Re: [PATCH 2/3] sched/core: Introduce nr_running percpu for each cookie

[cruzzhao <cruzzhao@linux.alibaba.com>]

在 2022/7/4 下午5:45, Peter Zijlstra 写道:
> On Tue, Jun 28, 2022 at 03:57:24PM +0800, Cruz Zhao wrote:
> 
>>  static unsigned long sched_core_alloc_cookie(void)
>>  {
>>  	struct sched_core_cookie *ck = kmalloc(sizeof(*ck), GFP_KERNEL);
>> +	int cpu;
>> +
>>  	if (!ck)
>>  		return 0;
>>  
>>  	refcount_set(&ck->refcnt, 1);
>> +
>> +	ck->nr_running = alloc_percpu(unsigned int);
> 
> 	if (!ck->nr_running)
> 		// do something
> 
>> +	for_each_possible_cpu(cpu)
>> +		*per_cpu_ptr(ck->nr_running, cpu) = 0;
> 
> So I really, as in *really* dislike how this blows up the size of
> cookies. Esp. with 100s of CPUs not actually being rare these days.

My idea is to get the distribution of cookie'd tasks on each runqueue
through ck->nr_running, so as to facilitate optimization of load
balance. Sorry for not stating this in the change log.

This does blow up the size of cookies in scenarios with a large number
of CPUs, and I'll try to get around this problem.

Many thanks for reviewing.
Best wishes,
Cruz Zhao