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.
Signed-off-by: Marcus Gelderie <[email protected]>
---
kernel/time/alarmtimer.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/kernel/time/alarmtimer.c b/kernel/time/alarmtimer.c
index f11d83b..81c8b31 100644
--- a/kernel/time/alarmtimer.c
+++ b/kernel/time/alarmtimer.c
@@ -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);
}
--
1.8.1.5
On 07/06/2013 09:37 PM, Marcus Gelderie wrote:
> 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
> ---
> kernel/time/alarmtimer.c | 6 +++++-
> 1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/kernel/time/alarmtimer.c b/kernel/time/alarmtimer.c
> index f11d83b..81c8b31 100644
> --- a/kernel/time/alarmtimer.c
> +++ b/kernel/time/alarmtimer.c
> @@ -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);
> }
>
Ok you're right, that prevents it.
Sorry, I missed this check somehow.
Cheers
Marcus
On 07/08/2013 07:31 AM, Srivatsa S. Bhat wrote:
> On 07/06/2013 09:37 PM, Marcus Gelderie wrote:
>> 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
>
>> ---
>> kernel/time/alarmtimer.c | 6 +++++-
>> 1 file changed, 5 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/time/alarmtimer.c b/kernel/time/alarmtimer.c
>> index f11d83b..81c8b31 100644
>> --- a/kernel/time/alarmtimer.c
>> +++ b/kernel/time/alarmtimer.c
>> @@ -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);
>> }
>>
>