move the related code from exit_notify() to exit_signals()
PATCH 1/1 move the related code from exit_notify() to exit_signals()
The previous bugfix was not optimal, we shouldn't care about group stop when we are the only thread or the group stop is in progress. In that case nothing special is needed, just set PF_EXITING and return.
Also, take the related "TIF_SIGPENDING re-targeting" code from exit_notify().
So, from the performance POV the only difference is that we don't trust !signal_pending() until we take ->siglock. But this in fact fixes another ___pure___ theoretical minor race. __group_complete_signal() finds the task without PF_EXITING and chooses it as the target for signal_wake_up(). But nothing prevents this task from exiting in between without noticing the pending signal and thus unpredictably delaying the actual delivery.
kernel/exit.c renamed from kernel/exit.c~2_unify_signal_exit +0 -18
@@ -743,24 +743,6 @@ static void exit_notify(struct task_strustruct task_struct *t;struct pid *pgrp; -if (signal_pending (tsk) && !(tsk->signal->flags &SIGNAL_GROUP_EXIT ) - && !thread_group_empty (tsk)) { -/* -* This occurs when there was a race between our exit -* syscall and a group signal choosing us as the one to -* wake up. It could be that we are the only thread -* alerted to check for pending signals, but another thread -* should be woken now to take the signal since we will not. -* Now we'll wake all the threads in the group just to make -* sure someone gets all the pending signals. -*/ -spin_lock_irq (&tsk->sighand->siglock); -for (t =next_thread (tsk); t != tsk; t =next_thread (t)) -if (!signal_pending (t) && !(t->flags &PF_EXITING )) -recalc_sigpending_and_wake (t); -spin_unlock_irq (&tsk->sighand->siglock); - } -/* * This does two things: *
kernel/signal.c renamed from kernel/signal.c~2_unify_signal_exit +22 -5
@@ -1871,19 +1871,36 @@ relock:void exit_signals (struct task_struct *tsk) {int group_stop =0 ; +struct task_struct *t; -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 ; +if (thread_group_empty (tsk) ||signal_group_exit (tsk->signal)) { + tsk->flags |=PF_EXITING ; +return ; } +spin_lock_irq (&tsk->sighand->siglock);/* * From now this task is not visible for group-wide signals, * see wants_signal(), do_signal_stop(). */ tsk->flags |=PF_EXITING ; +if (!signal_pending (tsk)) + goto out; + +/* It could be that __group_complete_signal() choose us to +* notify about group-wide signal. Another thread should be +* woken now to take the signal since we will not. +*/ +for (t = tsk; (t =next_thread (t)) != tsk; ) +if (!signal_pending (t) && !(t->flags &PF_EXITING )) +recalc_sigpending_and_wake (t); + +if (unlikely (tsk->signal->group_stop_count) && + !--tsk->signal->group_stop_count) { + tsk->signal->flags =SIGNAL_STOP_STOPPED ; + group_stop =1 ; + } +out:spin_unlock_irq (&tsk->sighand->siglock);if (unlikely (group_stop)) {
Quoted 4 lines · Oleg Nesterov
The previous bugfix was not optimal, we shouldn't care about group stop when we are the only thread or the group stop is in progress. In that case nothing special is needed, just set
PF_EXITINGand return.
Quoted 8 lines
Also, take the related "
TIF_SIGPENDINGre-targeting" code fromexit_notify().So, from the performance POV the only difference is that we don't trust !signal_pending() until we take ->siglock. But this in fact fixes another
___pure___theoretical minor race.__group_complete_signal()finds the task withoutPF_EXITINGand chooses it as the target forsignal_wake_up(). But nothing prevents this task from exiting in between without noticing the pending signal and thus unpredictably delaying the actual delivery.
For a second there, you got me confused, since it wasn't clear to me this patch was going over the previous one. When posting patches that are not in any tree, it may be wise to include the set they nest onto ;)
Quoted 7 lines
kernel/signal.c renamed from kernel/signal.c~2_unify_signal_exit +7
-0+if (!signal_pending (tsk)) + goto out; + +/* It could be that __group_complete_signal() choose us to +* notify about group-wide signal. Another thread should be +* woken now to take the signal since we will not. +*/
kernel/signal.c relock: · line 1903
Quoted 10 lines
kernel/signal.c renamed from kernel/signal.c~2_unify_signal_exit +10
-0+for (t = tsk; (t =next_thread (t)) != tsk; ) +if (!signal_pending (t) && !(t->flags &PF_EXITING )) +recalc_sigpending_and_wake (t); + +if (unlikely (tsk->signal->group_stop_count) && + !--tsk->signal->group_stop_count) { + tsk->signal->flags =SIGNAL_STOP_STOPPED ; + group_stop =1 ; + } +out:
Looks fine to me, even though the one above could simply be an:
if (signal_pending(tsk)) {
...
}(since the "out" label is used by that "if" only).
- Davide