If you are using PHP >= 7.1 then DO NOT use `declare(ticks=1)` instead use `pcntl_async_signals(true)`

There's no performance hit or overhead with `pcntl_async_signals()`. See this blog post https://blog.pascal-martin.fr/post/php71-en-other-new-things.html for a simple example of how to use this.

Remember that signal handlers are called immediately, so every time when your process receives registered signal blocking functions like sleep() and usleep() are interrupted (or more like ended) like it was their time to finish.

This is expected behaviour, but not mentioned in this documentation. Sleep interruption can be very unfortunate when for example you run a daemon that is supposed to execute some important task exactly every X seconds/minutes - this task could be called too early or too often and it would be hard to debug why it's happening exactly.

Simply remember that signal handlers might interfere with other parts of your code.
Here is example workaround ensuring that function gets executed every minute no matter how many times our loop gets interrupted by incoming signal (SIGUSR1 in this case).

<?php

  pcntl_async_signals(TRUE);

  pcntl_signal(SIGUSR1, function($signal) {
    // do something when signal is called
  });

  function everyMinute() {
    // do some important stuff
  }

  $wait = 60;
  $next = 0;

  while (TRUE) {
    $stamp = time();
    do {
      if ($stamp >= $next) { break; }
      $diff = $next - $stamp;
      sleep($diff);
      $stamp = time();
    } while ($stamp < $next);
    
    everyMinute();
    
    $next = $stamp + $wait;
    sleep($wait);
  }

?>

So in this infinite loop do {} while() calculates and adds missing sleep() to ensure that our everyMinute() function is not called too early. Both sleep() functions here are covered so everyMinute() will never be executed before it's time even if process receives multiple SIGUSR1 signals during it's runtime.

Remember that declaring a tick handler can become really expensive in terms of CPU cycles: Every n ticks the signal handling overhead will be executed. 

So instead of declaring tick=1, test if tick=100 can do the job. If so, you are likely to gain fivefold speed.

As your script might always might miss some signals due to blocking operations like cURL downloads, call pcntl_signal_dispatch() on vital spots, e.g. at the beginning of your main loop.

For PHP >= 5.3.0, instead of declare(ticks = 1), you should now use pcntl_ signal_ dispatch().

You cannot assign a signal handler for SIGKILL (kill -9).

Since php >= 5.3 handles Closures, you are now able to define the Callback directly.
Try this:

<?php
declare(ticks = 1);

pcntl_signal(SIGUSR1, function ($signal) {
            echo 'HANDLE SIGNAL ' . $signal . PHP_EOL;
});

posix_kill(posix_getpid(), SIGUSR1);

die;
?>

Some weird signal interactions going on here. I'm running PHP 4.3.9.

sleep() calls seem to be interrupted when any signal is received by the PHP script. But things get weird when you sleep() inside a signal handler.

Ordinarily, signal handlers are non-reentrant. That is, if the signal handler is running, sending another signal has no effect. However, sleep() seems to override PHP's signal handling. If you sleep() inside a signal handler, the signal is received and the sleep() is interrupted.

This can be worked around like this:

function handler($signal)
{
    // Ignore this signal
    pcntl_signal($signal, SIG_IGN);

    sleep(10);

    // Reinstall signal handler
    pcntl_signal($signal, __FUNCTION__);
}

I don't see any mention of this behavior in the documentation.

Tip: when using objects, don't set the signal handler from the constructor or even a method called from the constructor - your internal variables will be uninitialised.

I was having some issues with processing a signal using an object method I use for something else as well. In particular, I wanted to handle SIGCHLD with my own method "do_reap()" which I also call after a stream_select timeout and that uses a non-blocking pcntl_waitpid function.

The method was called when the signal was received but it couldn't reap the child.

The only way it worked was by creating a new handler that itself called do_reap().

So in other words, the following does not work:

<?php
class Parent {
  /* ... */
  private function do_reap(){
    $p = pcntl_waitpid(-1,$status,WNOHANG);
    if($p > 0){
      echo "\nReaped zombie child " . $p;
    }

   public function run(){
    /* ... */
    pcntl_signal(SIGCHLD,array(&$this,"do_reap"));
    $readable = @stream_select($read,$null,$null,5); // 5 sec timeout
    if($readable === 0){
      // we got a timeout
      $this->do_reap();
   }
}
?>

But this work:

<?php
class Parent {
  /* ... */
  private function do_reap(){
    $p = pcntl_waitpid(-1,$status,WNOHANG);
    if($p > 0){
      echo "\nReaped zombie child " . $p;
    }

   public function run(){
    /* ... */
    pcntl_signal(SIGCHLD,array(&$this,"child_died"));
    $readable = @stream_select($read,$null,$null,5); // 5 sec timeout
    if($readable === 0){
      // we got a timeout
      $this->do_reap();
   }

    private function child_died(){
      $this->do_reap();
    }
}
?>

A word of caution around the third parameter (restart_syscalls) in pcntl_signal(...).

I kept having a repeated issue where (seemingly randomly) my script would "exit unexpectedly" (_exit_, not crash: the exit code was always 0) while tracking forked children using signal handlers.

It appears that the signal handling is not at fault (indeed, PHP isn't "at fault").  Having "restart_syscalls" set to FALSE seemed to be the problem's cause.

I haven't debugged the issue extensively - except to observe that the issue was highly intermittent and seemed to relate to my use of usleep() in conjunction with restart_syscalls=FALSE.

My theory is the usleep() was wrongly tracking time - as described over here: http://man7.org/linux/man-pages/man2/restart_syscall.2.html

Long story short, I re-enabled (which is the default value) restart_syscalls, and the issue went away.

If you're curious - register_shutdown_function was still being handled correctly - so PHP definitely was _NOT_ crashing.  Interestingly however, my procedural code never "resumed" after the signal was handled.

I do not believe this is a bug.  I believe it was ignorant user error.  YMMV.

If you have a script that needs certain sections to not be interrupted by a signal (especially SIGTERM or SIGINT), but want to make your script ready to process that signal ASAP, there's only one way to do it. Flag the script as having received the signal, and wait for your script to say its ready to process it.

Here's a sample script:

<?
    $allow_exit = true;  // are we permitting exit?
    $force_exit = false; // do we need to exit?

    declare(ticks = 1);
    register_tick_function('check_exit');
    pcntl_signal(SIGTERM, 'sig_handler');
    pcntl_signal(SIGINT, 'sig_handler');

    function sig_handler () {
        global $allow_exit, $force_exit;

        if ($allow_exit)
            exit;
        else
            $force_exit = true;
    }

    function check_exit () {
        global $allow_exit, $force_exit;

        if ($force_exit && $allow_exit)
            exit;
    }

    $allow_exit = false;

    $i = 0;
    while (++$i) {
        echo "still going (${i})\n";
        if ($i == 10)
            $allow_exit = true;

        sleep(2);
    }
?>

You set $allow_exit to true at all times when it is perfectly acceptable that your script could exit without warning. In sections where you really need the script to continue running, you set $allow_exit to false. Any signals received while $allow_exit is false will not take effect until you set $allow_exit to true.

<?
    $allow_exit = true;

    // unimportant stuff here. exiting will not harm anything

    $allow_exit = false;

    // really important stuff not to be interrupted

    $allow_exit = true;

    // more unimportant stuff. if signal was received during
    // important processing above, script will exit here
?>

Looks like php uses  RealTime signals. This means that if one signal is currently processed, then other signals will not be lost.

just for example
<?php
pcntl_signal(SIGHUP, SIG_IGN)
?>

in strace log looks like this

"rt_sigaction(SIGHUP, {SIG_IGN, [], SA_RESTORER|SA_RESTART, 0x7f8caf83cc30}, {SIG_DFL, [], 0}, 8) = 0"

And code for testing

<?php
pcntl_signal(SIGHUP, function($signo){
  echo "1\n";
  sleep(2);
  echo "2\n";
});

while(true){
  sleep(1);
  pcntl_signal_dispatch();
}
?>

Run this and send how many SIGHUP signals as you want. All of them will be processed.

P.S.
I guess "all of them". I couldn't found real signal queue size. Will appreciate if someone points me.

To be clear, the phrase "pcntl_signal() doesn't stack the signal handlers, but replaces them. " means that you can still have different functions for different signals but only one function per signal.

In other words, this will work as expected

<?php

pcntl_async_signals(true);
pcntl_signal(SIGUSR1,function(){echo "received SIGUSR1";});
pcntl_signal(SIGUSR2,function(){echo "received SIGUSR2";});
posix_kill(posix_getpid(),SIGUSR1);
posix_kill(posix_getpid(),SIGUSR2);

// returns "received SIGUSR1" and then "received SIGUSR2"

?>

In at least version 5.1.4, the parameter passed to the handler is not a strict integer.

I have had such problems as trying to add the signal to an array, but the array is completely screwed up when viewed (but not viewed immediately after being added). This occurs when the handler is a method (array($this, 'methodname')), or a traditional functions.

To avoid this bug, typecast the parameter to an integer:
(note that each newline may appear to just be 'n'.)

<?php
print("pid= " . posix_getpid() . "\n");
declare(ticks=1);
$arrsignals = array();

function handler($nsig)
{
    global $arrsignals;
    $arrsignals[] = (int)$nsig;
    print("Signal caught and registered.\n");
    var_dump($arrsignals);
}

pcntl_signal(SIGTERM, 'handler');

// Wait for signals from the command-line (just a simple 'kill (pid)').
$n = 15;
while($n)
{
    sleep(1);
    $n--;
}

print("terminated.\n\n");
var_dump($arrsignals);
?>

Dustin

When you are running a script inside of a loop that checks a socket, and it hangs on that checking (Either by flaw or design), it can't handle signals until some data is received.

A suggested workaround would be to use the stream_set_blocking function, or stream_select on the offending reads.

Multiple children return less than the number of children exiting at a given moment SIGCHLD signals is normal behavior for Unix (POSIX) systems.  SIGCHLD might be read as "one or more children changed status -- go examine your children and harvest their status values".  Signals are implemented as bit masks in most Unix systems, so there can only be 1 SIGCHLD bit set in any given kernel tick for a process.

I have been having trouble reaping my child process. It seemed most of the time, children were reaped properly but *sometimes* they would stay as zombies. I was catching the CHLD signal to reap child processes with the following code:

<?php

function childFinished($signal)
{
  global $kids;
  $kids--;
  pcntl_waitpid(-1, $status);
}

$kids = 0;
pcntl_signal(SIGCHLD, "childFinished");
for ($i = 0; $i < 1000; $i++)
{
  while ($kids >= 50) sleep(1);
  
  $pid = pcntl_fork();
  if ($pid == -1) die('failed to fork :(');

  /* child process */
  if ($pid == 0) 
  {
    /* do something */
    exit(0);
  }

  /* parent */
  else { $kids++; }
}

/* when finished, just clean up the children */
print "Reaping $kids children...\n";
while ($kids) sleep(1);

print "Finished.\n";
?>

The problem was, $kids never became zero so it would effectively wait forever. After wracking my brains (UNIX forks are new to me) I finally read the Perl IPC docs and viola, a solution! It turns out that because signal handlers are not re-entrant, my handler will not be called again while it is in use. The scenario that caused me trouble was that one child would exit and call the signal handler, which would pcntl_waitpid() it and decrement the counter. Meanwhile, another child would exit while the first child was still being reaped,  so the second would never get to notify the parent! 

The solution was to continually reap children from the SIGCHLD handler, so long as there were children to reap. Here is the *fixed* childFinished function:

<?php

function childFinished($signal)
{
  global $kids;

  while( pcntl_waitpid(-1, $status, WNOHANG) > 0 )
    $kids--;
}

?>

This issue occurs in at least PHP 5.1.2.

When a SIGINT is sent via CTRL+C or CTRL+BREAK, the handler is called. If this handler sends a SIGTERM to other children, the signals are not received.

SIGINT can be sent via posix_kill() and it work exactly as expected-- This only applies when initiated via a hard break.

Process handling is not available when using a blocking socket! Bear this in mind.

I found out then when you use pcntl_signal in a 'deamon' script and you run it before you fork childs it does not work as expected.

instead you need to use pcntl_signal in the child code of the child you are forking

and if you want to cach signals for the 'deamon' part you should use pcntl_signal in the parent code.

There are two documents I consider reading:

  Unix Signals Programming
  http://users.actcom.co.il/~choo/lupg/tutorials/

  Beej's Guide to Unix Interprocess Communication 
  http://www.ecst.csuchico.edu/~beej/guide/ipc/

Also, have a look at the manpage:

  http://www.mcsr.olemiss.edu/cgi-bin/man-cgi?signal+5

I ran into an interesting problem. CLI 4.3.10 Linux

The parent forked a child.  The child did a bunch of stuff, and when it was done it sent SIGUSR1 to the parent and immediately exited.

Result: 
The child turned into a zombie process waiting for the parent to harvest it, as it should.

But the parent was unable to harvest the child because it was receiving an endless barrage of SIGUSR1s.  In fact, it logged over 200000 before I shut it down (SIGKILL).  The parent was not able to do anything else (respond to other events) during that time.

No, it wasn't a bug in my child code.  Apparently, after sending a signal there is some "behind the covers" work that needs to occur to acknowledge signal completion, and when you exit immediately it is not able to happen, so the system just keeps trying.

Solution:  I introduced a small delay in the child, after sending the signal, and before exiting.  
No more Sig loops...

----------

P.S.  With respect to the note below.  The whole point of the sleep function is to enable the processing of other events.  So, yes, your non-renterent code, will suddenly get re-entered when you do a sleep, because you have just handed off control to the next pending event.

Ignoring the signal is only an option if the signal is unimportant to your program....   The better way to approach it, is to not do lengthy processing inside of the signal event.  Instead set a global flag and get the heck out of there as fast as possible.  Allow another part of your program to check the flag and do the processing outside of the signal event.  Usually your program is in some kind of loop when it is receiving signals, so checking a flag regularly shouldn't be a problem.

you should use following code to prevent situation described by anxious2006 (children exit near simultaneously)

public function sig_handler($signo){
  switch ($signo) {
    case SIGCLD:
      while( ( $pid = pcntl_wait ( $signo, WNOHANG ) ) > 0 ){
        $signal = pcntl_wexitstatus ( $signo );
      }
      break;
  }
}

Under my setup (FreeBSD 6.2-RELEASE / PHP 5.2.4 CLI) I've noticed that when a child exits the SIGCHLD handler in the parent is not always invoked. It seems to happen when two children exit near simultaneously.

In this instance the child prints "EXIT" and the parent prints "SIGCHLD received":

- EXIT
- SIGCHLD received

This works as expected, but now look what happens when three exit in quick succession:

- EXIT
- EXIT
- EXIT
- SIGCHLD received
- SIGCHLD received

Because of this quirk any code which tries to limit the maximum number of children by incrementing on fork and decrementing on SIGCHLD will eventually end up with a single child (or no further forks), since the "active children" count is always above the maximum. I've noticed similar behaviour with using decrement after pcntl_wait(). Hopefully there's a workaround for this.

static class method to get the name of a process signal as string:

(self::$processSignalDescriptions is used to cache the results)

<?php
    public static function getPOSIXSignalText($signo) {

        try {
            if (is_null(self::$processSignalDescriptions)) {

                self::$processSignalDescriptions = array();

                $signal_list = explode(" ", trim(shell_exec("kill -l")));
                foreach ($signal_list as $key => $value) {
                    self::$processSignalDescriptions[$key+1] = "SIG".$value;
                }
            }

            return isset(self::$processSignalDescriptions[$signo])?self::$processSignalDescriptions[$signo]:"UNKNOWN";

        } catch (Exception $e) {}

        return "UNKNOWN";
    }
?>

<?php
pcntl_signal(SIGTERM,  function($signo) {
    echo "\n This signal is called. [$signo] \n";
    Status::$state = -1;
});

class Status{
    public static $state = 0;
}

$pid = pcntl_fork();
if ($pid == -1) {
    die('could not fork');
}

if($pid) {
    // parent
} else {
    while(true) {
        // Dispatching...
        pcntl_signal_dispatch();
        if(Status::$state == -1) {
            // Do something and end loop.
            break;
        }
        
        for($j = 0; $j < 2; $j++) {
            echo '.';
            sleep(1);
        }
        echo "\n";
    }
    
    echo "Finish \n";
    exit();
}

$n = 0;
while(true) {
    $res = pcntl_waitpid($pid, $status, WNOHANG);
    
    // If the child process ends, then end the main process.
    if(-1 == $res || $res > 0)
        break;
    
    // Send a signal after 5 seconds..
    if($n == 5)
        posix_kill($pid, SIGTERM);
    
    $n++;
    
    sleep(1);
}
?>

Be careful, when using an object in your callback. It seems this elevates the reference count. You may not want it to happen in repeated child processes.

Please be aware that declaring ticks or calling pcntl_signal_dispatch() in 5.3 and later is required to make pcntl_signal do anything useful. I wish the documentation made this more clear.