Re: [RFC PATCH] madvise: make madvise_cold_or_pageout_pte_range() support large folio

[Yin, Fengwei]

On 7/14/2023 11:41 PM, Yu Zhao wrote:
> On Thu, Jul 13, 2023 at 11:57 PM Yin, Fengwei <fengwei.yin@xxxxxxxxx> wrote:
>>
>>
>>>> -               if (pageout_anon_only_filter && !folio_test_anon(folio))
>>>> +               /* Do not interfere with other mappings of this folio */
>>>> +               if (folio_mapcount(folio) != 1)
>>>>                         continue;
>>>>
>>>> -               VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
>>>> -
>>>> -               if (pte_young(ptent)) {
>>>> -                       ptent = ptep_get_and_clear_full(mm, addr, pte,
>>>> -                                                       tlb->fullmm);
>>>> -                       ptent = pte_mkold(ptent);
>>>> -                       set_pte_at(mm, addr, pte, ptent);
>>>> -                       tlb_remove_tlb_entry(tlb, pte, addr);
>>>> -               }
>>>> -
>>>> -               /*
>>>> -                * We are deactivating a folio for accelerating reclaiming.
>>>> -                * VM couldn't reclaim the folio unless we clear PG_young.
>>>> -                * As a side effect, it makes confuse idle-page tracking
>>>> -                * because they will miss recent referenced history.
>>>> -                */
>>>> -               folio_clear_referenced(folio);
>>>> -               folio_test_clear_young(folio);
>>>> -               if (folio_test_active(folio))
>>>> -                       folio_set_workingset(folio);
>>>> +pageout_cold_folio:
>>>>                 if (pageout) {
>>>>                         if (folio_isolate_lru(folio)) {
>>>>                                 if (folio_test_unevictable(folio))
>>>> @@ -529,8 +542,30 @@ static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
>>>>                 arch_leave_lazy_mmu_mode();
>>>>                 pte_unmap_unlock(start_pte, ptl);
>>>>         }
>>>> -       if (pageout)
>>>> -               reclaim_pages(&folio_list);
>>>> +
>>>> +       if (pageout) {
>>>> +               LIST_HEAD(reclaim_list);
>>>> +
>>>> +               while (!list_empty(&folio_list)) {
>>>> +                       int refs;
>>>> +                       unsigned long flags;
>>>> +                       struct mem_cgroup *memcg = folio_memcg(folio);
>>>> +
>>>> +                       folio = lru_to_folio(&folio_list);
>>>> +                       list_del(&folio->lru);
>>>> +
>>>> +                       refs = folio_referenced(folio, 0, memcg, &flags);
>>>> +
>>>> +                       if ((flags & VM_LOCKED) || (refs == -1)) {
>>>> +                               folio_putback_lru(folio);
>>>> +                               continue;
>>>> +                       }
>>>> +
>>>> +                       folio_test_clear_referenced(folio);
>>>> +                       list_add(&folio->lru, &reclaim_list);
>>>> +               }
>>>> +               reclaim_pages(&reclaim_list);
>>>> +       }
>>>
>>> i overlooked the chunk above -- it's unnecessary: after we split the
>>> large folio (and splice the base folios onto the same LRU list), we
>>> continue at the position of the first base folio because of:
>>>
>>>   pte--;
>>>   addr -= PAGE_SIZE;
>>>   continue;
>>>
>>> And then we do pte_mkold(), which takes care of the A-bit.
>> This patch moves the A-bit clear out of the folio isolation loop. So
>> even the folio is split and loop restarts from the first base folio,
>> the A-bit is not cleared. A-bit is only cleared in reclaim loop.
>>
>> There is one option for A-bit clearing:
>>   - clear A-bit of base 4K page in isolation loop and leave large folio
>>     A-bit clearing to reclaim loop.
>>
>> This patch didn't use it because don't want to introduce A-bit clearing
>> in two places. But I am open about clearing base 4K page A-bit cleared in
>> isolation loop. Thanks.
> 
> Sorry but why are we trying to do multiple things in one patch that I
> assumed is supposed to simply fix madvise() for large anon folios? And
> none of those things seems to have a clear rationale behind it.
> 
> The only patch that makes sense at the moment (or the first patch of a
> series) is what I said before:
> 
> -  if (folio_mapcount(folio) != 1)
> +  if (folio_estimated_sharers(folio) != 1)
Definitely. As I replied to you, I will split the patch to two parts:
  - just bug fixing. Include the filio_mapcount() -> folio_estimated_shares().
    And using ptep_clear_flush_young_notify() to clear the young of PTEs.
  - refactor for large folio.
Let me know if this is OK. Thanks.


Regards
Yin, Fengwei

> 
> And probably clarify (preferrably in the comments above) this is an
> estimate because we think it's a better tradeoff if we do so (less
> code/overhead from checking the mapcounts of the rest of folios within
> the range).