Re: [PATCH v14 15/20] mm/swap: serialize memcg changes during pagevec_lru_move_fn

[Alex Shi]

å 2020/7/4 äå9:33, Matthew Wilcox åé:
> On Sat, Jul 04, 2020 at 09:12:46PM +0800, Alex Shi wrote:
>> å 2020/7/4 äå7:39, Matthew Wilcox åé:
>>> On Sat, Jul 04, 2020 at 07:34:59PM +0800, Alex Shi wrote:
>>>> That's a great idea! Guess what the new struct we need would be like this?
>>>> I like to try this. :)
>>>>
>>>>
>>>> diff --git a/include/linux/pagevec.h b/include/linux/pagevec.h
>>>> index 081d934eda64..d62778c8c184 100644
>>>> --- a/include/linux/pagevec.h
>>>> +++ b/include/linux/pagevec.h
>>>> @@ -20,7 +20,7 @@
>>>>  struct pagevec {
>>>>         unsigned char nr;
>>>>         bool percpu_pvec_drained;
>>>> -       struct page *pages[PAGEVEC_SIZE];
>>>> +       struct list_head veclist;
>>>>  };
>>>
>>> pagevecs are used not just for LRU.  If you want to use a list_head for
>>> LRU then define a new structure.
>>
>> yes, there are much page don't use page->lru, like slab etc. we need a new struct> 
> That's not what I mean.  Slab pages aren't on the LRU anyway.

Right. I mean, that's reason for a new struct if we change to list.

> 
> Consider the callers of page_cache_delete_batch().  These use a pagevec
> for a non-LRU purpose, and they will be much slower with a list_head than
> with an array.
> 

Thanks for the info.
If the list is slower than pagevec, maybe it's not worth to do the change.
Since pagevec could handle any kind of pages, anon/file, non-active/active, but one
list only fit for just one kind of list. 5 kinds of list adding would increase the
complexity. Consider this, I am wondering if it's worth?

Thanks
Alex