View on GitHub

Vorner's random stuff

Cleanup support in Signal Hook

It seems one of the duties maintainers have is making sure their libraries are known. A library nobody uses just dies of loneliness. One of the way is to write blog posts about the libraries, so here we go. This one is about Signal Hook and its new support for cleaning up the signals.

A little about the library

Signals are an old Unix API. It’s also pretty terrible. I guess it seemed a good idea back then and maybe even was a good idea back then. But the landscape has changed since.

As with all Unix APIs, this one gives you enough rope to hang yourself. In a bazillion different very subtle ways which lead to absolutely irreproducible but highly spectacular effects like a single-threaded applications deadlocking or corrupt memory allocator internals. The worst thing about it is it looks fine most of the time (except when you demo the application to the customer).

You can either take my word that the API is quite painful to work with or read my previous post about signals that goes into much more details.

The library exists to handle some of the challenges of using signals correctly and offers some common patterns ready to be used. If you stick to the usual patterns, you don’t even have to write unsafe code.

New features since the last post

There are three improvements worth mentioning since then.

Using the library

The most flexible possibility is using the register function directly. That one allows to run whatever code inside the signal handler. The function is marked as unsafe for a good reason, as it opens the floodgates to many of the mentioned traps.

Usually it is enough to only mark the signal in question happened and notify some other thread about it. The library provides three ways to do that without using any unsafe on the caller side:

These patterns are not perfect. For example, I’d really like to have a support for crossbeam-channel, as that one can multiplex multiple channels together on the receiver end ‒ one wouldn’t need to use a separate signal handling thread nor tokio. It seems non-trivial, though.

But they seem to be good enough for most cases I’ve met. If you want to see examples, follow the above links, the documentation contains them.

The problem

All of this, including the direct approach, has one downside. Imagine a bug in the application which causes it to somehow lock up ‒ maybe it enters an infinite loop or deadlocks.

The application doesn’t respond, so the user presses CTRL+C. Usually the application would use one of the above patterns, stop whatever it was doing and proceeded to do a graceful shutdown. But because ti is running in circles in the infinite loop, it doesn’t notice it should shutdown. The application still doesn’t react in any meaningful way, so the user presses CTRL+C again, with much the same effect. Eventually the user would either use kill -9 or reboot the computer, depending on their knowledge and experience.

The trick here is to reset the signal handler to its OS default after receiving the first signal. The default for many signals, including SIGINT (the one coming from CTRL+C) is to terminate the program.

If we do that than the stuck application will terminate (in non-graceful manner) on the second CTRL+C. But if it’s healthy and responding like it should, it would terminate correctly after the first one.

To support this trick, the cleanup module was added (it seems the link doesn’t work yet, has some problem ‒ it’ll hopefully catch up soon; until then you can still build the documentation locally). It allows resetting signal handlers to defaults, either manually or registering the cleanup into the current signal handler. The latter can of course be combined with whatever else signal handling is done by Signal Hook.

There are still some rough edges around this. Most important one is that the cleanup is a one-way irreversible change, at least for now. But a program should terminate only once during its lifetime anyway (🤔 that’s actually an interesting challenge, finding a way to terminate the same application multiple times 😈).

I believe this is a place for another thank you, this time to Justin Karneges. While it was me who wrote the code, I didn’t notice it needed writing in the first places and wouldn’t do it without the ticket. Feature requests are no less important than pull requests.

Closing thoughts

If you need to ask if you need this library, the chances are you don’t, because most programs are fine with not touching signals at all. But if you want to set them, be sure to give this library a look. It’ll save you a lot of reading documentation of fine details and hunting down weird crashes happening once in a blue moon.

And if you find you have an idea how to improve a library (not just this one), share it. Maintainers like pull requests and they like good ideas. Not everything gets accepted all the times, but the interest from other people is valuable feedback and fuel driving both the project and the maintainer forward.