Re: [PATCH 1/2] generic smp function call: add multiple queues for scaling

[Jeremy Fitzhardinge <jeremy@goop.org>]

Andi Kleen wrote:
> Ah I see the locking is here. Never mind the earlier comment.
>
>   
>> +#define	NQUEUES	CONFIG_GENERIC_SMP_QUEUES
>> +#else
>> +#define	NQUEUES	1
>> +#endif
>> +
>> static DEFINE_PER_CPU(struct call_single_queue, call_single_queue);
>> -static LIST_HEAD(call_function_queue);
>> -__cacheline_aligned_in_smp DEFINE_SPINLOCK(call_function_lock);
>> +struct queue {
>> +	struct list_head list;
>> +	spinlock_t lock;
>> +};
>> +
>> +static __cacheline_aligned_in_smp struct queue 
>> call_function_queues[NQUEUES];
>>     
>
> Hmm are you sure this aligns the individual elements and not the whole
> array? 
>   

Hm, that's a point.  I guess the __cacheline_aligned_in_smp should be on 
the struct definition.

>> +static int __init init_smp_function_call(void)
>> +{
>> +	int i;
>> +
>> +	for(i = 0; i < NQUEUES; i++) {
>> +		INIT_LIST_HEAD(&call_function_queues[i].list);
>> +		spin_lock_init(&call_function_queues[i].lock);
>> +	}
>> +
>> +	return 0;
>> +}
>> +early_initcall(init_smp_function_call);
>>     
>
> You can avoid all that init gunk by using the [0 ... NQUEUES] = ..
> gcc extension in the initializer.
>   

Are you sure?  I tried using it, but couldn't work out how.  Remember 
the list head init needs to point to itself, which means it's not 
constant across the array.

    J