Re: [PATCH 3 of 8] xen: let the (credit) scheduler know about `node affinity`

[Keir Fraser <keir.xen@gmail.com>]

On 09/10/2012 11:29, "Dario Faggioli" <dario.faggioli@citrix.com> wrote:

> On Fri, 2012-10-05 at 15:25 +0100, Jan Beulich wrote:
>>>>> On 05.10.12 at 16:08, Dario Faggioli <dario.faggioli@citrix.com> wrote:
>>> @@ -287,22 +344,26 @@ static inline void
>>>          }
>>>          else
>>>          {
>>> -            cpumask_t idle_mask;
>>> +            cpumask_t idle_mask, balance_mask;
>> 
>> Be _very_ careful about adding on-stack CPU mask variables
>> (also further below): each one of them grows the stack frame
>> by 512 bytes (when building for the current maximum of 4095
>> CPUs), which is generally too much; you may want to consider
>> pre-allocated scratch space instead.
>> 
> I see your point, and I think you're right... I wasn't "thinking that
> big". :-)
> 
> I'll look into all of these situations and see if I can move the masks
> off the stack. Any preference between global variables and members of
> one of the scheduler's data structures?

Since multiple instances of the scheduler can be active, across multiple cpu
pools, surely they have to be allocated in the per-scheduler-instance
structures? Or dynamically xmalloc'ed just in the scope they are needed.

 -- Keir

> Thanks and Regards,
> Dario