Skip to contentSkip to conversation

khugepaged: Optimize collapse_pte_mapped_thp() for large folios by PTE batching

PATCH In review Dev Jain 4 participants 9 messages Jun 18, 2025 – Jun 23, 2025

Subsystems (2)
lore
Share & tools
lore

Dev Jain

author reviewer
#0 lore raw

PATCH 1/1 khugepaged: Optimize collapse_pte_mapped_thp() for large folios by PTE batching

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

 mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------
 1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +25 -13
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 			    bool install_pmd)
 {
+	int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;
 	struct mmu_notifier_range range;
 	bool notified = false;
 	unsigned long haddr = addr & HPAGE_PMD_MASK;
+	unsigned long end = haddr + HPAGE_PMD_SIZE;
 	struct vm_area_struct *vma = vma_lookup(mm, haddr);
 	struct folio *folio;
 	pte_t *start_pte, *pte;
 	pmd_t *pmd, pgt_pmd;
 	spinlock_t *pml = NULL, *ptl;
-	int nr_ptes = 0, result = SCAN_FAIL;
 	int i;
 
 	mmap_assert_locked(mm);
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 	if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
 		goto abort;
 
+	i = 0, addr = haddr, pte = start_pte;
 	/* step 2: clear page table and adjust rmap */
-	for (i = 0, addr = haddr, pte = start_pte;
-	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+	do {
+		const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+		int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+		struct folio *this_folio;
 		struct page *page;
 		pte_t ptent = ptep_get(pte);
 
+		nr_batch_ptes = 1;
+
 		if (pte_none(ptent))
 			continue;
 		/*
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 			goto abort;
 		}
 		page = vm_normal_page(vma, addr, ptent);
+		this_folio = page_folio(page);
+		if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
+			nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, ptent,
+					max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
 		if (folio_page(folio, i) != page)
 			goto abort;
 
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 		 * TLB flush can be left until pmdp_collapse_flush() does it.
 		 * PTE dirty? Shmem page is already dirty; file is read-only.
 		 */
-		ptep_clear(mm, addr, pte);
-		folio_remove_rmap_pte(folio, page, vma);
-		nr_ptes++;
-	}
+		clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+		folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+		nr_mapped_ptes += nr_batch_ptes;
+	} while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+		 pte += nr_batch_ptes, i < HPAGE_PMD_NR);
 
 	if (!pml)
 		spin_unlock(ptl);
 
 	/* step 3: set proper refcount and mm_counters. */
-	if (nr_ptes) {
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+	if (nr_mapped_ptes) {
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
 	}
 
 	/* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 			: SCAN_SUCCEED;
 	goto drop_folio;
 abort:
-	if (nr_ptes) {
+	if (nr_mapped_ptes) {
 		flush_tlb_mm(mm);
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
 	}
 unlock:
 	if (start_pte)
Signature · 1 line
2.30.2
#1 lore raw

This series has a lot of duplication in it esp vs. your other series [0], but perhaps something we can tackle in a follow up.

It'd be nice if we could find a way to de-duplicate some of the near-identical code though.

But it's a 'maybe' on that because hey, the code in this file is hideous anyway and needs a mega rework in any case...

[0]: https://lore.kernel.org/all/[email protected]/

Quoted 18 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

Quoted 8 lines

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

Hmmm I say again - slow down a bit :) there's no need to shoot out multiple patches in a single day and you'd maybe avoid some of this kind of thing.

It's really preferable to avoid possible conflicts like this or at least reduce the chance by having review on one thing done first.

I mean, why not just put both of these in a series for the respin? Just a thought ;) in fact this is probably an ideal use of a series for that as you can ensure you deal with both if any conflicts arise.

mm/khugepaged.c static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, · line 1502

Quoted 13 lines
 mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------
 1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +1 -0
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
 int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 			    bool install_pmd)
 {
+	int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;

NIT: I don't know why you're moving this, and while y'know it's kind of the fun of subjective stuff I'd rather the assigned values and unassigned values be on different lines (yes I know this codebase violates this with the pml, ptl below but hey :P)

Quoted 14 lines
mm/khugepaged.c +1 -1
 	struct mmu_notifier_range range;
 	bool notified = false;
 	unsigned long haddr = addr & HPAGE_PMD_MASK;
+	unsigned long end = haddr + HPAGE_PMD_SIZE;
 	struct vm_area_struct *vma = vma_lookup(mm, haddr);
 	struct folio *folio;
 	pte_t *start_pte, *pte;
 	pmd_t *pmd, pgt_pmd;
 	spinlock_t *pml = NULL, *ptl;
-	int nr_ptes = 0, result = SCAN_FAIL;
 	int i;
 
 	mmap_assert_locked(mm);

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1624

Quoted 4 lines
mm/khugepaged.c +1 -0
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 	if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
 		goto abort;
 
+	i = 0, addr = haddr, pte = start_pte;

This is horrid, no absolutely not. This is not how we do assignment in arbitrary C code.

I don't know why we need a do/while here in general, I think the for loop should still work ok no?

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1629

Quoted 7 lines
mm/khugepaged.c +4 -2
 	/* step 2: clear page table and adjust rmap */
-	for (i = 0, addr = haddr, pte = start_pte;
-	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+	do {
+		const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+		int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+		struct folio *this_folio;

Hate this name. We are not C#... ;)

Just call it folio no? The 'this_' is redundant.

Quoted 62 lines
mm/khugepaged.c +18 -11
 		struct page *page;
 		pte_t ptent = ptep_get(pte);
 
+		nr_batch_ptes = 1;
+
 		if (pte_none(ptent))
 			continue;
 		/*
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 			goto abort;
 		}
 		page = vm_normal_page(vma, addr, ptent);
+		this_folio = page_folio(page);
+		if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
+			nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, ptent,
+					max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
 		if (folio_page(folio, i) != page)
 			goto abort;
 
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 		 * TLB flush can be left until pmdp_collapse_flush() does it.
 		 * PTE dirty? Shmem page is already dirty; file is read-only.
 		 */
-		ptep_clear(mm, addr, pte);
-		folio_remove_rmap_pte(folio, page, vma);
-		nr_ptes++;
-	}
+		clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+		folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+		nr_mapped_ptes += nr_batch_ptes;
+	} while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+		 pte += nr_batch_ptes, i < HPAGE_PMD_NR);
 
 	if (!pml)
 		spin_unlock(ptl);
 
 	/* step 3: set proper refcount and mm_counters. */
-	if (nr_ptes) {
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+	if (nr_mapped_ptes) {
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
 	}
 
 	/* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
 			: SCAN_SUCCEED;
 	goto drop_folio;
 abort:
-	if (nr_ptes) {
+	if (nr_mapped_ptes) {
 		flush_tlb_mm(mm);
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
 	}
 unlock:
 	if (start_pte)
--
Quoted 2 lines

2.30.2

Logic looks generally sane though... :)

author reviewer
#2 lore raw
Quoted 41 lines · Lorenzo Stoakes

This series has a lot of duplication in it esp vs. your other series [0], but perhaps something we can tackle in a follow up.

It'd be nice if we could find a way to de-duplicate some of the near-identical code though.

But it's a 'maybe' on that because hey, the code in this file is hideous anyway and needs a mega rework in any case...

[0]: https://lore.kernel.org/all/[email protected]/

Quoted 18 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

Quoted 8 lines

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

Hmmm I say again - slow down a bit :) there's no need to shoot out multiple patches in a single day and you'd maybe avoid some of this kind of thing.

Quoted 6 lines

It's really preferable to avoid possible conflicts like this or at least reduce the chance by having review on one thing done first.

I mean, why not just put both of these in a series for the respin? Just a thought ;) in fact this is probably an ideal use of a series for that as you can ensure you deal with both if any conflicts arise.

Sorry for this. I found these two patches independently with a couple of hours difference, and I posted this one yesterday because stupid me thought someone will, after seeing my first patch, point out that the optimization can be made at one more place. So I will send this and the other patch as a series anyway for v2.

Quoted 4 lines
Quoted 4 lines
  mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------
  1 file changed, 25 insertions(+), 13 deletions(-)
Quoted 13 lines

mm/khugepaged.c static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, · line 1502

Quoted 9 lines
mm/khugepaged.c +1 -0
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
  int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			    bool install_pmd)
  {
+	int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;

NIT: I don't know why you're moving this, and while y'know it's kind of the fun of subjective stuff I'd rather the assigned values and unassigned values be on different lines (yes I know this codebase violates this with the pml, ptl below but hey :P)

To maintain a reverse Xmas fashion. Now I know that the declarations are already not in an Xmas fashion, but I wanted to make sure the code I change maintains that for the part I am changing :)

Quoted 18 lines
Quoted 18 lines
mm/khugepaged.c +1 -1
  	struct mmu_notifier_range range;
  	bool notified = false;
  	unsigned long haddr = addr & HPAGE_PMD_MASK;
+	unsigned long end = haddr + HPAGE_PMD_SIZE;
  	struct vm_area_struct *vma = vma_lookup(mm, haddr);
  	struct folio *folio;
  	pte_t *start_pte, *pte;
  	pmd_t *pmd, pgt_pmd;
  	spinlock_t *pml = NULL, *ptl;
-	int nr_ptes = 0, result = SCAN_FAIL;
  	int i;
 
  	mmap_assert_locked(mm);
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  	if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
  		goto abort;
 
Quoted 6 lines

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1624

Quoted 1 line
mm/khugepaged.c +1 -0
+	i = 0, addr = haddr, pte = start_pte;

This is horrid, no absolutely not. This is not how we do assignment in arbitrary C code.

I don't know why we need a do/while here in general, I think the for loop should still work ok no?

I don't recall now and I cannot even find it, but I have been following this pattern for some time, by god I cannot remember which pattern I am copying from. Anyhow I also hate the do/while thingy so I will change this to a for loop.

Quoted 11 lines

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1629

Quoted 8 lines
mm/khugepaged.c +4 -2
  	/* step 2: clear page table and adjust rmap */
-	for (i = 0, addr = haddr, pte = start_pte;
-	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+	do {
+		const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+		int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+		struct folio *this_folio;

Hate this name. We are not C#... ;)

Just call it folio no? The 'this_' is redundant.

There is already a struct folio *folio which we retrieve from filemap_lock_folio. So wanted to differentiate, I didn't think hard about the name. How about mapped_folio?

Quoted 67 lines
Quoted 64 lines
mm/khugepaged.c +18 -11
  		struct page *page;
  		pte_t ptent = ptep_get(pte);
 
+		nr_batch_ptes = 1;
+
  		if (pte_none(ptent))
  			continue;
  		/*
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			goto abort;
  		}
  		page = vm_normal_page(vma, addr, ptent);
+		this_folio = page_folio(page);
+		if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
+			nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, ptent,
+					max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
  		if (folio_page(folio, i) != page)
  			goto abort;
 
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  		 * TLB flush can be left until pmdp_collapse_flush() does it.
  		 * PTE dirty? Shmem page is already dirty; file is read-only.
  		 */
-		ptep_clear(mm, addr, pte);
-		folio_remove_rmap_pte(folio, page, vma);
-		nr_ptes++;
-	}
+		clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+		folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+		nr_mapped_ptes += nr_batch_ptes;
+	} while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+		 pte += nr_batch_ptes, i < HPAGE_PMD_NR);
 
  	if (!pml)
  		spin_unlock(ptl);
 
  	/* step 3: set proper refcount and mm_counters. */
-	if (nr_ptes) {
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+	if (nr_mapped_ptes) {
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
  	}
 
  	/* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			: SCAN_SUCCEED;
  	goto drop_folio;
  abort:
-	if (nr_ptes) {
+	if (nr_mapped_ptes) {
  		flush_tlb_mm(mm);
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
  	}
  unlock:
  	if (start_pte)
--
Quoted 2 lines

2.30.2

Logic looks generally sane though... :)

#3 lore raw
Quoted 50 lines · Dev Jain
Quoted 41 lines · Lorenzo Stoakes

This series has a lot of duplication in it esp vs. your other series [0], but perhaps something we can tackle in a follow up.

It'd be nice if we could find a way to de-duplicate some of the near-identical code though.

But it's a 'maybe' on that because hey, the code in this file is hideous anyway and needs a mega rework in any case...

[0]: https://lore.kernel.org/all/[email protected]/

Quoted 18 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

Quoted 8 lines

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

Hmmm I say again - slow down a bit :) there's no need to shoot out multiple patches in a single day and you'd maybe avoid some of this kind of thing.

Quoted 6 lines

It's really preferable to avoid possible conflicts like this or at least reduce the chance by having review on one thing done first.

I mean, why not just put both of these in a series for the respin? Just a thought ;) in fact this is probably an ideal use of a series for that as you can ensure you deal with both if any conflicts arise.

Quoted 5 lines

Sorry for this. I found these two patches independently with a couple of hours difference, and I posted this one yesterday because stupid me thought someone will, after seeing my first patch, point out that the optimization can be made at one more place. So I will send this and the other patch as a series anyway for v2.

Sure, thanks, it'll just make life easier.

Quoted 19 lines
Quoted 5 lines
Quoted 5 lines

  mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------
  1 file changed, 25 insertions(+), 13 deletions(-)
Quoted 13 lines

mm/khugepaged.c static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, · line 1502

Quoted 9 lines
mm/khugepaged.c +1 -0
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
  int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			    bool install_pmd)
  {
+	int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;

NIT: I don't know why you're moving this, and while y'know it's kind of the fun of subjective stuff I'd rather the assigned values and unassigned values be on different lines (yes I know this codebase violates this with the pml, ptl below but hey :P)

Quoted 3 lines

To maintain a reverse Xmas fashion. Now I know that the declarations are already not in an Xmas fashion, but I wanted to make sure the code I change maintains that for the part I am changing :)

We have no such requirement in mm nor do we particularly want to establish any conventions around this.

I've already read enough stupid articles about unreasonable kernel devs insist on yada yada... let's just keep it sensible and logical! :)

Quoted 23 lines
Quoted 23 lines
Quoted 16 lines
mm/khugepaged.c +1 -1
  	struct mmu_notifier_range range;
  	bool notified = false;
  	unsigned long haddr = addr & HPAGE_PMD_MASK;
+	unsigned long end = haddr + HPAGE_PMD_SIZE;
  	struct vm_area_struct *vma = vma_lookup(mm, haddr);
  	struct folio *folio;
  	pte_t *start_pte, *pte;
  	pmd_t *pmd, pgt_pmd;
  	spinlock_t *pml = NULL, *ptl;
-	int nr_ptes = 0, result = SCAN_FAIL;
  	int i;
 
  	mmap_assert_locked(mm);

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1624

Quoted 4 lines
mm/khugepaged.c +1 -0
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  	if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
  		goto abort;
 
+	i = 0, addr = haddr, pte = start_pte;

This is horrid, no absolutely not. This is not how we do assignment in arbitrary C code.

Quoted 7 lines
Quoted 2 lines

I don't know why we need a do/while here in general, I think the for loop should still work ok no?

I don't recall now and I cannot even find it, but I have been following this pattern for some time, by god I cannot remember which pattern I am copying from. Anyhow I also hate the do/while thingy so I will change this to a for loop.

Thanks

Quoted 11 lines
Quoted 11 lines

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1629

Quoted 9 lines
mm/khugepaged.c +4 -2
  	/* step 2: clear page table and adjust rmap */
-	for (i = 0, addr = haddr, pte = start_pte;
-	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+	do {
+		const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+		int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+		struct folio *this_folio;

Hate this name. We are not C#... ;)

Quoted 4 lines
Quoted 1 line

Just call it folio no? The 'this_' is redundant.

There is already a struct folio *folio which we retrieve from filemap_lock_folio. So wanted to differentiate, I didn't think hard about the name. How about mapped_folio?

Ah damn, ok that sounds better, thanks!

#4 lore raw
Quoted 18 lines
Quoted 18 lines
Quoted 17 lines

mm/khugepaged.c static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr, · line 1502

Quoted 13 lines
   mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------
   1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +1 -0
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
   int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
   			    bool install_pmd)
   {
+	int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;

NIT: I don't know why you're moving this, and while y'know it's kind of the fun of subjective stuff I'd rather the assigned values and unassigned values be on different lines (yes I know this codebase violates this with the pml, ptl below but hey :P)

Quoted 6 lines
Quoted 3 lines

To maintain a reverse Xmas fashion. Now I know that the declarations are already not in an Xmas fashion, but I wanted to make sure the code I change maintains that for the part I am changing :)

We have no such requirement in mm nor do we particularly want to establish any conventions around this.

Well, if we already do have reverse xmas tree, we tend to maintain it

... or when adding new code.

So in MM it's actually very common to do that. But we avoid doing that just for the sake of it when touching some code.

Signature · 3 lines
Cheers,

David / dhildenb
reviewer
#5 lore raw
In reply to Dev Jain
Quoted 73 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

  mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------
  1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +9 -3
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct *vma, unsigned long addr,
  int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			    bool install_pmd)
  {
+	int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;
  	struct mmu_notifier_range range;
  	bool notified = false;
  	unsigned long haddr = addr & HPAGE_PMD_MASK;
+	unsigned long end = haddr + HPAGE_PMD_SIZE;
  	struct vm_area_struct *vma = vma_lookup(mm, haddr);
  	struct folio *folio;
  	pte_t *start_pte, *pte;
  	pmd_t *pmd, pgt_pmd;
  	spinlock_t *pml = NULL, *ptl;
-	int nr_ptes = 0, result = SCAN_FAIL;
  	int i;
 
  	mmap_assert_locked(mm);
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  	if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
  		goto abort;
 
+	i = 0, addr = haddr, pte = start_pte;
  	/* step 2: clear page table and adjust rmap */
-	for (i = 0, addr = haddr, pte = start_pte;
-	     i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+	do {
+		const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+		int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+		struct folio *this_folio;
  		struct page *page;
  		pte_t ptent = ptep_get(pte);
 
+		nr_batch_ptes = 1;
+
  		if (pte_none(ptent))
  			continue;
  		/*

mm/khugepaged.c int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr, · line 1654

Quoted 10 lines
mm/khugepaged.c +5 -0
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			goto abort;
  		}
  		page = vm_normal_page(vma, addr, ptent);
+		this_folio = page_folio(page);
+		if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
+			nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, ptent,
+					max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
  		if (folio_page(folio, i) != page)
  			goto abort;

IMO, 'this_folio' is always equal 'folio', right? Can't we just use 'folio'?

In addition, I think the folio_test_large() and max_nr_batch_ptes checks are redundant, since the 'folio' must be PMD-sized large folio after 'folio_page(folio, i) != page' check.

So I think we can move the 'nr_batch_ptes' calculation after the folio_page() check, then shoule be:

nr_batch_ptes = folio_pte_batch(folio, addr, pte, ptent,

			max_nr_batch_ptes, flags, NULL, NULL, NULL);
Quoted 41 lines
mm/khugepaged.c +11 -10
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  		 * TLB flush can be left until pmdp_collapse_flush() does it.
  		 * PTE dirty? Shmem page is already dirty; file is read-only.
  		 */
-		ptep_clear(mm, addr, pte);
-		folio_remove_rmap_pte(folio, page, vma);
-		nr_ptes++;
-	}
+		clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+		folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+		nr_mapped_ptes += nr_batch_ptes;
+	} while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+		 pte += nr_batch_ptes, i < HPAGE_PMD_NR);
 
  	if (!pml)
  		spin_unlock(ptl);
 
  	/* step 3: set proper refcount and mm_counters. */
-	if (nr_ptes) {
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+	if (nr_mapped_ptes) {
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
  	}
 
  	/* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
  			: SCAN_SUCCEED;
  	goto drop_folio;
  abort:
-	if (nr_ptes) {
+	if (nr_mapped_ptes) {
  		flush_tlb_mm(mm);
-		folio_ref_sub(folio, nr_ptes);
-		add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+		folio_ref_sub(folio, nr_mapped_ptes);
+		add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
  	}
  unlock:
  	if (start_pte)
author reviewer
#6 lore raw
Quoted 81 lines · Baolin Wang
Quoted 78 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

  mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------

  1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +7 -3
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct 
 *vma, unsigned long addr,
   int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
                   bool install_pmd)
   {
+    int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;
       struct mmu_notifier_range range;
       bool notified = false;
       unsigned long haddr = addr & HPAGE_PMD_MASK;
+    unsigned long end = haddr + HPAGE_PMD_SIZE;
       struct vm_area_struct *vma = vma_lookup(mm, haddr);
       struct folio *folio;
       pte_t *start_pte, *pte;
       pmd_t *pmd, pgt_pmd;
       spinlock_t *pml = NULL, *ptl;
-    int nr_ptes = 0, result = SCAN_FAIL;
       int i;
         mmap_assert_locked(mm);
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
       if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
           goto abort;
   +    i = 0, addr = haddr, pte = start_pte;
       /* step 2: clear page table and adjust rmap */
-    for (i = 0, addr = haddr, pte = start_pte;
-         i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+    do {
+        const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+        int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+        struct folio *this_folio;
           struct page *page;
           pte_t ptent = ptep_get(pte);
   +        nr_batch_ptes = 1;
+
           if (pte_none(ptent))
               continue;

          /*

mm/khugepaged.c +2 -0
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
               goto abort;
           }
           page = vm_normal_page(vma, addr, ptent);
+        this_folio = page_folio(page);
+        if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
Quoted 8 lines
Quoted 6 lines
mm/khugepaged.c +3 -0
+            nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, 
 ptent,
+                    max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
           if (folio_page(folio, i) != page)

              goto abort;

IMO, 'this_folio' is always equal 'folio', right? Can't we just use 'folio'?

I don't think so. What if we have mremapped some bytes of this PMD range

to point to another folio.

Quoted 2 lines

In addition, I think the folio_test_large() and max_nr_batch_ptes checks are redundant, since the 'folio' must be PMD-sized large folio after 'folio_page(folio, i) != page' check.

As an improvement we can at least do likely(folio_test_large()) since this is very likely.

Quoted 47 lines

So I think we can move the 'nr_batch_ptes' calculation after the folio_page() check, then shoule be:

nr_batch_ptes = folio_pte_batch(folio, addr, pte, ptent, max_nr_batch_ptes, flags, NULL, NULL, NULL);

Quoted 40 lines
mm/khugepaged.c +8 -7
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
            * TLB flush can be left until pmdp_collapse_flush() does it.
            * PTE dirty? Shmem page is already dirty; file is read-only.
            */
-        ptep_clear(mm, addr, pte);
-        folio_remove_rmap_pte(folio, page, vma);
-        nr_ptes++;
-    }
+        clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+        folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+        nr_mapped_ptes += nr_batch_ptes;
+    } while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+         pte += nr_batch_ptes, i < HPAGE_PMD_NR);
         if (!pml)
           spin_unlock(ptl);
         /* step 3: set proper refcount and mm_counters. */
-    if (nr_ptes) {
-        folio_ref_sub(folio, nr_ptes);
-        add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+    if (nr_mapped_ptes) {
+        folio_ref_sub(folio, nr_mapped_ptes);
+        add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
       }
         /* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
               : SCAN_SUCCEED;
       goto drop_folio;

  abort:
-    if (nr_ptes) {
+    if (nr_mapped_ptes) {
          flush_tlb_mm(mm);
-        folio_ref_sub(folio, nr_ptes);
-        add_mm_counter(mm, mm_counter_file(folio), -nr_ptes); +        folio_ref_sub(folio, nr_mapped_ptes);
+        add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes); }
  unlock:
      if (start_pte)

reviewer
#7 lore raw
Quoted 92 lines · Dev Jain
Quoted 81 lines · Baolin Wang
Quoted 78 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

  mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------

  1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +7 -3
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct vm_area_struct 
 *vma, unsigned long addr,
   int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
                   bool install_pmd)
   {
+    int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;
       struct mmu_notifier_range range;
       bool notified = false;
       unsigned long haddr = addr & HPAGE_PMD_MASK;
+    unsigned long end = haddr + HPAGE_PMD_SIZE;
       struct vm_area_struct *vma = vma_lookup(mm, haddr);
       struct folio *folio;
       pte_t *start_pte, *pte;
       pmd_t *pmd, pgt_pmd;
       spinlock_t *pml = NULL, *ptl;
-    int nr_ptes = 0, result = SCAN_FAIL;
       int i;
         mmap_assert_locked(mm);
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
       if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
           goto abort;
   +    i = 0, addr = haddr, pte = start_pte;
       /* step 2: clear page table and adjust rmap */
-    for (i = 0, addr = haddr, pte = start_pte;
-         i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+    do {
+        const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+        int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+        struct folio *this_folio;
           struct page *page;
           pte_t ptent = ptep_get(pte);
   +        nr_batch_ptes = 1;
+
           if (pte_none(ptent))
               continue;

          /*

mm/khugepaged.c +2 -0
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
               goto abort;
           }
           page = vm_normal_page(vma, addr, ptent);
+        this_folio = page_folio(page);
+        if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
Quoted 8 lines
Quoted 6 lines
mm/khugepaged.c +3 -0
+            nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, 
 ptent,
+                    max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
           if (folio_page(folio, i) != page)

              goto abort;

IMO, 'this_folio' is always equal 'folio', right? Can't we just use 'folio'?

Quoted 3 lines

I don't think so. What if we have mremapped some bytes of this PMD range

to point to another folio.

Then 'folio_page(folio, i) != page' can catch this, which is why I suggest you move the 'nr_batch_ptes' calculation after the folio_page() check.

Quoted 52 lines
Quoted 1 line

In addition, I think the folio_test_large() and max_nr_batch_ptes checks are redundant, since the 'folio' must be PMD-sized large folio after 'folio_page(folio, i) != page' check.

As an improvement we can at least do likely(folio_test_large()) since this is very likely.

Quoted 47 lines

So I think we can move the 'nr_batch_ptes' calculation after the folio_page() check, then shoule be:

nr_batch_ptes = folio_pte_batch(folio, addr, pte, ptent, max_nr_batch_ptes, flags, NULL, NULL, NULL);

Quoted 40 lines
mm/khugepaged.c +8 -7
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
            * TLB flush can be left until pmdp_collapse_flush() does it.
            * PTE dirty? Shmem page is already dirty; file is read-only.
            */
-        ptep_clear(mm, addr, pte);
-        folio_remove_rmap_pte(folio, page, vma);
-        nr_ptes++;
-    }
+        clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+        folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+        nr_mapped_ptes += nr_batch_ptes;
+    } while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+         pte += nr_batch_ptes, i < HPAGE_PMD_NR);
         if (!pml)
           spin_unlock(ptl);
         /* step 3: set proper refcount and mm_counters. */
-    if (nr_ptes) {
-        folio_ref_sub(folio, nr_ptes);
-        add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+    if (nr_mapped_ptes) {
+        folio_ref_sub(folio, nr_mapped_ptes);
+        add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
       }
         /* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
               : SCAN_SUCCEED;
       goto drop_folio;

  abort:
-    if (nr_ptes) {
+    if (nr_mapped_ptes) {
          flush_tlb_mm(mm);
-        folio_ref_sub(folio, nr_ptes);
-        add_mm_counter(mm, mm_counter_file(folio), -nr_ptes); +        folio_ref_sub(folio, nr_mapped_ptes);
+        add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes); }
  unlock:
      if (start_pte)

author reviewer
#8 lore raw
Quoted 95 lines · Baolin Wang
Quoted 92 lines · Dev Jain
Quoted 89 lines · Baolin Wang
Quoted 83 lines · Dev Jain

Use PTE batching to optimize collapse_pte_mapped_thp().

On arm64, suppose khugepaged is scanning a pte-mapped 2MB THP for collapse. Then, calling ptep_clear() for every pte will cause a TLB flush for every contpte block. Instead, clear_full_ptes() does a contpte_try_unfold_partial() which will flush the TLB only for the (if any) starting and ending contpte block, if they partially overlap with the range khugepaged is looking at.

For all arches, there should be a benefit due to batching atomic operations on mapcounts due to folio_remove_rmap_ptes().

Note that we do not need to make a change to the check
"if (folio_page(folio, i) != page)"; if i'th page of the folio is equal to the first page of our batch, then i + 1, .... i + nr_batch_ptes - 1 pages of the folio will be equal to the corresponding pages of our batch mapping consecutive pages.

No issues were observed with mm-selftests.

---

This is rebased on:
https://lore.kernel.org/all/[email protected]/
If there will be a v2 of either version I'll send them together.

  mm/khugepaged.c | 38 +++++++++++++++++++++++++-------------

  1 file changed, 25 insertions(+), 13 deletions(-)
mm/khugepaged.c +7 -3
index 649ccb2670f8..7d37058eda5b 100644
@@ -1499,15 +1499,16 @@ static int set_huge_pmd(struct 
 vm_area_struct *vma, unsigned long addr,
   int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
                   bool install_pmd)
   {
+    int nr_mapped_ptes = 0, nr_batch_ptes, result = SCAN_FAIL;
       struct mmu_notifier_range range;
       bool notified = false;
       unsigned long haddr = addr & HPAGE_PMD_MASK;
+    unsigned long end = haddr + HPAGE_PMD_SIZE;
       struct vm_area_struct *vma = vma_lookup(mm, haddr);
       struct folio *folio;
       pte_t *start_pte, *pte;
       pmd_t *pmd, pgt_pmd;
       spinlock_t *pml = NULL, *ptl;
-    int nr_ptes = 0, result = SCAN_FAIL;
       int i;
         mmap_assert_locked(mm);
@@ -1620,12 +1621,17 @@ int collapse_pte_mapped_thp(struct 
 mm_struct *mm, unsigned long addr,
       if (unlikely(!pmd_same(pgt_pmd, pmdp_get_lockless(pmd))))
           goto abort;
   +    i = 0, addr = haddr, pte = start_pte;
       /* step 2: clear page table and adjust rmap */
-    for (i = 0, addr = haddr, pte = start_pte;
-         i < HPAGE_PMD_NR; i++, addr += PAGE_SIZE, pte++) {
+    do {
+        const fpb_t flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
+        int max_nr_batch_ptes = (end - addr) >> PAGE_SHIFT;
+        struct folio *this_folio;
           struct page *page;
           pte_t ptent = ptep_get(pte);
   +        nr_batch_ptes = 1;
+
           if (pte_none(ptent))
               continue;

          /*

mm/khugepaged.c +5 -0
@@ -1639,6 +1645,11 @@ int collapse_pte_mapped_thp(struct mm_struct 
 *mm, unsigned long addr,
               goto abort;
           }
           page = vm_normal_page(vma, addr, ptent);
+        this_folio = page_folio(page);
+        if (folio_test_large(this_folio) && max_nr_batch_ptes != 1)
+            nr_batch_ptes = folio_pte_batch(this_folio, addr, pte, 
 ptent,
+                    max_nr_batch_ptes, flags, NULL, NULL, NULL);
+
           if (folio_page(folio, i) != page)
Quoted 1 line

              goto abort;

IMO, 'this_folio' is always equal 'folio', right? Can't we just use 'folio'?

Quoted 5 lines
Quoted 3 lines

I don't think so. What if we have mremapped some bytes of this PMD range

to point to another folio.

Then 'folio_page(folio, i) != page' can catch this, which is why I suggest you move the 'nr_batch_ptes' calculation after the folio_page() check.

Ah. page will be part of the folio, so the folio derived from page is equal to folio, so it is large. Thanks.

Quoted 54 lines
Quoted 53 lines
Quoted 2 lines

In addition, I think the folio_test_large() and max_nr_batch_ptes checks are redundant, since the 'folio' must be PMD-sized large folio after 'folio_page(folio, i) != page' check.

As an improvement we can at least do likely(folio_test_large()) since this is very likely.

Quoted 47 lines

So I think we can move the 'nr_batch_ptes' calculation after the folio_page() check, then shoule be:

nr_batch_ptes = folio_pte_batch(folio, addr, pte, ptent, max_nr_batch_ptes, flags, NULL, NULL, NULL);

Quoted 40 lines
mm/khugepaged.c +8 -7
@@ -1647,18 +1658,19 @@ int collapse_pte_mapped_thp(struct 
 mm_struct *mm, unsigned long addr,
            * TLB flush can be left until pmdp_collapse_flush() does it.
            * PTE dirty? Shmem page is already dirty; file is read-only.
            */
-        ptep_clear(mm, addr, pte);
-        folio_remove_rmap_pte(folio, page, vma);
-        nr_ptes++;
-    }
+        clear_full_ptes(mm, addr, pte, nr_batch_ptes, false);
+        folio_remove_rmap_ptes(folio, page, nr_batch_ptes, vma);
+        nr_mapped_ptes += nr_batch_ptes;
+    } while (i += nr_batch_ptes, addr += nr_batch_ptes * PAGE_SIZE,
+         pte += nr_batch_ptes, i < HPAGE_PMD_NR);
         if (!pml)
           spin_unlock(ptl);
         /* step 3: set proper refcount and mm_counters. */
-    if (nr_ptes) {
-        folio_ref_sub(folio, nr_ptes);
-        add_mm_counter(mm, mm_counter_file(folio), -nr_ptes);
+    if (nr_mapped_ptes) {
+        folio_ref_sub(folio, nr_mapped_ptes);
+        add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes);
       }
         /* step 4: remove empty page table */
@@ -1691,10 +1703,10 @@ int collapse_pte_mapped_thp(struct 
 mm_struct *mm, unsigned long addr,
               : SCAN_SUCCEED;
       goto drop_folio;

  abort:
-    if (nr_ptes) {
+    if (nr_mapped_ptes) {
          flush_tlb_mm(mm);
-        folio_ref_sub(folio, nr_ptes);
-        add_mm_counter(mm, mm_counter_file(folio), -nr_ptes); +        folio_ref_sub(folio, nr_mapped_ptes);
+        add_mm_counter(mm, mm_counter_file(folio), -nr_mapped_ptes); }
  unlock:
      if (start_pte)