Request: forward all (well, most) signals to child process

[brettdh]

I recently upgraded a Pyinstaller-built program to use Python 3.11. This resulted in the binary using a newer version of glibc than the target platform supported, which led me to discover staticx, without which this effort would be rather doomed. So first things first - thank you for building this. 🙂

Back when I first started using Pyinstaller, I noticed that most signals were not being forwarded to the child process. In particular, my program uses a somewhat arbitrary signal number to control the logging level of the child process. A more commonly-used example would be SIGHUP, which is often used to ask a daemon to reload its configuration from disk.

I made a change to the Pyinstaller bootloader back in 2018 whereby it forwards all signals (except SIGCLD/SIGCHLD, because they're special) to the child process. I think something similar would work for staticx's (grammar...? 😅 ) bootloader, but I haven't had a chance to try it out yet.

I might get some time to put together a PR for this, but I wanted to first open this issue so as to:

1. get your feedback on the idea; and
2. give you the chance to maybe get to it before I do 😁

Thanks again!

[JonathonReinhart]

Ah, you implemented that change to the PyInstaller bootloader after it had already influenced the staticx bootloader (167a585). No wonder it's not there! 😄

I agree, all signals should be forwarded, not just termination signals, with the exception of SIGCHLD as you indicate, and SIGKILL and SIGSTOP, since according to signal(7):

The signals SIGKILL and SIGSTOP cannot be caught, blocked, or ignored.

I also think there's a mistake in pyinstaller/pyinstaller@2ddb80d in the bounds of the for() loop: It should start at 1, not 0. And the upper bound is technically correct, but it is a bit misleading: I would use num_signals = 64, and use i <= num_signals.

I'm experimenting with an implementation now, but I'm seeing some strange differences between signal() and sigaction() on musl vs glibc... stay tuned.

[JonathonReinhart]

I've posted a draft PR in #242.

The "strange differences" I was referring to were the fact that SIGRTMIN starts at 35 on musl-libc (32-34 are reserved for pthreads implementation I believe), and thus sigaction() was failing.

I'm still not sure we want to forward all signals 1-31 with SIGCHLD, SIGKILL, and SIGSTOP in a denylist. If you read all of the signals listed on signal(7), I think there are plenty of others that we probably don't want to forward, too. Specifically ones that originate from hardware exceptions like SIGSEGV, SIGILL, SIGBUS, SIGFPE, etc.

I think we might consider instead an allowlist approach, including e.g.

• SIGABRT (maybe)
• SIGINT
• SIGQUIT
• SIGTERM
• SIGUSR1
• SIGUSR2

[brettdh]

So long as the allow list is configurable when running staticx (with some reasonable defaults as you listed), that sounds just fine to me. Thanks for looking into it!

[brettdh]

In my original case, I was wrapping a legacy application which arbitrarily used signal 42. (A lot has changed since 2018, though, and I could change this to something else now if needed.)