Skip to contentSkip to conversation

mm: Introduce and use folio_owner_ops

RFC PATCH In review Fuad Tabba 6 participants 23 messages Nov 8, 2024 – Nov 14, 2024

Subsystems (16)
lore
Share & tools
lore

Fuad Tabba

author
#0 lore raw

cover mm: Introduce and use folio_owner_ops

Some folios, such as hugetlb folios and zone device folios, require special handling when the folio's reference count reaches 0, before being freed. Moreover, guest_memfd folios will likely require special handling to notify it once a folio's reference count reaches 0, to facilitate shared to private folio conversion [*]. Currently, each usecase has a dedicated callback when the folio refcount reaches 0 to that effect. Adding yet more callbacks is not ideal.

This patch series introduces struct folio_owner_ops and uses it as a generic way to handle callbacks on freeing a folio. It also applies the callbacks to hugetlb and zone device folios.

A pointer to struct folio_owner_ops is overlaid on struct page compound_page, struct page/folio lru. To indicate that the folio uses the callback, this patch series sets bit 1 of the new field, similar to how bit 0 indicates a compound page.

Patches 1 to 6 rework the hugetlb code to allow us to reuse folio->lru for the owner ops as long as they are not isolated.

Patches 7 to 10 introduce struct folio_owner_ops, and apply the callbacks to zone device and hugetlb folios.

Cheers,
/fuad

[*] https://lore.kernel.org/all/CAGtprH_JP2w-4rq02h_Ugvq5KuHX7TUvegOS7xUs_iy5hriE7g@mail.gmail.com/

David Hildenbrand (6):
mm/hugetlb: rename isolate_hugetlb() to folio_isolate_hugetlb() mm/migrate: don't call folio_putback_active_hugetlb() on dst hugetlb
folio
mm/hugetlb: rename "folio_putback_active_hugetlb()" to
"folio_putback_hugetlb()"
mm/hugetlb-cgroup: convert hugetlb_cgroup_css_offline() to work on
folios
mm/hugetlb: use folio->lru int demote_free_hugetlb_folios() mm/hugetlb: use separate folio->_hugetlb_list for hugetlb-internals

Fuad Tabba (4):
mm: Introduce struct folio_owner_ops
mm: Use getters and setters to access page pgmap
mm: Use owner_ops on folio_put for zone device pages
mm: hugetlb: Use owner_ops on folio_put for hugetlb

 drivers/gpu/drm/nouveau/nouveau_dmem.c |   4 +-
 drivers/pci/p2pdma.c                   |   8 +-
 include/linux/hugetlb.h                |  10 +-
 include/linux/memremap.h               |  14 +-
 include/linux/mm_types.h               | 107 ++++++++++++++-
 lib/test_hmm.c                         |   2 +-
 mm/gup.c                               |   2 +-
 mm/hmm.c                               |   2 +-
 mm/hugetlb.c                           | 179 ++++++++++++++++++-------
 mm/hugetlb_cgroup.c                    |  19 ++-
 mm/hugetlb_vmemmap.c                   |   8 +-
 mm/internal.h                          |   1 -
 mm/memory.c                            |   2 +-
 mm/mempolicy.c                         |   2 +-
 mm/memremap.c                          |  49 +------
 mm/migrate.c                           |  20 +--
 mm/migrate_device.c                    |   4 +-
 mm/mm_init.c                           |  48 ++++++-
 mm/swap.c                              |  25 ++--
 19 files changed, 342 insertions(+), 164 deletions(-)

base-commit: beb2622b970047000fa3cae64c23585669b01fca

Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#1 lore raw

RFC PATCH 1/10 mm/hugetlb: rename isolate_hugetlb() to folio_isolate_hugetlb()

From: David Hildenbrand <[email protected]>

Let's make the function name match "folio_isolate_lru()", and add some kernel doc.

---
 include/linux/hugetlb.h |  4 ++--
 mm/gup.c                |  2 +-
 mm/hugetlb.c            | 23 ++++++++++++++++++++---
 mm/mempolicy.c          |  2 +-
 mm/migrate.c            |  6 +++---
 5 files changed, 27 insertions(+), 10 deletions(-)
include/linux/hugetlb.h +2 -2
index ae4fe8615bb6..b0cf8dbfeb6a 100644
@@ -153,7 +153,7 @@ bool hugetlb_reserve_pages(struct inode *inode, long from, long to,
 						vm_flags_t vm_flags);
 long hugetlb_unreserve_pages(struct inode *inode, long start, long end,
 						long freed);
-bool isolate_hugetlb(struct folio *folio, struct list_head *list);
+bool folio_isolate_hugetlb(struct folio *folio, struct list_head *list);
 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison);
 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
 				bool *migratable_cleared);
@@ -414,7 +414,7 @@ static inline pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr,
 	return NULL;
 }
 
-static inline bool isolate_hugetlb(struct folio *folio, struct list_head *list)
+static inline bool folio_isolate_hugetlb(struct folio *folio, struct list_head *list)
 {
 	return false;
 }
mm/gup.c +1 -1
index 28ae330ec4dd..40bbcffca865 100644
@@ -2301,7 +2301,7 @@ static unsigned long collect_longterm_unpinnable_folios(
 			continue;
 
 		if (folio_test_hugetlb(folio)) {
-			isolate_hugetlb(folio, movable_folio_list);
+			folio_isolate_hugetlb(folio, movable_folio_list);
 			continue;
 		}
 
mm/hugetlb.c +20 -3
index cec4b121193f..e17bb2847572 100644
@@ -2868,7 +2868,7 @@ static int alloc_and_dissolve_hugetlb_folio(struct hstate *h,
 		 * Fail with -EBUSY if not possible.
 		 */
 		spin_unlock_irq(&hugetlb_lock);
-		isolated = isolate_hugetlb(old_folio, list);
+		isolated = folio_isolate_hugetlb(old_folio, list);
 		ret = isolated ? 0 : -EBUSY;
 		spin_lock_irq(&hugetlb_lock);
 		goto free_new;
@@ -2953,7 +2953,7 @@ int isolate_or_dissolve_huge_page(struct page *page, struct list_head *list)
 	if (hstate_is_gigantic(h))
 		return -ENOMEM;
 
-	if (folio_ref_count(folio) && isolate_hugetlb(folio, list))
+	if (folio_ref_count(folio) && folio_isolate_hugetlb(folio, list))
 		ret = 0;
 	else if (!folio_ref_count(folio))
 		ret = alloc_and_dissolve_hugetlb_folio(h, folio, list);
@@ -7396,7 +7396,24 @@ __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
 
 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
 
-bool isolate_hugetlb(struct folio *folio, struct list_head *list)
+/**
+ * folio_isolate_hugetlb: try to isolate an allocated hugetlb folio
+ * @folio: the folio to isolate
+ * @list: the list to add the folio to on success
+ *
+ * Isolate an allocated (refcount > 0) hugetlb folio, marking it as
+ * isolated/non-migratable, and moving it from the active list to the
+ * given list.
+ *
+ * Isolation will fail if @folio is not an allocated hugetlb folio, or if
+ * it is already isolated/non-migratable.
+ *
+ * On success, an additional folio reference is taken that must be dropped
+ * using folio_putback_active_hugetlb() to undo the isolation.
+ *
+ * Return: True if isolation worked, otherwise False.
+ */
+bool folio_isolate_hugetlb(struct folio *folio, struct list_head *list)
 {
 	bool ret = true;
 
mm/mempolicy.c +1 -1
index bb37cd1a51d8..41bdff67757c 100644
@@ -647,7 +647,7 @@ static int queue_folios_hugetlb(pte_t *pte, unsigned long hmask,
 	 */
 	if ((flags & MPOL_MF_MOVE_ALL) ||
 	    (!folio_likely_mapped_shared(folio) && !hugetlb_pmd_shared(pte)))
-		if (!isolate_hugetlb(folio, qp->pagelist))
+		if (!folio_isolate_hugetlb(folio, qp->pagelist))
 			qp->nr_failed++;
 unlock:
 	spin_unlock(ptl);
mm/migrate.c +3 -3
index dfb5eba3c522..55585b5f57ec 100644
@@ -136,7 +136,7 @@ static void putback_movable_folio(struct folio *folio)
  *
  * This function shall be used whenever the isolated pageset has been
  * built from lru, balloon, hugetlbfs page. See isolate_migratepages_range()
- * and isolate_hugetlb().
+ * and folio_isolate_hugetlb().
  */
 void putback_movable_pages(struct list_head *l)
 {
@@ -177,7 +177,7 @@ bool isolate_folio_to_list(struct folio *folio, struct list_head *list)
 	bool isolated, lru;
 
 	if (folio_test_hugetlb(folio))
-		return isolate_hugetlb(folio, list);
+		return folio_isolate_hugetlb(folio, list);
 
 	lru = !__folio_test_movable(folio);
 	if (lru)
@@ -2208,7 +2208,7 @@ static int __add_folio_for_migration(struct folio *folio, int node,
 		return -EACCES;
 
 	if (folio_test_hugetlb(folio)) {
-		if (isolate_hugetlb(folio, pagelist))
+		if (folio_isolate_hugetlb(folio, pagelist))
 			return 1;
 	} else if (folio_isolate_lru(folio)) {
 		list_add_tail(&folio->lru, pagelist);
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#2 lore raw

RFC PATCH 2/10 mm/migrate: don't call folio_putback_active_hugetlb() on dst hugetlb folio

In reply to Fuad Tabba

From: David Hildenbrand <[email protected]>

We replaced a simple put_page() by a putback_active_hugepage() call in commit 3aaa76e125c1 (" mm: migrate: hugetlb: putback destination hugepage to active list"), to set the "active" flag on the dst hugetlb folio.

Nowadays, we decoupled the "active" list from the flag, by calling the flag "migratable".

Calling "putback" on something that wasn't allocated is weird and not future proof, especially if we might reach that path when migration failed and we just want to free the freshly allocated hugetlb folio.

Let's simply set the "migratable" flag in move_hugetlb_state(), where we know that allocation succeeded, and use simple folio_put() to return our reference.

Do we need the hugetlb_lock for setting that flag? Staring at other users of folio_set_hugetlb_migratable(), it does not look like it. After all, the dst folio should already be on the active list, and we are not modifying that list.

---
 mm/hugetlb.c | 5 +++++
 mm/migrate.c | 8 ++++----
 2 files changed, 9 insertions(+), 4 deletions(-)
mm/hugetlb.c +5 -0
index e17bb2847572..da3fe1840ab8 100644
@@ -7508,6 +7508,11 @@ void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int re
 		}
 		spin_unlock_irq(&hugetlb_lock);
 	}
+	/*
+	 * Our old folio is isolated and has "migratable" cleared until it
+	 * is putback. As migration succeeded, set the new folio "migratable".
+	 */
+	folio_set_hugetlb_migratable(new_folio);
 }
 
 static void hugetlb_unshare_pmds(struct vm_area_struct *vma,
mm/migrate.c +4 -4
index 55585b5f57ec..b129dc41c140 100644
@@ -1547,14 +1547,14 @@ static int unmap_and_move_huge_page(new_folio_t get_new_folio,
 		list_move_tail(&src->lru, ret);
 
 	/*
-	 * If migration was not successful and there's a freeing callback, use
-	 * it.  Otherwise, put_page() will drop the reference grabbed during
-	 * isolation.
+	 * If migration was not successful and there's a freeing callback,
+	 * return the folio to that special allocator. Otherwise, simply drop
+	 * our additional reference.
 	 */
 	if (put_new_folio)
 		put_new_folio(dst, private);
 	else
-		folio_putback_active_hugetlb(dst);
+		folio_put(dst);
 
 	return rc;
 }
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#3 lore raw

RFC PATCH 3/10 mm/hugetlb: rename "folio_putback_active_hugetlb()" to "folio_putback_hugetlb()"

In reply to Fuad Tabba

From: David Hildenbrand <[email protected]>

Now that folio_putback_hugetlb() is only called on folios that were previously isolated through folio_isolate_hugetlb(), let's rename it to match folio_putback_lru().

Add some kernel doc to clarify how this function is supposed to be used.

---
 include/linux/hugetlb.h |  4 ++--
 mm/hugetlb.c            | 15 +++++++++++++--
 mm/migrate.c            |  6 +++---
 3 files changed, 18 insertions(+), 7 deletions(-)
include/linux/hugetlb.h +2 -2
index b0cf8dbfeb6a..e846d7dac77c 100644
@@ -157,7 +157,7 @@ bool folio_isolate_hugetlb(struct folio *folio, struct list_head *list);
 int get_hwpoison_hugetlb_folio(struct folio *folio, bool *hugetlb, bool unpoison);
 int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
 				bool *migratable_cleared);
-void folio_putback_active_hugetlb(struct folio *folio);
+void folio_putback_hugetlb(struct folio *folio);
 void move_hugetlb_state(struct folio *old_folio, struct folio *new_folio, int reason);
 void hugetlb_fix_reserve_counts(struct inode *inode);
 extern struct mutex *hugetlb_fault_mutex_table;
@@ -430,7 +430,7 @@ static inline int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
 	return 0;
 }
 
-static inline void folio_putback_active_hugetlb(struct folio *folio)
+static inline void folio_putback_hugetlb(struct folio *folio)
 {
 }
 
mm/hugetlb.c +13 -2
index da3fe1840ab8..d58bd815fdf2 100644
@@ -7409,7 +7409,7 @@ __weak unsigned long hugetlb_mask_last_page(struct hstate *h)
  * it is already isolated/non-migratable.
  *
  * On success, an additional folio reference is taken that must be dropped
- * using folio_putback_active_hugetlb() to undo the isolation.
+ * using folio_putback_hugetlb() to undo the isolation.
  *
  * Return: True if isolation worked, otherwise False.
  */
@@ -7461,7 +7461,18 @@ int get_huge_page_for_hwpoison(unsigned long pfn, int flags,
 	return ret;
 }
 
-void folio_putback_active_hugetlb(struct folio *folio)
+/**
+ * folio_putback_hugetlb: unisolate a hugetlb folio
+ * @folio: the isolated hugetlb folio
+ *
+ * Putback/un-isolate the hugetlb folio that was previous isolated using
+ * folio_isolate_hugetlb(): marking it non-isolated/migratable and putting it
+ * back onto the active list.
+ *
+ * Will drop the additional folio reference obtained through
+ * folio_isolate_hugetlb().
+ */
+void folio_putback_hugetlb(struct folio *folio)
 {
 	spin_lock_irq(&hugetlb_lock);
 	folio_set_hugetlb_migratable(folio);
mm/migrate.c +3 -3
index b129dc41c140..89292d131148 100644
@@ -145,7 +145,7 @@ void putback_movable_pages(struct list_head *l)
 
 	list_for_each_entry_safe(folio, folio2, l, lru) {
 		if (unlikely(folio_test_hugetlb(folio))) {
-			folio_putback_active_hugetlb(folio);
+			folio_putback_hugetlb(folio);
 			continue;
 		}
 		list_del(&folio->lru);
@@ -1459,7 +1459,7 @@ static int unmap_and_move_huge_page(new_folio_t get_new_folio,
 
 	if (folio_ref_count(src) == 1) {
 		/* page was freed from under us. So we are done. */
-		folio_putback_active_hugetlb(src);
+		folio_putback_hugetlb(src);
 		return MIGRATEPAGE_SUCCESS;
 	}
 
@@ -1542,7 +1542,7 @@ static int unmap_and_move_huge_page(new_folio_t get_new_folio,
 	folio_unlock(src);
 out:
 	if (rc == MIGRATEPAGE_SUCCESS)
-		folio_putback_active_hugetlb(src);
+		folio_putback_hugetlb(src);
 	else if (rc != -EAGAIN)
 		list_move_tail(&src->lru, ret);
 
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#4 lore raw

RFC PATCH 4/10 mm/hugetlb-cgroup: convert hugetlb_cgroup_css_offline() to work on folios

In reply to Fuad Tabba

From: David Hildenbrand <[email protected]>

Let's convert hugetlb_cgroup_css_offline() and hugetlb_cgroup_move_parent() to work on folios. hugepage_activelist contains folios, not pages.

While at it, rename page_hcg simply to hcg, removing most of the "page" terminology.

---
 mm/hugetlb_cgroup.c | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)
mm/hugetlb_cgroup.c +8 -9
index d8d0e665caed..1bdeaf25f640 100644
@@ -195,24 +195,23 @@ static void hugetlb_cgroup_css_free(struct cgroup_subsys_state *css)
  * cannot fail.
  */
 static void hugetlb_cgroup_move_parent(int idx, struct hugetlb_cgroup *h_cg,
-				       struct page *page)
+				       struct folio *folio)
 {
 	unsigned int nr_pages;
 	struct page_counter *counter;
-	struct hugetlb_cgroup *page_hcg;
+	struct hugetlb_cgroup *hcg;
 	struct hugetlb_cgroup *parent = parent_hugetlb_cgroup(h_cg);
-	struct folio *folio = page_folio(page);
 
-	page_hcg = hugetlb_cgroup_from_folio(folio);
+	hcg = hugetlb_cgroup_from_folio(folio);
 	/*
 	 * We can have pages in active list without any cgroup
 	 * ie, hugepage with less than 3 pages. We can safely
 	 * ignore those pages.
 	 */
-	if (!page_hcg || page_hcg != h_cg)
+	if (!hcg || hcg != h_cg)
 		goto out;
 
-	nr_pages = compound_nr(page);
+	nr_pages = folio_nr_pages(folio);
 	if (!parent) {
 		parent = root_h_cgroup;
 		/* root has no limit */
@@ -235,13 +234,13 @@ static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
 {
 	struct hugetlb_cgroup *h_cg = hugetlb_cgroup_from_css(css);
 	struct hstate *h;
-	struct page *page;
+	struct folio *folio;
 
 	do {
 		for_each_hstate(h) {
 			spin_lock_irq(&hugetlb_lock);
-			list_for_each_entry(page, &h->hugepage_activelist, lru)
-				hugetlb_cgroup_move_parent(hstate_index(h), h_cg, page);
+			list_for_each_entry(folio, &h->hugepage_activelist, lru)
+				hugetlb_cgroup_move_parent(hstate_index(h), h_cg, folio);
 
 			spin_unlock_irq(&hugetlb_lock);
 		}
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#5 lore raw

RFC PATCH 5/10 mm/hugetlb: use folio->lru int demote_free_hugetlb_folios()

In reply to Fuad Tabba

From: David Hildenbrand <[email protected]>

Let's avoid messing with pages.

---
 mm/hugetlb.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)
mm/hugetlb.c +4 -2
index d58bd815fdf2..a64852280213 100644
@@ -3806,13 +3806,15 @@ static long demote_free_hugetlb_folios(struct hstate *src, struct hstate *dst,
 
 		for (i = 0; i < pages_per_huge_page(src); i += pages_per_huge_page(dst)) {
 			struct page *page = folio_page(folio, i);
+			struct folio *new_folio;
 
 			page->mapping = NULL;
 			clear_compound_head(page);
 			prep_compound_page(page, dst->order);
+			new_folio = page_folio(page);
 
-			init_new_hugetlb_folio(dst, page_folio(page));
-			list_add(&page->lru, &dst_list);
+			init_new_hugetlb_folio(dst, new_folio);
+			list_add(&new_folio->lru, &dst_list);
 		}
 	}
 
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#6 lore raw

RFC PATCH 6/10 mm/hugetlb: use separate folio->_hugetlb_list for hugetlb-internals

In reply to Fuad Tabba

From: David Hildenbrand <[email protected]>

Let's use a separate list head in the folio, as long as hugetlb folios are not isolated. This way, we can reuse folio->lru for different purpose (e.g., owner_ops) as long as they are not isolated.

Consequently, folio->lru will only be used while there is an additional folio reference that cannot be dropped until putback/un-isolated.

---
 include/linux/mm_types.h | 18 +++++++++
 mm/hugetlb.c             | 81 +++++++++++++++++++++-------------------
 mm/hugetlb_cgroup.c      |  4 +-
 mm/hugetlb_vmemmap.c     |  8 ++--
 4 files changed, 66 insertions(+), 45 deletions(-)
include/linux/mm_types.h +18 -0
index 80fef38d9d64..365c73be0bb4 100644
@@ -310,6 +310,7 @@ typedef struct {
  * @_hugetlb_cgroup: Do not use directly, use accessor in hugetlb_cgroup.h.
  * @_hugetlb_cgroup_rsvd: Do not use directly, use accessor in hugetlb_cgroup.h.
  * @_hugetlb_hwpoison: Do not use directly, call raw_hwp_list_head().
+ * @_hugetlb_list: To be used in hugetlb core code only.
  * @_deferred_list: Folios to be split under memory pressure.
  * @_unused_slab_obj_exts: Placeholder to match obj_exts in struct slab.
  *
@@ -397,6 +398,17 @@ struct folio {
 		};
 		struct page __page_2;
 	};
+	union {
+		struct {
+			unsigned long _flags_3;
+			unsigned long _head_3;
+	/* public: */
+			struct list_head _hugetlb_list;
+	/* private: the union with struct page is transitional */
+		};
+		struct page __page_3;
+	};
+
 };
 
 #define FOLIO_MATCH(pg, fl)						\
@@ -433,6 +445,12 @@ FOLIO_MATCH(compound_head, _head_2);
 FOLIO_MATCH(flags, _flags_2a);
 FOLIO_MATCH(compound_head, _head_2a);
 #undef FOLIO_MATCH
+#define FOLIO_MATCH(pg, fl)						\
+	static_assert(offsetof(struct folio, fl) ==			\
+			offsetof(struct page, pg) + 3 * sizeof(struct page))
+FOLIO_MATCH(flags, _flags_3);
+FOLIO_MATCH(compound_head, _head_3);
+#undef FOLIO_MATCH
 
 /**
  * struct ptdesc -    Memory descriptor for page tables.
mm/hugetlb.c +42 -39

292 lines not on this page · raw message

mm/hugetlb_cgroup.c +2 -2
index 1bdeaf25f640..ee720eeaf6b1 100644
@@ -239,7 +239,7 @@ static void hugetlb_cgroup_css_offline(struct cgroup_subsys_state *css)
 	do {
 		for_each_hstate(h) {
 			spin_lock_irq(&hugetlb_lock);
-			list_for_each_entry(folio, &h->hugepage_activelist, lru)
+			list_for_each_entry(folio, &h->hugepage_activelist, _hugetlb_list)
 				hugetlb_cgroup_move_parent(hstate_index(h), h_cg, folio);
 
 			spin_unlock_irq(&hugetlb_lock);
@@ -933,7 +933,7 @@ void hugetlb_cgroup_migrate(struct folio *old_folio, struct folio *new_folio)
 	/* move the h_cg details to new cgroup */
 	set_hugetlb_cgroup(new_folio, h_cg);
 	set_hugetlb_cgroup_rsvd(new_folio, h_cg_rsvd);
-	list_move(&new_folio->lru, &h->hugepage_activelist);
+	list_move(&new_folio->_hugetlb_list, &h->hugepage_activelist);
 	spin_unlock_irq(&hugetlb_lock);
 	return;
 }
mm/hugetlb_vmemmap.c +4 -4
index 57b7f591eee8..b2cb8d328aac 100644
@@ -519,7 +519,7 @@ long hugetlb_vmemmap_restore_folios(const struct hstate *h,
 	long ret = 0;
 	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH | VMEMMAP_SYNCHRONIZE_RCU;
 
-	list_for_each_entry_safe(folio, t_folio, folio_list, lru) {
+	list_for_each_entry_safe(folio, t_folio, folio_list, _hugetlb_list) {
 		if (folio_test_hugetlb_vmemmap_optimized(folio)) {
 			ret = __hugetlb_vmemmap_restore_folio(h, folio, flags);
 			/* only need to synchronize_rcu() once for each batch */
@@ -531,7 +531,7 @@ long hugetlb_vmemmap_restore_folios(const struct hstate *h,
 		}
 
 		/* Add non-optimized folios to output list */
-		list_move(&folio->lru, non_hvo_folios);
+		list_move(&folio->_hugetlb_list, non_hvo_folios);
 	}
 
 	if (restored)
@@ -651,7 +651,7 @@ void hugetlb_vmemmap_optimize_folios(struct hstate *h, struct list_head *folio_l
 	LIST_HEAD(vmemmap_pages);
 	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH | VMEMMAP_SYNCHRONIZE_RCU;
 
-	list_for_each_entry(folio, folio_list, lru) {
+	list_for_each_entry(folio, folio_list, _hugetlb_list) {
 		int ret = hugetlb_vmemmap_split_folio(h, folio);
 
 		/*
@@ -666,7 +666,7 @@ void hugetlb_vmemmap_optimize_folios(struct hstate *h, struct list_head *folio_l
 
 	flush_tlb_all();
 
-	list_for_each_entry(folio, folio_list, lru) {
+	list_for_each_entry(folio, folio_list, _hugetlb_list) {
 		int ret;
 
 		ret = __hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, flags);
Signature · 1 line
2.47.0.277.g8800431eea-goog
#7 lore raw

Re:[RFC PATCH v1 06/10] mm/hugetlb: use separate folio->_hugetlb_list for hugetlb-internals

---

Quoted 19 lines
include/linux/mm_types.h +4 -0
index 80fef38d9d64..365c73be0bb4 100644
@@ -310,6 +310,7 @@ typedef struct {
 * @_hugetlb_cgroup: Do not use directly, use accessor in hugetlb_cgroup.h.
 * @_hugetlb_cgroup_rsvd: Do not use directly, use accessor in hugetlb_cgroup.h.
 * @_hugetlb_hwpoison: Do not use directly, call raw_hwp_list_head().
+ * @_hugetlb_list: To be used in hugetlb core code only.
 * @_deferred_list: Folios to be split under memory pressure.
 * @_unused_slab_obj_exts: Placeholder to match obj_exts in struct slab.
 *
@@ -397,6 +398,17 @@ struct folio {
 		};
 		struct page __page_2;
 	};
+	union {
+		struct {
+			unsigned long _flags_3;

include/linux/mm_types.h struct folio { · line 413

Quoted 10 lines
include/linux/mm_types.h +8 -0
+			unsigned long _head_3;
+	/* public: */
+			struct list_head _hugetlb_list;
+	/* private: the union with struct page is transitional */
+		};
+		struct page __page_3;
+	};
+
 };
 

As far as I know, increasing the size of folio maybe decrease the revenue of HVO, do you measure it?

Quoted 14 lines
include/linux/mm_types.h +6 -0
 #define FOLIO_MATCH(pg, fl)						\
@@ -433,6 +445,12 @@ FOLIO_MATCH(compound_head, _head_2);
 FOLIO_MATCH(flags, _flags_2a);
 FOLIO_MATCH(compound_head, _head_2a);
 #undef FOLIO_MATCH
+#define FOLIO_MATCH(pg, fl)						\
+	static_assert(offsetof(struct folio, fl) ==			\
+			offsetof(struct page, pg) + 3 * sizeof(struct page))
+FOLIO_MATCH(flags, _flags_3);
+FOLIO_MATCH(compound_head, _head_3);
+#undef FOLIO_MATCH
 
 /**
 * struct ptdesc -    Memory descriptor for page tables.
Signature · 1 line
2.25.1
#8 lore raw

Re: [RFC PATCH v1 06/10] mm/hugetlb: use separate folio->_hugetlb_list for hugetlb-internals

Quoted 25 lines · wang wei

---

Quoted 23 lines
include/linux/mm_types.h +8 -0
index 80fef38d9d64..365c73be0bb4 100644
@@ -310,6 +310,7 @@ typedef struct {
  * @_hugetlb_cgroup: Do not use directly, use accessor in hugetlb_cgroup.h.
  * @_hugetlb_cgroup_rsvd: Do not use directly, use accessor in hugetlb_cgroup.h.
  * @_hugetlb_hwpoison: Do not use directly, call raw_hwp_list_head().
+ * @_hugetlb_list: To be used in hugetlb core code only.
  * @_deferred_list: Folios to be split under memory pressure.
  * @_unused_slab_obj_exts: Placeholder to match obj_exts in struct slab.
  *
@@ -397,6 +398,17 @@ struct folio {
 		};
 		struct page __page_2;
 	};
+	union {
+		struct {
+			unsigned long _flags_3;
+			unsigned long _head_3;
+	/* public: */
+			struct list_head _hugetlb_list;
+	/* private: the union with struct page is transitional */
Quoted 8 lines

include/linux/mm_types.h struct folio { · line 413

Quoted 6 lines
include/linux/mm_types.h +4 -0
+		};
+		struct page __page_3;
+	};
+
 };
 

As far as I know, increasing the size of folio maybe decrease the revenue of HVO, do you measure it?

We should always have a writable page2, even with HVO. So nothing should change at this point. Thanks!

Signature · 3 lines
Cheers,

David / dhildenb
author
#9 lore raw

RFC PATCH 7/10 mm: Introduce struct folio_owner_ops

In reply to Fuad Tabba

Introduce struct folio_owner_ops, a method table that contains callbacks to owners of folios that need special handling for certain operations. For now, it only contains a callback for folio free(), which is called immediately after the folio refcount drops to 0.

Add a pointer to this struct overlaid on struct page compound_head, pgmap, and struct page/folio lru. The users of this struct either will not use lru (e.g., zone device), or would be able to easily isolate when lru is being used (e.g., hugetlb) and handle it accordingly. While folios are isolated, they cannot get freed and the owner_ops are unstable. This is sufficient for the current use case of returning these folios to a custom allocator.

To identify that a folio has owner_ops, we set bit 1 of the field, in a similar way to that bit 0 of compound_head is used to identify compound pages.

---
 include/linux/mm_types.h | 64 +++++++++++++++++++++++++++++++++++++---
 mm/swap.c                | 19 ++++++++++++
 2 files changed, 79 insertions(+), 4 deletions(-)
include/linux/mm_types.h +60 -4
index 365c73be0bb4..6e06286f44f1 100644
@@ -41,10 +41,12 @@ struct mem_cgroup;
  *
  * If you allocate the page using alloc_pages(), you can use some of the
  * space in struct page for your own purposes.  The five words in the main
- * union are available, except for bit 0 of the first word which must be
- * kept clear.  Many users use this word to store a pointer to an object
- * which is guaranteed to be aligned.  If you use the same storage as
- * page->mapping, you must restore it to NULL before freeing the page.
+ * union are available, except for bit 0 (used for compound_head pages)
+ * and bit 1 (used for owner_ops) of the first word, which must be kept
+ * clear and used with care.  Many users use this word to store a pointer
+ * to an object which is guaranteed to be aligned.  If you use the same
+ * storage as page->mapping, you must restore it to NULL before freeing
+ * the page.
  *
  * The mapcount field must not be used for own purposes.
  *
@@ -283,10 +285,16 @@ typedef struct {
 	unsigned long val;
 } swp_entry_t;
 
+struct folio_owner_ops;
+
 /**
  * struct folio - Represents a contiguous set of bytes.
  * @flags: Identical to the page flags.
  * @lru: Least Recently Used list; tracks how recently this folio was used.
+ * @owner_ops: Pointer to callback operations of the folio owner. Valid if bit 1
+ *    is set.
+ *    NOTE: Cannot be used with lru, since it is overlaid with it. To use lru,
+ *          owner_ops must be cleared first, and restored once done with lru.
  * @mlock_count: Number of times this folio has been pinned by mlock().
  * @mapping: The file this page belongs to, or refers to the anon_vma for
  *    anonymous memory.
@@ -330,6 +338,7 @@ struct folio {
 			unsigned long flags;
 			union {
 				struct list_head lru;
+				const struct folio_owner_ops *owner_ops; /* Bit 1 is set */
 	/* private: avoid cluttering the output */
 				struct {
 					void *__filler;
@@ -417,6 +426,7 @@ FOLIO_MATCH(flags, flags);
 FOLIO_MATCH(lru, lru);
 FOLIO_MATCH(mapping, mapping);
 FOLIO_MATCH(compound_head, lru);
+FOLIO_MATCH(compound_head, owner_ops);
 FOLIO_MATCH(index, index);
 FOLIO_MATCH(private, private);
 FOLIO_MATCH(_mapcount, _mapcount);
@@ -452,6 +462,13 @@ FOLIO_MATCH(flags, _flags_3);
 FOLIO_MATCH(compound_head, _head_3);
 #undef FOLIO_MATCH
 
+struct folio_owner_ops {
+	/*
+	 * Called once the folio refcount reaches 0.
+	 */
+	void (*free)(struct folio *folio);
+};
+
 /**
  * struct ptdesc -    Memory descriptor for page tables.
  * @__page_flags:     Same as page flags. Powerpc only.
@@ -560,6 +577,45 @@ static inline void *folio_get_private(struct folio *folio)
 	return folio->private;
 }
 
+/*
+ * Use bit 1, since bit 0 is used to indicate a compound page in compound_head,
+ * which owner_ops is overlaid with.
+ */
+#define FOLIO_OWNER_OPS_BIT    1UL
+#define FOLIO_OWNER_OPS        (1UL << FOLIO_OWNER_OPS_BIT)
+
+/*
+ * Set the folio owner_ops as well as bit 1 of the pointer to indicate that the
+ * folio has owner_ops.
+ */
+static inline void folio_set_owner_ops(struct folio *folio, const struct folio_owner_ops *owner_ops)
+{
+	owner_ops = (const struct folio_owner_ops *)((unsigned long)owner_ops | FOLIO_OWNER_OPS);
+	folio->owner_ops = owner_ops;
+}
+
+/*
+ * Clear the folio owner_ops including bit 1 of the pointer.
+ */
+static inline void folio_clear_owner_ops(struct folio *folio)
+{
+	folio->owner_ops = NULL;
+}
+
+/*
+ * Return the folio's owner_ops if it has them, otherwise, return NULL.
+ */
+static inline const struct folio_owner_ops *folio_get_owner_ops(struct folio *folio)
+{
+	const struct folio_owner_ops *owner_ops = folio->owner_ops;
+
+	if (!((unsigned long)owner_ops & FOLIO_OWNER_OPS))
+		return NULL;
+
+	owner_ops = (const struct folio_owner_ops *)((unsigned long)owner_ops & ~FOLIO_OWNER_OPS);
+	return owner_ops;
+}
+
 struct page_frag_cache {
 	void * va;
 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
mm/swap.c +19 -0
index 638a3f001676..767ff6d8f47b 100644
@@ -110,6 +110,13 @@ static void page_cache_release(struct folio *folio)
 
 void __folio_put(struct folio *folio)
 {
+	const struct folio_owner_ops *owner_ops = folio_get_owner_ops(folio);
+
+	if (unlikely(owner_ops)) {
+		owner_ops->free(folio);
+		return;
+	}
+
 	if (unlikely(folio_is_zone_device(folio))) {
 		free_zone_device_folio(folio);
 		return;
@@ -929,10 +936,22 @@ void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 	for (i = 0, j = 0; i < folios->nr; i++) {
 		struct folio *folio = folios->folios[i];
 		unsigned int nr_refs = refs ? refs[i] : 1;
+		const struct folio_owner_ops *owner_ops;
 
 		if (is_huge_zero_folio(folio))
 			continue;
 
+		owner_ops = folio_get_owner_ops(folio);
+		if (unlikely(owner_ops)) {
+			if (lruvec) {
+				unlock_page_lruvec_irqrestore(lruvec, flags);
+				lruvec = NULL;
+			}
+			if (folio_ref_sub_and_test(folio, nr_refs))
+				owner_ops->free(folio);
+			continue;
+		}
+
 		if (folio_is_zone_device(folio)) {
 			if (lruvec) {
 				unlock_page_lruvec_irqrestore(lruvec, flags);
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#10 lore raw

RFC PATCH 8/10 mm: Use getters and setters to access page pgmap

In reply to Fuad Tabba

The pointer to pgmap in struct page is overlaid with folio owner_ops. To indicate that a page/folio has owner ops, bit 1 is set. Therefore, before we can start to using owner_ops, we need to ensure that all accesses to page pgmap sanitize the pointer value.

This patch introduces the accessors, which will be modified in the following patch to sanitize the pointer values.

No functional change intended.

---
 drivers/gpu/drm/nouveau/nouveau_dmem.c |  4 +++-
 drivers/pci/p2pdma.c                   |  8 +++++---
 include/linux/memremap.h               |  6 +++---
 include/linux/mm_types.h               | 13 +++++++++++++
 lib/test_hmm.c                         |  2 +-
 mm/hmm.c                               |  2 +-
 mm/memory.c                            |  2 +-
 mm/memremap.c                          | 19 +++++++++++--------
 mm/migrate_device.c                    |  4 ++--
 mm/mm_init.c                           |  2 +-
 10 files changed, 41 insertions(+), 21 deletions(-)
drivers/gpu/drm/nouveau/nouveau_dmem.c +3 -1
index 1a072568cef6..d7d9d9476bb0 100644
@@ -88,7 +88,9 @@ struct nouveau_dmem {
 
 static struct nouveau_dmem_chunk *nouveau_page_to_chunk(struct page *page)
 {
-	return container_of(page->pgmap, struct nouveau_dmem_chunk, pagemap);
+	struct dev_pagemap *pgmap = page_get_pgmap(page);
+
+	return container_of(pgmap, struct nouveau_dmem_chunk, pagemap);
 }
 
 static struct nouveau_drm *page_to_drm(struct page *page)
drivers/pci/p2pdma.c +5 -3
index 4f47a13cb500..19519bb4ba56 100644
@@ -193,7 +193,7 @@ static const struct attribute_group p2pmem_group = {
 
 static void p2pdma_page_free(struct page *page)
 {
-	struct pci_p2pdma_pagemap *pgmap = to_p2p_pgmap(page->pgmap);
+	struct pci_p2pdma_pagemap *pgmap = to_p2p_pgmap(page_get_pgmap(page));
 	/* safe to dereference while a reference is held to the percpu ref */
 	struct pci_p2pdma *p2pdma =
 		rcu_dereference_protected(pgmap->provider->p2pdma, 1);
@@ -1016,8 +1016,10 @@ enum pci_p2pdma_map_type
 pci_p2pdma_map_segment(struct pci_p2pdma_map_state *state, struct device *dev,
 		       struct scatterlist *sg)
 {
-	if (state->pgmap != sg_page(sg)->pgmap) {
-		state->pgmap = sg_page(sg)->pgmap;
+	struct dev_pagemap *pgmap = page_get_pgmap(sg_page(sg));
+
+	if (state->pgmap != pgmap) {
+		state->pgmap = pgmap;
 		state->map = pci_p2pdma_map_type(state->pgmap, dev);
 		state->bus_off = to_p2p_pgmap(state->pgmap)->bus_offset;
 	}
include/linux/memremap.h +3 -3
index 3f7143ade32c..060e27b6aee0 100644
@@ -161,7 +161,7 @@ static inline bool is_device_private_page(const struct page *page)
 {
 	return IS_ENABLED(CONFIG_DEVICE_PRIVATE) &&
 		is_zone_device_page(page) &&
-		page->pgmap->type == MEMORY_DEVICE_PRIVATE;
+		page_get_pgmap(page)->type == MEMORY_DEVICE_PRIVATE;
 }
 
 static inline bool folio_is_device_private(const struct folio *folio)
@@ -173,13 +173,13 @@ static inline bool is_pci_p2pdma_page(const struct page *page)
 {
 	return IS_ENABLED(CONFIG_PCI_P2PDMA) &&
 		is_zone_device_page(page) &&
-		page->pgmap->type == MEMORY_DEVICE_PCI_P2PDMA;
+		page_get_pgmap(page)->type == MEMORY_DEVICE_PCI_P2PDMA;
 }
 
 static inline bool is_device_coherent_page(const struct page *page)
 {
 	return is_zone_device_page(page) &&
-		page->pgmap->type == MEMORY_DEVICE_COHERENT;
+		page_get_pgmap(page)->type == MEMORY_DEVICE_COHERENT;
 }
 
 static inline bool folio_is_device_coherent(const struct folio *folio)
include/linux/mm_types.h +13 -0
index 6e06286f44f1..27075ea24e67 100644
@@ -616,6 +616,19 @@ static inline const struct folio_owner_ops *folio_get_owner_ops(struct folio *fo
 	return owner_ops;
 }
 
+/*
+ * Get the page dev_pagemap pgmap pointer.
+ */
+#define page_get_pgmap(page)	((page)->pgmap)
+
+/*
+ * Set the page dev_pagemap pgmap pointer.
+ */
+static inline void page_set_pgmap(struct page *page, struct dev_pagemap *pgmap)
+{
+	page->pgmap = pgmap;
+}
+
 struct page_frag_cache {
 	void * va;
 #if (PAGE_SIZE < PAGE_FRAG_CACHE_MAX_SIZE)
lib/test_hmm.c +1 -1
index 056f2e411d7b..d3e3843f57dd 100644
@@ -195,7 +195,7 @@ static int dmirror_fops_release(struct inode *inode, struct file *filp)
 
 static struct dmirror_chunk *dmirror_page_to_chunk(struct page *page)
 {
-	return container_of(page->pgmap, struct dmirror_chunk, pagemap);
+	return container_of(page_get_pgmap(page), struct dmirror_chunk, pagemap);
 }
 
 static struct dmirror_device *dmirror_page_to_device(struct page *page)
mm/hmm.c +1 -1
index 7e0229ae4a5a..b5f5ac218fda 100644
@@ -248,7 +248,7 @@ static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
 		 * just report the PFN.
 		 */
 		if (is_device_private_entry(entry) &&
-		    pfn_swap_entry_to_page(entry)->pgmap->owner ==
+		    page_get_pgmap(pfn_swap_entry_to_page(entry))->owner ==
 		    range->dev_private_owner) {
 			cpu_flags = HMM_PFN_VALID;
 			if (is_writable_device_private_entry(entry))
mm/memory.c +1 -1
index 80850cad0e6f..5853fa5767c7 100644
@@ -4276,7 +4276,7 @@ vm_fault_t do_swap_page(struct vm_fault *vmf)
 			 */
 			get_page(vmf->page);
 			pte_unmap_unlock(vmf->pte, vmf->ptl);
-			ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
+			ret = page_get_pgmap(vmf->page)->ops->migrate_to_ram(vmf);
 			put_page(vmf->page);
 		} else if (is_hwpoison_entry(entry)) {
 			ret = VM_FAULT_HWPOISON;
mm/memremap.c +11 -8
index 40d4547ce514..931bc85da1df 100644
@@ -458,8 +458,9 @@ EXPORT_SYMBOL_GPL(get_dev_pagemap);
 
 void free_zone_device_folio(struct folio *folio)
 {
-	if (WARN_ON_ONCE(!folio->page.pgmap->ops ||
-			!folio->page.pgmap->ops->page_free))
+	struct dev_pagemap *pgmap = page_get_pgmap(&folio->page);
+
+	if (WARN_ON_ONCE(!pgmap->ops || !pgmap->ops->page_free))
 		return;
 
 	mem_cgroup_uncharge(folio);
@@ -486,17 +487,17 @@ void free_zone_device_folio(struct folio *folio)
 	 * to clear folio->mapping.
 	 */
 	folio->mapping = NULL;
-	folio->page.pgmap->ops->page_free(folio_page(folio, 0));
+	pgmap->ops->page_free(folio_page(folio, 0));
 
-	if (folio->page.pgmap->type != MEMORY_DEVICE_PRIVATE &&
-	    folio->page.pgmap->type != MEMORY_DEVICE_COHERENT)
+	if (pgmap->type != MEMORY_DEVICE_PRIVATE &&
+	    pgmap->type != MEMORY_DEVICE_COHERENT)
 		/*
 		 * Reset the refcount to 1 to prepare for handing out the page
 		 * again.
 		 */
 		folio_set_count(folio, 1);
 	else
-		put_dev_pagemap(folio->page.pgmap);
+		put_dev_pagemap(pgmap);
 }
 
 void zone_device_page_init(struct page *page)
@@ -505,7 +506,7 @@ void zone_device_page_init(struct page *page)
 	 * Drivers shouldn't be allocating pages after calling
 	 * memunmap_pages().
 	 */
-	WARN_ON_ONCE(!percpu_ref_tryget_live(&page->pgmap->ref));
+	WARN_ON_ONCE(!percpu_ref_tryget_live(&page_get_pgmap(page)->ref));
 	set_page_count(page, 1);
 	lock_page(page);
 }
@@ -514,7 +515,9 @@ EXPORT_SYMBOL_GPL(zone_device_page_init);
 #ifdef CONFIG_FS_DAX
 bool __put_devmap_managed_folio_refs(struct folio *folio, int refs)
 {
-	if (folio->page.pgmap->type != MEMORY_DEVICE_FS_DAX)
+	struct dev_pagemap *pgmap = page_get_pgmap(&folio->page);
+
+	if (pgmap->type != MEMORY_DEVICE_FS_DAX)
 		return false;
 
 	/*
mm/migrate_device.c +2 -2
index 9cf26592ac93..368def358d02 100644
@@ -135,7 +135,7 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,
 			page = pfn_swap_entry_to_page(entry);
 			if (!(migrate->flags &
 				MIGRATE_VMA_SELECT_DEVICE_PRIVATE) ||
-			    page->pgmap->owner != migrate->pgmap_owner)
+			    page_get_pgmap(page)->owner != migrate->pgmap_owner)
 				goto next;
 
 			mpfn = migrate_pfn(page_to_pfn(page)) |
@@ -156,7 +156,7 @@ static int migrate_vma_collect_pmd(pmd_t *pmdp,
 				goto next;
 			else if (page && is_device_coherent_page(page) &&
 			    (!(migrate->flags & MIGRATE_VMA_SELECT_DEVICE_COHERENT) ||
-			     page->pgmap->owner != migrate->pgmap_owner))
+			     page_get_pgmap(page)->owner != migrate->pgmap_owner))
 				goto next;
 			mpfn = migrate_pfn(pfn) | MIGRATE_PFN_MIGRATE;
 			mpfn |= pte_write(pte) ? MIGRATE_PFN_WRITE : 0;
mm/mm_init.c +1 -1
index 1c205b0a86ed..279cdaebfd2b 100644
@@ -995,7 +995,7 @@ static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
 	 * and zone_device_data.  It is a bug if a ZONE_DEVICE page is
 	 * ever freed or placed on a driver-private list.
 	 */
-	page->pgmap = pgmap;
+	page_set_pgmap(page, pgmap);
 	page->zone_device_data = NULL;
 
 	/*
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#11 lore raw

RFC PATCH 9/10 mm: Use owner_ops on folio_put for zone device pages

In reply to Fuad Tabba

Now that we have the folio_owner_ops callback, use it for zone device pages instead of using a dedicated callback.

Note that struct dev_pagemap (pgmap) in struct page is overlaid with struct folio owner_ops. Therefore, make struct dev_pagemap contain an instance of struct folio_owner_ops, to handle it the same way as struct folio_owner_ops.

Also note that, although struct dev_pagemap_ops has a page_free() function, it does not have the same intention as the folio_owner_ops free() callback nor does it have the same behavior. The page_free() function is used as an optional callback to drivers that use zone device to inform them of the freeing of the page.

---
 include/linux/memremap.h |  8 +++++++
 include/linux/mm_types.h | 16 ++++++++++++--
 mm/internal.h            |  1 -
 mm/memremap.c            | 44 --------------------------------------
 mm/mm_init.c             | 46 ++++++++++++++++++++++++++++++++++++++++
 mm/swap.c                | 18 ++--------------
 6 files changed, 70 insertions(+), 63 deletions(-)
include/linux/memremap.h +8 -0
index 060e27b6aee0..5b68bbc588a3 100644
@@ -106,6 +106,7 @@ struct dev_pagemap_ops {
 
 /**
  * struct dev_pagemap - metadata for ZONE_DEVICE mappings
+ * @folio_ops: method table for folio operations.
  * @altmap: pre-allocated/reserved memory for vmemmap allocations
  * @ref: reference count that pins the devm_memremap_pages() mapping
  * @done: completion for @ref
@@ -125,6 +126,7 @@ struct dev_pagemap_ops {
  * @ranges: array of ranges to be mapped when nr_range > 1
  */
 struct dev_pagemap {
+	struct folio_owner_ops folio_ops;
 	struct vmem_altmap altmap;
 	struct percpu_ref ref;
 	struct completion done;
@@ -140,6 +142,12 @@ struct dev_pagemap {
 	};
 };
 
+/*
+ * The folio_owner_ops structure needs to be first since pgmap in struct page is
+ * overlaid with owner_ops in struct folio.
+ */
+static_assert(offsetof(struct dev_pagemap, folio_ops) == 0);
+
 static inline bool pgmap_has_memory_failure(struct dev_pagemap *pgmap)
 {
 	return pgmap->ops && pgmap->ops->memory_failure;
include/linux/mm_types.h +14 -2
index 27075ea24e67..a72fda20d5e9 100644
@@ -427,6 +427,7 @@ FOLIO_MATCH(lru, lru);
 FOLIO_MATCH(mapping, mapping);
 FOLIO_MATCH(compound_head, lru);
 FOLIO_MATCH(compound_head, owner_ops);
+FOLIO_MATCH(pgmap, owner_ops);
 FOLIO_MATCH(index, index);
 FOLIO_MATCH(private, private);
 FOLIO_MATCH(_mapcount, _mapcount);
@@ -618,15 +619,26 @@ static inline const struct folio_owner_ops *folio_get_owner_ops(struct folio *fo
 
 /*
  * Get the page dev_pagemap pgmap pointer.
+ *
+ * The page pgmap is overlaid with the folio owner_ops, where bit 1 is used to
+ * indicate that the page/folio has owner ops. The dev_pagemap contains
+ * owner_ops and is handled the same way. The getter returns a sanitized
+ * pointer.
  */
-#define page_get_pgmap(page)	((page)->pgmap)
+#define page_get_pgmap(page)							\
+	((struct dev_pagemap *)((unsigned long)(page)->pgmap & ~FOLIO_OWNER_OPS))
 
 /*
  * Set the page dev_pagemap pgmap pointer.
+ *
+ * The page pgmap is overlaid with the folio owner_ops, where bit 1 is used to
+ * indicate that the page/folio has owner ops. The dev_pagemap contains
+ * owner_ops and is handled the same way. The setter sets bit 1 to indicate
+ * that the page owner_ops.
  */
 static inline void page_set_pgmap(struct page *page, struct dev_pagemap *pgmap)
 {
-	page->pgmap = pgmap;
+	page->pgmap = (struct dev_pagemap *)((unsigned long)pgmap | FOLIO_OWNER_OPS);
 }
 
 struct page_frag_cache {
mm/internal.h +0 -1
index 5a7302baeed7..a041247bed10 100644
@@ -1262,7 +1262,6 @@ int numa_migrate_check(struct folio *folio, struct vm_fault *vmf,
 		      unsigned long addr, int *flags, bool writable,
 		      int *last_cpupid);
 
-void free_zone_device_folio(struct folio *folio);
 int migrate_device_coherent_folio(struct folio *folio);
 
 struct vm_struct *__get_vm_area_node(unsigned long size,
mm/memremap.c +0 -44

50 lines not on this page · raw message

mm/mm_init.c +46 -0

58 lines not on this page · raw message

mm/swap.c +2 -16
index 767ff6d8f47b..d2578465e270 100644
@@ -117,11 +117,6 @@ void __folio_put(struct folio *folio)
 		return;
 	}
 
-	if (unlikely(folio_is_zone_device(folio))) {
-		free_zone_device_folio(folio);
-		return;
-	}
-
 	if (folio_test_hugetlb(folio)) {
 		free_huge_folio(folio);
 		return;
@@ -947,20 +942,11 @@ void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 				unlock_page_lruvec_irqrestore(lruvec, flags);
 				lruvec = NULL;
 			}
-			if (folio_ref_sub_and_test(folio, nr_refs))
-				owner_ops->free(folio);
-			continue;
-		}
-
-		if (folio_is_zone_device(folio)) {
-			if (lruvec) {
-				unlock_page_lruvec_irqrestore(lruvec, flags);
-				lruvec = NULL;
-			}
+			/* fenced by folio_is_zone_device() */
 			if (put_devmap_managed_folio_refs(folio, nr_refs))
 				continue;
 			if (folio_ref_sub_and_test(folio, nr_refs))
-				free_zone_device_folio(folio);
+				owner_ops->free(folio);
 			continue;
 		}
 
Signature · 1 line
2.47.0.277.g8800431eea-goog
author
#12 lore raw

RFC PATCH 10/10 mm: hugetlb: Use owner_ops on folio_put for hugetlb

In reply to Fuad Tabba

Now that we have the folio_owner_ops callback, use it for hugetlb pages instead of using a dedicated callback.

Since owner_ops is overlaid with lru, we need to unset owner_ops to allow the use of lru when its isolated. At that point we know that the reference count is elevated, will not reach 0, and thus not trigger a callback. Therefore, it is safe to do so provided we restore it before we put the folio back.

---
 include/linux/hugetlb.h |  2 --
 mm/hugetlb.c            | 57 +++++++++++++++++++++++++++++++++--------
 mm/swap.c               | 14 ----------
 3 files changed, 47 insertions(+), 26 deletions(-)
include/linux/hugetlb.h +0 -2
index e846d7dac77c..500848862702 100644
@@ -20,8 +20,6 @@ struct user_struct;
 struct mmu_gather;
 struct node;
 
-void free_huge_folio(struct folio *folio);
-
 #ifdef CONFIG_HUGETLB_PAGE
 
 #include <linux/pagemap.h>
mm/hugetlb.c +47 -10

129 lines not on this page · raw message

mm/swap.c +0 -14
index d2578465e270..9798ca47f26a 100644
@@ -117,11 +117,6 @@ void __folio_put(struct folio *folio)
 		return;
 	}
 
-	if (folio_test_hugetlb(folio)) {
-		free_huge_folio(folio);
-		return;
-	}
-
 	page_cache_release(folio);
 	folio_unqueue_deferred_split(folio);
 	mem_cgroup_uncharge(folio);
@@ -953,15 +948,6 @@ void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 		if (!folio_ref_sub_and_test(folio, nr_refs))
 			continue;
 
-		/* hugetlb has its own memcg */
-		if (folio_test_hugetlb(folio)) {
-			if (lruvec) {
-				unlock_page_lruvec_irqrestore(lruvec, flags);
-				lruvec = NULL;
-			}
-			free_huge_folio(folio);
-			continue;
-		}
 		folio_unqueue_deferred_split(folio);
 		__page_cache_release(folio, &lruvec, &flags);
 
Signature · 1 line
2.47.0.277.g8800431eea-goog
maintainer
#13 lore raw
In reply to Fuad Tabba
Quoted 8 lines · Fuad Tabba

Some folios, such as hugetlb folios and zone device folios, require special handling when the folio's reference count reaches 0, before being freed. Moreover, guest_memfd folios will likely require special handling to notify it once a folio's reference count reaches 0, to facilitate shared to private folio conversion [*]. Currently, each usecase has a dedicated callback when the folio refcount reaches 0 to that effect. Adding yet more callbacks is not ideal.

Honestly, I question this thesis. How complex would it be to have 'yet more callbacks'? Is the challenge really that the mm can't detect when guestmemfd is the owner of the page because the page will be ZONE_NORMAL?

So the point of this is really to allow ZONE_NORMAL pages to have a per-allocator callback?

But this is also why I suggested to shift them to ZONE_DEVICE for guestmemfd, because then you get these things for free from the pgmap.

(this is not a disagreement this is a valid solution, but a request you explain much more about what it is you actually need and compare it with the other existing options)

Jason

#14 lore raw
Quoted 10 lines · Jason Gunthorpe
Quoted 8 lines · Fuad Tabba

Some folios, such as hugetlb folios and zone device folios, require special handling when the folio's reference count reaches 0, before being freed. Moreover, guest_memfd folios will likely require special handling to notify it once a folio's reference count reaches 0, to facilitate shared to private folio conversion [*]. Currently, each usecase has a dedicated callback when the folio refcount reaches 0 to that effect. Adding yet more callbacks is not ideal.

Thanks for having a look!

Replying to clarify some things. Fuad, feel free to add additional information.

Quoted 4 lines

Honestly, I question this thesis. How complex would it be to have 'yet more callbacks'? Is the challenge really that the mm can't detect when guestmemfd is the owner of the page because the page will be ZONE_NORMAL?

Fuad might have been a bit imprecise here: We don't want an ever growing list of checks+callbacks on the page freeing fast path.

This series replaces the two cases we have by a single generic one, which is nice independent of guest_memfd I think.

Quoted 3 lines

So the point of this is really to allow ZONE_NORMAL pages to have a per-allocator callback?

To intercept the refcount going to zero independent of any zones or magic page types, without as little overhead in the common page freeing path.

It can be used to implement custom allocators, like factored out for hugetlb in this series. It's not necessarily limited to that, though. It can be used as a form of "asynchronous page ref freezing", where you get notified once all references are gone.

(I might have another use case with PageOffline, where we want to prevent virtio-mem ones of them from getting accidentally leaked into the buddy during memory offlining with speculative references -- virtio_mem_fake_offline_going_offline() contains the interesting bits. But I did not look into the dirty details yet, just some thought where we'd want to intercept the refcount going to 0.)

Quoted 3 lines

But this is also why I suggested to shift them to ZONE_DEVICE for guestmemfd, because then you get these things for free from the pgmap.

With this series even hugetlb gets it for "free", and hugetlb is not quite the nail for the ZONE_DEVICE hammer IMHO :)

For things we can statically set aside early during boot and never really want to return to the buddy/another allocator, I would agree that static ZONE_DEVICE would have possible.

Whenever the buddy or other allocators are involved, and we might have granularity as a handful of pages (e.g., taken from the buddy), getting ZONE_DEVICE involved is not a good (or even feasible) approach.

After all, all we want is intercept the refcount going to 0.

Signature · 3 lines
Cheers,

David / dhildenb
author
#15 lore raw

Hi Jason and David,

Quoted 13 lines · David Hildenbrand
Quoted 10 lines · Jason Gunthorpe
Quoted 8 lines · Fuad Tabba

Some folios, such as hugetlb folios and zone device folios, require special handling when the folio's reference count reaches 0, before being freed. Moreover, guest_memfd folios will likely require special handling to notify it once a folio's reference count reaches 0, to facilitate shared to private folio conversion [*]. Currently, each usecase has a dedicated callback when the folio refcount reaches 0 to that effect. Adding yet more callbacks is not ideal.

Quoted 4 lines

Thanks for having a look!

Replying to clarify some things. Fuad, feel free to add additional information.

Thanks for your comments Jason, and for clarifying my cover letter David. I think David has covered everything, and I'll make sure to clarify this in the cover letter when I respin.

Cheers,
/fuad

Quoted 54 lines
Quoted 5 lines

Honestly, I question this thesis. How complex would it be to have 'yet more callbacks'? Is the challenge really that the mm can't detect when guestmemfd is the owner of the page because the page will be ZONE_NORMAL?

Fuad might have been a bit imprecise here: We don't want an ever growing list of checks+callbacks on the page freeing fast path.

This series replaces the two cases we have by a single generic one, which is nice independent of guest_memfd I think.

Quoted 3 lines

So the point of this is really to allow ZONE_NORMAL pages to have a per-allocator callback?

To intercept the refcount going to zero independent of any zones or magic page types, without as little overhead in the common page freeing path.

It can be used to implement custom allocators, like factored out for hugetlb in this series. It's not necessarily limited to that, though. It can be used as a form of "asynchronous page ref freezing", where you get notified once all references are gone.

(I might have another use case with PageOffline, where we want to prevent virtio-mem ones of them from getting accidentally leaked into the buddy during memory offlining with speculative references -- virtio_mem_fake_offline_going_offline() contains the interesting bits. But I did not look into the dirty details yet, just some thought where we'd want to intercept the refcount going to 0.)

Quoted 3 lines

But this is also why I suggested to shift them to ZONE_DEVICE for guestmemfd, because then you get these things for free from the pgmap.

With this series even hugetlb gets it for "free", and hugetlb is not quite the nail for the ZONE_DEVICE hammer IMHO :)

For things we can statically set aside early during boot and never really want to return to the buddy/another allocator, I would agree that static ZONE_DEVICE would have possible.

Whenever the buddy or other allocators are involved, and we might have granularity as a handful of pages (e.g., taken from the buddy), getting ZONE_DEVICE involved is not a good (or even feasible) approach.

After all, all we want is intercept the refcount going to 0.

--
Cheers,

David / dhildenb

#16 lore raw
Quoted 3 lines · Fuad Tabba

Thanks for your comments Jason, and for clarifying my cover letter David. I think David has covered everything, and I'll make sure to clarify this in the cover letter when I respin.

I don't want you to respin. I think this is a bad idea.

#17 lore raw
Quoted 6 lines · Matthew Wilcox
Quoted 3 lines · Fuad Tabba

Thanks for your comments Jason, and for clarifying my cover letter David. I think David has covered everything, and I'll make sure to clarify this in the cover letter when I respin.

I don't want you to respin. I think this is a bad idea.

I'm hoping you'll find some more time to explain what exactly you don't like, because this series only refactors what we already have.

I enjoy seeing the special casing (especially hugetlb) gone from mm/swap.c.

I don't particularly enjoy overlaying folio->lru, primarily because we have to temporarily "evacuate" it when someone wants to make use of folio->lru (e.g., hugetlb isolation). So it's not completely "sticky", at least for hugetlb.

Overlaying folio->mapping, similar to how "struct movable_operations" overlay page->mapping is not an option, because folio->mapping will be used for other purposes.

We'd need some sticky and reliable way to tell folio freeing code that someone wants to intercept when the refcount of that folio goes to 0, and identify who to notify.

Maybe folio->private/page->private could be overlayed? hugetlb only uses folio->private for flags, which we could move to some other tail page (e.g., simply putting them into flags1).

Signature · 3 lines
Cheers,

David / dhildenb
maintainer
#18 lore raw
Quoted 11 lines · David Hildenbrand
Quoted 6 lines · Matthew Wilcox
Quoted 3 lines · Fuad Tabba

Thanks for your comments Jason, and for clarifying my cover letter David. I think David has covered everything, and I'll make sure to clarify this in the cover letter when I respin.

I don't want you to respin. I think this is a bad idea.

I'm hoping you'll find some more time to explain what exactly you don't like, because this series only refactors what we already have.

Quoted 6 lines

I enjoy seeing the special casing (especially hugetlb) gone from mm/swap.c.

I don't particularly enjoy overlaying folio->lru, primarily because we have to temporarily "evacuate" it when someone wants to make use of folio->lru (e.g., hugetlb isolation). So it's not completely "sticky", at least for hugetlb.

This is really the worst part of it though

And, IMHO, seems like overkill. We have only a handful of cases - maybe we shouldn't be trying to get to full generality but just handle a couple of cases directly? I don't really think it is such a bad thing to have an if ladder on the free path if we have only a couple things. Certainly it looks good instead of doing overlaying tricks.

Also how does this translate to Matthew's memdesc world?

Jason

#19 lore raw
Quoted 12 lines · Jason Gunthorpe
Quoted 11 lines · David Hildenbrand
Quoted 6 lines · Matthew Wilcox
Quoted 3 lines · Fuad Tabba

Thanks for your comments Jason, and for clarifying my cover letter David. I think David has covered everything, and I'll make sure to clarify this in the cover letter when I respin.

I don't want you to respin. I think this is a bad idea.

I'm hoping you'll find some more time to explain what exactly you don't like, because this series only refactors what we already have.

Quoted 8 lines
Quoted 6 lines

I enjoy seeing the special casing (especially hugetlb) gone from mm/swap.c.

I don't particularly enjoy overlaying folio->lru, primarily because we have to temporarily "evacuate" it when someone wants to make use of folio->lru (e.g., hugetlb isolation). So it's not completely "sticky", at least for hugetlb.

This is really the worst part of it though

Yes.

Quoted 6 lines

And, IMHO, seems like overkill. We have only a handful of cases - maybe we shouldn't be trying to get to full generality but just handle a couple of cases directly? I don't really think it is such a bad thing to have an if ladder on the free path if we have only a couple things. Certainly it looks good instead of doing overlaying tricks.

I'd really like to abstract hugetlb handling if possible. The way it stands it's just very odd.

We'll need some reliable way to identify these folios that need care. guest_memfd will be using folio->mapcount for now, so for now we couldn't set a page type like hugetlb does.

Quoted 1 line

Also how does this translate to Matthew's memdesc world?

guest_memfd and hugetlb would be operating on folios (at least for now), which contain the refcount,lru,private, ... so nothing special there.

Once we actually decoupled "struct folio" from "struct page", we *might* have to play less tricks, because we could just have a callback pointer there. But well, maybe we also want to save some space in there.

Do we want dedicated memdescs for hugetlb/guest_memfd that extend folios in the future? I don't know, maybe.

I'm currently wondering if we can use folio->private for the time being. Either

(a) If folio->private is still set once the refcount drops to 0, it indicates that there is a freeing callback/owner_ops. We'll have to make hugetlb not use folio->private and convert others to clear folio->private before freeing.

(b) Use bitX of folio->private to indicate that this has "owner_ops" meaning. We'll have to make hugetlb not use folio->private and make others not use bitX. Might be harder and overkill, because right now we only really need the callback when refcount==0.

(c) Use some other indication that folio->private contains folio_ops.

Signature · 3 lines
Cheers,

David / dhildenb
#20 lore raw
Quoted 9 lines · David Hildenbrand
Quoted 8 lines · Jason Gunthorpe
Quoted 7 lines · David Hildenbrand
Quoted 1 line · Matthew Wilcox

I don't want you to respin. I think this is a bad idea.

I'm hoping you'll find some more time to explain what exactly you don't like, because this series only refactors what we already have.

I enjoy seeing the special casing (especially hugetlb) gone from mm/swap.c.

I don't. The list of 'if's is better than the indirect function call. That's terribly expensive, and the way we reuse the lru.next field is fragile. Not to mention that it introduces a new thing for the hardening people to fret over.

Quoted 8 lines
Quoted 5 lines

And, IMHO, seems like overkill. We have only a handful of cases - maybe we shouldn't be trying to get to full generality but just handle a couple of cases directly? I don't really think it is such a bad thing to have an if ladder on the free path if we have only a couple things. Certainly it looks good instead of doing overlaying tricks.

I'd really like to abstract hugetlb handling if possible. The way it stands it's just very odd.

There might be ways to make that better. I haven't really been looking too hard at making that special handling go away.

Quoted 3 lines

We'll need some reliable way to identify these folios that need care. guest_memfd will be using folio->mapcount for now, so for now we couldn't set a page type like hugetlb does.

If hugetlb can set lru.next at a certain point, then guestmemfd could set a page type at a similar point, no?

Quoted 1 line
Quoted 1 line

Also how does this translate to Matthew's memdesc world?

In a memdesc world, pages no longer have a refcount. We might still have put_page() which will now be a very complicated (and out-of-line) function that looks up what kind of memdesc it is and operates on the memdesc's refcount ... if it has one. I don't know if it'll be exported to modules; I can see uses in the mm code, but I'm not sure if modules will have a need.

Each memdesc type will have its own function to call to free the memdesc. So we'll still have folio_put(). But slab does not have, need nor want a refcount, so it'll just slab_free(). I expect us to keep around a list of recently-freed memdescs of a particular type with their pages still attached so that we can allocate them again quickly (or reclaim them under memory pressure). Once that freelist overflows, we'll free a batch of them to the buddy allocator (for the pages) and the slab allocator (for the memdesc itself).

Quoted 9 lines

guest_memfd and hugetlb would be operating on folios (at least for now), which contain the refcount,lru,private, ... so nothing special there.

Once we actually decoupled "struct folio" from "struct page", we *might* have to play less tricks, because we could just have a callback pointer there. But well, maybe we also want to save some space in there.

Do we want dedicated memdescs for hugetlb/guest_memfd that extend folios in the future? I don't know, maybe.

I've certainly considered going so far as a per-fs folio. So we'd have an ext4_folio, an btrfs_folio, an iomap_folio, etc. That'd let us get rid of folio->private, but I'm not sure that C's type system can really handle this nicely. Maybe in a Rust world ;-)

What I'm thinking about is that I'd really like to be able to declare that all the functions in ext4_aops only accept pointers to ext4_folio, so ext4_dirty_folio() can't be called with pointers to _any_ folio, but specifically folios which were previously allocated for ext4.

I don't know if Rust lets you do something like that.

Quoted 8 lines

I'm currently wondering if we can use folio->private for the time being. Either

(a) If folio->private is still set once the refcount drops to 0, it indicates that there is a freeing callback/owner_ops. We'll have to make hugetlb not use folio->private and convert others to clear folio->private before freeing.

Quoted 6 lines

(b) Use bitX of folio->private to indicate that this has "owner_ops" meaning. We'll have to make hugetlb not use folio->private and make others not use bitX. Might be harder and overkill, because right now we only really need the callback when refcount==0.

(c) Use some other indication that folio->private contains folio_ops.

I really don't want to use folio_ops / folio_owner_ops. I read
https://lore.kernel.org/all/CAGtprH_JP2w-4rq02h_Ugvq5KuHX7TUvegOS7xUs_iy5hriE7g@mail.gmail.com/
and I still don't understand what you're trying to do.

Would it work to use aops->free_folio() to notify you when the folio is being removed from the address space?

#21 lore raw
Quoted 9 lines · Matthew Wilcox
Quoted 8 lines · David Hildenbrand
Quoted 7 lines · Jason Gunthorpe
Quoted 6 lines · David Hildenbrand
Quoted 1 line · Matthew Wilcox

I don't want you to respin. I think this is a bad idea.

I'm hoping you'll find some more time to explain what exactly you don't like, because this series only refactors what we already have.

Quoted 6 lines
Quoted 1 line
Quoted 1 line
Quoted 1 line

I enjoy seeing the special casing (especially hugetlb) gone from mm/swap.c.

I don't. The list of 'if's is better than the indirect function call. That's terribly expensive, and the way we reuse the lru.next field is fragile. Not to mention that it introduces a new thing for the hardening people to fret over.

Right, indirect calls are nasty and probably more fragile/insecure, but this is really what ZONE_DEVICE is already using internally ... but I agree that is less desirable to abstract that.

[...]

Quoted 6 lines
Quoted 3 lines

We'll need some reliable way to identify these folios that need care. guest_memfd will be using folio->mapcount for now, so for now we couldn't set a page type like hugetlb does.

If hugetlb can set lru.next at a certain point, then guestmemfd could set a page type at a similar point, no?

The main problem is that we will have to set it on small folios that can be mapped to user space. As long as mapcount overlays page_type, that's ... not going to work.

We won't be truncating+freeing the folio as long as it is mapped to user space, though. So one workaround would be to set the page type only as long as it isn't mapped to user space.

Won't win a beauty price, but could be one workaround until we decoupled the type from the mapcount.

[...]

Quoted 8 lines
Quoted 8 lines

guest_memfd and hugetlb would be operating on folios (at least for now), which contain the refcount,lru,private, ... so nothing special there.

Once we actually decoupled "struct folio" from "struct page", we *might* have to play less tricks, because we could just have a callback pointer there. But well, maybe we also want to save some space in there.

Quoted 7 lines
Quoted 2 lines

Do we want dedicated memdescs for hugetlb/guest_memfd that extend folios in the future? I don't know, maybe.

I've certainly considered going so far as a per-fs folio. So we'd have an ext4_folio, an btrfs_folio, an iomap_folio, etc. That'd let us get rid of folio->private, but I'm not sure that C's type system can really handle this nicely. Maybe in a Rust world ;-)

:)

Quoted 5 lines

What I'm thinking about is that I'd really like to be able to declare that all the functions in ext4_aops only accept pointers to ext4_folio, so ext4_dirty_folio() can't be called with pointers to _any_ folio, but specifically folios which were previously allocated for ext4.

Yes, that sounds reasonable. hugetlb definetly might be another such candidate.

Quoted 11 lines

I don't know if Rust lets you do something like that.

Quoted 8 lines

I'm currently wondering if we can use folio->private for the time being. Either

(a) If folio->private is still set once the refcount drops to 0, it indicates that there is a freeing callback/owner_ops. We'll have to make hugetlb not use folio->private and convert others to clear folio->private before freeing.

Quoted 8 lines
Quoted 6 lines

(b) Use bitX of folio->private to indicate that this has "owner_ops" meaning. We'll have to make hugetlb not use folio->private and make others not use bitX. Might be harder and overkill, because right now we only really need the callback when refcount==0.

(c) Use some other indication that folio->private contains folio_ops.

I really don't want to use folio_ops / folio_owner_ops.

Yes, and I understand your reasoning. It was one approach to work around the page_type vs. mapcount issue and avoiding more checks on the freeing path.

If we manage to make the page type fly, the following could work and leave the ordinary folio freeing path as fast as before (and allow me to add the PGTY_offline handling :) ):

From 5a55e4bcf4d6cfa64d3383a7cf6649042cedcecb Mon Sep 17 00:00:00 2001
From: David Hildenbrand <[email protected]>
Date: Wed, 13 Nov 2024 12:20:58 +0100
Subject: [PATCH] tmp

---
 include/linux/page-flags.h | 11 +++++++++++
 mm/swap.c                  | 27 ++++++++++++++++++++++-----
 2 files changed, 33 insertions(+), 5 deletions(-)
include/linux/page-flags.h +11 -0
index e80665bc51fac..ebf89075eeb5f 100644
@@ -950,6 +950,7 @@ enum pagetype {
 	PGTY_slab	= 0xf5,
 	PGTY_zsmalloc	= 0xf6,
 	PGTY_unaccepted	= 0xf7,
+	PGTY_guestmem	= 0xf8,
 
 	PGTY_mapcount_underflow = 0xff
 };
@@ -970,6 +971,16 @@ static inline bool page_has_type(const struct page *page)
 	return page_mapcount_is_type(data_race(page->page_type));
 }
 
+static inline bool folio_has_type(const struct folio *folio)
+{
+	return page_has_type(&folio->page);
+}
+
+static inline int folio_get_type(const struct folio *folio)
+{
+	return folio->page.page_type >> 24;
+}
+
 #define FOLIO_TYPE_OPS(lname, fname)					\
 static __always_inline bool folio_test_##fname(const struct folio *folio) \
 {									\
mm/swap.c +22 -5
index 10decd9dffa17..bf4efc7bba18a 100644
@@ -94,6 +94,22 @@ static void page_cache_release(struct folio *folio)
 		unlock_page_lruvec_irqrestore(lruvec, flags);
 }
 
+static void free_typed_folio(struct folio *folio)
+{
+	switch (folio_get_type(folio)) {
+	case PGTY_hugetlb:
+		free_huge_folio(folio);
+		return;
+	case PGTY_offline:
+		/* Nothing to do, it's offline. */
+		return;
+	case PGTY_guestmem:
+		// free_guestmem_folio(folio);
+	default:
+		WARN_ON_ONCE(1);
+	}
+}
+
 void __folio_put(struct folio *folio)
 {
 	if (unlikely(folio_is_zone_device(folio))) {
@@ -101,8 +117,8 @@ void __folio_put(struct folio *folio)
 		return;
 	}
 
-	if (folio_test_hugetlb(folio)) {
-		free_huge_folio(folio);
+	if (unlikely(folio_has_type(folio))) {
+		free_typed_folio(folio);
 		return;
 	}
 
@@ -934,15 +950,16 @@ void folios_put_refs(struct folio_batch *folios, unsigned int *refs)
 		if (!folio_ref_sub_and_test(folio, nr_refs))
 			continue;
 
-		/* hugetlb has its own memcg */
-		if (folio_test_hugetlb(folio)) {
+		if (unlikely(folio_has_type(folio))) {
+			/* typed folios have their own memcg, if any */
 			if (lruvec) {
 				unlock_page_lruvec_irqrestore(lruvec, flags);
 				lruvec = NULL;
 			}
-			free_huge_folio(folio);
+			free_typed_folio(folio);
 			continue;
 		}
+
 		folio_unqueue_deferred_split(folio);
 		__page_cache_release(folio, &lruvec, &flags);
 
Signature · 5 lines
2.47.0



I read
Quoted 2 lines

https://lore.kernel.org/all/CAGtprH_JP2w-4rq02h_Ugvq5KuHX7TUvegOS7xUs_iy5hriE7g@mail.gmail.com/
and I still don't understand what you're trying to do.

It's a bunch of related issues (e.g., one idea is also having another "global" pool of large pages that are not owned by hugetlb), but the biggest thing we want to sort out is the following:

We allocated a large folio from somewhere (hugetlb, another global pool). We might have to split this folio internally in some cases (for example, because we need precise per-page refcounts/mapcounts) and we:

(a) Really have to return the whole re-collapsed folio to the original allocator.

(b) Might have to clean up some stuff (e.g., restore original page type) before handing the folio back to the original allocator.

(c) Want to re-collapse (parts of?) the folio when we don't need the split anymore.

Re-collapsing is one problem because of possible GUP pins (for parts that were exposed to user space) and speculative references.

As an example during (a): we might have truncated the whole split thing, but some reference on a split folio prevents us from re-collapsing the whole thing synchronously during truncation. As soon as that last reference is put, we can collapse the thing and return it to the original allocator.

Similar during (c), just that we want to "freeze" all refcounts as the last references are put, so we can just collapse it once we are notified that it is now possible -- and that no other speculative references can show up as the refcount is 0.

Quoted 3 lines

Would it work to use aops->free_folio() to notify you when the folio is being removed from the address space?

We really have to know when the last reference is gone, and intercept that to cleanup and prepare the large folio to go back to its original allocator.

As always, thanks for taking your time and the detailed response Willy!

Signature · 3 lines
Cheers,

David / dhildenb
reviewer
#22 lore raw
In reply to Matthew Wilcox
Quoted 4 lines · Matthew Wilcox
Quoted 3 lines · David Hildenbrand
Quoted 2 lines · Jason Gunthorpe
Quoted 1 line · David Hildenbrand

On 12.11.24 06:26, Matthew Wilcox wrote:

...

Quoted 5 lines

I've certainly considered going so far as a per-fs folio. So we'd have an ext4_folio, an btrfs_folio, an iomap_folio, etc. That'd let us get rid of folio->private, but I'm not sure that C's type system can really handle this nicely. Maybe in a Rust world ;-)

Quoted 7 lines

What I'm thinking about is that I'd really like to be able to declare that all the functions in ext4_aops only accept pointers to ext4_folio, so ext4_dirty_folio() can't be called with pointers to _any_ folio, but specifically folios which were previously allocated for ext4.

I don't know if Rust lets you do something like that.

As Rust-for-Linux student, I can answer that one: "yes".

Some combination of "newtypes" and Traits will provide exactly what you need here. newtypes provide a zero-overhead type safe way of specifying a type, and Traits can be used to require that only types that support specific operations are accepted in foo().

(Rust at the language level looks a lot more like a replacement for C++, than a replacement for C, imho. By which I mean, it has lots of goodies for expressing things, built right into the language.)

thanks,

Signature · 1 line
John Hubbard