Re: [PATCH v5 5/6] mm: vmscan: Avoid split during shrink_folio_list()

From: Ryan Roberts
Date: Fri Apr 05 2024 - 03:28:51 EST

On 05/04/2024 05:06, Barry Song wrote:
> On Wed, Apr 3, 2024 at 2:10 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>>
>> On 28/03/2024 08:18, Barry Song wrote:
>>> On Thu, Mar 28, 2024 at 3:45 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote:
>>>>
>>>> Now that swap supports storing all mTHP sizes, avoid splitting large
>>>> folios before swap-out. This benefits performance of the swap-out path
>>>> by eliding split_folio_to_list(), which is expensive, and also sets us
>>>> up for swapping in large folios in a future series.
>>>>
>>>> If the folio is partially mapped, we continue to split it since we want
>>>> to avoid the extra IO overhead and storage of writing out pages
>>>> uneccessarily.
>>>>
>>>> Reviewed-by: David Hildenbrand <david@xxxxxxxxxx>
>>>> Reviewed-by: Barry Song <v-songbaohua@xxxxxxxx>
>>>> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx>
>>>> ---
>>>> mm/vmscan.c | 9 +++++----
>>>> 1 file changed, 5 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>>>> index 00adaf1cb2c3..293120fe54f3 100644
>>>> --- a/mm/vmscan.c
>>>> +++ b/mm/vmscan.c
>>>> @@ -1223,11 +1223,12 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
>>>> if (!can_split_folio(folio, NULL))
>>>> goto activate_locked;
>>>> /*
>>>> - * Split folios without a PMD map right
>>>> - * away. Chances are some or all of the
>>>> - * tail pages can be freed without IO.
>>>> + * Split partially mapped folios right
>>>> + * away. We can free the unmapped pages
>>>> + * without IO.
>>>> */
>>>> - if (!folio_entire_mapcount(folio) &&
>>>> + if (data_race(!list_empty(
>>>> + &folio->_deferred_list)) &&
>>>> split_folio_to_list(folio,
>>>> folio_list))
>>>> goto activate_locked;
>>>
>>> Hi Ryan,
>>>
>>> Sorry for bringing up another minor issue at this late stage.
>>
>> No problem - I'd rather take a bit longer and get it right, rather than rush it
>> and get it wrong!
>>
>>>
>>> During the debugging of thp counter patch v2, I noticed the discrepancy between
>>> THP_SWPOUT_FALLBACK and THP_SWPOUT.
>>>
>>> Should we make adjustments to the counter?
>>
>> Yes, agreed; we want to be consistent here with all the other existing THP
>> counters; they only refer to PMD-sized THP. I'll make the change for the next
>> version.
>>
>> I guess we will eventually want equivalent counters for per-size mTHP using the
>> framework you are adding.
>
> Hi Ryan,
>
> Today, I created counters for per-order SWPOUT and SWPOUT_FALLBACK.
> I'd appreciate any
> suggestions you might have before I submit this as patch 2/2 of my
> mTHP counters series.

Amazing - this is going to be very useful!

>
> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
> index cc13fa14aa32..762a6d8759b9 100644
> --- a/include/linux/huge_mm.h
> +++ b/include/linux/huge_mm.h
> @@ -267,6 +267,8 @@ unsigned long thp_vma_allowable_orders(struct
> vm_area_struct *vma,
> enum thp_stat_item {
> THP_STAT_ANON_ALLOC,
> THP_STAT_ANON_ALLOC_FALLBACK,
> + THP_STAT_ANON_SWPOUT,
> + THP_STAT_ANON_SWPOUT_FALLBACK,
> __THP_STAT_COUNT
> };
>
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index e704b4408181..7f2b5d2852cc 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -554,10 +554,14 @@ static struct kobj_attribute _name##_attr =
> __ATTR_RO(_name)
>
> THP_STATE_ATTR(anon_alloc, THP_STAT_ANON_ALLOC);
> THP_STATE_ATTR(anon_alloc_fallback, THP_STAT_ANON_ALLOC_FALLBACK);
> +THP_STATE_ATTR(anon_swpout, THP_STAT_ANON_SWPOUT);
> +THP_STATE_ATTR(anon_swpout_fallback, THP_STAT_ANON_SWPOUT_FALLBACK);
>
> static struct attribute *stats_attrs[] = {
> &anon_alloc_attr.attr,
> &anon_alloc_fallback_attr.attr,
> + &anon_swpout_attr.attr,
> + &anon_swpout_fallback_attr.attr,
> NULL,
> };
>
> diff --git a/mm/page_io.c b/mm/page_io.c
> index a9a7c236aecc..be4f822b39f8 100644
> --- a/mm/page_io.c
> +++ b/mm/page_io.c
> @@ -212,13 +212,16 @@ int swap_writepage(struct page *page, struct
> writeback_control *wbc)
>
> static inline void count_swpout_vm_event(struct folio *folio)
> {
> + long nr_pages = folio_nr_pages(folio);
> #ifdef CONFIG_TRANSPARENT_HUGEPAGE
> if (unlikely(folio_test_pmd_mappable(folio))) {
> count_memcg_folio_events(folio, THP_SWPOUT, 1);
> count_vm_event(THP_SWPOUT);
> }
> + if (nr_pages > 0 && nr_pages <= HPAGE_PMD_NR)

The guard is a bit ugly; I wonder if we should at least check that order is in
bounds in count_thp_state(), since all callers could benefit? Then we only have
to care about the nr_pages > 0 condition here. Just a thought...

> + count_thp_state(folio_order(folio), THP_STAT_ANON_SWPOUT);

So you're counting THPs, not pages; I agree with that approach.

> #endif
> - count_vm_events(PSWPOUT, folio_nr_pages(folio));
> + count_vm_events(PSWPOUT, nr_pages);
> }
>
> #if defined(CONFIG_MEMCG) && defined(CONFIG_BLK_CGROUP)
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index ffc4553c8615..b7c5fbd830b6 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -1247,6 +1247,10 @@ static unsigned int shrink_folio_list(struct
> list_head *folio_list,
> count_vm_event(
> THP_SWPOUT_FALLBACK);
> }
> + if (nr_pages > 0 && nr_pages
> <= HPAGE_PMD_NR)
> +
> count_thp_state(folio_order(folio),
> +
> THP_STAT_ANON_SWPOUT_FALLBACK);
> +
> #endif
> if (!add_to_swap(folio))
> goto activate_locked_split;
>
>
> Thanks
> Barry