What to do when you get an error

Started by
6 comments, last by howie_007 11 years, 5 months ago
I'm writing a TCP client/server Boost::asio wrapper.

Say I get an error during a send or receive. The simple thing to do would be to drop the connection. I guess if I get an error during a send, I could try a few more times before I drop the connection. Would that be the logical way to handle a send error?

Say I get an error during a "receive". Since we are talking TCP, I'm not sure how I would re-sync with the client or server since the received data can be part of a message or a combination of many. Perhaps send a message that indicates an error has happened so stop sending data and re-sync messages.

Would that be a proper way to handle errors?
Advertisement
The base recv/send log on errno what kind of error just happened. Have you tried checking if there is some extra info? Some errors may be handled (for instance, on linux, EINTR), others indicate that something bad is happening . So I guess a good start would be checking for extra information.

Since you are using TCP, dropping the connection is a fairly decent way to handle the error, but again, it will depend on the error. If it is a timeout, you may try to resend the data other than that my approach would be log the error and drop the connection.

Currently working on a scene editor for ORX (http://orx-project.org), using kivy (http://kivy.org).

Thanks KnolanCross. Since I'm new to networking, I'm not sure what to expect as far as errors are concerned. Log the error and drop the connections sounds like a straight forward solution.
If you've gotten to the point of an error in sending, there's no use re-trying IMO. The kernel has already done all it can. You might as well give up.
Similarly, you'll want to keep a maximum buffer size for pending data to a client, and if it goes above a certain threshold, drop the client, as the client is too slow to keep up and thus will have a terrible interactive experience anyway.
So, it's my opinion that you typically should detect errors, log the specifics (operation, IP, error code, perhaps other data,) and mark the connection as needs-disconnection, and then handle it as such. If there turns out to be particular errors that are common, and can be worked around (your logging might let you analyze this,) then you can fine-tune the behavior at that point.
enum Bool { True, False, FileNotFound };

...keep a maximum buffer size for pending data to a client, and if it goes above a certain threshold, drop the client...


I never thought of that. Good idea! When I try to do a send and if the send is currently busy with a previous send, I store that message in a deque to try in the next cycle. So if the deque starts backing up, that would be a good sign to drop the connection.

What do you think would be the tipping point for messages piling up in the deque to warrant a disconnect? 5, 10, 20? I realize this might be game specific but in general, considering the messages will be small.
There was a very good analysis done a few years ago about network connections over the internet. They determined that 10% of packets get dropped. But it's not that 1 in 10 of your packets disappears. It's that 10% of the time you can send nothing... so you need to be able to detect complete stoppages of half a second, kill the connection (to drop all the pending data in your local queue which is unsent, but probably out of date) and then re-init the connections with whatever passes for a keyframe in your system (because you don't know what deltas made it through).

What do you think would be the tipping point for messages piling up in the deque to warrant a disconnect? 5, 10, 20? I realize this might be game specific but in general, considering the messages will be small.


For an action game, a backlog of 5 seconds might be sufficient to drop the client. For a RPG, you (or, rather, players) might tolerate up to 30 seconds. The number of packets then depends on how often you send them :-)

Also note that the kernel will have a send buffer, that you need to include in this mechanism. It may be that the kernel buffer is big enough, that if you fail to enqueue even a single byte, it's time to drop the client. You can set the size of the kernel buffer with the SO_SNDBUF socket option.

enum Bool { True, False, FileNotFound };
Thanks guys! This has been very helpful!

This topic is closed to new replies.

Advertisement