I have a C++ networking application that accepts TCP connections from clients and then waits on the socket until the client decides to send data (sometimes they won't send anything for a long time and thats OK).
It mostly detects error conditions when clients crash or machines are turned off, but it takes many minutes to notice when the network cable to the client has been unplugged and I would prefer it to notice this condition as soon as possible.
I don't have control over the clients and I can't make them send something like a "ping". My server does send out a "ping" packet to the to the clients (but they won't send a response), but even when the cable is unplugged write() returns the correct number of bytes (I see the TCP stack sending retry packets in Wireshark).
What is the best way to notice the loss of connection ? It would be most convenient if I could detect it on the write() call.
I need this to work on Windows and on Linux.
Unfortunately, there is no way to distinguish the cable being pulled out at the other end from any other reason for packet loss. Having said that, you can approximate loss of connectivity at the other end as "indefinite packet loss" occurring over a sufficiently long period of time (say T). TCP tracks packet loss, so the general approach for doing this would be:
- Get the number of unacked bytes in the connection (say it's B)
- send data, size = N
- Set a timeout = T, when it fires, check the number of unacked bytes again. If it's B+N, then assume that the other side has lost connectivity. At this point, you could try ICMP echo to verify your assumption.
Getting TCP-specific information for a connection is not a standard interface on UNIX, and definitely not something portable to Windows. On Linux, there's a socket option called TCP_INFO, which you can call via getsockopt(). Google should give you some examples. I don't know if there's an equivalent option on Windows.
Another way to do this (i.e. approximate tracking of connectivity loss) is via RAW sockets. Open a RAW socket and filter it to receive only TCP traffic for your connection. Then rather than fetching information from TCP to determine if you are getting anything from the other end, simply wait to receive any packet from the other side. If you get something in the stipulated period, then it means the peer is still up.