Beyond Ctrl-C: The dark corners of Unix signal handling

Beyond Ctrl-C: The dark corners of Unix signal handling(sunshowers.io)

165 points by PuercoPop 1 year ago | 71 comments

chrsig 1 year ago |

My favorite signal surprise was running nginx and/or httpd in the foreground and wondering why on earth it quit whenver i resized the window.

Turns out, they use SIGWINCH (which is sent on WINdow CHange) for graceful shutdown.

It's a silly, silly problem.

eadmund 1 year ago | |

> Turns out, they use SIGWINCH (which is sent on WINdow CHange) for graceful shutdown.

That’s … that’s even worse than people who send errors with an HTTP 200 response code.

aunderscored 1 year ago | | |

Disagree. Annoyingly there is a reasonable case for 200 but with an error, if http is your transport but not your application, then 200 says "yes, the message was transfered and understood correctly, here is your response" which may be an error response from the application

thezilch 1 year ago | | |

That's ... not what most people are doing. People send _application_ errors on HTTP 200 response codes, because HTTP response codes are for HTTP and not applications. Most "REST" libraries and webdev get this wrong, building ever more fragile web services.

chrsig 1 year ago | | |

y'know...what really is an error, anyway?

thayne 1 year ago | |

Why? That's what SIGTERM is for.

chrsig 1 year ago | | |

No clue what the decision making process was.

There's a bug report for httpd dating back to 2011[0]. The nginx mailling list also has a grumpy person contemporary with that[1].

My guess is someone thought "httpd is a server running somewhere without a monitor attached, why on earth would it get a SIGWINCH!? surely it's available to use for something completely different", not considering users running it in the foreground during development. Nginx probably followed suit for convention, but that's pure speculation on my part.

Also that was before docker really took off (I'm not sure if it was around in 2011 yet; still in it's infancy maybe). Running it in the foreground didn't happen as much yet. People were still using wamp or installing it via apt and restarting via sudo.

[0] https://bz.apache.org/bugzilla/show_bug.cgi?id=50669

[1] https://mailman.nginx.org/pipermail/nginx/2011-August/028640...

lolinder 1 year ago | | |

They use SIGWINCH for gracefully shutting down workers but not the main process [0]. SIGQUIT is used for a graceful shutdown and SIGTERM for a sort of graceful shutdown (with timeouts).

SIGWINCH is apparently used for an online upgrade [1]. Because it only shuts the workers down you can quickly transition back to the old binary and old configuration if there's a problem, even after upgrading the binary or config stored on the hard drive.

I'm sure there are other ways to get a similar capability, but this set of signals is apparently what they came up with.

[0] http://nginx.org/en/docs/dev/development_guide.html#processe...

[1] https://www.digitalocean.com/community/tutorials/how-to-upgr...

ibash 1 year ago | | |

I tried to find out why.

Unfortunately the change that introduces it predates the official release by a few months. And predates the mailing list by about a year:

https://trac.nginx.org/nginx/changeset/5238e93961a189c13eeff...

ykonstant 1 year ago | |

I don't know whether to laugh or cry.

chrsig 1 year ago | | |

definitely laugh! life's too short, you'll never get out alive :)

layer8 1 year ago |

> Another common extension is to use what is sometimes called a double Ctrl-C pattern. The first time the user hits Ctrl-C, you attempt to shut down the database cleanly, but the second time you encounter it, you give up and exit immediately.

This is a terrible behavior, because users tend to hit Ctrl-C multiple times without intending anything different than on a single hit (not to mention bouncing key mechanics and short key repeat delays). Unclean exits should be reserved for SIGQUIT (Ctrl-\) and SIGKILL (by definition).

cperciva 1 year ago |

The article doesn't mention the most useful of all signals: SIGINFO, aka "please print to stderr your current status". Very useful for tools like dd and tar.

Probably because Linux doesn't implement it. Worst mistake Linus ever made.

Also, it talks about self-pipe but doesn't mention that self-socket is much better since you can't select on a pipe.

epcoa 1 year ago | |

> self-socket is much better since you can't select on a pipe.

This needs further explanation. Why can’t you select on a pipe? You certainly can use select/poll on pipes in general and I’m not sure of any reason in particular they won’t work for the self pipe notification.

Its even right in the original: https://cr.yp.to/docs/selfpipe.html

efxhoy 1 year ago |

I recently wrote a little data transfer service in python that runs in ECS. When developing it locally it was easy to handle SIGINT: try write a batch, except KeyboardInterrupt, if caught mark the transfer as incomplete and finally commit the change and shut down.

But there’s no exception in python to catch for a SIGTERM, which is what ECS and other service mangers send when it’s time to shut down. So I had to add a signal handler. Would have been neat if SIGTERM could be caught like SIGINT with a “native” exception.

mananaysiempre 1 year ago | |

  from signal import SIGTERM, raise_signal, signal
  import sys # for excepthook
  class Terminate(BaseException):
      pass
  def _excepthook(type, value, traceback):
      if not issubclass(type, Terminate):
          return _prevhook(type, value, traceback)
      # If a Terminate went unhandled, make sure we are killed
      # by SIGTERM as far as wait(2) and friends are concerned.
      signal(SIGTERM, _prevterm)
      raise_signal(SIGTERM)
  _prevhook, sys.excepthook = sys.excepthook, _excepthook
  def terminate(signo=SIGTERM, frame=None):
      signal(SIGTERM, _prevterm)
      raise Terminate
  _prevterm = signal(SIGTERM, terminate)

Spivak 1 year ago | |

I mean you can just have the signal handler throw StopRequested in your Python boilerplate and never think about it again.

One common pattern is raising KeyboardInterrupt from your handler so it's all handled the same.

To restart the X server (Xsgi) once, do any one of the following (in increasing order of brutality): - killall -TERM Xsgi - hold down the left-Control, left-Shift, F12 and keypad slash keys (this is fondly known as the "Vulcan Death Grip") - /usr/gfx/stopgfx; /usr/gfx/startgfx - reboot To restart the X server every time someone logs out of the console, edit /var/X11/xdm/xdm-config, change the setting of "DisplayManager._0.terminateServer" from "False" to "True" and do 'killall -HUP xdm'.