There is a small window between setting t->task to NULL and waking the
task up (which would set TASK_RUNNING). So the timer would fire, run and
set ->task to NULL while the other side/do_nanosleep() wouldn't enter
freezable_schedule(). After all we are peemptible here (in
do_nanosleep() and on the timer wake up path) and on KVM/virt the
virt-CPU might get preempted.
So do_nanosleep() wouldn't enter freezable_schedule() but cancel the
timer which is still running and wait for it via
hrtimer_wait_for_timer(). Then wait_event()/might_sleep() would complain
that it is invoked with state != TASK_RUNNING.
This isn't a problem since it would be reset to TASK_RUNNING later
anyway and we don't rely on the previous state.
Move the state update to TASK_RUNNING before hrtimer_cancel() so there
are no complains from might_sleep() about wrong state.
Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
---
kernel/time/hrtimer.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/kernel/time/hrtimer.c b/kernel/time/hrtimer.c
index bf0ec3fbbe797..851b2134e77f4 100644
--- a/kernel/time/hrtimer.c
+++ b/kernel/time/hrtimer.c
@@ -1785,12 +1785,12 @@ static int __sched do_nanosleep(struct hrtimer_sleeper *t, enum hrtimer_mode mod
if (likely(t->task))
freezable_schedule();
+ __set_current_state(TASK_RUNNING);
hrtimer_cancel(&t->timer);
mode = HRTIMER_MODE_ABS;
} while (t->task && !signal_pending(current));
- __set_current_state(TASK_RUNNING);
if (!t->task)
return 0;
--
2.20.0.rc2
On 12/6/18 11:15 AM, Sebastian Andrzej Siewior wrote:
> There is a small window between setting t->task to NULL and waking the
> task up (which would set TASK_RUNNING). So the timer would fire, run and
> set ->task to NULL while the other side/do_nanosleep() wouldn't enter
> freezable_schedule(). After all we are peemptible here (in
> do_nanosleep() and on the timer wake up path) and on KVM/virt the
> virt-CPU might get preempted.
> So do_nanosleep() wouldn't enter freezable_schedule() but cancel the
> timer which is still running and wait for it via
> hrtimer_wait_for_timer(). Then wait_event()/might_sleep() would complain
> that it is invoked with state != TASK_RUNNING.
> This isn't a problem since it would be reset to TASK_RUNNING later
> anyway and we don't rely on the previous state.
>
> Move the state update to TASK_RUNNING before hrtimer_cancel() so there
> are no complains from might_sleep() about wrong state.
>
> Signed-off-by: Sebastian Andrzej Siewior <[email protected]>
Reviewed-by: Daniel Bristot de Oliveira <[email protected]>
Thanks!
-- Daniel