After closing client socket on sever side and exit application, socket still open for some time.
I can see it via netstat
Every 0.1s: netstat -tuplna | grep 6676
tcp 0 0 127.0.0.1:6676 127.0.0.1:36065 TIME_WAIT -
I use log4cxx logging and telnet appender. log4cxx use apr sockets.
Socket::close() method looks like that:
void Socket::close() {
if (socket != 0) {
apr_status_t status = apr_socket_close(socket);
if (status != APR_SUCCESS) {
throw SocketException(status);
}
socket = 0;
}
}
And it's successfully processed. But after program is finished I can see opened socket via netstat, and if it starts again log4cxx unable to open 6676 port, because it is busy.
I tries to modify log4cxx.
Shutdown socket before close:
void Socket::close() {
if (socket != 0) {
apr_status_t shutdown_status = apr_socket_shutdown(socket, APR_SHUTDOWN_READWRITE);
printf("Socket::close shutdown_status %d\n", shutdown_status);
if (shutdown_status != APR_SUCCESS) {
printf("Socket::close WTF %d\n", shutdown_status != APR_SUCCESS);
throw SocketException(shutdown_status);
}
apr_status_t close_status = apr_socket_close(socket);
printf("Socket::close close_status %d\n", close_status);
if (close_status != APR_SUCCESS) {
printf("Socket::close WTF %d\n", close_status != APR_SUCCESS);
throw SocketException(close_status);
}
socket = 0;
}
}
But it didn't helped, bug still reproduced.
This is not a bug. Time Wait (and Close Wait) is by design for safety purpose. You may however adjust the wait time. In any case, on server's perspective the socket is closed and you are relax by the ulimit counter, it has not much visible impact unless you are doing stress test.
As noted by Calvin this isn't a bug, it's a feature. Time Wait is a socket state that says, this socket isn't in use any more but nevertheless can't be reused quite yet.
Imagine you have a socket open and some client is sending data. The data may be backed up in the network or be in-flight when the server closes its socket.
Now imagine you start the service again or start some new service. The packets on the wire aren't aware that its a new service and the service can't know the packets were destined for a service that's gone. The new service may try to parse the packets and fail because they're in some odd format or the client may get an unrelated error back and keep trying to send, maybe because the sequence numbers don't match and the receiving host will get some odd error. With timed wait the client will get notified that the socket is closed and the server won't potentially get odd data. A win-win. The time it waits should be sofficient for all in-transit data to be flused from the system.
Take a look at this post for some additional info: Socket options SO_REUSEADDR and SO_REUSEPORT, how do they differ? Do they mean the same across all major operating systems?
TIME_WAIT is a socket state to allow all in travel packets that could remain from the connection to arrive or dead before the connection parameters (source address, source port, desintation address, destination port) can be reused again. The kernel simply sets a timer to wait for this time to elapse, before allowing you to reuse that socket again. But you cannot shorten it (even if you can, you had better not to do it), because you have no possibility to know if there are still packets travelling or to accelerate or kill them. The only possibility you have is to wait for a socket bound to that port to timeout and pass from the state TIME_WAIT to the CLOSED state.
If you were allowed to reuse the connection (I think there's an option or something can be done in the linux kernel) and you receive an old connection packet, you can get a connection reset due to the received packet. This can lead to more problems in the new connection. These are solved making you wait for all traffic belonging to the old connection to die or reach destination, before you use that socket again.
Related
I have a system that can start multiple instances.
Every instance has a client and a server.
They are connected over socket/TCP
Every instance is started by starting a client.
The client starts (checks if IP is available, if not increase the IP by 1, checks again ...) -
The client starts the server with the free IP and connects to it. (for legacy reasons has to be like this)
Instance numbers 2, 3, 4, 5 work without issues.
...
Instance number 6. -> Fails on checking if the first IP in the range is available.
To check if IP is already in use, I do not close the socket on the server side so that it can accept the additional connection.
On the client-side, I check if I can connect to the server-side with the following code:
bool CheckIPInUse(char *ip)
{
bool ret = false;
int port = 12345;
int sock;
struct sockaddr_in serv_addr;
serv_addr.sin_family = AF_INET;
serv_addr.sin_port = htons(port);
// **non blocking** because I want the check to be fast.
sock = socket(AF_INET, SOCK_STREAM | SOCK_NONBLOCK, 0);
inet_pton(AF_INET, ip, &serv_addr.sin_addr);
int ret_conn = connect(sock, (struct sockaddr *)&serv_addr, sizeof(serv_addr));
if (ret_conn == 0){
fprintf(stdout, "connected");
ret = true;
}
else if (ret_conn < 0 && (errno != EINPROGRESS)){
fprintf(stdout, "failed to connect");
}
else
{
int check_if_connected = 10;
while (check_if_connected--)
{
socklen_t len = sizeof(serv_addr);
int ret_getpeer = getpeername(sock, (struct sockaddr *)&serv_addr, &len);
if (ret_getpeer == 0)
{
fprintf(stdout, "connected");
ret = true;
break;
}
usleep(100000);
}
}
close(sock);
return ret;
}
This works for the first 5 instances.
6th instance fails to connect to the first IP in range and tries to start the server with IP which is already in use. (always the 6th).
Is there any better way to check programmatically if IP/Port is already busy?
Any ideas on what to check. for failure in the instance number 6?
The only way to check if an ip/port on a server is available is to bind() to it. If it worked, it was available (but not any more).
Any approach that involves a test connect()ion first, to see if it fails, or anything along the lines of poking somewhere in /proc to see which IPs and ports are in use -- nothing along these lines will ever be 100% foolproof. That's because even if you reach the conclusion that the port is available, it may no longer be by the time you get around to try to bind() to it.
Now, you can take, as a starting position, that a particular IP and/or port range is reserved for your application's use, and you only wish to arbitrate IP/port allocation between different instances of your application. In that case you can do that pretty much whatever you want, you're not limited to attempting to actually start instances of your application, and hope for the best. One simplistic approach is to use lock files in /var/tmp to represent all possible IP/port combination, and have your application try, in turn, to acquire a lock on the corresponding lock file, first, and once it's official, and the lock file is acquired, then the corresponding IP/port then can be established at your leisure, but the lock file must remain locked until the IP/port is no longer in use.
But in terms of attempting to check if a socket port is available, or not, the only way to do it is to bind() it, because that, by definition, is what it does. You could attempt to implement a multi-layered approach, like trying to connect() first, and then attempt to bind() it, and if the bind() fails, then keep looking for a free port. But that's creating extra complexity, without much of a benefit.
Did you check that the server did not meet its maximum backlog length ?
You may be getting "connection refused" if the server you are trying to connect to
has more pending connections then the defined backlog.
So if multiple clients are testing at the same time, one of them may encounter this.
The most probable cause of your problem is that your client is getting a connect from the server due to the listen queue. The best way to avoid this problem is to close the socket on which you call accept(2) once all the instances are in use, and reopen it again when any of the server instances are finished.
The listen queue makes the kernel to accept (send the SYN/ACK segment) connections on the otherwise not yet open socket waiting, and this will make the connection establishment quicker for the next server instances if many such connections are entering in the system. All those connections are handled in the accept(2) socket, so the best way to accept five such connections is to close the accept socket as soon as the last connection has been established (this will not avoid the problem if a connection happens to enter the server in the time between one accept(2) and the next, but the connection so established will be closed as soon as the accept socket is still open)
In my opinion, you should have a master server process that forks new processes to handle the different connection and closes the accept socket as soon as it reaches the full capacity. Once one of the servers attending the connections closes one of them, it should reopen the accept socket and accept a new connection.
IMHO, also the most robust way of implementing such a system is to allow the extra connections to get in, but not attend them, so the connection remains open in case a new client happens to enter, and it can close it if the server doesn't attend it in a timeout interval. Having a sixth client already connected, but waiting for the server to say hello, will leave you in a state in which you can start talking to the server as soon as the last service ends.
When you use the simple ZeroMQ REQ/REP pattern you depend on a fixed send()->recv() / recv()->send() sequence.
As this article describes you get into trouble when a participant disconnects in the middle of a request because then you can't just start over with receiving the next request from another connection but the state machine would force you to send a request to the disconnected one.
Has there emerged a more elegant way to solve this since the mentioned article has been written?
Is reconnecting the only way to solve this (apart from not using REQ/REP but use another pattern)
As the accepted answer seem so terribly sad to me, I did some research and have found that everything we need was actually in the documentation.
The .setsockopt() with the correct parameter can help you resetting your socket state-machine without brutally destroy it and rebuild another on top of the previous one dead body.
(yeah I like the image).
ZMQ_REQ_CORRELATE: match replies with requests
The default behaviour of REQ sockets is to rely on the ordering of messages to match requests and responses and that is usually sufficient. When this option is set to 1, the REQ socket will prefix outgoing messages with an extra frame containing a request id. That means the full message is (request id, 0, user frames…). The REQ socket will discard all incoming messages that don't begin with these two frames.
Option value type int
Option value unit 0, 1
Default value 0
Applicable socket types ZMQ_REQ
ZMQ_REQ_RELAXED: relax strict alternation between request and reply
By default, a REQ socket does not allow initiating a new request with zmq_send(3) until the reply to the previous one has been received. When set to 1, sending another message is allowed and has the effect of disconnecting the underlying connection to the peer from which the reply was expected, triggering a reconnection attempt on transports that support it. The request-reply state machine is reset and a new request is sent to the next available peer.
If set to 1, also enable ZMQ_REQ_CORRELATE to ensure correct matching of requests and replies. Otherwise a late reply to an aborted request can be reported as the reply to the superseding request.
Option value type int
Option value unit 0, 1
Default value 0
Applicable socket types ZMQ_REQ
A complete documentation is here
The good news is that, as of ZMQ 3.0 and later (the modern era), you can set a timeout on a socket. As others have noted elsewhere, you must do this after you have created the socket, but before you connect it:
zmq_req_socket.setsockopt( zmq.RCVTIMEO, 500 ) # milliseconds
Then, when you actually try to receive the reply (after you have sent a message to the REP socket), you can catch the error that will be asserted if the timeout is exceeded:
try:
send( message, 0 )
send_failed = False
except zmq.Again:
logging.warning( "Image send failed." )
send_failed = True
However! When this happens, as observed elsewhere, your socket will be in a funny state, because it will still be expecting the response. At this point, I cannot find anything that works reliably other than just restarting the socket. Note that if you disconnect() the socket and then re connect() it, it will still be in this bad state. Thus you need to
def reset_my_socket:
zmq_req_socket.close()
zmq_req_socket = zmq_context.socket( zmq.REQ )
zmq_req_socket.setsockopt( zmq.RCVTIMEO, 500 ) # milliseconds
zmq_req_socket.connect( zmq_endpoint )
You will also notice that because I close()d the socket, the receive timeout option was "lost", so it is important set that on the new socket.
I hope this helps. And I hope that this does not turn out to be the best answer to this question. :)
There is one solution to this and that is adding timeouts to all calls. Since ZeroMQ by itself does not really provide simple timeout functionality I recommend using a subclass of the ZeroMQ socket that adds a timeout parameter to all important calls.
So, instead of calling s.recv() you would call s.recv(timeout=5.0) and if a response does not come back within that 5 second window it will return None and stop blocking. I had made a futile attempt at this when I run into this problem.
I'm actually looking into this at the moment, because I am retro fitting a legacy system.
I am coming across code constantly that "needs" to know about the state of the connection. However the thing is I want to move to the message passing paradigm that the library promotes.
I found the following function : zmq_socket_monitor
What it does is monitor the socket passed to it and generate events that are then passed to an "inproc" endpoint - at that point you can add handling code to actually do something.
There is also an example (actually test code) here : github
I have not got any specific code to give at the moment (maybe at the end of the week) but my intention is to respond to the connect and disconnects such that I can actually perform any resetting of logic required.
Hope this helps, and despite quoting 4.2 docs, I am using 4.0.4 which seems to have the functionality
as well.
Note I notice you talk about python above, but the question is tagged C++ so that's where my answer is coming from...
Update: I'm updating this answer with this excellent resource here: https://blog.cloudflare.com/when-tcp-sockets-refuse-to-die/ Socket programming is complicated so do checkout the references in this post.
None of the answers here seem accurate or useful. The OP is not looking for information on BSD socket programming. He is trying to figure out how to robustly handle accept()ed client-socket failures in ZMQ on the REP socket to prevent the server from hanging or crashing.
As already noted -- this problem is complicated by the fact that ZMQ tries to pretend that the servers listen()ing socket is the same as an accept()ed socket (and there is no where in the documentation that describes how to set basic timeouts on such sockets.)
My answer:
After doing a lot of digging through the code, the only relevant socket options passed along to accept()ed socks seem to be keep alive options from the parent listen()er. So the solution is to set the following options on the listen socket before calling send or recv:
void zmq_setup(zmq::context_t** context, zmq::socket_t** socket, const char* endpoint)
{
// Free old references.
if(*socket != NULL)
{
(**socket).close();
(**socket).~socket_t();
}
if(*context != NULL)
{
// Shutdown all previous server client-sockets.
zmq_ctx_destroy((*context));
(**context).~context_t();
}
*context = new zmq::context_t(1);
*socket = new zmq::socket_t(**context, ZMQ_REP);
// Enable TCP keep alive.
int is_tcp_keep_alive = 1;
(**socket).setsockopt(ZMQ_TCP_KEEPALIVE, &is_tcp_keep_alive, sizeof(is_tcp_keep_alive));
// Only send 2 probes to check if client is still alive.
int tcp_probe_no = 2;
(**socket).setsockopt(ZMQ_TCP_KEEPALIVE_CNT, &tcp_probe_no, sizeof(tcp_probe_no));
// How long does a con need to be "idle" for in seconds.
int tcp_idle_timeout = 1;
(**socket).setsockopt(ZMQ_TCP_KEEPALIVE_IDLE, &tcp_idle_timeout, sizeof(tcp_idle_timeout));
// Time in seconds between individual keep alive probes.
int tcp_probe_interval = 1;
(**socket).setsockopt(ZMQ_TCP_KEEPALIVE_INTVL, &tcp_probe_interval, sizeof(tcp_probe_interval));
// Discard pending messages in buf on close.
int is_linger = 0;
(**socket).setsockopt(ZMQ_LINGER, &is_linger, sizeof(is_linger));
// TCP user timeout on unacknowledged send buffer
int is_user_timeout = 2;
(**socket).setsockopt(ZMQ_TCP_MAXRT, &is_user_timeout, sizeof(is_user_timeout));
// Start internal enclave event server.
printf("Host: Starting enclave event server\n");
(**socket).bind(endpoint);
}
What this does is tell the operating system to aggressively check the client socket for timeouts and reap them for cleanup when a client doesn't return a heart beat in time. The result is that the OS will send a SIGPIPE back to your program and socket errors will bubble up to send / recv - fixing a hung server. You then need to do two more things:
1. Handle SIGPIPE errors so the program doesn't crash
#include <signal.h>
#include <zmq.hpp>
// zmq_setup def here [...]
int main(int argc, char** argv)
{
// Ignore SIGPIPE signals.
signal(SIGPIPE, SIG_IGN);
// ... rest of your code after
// (Could potentially also restart the server
// sock on N SIGPIPEs if you're paranoid.)
// Start server socket.
const char* endpoint = "tcp://127.0.0.1:47357";
zmq::context_t* context;
zmq::socket_t* socket;
zmq_setup(&context, &socket, endpoint);
// Message buffers.
zmq::message_t request;
zmq::message_t reply;
// ... rest of your socket code here
}
2. Check for -1 returned by send or recv and catch ZMQ errors.
// E.g. skip broken accepted sockets (pseudo-code.)
while (1):
{
try
{
if ((*socket).recv(&request)) == -1)
throw -1;
}
catch (...)
{
// Prevent any endless error loops killing CPU.
sleep(1)
// Reset ZMQ state machine.
try
{
zmq::message_t blank_reply = zmq::message_t();
(*socket).send (blank_reply);
}
catch (...)
{
1;
}
continue;
}
Notice the weird code that tries to send a reply on a socket failure? In ZMQ, a REP server "socket" is an endpoint to another program making a REQ socket to that server. The result is if you go do a recv on a REP socket with a hung client, the server sock becomes stuck in a broken receive loop where it will wait forever to receive a valid reply.
To force an update on the state machine, you try send a reply. ZMQ detects that the socket is broken, and removes it from its queue. The server socket becomes "unstuck", and the next recv call returns a new client from the queue.
To enable timeouts on an async client (in Python 3), the code would look something like this:
import asyncio
import zmq
import zmq.asyncio
#asyncio.coroutine
def req(endpoint):
ms = 2000 # In milliseconds.
sock = ctx.socket(zmq.REQ)
sock.setsockopt(zmq.SNDTIMEO, ms)
sock.setsockopt(zmq.RCVTIMEO, ms)
sock.setsockopt(zmq.LINGER, ms) # Discard pending buffered socket messages on close().
sock.setsockopt(zmq.CONNECT_TIMEOUT, ms)
# Connect the socket.
# Connections don't strictly happen here.
# ZMQ waits until the socket is used (which is confusing, I know.)
sock.connect(endpoint)
# Send some bytes.
yield from sock.send(b"some bytes")
# Recv bytes and convert to unicode.
msg = yield from sock.recv()
msg = msg.decode(u"utf-8")
Now you have some failure scenarios when something goes wrong.
By the way -- if anyone's curious -- the default value for TCP idle timeout in Linux seems to be 7200 seconds or 2 hours. So you would be waiting a long time for a hung server to do anything!
Sources:
https://github.com/zeromq/libzmq/blob/84dc40dd90fdc59b91cb011a14c1abb79b01b726/src/tcp_listener.cpp#L82 TCP keep alive options preserved for client sock
http://www.tldp.org/HOWTO/html_single/TCP-Keepalive-HOWTO/ How does keep alive work
https://github.com/zeromq/libzmq/blob/master/builds/zos/README.md Handling sig pipe errors
https://github.com/zeromq/libzmq/issues/2586 for information on closing sockets
https://blog.cloudflare.com/when-tcp-sockets-refuse-to-die/
https://github.com/zeromq/libzmq/issues/976
Disclaimer:
I've tested this code and it seems to be working, but ZMQ does complicate testing this a fair bit because the client re-connects on failure? If anyone wants to use this solution in production, I recommend writing some basic unit tests, first.
The server code could also be improved a lot with threading or polling to be able to handle multiple clients at once. As it stands, a malicious client can temporarily take up resources from the server (3 second timeout) which isn't ideal.
The server (192.168.1.5:3001), is running Linux 3.2, and is designed to only accept one connection at a time.
The client (192.168.1.18), is running Windows 7. The connection is a wireless connection. Both programs are written in C++.
It works great 9 in 10 connect/disconnect cycles. The tenth-ish (randomly happens) connection has the server accept the connection, then when it later actually writes to it (typically 30+s later), according to Wireshark (see screenshot) it looks like it's writing to an old stale connection, with a port number that the client has FINed (a while ago), but the server hasn't yet FINed. So the client and server connections seems to get out of sync - the client makes new connections, and the server tries writing to the previous one. Every subsequent connection attempt fails once it gets in this broken state. The broken state can be initiated by going beyond the maximum wireless range for a half a minute (as before 9 in 10 cases this works, but it sometimes causes the broken state).
Wireshark screenshot behind link
The red arrows in the screenshot indicate when the server started sending data (Len != 0), which is the point when the client rejects it and sends a RST to the server. The coloured dots down the right edge indicate a single colour for each of the client port numbers used. Note how one or two dots appear well after the rest of the dots of that colour were (and note the time column).
The problem looks like it's on the server's end, since if you kill the server process and restart, it resolves itself (until next time it occurs).
The code is hopefully not too out-of-the-ordinary. I set the queue size parameter in listen() to 0, which I think means it only allows one current connection and no pending connections (I tried 1 instead, but the problem was still there). None of the errors appear as trace prints where "// error" is shown in the code.
// Server code
mySocket = ::socket(AF_INET, SOCK_STREAM, 0);
if (mySocket == -1)
{
// error
}
// Set non-blocking
const int saveFlags = ::fcntl(mySocket, F_GETFL, 0);
::fcntl(mySocket, F_SETFL, saveFlags | O_NONBLOCK);
// Bind to port
// Union to work around pointer aliasing issues.
union SocketAddress
{
sockaddr myBase;
sockaddr_in myIn4;
};
SocketAddress address;
::memset(reinterpret_cast<Tbyte*>(&address), 0, sizeof(address));
address.myIn4.sin_family = AF_INET;
address.myIn4.sin_port = htons(Port);
address.myIn4.sin_addr.s_addr = INADDR_ANY;
if (::bind(mySocket, &address.myBase, sizeof(address)) != 0)
{
// error
}
if (::listen(mySocket, 0) != 0)
{
// error
}
// main loop
{
...
// Wait for a connection.
fd_set readSet;
FD_ZERO(&readSet);
FD_SET(mySocket, &readSet);
const int aResult = ::select(getdtablesize(), &readSet, NULL, NULL, NULL);
if (aResult != 1)
{
continue;
}
// A connection is definitely waiting.
const int fileDescriptor = ::accept(mySocket, NULL, NULL);
if (fileDescriptor == -1)
{
// error
}
// Set non-blocking
const int saveFlags = ::fcntl(fileDescriptor, F_GETFL, 0);
::fcntl(fileDescriptor, F_SETFL, saveFlags | O_NONBLOCK);
...
// Do other things for 30+ seconds.
...
const int bytesWritten = ::write(fileDescriptor, buffer, bufferSize);
if (bytesWritten < 0)
{
// THIS FAILS!! (but succeeds the first ~9 times)
}
// Finished with the connection.
::shutdown(fileDescriptor, SHUT_RDWR);
while (::close(fileDescriptor) == -1)
{
switch(errno)
{
case EINTR:
// Break from the switch statement. Continue in the loop.
break;
case EIO:
case EBADF:
default:
// error
return;
}
}
}
So somewhere between the accept() call (assuming that is exactly the point when the SYN packet is sent), and the write() call, the client's port gets changed to the previously-used client port.
So the question is: how can it be that the server accepts a connection (and thus opens a file descriptor), and then sends data through a previous (now stale and dead) connection/file descriptor? Does it need some sort of option in a system call that's missing?
I'm submitting an answer to summarize what we've figured out in the comments, even though it's not a finished answer yet. It does cover the important points, I think.
You have a server that handles clients one at a time. It accepts a connection, prepares some data for the client, writes the data, and closes the connection. The trouble is that the preparing-the-data step sometimes takes longer than the client is willing to wait. While the server is busy preparing the data, the client gives up.
On the client side, when the socket is closed, a FIN is sent notifying the server that the client has no more data to send. The client's socket now goes into FIN_WAIT1 state.
The server receives the FIN and replies with an ACK. (ACKs are done by the kernel without any help from the userspace process.) The server socket goes into the CLOSE_WAIT state. The socket is now readable, but the server process doesn't notice because it's busy with its data-preparation phase.
The client receives the ACK of the FIN and goes into FIN_WAIT2 state. I don't know what's happening in userspace on the client since you haven't shown the client code, but I don't think it matters.
The server process is still preparing data for a client that has hung up. It's oblivious to everything else. Meanwhile, another client connects. The kernel completes the handshake. This new client will not be getting any attention from the server process for a while, but at the kernel level the second connection is now ESTABLISHED on both ends.
Eventually, the server's data preparation (for the first client) is complete. It attempts to write(). The server's kernel doesn't know that the first client is no longer willing to receive data because TCP doesn't communicate that information! So the write succeeds and the data is sent out (packet 10711 in your wireshark listing).
The client gets this packet and its kernel replies with RST because it knows what the server didn't know: the client socket has already been shut down for both reading and writing, probably closed, and maybe forgotten already.
In the wireshark trace it appears that the server only wanted to send 15 bytes of data to the client, so it probably completed the write() successfully. But the RST arrived quickly, before the server got a chance to do its shutdown() and close() which would have sent a FIN. Once the RST is received, the server won't send any more packets on that socket. The shutdown() and close() are now executed, but don't have any on-the-wire effect.
Now the server is finally ready to accept() the next client. It begins another slow preparation step, and it's falling further behind schedule because the second client has been waiting a while already. The problem will keep getting worse until the rate of client connections slows down to something the server can handle.
The fix will have to be for you to make the server process notice when a client hangs up during the preparation step, and immediately close the socket and move on to the next client. How you will do it depends on what the data preparation code actually looks like. If it's just a big CPU-bound loop, you have to find some place to insert a periodic check of the socket. Or create a child process to do the data preparation and writing, while the parent process just watches the socket - and if the client hangs up before the child exits, kill the child process. Other solutions are possible (like F_SETOWN to have a signal sent to the process when something happens on the socket).
Aha, success! It turns out the server was receiving the client's SYN, and the server's kernel was automatically completing the connection with another SYN, before the accept() had been called. So there definitely a listening queue, and having two connections waiting on the queue was half of the cause.
The other half of the cause was to do with information which was omitted from the question (I thought it was irrelevant because of the false assumption above). There was a primary connection port (call it A), and the secondary, troublesome connection port which this question is all about (call it B). The proper connection order is A establishes a connection (A1), then B attempts to establish a connection (which would become B1)... within a time frame of 200ms (I already doubled the timeout from 100ms which was written ages ago, so I thought I was being generous!). If it doesn't get a B connection within 200ms, then it drops A1. So then B1 establishes a connection with the server's kernel, waiting to be accepted. It only gets accepted on the next connection cycle when A2 establishes a connection, and the client also sends a B2 connection. The server accepts the A2 connection, then gets the first connection on the B queue, which is B1 (hasn't been accepted yet - the queue looked like B1, B2). That is why the server didn't send a FIN for B1 when the client had disconnected B1. So the two connections the server has are A2 and B1, which are obviously out of sync. It tries writing to B1, which is a dead connection, so it drops A2 and B1. Then the next pair are A3 and B2, which are also invalid pairs. They never recover from being out of sync until the server process is killed and the TCP connections are all reset.
So the solution was to just change a timeout for waiting on the B socket from 200ms to 5s. Such a simple fix that had me scratching my head for days (and fixed it within 24 hours of putting it on stackoverflow)! I also made it recover from stray B connections by adding socket B to the main select() call, and then accept()ing it and close()ing it immediately (which would only happen if the B connection took longer than 5s to establish). Thanks #AlanCurry for the suggestion of adding it to the select() and adding the puzzle piece about the listen() backlog parameter being a hint.
I have written complex library for managing network communication based on iocp mechanism. Problem is that when server closes the connection by calling API method closesocket() this information is sometimes transmitted to client delayed by seconds or even minutes. My code for detecting connection closure looks like this (simplified):
ok = GetQueuedCompletionStatus(completion_port, &io_size, (PULONG_PTR)&context, &overlapped, 40);
if (!ok) {
// something went broken
DWORD err = GetLastError();
if (err == ERROR_CONNECTION_REFUSED) {
// connection failed
} else if (err == ERROR_SEM_TIMEOUT) {
// connection timeout
} else if (err == ERROR_NETNAME_DELETED) {
// connection closure - point of interest
} else if (err != WAIT_TIMEOUT) {
// unknown error
}
} else {
// process incomming or outgoing data
}
Why is this happening? I need to know about connection closure immediately to be able to connect to backup server (not so heavily loaded - disconnect is happening because of this).
How are you closing the connection?
If you're just calling closesocket() then you are initiating a shutdown sequence which will attempt to ensure that all data that is currently pending will reach the destination. This can take time, especially if the network connection has been overloaded and datagrams have been lost and TCP retransmission is occurring.
If you want to close the connection straight away, and lose any pending data, then set linger to 0 and then close the socket. This will issue an RST on the connection and you'll get that much quicker.
I tried to experiment with linger parameter as Len wrote but this did not help. Adding call of shutdown() function just before closesocket() helped me. After analyzing packets reaching network interface on client (with WireShark) I have found that RST packet was replaced by FIN packet. Curiously that RST packet was not delayed. So operating system knew that connection was closed but by some unknown reason this information was transferred to application layer very delayed. I measured delays between 10 seconds and 4 minutes.
I am writing an XMLRPC client in c++ that is intended to talk to a python XMLRPC server.
Unfortunately, at this time, the python XMLRPC server is only capable of fielding one request on a connection, then it shuts down, I discovered this thanks to mhawke's response to my previous query about a related subject
Because of this, I have to create a new socket connection to my python server every time I want to make an XMLRPC request. This means the creation and deletion of a lot of sockets. Everything works fine, until I approach ~4000 requests. At this point I get socket error 10048, Socket in use.
I've tried sleeping the thread to let winsock fix its file descriptors, a trick that worked when a python client of mine had an identical issue, to no avail.
I've tried the following
int err = setsockopt(s_,SOL_SOCKET,SO_REUSEADDR,(char*)TRUE,sizeof(BOOL));
with no success.
I'm using winsock 2.0, so WSADATA::iMaxSockets shouldn't come into play, and either way, I checked and its set to 0 (I assume that means infinity)
4000 requests doesn't seem like an outlandish number of requests to make during the run of an application. Is there some way to use SO_KEEPALIVE on the client side while the server continually closes and reopens?
Am I totally missing something?
The problem is being caused by sockets hanging around in the TIME_WAIT state which is entered once you close the client's socket. By default the socket will remain in this state for 4 minutes before it is available for reuse. Your client (possibly helped by other processes) is consuming them all within a 4 minute period. See this answer for a good explanation and a possible non-code solution.
Windows dynamically allocates port numbers in the range 1024-5000 (3977 ports) when you do not explicitly bind the socket address. This Python code demonstrates the problem:
import socket
sockets = []
while True:
s = socket.socket()
s.connect(('some_host', 80))
sockets.append(s.getsockname())
s.close()
print len(sockets)
sockets.sort()
print "Lowest port: ", sockets[0][1], " Highest port: ", sockets[-1][1]
# on Windows you should see something like this...
3960
Lowest port: 1025 Highest port: 5000
If you try to run this immeditaely again, it should fail very quickly since all dynamic ports are in the TIME_WAIT state.
There are a few ways around this:
Manage your own port assignments and
use bind() to explicitly bind your
client socket to a specific port
that you increment each time your
create a socket. You'll still have
to handle the case where a port is
already in use, but you will not be
limited to dynamic ports. e.g.
port = 5000
while True:
s = socket.socket()
s.bind(('your_host', port))
s.connect(('some_host', 80))
s.close()
port += 1
Fiddle with the SO_LINGER socket
option. I have found that this
sometimes works in Windows (although
not exactly sure why):
s.setsockopt(socket.SOL_SOCKET,
socket.SO_LINGER, 1)
I don't know if this will help in
your particular application,
however, it is possible to send
multiple XMLRPC requests over the
same connection using the
multicall method. Basically
this allows you to accumulate
several requests and then send them
all at once. You will not get any
responses until you actually send
the accumulated requests, so you can
essentially think of this as batch
processing - does this fit in with
your application design?
Update:
I tossed this into the code and it seems to be working now.
if(::connect(s_, (sockaddr *) &addr, sizeof(sockaddr)))
{
int err = WSAGetLastError();
if(err == 10048) //if socket in user error, force kill and reopen socket
{
closesocket(s_);
WSACleanup();
WSADATA info;
WSAStartup(MAKEWORD(2,0), &info);
s_ = socket(AF_INET,SOCK_STREAM,0);
setsockopt(s_,SOL_SOCKET,SO_REUSEADDR,(char*)&x,sizeof(BOOL));
}
}
Basically, if you encounter the 10048 error (socket in use), you can simply close the socket, call cleanup, and restart WSA, the reset the socket and its sockopt
(the last sockopt may not be necessary)
i must have been missing the WSACleanup/WSAStartup calls before, because closesocket() and socket() were definitely being called
this error only occurs once every 4000ish calls.
I am curious as to why this may be, even though this seems to fix it.
If anyone has any input on the subject i would be very curious to hear it
Do you close the sockets after using it?