I just noticed we have generic semaphores, a couple of questions.
down():
spin_lock_irqsave(&sem->lock, flags);
...
__down(sem);
Why _irqsave ? we must not do down() with irqs disabled, and of course
__down() restores/clears irqs unconditionally.
Another question,
__down_common(TASK_KILLABLE):
if (state == TASK_KILLABLE && fatal_signal_pending(task))
goto interrupted;
/* --- WINDOW --- */
__set_task_state(task, TASK_KILLABLE);
schedule_timeout(timeout);
This looks racy. If SIGKILL comes in the WINDOW above, the event is lost.
The task will wait for up() or timeout with the fatal signal pending, and
it is not possible to wakeup it via kill() again.
This is easy to fix, but I wonder if we should change schedule() instead.
Note that __down_common() does 2 checks,
if (state == TASK_INTERRUPTIBLE && signal_pending(task))
goto interrupted;
if (state == TASK_KILLABLE && fatal_signal_pending(task))
goto interrupted;
they look very symmetrical, but the first one is OK, and the second is racy.
Also, I think we have the similar issues with lock_page_killable().
How about something like
int signal_pending_state(struct task_struct *tsk)
{
if (!(state & (TASK_INTERRUPTIBLE | TASK_WAKEKILL)))
return 0;
if (signal_pending(tsk))
return 0;
return (state & TASK_INTERRUPTIBLE) ||
__fatal_signal_pending(tsk);
}
now,
--- kernel/sched.c
+++ kernel/sched.c
@@ -4510,8 +4510,7 @@ need_resched_nonpreemptible:
clear_tsk_need_resched(prev);
if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
- if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
- signal_pending(prev))) {
+ if (unlikely(signal_pending_state(prev))) {
prev->state = TASK_RUNNING;
} else {
deactivate_task(rq, prev, 1);
Thoughts?
Oleg.
On Tue, Jun 03, 2008 at 04:33:09PM +0400, Oleg Nesterov wrote:
> I just noticed we have generic semaphores, a couple of questions.
>
> down():
>
> spin_lock_irqsave(&sem->lock, flags);
> ...
> __down(sem);
>
> Why _irqsave ? we must not do down() with irqs disabled, and of course
> __down() restores/clears irqs unconditionally.
How about reading the fine comments?
I would paste it, but Debian has fucked up my X copy and paste. Line 13
of kernel/semaphore.c.
> __down_common(TASK_KILLABLE):
>
> if (state == TASK_KILLABLE && fatal_signal_pending(task))
> goto interrupted;
>
> /* --- WINDOW --- */
>
> __set_task_state(task, TASK_KILLABLE);
> schedule_timeout(timeout);
>
> This looks racy. If SIGKILL comes in the WINDOW above, the event is lost.
> The task will wait for up() or timeout with the fatal signal pending, and
> it is not possible to wakeup it via kill() again.
Hmmm. I think you're right. But mutex.c has the same problem, then.
The wait_event_* macros get this right -- they set the task state before
they check for a signal.
> This is easy to fix, but I wonder if we should change schedule() instead.
> Note that __down_common() does 2 checks,
>
> if (state == TASK_INTERRUPTIBLE && signal_pending(task))
> goto interrupted;
> if (state == TASK_KILLABLE && fatal_signal_pending(task))
> goto interrupted;
>
> they look very symmetrical, but the first one is OK, and the second is racy.
Oh, because of the special casing in sched.c. Why not just move the
__set_task_state before the checks for signals pending? We'd have to
reset to TASK_RUNNING at the 'interrupted:' label, but that's OK.
> Also, I think we have the similar issues with lock_page_killable().
I don't think so because __wait_on_bit_lock sets the state before
checking the 'action' (sync_page_killable).
> How about something like
>
> int signal_pending_state(struct task_struct *tsk)
> {
> if (!(state & (TASK_INTERRUPTIBLE | TASK_WAKEKILL)))
> return 0;
> if (signal_pending(tsk))
> return 0;
>
> return (state & TASK_INTERRUPTIBLE) ||
> __fatal_signal_pending(tsk);
> }
>
> now,
>
> --- kernel/sched.c
> +++ kernel/sched.c
> @@ -4510,8 +4510,7 @@ need_resched_nonpreemptible:
> clear_tsk_need_resched(prev);
>
> if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
> - if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
> - signal_pending(prev))) {
> + if (unlikely(signal_pending_state(prev))) {
> prev->state = TASK_RUNNING;
> } else {
> deactivate_task(rq, prev, 1);
>
> Thoughts?
That might be worth doing anyway, but I'd leave that up to Ingo.
--
Intel are signing my paycheques ... these opinions are still mine
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours. We can't possibly take such
a retrograde step."
On 06/03, Matthew Wilcox wrote:
>
> On Tue, Jun 03, 2008 at 04:33:09PM +0400, Oleg Nesterov wrote:
> >
> > Why _irqsave ? we must not do down() with irqs disabled, and of course
> > __down() restores/clears irqs unconditionally.
>
> How about reading the fine comments?
Thanks,
> I would paste it, but Debian has fucked up my X copy and paste. Line 13
> of kernel/semaphore.c.
>
> > __down_common(TASK_KILLABLE):
> >
> > if (state == TASK_KILLABLE && fatal_signal_pending(task))
> > goto interrupted;
> >
> > /* --- WINDOW --- */
> >
> > __set_task_state(task, TASK_KILLABLE);
> > schedule_timeout(timeout);
> >
> > This looks racy. If SIGKILL comes in the WINDOW above, the event is lost.
> > The task will wait for up() or timeout with the fatal signal pending, and
> > it is not possible to wakeup it via kill() again.
>
> Hmmm. I think you're right. But mutex.c has the same problem, then.
and do_wait_for_common()
> > This is easy to fix, but I wonder if we should change schedule() instead.
> > Note that __down_common() does 2 checks,
> >
> > if (state == TASK_INTERRUPTIBLE && signal_pending(task))
> > goto interrupted;
> > if (state == TASK_KILLABLE && fatal_signal_pending(task))
> > goto interrupted;
> >
> > they look very symmetrical, but the first one is OK, and the second is racy.
>
> Oh, because of the special casing in sched.c. Why not just move the
> __set_task_state before the checks for signals pending?
Yes sure, this all is fixeable (we need set_task_state() of course).
But please compare
current->state = TASK_INTERRUPTIBLE;
schedule();
and
current->state = TASK_KILLABLE;
schedule();
it seems to me that it is not nice they behave "differently".
> > Also, I think we have the similar issues with lock_page_killable().
>
> I don't think so because __wait_on_bit_lock sets the state before
> checking the 'action' (sync_page_killable).
You are right.
> > int signal_pending_state(struct task_struct *tsk)
> > {
> > if (!(state & (TASK_INTERRUPTIBLE | TASK_WAKEKILL)))
> > return 0;
> > if (signal_pending(tsk))
> > return 0;
> >
> > return (state & TASK_INTERRUPTIBLE) ||
> > __fatal_signal_pending(tsk);
> > }
> >
> > now,
> >
> > --- kernel/sched.c
> > +++ kernel/sched.c
> > @@ -4510,8 +4510,7 @@ need_resched_nonpreemptible:
> > clear_tsk_need_resched(prev);
> >
> > if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
> > - if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
> > - signal_pending(prev))) {
> > + if (unlikely(signal_pending_state(prev))) {
> > prev->state = TASK_RUNNING;
> > } else {
> > deactivate_task(rq, prev, 1);
> >
> > Thoughts?
>
> That might be worth doing anyway, but I'd leave that up to Ingo.
Yes, we need Ingo's opinion ;)
Oleg.
2008/6/3 Oleg Nesterov <[email protected]>:
> I just noticed we have generic semaphores, a couple of questions.
>
> down():
>
> spin_lock_irqsave(&sem->lock, flags);
> ...
> __down(sem);
>
> Why _irqsave ? we must not do down() with irqs disabled, and of course
> __down() restores/clears irqs unconditionally.
>
>
> Another question,
>
> __down_common(TASK_KILLABLE):
>
> if (state == TASK_KILLABLE && fatal_signal_pending(task))
> goto interrupted;
>
> /* --- WINDOW --- */
>
> __set_task_state(task, TASK_KILLABLE);
> schedule_timeout(timeout);
>
> This looks racy. If SIGKILL comes in the WINDOW above, the event is lost.
> The task will wait for up() or timeout with the fatal signal pending, and
> it is not possible to wakeup it via kill() again.
>
> This is easy to fix, but I wonder if we should change schedule() instead.
[ for what it's worth ] I think, you are definitely right here.
The schedule() would be the right place to fix it. At the very least,
because otherwise callers are obliged to always check for
fatal_signal_pending(task) before scheduling with state ==
TASK_KILLABLE. e.g. schedule_timeout_killable().
Not very nice, IMHO.
> int signal_pending_state(struct task_struct *tsk)
> {
> if (!(state & (TASK_INTERRUPTIBLE | TASK_WAKEKILL)))
> return 0;
> if (signal_pending(tsk))
> return 0;
I guess, it should be ! signal_pending(tsk).
>
> return (state & TASK_INTERRUPTIBLE) ||
> __fatal_signal_pending(tsk);
> }
>
> if (state == TASK_INTERRUPTIBLE && signal_pending(task))
> goto interrupted;
> if (state == TASK_KILLABLE && fatal_signal_pending(task))
>
> Oleg.
>
--
Best regards,
Dmitry Adamushko
* Dmitry Adamushko <[email protected]> wrote:
> > This looks racy. If SIGKILL comes in the WINDOW above, the event is
> > lost. The task will wait for up() or timeout with the fatal signal
> > pending, and it is not possible to wakeup it via kill() again.
> >
> > This is easy to fix, but I wonder if we should change schedule()
> > instead.
>
> [ for what it's worth ] I think, you are definitely right here.
>
> The schedule() would be the right place to fix it. At the very least,
> because otherwise callers are obliged to always check for
> fatal_signal_pending(task) before scheduling with state ==
> TASK_KILLABLE. e.g. schedule_timeout_killable().
>
> Not very nice, IMHO.
i guess we should fix this in schedule() - is there a patch i could try?
Ingo