In tick_handle_oneshot_broadcast, it doesn't reprogram tick-broadcast when no more event to set. In hardware it's true, but the next event value on broadcast device is expired, which will be used in idle broadcast enter. It may cause no more timer set in system in below case.
nte_l: next_timer_event on local timer nte_b: next_timer_event on broadcast timer time: in ms, assume HZ = 128 core0 core1 broadcast timer 990: nte_l = 1004 nte_l = 1004 nte_b = 1000 enter idle enter idle nte_b = 1000 ----------------------------------------------------------- 996: non-timer interrupt wake up both core0 & core1 ----------------------------------------------------------- 996: core0/1 exit idle, clear broadcast oneshot mask, set local timer next event(nte_l = 996 + 8 = 1004) 1000:broadcast interrupt comes, one shot mask is 0. do not program broadcast next time event.(nte_b = 1000) 1002:core0/1 enter idle, shut down local timer. since nte_l is larger than nte_b, do not program broadcast timer. 1004:no more timer interrupt, not report rcu quiescent state completion. no new timer is set in tick_nohz_idle_enter when core0/1 enter/exit idle. at last system will crash. tick_program_event(dev->next_event, 1) is not guaranteed to generate local timer interrupt when exiting idle. When local timer is shutdown, we need to compare local next event to updated broadcast next event. So when there is no broadcast next event to set on hardware, we need to set software next_event to KTIME_MAX. Signed-off-by: Qiao Zhou <qiaoz...@asrmicro.com> --- kernel/time/tick-broadcast.c | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/kernel/time/tick-broadcast.c b/kernel/time/tick-broadcast.c index b398c2e..7648326 100644 --- a/kernel/time/tick-broadcast.c +++ b/kernel/time/tick-broadcast.c @@ -661,6 +661,14 @@ static void tick_handle_oneshot_broadcast(struct clock_event_device *dev) */ if (next_event != KTIME_MAX) tick_broadcast_set_event(dev, next_cpu, next_event); + else + /* + * not program broadcast timer, but set the value to show that + * the next event is not set, which is used in + * __tick_broadcast_oneshot_control in idle tick_broadcast + * enter/exit control flow. + */ + dev->next_event.tv64 = KTIME_MAX; raw_spin_unlock(&tick_broadcast_lock); -- 2.7.4