Skip to contentSkip to conversation

kernel:time fix race condition in alaramtimer.c

PATCH In review Marcus Gelderie 2 participants 3 messages Jul 6, 2013 – Jul 8, 2013 Timekeeping, clocksource core, NTP, alarmtimer

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 kernel:time fix race condition in alaramtimer.c

This patch fixes a race condition whereby the process can be caused to sleep indefinitely. The problem occurs when the process is preempted after having set its state to TASK_INTERRUPTIBLE but before starting the alarm.

Disabling preemption until the alarm has been set solves this problem.

A second problem may occur if the process reaches the alarm_cancel call while still being in TASK_INTERRUPTIBLE. I do not see how this could possibly happen (if schedule returns, the state is TASK_RUNNING; if schedule is never called, the wakeup callback must have woken the process (data==NULL) thereby setting its state to TASK_RUNNING) but the code is clearly written under the assumption that it can happen. In this case, the state must be set to TASK_RUNNING after the alarm has been canceled for the same reasons as above.

---
 kernel/time/alarmtimer.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
kernel/time/alarmtimer.c +5 -1
index f11d83b..81c8b31 100644
@@ -585,15 +585,19 @@ static int alarmtimer_do_nsleep(struct alarm *alarm, ktime_t absexp)
 {
 	alarm->data = (void *)current;
 	do {
+		preempt_disable();
 		set_current_state(TASK_INTERRUPTIBLE);
 		alarm_start(alarm, absexp);
+		preempt_enable_no_resched();
 		if (likely(alarm->data))
 			schedule();
 
+		preempt_disable();
 		alarm_cancel(alarm);
+		__set_current_state(TASK_RUNNING);
+		preempt_enable_no_resched();
 	} while (alarm->data && !signal_pending(current));
 
-	__set_current_state(TASK_RUNNING);
 
 	return (alarm->data == NULL);
 }
Signature · 1 line
1.8.1.5
#1 lore raw
Quoted 5 lines · Marcus Gelderie

This patch fixes a race condition whereby the process can be caused to sleep indefinitely. The problem occurs when the process is preempted after having set its state to TASK_INTERRUPTIBLE but before starting the alarm.

I don't think there is any such problem. The __schedule() function does this check before dequeuing a task:

if (prev->state && !(preempt_count() & PREEMPT_ACTIVE))

The latter part helps prevent the kind of indefinite sleeping you described above. Also, take a look at preempt_schedule(), preempt_schedule_irq() and friends to see how in-kernel preemption is dealt with, by using PREEMPT_ACTIVE. That would clarify why there is no bug here.

Regards,
Srivatsa S. Bhat

Quoted 30 lines
---
 kernel/time/alarmtimer.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
kernel/time/alarmtimer.c +5 -1
index f11d83b..81c8b31 100644
@@ -585,15 +585,19 @@ static int alarmtimer_do_nsleep(struct alarm *alarm, ktime_t absexp)
 {
 	alarm->data = (void *)current;
 	do {
+		preempt_disable();
 		set_current_state(TASK_INTERRUPTIBLE);
 		alarm_start(alarm, absexp);
+		preempt_enable_no_resched();
 		if (likely(alarm->data))
 			schedule();
 
+		preempt_disable();
 		alarm_cancel(alarm);
+		__set_current_state(TASK_RUNNING);
+		preempt_enable_no_resched();
 	} while (alarm->data && !signal_pending(current));
 
-	__set_current_state(TASK_RUNNING);
 
 	return (alarm->data == NULL);
 }
#2 lore raw

Ok you're right, that prevents it.

Sorry, I missed this check somehow.

Cheers
Marcus

Quoted 51 lines · Srivatsa S. Bhat
Quoted 5 lines · Marcus Gelderie

This patch fixes a race condition whereby the process can be caused to sleep indefinitely. The problem occurs when the process is preempted after having set its state to TASK_INTERRUPTIBLE but before starting the alarm.

I don't think there is any such problem. The __schedule() function does this check before dequeuing a task:

if (prev->state && !(preempt_count() & PREEMPT_ACTIVE))

The latter part helps prevent the kind of indefinite sleeping you described above. Also, take a look at preempt_schedule(), preempt_schedule_irq() and friends to see how in-kernel preemption is dealt with, by using PREEMPT_ACTIVE. That would clarify why there is no bug here.

Regards,
Srivatsa S. Bhat

Quoted 30 lines
---
 kernel/time/alarmtimer.c | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
kernel/time/alarmtimer.c +5 -1
index f11d83b..81c8b31 100644
@@ -585,15 +585,19 @@ static int alarmtimer_do_nsleep(struct alarm *alarm, ktime_t absexp)
 {
 	alarm->data = (void *)current;
 	do {
+		preempt_disable();
 		set_current_state(TASK_INTERRUPTIBLE);
 		alarm_start(alarm, absexp);
+		preempt_enable_no_resched();
 		if (likely(alarm->data))
 			schedule();
 
+		preempt_disable();
 		alarm_cancel(alarm);
+		__set_current_state(TASK_RUNNING);
+		preempt_enable_no_resched();
 	} while (alarm->data && !signal_pending(current));
 
-	__set_current_state(TASK_RUNNING);
 
 	return (alarm->data == NULL);
 }