Prevent crashing due to SIGPIPE

[briancroom]

By default, calling send on a socket which has been closed on the remote end causes SIGPIPE to be sent to the process, which terminates it. These changes change the behavior to disable the signal, and instead make send return an error code if the socket has been closed.

This situation arises frequently when deploying to Cloud Foundry, as its healthcheck process opens a connection to the app and then closes without sending any data.

[kylef]

Thanks for the pull request, unfortunately this will only apply to Linux and not OS X since SO_NOSIGPIPE is not portable.

I'm wondering if a better approach could be just to override the signal handler for SIGPIPE and "do nothing", then send() should error from what I understand. This would at least work on all platforms.

What do you think @briancroom?

[briancroom]

Hey @kylef, the first chunk in this patch, enabling SO_NOSIGPIPE handles OS X, while the second chunk with MSG_NOSIGNAL handles Linux. (I've been looking at e.g. here and here)

Now that you mention it, it does actually seem that it would be simpler to change the disposition of SIGPIPE to just ignore it (signal(SIGPIPE, SIG_IGN)), although I hesitate because it may not be appropriate for Curassow to mess with state that affects the entire process.

I guess I'd lean towards sticking with the current implementation, but I'd be happy to switch it over if you'd prefer!

[kylef]

Okay that's great, I misread it initially and though they we're both #if os(Linux).

Thanks @briancroom!