PHP 8.2 FPM Leaks processes on pcntl_fork

[Xmister]

Description

OS: xenial
PHP tested: 8.2.7, 8.2.10

The following drupal 10 index.php:

<?php

/**
 * @file
 * The PHP page that serves all page requests on a Drupal installation.
 *
 * All Drupal code is released under the GNU General Public License.
 * See COPYRIGHT.txt and LICENSE.txt files in the "core" directory.
 */

use Drupal\Core\DrupalKernel;
use Symfony\Component\HttpFoundation\Request;

$autoloader = require_once 'autoload.php';

$kernel = new DrupalKernel('prod', $autoloader);

$request = Request::createFromGlobals();
$response = $kernel->handle($request);
$response->send();
$pid = pcntl_fork();
    if ( $pid !== 0  ) {
        // we are the parent process
        return;
    }
    sleep(10);

$kernel->terminate($request, $response);

After doing ~5 requests to the site, results in the following sub-processes to the FPM master:

root     15873  0.0  0.8 464188 17200 ?        Ss   13:08   0:00 php-fpm: master process (/var/www/site-fpm/drush12test.prod/php-fpm-8.2.conf)
drush12+ 16518  0.7  1.4 466300 29860 ?        S    13:11   0:00  \_ php-fpm: pool drush12test
drush12+ 16557  0.0  0.6 466300 13068 ?        S    13:11   0:00      \_ php-fpm: pool drush12test
drush12+ 16559  0.0  0.6 466300 12640 ?        S    13:11   0:00      \_ php-fpm: pool drush12test
drush12+ 16561  0.0  0.6 466300 12640 ?        S    13:11   0:00      \_ php-fpm: pool drush12test
drush12+ 16563  0.0  0.6 466300 12640 ?        S    13:11   0:00      \_ php-fpm: pool drush12test

But then after a few seconds, when it should clean them up, they will instead detach and linger around:

root     15873  0.0  0.8 464188 17188 ?        Ss   13:08   0:00 php-fpm: master process (/var/www/site-fpm/drush12test.prod/php-fpm-8.2.conf)
drush12+ 16557  0.0  1.1 464252 23364 ?        S    13:11   0:00 php-fpm: pool drush12test
drush12+ 16559  0.0  1.1 464252 23028 ?        S    13:11   0:00 php-fpm: pool drush12test
drush12+ 16561  0.0  1.1 464252 23028 ?        S    13:11   0:00 php-fpm: pool drush12test
drush12+ 16563  0.0  1.1 466300 23432 ?        S    13:11   0:00 php-fpm: pool drush12test

The same doesn't happen with PHP 7.4. In that case the processes get cleaned up properly.

PHP Version

PHP 8.2.10

Operating System

Ubuntu 16.04

[KapitanOczywisty]

FPM is not designed to support user-forks, and the whole PCNTL is intended to be used in cli environment (or at least single request SAPIs).

[Xmister]

@KapitanOczywisty The issue is that this information is not shown on https://www.php.net/manual/en/function.pcntl-fork.php, so we can't expect our users to know that.
And it still seems like a regression over 7.4 (possibly over later versions as well).
If we now disable this for our customers, we are ripping them of a functionality that was working fine before, just so they can run the latest PHP version, which will probably reduce their willingness to do so.

[iluuu1994]

Pinging @bukka. I'm not sure if this is something we previously supported. pcntl_fork is a direct call to fork with little additional logic/syncing going on. Forking a worker process means the process will stick around after the request ends. It seems that if fork is enabled, it's the responsibility of the forked process to terminate itself. After all, there's no way to know whether we should terminate the forked process when the request ends. E.g. you might fork the process to perform some background task that outlives the request, and then terminate once finished.

In general, it's probably not a great idea to fork with the FPM sapi. If your application crashes, it might be hard to guarantee that the process terminates. If it doesn't, it will "leak" and stay alive. EDIT: You could probably use prctl(PR_SET_PDEATHSIG, SIGTERM); for this though, but we don't currently expose that. Nevermind, that would of course only terminate once the worker terminates.

[thg2k]

I think the user's expectation is that pcntl_fork() does a bit more than just calling system's fork(). Including the fact that if you call exit() on the child side of the script, the process terminates.

[iluuu1994]

I suppose it would be ok for the child process to terminate itself once its request processing is finished. I'm still not sure if this is a good idea, inheriting file descriptors, sockets, etc. that might interfere with the worker process. Anyway, let's see what Jakub thinks.

[bukka]

pcntl_fork (as well as some other pcntl functions) is not really supported by FPM so doing fork there is on your own risk and you need to make sure it is terminated before process exits. This is really a documentation issue at this stage. It's actaully not just pcntl but executing shell and spinning subprocesses there will result in the same issue.

I'm not sure we can fully address this but we could maybe introduce a way how to kill direct child process on request exit. It's not that simple task though. It would be a breaking change but having some optional disabling of it should probably suffice.

[Xmister]

@iluuu1994 @thg2k Just to note here that the behavior doesn't change when adding exit(); at the end of the child process.

[iluuu1994]

Indeed, exit is not an actual call to exit(). I'm not sure there's an option at all, other than posix_kill.