Re: [PATCH v3 10/45] smp: Use get/put_online_cpus_atomic() to prevent CPU offline

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 07/02/2013 02:17 PM, Michael Wang wrote:
> On 07/02/2013 04:25 PM, Srivatsa S. Bhat wrote:
>> Hi Michael,
>>
>> On 07/02/2013 11:02 AM, Michael Wang wrote:
>>> Hi, Srivatsa
>>>
>>> On 06/28/2013 03:54 AM, Srivatsa S. Bhat wrote:
>>> [snip]
>>>> @@ -625,8 +632,9 @@ EXPORT_SYMBOL(on_each_cpu_mask);
>>>>   * The function might sleep if the GFP flags indicates a non
>>>>   * atomic allocation is allowed.
>>>>   *
>>>> - * Preemption is disabled to protect against CPUs going offline but not online.
>>>> - * CPUs going online during the call will not be seen or sent an IPI.
>>>> + * We use get/put_online_cpus_atomic() to protect against CPUs going
>>>> + * offline but not online. CPUs going online during the call will
>>>> + * not be seen or sent an IPI.
>>>
>>> I was a little confused about this comment, if the offline-cpu still
>>> have chances to become online, then there is chances that we will pick
>>> it from for_each_online_cpu(), isn't it? Did I miss some point?
>>>
>>
>> Whether or not the newly onlined CPU is observed in our for_each_online_cpu()
>> loop, is dependent on timing. On top of that, there are 2 paths in the code:
>> one which uses a temporary cpumask and the other which doesn't. In the former
>> case, if a CPU comes online _after_ we populate the temporary cpumask, then
>> we won't send an IPI to that cpu, since the temporary cpumask doesn't contain
>> that CPU. Whereas, if we observe the newly onlined CPU in the for_each_online_cpu()
>> loop itself (either in the former or the latter case), then yes, we will send
>> the IPI to that CPU.
> 
> So it is not 'during the call' but 'during the call of
> on_each_cpu_mask()', correct?
> 

Well, as I said, its timing dependent. We might miss the newly onlined CPU in
the for_each_online_cpu() loop itself, based on when exactly the CPU was added
to the cpu_online_mask. So you can't exactly pin-point the places where you'll
miss the CPU and where you won't. Besides, is it _that_ important? It is after
all unpredictable..

> The comment position seems like it declaim that during the call of this
> func, online-cpu won't be seem and send IPI...
>

Doesn't matter, AFAICS. The key take-away from that whole comment is: nothing is
done to prevent CPUs from coming online while the function is running, whereas
the online CPUs are guaranteed to remain online throughout the function. In other
words, its a weaker form of get_online_cpus()/put_online_cpus(), providing a
one-way synchronization (CPU offline).

As long as that idea is conveyed properly, the purpose of that comment is served,
IMHO.

Regards,
Srivatsa S. Bhat


>>>>   *
>>>>   * You must not call this function with disabled interrupts or
>>>>   * from a hardware interrupt handler or from a bottom half handler.
>>>> @@ -641,26 +649,26 @@ void on_each_cpu_cond(bool (*cond_func)(int cpu, void *info),
>>>>  	might_sleep_if(gfp_flags & __GFP_WAIT);
>>>>
>>>>  	if (likely(zalloc_cpumask_var(&cpus, (gfp_flags|__GFP_NOWARN)))) {
>>>> -		preempt_disable();
>>>> +		get_online_cpus_atomic();
>>>>  		for_each_online_cpu(cpu)
>>>>  			if (cond_func(cpu, info))
>>>>  				cpumask_set_cpu(cpu, cpus);
>>>>  		on_each_cpu_mask(cpus, func, info, wait);
>>>> -		preempt_enable();
>>>> +		put_online_cpus_atomic();
>>>>  		free_cpumask_var(cpus);
>>>>  	} else {
>>>>  		/*
>>>>  		 * No free cpumask, bother. No matter, we'll
>>>>  		 * just have to IPI them one by one.
>>>>  		 */
>>>> -		preempt_disable();
>>>> +		get_online_cpus_atomic();
>>>>  		for_each_online_cpu(cpu)
>>>>  			if (cond_func(cpu, info)) {
>>>>  				ret = smp_call_function_single(cpu, func,
>>>>  								info, wait);
>>>>  				WARN_ON_ONCE(!ret);
>>>>  			}
>>>> -		preempt_enable();
>>>> +		put_online_cpus_atomic();
>>>>  	}
>>>>  }
>>>>  EXPORT_SYMBOL(on_each_cpu_cond);
>>>>
>>>
>>