child_process.exec is not abortable in Linux

[MadaraUchiha]

• Version: 15.8.0 (but also 12.21.0, and 16.0.0-pre)
• Platform: Linux madara-laptop-linux 5.4.0-65-generic The README is oriented towards Node #73-Ubuntu SMP Mon Jan 18 17:25:17 UTC 2021 x86_64 x86_64 x86_64 GNU/Linux
• Subsystem: child_process

What steps will reproduce the bug?

1. Create a script that will run forever eg. keep-alive.js: setInterval(() => {}, 10_000);
2. Create a script that will exec that file and abort it within a few seconds:

// keep-alive-test.js
const { exec } = require('child_process');

const ac = new AbortController();
const { signal } = ac;
const child = child_process.exec(`${process.execPath} ./keep-alive.js`, { signal ), console.log);

setTimeout(() => ac.abort(), 10_000);

3. Run node keep-alive-test.js &
4. Examine running processes with ps awwx | grep alive | grep -v grep | cat
5. You will note 3 running processes, node keep-alive-test.js, sh -c node keep-alive.js and node keep-alive.js
6. After 10 seconds, the abort is invoked.
7. Examine running processes again with ps awwx | grep alive | grep -v grep | cat
8. You will note 2 running processes, node keep-alive-test.js and node keep-alive.js (sh was terminated).
9. Both processes will live forever until explicitly terminated by the user (fg -> CTRL+C, or killall node etc)

How often does it reproduce? Is there a required condition?

Every time.

What is the expected behavior?

Expected behavior for ac.abort() was to abort the child process.

What do you see instead?

Child process does not get aborted, only the shell does.

Additional information

Internally, the abort handler for all child_process methods use child.kill() to terminate the underlying process. However, it is documented that this does not work "as expected" in Linux.

This makes ac.abort() quite useless on execed children, since none of them would actually get aborted at all.

This will probably happen if one uses child_process.spawn() to similarly spawn a shell to invoke a process, though I have not tested.

[MadaraUchiha]

cc @Linkgoron @benjamingr

[RaisinTen]

@MadaraUchiha Is this the same problem as reported in #37273?
exec calls execFile which calls spawn internally with the signal here:

node/lib/child_process.js

Lines 254 to 263 in 5c3bc21

	const child = spawn(file, args, {
	cwd: options.cwd,
	env: options.env,
	gid: options.gid,
	shell: options.shell,
	signal: options.signal,
	uid: options.uid,
	windowsHide: !!options.windowsHide,
	windowsVerbatimArguments: !!options.windowsVerbatimArguments
	});

and that's being fixed in #37325.

[benjamingr]

Fixed in #37273 - thanks for the report :)

[benjamingr]

Apparently not 😅

[MadaraUchiha]

and that's being fixed in #37325.

@RaisinTen I don't see how. ac.abort() still calls the underlying child.kill(), and child.kill() does not work as expected in Linux, when child is sh.

In fact, the fact that @Linkgoron uses sleep instead of the alive-forever fixture for the exec test indicates that they specifically want to use something that's long enough to not immediately terminate and affect the test, but short-lived enough as to not leak resources in the test.

[richardlau]

FTR I don't think this is limited to Linux. I've been able to reproduce this on our SmartOS CI instances while investigating nodejs/build#3154.

[BlackGlory]

As I recall exec is just an advanced wrapper for spawn, so once you use shell: true, the problem exists.

My current solution is to kill the whole process group via process.kill(-pid)(you also need detached: true):

signal?.addEventListener('abort', () => {
  if (childProcess.pid !== undefined) {
    process.kill(-childProcess.pid)
  }
})

[bnoordhuis]

I think this issue is a case of wrong expectations. Maybe the documentation can be clarified but I don't consider it a node bug, it's just how the UNIX process model works. Either avoid shell: true or use @BlackGlory's solution.

I'm going to close this but if someone wants to send a documentation pull request, please do.