[PATCH v2 5/6] mm: support batched checking of the young flag for MGLRU
Baolin Wang
baolin.wang at linux.alibaba.com
Mon Mar 2 22:03:35 PST 2026
On 3/2/26 5:57 PM, David Hildenbrand (Arm) wrote:
> On 2/27/26 10:44, Baolin Wang wrote:
>> Use the batched helper test_and_clear_young_ptes_notify() to check and clear
>> the young flag to improve the performance during large folio reclamation when
>> MGLRU is enabled.
>>
>> Meanwhile, we can also support batched checking the young and dirty flag
>> when MGLRU walks the mm's pagetable to update the folios' generation
>> counter. Since MGLRU also checks the PTE dirty bit, use folio_pte_batch_flags()
>> with FPB_MERGE_YOUNG_DIRTY set to detect batches of PTEs for a large folio.
>>
>> Then we can remove the ptep_test_and_clear_young_notify() since it has
>> no users now.
>>
>> Note that we also update the 'young' counter and 'mm_stats[MM_LEAF_YOUNG]' counter
>> with the batched count in the lru_gen_look_around() and walk_pte_range(). However,
>> the batched operations may inflate these two counters, because in a large folio not
>> all PTEs may have been accessed. (Additionally, tracking how many PTEs have been
>> accessed within a large folio is not very meaningful, since the mm core actually
>> tracks access/dirty on a per-folio basis, not per page). The impact analysis is as
>> follows:
>>
>> 1. The 'mm_stats[MM_LEAF_YOUNG]' counter has no functional impact and is mainly for
>> debugging.
>>
>> 2. The 'young' counter is used to decide whether to place the current PMD entry into the
>> bloom filters by suitable_to_scan() (so that next time we can check whether it has been
>> accessed again), which may set the hash bit in the bloom filters for a PMD entry that
>> hasn’t seen much access. However, bloom filters inherently allow some error, so this
>> effect appears negligible.
>
>
> Doesn't checkpatch complain about long lines in the patch description?
Will update the commit message.
>> Reviewed-by: Rik van Riel <riel at surriel.com>
>> Signed-off-by: Baolin Wang <baolin.wang at linux.alibaba.com>
>> ---
>
> ...
>
>> index a5f0a264ad56..a1b3967afe41 100644
>> --- a/mm/internal.h
>> +++ b/mm/internal.h
>> @@ -1843,10 +1843,4 @@ static inline int pmdp_test_and_clear_young_notify(struct vm_area_struct *vma,
>>
>> #endif /* CONFIG_MMU_NOTIFIER */
>>
>> -static inline int ptep_test_and_clear_young_notify(struct vm_area_struct *vma,
>> - unsigned long addr, pte_t *ptep)
>> -{
>> - return test_and_clear_young_ptes_notify(vma, addr, ptep, 1);
>> -}
>> -
>> #endif /* __MM_INTERNAL_H */
>> diff --git a/mm/rmap.c b/mm/rmap.c
>> index 11cc6171344f..beb423f3e8ec 100644
>> --- a/mm/rmap.c
>> +++ b/mm/rmap.c
>> @@ -958,25 +958,21 @@ static bool folio_referenced_one(struct folio *folio,
>> return false;
>> }
>>
>> + if (pvmw.pte && folio_test_large(folio)) {
>> + const unsigned long end_addr = pmd_addr_end(address, vma->vm_end);
>> + const unsigned int max_nr = (end_addr - address) >> PAGE_SHIFT;
>> + pte_t pteval = ptep_get(pvmw.pte);
>> +
>> + nr = folio_pte_batch(folio, pvmw.pte, pteval, max_nr);
>> + ptes += nr;
>
> Could we move that "ptes += nr;" just before the "pra->mapcount -= nr;"?
>
> Would make the whole thing look less weird (only incrementing "ptes" with large folios).
OK. Sounds reasonable.
>> + }
>> +
>> if (lru_gen_enabled() && pvmw.pte) {
>> - if (lru_gen_look_around(&pvmw))
>> + if (lru_gen_look_around(&pvmw, nr))
>> referenced++;
>> } else if (pvmw.pte) {
>> - if (folio_test_large(folio)) {
>> - unsigned long end_addr = pmd_addr_end(address, vma->vm_end);
>> - unsigned int max_nr = (end_addr - address) >> PAGE_SHIFT;
>> - pte_t pteval = ptep_get(pvmw.pte);
>> -
>> - nr = folio_pte_batch(folio, pvmw.pte,
>> - pteval, max_nr);
>> - }
>> -
>> - ptes += nr;
>> if (clear_flush_young_ptes_notify(vma, address, pvmw.pte, nr))
>> referenced++;
>> - /* Skip the batched PTEs */
>> - pvmw.pte += nr - 1;
>> - pvmw.address += (nr - 1) * PAGE_SIZE;
>> } else if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE)) {
>> if (pmdp_clear_flush_young_notify(vma, address,
>> pvmw.pmd))
>> @@ -995,6 +991,10 @@ static bool folio_referenced_one(struct folio *folio,
>> page_vma_mapped_walk_done(&pvmw);
>> break;
>> }
>> +
>> + /* Skip the batched PTEs */
>> + pvmw.pte += nr - 1;
>> + pvmw.address += (nr - 1) * PAGE_SIZE;
>> }
>>
>> if (referenced)
>> diff --git a/mm/vmscan.c b/mm/vmscan.c
>> index 0a5622420987..7457b3c06fa3 100644
>> --- a/mm/vmscan.c
>> +++ b/mm/vmscan.c
>> @@ -3474,6 +3474,7 @@ static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
>> struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
>> DEFINE_MAX_SEQ(walk->lruvec);
>> int gen = lru_gen_from_seq(max_seq);
>> + unsigned int nr;
>> pmd_t pmdval;
>>
>> pte = pte_offset_map_rw_nolock(args->mm, pmd, start & PMD_MASK, &pmdval, &ptl);
>> @@ -3492,11 +3493,13 @@ static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
>>
>> lazy_mmu_mode_enable();
>> restart:
>> - for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
>> + for (i = pte_index(start), addr = start; addr != end; i += nr, addr += nr * PAGE_SIZE) {
>> unsigned long pfn;
>> struct folio *folio;
>> - pte_t ptent = ptep_get(pte + i);
>> + pte_t *cur_pte = pte + i;
>> + pte_t ptent = ptep_get(cur_pte);
>>
>> + nr = 1;
>
>
> Looking at this again, we should get rid of "i" completely and instead
> * rename pte to start_pte
> * Add a new pte, which we increment in the loop
>
> So we end up with something like
>
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index 54cf4924d223..150cbb2253b9 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -3486,9 +3486,8 @@ static void walk_update_folio(struct lru_gen_mm_walk *walk, struct folio *folio,
> static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
> struct mm_walk *args)
> {
> - int i;
> bool dirty;
> - pte_t *pte;
> + pte_t *start_pte, pte;
> spinlock_t *ptl;
> unsigned long addr;
> int total = 0;
> @@ -3499,29 +3498,31 @@ static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
> struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec);
> DEFINE_MAX_SEQ(walk->lruvec);
> int gen = lru_gen_from_seq(max_seq);
> + unsigned int nr;
> pmd_t pmdval;
>
> - pte = pte_offset_map_rw_nolock(args->mm, pmd, start & PMD_MASK, &pmdval, &ptl);
> - if (!pte)
> + start_pte = pte_offset_map_rw_nolock(args->mm, pmd, start & PMD_MASK, &pmdval, &ptl);
> + if (!start_pte)
> return false;
>
> if (!spin_trylock(ptl)) {
> - pte_unmap(pte);
> + pte_unmap(start_pte);
> return true;
> }
>
> if (unlikely(!pmd_same(pmdval, pmdp_get_lockless(pmd)))) {
> - pte_unmap_unlock(pte, ptl);
> + pte_unmap_unlock(start_pte, ptl);
> return false;
> }
>
> lazy_mmu_mode_enable();
> restart:
> - for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) {
> + for (addr = start, pte = start_pte; addr != end; addr += nr * PAGE_SIZE, pte += nr) {
> unsigned long pfn;
> struct folio *folio;
> - pte_t ptent = ptep_get(pte + i);
> + pte_t ptent = ptep_get(pte);
>
> + nr = 1;
> total++;
> walk->mm_stats[MM_LEAF_TOTAL]++;
>
> @@ -3533,7 +3534,16 @@ static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end,
> if (!folio)
> continue;
Thanks for the code. I considered simplifying the logic this way, but
this logic would be incorrect. Because the 'start' can be updated by the
get_next_vma() below, when we goto restart, we will get a new 'i' value
by 'pte_index(start)', so I kept the 'i' variable.
>> walk_update_folio(walk, last, gen, dirty);
>> @@ -4166,7 +4178,7 @@ static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc)
>> * the PTE table to the Bloom filter. This forms a feedback loop between the
>> * eviction and the aging.
>> */
>> -bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
>> +bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw, unsigned int nr)
>> {
>> int i;
>> bool dirty;
>> @@ -4184,12 +4196,13 @@ bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
>> struct lruvec *lruvec;
>> struct lru_gen_mm_state *mm_state;
>> unsigned long max_seq;
>> + pte_t *cur_pte;
>> int gen;
>>
>> lockdep_assert_held(pvmw->ptl);
>> VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio);
>>
>> - if (!ptep_test_and_clear_young_notify(vma, addr, pte))
>> + if (!test_and_clear_young_ptes_notify(vma, addr, pte, nr))
>> return false;
>>
>> if (spin_is_contended(pvmw->ptl))
>> @@ -4229,10 +4242,12 @@ bool lru_gen_look_around(struct page_vma_mapped_walk *pvmw)
>>
>> pte -= (addr - start) / PAGE_SIZE;
>>
>> - for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) {
>> + for (i = 0, addr = start, cur_pte = pte; addr != end;
>> + i += nr, cur_pte += nr, addr += nr * PAGE_SIZE) {
>> unsigned long pfn;
>> - pte_t ptent = ptep_get(pte + i);
>> + pte_t ptent = ptep_get(cur_pte);
>>
>> + nr = 1;
>
> Can't you just use pte and increment that, right?
>
> "pte" is not used afterwards.
Yes. Will do. Thanks for your comments.
More information about the linux-arm-kernel
mailing list