Re: [PATCH v2] rcu-tasks: Make rude RCU-Tasks work well with CPU hotplug

[Joel Fernandes <joel@joelfernandes.org>]

> On Nov 29, 2022, at 2:18 PM, Paul E. McKenney <paulmck@kernel.org> wrote:
> 
> On Tue, Nov 29, 2022 at 11:00:05AM -0500, Joel Fernandes wrote:
>> 
>> 
>>>> On Nov 29, 2022, at 10:18 AM, Paul E. McKenney <paulmck@kernel.org> wrote:
>>> 
>>> On Tue, Nov 29, 2022 at 06:25:04AM +0000, Zhang, Qiang1 wrote:
>>>>>> On Nov 28, 2022, at 11:54 PM, Zhang, Qiang1 <qiang1.zhang@intel.com> wrote:
>>>>> 
>>>>> On Mon, Nov 28, 2022 at 10:34:28PM +0800, Zqiang wrote:
>>>>>> Currently, invoke rcu_tasks_rude_wait_gp() to wait one rude
>>>>>> RCU-tasks grace period, if __num_online_cpus == 1, will return
>>>>>> directly, indicates the end of the rude RCU-task grace period.
>>>>>> suppose the system has two cpus, consider the following scenario:
>>>>>> 
>>>>>>  CPU0                                   CPU1 (going offline)
>>>>>>                        migration/1 task:
>>>>>>                                    cpu_stopper_thread
>>>>>>                                     -> take_cpu_down
>>>>>>                                        -> _cpu_disable
>>>>>>                             (dec __num_online_cpus)
>>>>>>                                        ->cpuhp_invoke_callback
>>>>>>                                              preempt_disable
>>>>>>                      access old_data0
>>>>>>         task1
>>>>>> del old_data0                                  .....
>>>>>> synchronize_rcu_tasks_rude()
>>>>>> task1 schedule out
>>>>>> ....
>>>>>> task2 schedule in
>>>>>> rcu_tasks_rude_wait_gp()
>>>>>>   ->__num_online_cpus == 1
>>>>>>     ->return
>>>>>> ....
>>>>>> task1 schedule in
>>>>>> ->free old_data0
>>>>>>                                              preempt_enable
>>>>>> 
>>>>>> when CPU1 dec __num_online_cpus and __num_online_cpus is equal one,
>>>>>> the CPU1 has not finished offline, stop_machine task(migration/1)
>>>>>> still running on CPU1, maybe still accessing 'old_data0', but the
>>>>>> 'old_data0' has freed on CPU0.
>>>>>> 
>>>>>> This commit add cpus_read_lock/unlock() protection before accessing
>>>>>> __num_online_cpus variables, to ensure that the CPU in the offline
>>>>>> process has been completed offline.
>>>>>> 
>>>>>> Signed-off-by: Zqiang <qiang1.zhang@intel.com>
>>>>>> 
>>>>>> First, good eyes and good catch!!!
>>>>>> 
>>>>>> The purpose of that check for num_online_cpus() is not performance
>>>>>> on single-CPU systems, but rather correct operation during early boot.
>>>>>> So a simpler way to make that work is to check for RCU_SCHEDULER_RUNNING,
>>>>>> for example, as follows:
>>>>>> 
>>>>>>  if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING &&
>>>>>>      num_online_cpus() <= 1)
>>>>>>      return;    // Early boot fastpath for only one CPU.
>>>>> 
>>>>> Hi Paul
>>>>> 
>>>>> During system startup, because the RCU_SCHEDULER_RUNNING is set after starting other CPUs, 
>>>>> 
>>>>>            CPU0                                                                       CPU1                                                                 
>>>>> 
>>>>> if (rcu_scheduler_active !=                                    
>>>>>  RCU_SCHEDULER_RUNNING &&
>>>>>         __num_online_cpus  == 1)                                               
>>>>>  return;                                                                         inc  __num_online_cpus
>>>>>                          (__num_online_cpus == 2)
>>>>> 
>>>>> CPU0 didn't notice the update of the __num_online_cpus variable by CPU1 in time
>>>>> Can we move rcu_set_runtime_mode() before smp_init()
>>>>> any thoughts?
>>>>> 
>>>>> Is anyone expected to do rcu-tasks operation before the scheduler is running? 
>>>> 
>>>> Not sure if such a scenario exists.
>>>> 
>>>>> Typically this requires the tasks to context switch which is a scheduler operation.
>>>>> 
>>>>> If the scheduler is not yet running, then I don’t think missing an update the __num_online_cpus matters since no one does a tasks-RCU synchronize.
>>>> 
>>>> Hi Joel
>>>> 
>>>> After the kernel_init task runs, before calling smp_init() to starting other CPUs, 
>>>> the scheduler haven been initialization, task context switching can occur.
>>> 
>>> Good catch, thank you both.  For some reason, I was thinking that the
>>> additional CPUs did not come online until later.
>>> 
>>> So how about this?
>>> 
>>>   if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
>>>       return;    // Early boot fastpath.
>>> 
>>> If this condition is true, there is only one CPU and no scheduler,
>>> thus no preemption.
>> 
>> Agreed. I was going to suggest exactly this :)
>> 
>> Ack.
>> (Replying by phone but feel free to add my reviewed by tag).
> 
> I should add that the downside of this approach is that there is a short
> time between the scheduler initializing and workqueues fully initializing
> where a critical-path call to synchronize_rcu_tasks() will hang the
> system.  I do -not- consider this to be a real problem because RCU had
> some hundreds of calls to synchronize_rcu() before this became an issue.
> 
> So this should be fine, but please recall this for when/if someone does
> stick a synchronize_rcu_tasks() into that short time.  ;-)

Thanks Paul, but why would anyone want to do sync rcu tasks, before the scheduler is fully initialized? 
Maybe we can add a warning here in the if-early-return path, to make sure no such usage slips. And then we can look into someone using it that way, if they ever start using it.

Thanks,

- Joel

> 
>                            Thanx, Paul
> 
>> - Joel
>> 
>> 
>>> 
>>>                       Thanx, Paul
>>> 
>>>> Thanks
>>>> Zqiang
>>>> 
>>>>> 
>>>>> Or did I miss something?
>>>>> 
>>>>> Thanks.
>>>>> 
>>>>> 
>>>>> 
>>>>> 
>>>>> Thanks
>>>>> Zqiang
>>>>> 
>>>>>> 
>>>>>> This works because rcu_scheduler_active is set to RCU_SCHEDULER_RUNNING
>>>>>> long before it is possible to offline CPUs.
>>>>>> 
>>>>>> Yes, schedule_on_each_cpu() does do cpus_read_lock(), again, good eyes,
>>>>>> and it also unnecessarily does the schedule_work_on() the current CPU,
>>>>>> but the code calling synchronize_rcu_tasks_rude() is on high-overhead
>>>>>> code paths, so this overhead is down in the noise.
>>>>>> 
>>>>>> Until further notice, anyway.
>>>>>> 
>>>>>> So simplicity is much more important than performance in this code.
>>>>>> So just adding the check for RCU_SCHEDULER_RUNNING should fix this,
>>>>>> unless I am missing something (always possible!).
>>>>>> 
>>>>>>                          Thanx, Paul
>>>>>> 
>>>>>> ---
>>>>>> kernel/rcu/tasks.h | 20 ++++++++++++++++++--
>>>>>> 1 file changed, 18 insertions(+), 2 deletions(-)
>>>>>> 
>>>>>> diff --git a/kernel/rcu/tasks.h b/kernel/rcu/tasks.h
>>>>>> index 4a991311be9b..08e72c6462d8 100644
>>>>>> --- a/kernel/rcu/tasks.h
>>>>>> +++ b/kernel/rcu/tasks.h
>>>>>> @@ -1033,14 +1033,30 @@ static void rcu_tasks_be_rude(struct work_struct *work)
>>>>>> {
>>>>>> }
>>>>>> 
>>>>>> +static DEFINE_PER_CPU(struct work_struct, rude_work);
>>>>>> +
>>>>>> // Wait for one rude RCU-tasks grace period.
>>>>>> static void rcu_tasks_rude_wait_gp(struct rcu_tasks *rtp)
>>>>>> {
>>>>>> +    int cpu;
>>>>>> +    struct work_struct *work;
>>>>>> +
>>>>>> +    cpus_read_lock();
>>>>>>  if (num_online_cpus() <= 1)
>>>>>> -        return;    // Fastpath for only one CPU.
>>>>>> +        goto end;// Fastpath for only one CPU.
>>>>>> 
>>>>>>  rtp->n_ipis += cpumask_weight(cpu_online_mask);
>>>>>> -    schedule_on_each_cpu(rcu_tasks_be_rude);
>>>>>> +    for_each_online_cpu(cpu) {
>>>>>> +        work = per_cpu_ptr(&rude_work, cpu);
>>>>>> +        INIT_WORK(work, rcu_tasks_be_rude);
>>>>>> +        schedule_work_on(cpu, work);
>>>>>> +    }
>>>>>> +
>>>>>> +    for_each_online_cpu(cpu)
>>>>>> +        flush_work(per_cpu_ptr(&rude_work, cpu));
>>>>>> +
>>>>>> +end:
>>>>>> +    cpus_read_unlock();
>>>>>> }
>>>>>> 
>>>>>> void call_rcu_tasks_rude(struct rcu_head *rhp, rcu_callback_t func);
>>>>>> -- 
>>>>>> 2.25.1
>>>>>>