Hello,
one friend noticed that multithreaded programs (VMware...) behave erraticaly
when started from rxvt or xterm with -e option. We've traced it down to some
strange SIGHUP delivery.
When process is session leader, is it supposed to receive SIGHUP when child
thread terminates? It did not receive SIGHUP under non-NPTL library, and IMHO
it was correct behavior. Now all exiting threads cause SIGHUP to be delivered
to the parent, and parent (I'd say correctly) assumes that connection to the
program was broken and terminates.
Probably disassociate_ctty(1) in do_exit() should not be invoked for all
tsk->signal->leader, but only for those with thread_group_empty() == 1, i.e.
when this process is really going away ? Or only for thread group leader?
Thanks,
Petr Vandrovec
/* Build with:
gcc -W -Wall -O2 -o testhup testhup.c -lpthread -lutil
vana:~# ./testhup
Got SIGHUP!
vana:~# LD_ASSUME_KERNEL=2.4.1 ./testhup
vana:~#
*/
#include <pthread.h>
#include <unistd.h>
#include <signal.h>
#include <stdio.h>
#include <fcntl.h>
#include <pty.h>
#include <stdlib.h>
#include <sys/wait.h>
static void sighup(int sig) {
printf("Got SIGHUP!\n");
}
static void* child(void* arg) {
return arg;
}
int main(void) {
int master, slave;
char name[100];
openpty(&master, &slave, name, NULL, NULL);
if (fork() == 0) {
pthread_t pth;
signal(SIGHUP, sighup);
/* Become session leader. */
if (setsid() == -1) {
perror("SetSID");
exit(1);
}
/* Assign controlling terminal. */
if (open(name, O_RDWR) == -1) {
perror("Open PTY");
exit(2);
}
/* In reality code above is xterm/rxvt, code below is an
* innocent MT application execed by xterm/rxvt */
/* Die with SIGHUP after child thread terminates. */
pthread_create(&pth, NULL, child, NULL);
pthread_join(pth, NULL);
exit(3);
} else {
wait(NULL);
}
return 0;
}
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1
Petr Vandrovec wrote:
> When process is session leader, is it supposed to receive SIGHUP when child
> thread terminates?
No, it's not. I hope somebody knowing the signal code will look at
this. I've forwarded the mail to Roland.
- --
➧ Ulrich Drepper ➧ Red Hat, Inc. ➧ 444 Castro St ➧ Mountain View, CA ❖
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.6 (GNU/Linux)
iD8DBQFBeEHt2ijCOnn/RHQRAsSmAJ9RXjR71D/rLDEG2IxKt5++VYfRGACdELgM
uxry9CwYy5/j/a2dHGyKdX0=
=V+53
-----END PGP SIGNATURE-----
> When process is session leader, is it supposed to receive SIGHUP when child
> thread terminates?
Nope, this is certainly wrong.
> Probably disassociate_ctty(1) in do_exit() should not be invoked for all
> tsk->signal->leader, but only for those with thread_group_empty() == 1, i.e.
> when this process is really going away ? Or only for thread group leader?
Indeed, but that particular kind of test is subject to race conditions. I
think the easiest way to fix this is to use the bookkeeping I added to get
process accounting correct (another thing that has to happen on the last
thread's death). Start with this patch:
http://www.kernel.org/pub/linux/kernel/people/akpm/patches/2.6/2.6.9/2.6.9-mm1/broken-out/acct-report-single-record-for-multithreaded-process.patch
Then throw this one on top.
Thanks,
Roland
---
The session leader should disassociate from its controlling terminal and
send SIGHUP signals only when the whole session leader process dies.
Currently, this gets done when any thread in that process dies, which is
wrong. This patch fixes it.
Signed-off-by: Roland McGrath <[email protected]>
--- linux-2.6/kernel/exit.c.acct
+++ linux-2.6/kernel/exit.c
@@ -783,6 +783,7 @@ static void exit_notify(struct task_stru
asmlinkage NORET_TYPE void do_exit(long code)
{
struct task_struct *tsk = current;
+ int group_dead;
profile_task_exit(tsk);
@@ -807,7 +808,8 @@ asmlinkage NORET_TYPE void do_exit(long
ptrace_notify((PTRACE_EVENT_EXIT << 8) | SIGTRAP);
}
- if (atomic_dec_and_test(&tsk->signal->live))
+ group_dead = atomic_dec_and_test(&tsk->signal->live);
+ if (group_dead)
acct_process(code);
__exit_mm(tsk);
@@ -818,7 +820,7 @@ asmlinkage NORET_TYPE void do_exit(long
exit_thread();
exit_keys(tsk);
- if (tsk->signal->leader)
+ if (group_dead && tsk->signal->leader)
disassociate_ctty(1);
module_put(tsk->thread_info->exec_domain->module);