an error.
The receive by itself never causes an error because the client side doesn't need to do anything (I mean TCP protocol-wise) for a receive; it's just idly waiting. But once you send some data, you've forced the server side's hand: it either has completed the session establishment (in which case it can accept the data) or it must send a reset (my guess here that it can't "hold" undelivered data on a session that isn't fully established).
I have a LPD server running on vxworks 6.3. The client application (over which I have no control) is sending me a LPQ query every tenth of a second. After 235 requests, the client receives a RST when trying to connect. After a time device will again accept some queries (about 300), until it again starts sending out RST.
I have confirmed that it is the TCP stack that is causing the RST. There are some things that I have noticed.
1) I can somewhat change the number of sockets that will accepted if I change the number of other applications that are running. For example, I freed up 4 sockets thereby changing the number accepted from 235 to 239.
2) If I send requests to lpr (port 515) and another port (say, port 80), the total number of connections that are accepted before the RST start happening stays constant at 235.
3) There are lots of sockets sitting TIME_WAIT.
4) I have a mock version of the client. If I slow the client down to one request every quarter second, the server doesn't reject the connections.
5) If I slow down the server's responses, I don't have any connections rejected.
So my theory is that there is some share resource (my top guess is total number of socket handles) that VxWorks can have consumed at a given time. I'm also guessing that this number tops out at 255.
Does anyone know how I can get VxWorks to accept more connections, and leave them in TIME_WAIT when closed? I have looked through the kernel configuration and changed all the values that looked remotely likely, but I have not been able change the number.
We know that we could set SO_LINGER but this not an acceptable solution. However, this does prevent the client connections from getting rejected. We have also tried changed the timeout value for SO_LINGER. This does not appear to be supported in VxWorks. It's either on or off.
Thanks!
Gail
To me it sounds like you are making a new connection for every LPQ query, and after the query is done you aren't closing the connection. In my opinion the correct thing to do is to accept one TCP connection and then use that to get all of the LPQ queries, however that may require mods to the client application. To avoid mods to the client, you should just close the TCP connection after each LPQ query.
Furthermore you can set the max number of FDs open in vxworks by adjusting the #define NUM_FILES config.h (or configall.h or one of those files), but that will just postpone an error if you have a FD leak, which you probably do.
I have a socket server which accepts multiple connections from various clients. I'm testing it on localhost with a client application which connects - sends data and closes connection 10 times every 10 ms. Some times the test client raises an error: Connection refused by the remote server or something similar.
Is this a normal behavior of the server application ?
10 connects every 10mS is one connection per millisecond, which seems a rather fast rate. Are these connection attempts being made in parallel? If so, perhaps you are filling up the server's listen() backlog-queue; IIRC clients who try to connect while the backlog-queue is full will get a connection-refused error.
To test that hypothesis, try passing in larger or smaller numbers as the second argument to listen() on your server, and see if that makes the connection-refused error occur more or less often.
I'm with Jeremy. You didn't mention the protocol, but I assume it's SOCK_STREAM. It will take longer than 10ms to do the tcp handshake on anything but the most local connection, eventually causing a backlog (and subsequent connection refused error) no matter how high you set your listen backlog to.
You'd be way ahead if you could keep the connection open, and not close it down during each of your computation cycles.
I think I understand the formal meaning of the option. In some legacy code I'm handling now, the option is used. The customer complains about RST as response to FIN from its side on connection close from its side.
I am not sure I can remove it safely, since I don't understand when it should be used.
Can you please give an example of when the option would be required?
For my suggestion, please read the last section: “When to use SO_LINGER with timeout 0”.
Before we come to that a little lecture about:
Normal TCP termination
TIME_WAIT
FIN, ACK and RST
Normal TCP termination
The normal TCP termination sequence looks like this (simplified):
We have two peers: A and B
A calls close()
A sends FIN to B
A goes into FIN_WAIT_1 state
B receives FIN
B sends ACK to A
B goes into CLOSE_WAIT state
A receives ACK
A goes into FIN_WAIT_2 state
B calls close()
B sends FIN to A
B goes into LAST_ACK state
A receives FIN
A sends ACK to B
A goes into TIME_WAIT state
B receives ACK
B goes to CLOSED state – i.e. is removed from the socket tables
TIME_WAIT
So the peer that initiates the termination – i.e. calls close() first – will end up in the TIME_WAIT state.
To understand why the TIME_WAIT state is our friend, please read section 2.7 in "UNIX Network Programming" third edition by Stevens et al (page 43).
However, it can be a problem with lots of sockets in TIME_WAIT state on a server as it could eventually prevent new connections from being accepted.
To work around this problem, I have seen many suggesting to set the SO_LINGER socket option with timeout 0 before calling close(). However, this is a bad solution as it causes the TCP connection to be terminated with an error.
Instead, design your application protocol so the connection termination is always initiated from the client side. If the client always knows when it has read all remaining data it can initiate the termination sequence. As an example, a browser knows from the Content-Length HTTP header when it has read all data and can initiate the close. (I know that in HTTP 1.1 it will keep it open for a while for a possible reuse, and then close it.)
If the server needs to close the connection, design the application protocol so the server asks the client to call close().
When to use SO_LINGER with timeout 0
Again, according to "UNIX Network Programming" third edition page 202-203, setting SO_LINGER with timeout 0 prior to calling close() will cause the normal termination sequence not to be initiated.
Instead, the peer setting this option and calling close() will send a RST (connection reset) which indicates an error condition and this is how it will be perceived at the other end. You will typically see errors like "Connection reset by peer".
Therefore, in the normal situation it is a really bad idea to set SO_LINGER with timeout 0 prior to calling close() – from now on called abortive close – in a server application.
However, certain situation warrants doing so anyway:
If a client of your server application misbehaves (times out, returns invalid data, etc.) an abortive close makes sense to avoid being stuck in CLOSE_WAIT or ending up in the TIME_WAIT state.
If you must restart your server application which currently has thousands of client connections you might consider setting this socket option to avoid thousands of server sockets in TIME_WAIT (when calling close() from the server end) as this might prevent the server from getting available ports for new client connections after being restarted.
On page 202 in the aforementioned book it specifically says: "There are certain circumstances which warrant using this feature to send an abortive close. One example is an RS-232 terminal server, which might hang forever in CLOSE_WAIT trying to deliver data to a stuck terminal port, but would properly reset the stuck port if it got an RST to discard the pending data."
I would recommend this long article which I believe gives a very good answer to your question.
The typical reason to set a SO_LINGER timeout of zero is to avoid large numbers of connections sitting in the TIME_WAIT state, tying up all the available resources on a server.
When a TCP connection is closed cleanly, the end that initiated the close ("active close") ends up with the connection sitting in TIME_WAIT for several minutes. So if your protocol is one where the server initiates the connection close, and involves very large numbers of short-lived connections, then it might be susceptible to this problem.
This isn't a good idea, though - TIME_WAIT exists for a reason (to ensure that stray packets from old connections don't interfere with new connections). It's a better idea to redesign your protocol to one where the client initiates the connection close, if possible.
When linger is on but the timeout is zero the TCP stack doesn't wait for pending data to be sent before closing the connection. Data could be lost due to this but by setting linger this way you're accepting this and asking that the connection be reset straight away rather than closed gracefully. This causes an RST to be sent rather than the usual FIN.
Thanks to EJP for his comment, see here for details.
Whether you can remove the linger in your code safely or not depends on the type of your application: is it a „client“ (opening TCP connections and actively closing it first) or is it a „server“ (listening to a TCP open and closing it after the other side initiated the close)?
If your application has the flavor of a „client“ (closing first) AND you initiate & close a huge number of connections to different servers (e.g. when your app is a monitoring app supervising the reachability of a huge number of different servers) your app has the problem that all your client connections are stuck in TIME_WAIT state. Then, I would recommend to shorten the timeout to a smaller value than the default to still shutdown gracefully but free up the client connections resources earlier. I would not set the timeout to 0, as 0 does not shutdown gracefully with FIN but abortive with RST.
If your application has the flavor of a „client“ and has to fetch a huge amount of small files from the same server, you should not initiate a new TCP connection per file and end up in a huge amount of client connections in TIME_WAIT, but keep the connection open and fetch all data over the same connection. Linger option can and should be removed.
If your application is a „server“ (close second as reaction to peer‘s close), on close() your connection is shutdown gracefully and resources are freed up as you don‘t enter TIME_WAIT state. Linger should not be used. But if your sever app has a supervisory process detecting inactive open connections idleing for a long time („long“ is to be defined) you can shutdown this inactive connection from your side - see it as kind of error handling - with an abortive shutdown. This is done by setting linger timeout to 0. close() will then send a RST to the client, telling him that you are angry :-)
I just saw that in the websockets RFC (RFC 6455), it explicitly states that the server should call close() on the TCP socket first(!)
I was in awe, as I hold the answer/posts by #mgd in this thread as de facto, and the RFC clearly goes against that. But, perhaps this would be a case where setting a linger time of 0 would be acceptable.
The underlying TCP connection, in most normal cases, SHOULD be closed
first by the server, so that it holds the TIME_WAIT state and not the
client
I'm very interested to hear any thoughts/insight on this.
In servers, you may like to send RST instead of FIN when disconnecting misbehaving clients. That skips FIN-WAIT followed by TIME-WAIT socket states in the server, which prevents from depleting server resources, and, hence, protects from this kind of denial-of-service attack.
I like Maxim's observation that DOS attacks can exhaust server resources. It also happens without an actually malicious adversary.
Some servers have to deal with the 'unintentional DOS attack' which occurs when the client app has a bug with connection leak, where they keep creating a new connection for every new command they send to your server. And then perhaps eventually closing their connections if they hit GC pressure, or perhaps the connections eventually time out.
Another scenario is when 'all clients have the same TCP address' scenario. Then client connections are distinguishable only by port numbers (if they connect to a single server). And if clients start rapidly cycling opening/closing connections for any reason they can exhaust the (client addr+port, server IP+port) tuple-space.
So I think servers may be best advised to switch to the Linger-Zero strategy when they see a high number of sockets in the TIME_WAIT state - although it doesn't fix the client behavior, it might reduce the impact.
The listen socket on a server can use linger with time 0 to have access to binding back to the socket immediately and to reset any clients whose connections are not yet finished connecting. TIME_WAIT is something that is only interesting when you have a multi-path network and can end up with miss-ordered packets or otherwise are dealing with odd network packet ordering/arrival-timing.