Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1476140 > unrolled thread

[PATCH] sched/core: simpler function for sched_exec migration

Back to article view | Back to linux.kernel

Contents

  [PATCH] sched/core: simpler function for sched_exec migration cheng chao <chengchao@kedacom.com> - 2016-09-05 08:30 +0200
    Re: [PATCH] sched/core: simpler function for sched_exec migration Oleg Nesterov <oleg@redhat.com> - 2016-09-05 15:20 +0200
      Re: [PATCH] sched/core: simpler function for sched_exec migration chengchao <chengchao@kedacom.com> - 2016-09-06 04:20 +0200
        Re: [PATCH] sched/core: simpler function for sched_exec migration Oleg Nesterov <oleg@redhat.com> - 2016-09-06 17:30 +0200
          Re: [PATCH] sched/core: simpler function for sched_exec migration chengchao <chengchao@kedacom.com> - 2016-09-07 05:30 +0200
            Re: [PATCH] sched/core: simpler function for sched_exec migration Oleg Nesterov <oleg@redhat.com> - 2016-09-07 14:40 +0200
              Re: [PATCH] sched/core: simpler function for sched_exec migration chengchao <chengchao@kedacom.com> - 2016-09-08 04:20 +0200
              Re: [PATCH] sched/core: simpler function for sched_exec migration Peter Zijlstra <peterz@infradead.org> - 2016-09-09 12:10 +0200

#1476140 — [PATCH] sched/core: simpler function for sched_exec migration

when sched_exec needs migration and CONFIG_PREEMPT_NONE=y,
migration_cpu_stop almost does nothing due to
the caller is !task_on_rq_queued().

currently CONFIG_PREEMPT and CONFIG_PREEMPT_VOLUNTARY work well because
the caller keeps task_on_rq_queued():

1. when CONFIG_PREEMPT=y
  stop_one_cpu
  ->cpu_stop_queue_work
    ->spin_unlock_irqrestore (preempt_enable calls __preempt_schedule)

2. when CONFIG_PREEMPT_VOLUNTARY=y
  stop_one_cpu
  ->wait_for_completion
    ->...
      -->might_sleep() (calls _cond_resched()

stop_one_cpu_sync is introduced here to address this problem,more further
it makes more simpler for CONFIG_PREEMPT=y or CONFIG_PREEMPT_VOLUNTARY=y
when sched_exec needs migration.

Signed-off-by: cheng chao <chengchao@kedacom.com>
---
 include/linux/stop_machine.h |  1 +
 kernel/sched/core.c          |  2 +-
 kernel/stop_machine.c        | 21 +++++++++++++++++++++
 3 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/include/linux/stop_machine.h b/include/linux/stop_machine.h
index 3cc9632..e4e7d42 100644
--- a/include/linux/stop_machine.h
+++ b/include/linux/stop_machine.h
@@ -28,6 +28,7 @@ struct cpu_stop_work {
 };
 
 int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg);
+void stop_one_cpu_sync(unsigned int cpu, cpu_stop_fn_t fn, void *arg);
 int stop_two_cpus(unsigned int cpu1, unsigned int cpu2, cpu_stop_fn_t fn, void *arg);
 bool stop_one_cpu_nowait(unsigned int cpu, cpu_stop_fn_t fn, void *arg,
 			 struct cpu_stop_work *work_buf);
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 556cb07..2fd71e6 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -2958,7 +2958,7 @@ void sched_exec(void)
 		struct migration_arg arg = { p, dest_cpu };
 
 		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
-		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
+		stop_one_cpu_sync(task_cpu(p), migration_cpu_stop, &arg);
 		return;
 	}
 unlock:
diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
index 4a1ca5f..24f8637 100644
--- a/kernel/stop_machine.c
+++ b/kernel/stop_machine.c
@@ -130,6 +130,27 @@ int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
 	return done.ret;
 }
 
+/**
+ * the caller keeps task_on_rq_queued, so it's more suitable for
+ * sched_exec on the case when needs migration
+ */
+void stop_one_cpu_sync(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
+{
+	struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = NULL };
+
+	if (!cpu_stop_queue_work(cpu, &work))
+		return;
+
+#if defined(CONFIG_PREEMPT_NONE) || defined(CONFIG_PREEMPT_VOLUNTARY)
+	/*
+	 * CONFIG_PREEMPT doesn't need call schedule here, because
+	 * preempt_enable already does the similar thing when call
+	 * cpu_stop_queue_work
+	 */
+	schedule();
+#endif
+}
+
 /* This controls the threads on each CPU. */
 enum multi_stop_state {
 	/* Dummy starting state for thread. */
-- 
2.4.11

[toc] | [next] | [standalone]

#1476410

On 09/05, cheng chao wrote:
>
> @@ -2958,7 +2958,7 @@ void sched_exec(void)
>  		struct migration_arg arg = { p, dest_cpu };
>  
>  		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
> -		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
> +		stop_one_cpu_sync(task_cpu(p), migration_cpu_stop, &arg);
>  		return;
>  	}
>  unlock:
> diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
> index 4a1ca5f..24f8637 100644
> --- a/kernel/stop_machine.c
> +++ b/kernel/stop_machine.c
> @@ -130,6 +130,27 @@ int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
>  	return done.ret;
>  }
>  
> +/**
> + * the caller keeps task_on_rq_queued, so it's more suitable for
> + * sched_exec on the case when needs migration
> + */
> +void stop_one_cpu_sync(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
> +{
> +	struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = NULL };
> +
> +	if (!cpu_stop_queue_work(cpu, &work))
> +		return;
> +
> +#if defined(CONFIG_PREEMPT_NONE) || defined(CONFIG_PREEMPT_VOLUNTARY)
> +	/*
> +	 * CONFIG_PREEMPT doesn't need call schedule here, because
> +	 * preempt_enable already does the similar thing when call
> +	 * cpu_stop_queue_work
> +	 */
> +	schedule();
> +#endif
> +}

Honestly, I don't really understand the changelog, but this looks wrong.

stop_one_cpu_sync() assumes that cpu == smp_processor_id/task_cpu(current),
and thus the stopper thread should preempt us at least after schedule()
(if CONFIG_PREEMPT_NONE), so we do not need to synchronize.

But this is not necessarily true? This task can migrate to another CPU
before cpu_stop_queue_work() ?

Oleg.

[toc] | [prev] | [next] | [standalone]

#1477021

Oleg, thank you.

the key point is for CONFIG_PREEMPT_NONE=y, if sched_exec needs migrate the current, 
migration_cpu_stop doesn't migrate the task(current) at all, it means that the stopper thread does some 
unuseful works in this scenario.
finally,the stopper thread calls cpu_stop_signal_done() to wake up this task, it calls select_task_rq() again, 
maybe select another different cpu. totally calls select_task_rq() two times(first at sched_exec())
plus one time(wake_up_new_task() also calls select_task_rq()).

it is too much overhead for one task(fork()+exec()), isn't it?



1.
sched_exec()
  ->stop_one_cpu()
    ->wait_for_completion().
     wait_for_completion() makes the current TASK_UNINTERRUPTIBLE and call schedule_timeout()

schedule_timeout(timeout)  timeout is MAX_SCHEDULE_TIMEOUT.
  ->schedule()
	deactivate_task(rq, current, DEQUEUE_SLEEP);
	current->on_rq = 0;

2.
migration_cpu_stop() checks the task_on_rq_queued(p), but the task p->on_rq is 0.

#define TASK_ON_RQ_QUEUED       1

static inline int task_on_rq_queued(struct task_struct *p)
{
        return p->on_rq == TASK_ON_RQ_QUEUED;
}


migration_cpu_stop()
...
        if (task_rq(p) == rq && task_on_rq_queued(p))   
                rq = __migrate_task(rq, p, arg->dest_cpu);
...


thanks again, any suggestions and more reviews are welcome.


on 09/05/2016 09:11 PM, Oleg Nesterov wrote:
> On 09/05, cheng chao wrote:
>>
>> @@ -2958,7 +2958,7 @@ void sched_exec(void)
>>  		struct migration_arg arg = { p, dest_cpu };
>>  
>>  		raw_spin_unlock_irqrestore(&p->pi_lock, flags);
>> -		stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
>> +		stop_one_cpu_sync(task_cpu(p), migration_cpu_stop, &arg);
>>  		return;
>>  	}
>>  unlock:
>> diff --git a/kernel/stop_machine.c b/kernel/stop_machine.c
>> index 4a1ca5f..24f8637 100644
>> --- a/kernel/stop_machine.c
>> +++ b/kernel/stop_machine.c
>> @@ -130,6 +130,27 @@ int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
>>  	return done.ret;
>>  }
>>  
>> +/**
>> + * the caller keeps task_on_rq_queued, so it's more suitable for
>> + * sched_exec on the case when needs migration
>> + */
>> +void stop_one_cpu_sync(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
>> +{
>> +	struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = NULL };
>> +
>> +	if (!cpu_stop_queue_work(cpu, &work))
>> +		return;
>> +
>> +#if defined(CONFIG_PREEMPT_NONE) || defined(CONFIG_PREEMPT_VOLUNTARY)
>> +	/*
>> +	 * CONFIG_PREEMPT doesn't need call schedule here, because
>> +	 * preempt_enable already does the similar thing when call
>> +	 * cpu_stop_queue_work
>> +	 */
>> +	schedule();
>> +#endif
>> +}
> 
> Honestly, I don't really understand the changelog, but this looks wrong.
> 
> stop_one_cpu_sync() assumes that cpu == smp_processor_id/task_cpu(current),
> and thus the stopper thread should preempt us at least after schedule()
> (if CONFIG_PREEMPT_NONE), so we do not need to synchronize.
> 
   yes. the stop_one_cpu_sync is not a good name, stop_one_cpu_schedule is better?  
there is nothing about synchronization.

> But this is not necessarily true? This task can migrate to another CPU
> before cpu_stop_queue_work() ?
>
  before sched_exec() calls stop_one_cpu()/cpu_stop_queue_work(), this task(current) cannot migrate 
to another cpu,because this task is running on the cpu.

 
> Oleg.
> 
> 

[toc] | [prev] | [next] | [standalone]

#1477533

On 09/06, chengchao wrote:
>
> the key point is for CONFIG_PREEMPT_NONE=y,
> ...
> it is too much overhead for one task(fork()+exec()), isn't it?

Yes, yes, I see, this is suboptimal. Not sure we actually do care,
but yes, perhaps another helper which migrates the current task makes
sense, I dunno.

But,

> > stop_one_cpu_sync() assumes that cpu == smp_processor_id/task_cpu(current),
> > and thus the stopper thread should preempt us at least after schedule()
> > (if CONFIG_PREEMPT_NONE), so we do not need to synchronize.
> >
>    yes. the stop_one_cpu_sync is not a good name, stop_one_cpu_schedule is better?  
> there is nothing about synchronization.

We need to synchronize with the stopper to ensure it can't touch
cpu_stop_work on stack after stop_one_cpu_sync() returns, and

> > But this is not necessarily true? This task can migrate to another CPU
> > before cpu_stop_queue_work() ?
> >
>   before sched_exec() calls stop_one_cpu()/cpu_stop_queue_work(), this
> task(current) cannot migrate  to another cpu,because this task is running
> on the cpu.

Why? The running task can migrate to another CPU at any moment. Unless it
runs with preemption disabled or CONFIG_PREEMPT_NONE=y.

And this means that cpu_stop_queue_work() can queue the work on another
CPU != smp_processor_id(), and in this case the kernel can crash because
the pending cpu_stop_work can be overwritten right after return.

So you need something like

	void stop_one_cpu_sync(cpu_stop_fn_t fn, void *arg)
	{
		struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = NULL };

		preempt_disable();
		cpu_stop_queue_work(raw_smp_processor_id(), &work);
		preempt_enable_no_resched();
		schedule();
	}

or I am totally confused. Note that it doesn't (and shouldn't) have
the "int cpu" argument.

Oleg.

[toc] | [prev] | [next] | [standalone]

#1477945

Oleg, thank you very much.

on 09/06/2016 11:22 PM, Oleg Nesterov wrote:
> On 09/06, chengchao wrote:
>>
>> the key point is for CONFIG_PREEMPT_NONE=y,
>> ...
>> it is too much overhead for one task(fork()+exec()), isn't it?
> 
> Yes, yes, I see, this is suboptimal. Not sure we actually do care,
> but yes, perhaps another helper which migrates the current task makes
> sense, I dunno.

for CONFIG_PREEMPT_NONE=y, this patch wants the stopper thread can migrate the current 
successfully instead of doing nothing.

> 
> But,
> 
>>> stop_one_cpu_sync() assumes that cpu == smp_processor_id/task_cpu(current),
>>> and thus the stopper thread should preempt us at least after schedule()
>>> (if CONFIG_PREEMPT_NONE), so we do not need to synchronize.
>>>
>>    yes. the stop_one_cpu_sync is not a good name, stop_one_cpu_schedule is better?  
>> there is nothing about synchronization.
> 
> We need to synchronize with the stopper to ensure it can't touch
> cpu_stop_work on stack after stop_one_cpu_sync() returns, and

yes, you are right.

> 
>>> But this is not necessarily true? This task can migrate to another CPU
>>> before cpu_stop_queue_work() ?
>>>
>>   before sched_exec() calls stop_one_cpu()/cpu_stop_queue_work(), this
>> task(current) cannot migrate  to another cpu,because this task is running
>> on the cpu.
> 
> Why? The running task can migrate to another CPU at any moment. Unless it
> runs with preemption disabled or CONFIG_PREEMPT_NONE=y.

yes, this patch focused the CONFIG_PREEMPT_NONE=y at the beginning, so I didn't
pay more attention to the CONFIG_PREEMPT=y and CONFIG_PREEMPT_VOLUNTARY=y.

> 
> And this means that cpu_stop_queue_work() can queue the work on another
> CPU != smp_processor_id(), and in this case the kernel can crash because
> the pending cpu_stop_work can be overwritten right after return.
> 
> So you need something like
> 
> 	void stop_one_cpu_sync(cpu_stop_fn_t fn, void *arg)
> 	{
> 		struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = NULL };
> 
> 		preempt_disable();
> 		cpu_stop_queue_work(raw_smp_processor_id(), &work);
> 		preempt_enable_no_resched();
> 		schedule();
> 	}
>
 
> or I am totally confused. Note that it doesn't (and shouldn't) have
> the "int cpu" argument.
> 


if preempt happens after preempt_enable_no_resched(), there is still risky that the 
stop_one_cpu_sync() returns before the stopper thread can use cpu_stop_work safely.
as you said previously.

thus, I modify the patch:

int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
{
        struct cpu_stop_done done;
        struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };

        cpu_stop_init_done(&done, 1);
        if (!cpu_stop_queue_work(cpu, &work))
                return -ENOENT;

#if defined(CONFIG_PREEMPT_NONE)
	/*
         * let the stopper thread runs as soon as possible,
         * and keep current TASK_RUNNING.
         */
	scheudle();
#endif	
        wait_for_completion(&done.completion);
        return done.ret;
}

remove the new function stop_one_cpu_sync(). When I posted this patch, I didn't want to
modify the stop_one_cpu(), because there are many functions to call the stop_one_cpu().
but now, I think it's good place to modify.

Any suggestions? thanks again.


> Oleg.
> 
> 

[toc] | [prev] | [next] | [standalone]

#1478253

On 09/07, chengchao wrote:
>
> Oleg, thank you very much.
>
> on 09/06/2016 11:22 PM, Oleg Nesterov wrote:
> > On 09/06, chengchao wrote:
> >>
> >> the key point is for CONFIG_PREEMPT_NONE=y,
> >> ...
> >> it is too much overhead for one task(fork()+exec()), isn't it?
> >
> > Yes, yes, I see, this is suboptimal. Not sure we actually do care,
> > but yes, perhaps another helper which migrates the current task makes
> > sense, I dunno.
>
> for CONFIG_PREEMPT_NONE=y, this patch wants the stopper thread can migrate the current
> successfully instead of doing nothing.

I understand the intent. But I am not sure this optimization makes
sense.

> > So you need something like
> >
> > 	void stop_one_cpu_sync(cpu_stop_fn_t fn, void *arg)
> > 	{
> > 		struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = NULL };
> >
> > 		preempt_disable();
> > 		cpu_stop_queue_work(raw_smp_processor_id(), &work);
> > 		preempt_enable_no_resched();
> > 		schedule();
> > 	}
> >
>
> > or I am totally confused. Note that it doesn't (and shouldn't) have
> > the "int cpu" argument.
> >
>
>
> if preempt happens after preempt_enable_no_resched(),

This doesn't differ from explicit schedule() call. Either way the
stopper thread will preempt us on the same CPU.

> there is still risky that the
> stop_one_cpu_sync() returns before the stopper thread can use cpu_stop_work safely.
> as you said previously.

No.


However, there is another problem. It can race with another
stop_one_cpu(migration_cpu_stop) which comes between preempt_disable()
and cpu_stop_queue_work(). So the caller still can migrate to another
CPU right after after preempt_enable_no_resched() and run before the
stopper thread completes the cpu_stop_work queued by us.

> int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
> {
>         struct cpu_stop_done done;
>         struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
>
>         cpu_stop_init_done(&done, 1);
>         if (!cpu_stop_queue_work(cpu, &work))
>                 return -ENOENT;
>
> #if defined(CONFIG_PREEMPT_NONE)
> 	/*
>          * let the stopper thread runs as soon as possible,
>          * and keep current TASK_RUNNING.
>          */
> 	scheudle();
> #endif
>         wait_for_completion(&done.completion);
>         return done.ret;
> }

Agreed this looks better, although I'd suggest _cond_resche().

Again, I am not sure this makes sense, I leave this to maintainers.

Oleg.

[toc] | [prev] | [next] | [standalone]

#1478735

Oled, thank you for moving this patch on.

on 09/07/2016 08:35 PM, Oleg Nesterov wrote:
> On 09/07, chengchao wrote:
>>
>> Oleg, thank you very much.
>>
>> on 09/06/2016 11:22 PM, Oleg Nesterov wrote:
>>> On 09/06, chengchao wrote:
>>>>
>>>> the key point is for CONFIG_PREEMPT_NONE=y,
>>>> ...
>>>> it is too much overhead for one task(fork()+exec()), isn't it?
>>>
>>> Yes, yes, I see, this is suboptimal. Not sure we actually do care,
>>> but yes, perhaps another helper which migrates the current task makes
>>> sense, I dunno.
>>
>> for CONFIG_PREEMPT_NONE=y, this patch wants the stopper thread can migrate the current
>> successfully instead of doing nothing.
> 
> I understand the intent. But I am not sure this optimization makes
> sense.
> 

For CONFIG_PREEMPT_NONE=y, when sched_exec() needs migration, sched_exec() calls
stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg).

If stopper thread can not migrate for us,why should we call stop_one_cpu() here? 
It just makes the task TASK_UNINTERRUPTIBLE, wakes up the stopper thread, executes the 
migration_cpu_stop, and the stopper thread wakes up the task.

But in fact, all above works are almost unuseful, the reason is that the migration_cpu_stop
doesn't migrate for us. why? the migration_cpu_stop() needs the task is TASK_ON_RQ_QUEUED 
before it calls __migrate_task().

This patch can make the task TASK_RUNNING instead of TASK_UNINTERRUPTIBLE,
so the migration_cpu_stop() can migrate happily.

Does this optimization make sense now?

Any different opinions are always welcome.

>> int stop_one_cpu(unsigned int cpu, cpu_stop_fn_t fn, void *arg)
>> {
>>         struct cpu_stop_done done;
>>         struct cpu_stop_work work = { .fn = fn, .arg = arg, .done = &done };
>>
>>         cpu_stop_init_done(&done, 1);
>>         if (!cpu_stop_queue_work(cpu, &work))
>>                 return -ENOENT;
>>
>> #if defined(CONFIG_PREEMPT_NONE)
>> 	   /*
>>          * let the stopper thread runs as soon as possible,
>>          * and keep current TASK_RUNNING.
>>          */
>> 	scheudle();
>> #endif
>>         wait_for_completion(&done.completion);
>>         return done.ret;
>> }
> 
> Agreed this looks better, although I'd suggest _cond_resche().
> 
> Again, I am not sure this makes sense, I leave this to maintainers.
> 

You have done much works for this patch. Thanks again. 

> Oleg.
> 
> 

[toc] | [prev] | [next] | [standalone]

#1479821

On Wed, Sep 07, 2016 at 02:35:11PM +0200, Oleg Nesterov wrote:
> Again, I am not sure this makes sense, I leave this to maintainers.

FWIW, I only received the emails from Oleg. So if one wants a maintainer
to look at this one has to fix his mail setup such that the emails
arrive at his Inbox.

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web