Re: [PATCH 7/7] mm/khugepaged: try to free transhuge swapcache when possible

[Miaohe Lin]

On 2022/6/18 0:35, Yang Shi wrote:
> On Thu, Jun 16, 2022 at 7:27 PM Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:
>>
>> On 2022/6/16 23:53, Yang Shi wrote:
>>> On Thu, Jun 16, 2022 at 12:42 AM Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:
>>>>
>>>> On 2022/6/16 7:58, Yang Shi wrote:
>>>>> On Sat, Jun 11, 2022 at 1:47 AM Miaohe Lin <linmiaohe@xxxxxxxxxx> wrote:
>>>>>>
>>>>>> Transhuge swapcaches won't be freed in __collapse_huge_page_copy().
>>>>>> It's because release_pte_page() is not called for these pages and
>>>>>> thus free_page_and_swap_cache can't grab the page lock. These pages
>>>>>> won't be freed from swap cache even if we are the only user until
>>>>>> next time reclaim. It shouldn't hurt indeed, but we could try to
>>>>>> free these pages to save more memory for system.
>>>>>
>>>>>
>>>>>>
>>>>>> Signed-off-by: Miaohe Lin <linmiaohe@xxxxxxxxxx>
>>>>>> ---
>>>>>>  include/linux/swap.h | 5 +++++
>>>>>>  mm/khugepaged.c      | 1 +
>>>>>>  mm/swap.h            | 5 -----
>>>>>>  3 files changed, 6 insertions(+), 5 deletions(-)
>>>>>>
>>>>>> diff --git a/include/linux/swap.h b/include/linux/swap.h
>>>>>> index 8672a7123ccd..ccb83b12b724 100644
>>>>>> --- a/include/linux/swap.h
>>>>>> +++ b/include/linux/swap.h
>>>>>> @@ -456,6 +456,7 @@ static inline unsigned long total_swapcache_pages(void)
>>>>>>         return global_node_page_state(NR_SWAPCACHE);
>>>>>>  }
>>>>>>
>>>>>> +extern void free_swap_cache(struct page *page);
>>>>>>  extern void free_page_and_swap_cache(struct page *);
>>>>>>  extern void free_pages_and_swap_cache(struct page **, int);
>>>>>>  /* linux/mm/swapfile.c */
>>>>>> @@ -540,6 +541,10 @@ static inline void put_swap_device(struct swap_info_struct *si)
>>>>>>  /* used to sanity check ptes in zap_pte_range when CONFIG_SWAP=0 */
>>>>>>  #define free_swap_and_cache(e) is_pfn_swap_entry(e)
>>>>>>
>>>>>> +static inline void free_swap_cache(struct page *page)
>>>>>> +{
>>>>>> +}
>>>>>> +
>>>>>>  static inline int add_swap_count_continuation(swp_entry_t swp, gfp_t gfp_mask)
>>>>>>  {
>>>>>>         return 0;
>>>>>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>>>>>> index ee0a719c8be9..52109ad13f78 100644
>>>>>> --- a/mm/khugepaged.c
>>>>>> +++ b/mm/khugepaged.c
>>>>>> @@ -756,6 +756,7 @@ static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
>>>>>>         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
>>>>>>                 list_del(&src_page->lru);
>>>>>>                 release_pte_page(src_page);
>>>>>> +               free_swap_cache(src_page);
>>>>>
>>>>> Will this really work? The free_swap_cache() will just dec refcounts
>>>>> without putting the page back to buddy. So the hugepage is not
>>>>> actually freed at all. Am I missing something?
>>>>
>>>> Thanks for catching this! If page is on percpu lru_pvecs cache, page will
>>>> be released when lru_pvecs are drained. But if not, free_swap_cache() won't
>>>> free the page as it assumes the caller has a reference on the page and thus
>>>> only does page_ref_sub(). Does the below change looks sense for you?
>>>
>>> THP gets drained immediately so they won't stay in pagevecs.
>>
>> Yes, you're right. I missed this.
>>
>>>
>>>>
>>>> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
>>>> index 52109ad13f78..b8c96e33591d 100644
>>>> --- a/mm/khugepaged.c
>>>> +++ b/mm/khugepaged.c
>>>> @@ -755,8 +755,12 @@ static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
>>>>
>>>>         list_for_each_entry_safe(src_page, tmp, compound_pagelist, lru) {
>>>>                 list_del(&src_page->lru);
>>>> -               release_pte_page(src_page);
>>>> +               mod_node_page_state(page_pgdat(src_page),
>>>> +                                   NR_ISOLATED_ANON + page_is_file_lru(src_page),
>>>> +                                   -compound_nr(src_page));
>>>> +               unlock_page(src_page);
>>>>                 free_swap_cache(src_page);
>>>> +               putback_lru_page(src_page);
>>>
>>> I'm not sure if it is worth it or not for a rare corner case since THP
>>> should not stay in swapcache unless try_to_unmap() in vmscan fails
>>
>> IIUC, even if try_to_unmap() in vmscan succeeds, THP might be still in the
>> swapcache if shrink_page_list is not called for this THP again after writeback
>> is done, e.g. when shrink_page_list is called from madvise, so there might be
> 
> I don't get, doesn't __remove_mapping() delete the page from swap cache?

Sorry for making confusion. :(
IIUC, __remove_mapping() is only called when page is clean and page writeback is done in
shrink_page_list(). So for the first round of shrink_page_list(), the THP is under writeback
and __remove_mapping() won't be called. THP will be removed from swapcache via __remove_mapping()
in next round of shrink_page_list() if THP is clean and not under writeback. So THP should be in
the swapcache until next round of shrink_page_list().

And if shrink_page_list is called from madvise, the next round of shrink_page_list() for this
THP won't arrive if there is no memory pressure because madvise can't shrink pages that are
already in the swapcache (!pte_present case is ignored in madvise_cold_or_pageout_pte_range()).
So the THP might stay in swapcache for a long time.

Does this make sense for you?

> 
>> no memory pressure, or do_swap_page puts the THP into page table again. Also THP
> 
> do_swap_page() just swaps in base page, never THP.

If the THP is not removed from swapcache, do_swap_cache can found it via lookup_swap_cache().
So we "swap in" the THP.

> 
>> might not be splited when deferred_split_shrinker is not called, e.g. due to
> 
> I don't see how deferred split is related to this.

What I mean is that if the THP is splitted, khugepaged won't found that THP. So deferred split
should be considered?

> 
>> not lacking of memory. Even if there is memory pressure, the THP will stay in
>> swapcache until next round page reclaim for this THP is done. So there should
>> be a non-negligible window that THP will stay in the swapcache.
>> Or am I miss something?
> 
> I guess you may misunderstand what I meant. This patch is trying to
> optimize freeing THP in swapcache. But it should be very rare that
> khugepaged sees THP from swap cache. The only case I could think of is
> try_to_unmap() in vmscan fails. That might leave THP in swap cache so
> that khugepaged could see it.

I was trying to show you that how a THP can stay in the swapcache. If it's not
removed from swapcache via __remove_mapping, not splitted or swapped in before
it's removed, it will stay in the swapcache. Or am I miss something?

I hope I make my point clear this time. ;)

> 
> 
>>
>>> IIUC. And it is not guaranteed that free_swap_cache() will get the
>>> page lock.
>>
>> IMHO, we're not guaranteed that free_swap_cache() will get the page lock for the normal
>> page anyway.
>>
>> Thanks!
>>
>>>
>>>>         }
>>>>  }
>>>>
>>>> Thanks!
>>>>
>>>>>
>>>>>>         }
>>>>>>  }
>>>>>>
>>>>>> diff --git a/mm/swap.h b/mm/swap.h
>>>>>> index 0193797b0c92..863f6086c916 100644
>>>>>> --- a/mm/swap.h
>>>>>> +++ b/mm/swap.h
>>>>>> @@ -41,7 +41,6 @@ void __delete_from_swap_cache(struct page *page,
>>>>>>  void delete_from_swap_cache(struct page *page);
>>>>>>  void clear_shadow_from_swap_cache(int type, unsigned long begin,
>>>>>>                                   unsigned long end);
>>>>>> -void free_swap_cache(struct page *page);
>>>>>>  struct page *lookup_swap_cache(swp_entry_t entry,
>>>>>>                                struct vm_area_struct *vma,
>>>>>>                                unsigned long addr);
>>>>>> @@ -81,10 +80,6 @@ static inline struct address_space *swap_address_space(swp_entry_t entry)
>>>>>>         return NULL;
>>>>>>  }
>>>>>>
>>>>>> -static inline void free_swap_cache(struct page *page)
>>>>>> -{
>>>>>> -}
>>>>>> -
>>>>>>  static inline void show_swap_cache_info(void)
>>>>>>  {
>>>>>>  }
>>>>>> --
>>>>>> 2.23.0
>>>>>>
>>>>>>
>>>>> .
>>>>>
>>>>
>>> .
>>>
>>
> .
>