On Tue, 2019-09-24 at 17:25 +0200, Sebastian Andrzej Siewior wrote: > On 2019-09-24 08:53:43 [-0500], Scott Wood wrote: > > As I pointed out in the "[PATCH RT 6/8] sched: migrate_enable: Set state > > to > > TASK_RUNNING" discussion, we can get here inside the rtmutex code (e.g. > > from > > debug_rt_mutex_print_deadlock) where saved_state is already holding > > something -- plus, the waker won't have WF_LOCK_SLEEPER and therefore > > saved_state will get cleared anyway. > > So let me drop the saved_state pieces and get back to it once I get to > the other thread (which you replied and I didn't realised until now). > > Regarding the WF_LOCK_SLEEPER part. I think this works as expected. > Imagine: > > CPU0 CPU1 > spin_lock(); > set_current_state(TASK_UNINTERRUPTIBLE); > … > spin_unlock() > -> migrate_enable(); > -> stop_one_cpu(); <-- A) > other_func(); <-- B) > schedule(); > > So. With only CPU0 we enter schedule() with TASK_UNINTERRUPTIBLE because > the state gets preserved with the change I added (which is expected). > If CPU1 sends a wake_up() at A) then the saved_state gets overwritten > and we enter schedule() with TASK_RUNNING. Same happens if it is sent at > B) point which is outside of any migrate/spin lock related code. > > Was this clear or did I miss the point? When the stop machine finishes it will do a wake_up_process() via complete(). Since this does not pass WF_LOCK_SLEEPER, saved_state will be cleared, and you'll have TASK_RUNNING when you get to other_func() and schedule(), regardless of whether CPU1 sends wake_up() -- so this change doesn't actually accomplish anything. While as noted in the other thread I don't think these spurious wakeups are a huge problem, we could avoid them by doing stop_one_cpu_nowait() and then schedule() without messing with task state. Since we're stopping our own cpu, it should be guaranteed that the stopper has finished by the time we exit schedule(). -Scott
