Re: Race condition when replacing the broadcast timer

[Thomas Gleixner]

On Wed, Jun 26 2024 at 02:17, 朱恺乾 wrote:
> We find a possible race condition when replacing the broadcast
> timer. Here is how the race happend,

> 1. In thread 0, ___tick_broadcast_oneshot_control, timer 0 as a
> broadcast timer is updating the next_event.

> 2. In thread 1, tick_install_broadcast_device, timer 0 is going to be
> replaced by a new timer 1.

> 3. If thread 0 gets the broadcast timer first, it would have the old
> timer returned (timer 0). When thread 1 shuts the old timer down and
> marks it as detached, Thread 0 still have the chance to re-enable the
> old timer with a noop handler if it executes slower than thread 1.

> 4. As the old timer is binded to a CPU, when plug out that CPU, kernel
> fails at clockevents.c:653

Clearly tick_install_broadcast_device() lacks serialization.

The untested patch below should cure that.

Thanks,

        tglx
---
 kernel/time/clockevents.c    |   31 +++++++++++++++++++------------
 kernel/time/tick-broadcast.c |   36 ++++++++++++++++++++++--------------
 kernel/time/tick-internal.h  |    2 ++
 3 files changed, 43 insertions(+), 26 deletions(-)

--- a/kernel/time/clockevents.c
+++ b/kernel/time/clockevents.c
@@ -557,23 +557,14 @@ void clockevents_handle_noop(struct cloc
 {
 }
 
-/**
- * clockevents_exchange_device - release and request clock devices
- * @old:	device to release (can be NULL)
- * @new:	device to request (can be NULL)
- *
- * Called from various tick functions with clockevents_lock held and
- * interrupts disabled.
- */
-void clockevents_exchange_device(struct clock_event_device *old,
-				 struct clock_event_device *new)
+void __clockevents_exchange_device(struct clock_event_device *old,
+				   struct clock_event_device *new)
 {
 	/*
 	 * Caller releases a clock event device. We queue it into the
 	 * released list and do a notify add later.
 	 */
 	if (old) {
-		module_put(old->owner);
 		clockevents_switch_state(old, CLOCK_EVT_STATE_DETACHED);
 		list_move(&old->list, &clockevents_released);
 	}
@@ -585,6 +576,22 @@ void clockevents_exchange_device(struct
 }
 
 /**
+ * clockevents_exchange_device - release and request clock devices
+ * @old:	device to release (can be NULL)
+ * @new:	device to request (can be NULL)
+ *
+ * Called from various tick functions with clockevents_lock held and
+ * interrupts disabled.
+ */
+void clockevents_exchange_device(struct clock_event_device *old,
+				 struct clock_event_device *new)
+{
+	if (old)
+		module_put(old->owner);
+	__clockevents_exchange_device(old, new);
+}
+
+/**
  * clockevents_suspend - suspend clock devices
  */
 void clockevents_suspend(void)
@@ -650,7 +657,7 @@ void tick_cleanup_dead_cpu(int cpu)
 		if (cpumask_test_cpu(cpu, dev->cpumask) &&
 		    cpumask_weight(dev->cpumask) == 1 &&
 		    !tick_is_broadcast_device(dev)) {
-			BUG_ON(!clockevent_state_detached(dev));
+			WARN_ON(!clockevent_state_detached(dev));
 			list_del(&dev->list);
 		}
 	}
--- a/kernel/time/tick-broadcast.c
+++ b/kernel/time/tick-broadcast.c
@@ -162,23 +162,31 @@ static bool tick_set_oneshot_wakeup_devi
  */
 void tick_install_broadcast_device(struct clock_event_device *dev, int cpu)
 {
-	struct clock_event_device *cur = tick_broadcast_device.evtdev;
+	struct clock_event_device *cur;
 
-	if (tick_set_oneshot_wakeup_device(dev, cpu))
-		return;
+	scoped_guard(raw_spinlock_irqsave, &tick_broadcast_lock) {
 
-	if (!tick_check_broadcast_device(cur, dev))
-		return;
+		if (tick_set_oneshot_wakeup_device(dev, cpu))
+			return;
 
-	if (!try_module_get(dev->owner))
-		return;
+		cur = tick_broadcast_device.evtdev;
+		if (!tick_check_broadcast_device(cur, dev))
+			return;
 
-	clockevents_exchange_device(cur, dev);
+		if (!try_module_get(dev->owner))
+			return;
+
+		__clockevents_exchange_device(cur, dev);
+		if (cur)
+			cur->event_handler = clockevents_handle_noop;
+		WRITE_ONCE(tick_broadcast_device.evtdev, dev);
+		if (!cpumask_empty(tick_broadcast_mask))
+			tick_broadcast_start_periodic(dev);
+	}
+
+	/* Module release must be outside of the lock */
 	if (cur)
-		cur->event_handler = clockevents_handle_noop;
-	tick_broadcast_device.evtdev = dev;
-	if (!cpumask_empty(tick_broadcast_mask))
-		tick_broadcast_start_periodic(dev);
+		module_put(old->owner);
 
 	if (!(dev->features & CLOCK_EVT_FEAT_ONESHOT))
 		return;
@@ -1185,7 +1193,7 @@ int tick_broadcast_oneshot_active(void)
  */
 bool tick_broadcast_oneshot_available(void)
 {
-	struct clock_event_device *bc = tick_broadcast_device.evtdev;
+	struct clock_event_device *bc = READ_ONCE(tick_broadcast_device.evtdev);
 
 	return bc ? bc->features & CLOCK_EVT_FEAT_ONESHOT : false;
 }
@@ -1193,7 +1201,7 @@ bool tick_broadcast_oneshot_available(vo
 #else
 int __tick_broadcast_oneshot_control(enum tick_broadcast_state state)
 {
-	struct clock_event_device *bc = tick_broadcast_device.evtdev;
+	struct clock_event_device *bc = READ_ONCE(tick_broadcast_device.evtdev);
 
 	if (!bc || (bc->features & CLOCK_EVT_FEAT_HRTIMER))
 		return -EBUSY;
--- a/kernel/time/tick-internal.h
+++ b/kernel/time/tick-internal.h
@@ -53,6 +53,8 @@ static inline void clockevent_set_state(
 }
 
 extern void clockevents_shutdown(struct clock_event_device *dev);
+extern void __clockevents_exchange_device(struct clock_event_device *old,
+					  struct clock_event_device *new);
 extern void clockevents_exchange_device(struct clock_event_device *old,
 					struct clock_event_device *new);
 extern void clockevents_switch_state(struct clock_event_device *dev,