Re: [RFD 1/9] Change cpustat fields to an array.

[Glauber Costa <glommer@parallels.com>]

On 09/28/2011 11:09 PM, Peter Zijlstra wrote:
> On Wed, 2011-09-28 at 15:19 -0300, Glauber Costa wrote:
>> On 09/27/2011 06:00 PM, Peter Zijlstra wrote:
>>> On Fri, 2011-09-23 at 19:20 -0300, Glauber Costa wrote:
>>>>    /* Must have preemption disabled for this to be meaningful. */
>>>> -#define kstat_this_cpu __get_cpu_var(kstat)
>>>> +#define kstat_this_cpu this_cpu_ptr(task_group_kstat(current))
>>>
>>> This just lost you a debug check, the former would whinge when called
>>> without preemption, the new one wont. Its part of the this_cpu feature
>>> set to make debugging impossible.
>>>
>>>> +#else
>>>> +#define kstat_cpu(cpu) per_cpu(kstat, cpu)
>>>> +#define kstat_this_cpu (&__get_cpu_var(kstat))
>>>> +#endif
>>>>
>>>>    extern unsigned long long nr_context_switches(void);
>>>>
>>>> @@ -52,8 +62,8 @@ struct irq_desc;
>>>>    static inline void kstat_incr_irqs_this_cpu(unsigned int irq,
>>>>                                               struct irq_desc *desc)
>>>>    {
>>>> -       __this_cpu_inc(kstat.irqs[irq]);
>>>> -       __this_cpu_inc(kstat.irqs_sum);
>>>> +       kstat_this_cpu->irqs[irq]++;
>>>> +       kstat_this_cpu->irqs_sum++;
>>>
>>> It might be worth looking at the asm output of that, I think you made it
>>> worse, but I'm not quite sure how smart gcc is, it might just figure out
>>> what you meant.
>>
>> I'd say leave it alone.
>> The biggest difference is that we don't have access to task_group(), or
>> any of the fields in struct task_group. Because of that, we end up
>> having to export a function to do the job of dealing with it.
>>
>> Users inside sched.c won't have this problem. Outside of it, we'll add a
>> call to some paths. True, mostly handle_irq paths, but I don't think
>> that's what's going to kill us.
>>
>> Now if we really really want to save it, we'd have to move struct
>> task_group and its friends to a more visible location like a header...
>
> I'm not quite getting how task_group is relevant here.
>
> The above will do something like:
>
> 	mov gs:$per-cpu-offset-of-kstat, reg
> 	inc reg + idx*8
>
> whereas __this_cpu_inc() could end up like:
>
> 	inc gs:$per-cpu-offset-of-kstat + idx*8
>
> or whatnot. Now clearly gcc could be smart and optimize the temporary
> reg thing away in the earlier case, or it might not, I really don't know
> how smart that thing is.
Btw, asm output with CGROUP_SCHED disabled seem to be no worse than
what is in now.