________________________________________
发件人: Uladzislau Rezki <urezki@xxxxxxxxx>
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@xxxxxxxxx>
> >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 rq's 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.
sched_cpu_wait_empty()
{
rcuwait_wait_event(&rq->hotplug_wait,
rq->nr_running == 1 && !rq_has_pinned_tasks(rq),
TASK_UNINTERRUPTIBLE);
}
>
> How about the following changes:
>
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index e7a226abff0d..2aa19537ac7c 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -3488,12 +3488,10 @@ add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
> (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
> bnode = get_cached_bnode(*krcp);
> if (!bnode && can_alloc) {
> - migrate_disable();
> krc_this_cpu_unlock(*krcp, *flags);
> bnode = (struct kvfree_rcu_bulk_data *)
> __get_free_page(GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOMEMALLOC | __GFP_NOWARN);
> - *krcp = krc_this_cpu_lock(flags);
> - migrate_enable();
> + raw_spin_lock_irqsave(&(*krcp)->lock, *flags);
>
Hm.. Taking the former lock can lead to a pointer leaking, i mean a CPU associated
with "krcp" might go offline during a page request process, so a queuing occurs on
off-lined CPU. Apat of that, acquiring a former lock still does not solve:
- CPU1 in process of page allocation;
- CPU1 gets migrated to CPU2;
- another task running on CPU1 also allocate a page;
- both bnodes are added to krcp associated with CPU1.
I agree that such scenario probably will never happen or i would say, can be
considered as a corner case. We can drop the:
[PATCH 3/3] kvfree_rcu: use migrate_disable/enable()
and live with: an allocated bnode can be queued to another CPU, so its prev.
"bnode" can be underutilized. What is also can be considered as a corner case.
According to my tests, it is hard to achieve:
Running kvfree_rcu() simultaneously in a tight loop, 1 000 000 allocations/freeing:
- 64 CPUs and 64 threads showed 1 migration;
- 64 CPUs and 128 threads showed 0 migrations;
- 64 CPUs and 32 threads showed 0 migration.
Thoughts?
Thank you for your comments!
--
Vlad Rezki
