Re: [PATCH rcu 3/3] rcu: Allow expedited RCU grace periods on incoming CPUs

[Neeraj Upadhyay <quic_neeraju@quicinc.com>]

Hi Mukesh,

On 2/12/2022 2:17 PM, Mukesh Ojha wrote:
> 
> On 2/12/2022 3:44 AM, Paul E. McKenney wrote:
>> On Sat, Feb 12, 2022 at 12:14:20AM +0530, Mukesh Ojha wrote:
>>> On 2/10/2022 3:36 AM, Paul E. McKenney wrote:
>>>> On Wed, Feb 09, 2022 at 11:53:33PM +0530, Mukesh Ojha wrote:
>>>>> On 2/5/2022 4:25 AM, Paul E. McKenney wrote:
>>>>>> Although it is usually safe to invoke synchronize_rcu_expedited() 
>>>>>> from a
>>>>>> preemption-enabled CPU-hotplug notifier, if it is invoked from a 
>>>>>> notifier
>>>>>> between CPUHP_AP_RCUTREE_ONLINE and CPUHP_AP_ACTIVE, its attempts to
>>>>>> invoke a workqueue handler will hang due to RCU waiting on a CPU that
>>>>>> the scheduler is not paying attention to.  This commit therefore 
>>>>>> expands
>>>>>> use of the existing workqueue-independent synchronize_rcu_expedited()
>>>>>> from early boot to also include CPUs that are being hotplugged.
>>>>>>
>>>>>> Link: 
>>>>>> https://lore.kernel.org/lkml/7359f994-8aaf-3cea-f5cf-c0d3929689d6@quicinc.com/ 
>>>>>>
>>>>>> Reported-by: Mukesh Ojha <quic_mojha@quicinc.com>
>>>>>> Cc: Tejun Heo <tj@kernel.org>
>>>>>> Signed-off-by: Paul E. McKenney <paulmck@kernel.org>
>>>>>> ---
>>>>>>     kernel/rcu/tree_exp.h | 14 ++++++++++----
>>>>>>     1 file changed, 10 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
>>>>>> index 60197ea24ceb9..1a45667402260 100644
>>>>>> --- a/kernel/rcu/tree_exp.h
>>>>>> +++ b/kernel/rcu/tree_exp.h
>>>>>> @@ -816,7 +816,7 @@ static int rcu_print_task_exp_stall(struct 
>>>>>> rcu_node *rnp)
>>>>>>      */
>>>>>>     void synchronize_rcu_expedited(void)
>>>>>>     {
>>>>>> -    bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
>>>>>> +    bool no_wq;
>>>>>>         struct rcu_exp_work rew;
>>>>>>         struct rcu_node *rnp;
>>>>>>         unsigned long s;
>>>>>> @@ -841,9 +841,15 @@ void synchronize_rcu_expedited(void)
>>>>>>         if (exp_funnel_lock(s))
>>>>>>             return;  /* Someone else did our work for us. */
>>>>>> +    /* Don't use workqueue during boot or from an incoming CPU. */
>>>>>> +    preempt_disable();
>>>>>> +    no_wq = rcu_scheduler_active == RCU_SCHEDULER_INIT ||
>>>>>> +        !cpumask_test_cpu(smp_processor_id(), cpu_active_mask);
>>>>>> +    preempt_enable();
>>>>>> +
>>>>>>         /* Ensure that load happens before action based on it. */
>>>>>> -    if (unlikely(boottime)) {
>>>>>> -        /* Direct call during scheduler init and 
>>>>>> early_initcalls(). */
>>>>>> +    if (unlikely(no_wq)) {
>>>>>> +        /* Direct call for scheduler init, early_initcall()s, and 
>>>>>> incoming CPUs. */
>>>>>>             rcu_exp_sel_wait_wake(s);
>>>>>>         } else {
>>>>>>             /* Marshall arguments & schedule the expedited grace 
>>>>>> period. */
>>>>>> @@ -861,7 +867,7 @@ void synchronize_rcu_expedited(void)
>>>>>>         /* Let the next expedited grace period start. */
>>>>>>         mutex_unlock(&rcu_state.exp_mutex);
>>>>>> -    if (likely(!boottime))
>>>>>> +    if (likely(!no_wq))
>>>>>>             destroy_work_on_stack(&rew.rew_work);
>>>>>>     }
>>>>>>     EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
>>>>> Can we reach a condition after this change where no_wq = true and 
>>>>> during
>>>>> rcu_stall report where exp_task = 0 list and exp_mask contain only 
>>>>> this cpu
>>>>> ?
>>>> Hello, Mukesh, and thank you for looking this over!
>>>>
>>>> At first glance, I do not believe that this can happen because the
>>>> expedited grace-period machinery avoids waiting on the current CPU.
>>>> (See sync_rcu_exp_select_node_cpus(), both the raw_smp_processor_id()
>>>> early in the function and the get_cpu() later in the function.)
>>>>
>>>> But please let me know if I am missing something here.
>>>>
>>>> But suppose that we could in fact reach this condition.  What bad thing
>>>> would happen?  Other than a resched_cpu() having been invoked several
>>>> times on a not-yet-online CPU, of course.  ;-)
>>>
>>> I thought more about this, what if  synchronize_rcu_expedited thread got
>>> schedule out and run on some other cpu
>>> and we clear out cpu on which it ran next from exp_mask.
>>>
>>> Queuing the work on same cpu ensures that it will always be right cpu to
>>> clear out.
>>> Do you think this can happen ?
>> Indeed it might.
>>
>> But if it did, the scheduler would invoke RCU's hook, which is named
>> rcu_note_context_switch(), and do so on the pre-switch CPU.  There are
>> two implementations for this function, one for CONFIG_PREEMPT=y
>> and another for CONFIG_PREEMPT=n.  Both look to me like they invoke
>> rcu_report_exp_rdp() when needed, one directly and the other via the
>> CONFIG_PREEMPT=n variant of rcu_qs().
>>
>> Am I missing something?
>>
>>
> 
> There is a issue we are facing where exp_mask is not getting cleared and 
> rcu_stall report that
> the cpu we are waiting on sometime in idle and sometime executing some 
> other task but
> it is not clearing itself from exp_mask from a very long time and in all 
> the instances exp_task list is NULL.

Can you please check whether [1] is present in your tree?



Thanks
Neeraj

[1] 
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/kernel/rcu/tree_exp.h?h=v5.17-rc3&id=81f6d49cce2d2fe507e3fddcc4a6db021d9c2e7b
> 
>     expmask = 8,     ==> cpu3
> 
> [80235.522440][T12441] rcu: INFO: rcu_preempt detected expedited stalls 
> on CPUs/tasks: { 3-... } 9163622 jiffies s: 634705 root: 0x8/.
> [80235.534757][T12441] rcu: blocking rcu_node structures:
> [80235.540102][T12441] Task dump for CPU 3:
> [80235.540118][T12441] task:core_ctl        state:D stack:    0 pid: 172 
> ppid:     2 flags:0x00000008
> [80235.540150][T12441] Call trace:
> [80235.540178][T12441]  __switch_to+0x2a8/0x3ac
> [80235.540207][T12441]  rcu_state+0x11b0/0x1480
> 
> 
> [80299.010105][T12441] rcu: INFO: rcu_preempt detected expedited stalls 
> on CPUs/tasks: { 3-... } 9179494 jiffies s: 634705 root: 0x8/.
> [80299.022623][T12441] rcu: blocking rcu_node structures:
> [80299.027924][T12441] Task dump for CPU 3:
> [80299.027942][T12441] task:swapper/3       state:R  running task 
> stack:    0 pid:    0 ppid:     1 flags:0x00000008
> [80299.027993][T12441] Call trace:
> [80299.028025][T12441]  __switch_to+0x2a8/0x3ac
> [80299.028051][T12441]  0xffffffc010113eb4
> 
> 
> As we were not seeing this earlier.
> Below is compile tested patch, can we do something like this  ?
> 
> ==========================================><==================================================== 
> 
> 
> diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h
> index 6453ac5..f0332e4 100644
> --- a/kernel/rcu/tree_exp.h
> +++ b/kernel/rcu/tree_exp.h
> @@ -812,10 +812,12 @@ static int rcu_print_task_exp_stall(struct 
> rcu_node *rnp)
>    */
>   void synchronize_rcu_expedited(void)
>   {
> -    bool no_wq;
> +    bool no_wq = (rcu_scheduler_active == RCU_SCHEDULER_INIT);
> +    bool is_active;
>       struct rcu_exp_work rew;
>       struct rcu_node *rnp;
>       unsigned long s;
> +    int next_cpu;
> 
>       RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
>                lock_is_held(&rcu_lock_map) ||
> @@ -837,19 +839,28 @@ void synchronize_rcu_expedited(void)
>       if (exp_funnel_lock(s))
>           return;  /* Someone else did our work for us. */
> 
> -    /* Don't use workqueue during boot or from an incoming CPU. */
> -    preempt_disable();
> -    no_wq = rcu_scheduler_active == RCU_SCHEDULER_INIT ||
> -        !cpumask_test_cpu(smp_processor_id(), cpu_active_mask);
> -    preempt_enable();
> -
>       /* Ensure that load happens before action based on it. */
>       if (unlikely(no_wq)) {
> -        /* Direct call during scheduler init, early_initcalls() and 
> incoming CPUs. */
> +        /* Direct call during scheduler init, early_initcalls(). */
>           rcu_exp_sel_wait_wake(s);
> +        mutex_unlock(&rcu_state.exp_mutex);
> +        return;
> +    }
> +
> +    preempt_disable();
> +    is_active = cpumask_test_cpu(smp_processor_id(), cpu_active_mask);
> +    preempt_enable();
> +
> +    rew.rew_s = s;
> +    if (!is_active) {
> +        INIT_WORK(&rew.rew_work, wait_rcu_exp_gp);
> +        next_cpu = cpumask_next(smp_processor_id(), cpu_active_mask);
> +        if (next_cpu >= nr_cpu_ids)
> +            next_cpu = cpumask_first(cpu_active_mask);
> +
> +        queue_work_on(next_cpu, rcu_gp_wq, &rew.rew_work);
>       } else {
>           /* Marshall arguments & schedule the expedited grace period. */
> -        rew.rew_s = s;
>           INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp);
>           queue_work(rcu_gp_wq, &rew.rew_work);
>       }
> @@ -863,7 +874,9 @@ void synchronize_rcu_expedited(void)
>       /* Let the next expedited grace period start. */
>       mutex_unlock(&rcu_state.exp_mutex);
> 
> -    if (likely(!no_wq))
> +    if (likely(is_active))
>           destroy_work_on_stack(&rew.rew_work);
> +    else
> +        flush_work(&rew.rew_work);
>   }
>   EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);