Re: [PATCH] Avoid mask based num_possible_cpus and num_online_cpus

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

On 01/21/2012 05:25 AM, Venki Pallipadi wrote:

> On Fri, Jan 20, 2012 at 3:45 PM, KOSAKI Motohiro
> <kosaki.motohiro@gmail.com> wrote:
>>>>> +int nr_online_cpus __read_mostly;
>>>>> +EXPORT_SYMBOL(nr_online_cpus);
>>>>> +
>>>>>  void set_cpu_possible(unsigned int cpu, bool possible)
>>>>>  {
>>>>>       if (possible)
>>>>
>>>>
>>>> Did you forget to add:
>>>>
>>>>        nr_possible_cpus = cpumask_weight(cpu_possible_mask);
>>>>
>>>> inside set_cpu_possible() ?
>>>
>>> No. That was intentional as I have that coupled with nr_cpu_ids and
>>> set once after all the bits are set in setup_nr_cpu_ids() instead of
>>> doing for each bit set.
>>
>> But, Srivatsa's way seems more safer, no? Is there any advantage to make couple
>> with nr_cpu_ids?
> 
> I think it is a tradeoff between safer and cleaner :). infact, that's
> how I had coded the patch first. But, then I changed it to be in sync
> with nr_cpu_ids as it seemed a bit cleaner (and also to make sure 2048
> CPU guys won't come after me for doing the mask calculation 2048 times
> during the boot).
> 


I knew you were trying to optimize further when I saw your patch. That's
precisely the reason I cross-checked the code to ensure that the optimization
didn't go beyond correctness :)

And this is what I found:

start_kernel()
  setup_nr_cpu_ids() // This is not the end of setting up cpu_possible_mask
  rest_init()
    kernel_init()
      smp_prepare_cpus();

And on x86, this becomes:
      native_smp_prepare_cpus();
        smp_sanity_check(); // cpu_possible_mask & nr_cpu_ids can change here!
           ^^^^^^^^^

And there is another place where things can change:
prefill_possible_map(). But this is called in setup_arch(), which is called
before setup_nr_cpu_ids(). So we need not worry about this.

(Btw, I checked only the x86 arch. Not sure how other architectures handle
things.)

Regards,
Srivatsa S. Bhat
IBM Linux Technology Center