On 2019-07-27 00:56:36 [-0500], Scott Wood wrote:
> If migrate_enable() is called while a task is preparing to sleep
> (state != TASK_RUNNING), that triggers a debug check in stop_one_cpu().
> Explicitly reset state to acknowledge that we're accepting the spurious
> wakeup.
>
> Signed-off-by: Scott Wood <sw...@redhat.com>
> ---
> kernel/sched/core.c | 8 ++++++++
> 1 file changed, 8 insertions(+)
>
> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> index 38a9a9df5638..eb27a9bf70d7 100644
> --- a/kernel/sched/core.c
> +++ b/kernel/sched/core.c
> @@ -7396,6 +7396,14 @@ void migrate_enable(void)
> unpin_current_cpu();
> preempt_lazy_enable();
> preempt_enable();
> +
> + /*
> + * Avoid sleeping with an existing non-running
> + * state. This will result in a spurious wakeup
> + * for the calling context.
> + */
> + __set_current_state(TASK_RUNNING);
Do you have an example for this? I'm not too sure if we are not using a
state by doing this. Actually we were losing it and get yelled at. We do:
|rt_spin_unlock()
|{
| rt_spin_lock_fastunlock();
| migrate_enable();
|}
So save the state as part of the locking process and lose it as part of
migrate_enable() if the CPU mask was changed. I *think* we need to
preserve that state until after the migration to the other CPU.
> sleeping_lock_inc();
> stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
> sleeping_lock_dec();
Sebastian
