Re: [RFC PATCH 2/3] sched: membarrier: cover kthread_use_mm (v3)

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

----- On Oct 7, 2020, at 11:07 AM, Peter Zijlstra peterz@infradead.org wrote:

> On Thu, Sep 24, 2020 at 01:25:07PM -0400, Mathieu Desnoyers wrote:
> 
>> diff --git a/kernel/sched/core.c b/kernel/sched/core.c
>> index 2d95dc3f4644..bab6f4f2809f 100644
>> --- a/kernel/sched/core.c
>> +++ b/kernel/sched/core.c
>> @@ -3736,6 +3736,8 @@ context_switch(struct rq *rq, struct task_struct *prev,
>>  	 */
>>  	arch_start_context_switch(prev);
>>  
>> +	membarrier_switch_mm(rq, prev->mm, next->mm);
>> +
>>  	/*
>>  	 * kernel -> kernel   lazy + transfer active
>>  	 *   user -> kernel   lazy + mmgrab() active
>> @@ -3752,7 +3754,6 @@ context_switch(struct rq *rq, struct task_struct *prev,
>>  		else
>>  			prev->active_mm = NULL;
>>  	} else {                                        // to user
>> -		membarrier_switch_mm(rq, prev->active_mm, next->mm);
>>  		/*
>>  		 * sys_membarrier() requires an smp_mb() between setting
>>  		 * rq->curr / membarrier_switch_mm() and returning to userspace.
> 
> I was thinking... do we need the above, when:
> 
>> diff --git a/kernel/sched/membarrier.c b/kernel/sched/membarrier.c
>> index 8bc8b8a888b7..e5246580201b 100644
>> --- a/kernel/sched/membarrier.c
>> +++ b/kernel/sched/membarrier.c
>> @@ -112,13 +112,9 @@ static int membarrier_global_expedited(void)
>>  		    MEMBARRIER_STATE_GLOBAL_EXPEDITED))
>>  			continue;
>>  
>> -		/*
>> -		 * Skip the CPU if it runs a kernel thread. The scheduler
>> -		 * leaves the prior task mm in place as an optimization when
>> -		 * scheduling a kthread.
>> -		 */
>> +		/* Skip the CPU if it runs the idle thread. */
>>  		p = rcu_dereference(cpu_rq(cpu)->curr);
>> -		if (p->flags & PF_KTHREAD)
> 
> We retain this in the form:
> 
>		if ((p->flags & PF_KTHREAD) && !p-mm)
>			continue;
> 
>> +		if (is_idle_task(p))
>>  			continue;
>>  
>>  		__cpumask_set_cpu(cpu, tmpmask);
> 
> Specifically, we only care about kthreads when they're between
> kthread_use_mm() / kthread_unuse_mm(), and in that case they will have
> updated state already.
> 
> It's too late in the day to be sure about the memory ordering though;
> but if we see !->mm, they'll do/have-done switch_mm() which implies
> sufficient barriers().
> 
> Hmm?

Interesting. There are two things we want to ensure here:

1) That we issue an IPI or have the kthread issue the proper barriers when a kthread is
   using/unusing a mm,
2) That we don't issue an IPI to kthreads with NULL mm, so we don't disturb them.

Moving the membarrier_switch_mm to cover kthread cases was to ensure (2), but if we
add a p->mm NULL check in the global expedited iteration, I think we would be OK
leaving the stale runqueue's membarrier state while in lazy tlb state.

As far as (1) is concerned, I think your idea would work, because as you say we will
have the proper barriers in kthread use/unuse mm.

I just wonder whether having this stale membarrier state for lazy tlb is warranted
performance-wise, as it adds complexity: the rq membarrier state will therefore not be
relevant when we are in lazy tlb mode.

Thoughts ?

Thanks,

Mathieu


-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com