Skip to contentSkip to conversation

Reclaim lazyfree THP without splitting

PATCH In review Lance Yang 3 participants 11 messages Jun 14, 2024 – Jun 22, 2024

Subsystems (3)
lore
Share & tools
lore

Lance Yang

author
#0 lore raw

v8 cover Reclaim lazyfree THP without splitting

Hi all,

This series adds support for reclaiming PMD-mapped THP marked as lazyfree without needing to first split the large folio via split_huge_pmd_address().

When the user no longer requires the pages, they would use madvise(MADV_FREE) to mark the pages as lazy free. Subsequently, they typically would not re-write to that memory again.

During memory reclaim, if we detect that the large folio and its PMD are both still marked as clean and there are no unexpected references(such as GUP), so we can just discard the memory lazily, improving the efficiency of memory reclamation in this case.

Performance Testing
===================

On an Intel i5 CPU, reclaiming 1GiB of lazyfree THPs using mem_cgroup_force_empty() results in the following runtimes in seconds (shorter is better):

--------------------------------------------
| Old | New | Change |
--------------------------------------------
| 0.683426 | 0.049197 | -92.80% |
--------------------------------------------

---

Changes since v7 [7]
====================
- mm/rmap: remove duplicated exit code in pagewalk loop
- Pick RB from Barry - thanks!
- Rename walk_done_err to walk_abort (per Barry and David)
- mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop
- Make page_vma_mapped_walk_restart() more general (per David)
- Squash page_vma_mapped_walk_restart() into this patch (per David)
- mm/vmscan: avoid split lazyfree THP during shrink_folio_list()
- Don't unmark a PMD-mapped folio as lazyfree in unmap_huge_pmd_locked()
- Drop the unused "pmd_mapped" variable (per Baolin)

Changes since v6 [6]
====================
- mm/rmap: remove duplicated exit code in pagewalk loop
- Pick RB from David - thanks!
- mm/rmap: add helper to restart pgtable walk on changes
- Add the page_vma_mapped_walk_restart() helper to handle scenarios where the page table walk needs to be restarted due to changes in the page table (suggested by David)
- mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop
- Pass 'pvmw.address' to split_huge_pmd_locked() (per David)
- Drop the check for PMD-mapped THP that is missing the mlock (per David)
- mm/vmscan: avoid split lazyfree THP during shrink_folio_list()
- Rename the function __discard_trans_pmd_locked() to __discard_anon_folio_pmd_locked() (per David)

Changes since v5 [5]
====================
- mm/rmap: remove duplicated exit code in pagewalk loop
- Pick RB from Baolin Wang - thanks!
- mm/mlock: check for THP missing the mlock in try_to_unmap_one()
- Merge this patch into patch 2 (per Baolin Wang)
- mm/vmscan: avoid split lazyfree THP during shrink_folio_list()
- Mark a folio as being backed by swap space if the folio or its PMD was redirtied (per Baolin Wang)
- Use pmdp_huge_clear_flush() to get and flush a PMD entry (per Baolin Wang)

Changes since v4 [4]
====================
- mm/rmap: remove duplicated exit code in pagewalk loop
- Pick RB from Zi Yan - thanks!
- mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop
- Remove the redundant alignment (per Baolin Wang)
- Set pvmw.ptl to NULL after unlocking the PTL (per Baolin Wang)
- mm/mlock: check for THP missing the mlock in try_to_unmap_one()
- Check whether the mlock of PMD-mapped THP was missed (suggested by Baolin Wang)
- mm/vmscan: avoid split lazyfree THP during shrink_folio_list()
- No need to check the TTU_SPLIT_HUGE_PMD flag for unmap_huge_pmd_locked() (per Zi Yan)
- Drain the local mlock batch after folio_remove_rmap_pmd() (per Baolin Wang)

Changes since v3 [3]
====================
- mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop
- Resolve compilation errors by handling the case where CONFIG_PGTABLE_HAS_HUGE_LEAVES is undefined (thanks to SeongJae Park)
- mm/vmscan: avoid split lazyfree THP during shrink_folio_list()
- Remove the unnecessary conditional compilation directives (thanks to Barry Song)
- Resolve compilation errors due to undefined references to unmap_huge_pmd_locked and split_huge_pmd_locked (thanks to Barry)

Changes since v2 [2]
====================
- Update the changelog (thanks to David Hildenbrand)
- Support try_to_unmap_one() to unmap PMD-mapped folios (thanks a lot to David Hildenbrand and Zi Yan)

Changes since v1 [1]
====================
- Update the changelog
- Follow the exact same logic as in try_to_unmap_one() (per David Hildenbrand)
- Remove the extra code from rmap.c (per Matthew Wilcox)

[1] https://lore.kernel.org/linux-mm/[email protected]
[2] https://lore.kernel.org/linux-mm/[email protected]
[3] https://lore.kernel.org/linux-mm/[email protected]
[4] https://lore.kernel.org/linux-mm/[email protected]
[5] https://lore.kernel.org/linux-mm/[email protected]
[6] https://lore.kernel.org/linux-mm/[email protected]
[7] https://lore.kernel.org/linux-mm/[email protected]

Lance Yang (3):
mm/rmap: remove duplicated exit code in pagewalk loop mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop mm/vmscan: avoid split lazyfree THP during shrink_folio_list()

 include/linux/huge_mm.h |  15 +++++
 include/linux/rmap.h    |  24 ++++++++
 mm/huge_memory.c        | 118 +++++++++++++++++++++++++++++++++-------
 mm/rmap.c               |  68 ++++++++++++-----------
 4 files changed, 174 insertions(+), 51 deletions(-)

base-commit: fb8d20fa1a94f807336ed209d33da8ec15ae6c3a

Signature · 1 line
2.33.1
author
#1 lore raw

PATCH v8 1/3 mm/rmap: remove duplicated exit code in pagewalk loop

Introduce the labels walk_done and walk_abort as exit points to eliminate duplicated exit code in the pagewalk loop.

---
 mm/rmap.c | 40 +++++++++++++++-------------------------
 1 file changed, 15 insertions(+), 25 deletions(-)
mm/rmap.c +15 -25
index ae250b2b4d55..2d778725e4f5 100644
@@ -1681,9 +1681,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 			/* Restore the mlock which got missed */
 			if (!folio_test_large(folio))
 				mlock_vma_folio(folio, vma);
-			page_vma_mapped_walk_done(&pvmw);
-			ret = false;
-			break;
+			goto walk_abort;
 		}
 
 		pfn = pte_pfn(ptep_get(pvmw.pte));
@@ -1721,11 +1719,8 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 			 */
 			if (!anon) {
 				VM_BUG_ON(!(flags & TTU_RMAP_LOCKED));
-				if (!hugetlb_vma_trylock_write(vma)) {
-					page_vma_mapped_walk_done(&pvmw);
-					ret = false;
-					break;
-				}
+				if (!hugetlb_vma_trylock_write(vma))
+					goto walk_abort;
 				if (huge_pmd_unshare(mm, vma, address, pvmw.pte)) {
 					hugetlb_vma_unlock_write(vma);
 					flush_tlb_range(vma,
@@ -1740,8 +1735,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 					 * actual page and drop map count
 					 * to zero.
 					 */
-					page_vma_mapped_walk_done(&pvmw);
-					break;
+					goto walk_done;
 				}
 				hugetlb_vma_unlock_write(vma);
 			}
@@ -1813,9 +1807,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 			if (unlikely(folio_test_swapbacked(folio) !=
 					folio_test_swapcache(folio))) {
 				WARN_ON_ONCE(1);
-				ret = false;
-				page_vma_mapped_walk_done(&pvmw);
-				break;
+				goto walk_abort;
 			}
 
 			/* MADV_FREE page check */
@@ -1854,23 +1846,17 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 				 */
 				set_pte_at(mm, address, pvmw.pte, pteval);
 				folio_set_swapbacked(folio);
-				ret = false;
-				page_vma_mapped_walk_done(&pvmw);
-				break;
+				goto walk_abort;
 			}
 
 			if (swap_duplicate(entry) < 0) {
 				set_pte_at(mm, address, pvmw.pte, pteval);
-				ret = false;
-				page_vma_mapped_walk_done(&pvmw);
-				break;
+				goto walk_abort;
 			}
 			if (arch_unmap_one(mm, vma, address, pteval) < 0) {
 				swap_free(entry);
 				set_pte_at(mm, address, pvmw.pte, pteval);
-				ret = false;
-				page_vma_mapped_walk_done(&pvmw);
-				break;
+				goto walk_abort;
 			}
 
 			/* See folio_try_share_anon_rmap(): clear PTE first. */
@@ -1878,9 +1864,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 			    folio_try_share_anon_rmap_pte(folio, subpage)) {
 				swap_free(entry);
 				set_pte_at(mm, address, pvmw.pte, pteval);
-				ret = false;
-				page_vma_mapped_walk_done(&pvmw);
-				break;
+				goto walk_abort;
 			}
 			if (list_empty(&mm->mmlist)) {
 				spin_lock(&mmlist_lock);
@@ -1920,6 +1904,12 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 		if (vma->vm_flags & VM_LOCKED)
 			mlock_drain_local();
 		folio_put(folio);
+		continue;
+walk_abort:
+		ret = false;
+walk_done:
+		page_vma_mapped_walk_done(&pvmw);
+		break;
 	}
 
 	mmu_notifier_invalidate_range_end(&range);
Signature · 1 line
2.33.1
author
#2 lore raw

PATCH v8 2/3 mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop

In reply to Lance Yang

In preparation for supporting try_to_unmap_one() to unmap PMD-mapped folios, start the pagewalk first, then call split_huge_pmd_address() to split the folio.

---
 include/linux/huge_mm.h |  6 ++++++
 include/linux/rmap.h    | 24 +++++++++++++++++++++++
 mm/huge_memory.c        | 42 +++++++++++++++++++++--------------------
 mm/rmap.c               | 21 +++++++++++++++------
 4 files changed, 67 insertions(+), 26 deletions(-)
include/linux/huge_mm.h +6 -0
index 7ad41de5eaea..9f720b0731c4 100644
@@ -428,6 +428,9 @@ static inline bool thp_migration_supported(void)
 	return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 }
 
+void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
+			   pmd_t *pmd, bool freeze, struct folio *folio);
+
 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 
 static inline bool folio_test_pmd_mappable(struct folio *folio)
@@ -490,6 +493,9 @@ static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 		unsigned long address, bool freeze, struct folio *folio) {}
 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 		unsigned long address, bool freeze, struct folio *folio) {}
+static inline void split_huge_pmd_locked(struct vm_area_struct *vma,
+					 unsigned long address, pmd_t *pmd,
+					 bool freeze, struct folio *folio) {}
 
 #define split_huge_pud(__vma, __pmd, __address)	\
 	do { } while (0)
include/linux/rmap.h +24 -0
index 0fd9bebce54c..d1c5e2d694b2 100644
@@ -703,6 +703,30 @@ static inline void page_vma_mapped_walk_done(struct page_vma_mapped_walk *pvmw)
 		spin_unlock(pvmw->ptl);
 }
 
+/**
+ * page_vma_mapped_walk_restart - Restart the page table walk.
+ * @pvmw: Pointer to struct page_vma_mapped_walk.
+ *
+ * It restarts the page table walk when changes occur in the page
+ * table, such as splitting a PMD. Ensures that the PTL held during
+ * the previous walk is released and resets the state to allow for
+ * a new walk starting at the current address stored in pvmw->address.
+ */
+static inline void
+page_vma_mapped_walk_restart(struct page_vma_mapped_walk *pvmw)
+{
+	WARN_ON_ONCE(!pvmw->pmd && !pvmw->pte);
+
+	if (likely(pvmw->ptl))
+		spin_unlock(pvmw->ptl);
+	else
+		WARN_ON_ONCE(1);
+
+	pvmw->ptl = NULL;
+	pvmw->pmd = NULL;
+	pvmw->pte = NULL;
+}
+
 bool page_vma_mapped_walk(struct page_vma_mapped_walk *pvmw);
 
 /*
mm/huge_memory.c +22 -20

54 lines not on this page · raw message

mm/rmap.c +15 -6
index 2d778725e4f5..dacf24bc82f0 100644
@@ -1642,9 +1642,6 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 	if (flags & TTU_SYNC)
 		pvmw.flags = PVMW_SYNC;
 
-	if (flags & TTU_SPLIT_HUGE_PMD)
-		split_huge_pmd_address(vma, address, false, folio);
-
 	/*
 	 * For THP, we have to assume the worse case ie pmd for invalidation.
 	 * For hugetlb, it could be much worse if we need to do pud
@@ -1670,9 +1667,6 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 	mmu_notifier_invalidate_range_start(&range);
 
 	while (page_vma_mapped_walk(&pvmw)) {
-		/* Unexpected PMD-mapped THP? */
-		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
-
 		/*
 		 * If the folio is in an mlock()d vma, we must not swap it out.
 		 */
@@ -1684,6 +1678,21 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 			goto walk_abort;
 		}
 
+		if (!pvmw.pte && (flags & TTU_SPLIT_HUGE_PMD)) {
+			/*
+			 * We temporarily have to drop the PTL and start once
+			 * again from that now-PTE-mapped page table.
+			 */
+			split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
+					      false, folio);
+			flags &= ~TTU_SPLIT_HUGE_PMD;
+			page_vma_mapped_walk_restart(&pvmw);
+			continue;
+		}
+
+		/* Unexpected PMD-mapped THP? */
+		VM_BUG_ON_FOLIO(!pvmw.pte, folio);
+
 		pfn = pte_pfn(ptep_get(pvmw.pte));
 		subpage = folio_page(folio, pfn - folio_pfn(folio));
 		address = pvmw.address;
Signature · 1 line
2.33.1
Acked-by on v8 2/3
#3 lore raw

Re: [PATCH v8 2/3] mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop

Quoted 8 lines · Lance Yang

In preparation for supporting try_to_unmap_one() to unmap PMD-mapped folios, start the pagewalk first, then call split_huge_pmd_address() to split the folio.

---

Would have converted that VM_BUG_ON to a VM_WARN_ON_ONCE, but it's just moving code, so no big deal.

Thanks!

Signature · 3 lines
Cheers,

David / dhildenb
author
#4 lore raw

Re: [PATCH v8 2/3] mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop

Quoted 6 lines · David Hildenbrand
Quoted 4 lines · Lance Yang

In preparation for supporting try_to_unmap_one() to unmap PMD-mapped folios, start the pagewalk first, then call split_huge_pmd_address() to split the folio.

Quoted 7 lines
Quoted 4 lines

---

Would have converted that VM_BUG_ON to a VM_WARN_ON_ONCE, but it's just moving code, so no big deal.

OK, let’s leave it as is for now ;)

Quoted 4 lines

Thanks!

Thanks for taking time to review!
Lance

Quoted 6 lines

--
Cheers,

David / dhildenb

reviewer Acked-by on v8 2/3
#5 lore raw

Re: [PATCH v8 2/3] mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop

In reply to Lance Yang
Quoted 7 lines · Lance Yang

In preparation for supporting try_to_unmap_one() to unmap PMD-mapped folios, start the pagewalk first, then call split_huge_pmd_address() to split the folio.

Quoted 6 lines
---
 include/linux/huge_mm.h |  6 ++++++
 include/linux/rmap.h    | 24 +++++++++++++++++++++++
 mm/huge_memory.c        | 42 +++++++++++++++++++++--------------------
 mm/rmap.c               | 21 +++++++++++++++------
 4 files changed, 67 insertions(+), 26 deletions(-)

Thanks.

--
Best Regards,
Yan, Zi

author
#6 lore raw

Re: [PATCH v8 2/3] mm/rmap: integrate PMD-mapped folio splitting into pagewalk loop

Quoted 17 lines · Zi Yan
Quoted 7 lines · Lance Yang

In preparation for supporting try_to_unmap_one() to unmap PMD-mapped folios, start the pagewalk first, then call split_huge_pmd_address() to split the folio.

Quoted 6 lines
---
 include/linux/huge_mm.h |  6 ++++++
 include/linux/rmap.h    | 24 +++++++++++++++++++++++
 mm/huge_memory.c        | 42 +++++++++++++++++++++--------------------
 mm/rmap.c               | 21 +++++++++++++++------
 4 files changed, 67 insertions(+), 26 deletions(-)
Quoted 3 lines

Thanks.

Thanks for taking time to review!
Lance

Quoted 4 lines

--
Best Regards,
Yan, Zi

author
#7 lore raw

PATCH v8 3/3 mm/vmscan: avoid split lazyfree THP during shrink_folio_list()

In reply to Lance Yang

When the user no longer requires the pages, they would use madvise(MADV_FREE) to mark the pages as lazy free. Subsequently, they typically would not re-write to that memory again.

During memory reclaim, if we detect that the large folio and its PMD are both still marked as clean and there are no unexpected references (such as GUP), so we can just discard the memory lazily, improving the efficiency of memory reclamation in this case.

On an Intel i5 CPU, reclaiming 1GiB of lazyfree THPs using mem_cgroup_force_empty() results in the following runtimes in seconds (shorter is better):

--------------------------------------------
| Old | New | Change |
--------------------------------------------
| 0.683426 | 0.049197 | -92.80% |
--------------------------------------------

---
 include/linux/huge_mm.h |  9 +++++
 mm/huge_memory.c        | 76 +++++++++++++++++++++++++++++++++++++++++
 mm/rmap.c               | 27 +++++++++------
 3 files changed, 102 insertions(+), 10 deletions(-)
include/linux/huge_mm.h +9 -0
index 9f720b0731c4..212cca384d7e 100644
@@ -430,6 +430,8 @@ static inline bool thp_migration_supported(void)
 
 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
 			   pmd_t *pmd, bool freeze, struct folio *folio);
+bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
+			   pmd_t *pmdp, struct folio *folio);
 
 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 
@@ -497,6 +499,13 @@ static inline void split_huge_pmd_locked(struct vm_area_struct *vma,
 					 unsigned long address, pmd_t *pmd,
 					 bool freeze, struct folio *folio) {}
 
+static inline bool unmap_huge_pmd_locked(struct vm_area_struct *vma,
+					 unsigned long addr, pmd_t *pmdp,
+					 struct folio *folio)
+{
+	return false;
+}
+
 #define split_huge_pud(__vma, __pmd, __address)	\
 	do { } while (0)
 
mm/huge_memory.c +76 -0

82 lines not on this page · raw message

mm/rmap.c +17 -10
index dacf24bc82f0..7d97806f74cd 100644
@@ -1678,16 +1678,23 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 			goto walk_abort;
 		}
 
-		if (!pvmw.pte && (flags & TTU_SPLIT_HUGE_PMD)) {
-			/*
-			 * We temporarily have to drop the PTL and start once
-			 * again from that now-PTE-mapped page table.
-			 */
-			split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
-					      false, folio);
-			flags &= ~TTU_SPLIT_HUGE_PMD;
-			page_vma_mapped_walk_restart(&pvmw);
-			continue;
+		if (!pvmw.pte) {
+			if (unmap_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
+						  folio))
+				goto walk_done;
+
+			if (flags & TTU_SPLIT_HUGE_PMD) {
+				/*
+				 * We temporarily have to drop the PTL and start
+				 * once again from that now-PTE-mapped page
+				 * table.
+				 */
+				split_huge_pmd_locked(vma, pvmw.address,
+						      pvmw.pmd, false, folio);
+				flags &= ~TTU_SPLIT_HUGE_PMD;
+				page_vma_mapped_walk_restart(&pvmw);
+				continue;
+			}
 		}
 
 		/* Unexpected PMD-mapped THP? */
Signature · 1 line
2.33.1
#8 lore raw

Re: [PATCH v8 3/3] mm/vmscan: avoid split lazyfree THP during shrink_folio_list()

Sorry for taking so long to review ... getting there. Mostly nits.

Quoted 19 lines
include/linux/huge_mm.h +7 -0
@@ -497,6 +499,13 @@ static inline void split_huge_pmd_locked(struct vm_area_struct *vma,
  					 unsigned long address, pmd_t *pmd,
  					 bool freeze, struct folio *folio) {}
 
+static inline bool unmap_huge_pmd_locked(struct vm_area_struct *vma,
+					 unsigned long addr, pmd_t *pmdp,
+					 struct folio *folio)
+{
+	return false;
+}
+
  #define split_huge_pud(__vma, __pmd, __address)	\
  	do { } while (0)
 
mm/huge_memory.c +0 -0
index e766d3f3a302..425374ae06ed 100644

mm/huge_memory.c static void unmap_folio(struct folio *folio) · line 2696

Quoted 9 lines
mm/huge_memory.c +6 -0
@@ -2688,6 +2688,82 @@ static void unmap_folio(struct folio *folio)
  	try_to_unmap_flush();
  }
 
+static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,
+					    unsigned long addr, pmd_t *pmdp,
+					    struct folio *folio)
+{
+	VM_WARN_ON_FOLIO(folio_test_swapbacked(folio), folio);
+	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);

I would drop these (that's exactly what the single caller checks). In any case don't place them above the variable declaration ;)

mm/huge_memory.c static void unmap_folio(struct folio *folio) · line 2708

Quoted 12 lines
mm/huge_memory.c +12 -0
+
+	struct mm_struct *mm = vma->vm_mm;
+	int ref_count, map_count;
+	pmd_t orig_pmd = *pmdp;
+	struct page *page;
+
+	if (unlikely(!pmd_present(orig_pmd) || !pmd_trans_huge(orig_pmd)))
+		return false;
+
+	page = pmd_page(orig_pmd);
+	if (unlikely(page_folio(page) != folio))
+		return false;

I'm curious, how could that happen? And how could it happen that we have !pmd_trans_huge() ? Didn't rmap walking code make sure that this PMD maps the folio already, and we are holding the PTL?

Quoted 79 lines
mm/huge_memory.c +58 -0
+
+	if (folio_test_dirty(folio) || pmd_dirty(orig_pmd))
+		return false;
+
+	orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp);
+
+	/*
+	 * Syncing against concurrent GUP-fast:
+	 * - clear PMD; barrier; read refcount
+	 * - inc refcount; barrier; read PMD
+	 */
+	smp_mb();
+
+	ref_count = folio_ref_count(folio);
+	map_count = folio_mapcount(folio);
+
+	/*
+	 * Order reads for folio refcount and dirty flag
+	 * (see comments in __remove_mapping()).
+	 */
+	smp_rmb();
+
+	/*
+	 * If the folio or its PMD is redirtied at this point, or if there
+	 * are unexpected references, we will give up to discard this folio
+	 * and remap it.
+	 *
+	 * The only folio refs must be one from isolation plus the rmap(s).
+	 */
+	if (folio_test_dirty(folio) || pmd_dirty(orig_pmd) ||
+	    ref_count != map_count + 1) {
+		set_pmd_at(mm, addr, pmdp, orig_pmd);
+		return false;
+	}
+
+	folio_remove_rmap_pmd(folio, page, vma);
+	zap_deposited_table(mm, pmdp);
+	add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR);
+	if (vma->vm_flags & VM_LOCKED)
+		mlock_drain_local();
+	folio_put(folio);
+
+	return true;
+}
+
+bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
+			   pmd_t *pmdp, struct folio *folio)
+{
+	VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio);
+	VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
+	VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE));
+
+	if (folio_test_anon(folio) && !folio_test_swapbacked(folio))
+		return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio);
+
+	return false;
+}
+
  static void remap_page(struct folio *folio, unsigned long nr)
  {
  	int i = 0;
mm/rmap.c +0 -10
index dacf24bc82f0..7d97806f74cd 100644
@@ -1678,16 +1678,23 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
  			goto walk_abort;
  		}
 
-		if (!pvmw.pte && (flags & TTU_SPLIT_HUGE_PMD)) {
-			/*
-			 * We temporarily have to drop the PTL and start once
-			 * again from that now-PTE-mapped page table.
-			 */
-			split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
-					      false, folio);
-			flags &= ~TTU_SPLIT_HUGE_PMD;
-			page_vma_mapped_walk_restart(&pvmw);
-			continue;

mm/rmap.c static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, · line 1690

Quoted 10 lines
mm/rmap.c +10 -0
+		if (!pvmw.pte) {
+			if (unmap_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
+						  folio))
+				goto walk_done;
+
+			if (flags & TTU_SPLIT_HUGE_PMD) {
+				/*
+				 * We temporarily have to drop the PTL and start
+				 * once again from that now-PTE-mapped page
+				 * table.

Nit: it's not a PTE-mapped page table.

Maybe

"... restart so we can process the PTE-mapped THP."

mm/rmap.c static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, · line 1700

Quoted 3 lines
mm/rmap.c +0 -0
  		}
 
  		/* Unexpected PMD-mapped THP? */

Nothing else jumped at me :)

Signature · 3 lines
Cheers,

David / dhildenb
author
#9 lore raw

Re: [PATCH v8 3/3] mm/vmscan: avoid split lazyfree THP during shrink_folio_list()

Quoted 2 lines · David Hildenbrand

Sorry for taking so long to review ... getting there. Mostly nits.

No worries at all :)

Thanks for taking time to review!

Quoted 23 lines
Quoted 23 lines
include/linux/huge_mm.h +7 -0
@@ -497,6 +499,13 @@ static inline void split_huge_pmd_locked(struct vm_area_struct *vma,
                                       unsigned long address, pmd_t *pmd,
                                       bool freeze, struct folio *folio) {}
 
+static inline bool unmap_huge_pmd_locked(struct vm_area_struct *vma,
+                                      unsigned long addr, pmd_t *pmdp,
+                                      struct folio *folio)
+{
+     return false;
+}
+
  #define split_huge_pud(__vma, __pmd, __address)     \
      do { } while (0)
 
mm/huge_memory.c +0 -0
index e766d3f3a302..425374ae06ed 100644
@@ -2688,6 +2688,82 @@ static void unmap_folio(struct folio *folio)
      try_to_unmap_flush();
  }
 
Quoted 8 lines

mm/huge_memory.c static void unmap_folio(struct folio *folio) · line 2696

Quoted 6 lines
mm/huge_memory.c +6 -0
+static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,
+                                         unsigned long addr, pmd_t *pmdp,
+                                         struct folio *folio)
+{
+     VM_WARN_ON_FOLIO(folio_test_swapbacked(folio), folio);
+     VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);

I would drop these (that's exactly what the single caller checks). In

Agreed. I will drop these.

Quoted 1 line

any case don't place them above the variable declaration ;)

Yep, I see.

Quoted 17 lines

mm/huge_memory.c static void unmap_folio(struct folio *folio) · line 2708

Quoted 13 lines
mm/huge_memory.c +12 -0
+
+     struct mm_struct *mm = vma->vm_mm;
+     int ref_count, map_count;
+     pmd_t orig_pmd = *pmdp;
+     struct page *page;
+
+     if (unlikely(!pmd_present(orig_pmd) || !pmd_trans_huge(orig_pmd)))
+             return false;
+
+     page = pmd_page(orig_pmd);
+     if (unlikely(page_folio(page) != folio))
+             return false;

I'm curious, how could that happen? And how could it happen that we have !pmd_trans_huge() ? Didn't rmap walking code make sure that this PMD maps the folio already, and we are holding the PTL?

Makes sense to me. I was adding these just in case, but it's probably too much.

Let's drop them ;)

Quoted 93 lines
Quoted 80 lines
mm/huge_memory.c +58 -0
+
+     if (folio_test_dirty(folio) || pmd_dirty(orig_pmd))
+             return false;
+
+     orig_pmd = pmdp_huge_clear_flush(vma, addr, pmdp);
+
+     /*
+      * Syncing against concurrent GUP-fast:
+      * - clear PMD; barrier; read refcount
+      * - inc refcount; barrier; read PMD
+      */
+     smp_mb();
+
+     ref_count = folio_ref_count(folio);
+     map_count = folio_mapcount(folio);
+
+     /*
+      * Order reads for folio refcount and dirty flag
+      * (see comments in __remove_mapping()).
+      */
+     smp_rmb();
+
+     /*
+      * If the folio or its PMD is redirtied at this point, or if there
+      * are unexpected references, we will give up to discard this folio
+      * and remap it.
+      *
+      * The only folio refs must be one from isolation plus the rmap(s).
+      */
+     if (folio_test_dirty(folio) || pmd_dirty(orig_pmd) ||
+         ref_count != map_count + 1) {
+             set_pmd_at(mm, addr, pmdp, orig_pmd);
+             return false;
+     }
+
+     folio_remove_rmap_pmd(folio, page, vma);
+     zap_deposited_table(mm, pmdp);
+     add_mm_counter(mm, MM_ANONPAGES, -HPAGE_PMD_NR);
+     if (vma->vm_flags & VM_LOCKED)
+             mlock_drain_local();
+     folio_put(folio);
+
+     return true;
+}
+
+bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
+                        pmd_t *pmdp, struct folio *folio)
+{
+     VM_WARN_ON_FOLIO(!folio_test_pmd_mappable(folio), folio);
+     VM_WARN_ON_FOLIO(!folio_test_locked(folio), folio);
+     VM_WARN_ON_ONCE(!IS_ALIGNED(addr, HPAGE_PMD_SIZE));
+
+     if (folio_test_anon(folio) && !folio_test_swapbacked(folio))
+             return __discard_anon_folio_pmd_locked(vma, addr, pmdp, folio);
+
+     return false;
+}
+
  static void remap_page(struct folio *folio, unsigned long nr)
  {
      int i = 0;
mm/rmap.c +0 -10
index dacf24bc82f0..7d97806f74cd 100644
@@ -1678,16 +1678,23 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
                      goto walk_abort;
              }
 
-             if (!pvmw.pte && (flags & TTU_SPLIT_HUGE_PMD)) {
-                     /*
-                      * We temporarily have to drop the PTL and start once
-                      * again from that now-PTE-mapped page table.
-                      */
-                     split_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
-                                           false, folio);
-                     flags &= ~TTU_SPLIT_HUGE_PMD;
-                     page_vma_mapped_walk_restart(&pvmw);
-                     continue;

mm/rmap.c static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, · line 1690

Quoted 10 lines
mm/rmap.c +10 -0
+             if (!pvmw.pte) {
+                     if (unmap_huge_pmd_locked(vma, pvmw.address, pvmw.pmd,
+                                               folio))
+                             goto walk_done;
+
+                     if (flags & TTU_SPLIT_HUGE_PMD) {
+                             /*
+                              * We temporarily have to drop the PTL and start
+                              * once again from that now-PTE-mapped page
+                              * table.

Nit: it's not a PTE-mapped page table.

Quoted 3 lines

Maybe

"... restart so we can process the PTE-mapped THP."

Nice. Will adjust as you suggested.

Quoted 8 lines

mm/rmap.c static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma, · line 1700

Quoted 6 lines
mm/rmap.c +0 -0
              }
 
              /* Unexpected PMD-mapped THP? */

Nothing else jumped at me :)

Thanks again for your time!
Lance

Quoted 6 lines

--
Cheers,

David / dhildenb

author
#10 lore raw

Re: [PATCH v8 3/3] mm/vmscan: avoid split lazyfree THP during shrink_folio_list()

In reply to David Hildenbrand

Hi Andrew,

I made some minor changes suggested by David[1]. Could you please fold the following changes into this patch?

[1] https://lore.kernel.org/linux-mm/[email protected]/

mm/huge_memory.c +0 -10
index 4b2817bb2c7d..0cb52ae29259 100644
@@ -2693,21 +2693,11 @@ static bool __discard_anon_folio_pmd_locked(struct vm_area_struct *vma,
 					    unsigned long addr, pmd_t *pmdp,
 					    struct folio *folio)
 {
-	VM_WARN_ON_FOLIO(folio_test_swapbacked(folio), folio);
-	VM_WARN_ON_FOLIO(!folio_test_anon(folio), folio);
-
 	struct mm_struct *mm = vma->vm_mm;
 	int ref_count, map_count;
 	pmd_t orig_pmd = *pmdp;
 	struct page *page;
 
-	if (unlikely(!pmd_present(orig_pmd) || !pmd_trans_huge(orig_pmd)))
-		return false;
-
-	page = pmd_page(orig_pmd);
-	if (unlikely(page_folio(page) != folio))
-		return false;
-
 	if (folio_test_dirty(folio) || pmd_dirty(orig_pmd))
 		return false;
 
mm/rmap.c +2 -3
index df1a43295c85..b358501fb7e8 100644
@@ -1678,9 +1678,8 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
 
 			if (flags & TTU_SPLIT_HUGE_PMD) {
 				/*
-				 * We temporarily have to drop the PTL and start
-				 * once again from that now-PTE-mapped page
-				 * table.
+				 * We temporarily have to drop the PTL and
+				 * restart so we can process the PTE-mapped THP.
 				 */
 				split_huge_pmd_locked(vma, pvmw.address,
 						      pvmw.pmd, false, folio);
Signature · 2 lines
Thanks,
Lance