Skip to contentSkip to conversation

fix group stop with exit race

PATCH In review Oleg Nesterov 2 participants 3 messages Dec 5, 2007 – Dec 6, 2007

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 fix group stop with exit race

do_signal_stop() counts all sub-thread and sets ->group_stop_count accordingly. Every thread should decrement ->group_stop_count and stop, the last one should notify the parent.

However a sub-thread can exit before it notices the signal_pending(), or it may be somewhere in do_exit() already. In that case the group stop never finishes properly.

Note: this is a minimal fix, we can add some optimizations later. Say we can return quickly if thread_group_empty(). Also, we can move some signal related code from exit_notify() to exit_signals().

include/linux/signal.h renamed from include/linux/signal.h~1_stop_exit +1 -0
@@ -241,6 +241,7 @@ extern int show_unhandled_signals;
 
 struct pt_regs;
 extern int get_signal_to_deliver(siginfo_t *info, struct k_sigaction *return_ka, struct pt_regs *regs, void *cookie);
+extern void exit_signals(struct task_struct *tsk);
 
 extern struct kmem_cache *sighand_cachep;
 
kernel/signal.c renamed from kernel/signal.c~1_stop_exit +26 -1
@@ -1707,7 +1707,7 @@ static int do_signal_stop(int signr)
 			 * stop is always done with the siglock held,
 			 * so this check has no races.
 			 */
-			if (!t->exit_state &&
+			if (!(t->flags & PF_EXITING) &&
 			    !is_task_stopped_or_traced(t)) {
 				stop_count++;
 				signal_wake_up(t, 0);
@@ -1868,6 +1868,31 @@ relock:
 	return signr;
 }
 
+void exit_signals(struct task_struct *tsk)
+{
+	int group_stop = 0;
+
+	spin_lock_irq(&tsk->sighand->siglock);
+	if (unlikely(tsk->signal->group_stop_count) &&
+			!--tsk->signal->group_stop_count) {
+		tsk->signal->flags = SIGNAL_STOP_STOPPED;
+		group_stop = 1;
+	}
+
+	/*
+	 * From now this task is not visible for group-wide signals,
+	 * see wants_signal(), do_signal_stop().
+	 */
+	tsk->flags |= PF_EXITING;
+	spin_unlock_irq(&tsk->sighand->siglock);
+
+	if (unlikely(group_stop)) {
+		read_lock(&tasklist_lock);
+		do_notify_parent_cldstop(tsk, CLD_STOPPED);
+		read_unlock(&tasklist_lock);
+	}
+}
+
 EXPORT_SYMBOL(recalc_sigpending);
 EXPORT_SYMBOL_GPL(dequeue_signal);
 EXPORT_SYMBOL(flush_signals);
kernel/exit.c renamed from kernel/exit.c~1_stop_exit +1 -1
@@ -945,7 +945,7 @@ fastcall NORET_TYPE void do_exit(long co
 		schedule();
 	}
 
-	tsk->flags |= PF_EXITING;
+	exit_signals(tsk);  /* sets PF_EXITING */
 	/*
 	 * tsk->flags are checked in the futex code to protect against
 	 * an exiting task cleaning up the robust pi futexes.
Acked-by on v1 1/1
#1 lore raw
Quoted 8 lines · Oleg Nesterov

do_signal_stop() counts all sub-thread and sets ->group_stop_count accordingly. Every thread should decrement ->group_stop_count and stop, the last one should notify the parent.

However a sub-thread can exit before it notices the signal_pending(), or it may be somewhere in do_exit() already. In that case the group stop never finishes properly.

Quoted 5 lines

Note: this is a minimal fix, we can add some optimizations later. Say we can return quickly if thread_group_empty(). Also, we can move some signal related code from exit_notify() to exit_signals().

Looks OK for me, even though we're doing more work on the exit path. OTOH I don't see a non-racy way of doing it w/out grabbing the lock. Did you try to bench how much this change costs?
Anyway, looks sane to me...

- Davide

#2 lore raw
Quoted 15 lines · Davide Libenzi
Quoted 12 lines · Oleg Nesterov

do_signal_stop() counts all sub-thread and sets ->group_stop_count accordingly. Every thread should decrement ->group_stop_count and stop, the last one should notify the parent.

However a sub-thread can exit before it notices the signal_pending(), or it may be somewhere in do_exit() already. In that case the group stop never finishes properly.

Note: this is a minimal fix, we can add some optimizations later. Say we can return quickly if thread_group_empty(). Also, we can move some signal related code from exit_notify() to exit_signals().

Quoted 5 lines
Quoted 1 line

Looks OK for me, even though we're doing more work on the exit path. OTOH I don't see a non-racy way of doing it w/out grabbing the lock. Did you try to bench how much this change costs?

Yes, you are right, this unconditional spin_lock() is not good, especially for exit_group/exec.

But please look at the next patch I am sending, it removes the pessimization almost completely.

The only difference: when there is no group exit in progress, we are doing

	spin_lock_irq(siglock);
	if (!signal_pending()) {
		unlock and return
	}

while the current code does

	if (!signal_pending())
		return;
	spin_lock_irq(siglock);
	...

It would be nice to measure the difference, but I can't invent the test-case.

I tested (just in case) 100000 fork+exit 's

	perl -e 'fork ? wait : exit for 1 .. 100_000'

with and without the patch, and didn't notice any difference as expected.

Quoted 1 line

Thanks for looking at this!

Oleg.