PREEMPT_RT does not spin and wait until a running timer completes its
callback but instead it blocks on a sleeping lock to prevent a deadlock.
This blocking can not be done for workqueue's IRQ_SAFE timer which will
be canceled in an IRQ-off region. It has to happen to in IRQ-off region
because changing the PENDING bit and clearing the timer must not be
interrupted to avoid a busy-loop.
The callback invocation of IRQSAFE timer is not preempted on PREEMPT_RT
so there is no need to synchronize on timer_base::expiry_lock.
Don't acquire the timer_base::expiry_lock for TIMER_IRQSAFE flagged
timer.
Add a lockdep annotation to ensure that this function is always invoked
in preemptible context on PREEMPT_RT.
Reported-by: Mike Galbraith <efa...@gmx.de>
Signed-off-by: Sebastian Andrzej Siewior <bige...@linutronix.de>
---
kernel/time/timer.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/kernel/time/timer.c b/kernel/time/timer.c
index 7edc9fba34bbd..1e2403126ccfe 100644
--- a/kernel/time/timer.c
+++ b/kernel/time/timer.c
@@ -1288,7 +1288,7 @@ static void del_timer_wait_running(struct timer_list
*timer)
u32 tf;
tf = READ_ONCE(timer->flags);
- if (!(tf & TIMER_MIGRATING)) {
+ if (!(tf & (TIMER_MIGRATING | TIMER_IRQSAFE))) {
struct timer_base *base = get_timer_base(tf);
/*
@@ -1372,6 +1372,13 @@ int del_timer_sync(struct timer_list *timer)
*/
WARN_ON(in_irq() && !(timer->flags & TIMER_IRQSAFE));
+ /*
+ * Must be able to sleep on PREEMPT_RT because of the slowpath in
+ * del_timer_wait_running().
+ */
+ if (IS_ENABLED(CONFIG_PREEMPT_RT) && !(timer->flags & TIMER_IRQSAFE))
+ lockdep_assert_preemption_enabled();
+
do {
ret = try_to_del_timer_sync(timer);
--
2.29.1
