On 7/31/19 9:20 PM, luca abeni wrote:
> On Wed, 31 Jul 2019 18:32:47 +0100
> Dietmar Eggemann <[email protected]> wrote:
> [...]
>>>>>> static void dequeue_dl_entity(struct sched_dl_entity *dl_se)
>>>>>> {
>>>>>> + if (!on_dl_rq(dl_se))
>>>>>> + return;
>>>>>
>>>>> Why allow double dequeue instead of WARN?
>>>>
>>>> As I was saying to Valentin, it can currently happen that a task
>>>> could have already been dequeued by update_curr_dl()->throttle
>>>> called by dequeue_task_dl() before calling __dequeue_task_dl(). Do
>>>> you think we should check for this condition before calling into
>>>> dequeue_dl_entity()?
>>>
>>> Yes, that's what ->dl_throttled is for, right? And !->dl_throttled
>>> && !on_dl_rq() is a BUG.
>>
>> OK, I will add the following snippet to the patch.
>> Although it's easy to provoke a situation in which DL tasks are
>> throttled, I haven't seen a throttling happening when the task is
>> being dequeued.
>
> This is a not-so-common situation, that can happen with periodic tasks
> (a-la rt-app) blocking on clock_nanosleep() (or similar) after
> executing for an amount of time comparable with the SCHED_DEADLINE
> runtime.
>
> It might happen that the task consumed a little bit more than the
> remaining runtime (but has not been throttled yet, because the
> accounting happens at every tick)... So, when dequeue_task_dl() invokes
> update_task_dl() the runtime becomes negative and the task is throttled.
>
> This happens infrequently, but if you try rt-app tasksets with multiple
> tasks and execution times near to the runtime you will see it
> happening, sooner or later.
>
>
> [...]
>> @@ -1592,6 +1591,10 @@ static void __dequeue_task_dl(struct rq *rq,
>> struct task_struct *p) static void dequeue_task_dl(struct rq *rq,
>> struct task_struct *p, int flags) {
>> update_curr_dl(rq);
>> +
>> + if (p->dl.dl_throttled)
>> + return;
>
> Sorry, I missed part of the previous discussion, so maybe I am missing
> something... But I suspect this "return" might be wrong (you risk to
> miss a call to task_non_contending(), coming later in this function).
>
> Maybe you cound use
> if (!p->dl_throttled)
> __dequeue_task_dl(rq, p)
>
I see. With the following rt-app file on h960 (8 CPUs) I'm able to
recreate the situation relatively frequently.
...
"tasks" : {
"thread0" : {
"instance" : 12,
"run" : 11950,
"timer" : { "ref" : "unique", "period" : 100000, "mode" : "absolute"},
"dl-runtime" : 12000,
"dl-period" : 100000,
"dl-deadline" : 100000
}
}
...
[ 1912.086664] CPU1: p=[thread0-9 3070] throttled p->on_rq=0 flags=0x9
[ 1912.086726] CPU2: p=[thread0-10 3071] throttled p->on_rq=0 flags=0x9
[ 1924.738912] CPU6: p=[thread0-10 3149] throttled p->on_rq=0 flags=0x9
...
And the flag DEQUEUE_SLEEP is set so like you said
task_non_contending(p) should be called.
I'm going to use your proposal. Thank you for the help!
