Re: 回复: 回复: [PATCH 3/3] kvfree_rcu: use migrate_disable/enable()

[Uladzislau Rezki <urezki@gmail.com>]

> 
> ________________________________________
> 发件人: Uladzislau Rezki <urezki@gmail.com>
> 发送时间: 2021年1月25日 5:57
> 收件人: Zhang, Qiang
> 抄送: Uladzislau Rezki (Sony); LKML; RCU; Paul E . McKenney; Michael Ellerman; Andrew Morton; Daniel Axtens; Frederic Weisbecker; Neeraj Upadhyay; Joel Fernandes; Peter Zijlstra; Michal Hocko; Thomas Gleixner; Theodore Y . Ts'o; Sebastian Andrzej Siewior; Oleksiy Avramchenko
> 主题: Re: 回复: [PATCH 3/3] kvfree_rcu: use migrate_disable/enable()
> 
> >Hello, Zhang.
> 
> > >________________________________________
> > >发件人: Uladzislau Rezki (Sony) <urezki@gmail.com>
> > >发送时间: 2021年1月21日 0:21
> > >收件人: LKML; RCU; Paul E . McKenney; Michael Ellerman
> > >抄送: Andrew Morton; Daniel Axtens; Frederic Weisbecker; Neeraj >Upadhyay; Joel Fernandes; Peter Zijlstra; Michal Hocko; Thomas >Gleixner; Theodore Y . Ts'o; Sebastian Andrzej Siewior; Uladzislau >Rezki; Oleksiy Avramchenko
> > >主题: [PATCH 3/3] kvfree_rcu: use migrate_disable/enable()
> > >
> > >Since the page is obtained in a fully preemptible context, dropping
> > >the lock can lead to migration onto another CPU. As a result a prev.
> > >bnode of that CPU may be underutilised, because a decision has been
> > >made for a CPU that was run out of free slots to store a pointer.
> > >
> > >migrate_disable/enable() are now independent of RT, use it in order
> > >to prevent any migration during a page request for a specific CPU it
> > >is requested for.
> >
> >
> > Hello Rezki
> >
> > The critical migrate_disable/enable() area is not allowed to block, under RT and non RT.
> > There is such a description in preempt.h
> >
> >
> > * Notes on the implementation.
> >  *
> >  * The implementation is particularly tricky since existing code patterns
> >  * dictate neither migrate_disable() nor migrate_enable() is allowed to block.
> >  * This means that it cannot use cpus_read_lock() to serialize against hotplug,
> >  * nor can it easily migrate itself into a pending affinity mask change on
> >  * migrate_enable().
> >
> >How i interpret it is migrate_enable()/migrate_disable() are not allowed to
> >use any blocking primitives, such as rwsem/mutexes/etc. in order to mark a
> >current context as non-migratable.
> >
> >void migrate_disable(void)
> >{
> > struct task_struct *p = current;
> >
> > if (p->migration_disabled) {
> >  p->migration_disabled++;
> >  return;
> > }
> 
> > preempt_disable();
> > this_rq()->nr_pinned++;
> > p->migration_disabled = 1;
> > preempt_enable();
> >}
> >
> >It does nothing that prevents you from doing schedule() or even wait for any
> >event(mutex slow path behaviour), when the process is removed from the run-queue.
> >I mean after the migrate_disable() is invoked. Or i miss something?
> 
> Hello Rezki
> 
> Sorry, there's something wrong with the previous description.
> There are the following scenarios
> 
> Due to migrate_disable will increase  this_rq()->nr_pinned , after that
> if get_free_page be blocked, and this time, CPU going offline,
> the sched_cpu_wait_empty() be called in per-cpu "cpuhp/%d" task,
> and be blocked.
> 
But after the migrate_disable() is invoked a CPU can not be brought down.
If there are pinned tasks a "hotplug path" will be blocked on balance_hotplug_wait()
call.

> blocked:
> sched_cpu_wait_empty()
> {
>       struct rq *rq = this_rq();
>        rcuwait_wait_event(&rq->hotplug_wait,
>                            rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
>                            TASK_UNINTERRUPTIBLE);
> }
>
Exactly.

> wakeup:
> balance_push()
> {
>         if (is_per_cpu_kthread(push_task) || is_migration_disabled(push_task)) {
>               
>                 if (!rq->nr_running && !rq_has_pinned_tasks(rq) &&
>                     rcuwait_active(&rq->hotplug_wait)) {
>                         raw_spin_unlock(&rq->lock);
>                         rcuwait_wake_up(&rq->hotplug_wait);
>                         raw_spin_lock(&rq->lock);
>                 }
>                 return;
>         }
> }
> 
> One of the conditions for this function to wake up is "rq->nr_pinned  == 0"
> that is to say between migrate_disable/enable, if blocked will defect CPU going
> offline longer blocking time.
> 
Indeed, the hotplug time is affected. For example in case of waiting for
a mutex to be released, an owner will wakeup waiters. But this is expectable.

>
> I'm not sure that's a problem，and I didn't find it in the kernel code  between 
>  migrate_disable/enable possible sleep calls.
> 
For example z3fold.c:

/* Add to the appropriate unbuddied list */
static inline void add_to_unbuddied(struct z3fold_pool *pool,
				struct z3fold_header *zhdr)
{
	if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
			zhdr->middle_chunks == 0) {
		struct list_head *unbuddied;
		int freechunks = num_free_chunks(zhdr);

		migrate_disable();
		unbuddied = this_cpu_ptr(pool->unbuddied);
		spin_lock(&pool->lock);
		list_add(&zhdr->buddy, &unbuddied[freechunks]);
		spin_unlock(&pool->lock);
		zhdr->cpu = smp_processor_id();
		migrate_enable();
	}
}

for PREEMPT_RT kernel a spinlock is converted to rt-mutex, thus it can sleep.

--
Vlad Rezki