Skip to contentSkip to conversation

change kernel threads to ignore signals instead of blocking them

PATCH In review Oleg Nesterov 3 participants 5 messages Apr 13, 2007 – Apr 24, 2007 Scheduler

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 change kernel threads to ignore signals instead of blocking them

On top of Eric's

kthread-dont-depend-on-work-queues-take-2.patch

Currently kernel threads use sigprocmask(SIG_BLOCK) to protect against signals. This doesn't prevent the signal delivery, this only blocks signal_wake_up(). Every "killall -33 kthreadd" means a "struct siginfo" leak.

Change kthreadd_setup() to set all handlers to SIG_IGN instead of blocking them (make a new helper ignore_signals() for that). If the kernel thread needs some signal, it should use allow_signal() anyway, and in that case it should not use CLONE_SIGHAND.

Note that we can't change daemonize() (should die!) in the same way, because it can be used along with CLONE_SIGHAND. This means that allow_signal() still should unblock the signal to work correctly with daemonize()ed threads.

However, disallow_signal() doesn't block the signal any longer but ignores it.

NOTE: with or without this patch the kernel threads are not protected from handle_stop_signal(), this seems harmless, but not good.

 include/linux/sched.h |    1 +
 kernel/exit.c         |    2 +-
 kernel/kthread.c      |   17 +++--------------
 kernel/signal.c       |   10 ++++++++++
 4 files changed, 15 insertions(+), 15 deletions(-)
include/linux/sched.h renamed from include/linux/sched.h~1_SIGIGN +1 -0
@@ -1299,6 +1299,7 @@ extern int in_egroup_p(gid_t);
 
 extern void proc_caches_init(void);
 extern void flush_signals(struct task_struct *);
+extern void ignore_signals(struct task_struct *);
 extern void flush_signal_handlers(struct task_struct *, int force_default);
 extern int dequeue_signal(struct task_struct *tsk, sigset_t *mask, siginfo_t *info);
 
kernel/signal.c renamed from kernel/signal.c~1_SIGIGN +10 -0
@@ -329,6 +329,16 @@ void flush_signals(struct task_struct *t
 	spin_unlock_irqrestore(&t->sighand->siglock, flags);
 }
 
+void ignore_signals(struct task_struct *t)
+{
+	int i;
+
+	for (i = 0; i < _NSIG; ++i)
+		t->sighand->action[i].sa.sa_handler = SIG_IGN;
+
+	flush_signals(t);
+}
+
 /*
  * Flush all handlers for a task.
  */
kernel/kthread.c renamed from kernel/kthread.c~1_SIGIGN +3 -14
@@ -215,24 +215,13 @@ EXPORT_SYMBOL(kthread_stop);
 static __init void kthreadd_setup(void)
 {
 	struct task_struct *tsk = current;
-	struct k_sigaction sa;
-	sigset_t blocked;
 
 	set_task_comm(tsk, "kthreadd");
 
-	/* Block and flush all signals */
-	sigfillset(&blocked);
-	sigprocmask(SIG_BLOCK, &blocked, NULL);
-	flush_signals(tsk);
-
-	/* SIG_IGN makes children autoreap: see do_notify_parent(). */
-	sa.sa.sa_handler = SIG_IGN;
-	sa.sa.sa_flags = 0;
-	siginitset(&sa.sa.sa_mask, sigmask(SIGCHLD));
-	do_sigaction(SIGCHLD, &sa, (struct k_sigaction *)0);
+	ignore_signals(tsk);
 
-	set_user_nice(current, -5);
-	set_cpus_allowed(current, CPU_MASK_ALL);
+	set_user_nice(tsk, -5);
+	set_cpus_allowed(tsk, CPU_MASK_ALL);
 }
 
 int kthreadd(void *unused)
kernel/exit.c renamed from kernel/exit.c~1_SIGIGN +1 -1
@@ -348,7 +348,7 @@ int disallow_signal(int sig)
 		return -EINVAL;
 
 	spin_lock_irq(&current->sighand->siglock);
-	sigaddset(&current->blocked, sig);
+	current->sighand->action[(sig)-1].sa.sa_handler = SIG_IGN;
 	recalc_sigpending();
 	spin_unlock_irq(&current->sighand->siglock);
 	return 0;
Acked-by on v1 1/1
#1 lore raw
Quoted 13 lines · Oleg Nesterov

On top of Eric's

kthread-dont-depend-on-work-queues-take-2.patch

Currently kernel threads use sigprocmask(SIG_BLOCK) to protect against signals. This doesn't prevent the signal delivery, this only blocks signal_wake_up(). Every "killall -33 kthreadd" means a "struct siginfo" leak.

Change kthreadd_setup() to set all handlers to SIG_IGN instead of blocking them (make a new helper ignore_signals() for that). If the kernel thread needs some signal, it should use allow_signal() anyway, and in that case it should not use CLONE_SIGHAND.

Quoted 8 lines

Note that we can't change daemonize() (should die!) in the same way, because it can be used along with CLONE_SIGHAND. This means that allow_signal() still should unblock the signal to work correctly with daemonize()ed threads.

However, disallow_signal() doesn't block the signal any longer but ignores it.

NOTE: with or without this patch the kernel threads are not protected from handle_stop_signal(), this seems harmless, but not good.

Hmm. I like it all except for disallow_signal.

disallow_signal currently only has one user, jffs2. While jffs2 currently doesn't care, given the way jffs2 is using disallow_signal I would expect it would prefer to have the signal blocked.

Thinking about this some more if jffs2 or anyone else wants blocked signal behavior they can go ahead and block the signal. Keeping disallow_signal in sync with allow_signal seems to make sense.

Nothing in the kernel should break with this change so:

#2 lore raw
Quoted 26 lines · On Fri, 13 Apr 2007 08:13:32 -0600 [email protected] (Eric W. Biederman)…
Quoted 13 lines · Oleg Nesterov

On top of Eric's

kthread-dont-depend-on-work-queues-take-2.patch

Currently kernel threads use sigprocmask(SIG_BLOCK) to protect against signals. This doesn't prevent the signal delivery, this only blocks signal_wake_up(). Every "killall -33 kthreadd" means a "struct siginfo" leak.

Change kthreadd_setup() to set all handlers to SIG_IGN instead of blocking them (make a new helper ignore_signals() for that). If the kernel thread needs some signal, it should use allow_signal() anyway, and in that case it should not use CLONE_SIGHAND.

Quoted 8 lines

Note that we can't change daemonize() (should die!) in the same way, because it can be used along with CLONE_SIGHAND. This means that allow_signal() still should unblock the signal to work correctly with daemonize()ed threads.

However, disallow_signal() doesn't block the signal any longer but ignores it.

NOTE: with or without this patch the kernel threads are not protected from handle_stop_signal(), this seems harmless, but not good.

Hmm. I like it all except for disallow_signal.

Quoted 7 lines

disallow_signal currently only has one user, jffs2. While jffs2 currently doesn't care, given the way jffs2 is using disallow_signal I would expect it would prefer to have the signal blocked.

Thinking about this some more if jffs2 or anyone else wants blocked signal behavior they can go ahead and block the signal. Keeping disallow_signal in sync with allow_signal seems to make sense.

jffs2 actually wants its head examined. W. T. F. does it think it's doing in there?

Sigh.

#3 lore raw
Quoted 2 lines · Andrew Morton

jffs2 actually wants its head examined. W. T. F. does it think it's doing in there?

Good question, especially with respect to SIGHUP.

It is on my short list of very annoying kernel threads...

NFS and a few kernel threads others currently need a way to abort an interruptible sleep when the are terminated. Which has been a pain finding a simple way to allow that without too much trouble.

I'm within about a day of having everything converted over to kthreads, allowing me to kill daemonize. Hopefully things I don't run into any serious snags in testing...

Eric

#4 lore raw
In reply to Oleg Nesterov
Quoted 13 lines · On Fri, 13 Apr 2007 11:31:16 +0400 Oleg Nesterov <[email protected]> wrote:

On top of Eric's

kthread-dont-depend-on-work-queues-take-2.patch

Currently kernel threads use sigprocmask(SIG_BLOCK) to protect against signals. This doesn't prevent the signal delivery, this only blocks signal_wake_up(). Every "killall -33 kthreadd" means a "struct siginfo" leak.

Change kthreadd_setup() to set all handlers to SIG_IGN instead of blocking them (make a new helper ignore_signals() for that). If the kernel thread needs some signal, it should use allow_signal() anyway, and in that case it should not use CLONE_SIGHAND.

Quoted 8 lines

Note that we can't change daemonize() (should die!) in the same way, because it can be used along with CLONE_SIGHAND. This means that allow_signal() still should unblock the signal to work correctly with daemonize()ed threads.

However, disallow_signal() doesn't block the signal any longer but ignores it.

NOTE: with or without this patch the kernel threads are not protected from handle_stop_signal(), this seems harmless, but not good.

I'm seeing 500 zombied instances of khelper (from udev startup). It only happens when the utrace patches are applied. Presumably an interaction between utrace and one of these kthread changes.

I'll drop utrace for now. I don't think it's getting much help from being in -mm at present and it's getting increasingly painful to keep it merged against all the other stuff which is happening.

Roland, I'll squirt all the extra utrace patches which I have in your direction. Please merge them or hang on to them for later on.