Re: Re: [PATCH v4 1/2] bpf: avoid grabbing spin_locks of all cpus when no free elems

[Feng Zhou <zhoufeng.zf@bytedance.com>]

在 2022/6/1 下午7:35, Alexei Starovoitov 写道:
> On Wed, Jun 1, 2022 at 1:11 PM Feng Zhou <zhoufeng.zf@bytedance.com> wrote:
>> 在 2022/6/1 下午5:50, Alexei Starovoitov 写道:
>>> On Wed, Jun 1, 2022 at 10:42 AM Feng zhou <zhoufeng.zf@bytedance.com> wrote:
>>>>    static inline void ___pcpu_freelist_push(struct pcpu_freelist_head *head,
>>>> @@ -130,14 +134,19 @@ static struct pcpu_freelist_node *___pcpu_freelist_pop(struct pcpu_freelist *s)
>>>>           orig_cpu = cpu = raw_smp_processor_id();
>>>>           while (1) {
>>>>                   head = per_cpu_ptr(s->freelist, cpu);
>>>> +               if (READ_ONCE(head->is_empty))
>>>> +                       goto next_cpu;
>>>>                   raw_spin_lock(&head->lock);
>>>>                   node = head->first;
>>>>                   if (node) {
>>> extra bool is unnecessary.
>>> just READ_ONCE(head->first)
>> As for why to add is_empty instead of directly judging head->first, my
>> understanding is this, head->first is frequently modified during updating
>> map, which will lead to invalid other cpus's cache, and is_empty is after
>> freelist having no free elems will be changed, the performance will be
>> better.
> maybe. pls benchmark it.
> imo wasting a bool for the corner case is not a good trade off.

Yes, I will do and post the results as soon as possible, Thanks.