janet-lang / janet

A dynamic language and bytecode vm
https://janet-lang.org
MIT License
3.38k stars 217 forks source link

`os/sigaction` fails if a separate thread is waiting on `(file/read stdin :line)`. #1378

Closed amano-kenji closed 4 months ago

amano-kenji commented 5 months ago

Here's the example code.

(import spork/sh)

(defn action []
  (sh/rm "test")
  (print "Handled SIGINT!")
  (os/exit 1))

(ev/spawn-thread
  (let [server (net/listen :unix "test" :stream)]
    (forever
      (with [conn (net/accept server)]
        (print (ev/read conn :all))))))

(ev/spawn-thread
  (forever
    (print (file/read stdin :line))))

(defn main [_]
  (os/sigaction :int action true)
  (forever
    (ev/sleep 1)))

I had to press Ctrl+C multiple times to kill this script.

iacore commented 5 months ago

a signal is send to a random thread of the process group. use sigmask on the other threads.

amano-kenji commented 5 months ago

How can I use sigmask?

iacore commented 5 months ago
man sigprocmask
man sigwait

you can mask the selected signals on every thread, and use sigwait to listen for new signals

amano-kenji commented 5 months ago

So, you are saying this isn't an issue? Or?

bakpakin commented 5 months ago

There is definitely some improvement to be made here - we can probably have better behavior by default, and possibly expose sigprocmask for control over which threads handle signals. Regardless, the current behavior isn't great.

bakpakin commented 5 months ago

An aside - a quick check on my machine seems to indicate that using :term SIGTERM instead of :int SIGINT will have the desired effect.

EDIT: nevermind, it works trivially because sigint handler is never called.

bakpakin commented 5 months ago

Looking into this some more, I think the issue is some undefined behavior of what happens when you called exit in a program with pthread "detached" threads, of which all Janet's threads are. This is certainly an issue that needs some careful consideration - by default, calling exit will run atexit registered functions in each thread. One "fix" for this issue is to replace exit in janet's source code with _exit, which bypasses this atexit functionality that pthread presumably uses for cleanup.

I'm not sure this is an ideal solution, though.

iacore commented 5 months ago

some undefined behavior

On Linux, the process group is killed together. The atexit is a libc feature.

Does janet use atexit at all?

iacore commented 5 months ago

I found another issue while playing around with signal handler.

During the signal handler, the process signal mask is not changed (deviation from the POSIX behavior). So if you send SIGINT inside the Janet signal handler, the expected behavior is to not intercept the program.

Current behavior: Pressing Ctrl+C makes the following program go into a loop.

(import spork/sh)

(def kill (ffi/lookup (ffi/native) "kill"))
(def signature (ffi/signature :default :int :int :int))

(defn action []
  (sh/rm "test")
  (print "Handled SIGINT!")
  (ffi/call kill signature 0 2)
  (print "unreachable"))

(ev/spawn-thread
  (let [server (net/listen :unix "test" :stream)]
    (forever
      (with [conn (net/accept server)]
        (print (ev/read conn :all))))))

(ev/spawn-thread
  (forever
    (print (file/read stdin :line))))

(defn main [_]
  (os/sigaction :int action)
  (forever
    (ev/sleep 1)))
iacore commented 5 months ago

@amano-kenji here's a janky workaround you can use for now (on Linux)

(def kill (ffi/lookup (ffi/native) "kill"))
(def signature (ffi/signature :default :int :int :int))

(defn action []
  (sh/rm "test")
  (print "Handled SIGINT!")
  (ffi/call kill signature 0 15))
bakpakin commented 4 months ago

This has been fixed on the latest master with the addition of a force argument to exit that calls _exit instead. The cause of this is just that exit() flushed stdio streams before exiting, as per the C standard, so this is not an issue with the interpreter.

sogaiu commented 4 months ago

So since f6df8ff93, os/exit has the docstring:

    (os/exit &opt x force)

    Exit from janet with an exit code equal to x. If x is not an 
    integer, the exit with status equal the hash of x. If `force` is 
    truthy will exit immediately and skip cleanup code.
amano-kenji commented 4 months ago

skip cleanup code

Does this mean os/sigaction will be skippped? Does this mean files and streams aren't closed?

amano-kenji commented 4 months ago

I just tested the latest commit.

(os/exit 1 true) can be called in signal handlers, but (os/exit 1) skips signal handlers.