Re: 答复: 答复: 答复: 答复: 答复: [PATCH] page_pool: mark unbound node page as reusable pages

[Yunsheng Lin <linyunsheng@huawei.com>]

On 2019/12/5 11:18, Li,Rongqing wrote:
> 
>>>> [1] https://lore.kernel.org/patchwork/patch/1125789/
>>>
>>>
>>> What is status of this patch? I think you should fix your firmware or
>>> bios
>>
>> Have not reached a conclusion yet.
> 
> I think it will never be accepted
> 
>>
>>>
>>> Consider the below condition:
>>>
>>> there is two numa node, and NIC sites on node 2, but NUMA_NO_NODE is
>>> used, recycle will fail due to page_to_nid(page) == numa_mem_id(), and
>> reallocated pages maybe always from node 1, then the recycle will never
>> success.
>>
>> For page pool:
>>
>> 1. if the pool->p.nid != NUMA_NO_NODE, the recycle is always decided by
>> checking page_to_nid(page) == pool->p.nid.
>>
>> 2. only when the pool->p.nid == NUMA_NO_NODE, the numa_mem_id() is
>> checked
>>    to decide the recycle.
>>
>> Yes, If pool->p.nid == NUMA_NO_NODE, and the cpu that is doing recycling is
>> changing each time, the recycle may never success, but it is not common, 
> 
> Why can we ignore this condition, and accept your hardware with abnormal node
> information
> 
>> and
>> have its own performance penalty when changing the recycling cpu so often.
> 
> I have said if hardware takes care of numa node, it should be assign when page pool
> Is created, not depends on recycle.
> 
> If you insist your idea, you can submit you patch after this one

I am just arguing that rx recycle the page pool is doing should be consistent
with other driver that does its own recycle.

If the driver that does its own recycle begin to support the page pool,
then there may be different behavior when they are not consistent.

> 
> -RongQing
> 
>>
>>
>>>
>>> -RongQing
>>>
>>>
>