Re: [RFC PATCH 02/10] smp, cpu hotplug: Fix smp_call_function_*() to prevent CPU offline properly

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

On 12/05/2012 03:47 AM, Andrew Morton wrote:
> On Tue, 04 Dec 2012 14:24:28 +0530
> "Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com> wrote:
> 
>> From: Michael Wang <wangyun@linux.vnet.ibm.com>
>>
>> With stop_machine() gone from the CPU offline path, we can't depend on
>> preempt_disable() to prevent CPUs from going offline from under us.
>>
>> Use the get/put_online_cpus_stable_atomic() APIs to prevent CPUs from going
>> offline, while invoking from atomic context.
>>
>> ...
>>
>>  	 */
>> -	this_cpu = get_cpu();
>> +	get_online_cpus_stable_atomic();
>> +	this_cpu = smp_processor_id();
> 
> I wonder if get_online_cpus_stable_atomic() should return the local CPU
> ID.  Just as a little convenience thing.  Time will tell.
>

With the new version which doesn't use extra cpumasks, we won't have to
bother about this..
 
>>  	/*
>>  	 * Can deadlock when called with interrupts disabled.
>>
>> ...
>>
>> @@ -380,15 +383,15 @@ int smp_call_function_any(const struct cpumask *mask,
>>  	nodemask = cpumask_of_node(cpu_to_node(cpu));
>>  	for (cpu = cpumask_first_and(nodemask, mask); cpu < nr_cpu_ids;
>>  	     cpu = cpumask_next_and(cpu, nodemask, mask)) {
>> -		if (cpu_online(cpu))
>> +		if (cpu_online_stable(cpu))
>>  			goto call;
>>  	}
>>  
>>  	/* Any online will do: smp_call_function_single handles nr_cpu_ids. */
>> -	cpu = cpumask_any_and(mask, cpu_online_mask);
>> +	cpu = cpumask_any_and(mask, cpu_online_stable_mask);
>>  call:
>>  	ret = smp_call_function_single(cpu, func, info, wait);
>> -	put_cpu();
>> +	put_online_cpus_stable_atomic();
>>  	return ret;
>>  }
>>  EXPORT_SYMBOL_GPL(smp_call_function_any);
> 
> So smp_call_function_any() has no synchronization against CPUs coming
> online.  Hence callers of smp_call_function_any() are responsible for
> ensuring that CPUs which are concurrently coming online will adopt the
> required state?
>

Yes.
 
> I guess that has always been the case...
> 

Right.

Regards,
Srivatsa S. Bhat