Skip to contentSkip to conversation

setsid: don't fork

PATCH In review Phillip Susi 2 participants 5 messages Nov 14, 2013 – Nov 19, 2013 Util-Linux package

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 setsid: don't fork

2.24 added a switch to setsid to wait for the child to exit and return its exit status. I believe this was the wrong way to fix the underlying bug since the behavior of the program still differs depending on whether it is run as the group leader or not, and requires a switch to get the correct behavior. Instead of this --wait switch, just make sure the parent process is the one that execs the new program rather than the child process.

---
 sys-utils/setsid.1 |  3 ---
 sys-utils/setsid.c | 23 +++++++++--------------
 2 files changed, 9 insertions(+), 17 deletions(-)
sys-utils/setsid.1 +0 -3
index da8d648..ab1b20c 100644
@@ -16,9 +16,6 @@ runs a program in a new session.
 \fB\-c\fP, \fB\-\-ctty\fP
 Set the controlling terminal to the current one.
 .TP
-\fB\-w\fP, \fB\-\-wait\fP
-Wait the execution of the program to end, and return the exit value of
-the child as return value of the
 .BR setsid .
 .SH "SEE ALSO"
 .BR setsid (2)
sys-utils/setsid.c +9 -14
index 782de82..9810138 100644
@@ -47,12 +47,10 @@ int main(int argc, char **argv)
 {
 	int ch;
 	int ctty = 0;
-	pid_t pid;
-	int status = 0;
+	pid_t pid = 0;
 
 	static const struct option longopts[] = {
 		{"ctty", no_argument, NULL, 'c'},
-		{"wait", no_argument, NULL, 'w'},
 		{"version", no_argument, NULL, 'V'},
 		{"help", no_argument, NULL, 'h'},
 		{NULL, 0, NULL, 0}
@@ -71,9 +69,6 @@ int main(int argc, char **argv)
 		case 'c':
 			ctty=1;
 			break;
-		case 'w':
-			status = 1;
-			break;
 		case 'h':
 			usage(stdout);
 		default:
@@ -90,22 +85,22 @@ int main(int argc, char **argv)
 			err(EXIT_FAILURE, _("fork"));
 		case 0:
 			/* child */
+			sleep(5);
 			break;
 		default:
 			/* parent */
-			if (!status)
-				return EXIT_SUCCESS;
-			if (wait(&status) != pid)
-				err(EXIT_FAILURE, "wait");
-			if (WIFEXITED(status))
-				return WEXITSTATUS(status);
-			err(status, _("child %d did not exit normally"), pid);
+			if (setpgid(pid, 0) == -1)
+				err(EXIT_FAILURE, _("setpgid failed"));
+			if (setpgid(0, pid) == -1)
+				err(EXIT_FAILURE, _("setpgid failed"));
+
 		}
 	}
 	if (setsid() < 0)
 		/* cannot happen */
 		err(EXIT_FAILURE, _("setsid failed"));
-
+	if (pid)
+		kill(pid, SIGTERM);
 	if (ctty) {
 		if (ioctl(STDIN_FILENO, TIOCSCTTY, 1))
 			err(EXIT_FAILURE, _("failed to set the controlling terminal"));
Signature · 1 line
1.8.3.2
maintainer
#1 lore raw
Quoted 5 lines · Phillip Susi

2.24 added a switch to setsid to wait for the child to exit and return its exit status. I believe this was the wrong way to fix the underlying bug since the behavior of the program still differs depending on whether it is run as the group leader or not, and requires a switch to get the correct behavior.

Yep, it was our goal to not change the default behaviour that exists for years.

sys-utils/setsid.c int main(int argc, char **argv) · line 88

Quoted 7 lines

Instead of this --wait switch, just make sure the parent process is the one that execs the new program rather than the child process.

sys-utils/setsid.c +1 -0
 		case 0:
 			/* child */
+			sleep(5);

pause() ?

Quoted 13 lines
sys-utils/setsid.c +3 -7
 			break;
 		default:
 			/* parent */
-			if (!status)
-				return EXIT_SUCCESS;
-			if (wait(&status) != pid)
-				err(EXIT_FAILURE, "wait");
-			if (WIFEXITED(status))
-				return WEXITSTATUS(status);
-			err(status, _("child %d did not exit normally"), pid);
+			if (setpgid(pid, 0) == -1)
+				err(EXIT_FAILURE, _("setpgid failed"));
+			if (setpgid(0, pid) == -1)

sys-utils/setsid.c int main(int argc, char **argv) · line 103

Quoted 10 lines
sys-utils/setsid.c +4 -1
+				err(EXIT_FAILURE, _("setpgid failed"));
+
 		}
 	}
 	if (setsid() < 0)
 		/* cannot happen */
 		err(EXIT_FAILURE, _("setsid failed"));
-
+	if (pid)
+		kill(pid, SIGTERM);

It does not seem too elegant :-) If you really want to change the default behaviour than it would be better to make --wait default and exec() in child, then you don't need setpgid(), kill() and sleep/pause() at all.

Karel

Signature · 2 lines
 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com
#2 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 2 lines · Karel Zak

Yep, it was our goal to not change the default behaviour that exists for years.

It isn't really a default preference but a straight up bug since the behavior depends on how the shell invokes it. You only get the no wait behavior when it is a process group leader. When invoked not as a group leader, then you get the wait behavior whether you asked for it or not. So if you really want no wait to be the default behavior then you need to fork whether or not you're a group leader and the policy decision chooses whether to wait for the child or not.

Quoted 4 lines

It does not seem too elegant :-) If you really want to change the default behaviour than it would be better to make --wait default and exec() in child, then you don't need setpgid(), kill() and sleep/pause() at all.

What's not elegant about it? I think it's cleaner than keeping around a useless intermediary process that just waits for the child and returns its exit status. I'm still not sure why the kernel requires you to not be a group leader but this seems like a better workaround than forking off a lost child.

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.17 (MingW32)
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iQEcBAEBAgAGBQJSi31MAAoJEJrBOlT6nu75Hd4H/2zwoBVn/wFbKfbTCMKxXg/i 80+PwDicDX3hnsU4RgSfi7AvqWJkiMBabSYg6aPE8f1HpKCxRDhdOg3wGP2a3Xj6 wZJRMLmVfxVW+bC0IuWU9unnsGmLF35OGfUMM1oNc0f3pdMPtMG80nV3FTwWints cP4YoAaHa3DC+yV0e7hxtLDqJVwx6xlRVu5B9eo/uOk/GCnJ4Hv0Zj+4Ff/FF6LV 07dbSqt6glXLc1IIuja19qxpJL2xT/tQ0+723zNK3JiDdM6zSQ+0IGTRZw+/HhD+ kjtWQWxXSCRGj/bHcotJpwJ8qmNW7/gTGEpZk7lUNAEMH2OzNN9+2uIA7PD7hxk= =UgK/
-----END PGP SIGNATURE-----

maintainer
#3 lore raw
Quoted 4 lines · Phillip Susi

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 11/19/2013 8:24 AM, Karel Zak wrote:

Quoted 8 lines
Quoted 2 lines

Yep, it was our goal to not change the default behaviour that exists for years.

It isn't really a default preference but a straight up bug since the behavior depends on how the shell invokes it. You only get the no wait behavior when it is a process group leader. When invoked not as a group leader, then you get the wait behavior whether you asked for it or not. So if you really want no wait to be the default behavior

but why you don't want --wait behavior? It does not change anything, but fix the problem with group leader.

IMHO we can make it (--wait) the default (you almost convinced me that the default behaviour should be fixed).

Quoted 9 lines

then you need to fork whether or not you're a group leader and the policy decision chooses whether to wait for the child or not.

Quoted 4 lines

It does not seem too elegant :-) If you really want to change the default behaviour than it would be better to make --wait default and exec() in child, then you don't need setpgid(), kill() and sleep/pause() at all.

What's not elegant about it?

fork for sleep() and then kill it

Quoted 3 lines

I think it's cleaner than keeping around
a useless intermediary process that just waits for the child and returns its exit status.

it also fixes the problem with group leader as child process is never group leader (you have to call setpgid() to be leader).

I'm also not sure if remove --wait is the best way... what about people who already use it? (yep, it's release few weeks ago, but...)

Quoted 3 lines

I'm still not sure why the kernel requires
you to not be a group leader but this seems like a better workaround than forking off a lost child.

If I good understand the man page then setsid() caller is session and group leader of the *new* session and group. It's probably unwanted to make the original group orphaned (without leader), because processes in orphaned group get SIGCONT from kernel...

(BTW, SIGCONT is probably what will happen in your version after
kill(pid) as the sleeping child is group leader, right?)

Karel

Signature · 2 lines
 Karel Zak  <[email protected]>
 http://karelzak.blogspot.com
#4 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Quoted 2 lines · Karel Zak

but why you don't want --wait behavior? It does not change anything, but fix the problem with group leader.

I *do* want that behavior, I just think it was supposed to have been that way all along and which behavior you get should NOT depend on whether it is a group leader or not.

Quoted 2 lines

IMHO we can make it (--wait) the default (you almost convinced me that the default behaviour should be fixed).

Why even have the option then? Does anyone actually want the non wait behavior? If they have to add a switch to get it, they may as well just have the shell background the job. Simpler and cleaner than a switch.

Quoted 1 line

fork for sleep() and then kill it

Why is that worse than fork then exec? And I thought swapping the group leader so the parent process can call setsid() was pretty clever ;)

Quoted 3 lines

I'm also not sure if remove --wait is the best way... what about people who already use it? (yep, it's release few weeks ago, but...)

The longer it stays in, the more people will use it. Best to get rid of it now if it wasn't such a great idea, otherwise it will be there forever.

Quoted 7 lines

If I good understand the man page then setsid() caller is session and group leader of the *new* session and group. It's probably unwanted to make the original group orphaned (without leader), because processes in orphaned group get SIGCONT from kernel...

(BTW, SIGCONT is probably what will happen in your version after kill(pid) as the sleeping child is group leader, right?)

I think that happens the old way too: as soon as setsid exits after successfully forking the child, then the group is orphaned. In other words, both setpgrp() and exit() remove this process from being the group leader.

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.17 (MingW32)
Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/

iQEcBAEBAgAGBQJSi8DXAAoJEJrBOlT6nu75N+AH/1qyxjdJ2h+nJgN6BhIQ5WfN SaEbSmVPGFKEFKS3ltVnH+aGfS9WJiQO5DitNewrqOlek3VSvwG3T9WZRmDPjg0M gSaTPQMVFDIAcMcaUt71AXBEMxHLZUI5rEDLwXraaDT9ISGSGjCNRpfTrwSCwiHY jUrulunvS0z2WkqyJbHjQZK19Z0waOEvUO7KH8UK/BWqZ2ybCHvCmC1D/uq3K09A 1dmavup13NNbxF7mG80QqdWfm9nN54N1sQ8iPUrzrQyW/6b29aSfUr74JMKbbSzt Rodimqio/sYpf7iSbMDA4+Gmswh//IANOf9235h8s5LDb1EghSdVzl5ktu+2aLU= =ZK8O
-----END PGP SIGNATURE-----