php-fpm: graceful restart without blocking/losing requests

 ## Issue:

When php-fpm gets "graceful" reloaded it will stop accepting requests
while finishing reqests it already received. If those requests to finish
are of a long running kind like big uploads over a slow connection, this
will block new requests to be handled for an unacceptable long time.

 ### Current behavior

```
|
v  FPM (OLD) running
|
v  $ sudo kill -USR2 $(cat /run/php-fpm/php-fpm.pid)
|
v  OLD sends SIGQUIT to children
v  OLD stops handling requests
|
v  OLD keeps running until all children complete or time out
|
v  OLD exec()s
v  FPM (NEW) starts & re-uses listening socket
|
v  NEW handles incoming requests from now on
```

 ## Proposed Solution:

In the case of a reloading request, we immediately fork and exec a new
fpm instance to seamlessly handle new requests, while the old instance
keeps finishing those requests already accepted.

 ### New behavior

```
|
v  FPM (OLD) running
|
v  $ sudo kill -USR2 $(cat /run/php-fpm/php-fpm.pid)
|
v  OLD sends SIGQUIT to children
v  OLD fork()s & exec()s
|\
| \
|  v  FPM (NEW) starts & re-uses listening socket
|  v  NEW handles incoming requests from now on
|  |
v  |  OLD keeps running until all children complete or time out
v  |  OLD gone
-  |
  /
 /
v  NEW still going
|
```

We've been running this patch since PHP 5.6 (early 2017).

Author: m6w6

UPGRADING

@@ -112,6 +112,12 @@ PHP 7.4 UPGRADE NOTES
 3. Changes in SAPI modules
 ========================================
 
+- FPM:
+  . Graceful restart (SIGUSR2) behavior changed to immediately fork and exec a
+    new FPM instance while the old instance finishes the legacy of requests.
+    IMPORTANT NOTE: Existing start-stop scripts and/or daemon units may have to
+    be adopted to this behavior (two instances of FPM active while reloading).
+
 ========================================
 4. Deprecated Functionality
 ========================================

sapi/fpm/fpm/fpm_process_ctl.c

@@ -78,6 +78,16 @@ static void fpm_pctl_exit() /* {{{ */
 static void fpm_pctl_exec() /* {{{ */
 {
 
+	switch (fork()) {
+	case 0:
+		break;
+	case -1:
+		zlog(ZLOG_SYSERROR, "failed to reload: fork() failed");
+		/* no break */
+	default:
+		fpm_global_config.pid_file = NULL;
+		return;
+	}
 	zlog(ZLOG_NOTICE, "reloading: execvp(\"%s\", {\"%s\""
 			"%s%s%s" "%s%s%s" "%s%s%s" "%s%s%s" "%s%s%s"
 			"%s%s%s" "%s%s%s" "%s%s%s" "%s%s%s" "%s%s%s"
@@ -106,9 +116,8 @@ static void fpm_pctl_action_last() /* {{{ */
 {
 	switch (fpm_state) {
 		case FPM_PCTL_STATE_RELOADING:
-			fpm_pctl_exec();
-			break;
-
+			zlog(ZLOG_NOTICE, "exiting after reload");
+			exit(FPM_EXIT_OK);
 		case FPM_PCTL_STATE_FINISHING:
 		case FPM_PCTL_STATE_TERMINATING:
 			fpm_pctl_exit();
@@ -195,6 +204,10 @@ static void fpm_pctl_action_next() /* {{{ */
 	fpm_pctl_kill_all(sig);
 	fpm_signal_sent = sig;
 	fpm_pctl_timeout_set(timeout);
+
+	if (fpm_signal_sent == SIGQUIT && fpm_state == FPM_PCTL_STATE_RELOADING) {
+		fpm_pctl_exec();
+	}
 }
 /* }}} */
 

sapi/fpm/fpm/fpm_sockets.c

@@ -74,7 +74,17 @@ static void fpm_sockets_cleanup(int which, void *arg) /* {{{ */
 
 		if (which == FPM_CLEANUP_PARENT_EXIT_MAIN) {
 			if (ls->type == FPM_AF_UNIX) {
-				unlink(ls->key);
+				struct sockaddr_un sa_un;
+
+				memset(&sa_un, 0, sizeof(sa_un));
+				strlcpy(sa_un.sun_path, ls->key, sizeof(sa_un.sun_path));
+				sa_un.sun_family = AF_UNIX;
+
+				if (fpm_socket_unix_test_connect(&sa_un, sizeof(sa_un)) == 0) {
+					zlog(ZLOG_WARNING, "Keeping unix socket, another FPM instance seems to already listen on %s", ls->key);
+				} else {
+					unlink(ls->key);
+				}
 			}
 		}
 		free(ls->key);