[PATCHv3 00/15] mm: Eliminate fake head pages from vmemmap optimization

[PATCHv3 00/15] mm: Eliminate fake head pages from vmemmap optimization

[Kiryl Shutsemau]

This series removes "fake head pages" from the HugeTLB vmemmap
optimization (HVO) by changing how tail pages encode their relationship
to the head page.

It simplifies compound_head() and page_ref_add_unless(). Both are in the
hot path.

Background
==========

HVO reduces memory overhead by freeing vmemmap pages for HugeTLB pages
and remapping the freed virtual addresses to a single physical page.
Previously, all tail page vmemmap entries were remapped to the first
vmemmap page (containing the head struct page), creating "fake heads" -
tail pages that appear to have PG_head set when accessed through the
deduplicated vmemmap.

This required special handling in compound_head() to detect and work
around fake heads, adding complexity and overhead to a very hot path.

New Approach
============

For architectures/configs where sizeof(struct page) is a power of 2 (the
common case), this series changes how position of the head page is encoded
in the tail pages.

Instead of storing a pointer to the head page, the ->compound_info
(renamed from ->compound_head) now stores a mask.

The mask can be applied to any tail page's virtual address to compute
the head page address. Critically, all tail pages of the same order now
have identical compound_info values, regardless of which compound page
they belong to.

The key insight is that all tail pages of the same order now have
identical compound_info values, regardless of which compound page they
belong to. This allows a single page of tail struct pages to be shared
across all huge pages of the same order on a NUMA node.

Benefits
========

1. Simplified compound_head(): No fake head detection needed, can be
   implemented in a branchless manner.

2. Simplified page_ref_add_unless(): RCU protection removed since there's
   no race with fake head remapping.

3. Cleaner architecture: The shared tail pages are truly read-only and
   contain valid tail page metadata.

If sizeof(struct page) is not power-of-2, there are no functional changes.
HVO is not supported in this configuration.

I had hoped to see performance improvement, but my testing thus far has
shown either no change or only a slight improvement within the noise.

Series Organization
===================

Patch 1-2: Preparation - move MAX_FOLIO_ORDER to mmzone.h and build fix
	   for x86/vdso32
Patches 3-5: Refactoring - interface changes, field rename, code movement
Patch 6: Core change - new mask-based compound_head() encoding
Patch 7: Correctness fix - page_zonenum() must use head page
Patch 8: Add memmap alignment check for compound_info_has_mask()
Patch 9: Refactor vmemmap_walk for new design
Patch 10: Eliminate fake heads with shared tail pages
Patches 11-14: Cleanup - remove fake head infrastructure
Patch 15: Documentation update

Changes in v3:
==============
  - Fixed error recovery path in vmemmap_remap_free() to pass correct start
    address for TLB flush. (Muchun)

  - Wrapped the mask-based compound_info encoding within CONFIG_SPARSEMEM_VMEMMAP
    check via compound_info_has_mask(). For other memory models, alignment
    guarantees are harder to verify. (Muchun)

  - Updated vmemmap_dedup.rst documentation wording: changed "vmemmap_tail
    shared for the struct hstate" to "A single, per-node page frame shared
    among all hugepages of the same size". (Muchun)

  - Fixed build error with MAX_FOLIO_ORDER expanding to undefined PUD_ORDER
    in certain configurations. (kernel test robot)

Changes in v2:
==============

- Handle boot-allocated huge pages correctly. (Frank)

- Changed from per-hstate vmemmap_tail to per-node vmemmap_tails[] array
  in pglist_data. (Muchun)

- Added spin_lock(&hugetlb_lock) protection in vmemmap_get_tail() to fix
  a race condition where two threads could both allocate tail pages.
  The losing thread now properly frees its allocated page. (Usama)

- Add warning if memmap is not aligned to MAX_FOLIO_SIZE, which is
  required for the mask approach. (Muchun)

- Make page_zonenum() use head page - correctness fix since shared
  tail pages cannot have valid zone information. (Muchun)

- Added 'const' qualifier to head parameter in set_compound_head() and
  prep_compound_tail(). (Usama)

- Updated commit messages.

Kiryl Shutsemau (15):
  x86/vdso32: Prepare for <linux/pgtable.h> inclusion
  mm: Move MAX_FOLIO_ORDER definition to mmzone.h
  mm: Change the interface of prep_compound_tail()
  mm: Rename the 'compound_head' field in the 'struct page' to
    'compound_info'
  mm: Move set/clear_compound_head() next to compound_head()
  mm: Rework compound_head() for power-of-2 sizeof(struct page)
  mm: Make page_zonenum() use head page
  mm/sparse: Check memmap alignment for compound_info_has_mask()
  mm/hugetlb: Refactor code around vmemmap_walk
  mm/hugetlb: Remove fake head pages
  mm: Drop fake head checks
  hugetlb: Remove VMEMMAP_SYNCHRONIZE_RCU
  mm/hugetlb: Remove hugetlb_optimize_vmemmap_key static key
  mm: Remove the branch from compound_head()
  hugetlb: Update vmemmap_dedup.rst

 .../admin-guide/kdump/vmcoreinfo.rst          |   2 +-
 Documentation/mm/vmemmap_dedup.rst            |  62 ++--
 arch/x86/entry/vdso/vdso32/fake_32bit_build.h |   2 +
 include/linux/mm.h                            |  31 --
 include/linux/mm_types.h                      |  20 +-
 include/linux/mmzone.h                        |  48 ++++
 include/linux/page-flags.h                    | 167 +++++------
 include/linux/page_ref.h                      |   8 +-
 include/linux/types.h                         |   2 +-
 kernel/vmcore_info.c                          |   2 +-
 mm/hugetlb.c                                  |   8 +-
 mm/hugetlb_vmemmap.c                          | 270 +++++++++---------
 mm/internal.h                                 |  12 +-
 mm/mm_init.c                                  |   2 +-
 mm/page_alloc.c                               |   4 +-
 mm/slab.h                                     |   2 +-
 mm/sparse-vmemmap.c                           |  44 ++-
 mm/sparse.c                                   |   5 +
 mm/util.c                                     |  16 +-
 19 files changed, 363 insertions(+), 344 deletions(-)

-- 
2.51.2

[PATCHv3 01/15] x86/vdso32: Prepare for <linux/pgtable.h> inclusion

[Kiryl Shutsemau]

The 32-bit vDSO for 64-bit kernels is built by faking a 32-bit environment
through various #undefs and #defines in fake_32bit_build.h.

Upcoming change will include <linux/pgtable.h> in <linux/mmzone.h>.
Without preparation, it breaks build of 32-bit vDSO because of exposure
to more 64-bit things.

CONFIG_PHYS_ADDR_T_64BIT triggers "Missing MAX_POSSIBLE_PHYSMEM_BITS
definition" error in <linux/pgtable.h>.

And CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS leads to "shift count >=
width of type" errors in pte_flags_pkey().

Undefine CONFIG_PHYS_ADDR_T_64BIT and CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
in fake_32bit_build.h to fix the problem.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Cc: x86@kernel.org
---
 arch/x86/entry/vdso/vdso32/fake_32bit_build.h | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/arch/x86/entry/vdso/vdso32/fake_32bit_build.h b/arch/x86/entry/vdso/vdso32/fake_32bit_build.h
index db1b15f686e3..900cdcde1029 100644
--- a/arch/x86/entry/vdso/vdso32/fake_32bit_build.h
+++ b/arch/x86/entry/vdso/vdso32/fake_32bit_build.h
@@ -13,6 +13,8 @@
 #undef CONFIG_SPARSEMEM_VMEMMAP
 #undef CONFIG_NR_CPUS
 #undef CONFIG_PARAVIRT_XXL
+#undef CONFIG_PHYS_ADDR_T_64BIT
+#undef CONFIG_X86_INTEL_MEMORY_PROTECTION_KEYS
 
 #define CONFIG_X86_32 1
 #define CONFIG_PGTABLE_LEVELS 2
-- 
2.51.2

[PATCHv3 02/15] mm: Move MAX_FOLIO_ORDER definition to mmzone.h

[Kiryl Shutsemau]

Move MAX_FOLIO_ORDER definition from mm.h to mmzone.h.

This is preparation for adding the vmemmap_tails array to struct
pglist_data, which requires MAX_FOLIO_ORDER to be available in mmzone.h.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---
 include/linux/mm.h     | 31 -------------------------------
 include/linux/mmzone.h | 32 ++++++++++++++++++++++++++++++++
 2 files changed, 32 insertions(+), 31 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 7c79b3369b82..2c409f583569 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -26,7 +26,6 @@
 #include <linux/page-flags.h>
 #include <linux/page_ref.h>
 #include <linux/overflow.h>
-#include <linux/sizes.h>
 #include <linux/sched.h>
 #include <linux/pgtable.h>
 #include <linux/kasan.h>
@@ -2074,36 +2073,6 @@ static inline unsigned long folio_nr_pages(const struct folio *folio)
 	return folio_large_nr_pages(folio);
 }
 
-#if !defined(CONFIG_HAVE_GIGANTIC_FOLIOS)
-/*
- * We don't expect any folios that exceed buddy sizes (and consequently
- * memory sections).
- */
-#define MAX_FOLIO_ORDER		MAX_PAGE_ORDER
-#elif defined(CONFIG_SPARSEMEM) && !defined(CONFIG_SPARSEMEM_VMEMMAP)
-/*
- * Only pages within a single memory section are guaranteed to be
- * contiguous. By limiting folios to a single memory section, all folio
- * pages are guaranteed to be contiguous.
- */
-#define MAX_FOLIO_ORDER		PFN_SECTION_SHIFT
-#elif defined(CONFIG_HUGETLB_PAGE)
-/*
- * There is no real limit on the folio size. We limit them to the maximum we
- * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
- * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
- */
-#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
-#else
-/*
- * Without hugetlb, gigantic folios that are bigger than a single PUD are
- * currently impossible.
- */
-#define MAX_FOLIO_ORDER		PUD_ORDER
-#endif
-
-#define MAX_FOLIO_NR_PAGES	(1UL << MAX_FOLIO_ORDER)
-
 /*
  * compound_nr() returns the number of pages in this potentially compound
  * page.  compound_nr() can be called on a tail page, and is defined to
diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 7fb7331c5725..6a2f3696068e 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -21,8 +21,10 @@
 #include <linux/atomic.h>
 #include <linux/mm_types.h>
 #include <linux/page-flags.h>
+#include <linux/pgtable.h>
 #include <linux/local_lock.h>
 #include <linux/zswap.h>
+#include <linux/sizes.h>
 #include <asm/page.h>
 
 /* Free memory management - zoned buddy allocator.  */
@@ -61,6 +63,36 @@
  */
 #define PAGE_ALLOC_COSTLY_ORDER 3
 
+#if !defined(CONFIG_HAVE_GIGANTIC_FOLIOS)
+/*
+ * We don't expect any folios that exceed buddy sizes (and consequently
+ * memory sections).
+ */
+#define MAX_FOLIO_ORDER		MAX_PAGE_ORDER
+#elif defined(CONFIG_SPARSEMEM) && !defined(CONFIG_SPARSEMEM_VMEMMAP)
+/*
+ * Only pages within a single memory section are guaranteed to be
+ * contiguous. By limiting folios to a single memory section, all folio
+ * pages are guaranteed to be contiguous.
+ */
+#define MAX_FOLIO_ORDER		PFN_SECTION_SHIFT
+#elif defined(CONFIG_HUGETLB_PAGE)
+/*
+ * There is no real limit on the folio size. We limit them to the maximum we
+ * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
+ * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
+ */
+#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
+#else
+/*
+ * Without hugetlb, gigantic folios that are bigger than a single PUD are
+ * currently impossible.
+ */
+#define MAX_FOLIO_ORDER		PUD_ORDER
+#endif
+
+#define MAX_FOLIO_NR_PAGES	(1UL << MAX_FOLIO_ORDER)
+
 enum migratetype {
 	MIGRATE_UNMOVABLE,
 	MIGRATE_MOVABLE,
-- 
2.51.2

[PATCHv3 03/15] mm: Change the interface of prep_compound_tail()

[Kiryl Shutsemau]

Instead of passing down the head page and tail page index, pass the tail
and head pages directly, as well as the order of the compound page.

This is a preparation for changing how the head position is encoded in
the tail page.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 include/linux/page-flags.h |  4 +++-
 mm/hugetlb.c               |  8 +++++---
 mm/internal.h              | 12 ++++++------
 mm/mm_init.c               |  2 +-
 mm/page_alloc.c            |  2 +-
 5 files changed, 16 insertions(+), 12 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 0091ad1986bf..d4952573a4af 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -865,7 +865,9 @@ static inline bool folio_test_large(const struct folio *folio)
 	return folio_test_head(folio);
 }
 
-static __always_inline void set_compound_head(struct page *page, struct page *head)
+static __always_inline void set_compound_head(struct page *page,
+					      const struct page *head,
+					      unsigned int order)
 {
 	WRITE_ONCE(page->compound_head, (unsigned long)head + 1);
 }
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 0455119716ec..a55d638975bd 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -3212,6 +3212,7 @@ int __alloc_bootmem_huge_page(struct hstate *h, int nid)
 
 /* Initialize [start_page:end_page_number] tail struct pages of a hugepage */
 static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
+					struct hstate *h,
 					unsigned long start_page_number,
 					unsigned long end_page_number)
 {
@@ -3220,6 +3221,7 @@ static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
 	struct page *page = folio_page(folio, start_page_number);
 	unsigned long head_pfn = folio_pfn(folio);
 	unsigned long pfn, end_pfn = head_pfn + end_page_number;
+	unsigned int order = huge_page_order(h);
 
 	/*
 	 * As we marked all tail pages with memblock_reserved_mark_noinit(),
@@ -3227,7 +3229,7 @@ static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
 	 */
 	for (pfn = head_pfn + start_page_number; pfn < end_pfn; page++, pfn++) {
 		__init_single_page(page, pfn, zone, nid);
-		prep_compound_tail((struct page *)folio, pfn - head_pfn);
+		prep_compound_tail(page, &folio->page, order);
 		set_page_count(page, 0);
 	}
 }
@@ -3247,7 +3249,7 @@ static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
 	__folio_set_head(folio);
 	ret = folio_ref_freeze(folio, 1);
 	VM_BUG_ON(!ret);
-	hugetlb_folio_init_tail_vmemmap(folio, 1, nr_pages);
+	hugetlb_folio_init_tail_vmemmap(folio, h, 1, nr_pages);
 	prep_compound_head((struct page *)folio, huge_page_order(h));
 }
 
@@ -3304,7 +3306,7 @@ static void __init prep_and_add_bootmem_folios(struct hstate *h,
 			 * time as this is early in boot and there should
 			 * be no contention.
 			 */
-			hugetlb_folio_init_tail_vmemmap(folio,
+			hugetlb_folio_init_tail_vmemmap(folio, h,
 					HUGETLB_VMEMMAP_RESERVE_PAGES,
 					pages_per_huge_page(h));
 		}
diff --git a/mm/internal.h b/mm/internal.h
index 1561fc2ff5b8..f385370256b9 100644
--- a/mm/internal.h
+++ b/mm/internal.h
@@ -810,13 +810,13 @@ static inline void prep_compound_head(struct page *page, unsigned int order)
 		INIT_LIST_HEAD(&folio->_deferred_list);
 }
 
-static inline void prep_compound_tail(struct page *head, int tail_idx)
+static inline void prep_compound_tail(struct page *tail,
+				      const struct page *head,
+				      unsigned int order)
 {
-	struct page *p = head + tail_idx;
-
-	p->mapping = TAIL_MAPPING;
-	set_compound_head(p, head);
-	set_page_private(p, 0);
+	tail->mapping = TAIL_MAPPING;
+	set_compound_head(tail, head, order);
+	set_page_private(tail, 0);
 }
 
 void post_alloc_hook(struct page *page, unsigned int order, gfp_t gfp_flags);
diff --git a/mm/mm_init.c b/mm/mm_init.c
index 7712d887b696..87d1e0277318 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -1102,7 +1102,7 @@ static void __ref memmap_init_compound(struct page *head,
 		struct page *page = pfn_to_page(pfn);
 
 		__init_zone_device_page(page, pfn, zone_idx, nid, pgmap);
-		prep_compound_tail(head, pfn - head_pfn);
+		prep_compound_tail(page, head, order);
 		set_page_count(page, 0);
 	}
 	prep_compound_head(head, order);
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index ed82ee55e66a..fe77c00c99df 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -717,7 +717,7 @@ void prep_compound_page(struct page *page, unsigned int order)
 
 	__SetPageHead(page);
 	for (i = 1; i < nr_pages; i++)
-		prep_compound_tail(page, i);
+		prep_compound_tail(page + i, page, order);
 
 	prep_compound_head(page, order);
 }
-- 
2.51.2

[PATCHv3 04/15] mm: Rename the 'compound_head' field in the 'struct page' to 'compound_info'

[Kiryl Shutsemau]

The 'compound_head' field in the 'struct page' encodes whether the page
is a tail and where to locate the head page. Bit 0 is set if the page is
a tail, and the remaining bits in the field point to the head page.

As preparation for changing how the field encodes information about the
head page, rename the field to 'compound_info'.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 .../admin-guide/kdump/vmcoreinfo.rst          |  2 +-
 Documentation/mm/vmemmap_dedup.rst            |  6 +++---
 include/linux/mm_types.h                      | 20 +++++++++----------
 include/linux/page-flags.h                    | 18 ++++++++---------
 include/linux/types.h                         |  2 +-
 kernel/vmcore_info.c                          |  2 +-
 mm/page_alloc.c                               |  2 +-
 mm/slab.h                                     |  2 +-
 mm/util.c                                     |  2 +-
 9 files changed, 28 insertions(+), 28 deletions(-)

diff --git a/Documentation/admin-guide/kdump/vmcoreinfo.rst b/Documentation/admin-guide/kdump/vmcoreinfo.rst
index 404a15f6782c..7663c610fe90 100644
--- a/Documentation/admin-guide/kdump/vmcoreinfo.rst
+++ b/Documentation/admin-guide/kdump/vmcoreinfo.rst
@@ -141,7 +141,7 @@ nodemask_t
 The size of a nodemask_t type. Used to compute the number of online
 nodes.
 
-(page, flags|_refcount|mapping|lru|_mapcount|private|compound_order|compound_head)
+(page, flags|_refcount|mapping|lru|_mapcount|private|compound_order|compound_info)
 ----------------------------------------------------------------------------------
 
 User-space tools compute their values based on the offset of these
diff --git a/Documentation/mm/vmemmap_dedup.rst b/Documentation/mm/vmemmap_dedup.rst
index b4a55b6569fa..1863d88d2dcb 100644
--- a/Documentation/mm/vmemmap_dedup.rst
+++ b/Documentation/mm/vmemmap_dedup.rst
@@ -24,7 +24,7 @@ For each base page, there is a corresponding ``struct page``.
 Within the HugeTLB subsystem, only the first 4 ``struct page`` are used to
 contain unique information about a HugeTLB page. ``__NR_USED_SUBPAGE`` provides
 this upper limit. The only 'useful' information in the remaining ``struct page``
-is the compound_head field, and this field is the same for all tail pages.
+is the compound_info field, and this field is the same for all tail pages.
 
 By removing redundant ``struct page`` for HugeTLB pages, memory can be returned
 to the buddy allocator for other uses.
@@ -124,10 +124,10 @@ Here is how things look before optimization::
  |           |
  +-----------+
 
-The value of page->compound_head is the same for all tail pages. The first
+The value of page->compound_info is the same for all tail pages. The first
 page of ``struct page`` (page 0) associated with the HugeTLB page contains the 4
 ``struct page`` necessary to describe the HugeTLB. The only use of the remaining
-pages of ``struct page`` (page 1 to page 7) is to point to page->compound_head.
+pages of ``struct page`` (page 1 to page 7) is to point to page->compound_info.
 Therefore, we can remap pages 1 to 7 to page 0. Only 1 page of ``struct page``
 will be used for each HugeTLB page. This will allow us to free the remaining
 7 pages to the buddy allocator.
diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 90e5790c318f..a94683272869 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -125,14 +125,14 @@ struct page {
 			atomic_long_t pp_ref_count;
 		};
 		struct {	/* Tail pages of compound page */
-			unsigned long compound_head;	/* Bit zero is set */
+			unsigned long compound_info;	/* Bit zero is set */
 		};
 		struct {	/* ZONE_DEVICE pages */
 			/*
-			 * The first word is used for compound_head or folio
+			 * The first word is used for compound_info or folio
 			 * pgmap
 			 */
-			void *_unused_pgmap_compound_head;
+			void *_unused_pgmap_compound_info;
 			void *zone_device_data;
 			/*
 			 * ZONE_DEVICE private pages are counted as being
@@ -383,7 +383,7 @@ struct folio {
 	/* private: avoid cluttering the output */
 				/* For the Unevictable "LRU list" slot */
 				struct {
-					/* Avoid compound_head */
+					/* Avoid compound_info */
 					void *__filler;
 	/* public: */
 					unsigned int mlock_count;
@@ -484,7 +484,7 @@ struct folio {
 FOLIO_MATCH(flags, flags);
 FOLIO_MATCH(lru, lru);
 FOLIO_MATCH(mapping, mapping);
-FOLIO_MATCH(compound_head, lru);
+FOLIO_MATCH(compound_info, lru);
 FOLIO_MATCH(__folio_index, index);
 FOLIO_MATCH(private, private);
 FOLIO_MATCH(_mapcount, _mapcount);
@@ -503,7 +503,7 @@ FOLIO_MATCH(_last_cpupid, _last_cpupid);
 	static_assert(offsetof(struct folio, fl) ==			\
 			offsetof(struct page, pg) + sizeof(struct page))
 FOLIO_MATCH(flags, _flags_1);
-FOLIO_MATCH(compound_head, _head_1);
+FOLIO_MATCH(compound_info, _head_1);
 FOLIO_MATCH(_mapcount, _mapcount_1);
 FOLIO_MATCH(_refcount, _refcount_1);
 #undef FOLIO_MATCH
@@ -511,13 +511,13 @@ FOLIO_MATCH(_refcount, _refcount_1);
 	static_assert(offsetof(struct folio, fl) ==			\
 			offsetof(struct page, pg) + 2 * sizeof(struct page))
 FOLIO_MATCH(flags, _flags_2);
-FOLIO_MATCH(compound_head, _head_2);
+FOLIO_MATCH(compound_info, _head_2);
 #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);
+FOLIO_MATCH(compound_info, _head_3);
 #undef FOLIO_MATCH
 
 /**
@@ -583,8 +583,8 @@ struct ptdesc {
 #define TABLE_MATCH(pg, pt)						\
 	static_assert(offsetof(struct page, pg) == offsetof(struct ptdesc, pt))
 TABLE_MATCH(flags, pt_flags);
-TABLE_MATCH(compound_head, pt_list);
-TABLE_MATCH(compound_head, _pt_pad_1);
+TABLE_MATCH(compound_info, pt_list);
+TABLE_MATCH(compound_info, _pt_pad_1);
 TABLE_MATCH(mapping, __page_mapping);
 TABLE_MATCH(__folio_index, pt_index);
 TABLE_MATCH(rcu_head, pt_rcu_head);
diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index d4952573a4af..72c933a43b6a 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -213,7 +213,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 	/*
 	 * Only addresses aligned with PAGE_SIZE of struct page may be fake head
 	 * struct page. The alignment check aims to avoid access the fields (
-	 * e.g. compound_head) of the @page[1]. It can avoid touch a (possibly)
+	 * e.g. compound_info) of the @page[1]. It can avoid touch a (possibly)
 	 * cold cacheline in some cases.
 	 */
 	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
@@ -223,7 +223,7 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 		 * because the @page is a compound page composed with at least
 		 * two contiguous pages.
 		 */
-		unsigned long head = READ_ONCE(page[1].compound_head);
+		unsigned long head = READ_ONCE(page[1].compound_info);
 
 		if (likely(head & 1))
 			return (const struct page *)(head - 1);
@@ -281,7 +281,7 @@ static __always_inline int page_is_fake_head(const struct page *page)
 
 static __always_inline unsigned long _compound_head(const struct page *page)
 {
-	unsigned long head = READ_ONCE(page->compound_head);
+	unsigned long head = READ_ONCE(page->compound_info);
 
 	if (unlikely(head & 1))
 		return head - 1;
@@ -320,13 +320,13 @@ static __always_inline unsigned long _compound_head(const struct page *page)
 
 static __always_inline int PageTail(const struct page *page)
 {
-	return READ_ONCE(page->compound_head) & 1 || page_is_fake_head(page);
+	return READ_ONCE(page->compound_info) & 1 || page_is_fake_head(page);
 }
 
 static __always_inline int PageCompound(const struct page *page)
 {
 	return test_bit(PG_head, &page->flags.f) ||
-	       READ_ONCE(page->compound_head) & 1;
+	       READ_ONCE(page->compound_info) & 1;
 }
 
 #define	PAGE_POISON_PATTERN	-1l
@@ -348,7 +348,7 @@ static const unsigned long *const_folio_flags(const struct folio *folio,
 {
 	const struct page *page = &folio->page;
 
-	VM_BUG_ON_PGFLAGS(page->compound_head & 1, page);
+	VM_BUG_ON_PGFLAGS(page->compound_info & 1, page);
 	VM_BUG_ON_PGFLAGS(n > 0 && !test_bit(PG_head, &page->flags.f), page);
 	return &page[n].flags.f;
 }
@@ -357,7 +357,7 @@ static unsigned long *folio_flags(struct folio *folio, unsigned n)
 {
 	struct page *page = &folio->page;
 
-	VM_BUG_ON_PGFLAGS(page->compound_head & 1, page);
+	VM_BUG_ON_PGFLAGS(page->compound_info & 1, page);
 	VM_BUG_ON_PGFLAGS(n > 0 && !test_bit(PG_head, &page->flags.f), page);
 	return &page[n].flags.f;
 }
@@ -869,12 +869,12 @@ static __always_inline void set_compound_head(struct page *page,
 					      const struct page *head,
 					      unsigned int order)
 {
-	WRITE_ONCE(page->compound_head, (unsigned long)head + 1);
+	WRITE_ONCE(page->compound_info, (unsigned long)head + 1);
 }
 
 static __always_inline void clear_compound_head(struct page *page)
 {
-	WRITE_ONCE(page->compound_head, 0);
+	WRITE_ONCE(page->compound_info, 0);
 }
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
diff --git a/include/linux/types.h b/include/linux/types.h
index 6dfdb8e8e4c3..3a65f0ef4a73 100644
--- a/include/linux/types.h
+++ b/include/linux/types.h
@@ -234,7 +234,7 @@ struct ustat {
  *
  * This guarantee is important for few reasons:
  *  - future call_rcu_lazy() will make use of lower bits in the pointer;
- *  - the structure shares storage space in struct page with @compound_head,
+ *  - the structure shares storage space in struct page with @compound_info,
  *    which encode PageTail() in bit 0. The guarantee is needed to avoid
  *    false-positive PageTail().
  */
diff --git a/kernel/vmcore_info.c b/kernel/vmcore_info.c
index e066d31d08f8..782bc2050a40 100644
--- a/kernel/vmcore_info.c
+++ b/kernel/vmcore_info.c
@@ -175,7 +175,7 @@ static int __init crash_save_vmcoreinfo_init(void)
 	VMCOREINFO_OFFSET(page, lru);
 	VMCOREINFO_OFFSET(page, _mapcount);
 	VMCOREINFO_OFFSET(page, private);
-	VMCOREINFO_OFFSET(page, compound_head);
+	VMCOREINFO_OFFSET(page, compound_info);
 	VMCOREINFO_OFFSET(pglist_data, node_zones);
 	VMCOREINFO_OFFSET(pglist_data, nr_zones);
 #ifdef CONFIG_FLATMEM
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index fe77c00c99df..cecd6d89ff60 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -704,7 +704,7 @@ static inline bool pcp_allowed_order(unsigned int order)
  * The first PAGE_SIZE page is called the "head page" and have PG_head set.
  *
  * The remaining PAGE_SIZE pages are called "tail pages". PageTail() is encoded
- * in bit 0 of page->compound_head. The rest of bits is pointer to head page.
+ * in bit 0 of page->compound_info. The rest of bits is pointer to head page.
  *
  * The first tail page's ->compound_order holds the order of allocation.
  * This usage means that zero-order pages may not be compound.
diff --git a/mm/slab.h b/mm/slab.h
index 078daecc7cf5..b471877af296 100644
--- a/mm/slab.h
+++ b/mm/slab.h
@@ -104,7 +104,7 @@ struct slab {
 #define SLAB_MATCH(pg, sl)						\
 	static_assert(offsetof(struct page, pg) == offsetof(struct slab, sl))
 SLAB_MATCH(flags, flags);
-SLAB_MATCH(compound_head, slab_cache);	/* Ensure bit 0 is clear */
+SLAB_MATCH(compound_info, slab_cache);	/* Ensure bit 0 is clear */
 SLAB_MATCH(_refcount, __page_refcount);
 #ifdef CONFIG_MEMCG
 SLAB_MATCH(memcg_data, obj_exts);
diff --git a/mm/util.c b/mm/util.c
index 8989d5767528..cbf93cf3223a 100644
--- a/mm/util.c
+++ b/mm/util.c
@@ -1244,7 +1244,7 @@ void snapshot_page(struct page_snapshot *ps, const struct page *page)
 again:
 	memset(&ps->folio_snapshot, 0, sizeof(struct folio));
 	memcpy(&ps->page_snapshot, page, sizeof(*page));
-	head = ps->page_snapshot.compound_head;
+	head = ps->page_snapshot.compound_info;
 	if ((head & 1) == 0) {
 		ps->idx = 0;
 		foliop = (struct folio *)&ps->page_snapshot;
-- 
2.51.2

[PATCHv3 05/15] mm: Move set/clear_compound_head() next to compound_head()

[Kiryl Shutsemau]

Move set_compound_head() and clear_compound_head() to be adjacent to the
compound_head() function in page-flags.h.

These functions encode and decode the same compound_info field, so
keeping them together makes it easier to verify their logic is
consistent, especially when the encoding changes.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 include/linux/page-flags.h | 24 ++++++++++++------------
 1 file changed, 12 insertions(+), 12 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 72c933a43b6a..0de7db7efb00 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -290,6 +290,18 @@ static __always_inline unsigned long _compound_head(const struct page *page)
 
 #define compound_head(page)	((typeof(page))_compound_head(page))
 
+static __always_inline void set_compound_head(struct page *page,
+					      const struct page *head,
+					      unsigned int order)
+{
+	WRITE_ONCE(page->compound_info, (unsigned long)head + 1);
+}
+
+static __always_inline void clear_compound_head(struct page *page)
+{
+	WRITE_ONCE(page->compound_info, 0);
+}
+
 /**
  * page_folio - Converts from page to folio.
  * @p: The page.
@@ -865,18 +877,6 @@ static inline bool folio_test_large(const struct folio *folio)
 	return folio_test_head(folio);
 }
 
-static __always_inline void set_compound_head(struct page *page,
-					      const struct page *head,
-					      unsigned int order)
-{
-	WRITE_ONCE(page->compound_info, (unsigned long)head + 1);
-}
-
-static __always_inline void clear_compound_head(struct page *page)
-{
-	WRITE_ONCE(page->compound_info, 0);
-}
-
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 static inline void ClearPageCompound(struct page *page)
 {
-- 
2.51.2

[PATCHv3 06/15] mm: Rework compound_head() for power-of-2 sizeof(struct page)

[Kiryl Shutsemau]

For tail pages, the kernel uses the 'compound_info' field to get to the
head page. The bit 0 of the field indicates whether the page is a
tail page, and if set, the remaining bits represent a pointer to the
head page.

For cases when size of struct page is power-of-2, change the encoding of
compound_info to store a mask that can be applied to the virtual address
of the tail page in order to access the head page. It is possible
because struct page of the head page is naturally aligned with regards
to order of the page.

The significant impact of this modification is that all tail pages of
the same order will now have identical 'compound_info', regardless of
the compound page they are associated with. This paves the way for
eliminating fake heads.

The HugeTLB Vmemmap Optimization (HVO) creates fake heads and it is only
applied when the sizeof(struct page) is power-of-2. Having identical
tail pages allows the same page to be mapped into the vmemmap of all
pages, maintaining memory savings without fake heads.

If sizeof(struct page) is not power-of-2, there is no functional
changes.

Limit mask usage to SPARSEMEM_VMEMMAP where it makes a difference
because HVO. The approach with mask would work for any memory model,
but it requires validating that struct pages are naturally aligned for
all orders up to the MAX_FOLIO order, which can be tricky.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 include/linux/page-flags.h | 81 ++++++++++++++++++++++++++++++++++----
 mm/util.c                  | 16 ++++++--
 2 files changed, 85 insertions(+), 12 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 0de7db7efb00..e16a4bc82856 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -198,6 +198,29 @@ enum pageflags {
 
 #ifndef __GENERATING_BOUNDS_H
 
+/*
+ * For tail pages, if the size of struct page is power-of-2 ->compound_info
+ * encodes the mask that converts the address of the tail page address to
+ * the head page address.
+ *
+ * Otherwise, ->compound_info has direct pointer to head pages.
+ */
+static __always_inline bool compound_info_has_mask(void)
+{
+	/*
+	 * Limit mask usage to SPARSEMEM_VMEMMAP where it makes a difference
+	 * because of the HugeTLB vmemmap optimization (HVO).
+	 *
+	 * The approach with mask would work for any memory model, but it
+	 * requires validating that struct pages are naturally aligned for
+	 * all orders up to the MAX_FOLIO order, which can be tricky.
+	 */
+	if (!IS_ENABLED(CONFIG_SPARSEMEM_VMEMMAP))
+		return false;
+
+	return is_power_of_2(sizeof(struct page));
+}
+
 #ifdef CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
 DECLARE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 
@@ -210,6 +233,10 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 	if (!static_branch_unlikely(&hugetlb_optimize_vmemmap_key))
 		return page;
 
+	/* Fake heads only exists if compound_info_has_mask() is true */
+	if (!compound_info_has_mask())
+		return page;
+
 	/*
 	 * Only addresses aligned with PAGE_SIZE of struct page may be fake head
 	 * struct page. The alignment check aims to avoid access the fields (
@@ -223,10 +250,14 @@ static __always_inline const struct page *page_fixed_fake_head(const struct page
 		 * because the @page is a compound page composed with at least
 		 * two contiguous pages.
 		 */
-		unsigned long head = READ_ONCE(page[1].compound_info);
+		unsigned long info = READ_ONCE(page[1].compound_info);
 
-		if (likely(head & 1))
-			return (const struct page *)(head - 1);
+		/* See set_compound_head() */
+		if (likely(info & 1)) {
+			unsigned long p = (unsigned long)page;
+
+			return (const struct page *)(p & info);
+		}
 	}
 	return page;
 }
@@ -281,11 +312,26 @@ static __always_inline int page_is_fake_head(const struct page *page)
 
 static __always_inline unsigned long _compound_head(const struct page *page)
 {
-	unsigned long head = READ_ONCE(page->compound_info);
+	unsigned long info = READ_ONCE(page->compound_info);
 
-	if (unlikely(head & 1))
-		return head - 1;
-	return (unsigned long)page_fixed_fake_head(page);
+	/* Bit 0 encodes PageTail() */
+	if (!(info & 1))
+		return (unsigned long)page_fixed_fake_head(page);
+
+	/*
+	 * If compound_info_has_mask() is false, the rest of compound_info is
+	 * the pointer to the head page.
+	 */
+	if (!compound_info_has_mask())
+		return info - 1;
+
+	/*
+	 * If compoun_info_has_mask() is true the rest of the info encodes
+	 * the mask that converts the address of the tail page to the head page.
+	 *
+	 * No need to clear bit 0 in the mask as 'page' always has it clear.
+	 */
+	return (unsigned long)page & info;
 }
 
 #define compound_head(page)	((typeof(page))_compound_head(page))
@@ -294,7 +340,26 @@ static __always_inline void set_compound_head(struct page *page,
 					      const struct page *head,
 					      unsigned int order)
 {
-	WRITE_ONCE(page->compound_info, (unsigned long)head + 1);
+	unsigned int shift;
+	unsigned long mask;
+
+	if (!compound_info_has_mask()) {
+		WRITE_ONCE(page->compound_info, (unsigned long)head | 1);
+		return;
+	}
+
+	/*
+	 * If the size of struct page is power-of-2, bits [shift:0] of the
+	 * virtual address of compound head are zero.
+	 *
+	 * Calculate mask that can be applied to the virtual address of
+	 * the tail page to get address of the head page.
+	 */
+	shift = order + order_base_2(sizeof(struct page));
+	mask = GENMASK(BITS_PER_LONG - 1, shift);
+
+	/* Bit 0 encodes PageTail() */
+	WRITE_ONCE(page->compound_info, mask | 1);
 }
 
 static __always_inline void clear_compound_head(struct page *page)
diff --git a/mm/util.c b/mm/util.c
index cbf93cf3223a..f01a9655067f 100644
--- a/mm/util.c
+++ b/mm/util.c
@@ -1234,7 +1234,7 @@ static void set_ps_flags(struct page_snapshot *ps, const struct folio *folio,
  */
 void snapshot_page(struct page_snapshot *ps, const struct page *page)
 {
-	unsigned long head, nr_pages = 1;
+	unsigned long info, nr_pages = 1;
 	struct folio *foliop;
 	int loops = 5;
 
@@ -1244,8 +1244,8 @@ void snapshot_page(struct page_snapshot *ps, const struct page *page)
 again:
 	memset(&ps->folio_snapshot, 0, sizeof(struct folio));
 	memcpy(&ps->page_snapshot, page, sizeof(*page));
-	head = ps->page_snapshot.compound_info;
-	if ((head & 1) == 0) {
+	info = ps->page_snapshot.compound_info;
+	if ((info & 1) == 0) {
 		ps->idx = 0;
 		foliop = (struct folio *)&ps->page_snapshot;
 		if (!folio_test_large(foliop)) {
@@ -1256,7 +1256,15 @@ void snapshot_page(struct page_snapshot *ps, const struct page *page)
 		}
 		foliop = (struct folio *)page;
 	} else {
-		foliop = (struct folio *)(head - 1);
+		/* See compound_head() */
+		if (compound_info_has_mask()) {
+			unsigned long p = (unsigned long)page;
+
+			foliop = (struct folio *)(p & info);
+		} else {
+			foliop = (struct folio *)(info - 1);
+		}
+
 		ps->idx = folio_page_idx(foliop, page);
 	}
 
-- 
2.51.2

[PATCHv3 07/15] mm: Make page_zonenum() use head page

[Kiryl Shutsemau]

With the upcoming changes to HVO, a single page of tail struct pages
will be shared across all huge pages of the same order on a node. Since
huge pages on the same node may belong to different zones, the zone
information stored in shared tail page flags would be incorrect.

Always fetch zone information from the head page, which has unique and
correct zone flags for each compound page.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---
 include/linux/mmzone.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 6a2f3696068e..590d1a494c4c 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -1224,6 +1224,7 @@ static inline enum zone_type memdesc_zonenum(memdesc_flags_t flags)
 
 static inline enum zone_type page_zonenum(const struct page *page)
 {
+	page = compound_head(page);
 	return memdesc_zonenum(page->flags);
 }
 
-- 
2.51.2

[PATCHv3 08/15] mm/sparse: Check memmap alignment for compound_info_has_mask()

[Kiryl Shutsemau]

If page->compound_info encodes a mask, it is expected that memmap to be
naturally aligned to the maximum folio size.

Add a warning if it is not.

A warning is sufficient as MAX_FOLIO_ORDER is very rarely used, so the
kernel is still likely to be functional if this strict check fails.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---
 include/linux/mmzone.h | 1 +
 mm/sparse.c            | 5 +++++
 2 files changed, 6 insertions(+)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 590d1a494c4c..322ed4c42cfc 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -92,6 +92,7 @@
 #endif
 
 #define MAX_FOLIO_NR_PAGES	(1UL << MAX_FOLIO_ORDER)
+#define MAX_FOLIO_SIZE		(PAGE_SIZE << MAX_FOLIO_ORDER)
 
 enum migratetype {
 	MIGRATE_UNMOVABLE,
diff --git a/mm/sparse.c b/mm/sparse.c
index 17c50a6415c2..5f41a3edcc24 100644
--- a/mm/sparse.c
+++ b/mm/sparse.c
@@ -600,6 +600,11 @@ void __init sparse_init(void)
 	BUILD_BUG_ON(!is_power_of_2(sizeof(struct mem_section)));
 	memblocks_present();
 
+	if (compound_info_has_mask()) {
+		WARN_ON(!IS_ALIGNED((unsigned long)pfn_to_page(0),
+				    MAX_FOLIO_SIZE / sizeof(struct page)));
+	}
+
 	pnum_begin = first_present_section_nr();
 	nid_begin = sparse_early_nid(__nr_to_section(pnum_begin));
 
-- 
2.51.2

[PATCHv3 09/15] mm/hugetlb: Refactor code around vmemmap_walk

[Kiryl Shutsemau]

To prepare for removing fake head pages, the vmemmap_walk code is being reworked.

The reuse_page and reuse_addr variables are being eliminated. There will
no longer be an expectation regarding the reuse address in relation to
the operated range. Instead, the caller will provide head and tail
vmemmap pages, along with the vmemmap_start address where the head page
is located.

Currently, vmemmap_head and vmemmap_tail are set to the same page, but
this will change in the future.

The only functional change is that __hugetlb_vmemmap_optimize_folio()
will abandon optimization if memory allocation fails.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---
 mm/hugetlb_vmemmap.c | 198 ++++++++++++++++++-------------------------
 1 file changed, 83 insertions(+), 115 deletions(-)

diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index ba0fb1b6a5a8..2b19c2205091 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -24,8 +24,9 @@
  *
  * @remap_pte:		called for each lowest-level entry (PTE).
  * @nr_walked:		the number of walked pte.
- * @reuse_page:		the page which is reused for the tail vmemmap pages.
- * @reuse_addr:		the virtual address of the @reuse_page page.
+ * @vmemmap_start:	the start of vmemmap range, where head page is located
+ * @vmemmap_head:	the page to be installed as first in the vmemmap range
+ * @vmemmap_tail:	the page to be installed as non-first in the vmemmap range
  * @vmemmap_pages:	the list head of the vmemmap pages that can be freed
  *			or is mapped from.
  * @flags:		used to modify behavior in vmemmap page table walking
@@ -34,11 +35,14 @@
 struct vmemmap_remap_walk {
 	void			(*remap_pte)(pte_t *pte, unsigned long addr,
 					     struct vmemmap_remap_walk *walk);
+
 	unsigned long		nr_walked;
-	struct page		*reuse_page;
-	unsigned long		reuse_addr;
+	unsigned long		vmemmap_start;
+	struct page		*vmemmap_head;
+	struct page		*vmemmap_tail;
 	struct list_head	*vmemmap_pages;
 
+
 /* Skip the TLB flush when we split the PMD */
 #define VMEMMAP_SPLIT_NO_TLB_FLUSH	BIT(0)
 /* Skip the TLB flush when we remap the PTE */
@@ -140,14 +144,7 @@ static int vmemmap_pte_entry(pte_t *pte, unsigned long addr,
 {
 	struct vmemmap_remap_walk *vmemmap_walk = walk->private;
 
-	/*
-	 * The reuse_page is found 'first' in page table walking before
-	 * starting remapping.
-	 */
-	if (!vmemmap_walk->reuse_page)
-		vmemmap_walk->reuse_page = pte_page(ptep_get(pte));
-	else
-		vmemmap_walk->remap_pte(pte, addr, vmemmap_walk);
+	vmemmap_walk->remap_pte(pte, addr, vmemmap_walk);
 	vmemmap_walk->nr_walked++;
 
 	return 0;
@@ -207,18 +204,12 @@ static void free_vmemmap_page_list(struct list_head *list)
 static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 			      struct vmemmap_remap_walk *walk)
 {
-	/*
-	 * Remap the tail pages as read-only to catch illegal write operation
-	 * to the tail pages.
-	 */
-	pgprot_t pgprot = PAGE_KERNEL_RO;
 	struct page *page = pte_page(ptep_get(pte));
 	pte_t entry;
 
 	/* Remapping the head page requires r/w */
-	if (unlikely(addr == walk->reuse_addr)) {
-		pgprot = PAGE_KERNEL;
-		list_del(&walk->reuse_page->lru);
+	if (unlikely(addr == walk->vmemmap_start)) {
+		list_del(&walk->vmemmap_head->lru);
 
 		/*
 		 * Makes sure that preceding stores to the page contents from
@@ -226,9 +217,16 @@ static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
 		 * write.
 		 */
 		smp_wmb();
+
+		entry = mk_pte(walk->vmemmap_head, PAGE_KERNEL);
+	} else {
+		/*
+		 * Remap the tail pages as read-only to catch illegal write
+		 * operation to the tail pages.
+		 */
+		entry = mk_pte(walk->vmemmap_tail, PAGE_KERNEL_RO);
 	}
 
-	entry = mk_pte(walk->reuse_page, pgprot);
 	list_add(&page->lru, walk->vmemmap_pages);
 	set_pte_at(&init_mm, addr, pte, entry);
 }
@@ -255,16 +253,13 @@ static inline void reset_struct_pages(struct page *start)
 static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 				struct vmemmap_remap_walk *walk)
 {
-	pgprot_t pgprot = PAGE_KERNEL;
 	struct page *page;
 	void *to;
 
-	BUG_ON(pte_page(ptep_get(pte)) != walk->reuse_page);
-
 	page = list_first_entry(walk->vmemmap_pages, struct page, lru);
 	list_del(&page->lru);
 	to = page_to_virt(page);
-	copy_page(to, (void *)walk->reuse_addr);
+	copy_page(to, (void *)walk->vmemmap_start);
 	reset_struct_pages(to);
 
 	/*
@@ -272,7 +267,7 @@ static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
 	 * before the set_pte_at() write.
 	 */
 	smp_wmb();
-	set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
+	set_pte_at(&init_mm, addr, pte, mk_pte(page, PAGE_KERNEL));
 }
 
 /**
@@ -282,33 +277,29 @@ static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
  *             to remap.
  * @end:       end address of the vmemmap virtual address range that we want to
  *             remap.
- * @reuse:     reuse address.
- *
  * Return: %0 on success, negative error code otherwise.
  */
-static int vmemmap_remap_split(unsigned long start, unsigned long end,
-			       unsigned long reuse)
+static int vmemmap_remap_split(unsigned long start, unsigned long end)
 {
 	struct vmemmap_remap_walk walk = {
 		.remap_pte	= NULL,
+		.vmemmap_start	= start,
 		.flags		= VMEMMAP_SPLIT_NO_TLB_FLUSH,
 	};
 
-	/* See the comment in the vmemmap_remap_free(). */
-	BUG_ON(start - reuse != PAGE_SIZE);
-
-	return vmemmap_remap_range(reuse, end, &walk);
+	return vmemmap_remap_range(start, end, &walk);
 }
 
 /**
  * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
- *			to the page which @reuse is mapped to, then free vmemmap
- *			which the range are mapped to.
+ *			to use @vmemmap_head/tail, then free vmemmap which
+ *			the range are mapped to.
  * @start:	start address of the vmemmap virtual address range that we want
  *		to remap.
  * @end:	end address of the vmemmap virtual address range that we want to
  *		remap.
- * @reuse:	reuse address.
+ * @vmemmap_head: the page to be installed as first in the vmemmap range
+ * @vmemmap_tail: the page to be installed as non-first in the vmemmap range
  * @vmemmap_pages: list to deposit vmemmap pages to be freed.  It is callers
  *		responsibility to free pages.
  * @flags:	modifications to vmemmap_remap_walk flags
@@ -316,69 +307,40 @@ static int vmemmap_remap_split(unsigned long start, unsigned long end,
  * Return: %0 on success, negative error code otherwise.
  */
 static int vmemmap_remap_free(unsigned long start, unsigned long end,
-			      unsigned long reuse,
+			      struct page *vmemmap_head,
+			      struct page *vmemmap_tail,
 			      struct list_head *vmemmap_pages,
 			      unsigned long flags)
 {
 	int ret;
 	struct vmemmap_remap_walk walk = {
 		.remap_pte	= vmemmap_remap_pte,
-		.reuse_addr	= reuse,
+		.vmemmap_start	= start,
+		.vmemmap_head	= vmemmap_head,
+		.vmemmap_tail	= vmemmap_tail,
 		.vmemmap_pages	= vmemmap_pages,
 		.flags		= flags,
 	};
-	int nid = page_to_nid((struct page *)reuse);
-	gfp_t gfp_mask = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
+
+	ret = vmemmap_remap_range(start, end, &walk);
+	if (!ret || !walk.nr_walked)
+		return ret;
+
+	end = start + walk.nr_walked * PAGE_SIZE;
 
 	/*
-	 * Allocate a new head vmemmap page to avoid breaking a contiguous
-	 * block of struct page memory when freeing it back to page allocator
-	 * in free_vmemmap_page_list(). This will allow the likely contiguous
-	 * struct page backing memory to be kept contiguous and allowing for
-	 * more allocations of hugepages. Fallback to the currently
-	 * mapped head page in case should it fail to allocate.
+	 * vmemmap_pages contains pages from the previous vmemmap_remap_range()
+	 * call which failed.  These are pages which were removed from
+	 * the vmemmap. They will be restored in the following call.
 	 */
-	walk.reuse_page = alloc_pages_node(nid, gfp_mask, 0);
-	if (walk.reuse_page) {
-		copy_page(page_to_virt(walk.reuse_page),
-			  (void *)walk.reuse_addr);
-		list_add(&walk.reuse_page->lru, vmemmap_pages);
-		memmap_pages_add(1);
-	}
+	walk = (struct vmemmap_remap_walk) {
+		.remap_pte	= vmemmap_restore_pte,
+		.vmemmap_start	= start,
+		.vmemmap_pages	= vmemmap_pages,
+		.flags		= 0,
+	};
 
-	/*
-	 * In order to make remapping routine most efficient for the huge pages,
-	 * the routine of vmemmap page table walking has the following rules
-	 * (see more details from the vmemmap_pte_range()):
-	 *
-	 * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
-	 *   should be continuous.
-	 * - The @reuse address is part of the range [@reuse, @end) that we are
-	 *   walking which is passed to vmemmap_remap_range().
-	 * - The @reuse address is the first in the complete range.
-	 *
-	 * So we need to make sure that @start and @reuse meet the above rules.
-	 */
-	BUG_ON(start - reuse != PAGE_SIZE);
-
-	ret = vmemmap_remap_range(reuse, end, &walk);
-	if (ret && walk.nr_walked) {
-		end = reuse + walk.nr_walked * PAGE_SIZE;
-		/*
-		 * vmemmap_pages contains pages from the previous
-		 * vmemmap_remap_range call which failed.  These
-		 * are pages which were removed from the vmemmap.
-		 * They will be restored in the following call.
-		 */
-		walk = (struct vmemmap_remap_walk) {
-			.remap_pte	= vmemmap_restore_pte,
-			.reuse_addr	= reuse,
-			.vmemmap_pages	= vmemmap_pages,
-			.flags		= 0,
-		};
-
-		vmemmap_remap_range(reuse, end, &walk);
-	}
+	vmemmap_remap_range(start, end, &walk);
 
 	return ret;
 }
@@ -415,29 +377,27 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
  *		to remap.
  * @end:	end address of the vmemmap virtual address range that we want to
  *		remap.
- * @reuse:	reuse address.
  * @flags:	modifications to vmemmap_remap_walk flags
  *
  * Return: %0 on success, negative error code otherwise.
  */
 static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
-			       unsigned long reuse, unsigned long flags)
+			       unsigned long flags)
 {
 	LIST_HEAD(vmemmap_pages);
 	struct vmemmap_remap_walk walk = {
 		.remap_pte	= vmemmap_restore_pte,
-		.reuse_addr	= reuse,
+		.vmemmap_start	= start,
 		.vmemmap_pages	= &vmemmap_pages,
 		.flags		= flags,
 	};
 
-	/* See the comment in the vmemmap_remap_free(). */
-	BUG_ON(start - reuse != PAGE_SIZE);
+	start += HUGETLB_VMEMMAP_RESERVE_SIZE;
 
 	if (alloc_vmemmap_page_list(start, end, &vmemmap_pages))
 		return -ENOMEM;
 
-	return vmemmap_remap_range(reuse, end, &walk);
+	return vmemmap_remap_range(start, end, &walk);
 }
 
 DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
@@ -454,8 +414,7 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
 					   struct folio *folio, unsigned long flags)
 {
 	int ret;
-	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
-	unsigned long vmemmap_reuse;
+	unsigned long vmemmap_start, vmemmap_end;
 
 	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio), folio);
 	VM_WARN_ON_ONCE_FOLIO(folio_ref_count(folio), folio);
@@ -466,18 +425,16 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
 	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
 		synchronize_rcu();
 
+	vmemmap_start	= (unsigned long)folio;
 	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
-	vmemmap_reuse	= vmemmap_start;
-	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
 
 	/*
 	 * The pages which the vmemmap virtual address range [@vmemmap_start,
-	 * @vmemmap_end) are mapped to are freed to the buddy allocator, and
-	 * the range is mapped to the page which @vmemmap_reuse is mapped to.
+	 * @vmemmap_end) are mapped to are freed to the buddy allocator.
 	 * When a HugeTLB page is freed to the buddy allocator, previously
 	 * discarded vmemmap pages must be allocated and remapping.
 	 */
-	ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse, flags);
+	ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, flags);
 	if (!ret) {
 		folio_clear_hugetlb_vmemmap_optimized(folio);
 		static_branch_dec(&hugetlb_optimize_vmemmap_key);
@@ -565,9 +522,9 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 					    struct list_head *vmemmap_pages,
 					    unsigned long flags)
 {
-	int ret = 0;
-	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
-	unsigned long vmemmap_reuse;
+	unsigned long vmemmap_start, vmemmap_end;
+	struct page *vmemmap_head, *vmemmap_tail;
+	int nid, ret = 0;
 
 	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio), folio);
 	VM_WARN_ON_ONCE_FOLIO(folio_ref_count(folio), folio);
@@ -592,18 +549,31 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	 */
 	folio_set_hugetlb_vmemmap_optimized(folio);
 
+	nid = folio_nid(folio);
+	vmemmap_head = alloc_pages_node(nid, GFP_KERNEL, 0);
+
+	if (!vmemmap_head) {
+		ret = -ENOMEM;
+		goto out;
+	}
+
+	copy_page(page_to_virt(vmemmap_head), folio);
+	list_add(&vmemmap_head->lru, vmemmap_pages);
+	memmap_pages_add(1);
+
+	vmemmap_tail	= vmemmap_head;
+	vmemmap_start	= (unsigned long)folio;
 	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
-	vmemmap_reuse	= vmemmap_start;
-	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
 
 	/*
-	 * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
-	 * to the page which @vmemmap_reuse is mapped to.  Add pages previously
-	 * mapping the range to vmemmap_pages list so that they can be freed by
-	 * the caller.
+	 * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end).
+	 * Add pages previously mapping the range to vmemmap_pages list so that
+	 * they can be freed by the caller.
 	 */
-	ret = vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse,
+	ret = vmemmap_remap_free(vmemmap_start, vmemmap_end,
+				 vmemmap_head, vmemmap_tail,
 				 vmemmap_pages, flags);
+out:
 	if (ret) {
 		static_branch_dec(&hugetlb_optimize_vmemmap_key);
 		folio_clear_hugetlb_vmemmap_optimized(folio);
@@ -632,21 +602,19 @@ void hugetlb_vmemmap_optimize_folio(const struct hstate *h, struct folio *folio)
 
 static int hugetlb_vmemmap_split_folio(const struct hstate *h, struct folio *folio)
 {
-	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
-	unsigned long vmemmap_reuse;
+	unsigned long vmemmap_start, vmemmap_end;
 
 	if (!vmemmap_should_optimize_folio(h, folio))
 		return 0;
 
+	vmemmap_start	= (unsigned long)folio;
 	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
-	vmemmap_reuse	= vmemmap_start;
-	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
 
 	/*
 	 * Split PMDs on the vmemmap virtual address range [@vmemmap_start,
 	 * @vmemmap_end]
 	 */
-	return vmemmap_remap_split(vmemmap_start, vmemmap_end, vmemmap_reuse);
+	return vmemmap_remap_split(vmemmap_start, vmemmap_end);
 }
 
 static void __hugetlb_vmemmap_optimize_folios(struct hstate *h,
-- 
2.51.2

[PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau]

HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
vmemmap pages for huge pages and remapping the freed range to a single
page containing the struct page metadata.

With the new mask-based compound_info encoding (for power-of-2 struct
page sizes), all tail pages of the same order are now identical
regardless of which compound page they belong to. This means the tail
pages can be truly shared without fake heads.

Allocate a single page of initialized tail struct pages per NUMA node
per order in the vmemmap_tails[] array in pglist_data. All huge pages
of that order on the node share this tail page, mapped read-only into
their vmemmap. The head page remains unique per huge page.

This eliminates fake heads while maintaining the same memory savings,
and simplifies compound_head() by removing fake head detection.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---
 include/linux/mmzone.h | 16 ++++++++++++++-
 mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
 mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
 3 files changed, 93 insertions(+), 11 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 322ed4c42cfc..2ee3eb610291 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -82,7 +82,11 @@
  * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
  * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
  */
-#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
+#ifdef CONFIG_64BIT
+#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
+#else
+#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
+#endif
 #else
 /*
  * Without hugetlb, gigantic folios that are bigger than a single PUD are
@@ -1408,6 +1412,13 @@ struct memory_failure_stats {
 };
 #endif
 
+/*
+ * vmemmap optimization (like HVO) is only possible for page orders that fill
+ * two or more pages with struct pages.
+ */
+#define VMEMMAP_TAIL_MIN_ORDER (ilog2(2 * PAGE_SIZE / sizeof(struct page)))
+#define NR_VMEMMAP_TAILS (MAX_FOLIO_ORDER - VMEMMAP_TAIL_MIN_ORDER + 1)
+
 /*
  * On NUMA machines, each NUMA node would have a pg_data_t to describe
  * it's memory layout. On UMA machines there is a single pglist_data which
@@ -1556,6 +1567,9 @@ typedef struct pglist_data {
 #ifdef CONFIG_MEMORY_FAILURE
 	struct memory_failure_stats mf_stats;
 #endif
+#ifdef CONFIG_SPARSEMEM_VMEMMAP
+	unsigned long vmemmap_tails[NR_VMEMMAP_TAILS];
+#endif
 } pg_data_t;
 
 #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index 2b19c2205091..cbdca4684db1 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -18,6 +18,7 @@
 #include <asm/pgalloc.h>
 #include <asm/tlbflush.h>
 #include "hugetlb_vmemmap.h"
+#include "internal.h"
 
 /**
  * struct vmemmap_remap_walk - walk vmemmap page table
@@ -517,6 +518,41 @@ static bool vmemmap_should_optimize_folio(const struct hstate *h, struct folio *
 	return true;
 }
 
+static struct page *vmemmap_get_tail(unsigned int order, int node)
+{
+	unsigned long pfn;
+	unsigned int idx;
+	struct page *tail, *p;
+
+	idx = order - VMEMMAP_TAIL_MIN_ORDER;
+	pfn =  NODE_DATA(node)->vmemmap_tails[idx];
+	if (pfn)
+		return pfn_to_page(pfn);
+
+	tail = alloc_pages_node(node, GFP_KERNEL, 0);
+	if (!tail)
+		return NULL;
+
+	p = page_to_virt(tail);
+	for (int i = 0; i < PAGE_SIZE / sizeof(struct page); i++)
+		prep_compound_tail(p + i, NULL, order);
+
+	spin_lock(&hugetlb_lock);
+	if (!NODE_DATA(node)->vmemmap_tails[idx]) {
+		pfn = PHYS_PFN(virt_to_phys(p));
+		NODE_DATA(node)->vmemmap_tails[idx] = pfn;
+		tail = NULL;
+	} else {
+		pfn = NODE_DATA(node)->vmemmap_tails[idx];
+	}
+	spin_unlock(&hugetlb_lock);
+
+	if (tail)
+		__free_page(tail);
+
+	return pfn_to_page(pfn);
+}
+
 static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 					    struct folio *folio,
 					    struct list_head *vmemmap_pages,
@@ -532,6 +568,12 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	if (!vmemmap_should_optimize_folio(h, folio))
 		return ret;
 
+	nid = folio_nid(folio);
+
+	vmemmap_tail = vmemmap_get_tail(h->order, nid);
+	if (!vmemmap_tail)
+		return -ENOMEM;
+
 	static_branch_inc(&hugetlb_optimize_vmemmap_key);
 
 	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
@@ -549,7 +591,6 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	 */
 	folio_set_hugetlb_vmemmap_optimized(folio);
 
-	nid = folio_nid(folio);
 	vmemmap_head = alloc_pages_node(nid, GFP_KERNEL, 0);
 
 	if (!vmemmap_head) {
@@ -561,7 +602,6 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	list_add(&vmemmap_head->lru, vmemmap_pages);
 	memmap_pages_add(1);
 
-	vmemmap_tail	= vmemmap_head;
 	vmemmap_start	= (unsigned long)folio;
 	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
 
diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c
index dbd8daccade2..94b4e90fa00f 100644
--- a/mm/sparse-vmemmap.c
+++ b/mm/sparse-vmemmap.c
@@ -378,16 +378,45 @@ void vmemmap_wrprotect_hvo(unsigned long addr, unsigned long end,
 	}
 }
 
-/*
- * Populate vmemmap pages HVO-style. The first page contains the head
- * page and needed tail pages, the other ones are mirrors of the first
- * page.
- */
+static __meminit unsigned long vmemmap_get_tail(unsigned int order, int node)
+{
+	unsigned long pfn;
+	unsigned int idx;
+	struct page *p;
+
+	BUG_ON(order < VMEMMAP_TAIL_MIN_ORDER);
+	BUG_ON(order > MAX_FOLIO_ORDER);
+
+	idx = order - VMEMMAP_TAIL_MIN_ORDER;
+	pfn =  NODE_DATA(node)->vmemmap_tails[idx];
+	if (pfn)
+		return pfn;
+
+	p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
+	if (!p)
+		return 0;
+
+	for (int i = 0; i < PAGE_SIZE / sizeof(struct page); i++)
+		prep_compound_tail(p + i, NULL, order);
+
+	pfn = PHYS_PFN(virt_to_phys(p));
+	NODE_DATA(node)->vmemmap_tails[idx] = pfn;
+
+	return pfn;
+}
+
 int __meminit vmemmap_populate_hvo(unsigned long addr, unsigned long end,
 				       int node, unsigned long headsize)
 {
+	unsigned long maddr, len, tail_pfn;
+	unsigned int order;
 	pte_t *pte;
-	unsigned long maddr;
+
+	len = end - addr;
+	order = ilog2(len * sizeof(struct page) / PAGE_SIZE);
+	tail_pfn = vmemmap_get_tail(order, node);
+	if (!tail_pfn)
+		return -ENOMEM;
 
 	for (maddr = addr; maddr < addr + headsize; maddr += PAGE_SIZE) {
 		pte = vmemmap_populate_address(maddr, node, NULL, -1, 0);
@@ -398,8 +427,7 @@ int __meminit vmemmap_populate_hvo(unsigned long addr, unsigned long end,
 	/*
 	 * Reuse the last page struct page mapped above for the rest.
 	 */
-	return vmemmap_populate_range(maddr, end, node, NULL,
-					pte_pfn(ptep_get(pte)), 0);
+	return vmemmap_populate_range(maddr, end, node, NULL, tail_pfn, 0);
 }
 
 void __weak __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
-- 
2.51.2

[PATCHv3 11/15] mm: Drop fake head checks

[Kiryl Shutsemau]

With fake head pages eliminated in the previous commit, remove the
supporting infrastructure:

  - page_fixed_fake_head(): no longer needed to detect fake heads;
  - page_is_fake_head(): no longer needed;
  - page_count_writable(): no longer needed for RCU protection;
  - RCU read_lock in page_ref_add_unless(): no longer needed;

This substantially simplifies compound_head() and page_ref_add_unless(),
removing both branches and RCU overhead from these hot paths.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 include/linux/page-flags.h | 93 ++------------------------------------
 include/linux/page_ref.h   |  8 +---
 2 files changed, 4 insertions(+), 97 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index e16a4bc82856..660f9154a211 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -221,102 +221,15 @@ static __always_inline bool compound_info_has_mask(void)
 	return is_power_of_2(sizeof(struct page));
 }
 
-#ifdef CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
 DECLARE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
 
-/*
- * Return the real head page struct iff the @page is a fake head page, otherwise
- * return the @page itself. See Documentation/mm/vmemmap_dedup.rst.
- */
-static __always_inline const struct page *page_fixed_fake_head(const struct page *page)
-{
-	if (!static_branch_unlikely(&hugetlb_optimize_vmemmap_key))
-		return page;
-
-	/* Fake heads only exists if compound_info_has_mask() is true */
-	if (!compound_info_has_mask())
-		return page;
-
-	/*
-	 * Only addresses aligned with PAGE_SIZE of struct page may be fake head
-	 * struct page. The alignment check aims to avoid access the fields (
-	 * e.g. compound_info) of the @page[1]. It can avoid touch a (possibly)
-	 * cold cacheline in some cases.
-	 */
-	if (IS_ALIGNED((unsigned long)page, PAGE_SIZE) &&
-	    test_bit(PG_head, &page->flags.f)) {
-		/*
-		 * We can safely access the field of the @page[1] with PG_head
-		 * because the @page is a compound page composed with at least
-		 * two contiguous pages.
-		 */
-		unsigned long info = READ_ONCE(page[1].compound_info);
-
-		/* See set_compound_head() */
-		if (likely(info & 1)) {
-			unsigned long p = (unsigned long)page;
-
-			return (const struct page *)(p & info);
-		}
-	}
-	return page;
-}
-
-static __always_inline bool page_count_writable(const struct page *page, int u)
-{
-	if (!static_branch_unlikely(&hugetlb_optimize_vmemmap_key))
-		return true;
-
-	/*
-	 * The refcount check is ordered before the fake-head check to prevent
-	 * the following race:
-	 *   CPU 1 (HVO)                     CPU 2 (speculative PFN walker)
-	 *
-	 *   page_ref_freeze()
-	 *   synchronize_rcu()
-	 *                                   rcu_read_lock()
-	 *                                   page_is_fake_head() is false
-	 *   vmemmap_remap_pte()
-	 *   XXX: struct page[] becomes r/o
-	 *
-	 *   page_ref_unfreeze()
-	 *                                   page_ref_count() is not zero
-	 *
-	 *                                   atomic_add_unless(&page->_refcount)
-	 *                                   XXX: try to modify r/o struct page[]
-	 *
-	 * The refcount check also prevents modification attempts to other (r/o)
-	 * tail pages that are not fake heads.
-	 */
-	if (atomic_read_acquire(&page->_refcount) == u)
-		return false;
-
-	return page_fixed_fake_head(page) == page;
-}
-#else
-static inline const struct page *page_fixed_fake_head(const struct page *page)
-{
-	return page;
-}
-
-static inline bool page_count_writable(const struct page *page, int u)
-{
-	return true;
-}
-#endif
-
-static __always_inline int page_is_fake_head(const struct page *page)
-{
-	return page_fixed_fake_head(page) != page;
-}
-
 static __always_inline unsigned long _compound_head(const struct page *page)
 {
 	unsigned long info = READ_ONCE(page->compound_info);
 
 	/* Bit 0 encodes PageTail() */
 	if (!(info & 1))
-		return (unsigned long)page_fixed_fake_head(page);
+		return (unsigned long)page;
 
 	/*
 	 * If compound_info_has_mask() is false, the rest of compound_info is
@@ -397,7 +310,7 @@ static __always_inline void clear_compound_head(struct page *page)
 
 static __always_inline int PageTail(const struct page *page)
 {
-	return READ_ONCE(page->compound_info) & 1 || page_is_fake_head(page);
+	return READ_ONCE(page->compound_info) & 1;
 }
 
 static __always_inline int PageCompound(const struct page *page)
@@ -924,7 +837,7 @@ static __always_inline bool folio_test_head(const struct folio *folio)
 static __always_inline int PageHead(const struct page *page)
 {
 	PF_POISONED_CHECK(page);
-	return test_bit(PG_head, &page->flags.f) && !page_is_fake_head(page);
+	return test_bit(PG_head, &page->flags.f);
 }
 
 __SETPAGEFLAG(Head, head, PF_ANY)
diff --git a/include/linux/page_ref.h b/include/linux/page_ref.h
index 544150d1d5fd..490d0ad6e56d 100644
--- a/include/linux/page_ref.h
+++ b/include/linux/page_ref.h
@@ -230,13 +230,7 @@ static inline int folio_ref_dec_return(struct folio *folio)
 
 static inline bool page_ref_add_unless(struct page *page, int nr, int u)
 {
-	bool ret = false;
-
-	rcu_read_lock();
-	/* avoid writing to the vmemmap area being remapped */
-	if (page_count_writable(page, u))
-		ret = atomic_add_unless(&page->_refcount, nr, u);
-	rcu_read_unlock();
+	bool ret = atomic_add_unless(&page->_refcount, nr, u);
 
 	if (page_ref_tracepoint_active(page_ref_mod_unless))
 		__page_ref_mod_unless(page, nr, ret);
-- 
2.51.2

[PATCHv3 12/15] hugetlb: Remove VMEMMAP_SYNCHRONIZE_RCU

[Kiryl Shutsemau]

The VMEMMAP_SYNCHRONIZE_RCU flag triggered synchronize_rcu() calls to
prevent a race between HVO remapping and page_ref_add_unless(). The
race could occur when a speculative PFN walker tried to modify the
refcount on a struct page that was in the process of being remapped
to a fake head.

With fake heads eliminated, page_ref_add_unless() no longer needs RCU
protection.

Remove the flag and synchronize_rcu() calls.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 mm/hugetlb_vmemmap.c | 20 ++++----------------
 1 file changed, 4 insertions(+), 16 deletions(-)

diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index cbdca4684db1..5a802e292bac 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -48,8 +48,6 @@ struct vmemmap_remap_walk {
 #define VMEMMAP_SPLIT_NO_TLB_FLUSH	BIT(0)
 /* Skip the TLB flush when we remap the PTE */
 #define VMEMMAP_REMAP_NO_TLB_FLUSH	BIT(1)
-/* synchronize_rcu() to avoid writes from page_ref_add_unless() */
-#define VMEMMAP_SYNCHRONIZE_RCU		BIT(2)
 	unsigned long		flags;
 };
 
@@ -423,9 +421,6 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
 	if (!folio_test_hugetlb_vmemmap_optimized(folio))
 		return 0;
 
-	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
-		synchronize_rcu();
-
 	vmemmap_start	= (unsigned long)folio;
 	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
 
@@ -456,7 +451,7 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
  */
 int hugetlb_vmemmap_restore_folio(const struct hstate *h, struct folio *folio)
 {
-	return __hugetlb_vmemmap_restore_folio(h, folio, VMEMMAP_SYNCHRONIZE_RCU);
+	return __hugetlb_vmemmap_restore_folio(h, folio, 0);
 }
 
 /**
@@ -479,14 +474,11 @@ long hugetlb_vmemmap_restore_folios(const struct hstate *h,
 	struct folio *folio, *t_folio;
 	long restored = 0;
 	long ret = 0;
-	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH | VMEMMAP_SYNCHRONIZE_RCU;
+	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH;
 
 	list_for_each_entry_safe(folio, t_folio, folio_list, lru) {
 		if (folio_test_hugetlb_vmemmap_optimized(folio)) {
 			ret = __hugetlb_vmemmap_restore_folio(h, folio, flags);
-			/* only need to synchronize_rcu() once for each batch */
-			flags &= ~VMEMMAP_SYNCHRONIZE_RCU;
-
 			if (ret)
 				break;
 			restored++;
@@ -576,8 +568,6 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 
 	static_branch_inc(&hugetlb_optimize_vmemmap_key);
 
-	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
-		synchronize_rcu();
 	/*
 	 * Very Subtle
 	 * If VMEMMAP_REMAP_NO_TLB_FLUSH is set, TLB flushing is not performed
@@ -636,7 +626,7 @@ void hugetlb_vmemmap_optimize_folio(const struct hstate *h, struct folio *folio)
 {
 	LIST_HEAD(vmemmap_pages);
 
-	__hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, VMEMMAP_SYNCHRONIZE_RCU);
+	__hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, 0);
 	free_vmemmap_page_list(&vmemmap_pages);
 }
 
@@ -664,7 +654,7 @@ static void __hugetlb_vmemmap_optimize_folios(struct hstate *h,
 	struct folio *folio;
 	int nr_to_optimize;
 	LIST_HEAD(vmemmap_pages);
-	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH | VMEMMAP_SYNCHRONIZE_RCU;
+	unsigned long flags = VMEMMAP_REMAP_NO_TLB_FLUSH;
 
 	nr_to_optimize = 0;
 	list_for_each_entry(folio, folio_list, lru) {
@@ -717,8 +707,6 @@ static void __hugetlb_vmemmap_optimize_folios(struct hstate *h,
 		int ret;
 
 		ret = __hugetlb_vmemmap_optimize_folio(h, folio, &vmemmap_pages, flags);
-		/* only need to synchronize_rcu() once for each batch */
-		flags &= ~VMEMMAP_SYNCHRONIZE_RCU;
 
 		/*
 		 * Pages to be freed may have been accumulated.  If we
-- 
2.51.2

[PATCHv3 13/15] mm/hugetlb: Remove hugetlb_optimize_vmemmap_key static key

[Kiryl Shutsemau]

The hugetlb_optimize_vmemmap_key static key was used to guard fake head
detection in compound_head() and related functions. It allowed skipping
the fake head checks entirely when HVO was not in use.

With fake heads eliminated and the detection code removed, the static
key serves no purpose. Remove its definition and all increment/decrement
calls.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 include/linux/page-flags.h |  2 --
 mm/hugetlb_vmemmap.c       | 14 ++------------
 2 files changed, 2 insertions(+), 14 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index 660f9154a211..f89702e101e8 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -221,8 +221,6 @@ static __always_inline bool compound_info_has_mask(void)
 	return is_power_of_2(sizeof(struct page));
 }
 
-DECLARE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
-
 static __always_inline unsigned long _compound_head(const struct page *page)
 {
 	unsigned long info = READ_ONCE(page->compound_info);
diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index 5a802e292bac..5afd9f4231e3 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -399,9 +399,6 @@ static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
 	return vmemmap_remap_range(start, end, &walk);
 }
 
-DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
-EXPORT_SYMBOL(hugetlb_optimize_vmemmap_key);
-
 static bool vmemmap_optimize_enabled = IS_ENABLED(CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP_DEFAULT_ON);
 static int __init hugetlb_vmemmap_optimize_param(char *buf)
 {
@@ -431,10 +428,8 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
 	 * discarded vmemmap pages must be allocated and remapping.
 	 */
 	ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, flags);
-	if (!ret) {
+	if (!ret)
 		folio_clear_hugetlb_vmemmap_optimized(folio);
-		static_branch_dec(&hugetlb_optimize_vmemmap_key);
-	}
 
 	return ret;
 }
@@ -566,8 +561,6 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	if (!vmemmap_tail)
 		return -ENOMEM;
 
-	static_branch_inc(&hugetlb_optimize_vmemmap_key);
-
 	/*
 	 * Very Subtle
 	 * If VMEMMAP_REMAP_NO_TLB_FLUSH is set, TLB flushing is not performed
@@ -604,10 +597,8 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 				 vmemmap_head, vmemmap_tail,
 				 vmemmap_pages, flags);
 out:
-	if (ret) {
-		static_branch_dec(&hugetlb_optimize_vmemmap_key);
+	if (ret)
 		folio_clear_hugetlb_vmemmap_optimized(folio);
-	}
 
 	return ret;
 }
@@ -673,7 +664,6 @@ static void __hugetlb_vmemmap_optimize_folios(struct hstate *h,
 			register_page_bootmem_memmap(pfn_to_section_nr(spfn),
 					&folio->page,
 					HUGETLB_VMEMMAP_RESERVE_SIZE);
-			static_branch_inc(&hugetlb_optimize_vmemmap_key);
 			continue;
 		}
 
-- 
2.51.2

[PATCHv3 14/15] mm: Remove the branch from compound_head()

[Kiryl Shutsemau]

The compound_head() function is a hot path. For example, the zap path
calls it for every leaf page table entry.

Rewrite the helper function in a branchless manner to eliminate the risk
of CPU branch misprediction.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
Reviewed-by: Muchun Song <muchun.song@linux.dev>
---
 include/linux/page-flags.h | 27 +++++++++++++++++----------
 1 file changed, 17 insertions(+), 10 deletions(-)

diff --git a/include/linux/page-flags.h b/include/linux/page-flags.h
index f89702e101e8..95ac963b78af 100644
--- a/include/linux/page-flags.h
+++ b/include/linux/page-flags.h
@@ -224,25 +224,32 @@ static __always_inline bool compound_info_has_mask(void)
 static __always_inline unsigned long _compound_head(const struct page *page)
 {
 	unsigned long info = READ_ONCE(page->compound_info);
+	unsigned long mask;
+
+	if (!compound_info_has_mask()) {
+		/* Bit 0 encodes PageTail() */
+		if (info & 1)
+			return info - 1;
 
-	/* Bit 0 encodes PageTail() */
-	if (!(info & 1))
 		return (unsigned long)page;
-
-	/*
-	 * If compound_info_has_mask() is false, the rest of compound_info is
-	 * the pointer to the head page.
-	 */
-	if (!compound_info_has_mask())
-		return info - 1;
+	}
 
 	/*
 	 * If compoun_info_has_mask() is true the rest of the info encodes
 	 * the mask that converts the address of the tail page to the head page.
 	 *
 	 * No need to clear bit 0 in the mask as 'page' always has it clear.
+	 *
+	 * Let's do it in a branchless manner.
 	 */
-	return (unsigned long)page & info;
+
+	/* Non-tail: -1UL, Tail: 0 */
+	mask = (info & 1) - 1;
+
+	/* Non-tail: -1UL, Tail: info */
+	mask |= info;
+
+	return (unsigned long)page & mask;
 }
 
 #define compound_head(page)	((typeof(page))_compound_head(page))
-- 
2.51.2

[PATCHv3 15/15] hugetlb: Update vmemmap_dedup.rst

[Kiryl Shutsemau]

Update the documentation regarding vmemmap optimization for hugetlb to
reflect the changes in how the kernel maps the tail pages.

Fake heads no longer exist. Remove their description.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---
 Documentation/mm/vmemmap_dedup.rst | 60 +++++++++++++-----------------
 1 file changed, 26 insertions(+), 34 deletions(-)

diff --git a/Documentation/mm/vmemmap_dedup.rst b/Documentation/mm/vmemmap_dedup.rst
index 1863d88d2dcb..fca9d0ce282a 100644
--- a/Documentation/mm/vmemmap_dedup.rst
+++ b/Documentation/mm/vmemmap_dedup.rst
@@ -124,33 +124,35 @@ Here is how things look before optimization::
  |           |
  +-----------+
 
-The value of page->compound_info is the same for all tail pages. The first
-page of ``struct page`` (page 0) associated with the HugeTLB page contains the 4
-``struct page`` necessary to describe the HugeTLB. The only use of the remaining
-pages of ``struct page`` (page 1 to page 7) is to point to page->compound_info.
-Therefore, we can remap pages 1 to 7 to page 0. Only 1 page of ``struct page``
-will be used for each HugeTLB page. This will allow us to free the remaining
-7 pages to the buddy allocator.
+The first page of ``struct page`` (page 0) associated with the HugeTLB page
+contains the 4 ``struct page`` necessary to describe the HugeTLB. The remaining
+pages of ``struct page`` (page 1 to page 7) are tail pages.
+
+The optimization is only applied when the size of the struct page is a power-of-2
+In this case, all tail pages of the same order are identical. See
+compound_head(). This allows us to remap the tail pages of the vmemmap to a
+shared, read-only page. The head page is also remapped to a new page. This
+allows the original vmemmap pages to be freed.
 
 Here is how things look after remapping::
 
-    HugeTLB                  struct pages(8 pages)         page frame(8 pages)
- +-----------+ ---virt_to_page---> +-----------+   mapping to   +-----------+
- |           |                     |     0     | -------------> |     0     |
- |           |                     +-----------+                +-----------+
- |           |                     |     1     | ---------------^ ^ ^ ^ ^ ^ ^
- |           |                     +-----------+                  | | | | | |
- |           |                     |     2     | -----------------+ | | | | |
- |           |                     +-----------+                    | | | | |
- |           |                     |     3     | -------------------+ | | | |
- |           |                     +-----------+                      | | | |
- |           |                     |     4     | ---------------------+ | | |
- |    PMD    |                     +-----------+                        | | |
- |   level   |                     |     5     | -----------------------+ | |
- |  mapping  |                     +-----------+                          | |
- |           |                     |     6     | -------------------------+ |
- |           |                     +-----------+                            |
- |           |                     |     7     | ---------------------------+
+    HugeTLB                  struct pages(8 pages)                 page frame
+ +-----------+ ---virt_to_page---> +-----------+   mapping to   +----------------+
+ |           |                     |     0     | -------------> |       0        |
+ |           |                     +-----------+                +----------------+
+ |           |                     |     1     | ------┐
+ |           |                     +-----------+       |
+ |           |                     |     2     | ------┼        +----------------------------+
+ |           |                     +-----------+       |        | A single, per-node page    |
+ |           |                     |     3     | ------┼------> | frame shared among all     |
+ |           |                     +-----------+       |        | hugepages of the same size |
+ |           |                     |     4     | ------┼        +----------------------------+
+ |           |                     +-----------+       |
+ |           |                     |     5     | ------┼
+ |    PMD    |                     +-----------+       |
+ |   level   |                     |     6     | ------┼
+ |  mapping  |                     +-----------+       |
+ |           |                     |     7     | ------┘
  |           |                     +-----------+
  |           |
  |           |
@@ -172,16 +174,6 @@ The contiguous bit is used to increase the mapping size at the pmd and pte
 (last) level. So this type of HugeTLB page can be optimized only when its
 size of the ``struct page`` structs is greater than **1** page.
 
-Notice: The head vmemmap page is not freed to the buddy allocator and all
-tail vmemmap pages are mapped to the head vmemmap page frame. So we can see
-more than one ``struct page`` struct with ``PG_head`` (e.g. 8 per 2 MB HugeTLB
-page) associated with each HugeTLB page. The ``compound_head()`` can handle
-this correctly. There is only **one** head ``struct page``, the tail
-``struct page`` with ``PG_head`` are fake head ``struct page``.  We need an
-approach to distinguish between those two different types of ``struct page`` so
-that ``compound_head()`` can return the real head ``struct page`` when the
-parameter is the tail ``struct page`` but with ``PG_head``.
-
 Device DAX
 ==========
 
-- 
2.51.2

Re: [PATCHv3 02/15] mm: Move MAX_FOLIO_ORDER definition to mmzone.h

[David Hildenbrand (Red Hat)]

On 1/15/26 15:45, Kiryl Shutsemau wrote:
> Move MAX_FOLIO_ORDER definition from mm.h to mmzone.h.

Why should that belong to mmzone.h ? Semantically doesn't make sense, no?

-- 
Cheers

David

Re: [PATCHv3 02/15] mm: Move MAX_FOLIO_ORDER definition to mmzone.h

[David Hildenbrand (Red Hat)]

On 1/15/26 17:35, David Hildenbrand (Red Hat) wrote:
> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>> Move MAX_FOLIO_ORDER definition from mm.h to mmzone.h.
> 
> Why should that belong to mmzone.h ? Semantically doesn't make sense, no?

To clarify, I would have moved it to something more folio-type specific, 
not something buddy/zone specific.

I guess page-flags.h + mm.h are our dumping ground for folio handling. I 
would have moved it to the former.

Likely we want some folio.h in some distant future.

Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>

-- 
Cheers

David

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[David Hildenbrand (Red Hat)]

On 1/15/26 15:45, Kiryl Shutsemau wrote:
> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
> vmemmap pages for huge pages and remapping the freed range to a single
> page containing the struct page metadata.
> 
> With the new mask-based compound_info encoding (for power-of-2 struct
> page sizes), all tail pages of the same order are now identical
> regardless of which compound page they belong to. This means the tail
> pages can be truly shared without fake heads.
> 
> Allocate a single page of initialized tail struct pages per NUMA node
> per order in the vmemmap_tails[] array in pglist_data. All huge pages
> of that order on the node share this tail page, mapped read-only into
> their vmemmap. The head page remains unique per huge page.
> 
> This eliminates fake heads while maintaining the same memory savings,
> and simplifies compound_head() by removing fake head detection.
> 
> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> ---
>   include/linux/mmzone.h | 16 ++++++++++++++-
>   mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>   mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>   3 files changed, 93 insertions(+), 11 deletions(-)
> 
> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> index 322ed4c42cfc..2ee3eb610291 100644
> --- a/include/linux/mmzone.h
> +++ b/include/linux/mmzone.h
> @@ -82,7 +82,11 @@
>    * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>    * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>    */
> -#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
> +#ifdef CONFIG_64BIT
> +#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
> +#else
> +#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
> +#endif

Where do these magic values stem from, and how do they related to the 
comment above that clearly spells out 16G vs. 1G ?


-- 
Cheers

David

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau]

On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
> On 1/15/26 15:45, Kiryl Shutsemau wrote:
> > HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
> > vmemmap pages for huge pages and remapping the freed range to a single
> > page containing the struct page metadata.
> > 
> > With the new mask-based compound_info encoding (for power-of-2 struct
> > page sizes), all tail pages of the same order are now identical
> > regardless of which compound page they belong to. This means the tail
> > pages can be truly shared without fake heads.
> > 
> > Allocate a single page of initialized tail struct pages per NUMA node
> > per order in the vmemmap_tails[] array in pglist_data. All huge pages
> > of that order on the node share this tail page, mapped read-only into
> > their vmemmap. The head page remains unique per huge page.
> > 
> > This eliminates fake heads while maintaining the same memory savings,
> > and simplifies compound_head() by removing fake head detection.
> > 
> > Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> > ---
> >   include/linux/mmzone.h | 16 ++++++++++++++-
> >   mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
> >   mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
> >   3 files changed, 93 insertions(+), 11 deletions(-)
> > 
> > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > index 322ed4c42cfc..2ee3eb610291 100644
> > --- a/include/linux/mmzone.h
> > +++ b/include/linux/mmzone.h
> > @@ -82,7 +82,11 @@
> >    * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
> >    * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
> >    */
> > -#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
> > +#ifdef CONFIG_64BIT
> > +#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
> > +#else
> > +#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
> > +#endif
> 
> Where do these magic values stem from, and how do they related to the
> comment above that clearly spells out 16G vs. 1G ?

This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
is 1G. Subtract PAGE_SHIFT to get the order.

The change allows the value to be used to define NR_VMEMMAP_TAILS which
is used specify size of vmemmap_tails array.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 02/15] mm: Move MAX_FOLIO_ORDER definition to mmzone.h

[Kiryl Shutsemau]

On Thu, Jan 15, 2026 at 05:48:20PM +0100, David Hildenbrand (Red Hat) wrote:
> On 1/15/26 17:35, David Hildenbrand (Red Hat) wrote:
> > On 1/15/26 15:45, Kiryl Shutsemau wrote:
> > > Move MAX_FOLIO_ORDER definition from mm.h to mmzone.h.
> > 
> > Why should that belong to mmzone.h ? Semantically doesn't make sense, no?
> 
> To clarify, I would have moved it to something more folio-type specific, not
> something buddy/zone specific.
> 
> I guess page-flags.h + mm.h are our dumping ground for folio handling. I
> would have moved it to the former.

I moved it next to MAX_PAGE_ORDER definition which seemed to be logical.
Moving MAX_PAGE_ORDER around is much more risky (will likely generate a
lot of build bugs) and outside of the scope of the patchset.

I think wherever they will be moved, it should be moved together.

> Likely we want some folio.h in some distant future.

Sounds like a plan! :P

> Acked-by: David Hildenbrand (Red Hat) <david@kernel.org>

Thanks!

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[David Hildenbrand (Red Hat)]

On 1/15/26 18:23, Kiryl Shutsemau wrote:
> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
>>> vmemmap pages for huge pages and remapping the freed range to a single
>>> page containing the struct page metadata.
>>>
>>> With the new mask-based compound_info encoding (for power-of-2 struct
>>> page sizes), all tail pages of the same order are now identical
>>> regardless of which compound page they belong to. This means the tail
>>> pages can be truly shared without fake heads.
>>>
>>> Allocate a single page of initialized tail struct pages per NUMA node
>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
>>> of that order on the node share this tail page, mapped read-only into
>>> their vmemmap. The head page remains unique per huge page.
>>>
>>> This eliminates fake heads while maintaining the same memory savings,
>>> and simplifies compound_head() by removing fake head detection.
>>>
>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
>>> ---
>>>    include/linux/mmzone.h | 16 ++++++++++++++-
>>>    mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>>>    mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>>>    3 files changed, 93 insertions(+), 11 deletions(-)
>>>
>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>> index 322ed4c42cfc..2ee3eb610291 100644
>>> --- a/include/linux/mmzone.h
>>> +++ b/include/linux/mmzone.h
>>> @@ -82,7 +82,11 @@
>>>     * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>>>     * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>>>     */
>>> -#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
>>> +#ifdef CONFIG_64BIT
>>> +#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
>>> +#else
>>> +#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
>>> +#endif
>>
>> Where do these magic values stem from, and how do they related to the
>> comment above that clearly spells out 16G vs. 1G ?
> 
> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
> is 1G. Subtract PAGE_SHIFT to get the order.
> 
> The change allows the value to be used to define NR_VMEMMAP_TAILS which
> is used specify size of vmemmap_tails array.

get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G) should evaluate to 
a constant by the compiler.

See __builtin_constant_p handling in get_order().

If that is not working then we have to figure out why.

Was this only a specific config in where you ran into compile-time problems?

-- 
Cheers

David

Re: [PATCHv3 02/15] mm: Move MAX_FOLIO_ORDER definition to mmzone.h

[David Hildenbrand (Red Hat)]

On 1/15/26 18:26, Kiryl Shutsemau wrote:
> On Thu, Jan 15, 2026 at 05:48:20PM +0100, David Hildenbrand (Red Hat) wrote:
>> On 1/15/26 17:35, David Hildenbrand (Red Hat) wrote:
>>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>>> Move MAX_FOLIO_ORDER definition from mm.h to mmzone.h.
>>>
>>> Why should that belong to mmzone.h ? Semantically doesn't make sense, no?
>>
>> To clarify, I would have moved it to something more folio-type specific, not
>> something buddy/zone specific.
>>
>> I guess page-flags.h + mm.h are our dumping ground for folio handling. I
>> would have moved it to the former.
> 
> I moved it next to MAX_PAGE_ORDER definition which seemed to be logical.

MAX_PAGE_ORDER is a buddy concept. MAX_FOLIO_ORDER is not a buddy concept.

-- 
Cheers

David

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau]

On Thu, Jan 15, 2026 at 06:41:44PM +0100, David Hildenbrand (Red Hat) wrote:
> On 1/15/26 18:23, Kiryl Shutsemau wrote:
> > On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
> > > On 1/15/26 15:45, Kiryl Shutsemau wrote:
> > > > HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
> > > > vmemmap pages for huge pages and remapping the freed range to a single
> > > > page containing the struct page metadata.
> > > > 
> > > > With the new mask-based compound_info encoding (for power-of-2 struct
> > > > page sizes), all tail pages of the same order are now identical
> > > > regardless of which compound page they belong to. This means the tail
> > > > pages can be truly shared without fake heads.
> > > > 
> > > > Allocate a single page of initialized tail struct pages per NUMA node
> > > > per order in the vmemmap_tails[] array in pglist_data. All huge pages
> > > > of that order on the node share this tail page, mapped read-only into
> > > > their vmemmap. The head page remains unique per huge page.
> > > > 
> > > > This eliminates fake heads while maintaining the same memory savings,
> > > > and simplifies compound_head() by removing fake head detection.
> > > > 
> > > > Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> > > > ---
> > > >    include/linux/mmzone.h | 16 ++++++++++++++-
> > > >    mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
> > > >    mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
> > > >    3 files changed, 93 insertions(+), 11 deletions(-)
> > > > 
> > > > diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> > > > index 322ed4c42cfc..2ee3eb610291 100644
> > > > --- a/include/linux/mmzone.h
> > > > +++ b/include/linux/mmzone.h
> > > > @@ -82,7 +82,11 @@
> > > >     * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
> > > >     * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
> > > >     */
> > > > -#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
> > > > +#ifdef CONFIG_64BIT
> > > > +#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
> > > > +#else
> > > > +#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
> > > > +#endif
> > > 
> > > Where do these magic values stem from, and how do they related to the
> > > comment above that clearly spells out 16G vs. 1G ?
> > 
> > This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
> > is 1G. Subtract PAGE_SHIFT to get the order.
> > 
> > The change allows the value to be used to define NR_VMEMMAP_TAILS which
> > is used specify size of vmemmap_tails array.
> 
> get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G) should evaluate to a
> constant by the compiler.
>
> See __builtin_constant_p handling in get_order().
> 
> If that is not working then we have to figure out why.

asm-offsets.s compilation fails:

../include/linux/mmzone.h:1574:16: error: fields must have a constant size:
      'variable length array in structure' extension will never be supported
 1574 |         unsigned long vmemmap_tails[NR_VMEMMAP_TAILS];

Here's how preprocessor dump of vmemmap_tails looks like:

 unsigned long vmemmap_tails[(get_order(1 ? (0x400000000ULL) : 0x40000000) - (( __builtin_constant_p(2 * ((1UL) << 12) / sizeof(struct page)) ? ((2 * ((1UL) << 12) / sizeof(struct page)) < 2 ? 0 : 63 - __builtin_clzll(2 * ((1UL) << 12) / sizeof(struct page))) : (sizeof(2 * ((1UL) << 12) / sizeof(struct page)) <= 4) ? __ilog2_u32(2 * ((1UL) << 12) / sizeof(struct page)) : __ilog2_u64(2 * ((1UL) << 12) / sizeof(struct page)) )) + 1)];

And here's get_order():

static inline __attribute__((__gnu_inline__)) __attribute__((__unused__)) __attribute__((no_instrument_function)) __attribute__((__always_inline__)) __attribute__((__const__)) int get_order(unsigned long size)
{
 if (__builtin_constant_p(size)) {
  if (!size)
   return 64 - 12;

  if (size < (1UL << 12))
   return 0;

  return ( __builtin_constant_p((size) - 1) ? (((size) - 1) < 2 ? 0 : 63 - __builtin_clzll((size) - 1)) : (sizeof((size) - 1) <= 4) ? __ilog2_u32((size) - 1) : __ilog2_u64((size) - 1) ) - 12 + 1;
 }

 size--;
 size >>= 12;



 return fls64(size);

}

I am not sure why it is not compile-time constant. I have not dig
deeper.

Switching to ilog2(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G) - PAGE_SHIFT works,
but I personally find my variant more readable.

Do you want me to dig deeper to check if making get_order() work
possible?

> Was this only a specific config in where you ran into compile-time problems?

I am not aware about any particular config dependency. Seems to be
everywhere.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[David Hildenbrand (Red Hat)]

On 1/15/26 19:58, Kiryl Shutsemau wrote:
> On Thu, Jan 15, 2026 at 06:41:44PM +0100, David Hildenbrand (Red Hat) wrote:
>> On 1/15/26 18:23, Kiryl Shutsemau wrote:
>>> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
>>>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
>>>>> vmemmap pages for huge pages and remapping the freed range to a single
>>>>> page containing the struct page metadata.
>>>>>
>>>>> With the new mask-based compound_info encoding (for power-of-2 struct
>>>>> page sizes), all tail pages of the same order are now identical
>>>>> regardless of which compound page they belong to. This means the tail
>>>>> pages can be truly shared without fake heads.
>>>>>
>>>>> Allocate a single page of initialized tail struct pages per NUMA node
>>>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
>>>>> of that order on the node share this tail page, mapped read-only into
>>>>> their vmemmap. The head page remains unique per huge page.
>>>>>
>>>>> This eliminates fake heads while maintaining the same memory savings,
>>>>> and simplifies compound_head() by removing fake head detection.
>>>>>
>>>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
>>>>> ---
>>>>>     include/linux/mmzone.h | 16 ++++++++++++++-
>>>>>     mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>>>>>     mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>>>>>     3 files changed, 93 insertions(+), 11 deletions(-)
>>>>>
>>>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>>>> index 322ed4c42cfc..2ee3eb610291 100644
>>>>> --- a/include/linux/mmzone.h
>>>>> +++ b/include/linux/mmzone.h
>>>>> @@ -82,7 +82,11 @@
>>>>>      * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>>>>>      * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>>>>>      */
>>>>> -#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
>>>>> +#ifdef CONFIG_64BIT
>>>>> +#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
>>>>> +#else
>>>>> +#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
>>>>> +#endif
>>>>
>>>> Where do these magic values stem from, and how do they related to the
>>>> comment above that clearly spells out 16G vs. 1G ?
>>>
>>> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
>>> is 1G. Subtract PAGE_SHIFT to get the order.
>>>
>>> The change allows the value to be used to define NR_VMEMMAP_TAILS which
>>> is used specify size of vmemmap_tails array.
>>
>> get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G) should evaluate to a
>> constant by the compiler.
>>
>> See __builtin_constant_p handling in get_order().
>>
>> If that is not working then we have to figure out why.
> 
> asm-offsets.s compilation fails:
> 
> ../include/linux/mmzone.h:1574:16: error: fields must have a constant size:
>        'variable length array in structure' extension will never be supported
>   1574 |         unsigned long vmemmap_tails[NR_VMEMMAP_TAILS];
> 
> Here's how preprocessor dump of vmemmap_tails looks like:
> 
>   unsigned long vmemmap_tails[(get_order(1 ? (0x400000000ULL) : 0x40000000) - (( __builtin_constant_p(2 * ((1UL) << 12) / sizeof(struct page)) ? ((2 * ((1UL) << 12) / sizeof(struct page)) < 2 ? 0 : 63 - __builtin_clzll(2 * ((1UL) << 12) / sizeof(struct page))) : (sizeof(2 * ((1UL) << 12) / sizeof(struct page)) <= 4) ? __ilog2_u32(2 * ((1UL) << 12) / sizeof(struct page)) : __ilog2_u64(2 * ((1UL) << 12) / sizeof(struct page)) )) + 1)];
> 
> And here's get_order():
> 
> static inline __attribute__((__gnu_inline__)) __attribute__((__unused__)) __attribute__((no_instrument_function)) __attribute__((__always_inline__)) __attribute__((__const__)) int get_order(unsigned long size)
> {
>   if (__builtin_constant_p(size)) {
>    if (!size)
>     return 64 - 12;
> 
>    if (size < (1UL << 12))
>     return 0;
> 
>    return ( __builtin_constant_p((size) - 1) ? (((size) - 1) < 2 ? 0 : 63 - __builtin_clzll((size) - 1)) : (sizeof((size) - 1) <= 4) ? __ilog2_u32((size) - 1) : __ilog2_u64((size) - 1) ) - 12 + 1;
>   }
> 
>   size--;
>   size >>= 12;
> 
> 
> 
>   return fls64(size);
> 
> }
> 
> I am not sure why it is not compile-time constant. I have not dig
> deeper.

Very weird. Almost sounds like a bug given that get_order() ends up using ilog2.

But it gets even weirder:

diff --git a/include/linux/mm.h b/include/linux/mm.h
index 6f959d8ca4b42..a54445682ccc4 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -2281,6 +2281,9 @@ static inline unsigned long folio_nr_pages(const struct folio *folio)
   * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
   */
  #define MAX_FOLIO_ORDER                get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
+
+static_assert(__builtin_constant_p(MAX_FOLIO_ORDER));
+
  #else
  /*
   * Without hugetlb, gigantic folios that are bigger than a single PUD are

gives me


./include/linux/build_bug.h:78:41: error: static assertion failed: "__builtin_constant_p(MAX_FOLIO_ORDER)"
    78 | #define __static_assert(expr, msg, ...) _Static_assert(expr, msg)
       |                                         ^~~~~~~~~~~~~~
./include/linux/build_bug.h:77:34: note: in expansion of macro '__static_assert'
    77 | #define static_assert(expr, ...) __static_assert(expr, ##__VA_ARGS__, #expr)
       |                                  ^~~~~~~~~~~~~~~
./include/linux/mm.h:2285:1: note: in expansion of macro 'static_assert'
  2285 | static_assert(__builtin_constant_p(MAX_FOLIO_ORDER));
       | ^~~~~~~~~~~~~

And reversing the condition fixes it.

... so it is a constant? Huh?



Some history on the SZ change here: https://lore.kernel.org/all/a31e6d70-9275-4277-991b-9de1aea03cd7@csgroup.eu/

> 
> Switching to ilog2(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G) - PAGE_SHIFT works,
> but I personally find my variant more readable.
> 

Using SZ_16G/SZ_1G, is self-documenting. I'm fine with repeating the ilog2 like:


ifdef CONFIG_64BIT
#define MAX_FOLIO_ORDER		(ilog2(SZ_16G) - PAGE_SHIFT)
...


Also, make sure to spell that out in the patch description.


Figuring out why we don't get a constant would be even nicer ... or why this does something else
than expected.

-- 
Cheers

David

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[David Hildenbrand (Red Hat)]

On 1/15/26 20:33, David Hildenbrand (Red Hat) wrote:
> On 1/15/26 19:58, Kiryl Shutsemau wrote:
>> On Thu, Jan 15, 2026 at 06:41:44PM +0100, David Hildenbrand (Red Hat) wrote:
>>> On 1/15/26 18:23, Kiryl Shutsemau wrote:
>>>> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
>>>>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>>>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
>>>>>> vmemmap pages for huge pages and remapping the freed range to a single
>>>>>> page containing the struct page metadata.
>>>>>>
>>>>>> With the new mask-based compound_info encoding (for power-of-2 struct
>>>>>> page sizes), all tail pages of the same order are now identical
>>>>>> regardless of which compound page they belong to. This means the tail
>>>>>> pages can be truly shared without fake heads.
>>>>>>
>>>>>> Allocate a single page of initialized tail struct pages per NUMA node
>>>>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
>>>>>> of that order on the node share this tail page, mapped read-only into
>>>>>> their vmemmap. The head page remains unique per huge page.
>>>>>>
>>>>>> This eliminates fake heads while maintaining the same memory savings,
>>>>>> and simplifies compound_head() by removing fake head detection.
>>>>>>
>>>>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
>>>>>> ---
>>>>>>      include/linux/mmzone.h | 16 ++++++++++++++-
>>>>>>      mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>>>>>>      mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>>>>>>      3 files changed, 93 insertions(+), 11 deletions(-)
>>>>>>
>>>>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>>>>> index 322ed4c42cfc..2ee3eb610291 100644
>>>>>> --- a/include/linux/mmzone.h
>>>>>> +++ b/include/linux/mmzone.h
>>>>>> @@ -82,7 +82,11 @@
>>>>>>       * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>>>>>>       * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>>>>>>       */
>>>>>> -#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
>>>>>> +#ifdef CONFIG_64BIT
>>>>>> +#define MAX_FOLIO_ORDER		(34 - PAGE_SHIFT)
>>>>>> +#else
>>>>>> +#define MAX_FOLIO_ORDER		(30 - PAGE_SHIFT)
>>>>>> +#endif
>>>>>
>>>>> Where do these magic values stem from, and how do they related to the
>>>>> comment above that clearly spells out 16G vs. 1G ?
>>>>
>>>> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
>>>> is 1G. Subtract PAGE_SHIFT to get the order.
>>>>
>>>> The change allows the value to be used to define NR_VMEMMAP_TAILS which
>>>> is used specify size of vmemmap_tails array.
>>>
>>> get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G) should evaluate to a
>>> constant by the compiler.
>>>
>>> See __builtin_constant_p handling in get_order().
>>>
>>> If that is not working then we have to figure out why.
>>
>> asm-offsets.s compilation fails:
>>
>> ../include/linux/mmzone.h:1574:16: error: fields must have a constant size:
>>         'variable length array in structure' extension will never be supported
>>    1574 |         unsigned long vmemmap_tails[NR_VMEMMAP_TAILS];
>>
>> Here's how preprocessor dump of vmemmap_tails looks like:
>>
>>    unsigned long vmemmap_tails[(get_order(1 ? (0x400000000ULL) : 0x40000000) - (( __builtin_constant_p(2 * ((1UL) << 12) / sizeof(struct page)) ? ((2 * ((1UL) << 12) / sizeof(struct page)) < 2 ? 0 : 63 - __builtin_clzll(2 * ((1UL) << 12) / sizeof(struct page))) : (sizeof(2 * ((1UL) << 12) / sizeof(struct page)) <= 4) ? __ilog2_u32(2 * ((1UL) << 12) / sizeof(struct page)) : __ilog2_u64(2 * ((1UL) << 12) / sizeof(struct page)) )) + 1)];
>>
>> And here's get_order():
>>
>> static inline __attribute__((__gnu_inline__)) __attribute__((__unused__)) __attribute__((no_instrument_function)) __attribute__((__always_inline__)) __attribute__((__const__)) int get_order(unsigned long size)
>> {
>>    if (__builtin_constant_p(size)) {
>>     if (!size)
>>      return 64 - 12;
>>
>>     if (size < (1UL << 12))
>>      return 0;
>>
>>     return ( __builtin_constant_p((size) - 1) ? (((size) - 1) < 2 ? 0 : 63 - __builtin_clzll((size) - 1)) : (sizeof((size) - 1) <= 4) ? __ilog2_u32((size) - 1) : __ilog2_u64((size) - 1) ) - 12 + 1;
>>    }
>>
>>    size--;
>>    size >>= 12;
>>
>>
>>
>>    return fls64(size);
>>
>> }
>>
>> I am not sure why it is not compile-time constant. I have not dig
>> deeper.
> 
> Very weird. Almost sounds like a bug given that get_order() ends up using ilog2.
> 
> But it gets even weirder:
> 
> diff --git a/include/linux/mm.h b/include/linux/mm.h
> index 6f959d8ca4b42..a54445682ccc4 100644
> --- a/include/linux/mm.h
> +++ b/include/linux/mm.h
> @@ -2281,6 +2281,9 @@ static inline unsigned long folio_nr_pages(const struct folio *folio)
>     * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>     */
>    #define MAX_FOLIO_ORDER                get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
> +
> +static_assert(__builtin_constant_p(MAX_FOLIO_ORDER));
> +
>    #else
>    /*
>     * Without hugetlb, gigantic folios that are bigger than a single PUD are
> 
> gives me
> 
> 
> ./include/linux/build_bug.h:78:41: error: static assertion failed: "__builtin_constant_p(MAX_FOLIO_ORDER)"
>      78 | #define __static_assert(expr, msg, ...) _Static_assert(expr, msg)
>         |                                         ^~~~~~~~~~~~~~
> ./include/linux/build_bug.h:77:34: note: in expansion of macro '__static_assert'
>      77 | #define static_assert(expr, ...) __static_assert(expr, ##__VA_ARGS__, #expr)
>         |                                  ^~~~~~~~~~~~~~~
> ./include/linux/mm.h:2285:1: note: in expansion of macro 'static_assert'
>    2285 | static_assert(__builtin_constant_p(MAX_FOLIO_ORDER));
>         | ^~~~~~~~~~~~~
> 
> And reversing the condition fixes it.
> 
> ... so it is a constant? Huh?

I've been staring at the computer for too long, this is not BUILD_BUG 
semantics. So we don't get a constant.

For some reason :)

Even when I just use get_order(4096).

-- 
Cheers

David

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Muchun Song]

> On Jan 16, 2026, at 01:23, Kiryl Shutsemau <kas@kernel.org> wrote:
> 
> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
>>> vmemmap pages for huge pages and remapping the freed range to a single
>>> page containing the struct page metadata.
>>> 
>>> With the new mask-based compound_info encoding (for power-of-2 struct
>>> page sizes), all tail pages of the same order are now identical
>>> regardless of which compound page they belong to. This means the tail
>>> pages can be truly shared without fake heads.
>>> 
>>> Allocate a single page of initialized tail struct pages per NUMA node
>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
>>> of that order on the node share this tail page, mapped read-only into
>>> their vmemmap. The head page remains unique per huge page.
>>> 
>>> This eliminates fake heads while maintaining the same memory savings,
>>> and simplifies compound_head() by removing fake head detection.
>>> 
>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
>>> ---
>>>  include/linux/mmzone.h | 16 ++++++++++++++-
>>>  mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>>>  mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>>>  3 files changed, 93 insertions(+), 11 deletions(-)
>>> 
>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>> index 322ed4c42cfc..2ee3eb610291 100644
>>> --- a/include/linux/mmzone.h
>>> +++ b/include/linux/mmzone.h
>>> @@ -82,7 +82,11 @@
>>>   * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>>>   * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>>>   */
>>> -#define MAX_FOLIO_ORDER get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
>>> +#ifdef CONFIG_64BIT
>>> +#define MAX_FOLIO_ORDER (34 - PAGE_SHIFT)
>>> +#else
>>> +#define MAX_FOLIO_ORDER (30 - PAGE_SHIFT)
>>> +#endif
>> 
>> Where do these magic values stem from, and how do they related to the
>> comment above that clearly spells out 16G vs. 1G ?
> 
> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
> is 1G. Subtract PAGE_SHIFT to get the order.
> 
> The change allows the value to be used to define NR_VMEMMAP_TAILS which
> is used specify size of vmemmap_tails array.

How about allocate ->vmemmap_tails array dynamically? If sizeof of struct
page is not power of two, then we could optimize away this array. Besides,
the original MAX_FOLIO_ORDER could work as well.

> 
> -- 
>  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau]

On Fri, Jan 16, 2026 at 10:38:02AM +0800, Muchun Song wrote:
> 
> 
> > On Jan 16, 2026, at 01:23, Kiryl Shutsemau <kas@kernel.org> wrote:
> > 
> > On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
> >> On 1/15/26 15:45, Kiryl Shutsemau wrote:
> >>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
> >>> vmemmap pages for huge pages and remapping the freed range to a single
> >>> page containing the struct page metadata.
> >>> 
> >>> With the new mask-based compound_info encoding (for power-of-2 struct
> >>> page sizes), all tail pages of the same order are now identical
> >>> regardless of which compound page they belong to. This means the tail
> >>> pages can be truly shared without fake heads.
> >>> 
> >>> Allocate a single page of initialized tail struct pages per NUMA node
> >>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
> >>> of that order on the node share this tail page, mapped read-only into
> >>> their vmemmap. The head page remains unique per huge page.
> >>> 
> >>> This eliminates fake heads while maintaining the same memory savings,
> >>> and simplifies compound_head() by removing fake head detection.
> >>> 
> >>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> >>> ---
> >>>  include/linux/mmzone.h | 16 ++++++++++++++-
> >>>  mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
> >>>  mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
> >>>  3 files changed, 93 insertions(+), 11 deletions(-)
> >>> 
> >>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> >>> index 322ed4c42cfc..2ee3eb610291 100644
> >>> --- a/include/linux/mmzone.h
> >>> +++ b/include/linux/mmzone.h
> >>> @@ -82,7 +82,11 @@
> >>>   * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
> >>>   * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
> >>>   */
> >>> -#define MAX_FOLIO_ORDER get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
> >>> +#ifdef CONFIG_64BIT
> >>> +#define MAX_FOLIO_ORDER (34 - PAGE_SHIFT)
> >>> +#else
> >>> +#define MAX_FOLIO_ORDER (30 - PAGE_SHIFT)
> >>> +#endif
> >> 
> >> Where do these magic values stem from, and how do they related to the
> >> comment above that clearly spells out 16G vs. 1G ?
> > 
> > This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
> > is 1G. Subtract PAGE_SHIFT to get the order.
> > 
> > The change allows the value to be used to define NR_VMEMMAP_TAILS which
> > is used specify size of vmemmap_tails array.
> 
> How about allocate ->vmemmap_tails array dynamically? If sizeof of struct
> page is not power of two, then we could optimize away this array. Besides,
> the original MAX_FOLIO_ORDER could work as well.

This is tricky.

We need vmemmap_tails array to be around early, in
hugetlb_vmemmap_init_early(). By the time, we don't have slab
functional yet.

I think getting the array compile-time is the best shot.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

[PATCHv3.1 10/15] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau]

HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
vmemmap pages for huge pages and remapping the freed range to a single
page containing the struct page metadata.

With the new mask-based compound_info encoding (for power-of-2 struct
page sizes), all tail pages of the same order are now identical
regardless of which compound page they belong to. This means the tail
pages can be truly shared without fake heads.

Allocate a single page of initialized tail struct pages per NUMA node
per order in the vmemmap_tails[] array in pglist_data. All huge pages of
that order on the node share this tail page, mapped read-only into their
vmemmap. The head page remains unique per huge page.

Redefine MAX_FOLIO_ORDER using ilog2(). The define has to produce a
compile-constant as it is used to specify vmemmap_tail array size.
For some reason, compiler is not able to solve get_order() at
compile-time, but ilog2() works.

This eliminates fake heads while maintaining the same memory savings,
and simplifies compound_head() by removing fake head detection.

Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
---

v3.1:
 - Define MAX_FOLIO_ORDER using ilog2();
 - Update commit message;

---
 include/linux/mmzone.h | 16 ++++++++++++++-
 mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
 mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
 3 files changed, 93 insertions(+), 11 deletions(-)

diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
index 322ed4c42cfc..bc333546c2d3 100644
--- a/include/linux/mmzone.h
+++ b/include/linux/mmzone.h
@@ -82,7 +82,11 @@
  * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
  * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
  */
-#define MAX_FOLIO_ORDER		get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
+#ifdef CONFIG_64BIT
+#define MAX_FOLIO_ORDER		(ilog2(SZ_16G) - PAGE_SHIFT)
+#else
+#define MAX_FOLIO_ORDER		(ilog2(SZ_1G) - PAGE_SHIFT)
+#endif
 #else
 /*
  * Without hugetlb, gigantic folios that are bigger than a single PUD are
@@ -1408,6 +1412,13 @@ struct memory_failure_stats {
 };
 #endif
 
+/*
+ * vmemmap optimization (like HVO) is only possible for page orders that fill
+ * two or more pages with struct pages.
+ */
+#define VMEMMAP_TAIL_MIN_ORDER (ilog2(2 * PAGE_SIZE / sizeof(struct page)))
+#define NR_VMEMMAP_TAILS (MAX_FOLIO_ORDER - VMEMMAP_TAIL_MIN_ORDER + 1)
+
 /*
  * On NUMA machines, each NUMA node would have a pg_data_t to describe
  * it's memory layout. On UMA machines there is a single pglist_data which
@@ -1556,6 +1567,9 @@ typedef struct pglist_data {
 #ifdef CONFIG_MEMORY_FAILURE
 	struct memory_failure_stats mf_stats;
 #endif
+#ifdef CONFIG_SPARSEMEM_VMEMMAP
+	unsigned long vmemmap_tails[NR_VMEMMAP_TAILS];
+#endif
 } pg_data_t;
 
 #define node_present_pages(nid)	(NODE_DATA(nid)->node_present_pages)
diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index 2b19c2205091..cbdca4684db1 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c
@@ -18,6 +18,7 @@
 #include <asm/pgalloc.h>
 #include <asm/tlbflush.h>
 #include "hugetlb_vmemmap.h"
+#include "internal.h"
 
 /**
  * struct vmemmap_remap_walk - walk vmemmap page table
@@ -517,6 +518,41 @@ static bool vmemmap_should_optimize_folio(const struct hstate *h, struct folio *
 	return true;
 }
 
+static struct page *vmemmap_get_tail(unsigned int order, int node)
+{
+	unsigned long pfn;
+	unsigned int idx;
+	struct page *tail, *p;
+
+	idx = order - VMEMMAP_TAIL_MIN_ORDER;
+	pfn =  NODE_DATA(node)->vmemmap_tails[idx];
+	if (pfn)
+		return pfn_to_page(pfn);
+
+	tail = alloc_pages_node(node, GFP_KERNEL, 0);
+	if (!tail)
+		return NULL;
+
+	p = page_to_virt(tail);
+	for (int i = 0; i < PAGE_SIZE / sizeof(struct page); i++)
+		prep_compound_tail(p + i, NULL, order);
+
+	spin_lock(&hugetlb_lock);
+	if (!NODE_DATA(node)->vmemmap_tails[idx]) {
+		pfn = PHYS_PFN(virt_to_phys(p));
+		NODE_DATA(node)->vmemmap_tails[idx] = pfn;
+		tail = NULL;
+	} else {
+		pfn = NODE_DATA(node)->vmemmap_tails[idx];
+	}
+	spin_unlock(&hugetlb_lock);
+
+	if (tail)
+		__free_page(tail);
+
+	return pfn_to_page(pfn);
+}
+
 static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 					    struct folio *folio,
 					    struct list_head *vmemmap_pages,
@@ -532,6 +568,12 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	if (!vmemmap_should_optimize_folio(h, folio))
 		return ret;
 
+	nid = folio_nid(folio);
+
+	vmemmap_tail = vmemmap_get_tail(h->order, nid);
+	if (!vmemmap_tail)
+		return -ENOMEM;
+
 	static_branch_inc(&hugetlb_optimize_vmemmap_key);
 
 	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
@@ -549,7 +591,6 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	 */
 	folio_set_hugetlb_vmemmap_optimized(folio);
 
-	nid = folio_nid(folio);
 	vmemmap_head = alloc_pages_node(nid, GFP_KERNEL, 0);
 
 	if (!vmemmap_head) {
@@ -561,7 +602,6 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
 	list_add(&vmemmap_head->lru, vmemmap_pages);
 	memmap_pages_add(1);
 
-	vmemmap_tail	= vmemmap_head;
 	vmemmap_start	= (unsigned long)folio;
 	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
 
diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c
index dbd8daccade2..94b4e90fa00f 100644
--- a/mm/sparse-vmemmap.c
+++ b/mm/sparse-vmemmap.c
@@ -378,16 +378,45 @@ void vmemmap_wrprotect_hvo(unsigned long addr, unsigned long end,
 	}
 }
 
-/*
- * Populate vmemmap pages HVO-style. The first page contains the head
- * page and needed tail pages, the other ones are mirrors of the first
- * page.
- */
+static __meminit unsigned long vmemmap_get_tail(unsigned int order, int node)
+{
+	unsigned long pfn;
+	unsigned int idx;
+	struct page *p;
+
+	BUG_ON(order < VMEMMAP_TAIL_MIN_ORDER);
+	BUG_ON(order > MAX_FOLIO_ORDER);
+
+	idx = order - VMEMMAP_TAIL_MIN_ORDER;
+	pfn =  NODE_DATA(node)->vmemmap_tails[idx];
+	if (pfn)
+		return pfn;
+
+	p = vmemmap_alloc_block_zero(PAGE_SIZE, node);
+	if (!p)
+		return 0;
+
+	for (int i = 0; i < PAGE_SIZE / sizeof(struct page); i++)
+		prep_compound_tail(p + i, NULL, order);
+
+	pfn = PHYS_PFN(virt_to_phys(p));
+	NODE_DATA(node)->vmemmap_tails[idx] = pfn;
+
+	return pfn;
+}
+
 int __meminit vmemmap_populate_hvo(unsigned long addr, unsigned long end,
 				       int node, unsigned long headsize)
 {
+	unsigned long maddr, len, tail_pfn;
+	unsigned int order;
 	pte_t *pte;
-	unsigned long maddr;
+
+	len = end - addr;
+	order = ilog2(len * sizeof(struct page) / PAGE_SIZE);
+	tail_pfn = vmemmap_get_tail(order, node);
+	if (!tail_pfn)
+		return -ENOMEM;
 
 	for (maddr = addr; maddr < addr + headsize; maddr += PAGE_SIZE) {
 		pte = vmemmap_populate_address(maddr, node, NULL, -1, 0);
@@ -398,8 +427,7 @@ int __meminit vmemmap_populate_hvo(unsigned long addr, unsigned long end,
 	/*
 	 * Reuse the last page struct page mapped above for the rest.
 	 */
-	return vmemmap_populate_range(maddr, end, node, NULL,
-					pte_pfn(ptep_get(pte)), 0);
+	return vmemmap_populate_range(maddr, end, node, NULL, tail_pfn, 0);
 }
 
 void __weak __meminit vmemmap_set_pmd(pmd_t *pmd, void *p, int node,
-- 
2.51.2

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Muchun Song]

> On Jan 16, 2026, at 23:52, Kiryl Shutsemau <kas@kernel.org> wrote:
> 
> On Fri, Jan 16, 2026 at 10:38:02AM +0800, Muchun Song wrote:
>> 
>> 
>>> On Jan 16, 2026, at 01:23, Kiryl Shutsemau <kas@kernel.org> wrote:
>>> 
>>> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
>>>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
>>>>> vmemmap pages for huge pages and remapping the freed range to a single
>>>>> page containing the struct page metadata.
>>>>> 
>>>>> With the new mask-based compound_info encoding (for power-of-2 struct
>>>>> page sizes), all tail pages of the same order are now identical
>>>>> regardless of which compound page they belong to. This means the tail
>>>>> pages can be truly shared without fake heads.
>>>>> 
>>>>> Allocate a single page of initialized tail struct pages per NUMA node
>>>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
>>>>> of that order on the node share this tail page, mapped read-only into
>>>>> their vmemmap. The head page remains unique per huge page.
>>>>> 
>>>>> This eliminates fake heads while maintaining the same memory savings,
>>>>> and simplifies compound_head() by removing fake head detection.
>>>>> 
>>>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
>>>>> ---
>>>>> include/linux/mmzone.h | 16 ++++++++++++++-
>>>>> mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>>>>> mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>>>>> 3 files changed, 93 insertions(+), 11 deletions(-)
>>>>> 
>>>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>>>> index 322ed4c42cfc..2ee3eb610291 100644
>>>>> --- a/include/linux/mmzone.h
>>>>> +++ b/include/linux/mmzone.h
>>>>> @@ -82,7 +82,11 @@
>>>>>  * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>>>>>  * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>>>>>  */
>>>>> -#define MAX_FOLIO_ORDER get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
>>>>> +#ifdef CONFIG_64BIT
>>>>> +#define MAX_FOLIO_ORDER (34 - PAGE_SHIFT)
>>>>> +#else
>>>>> +#define MAX_FOLIO_ORDER (30 - PAGE_SHIFT)
>>>>> +#endif
>>>> 
>>>> Where do these magic values stem from, and how do they related to the
>>>> comment above that clearly spells out 16G vs. 1G ?
>>> 
>>> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
>>> is 1G. Subtract PAGE_SHIFT to get the order.
>>> 
>>> The change allows the value to be used to define NR_VMEMMAP_TAILS which
>>> is used specify size of vmemmap_tails array.
>> 
>> How about allocate ->vmemmap_tails array dynamically? If sizeof of struct
>> page is not power of two, then we could optimize away this array. Besides,
>> the original MAX_FOLIO_ORDER could work as well.
> 
> This is tricky.
> 
> We need vmemmap_tails array to be around early, in
> hugetlb_vmemmap_init_early(). By the time, we don't have slab
> functional yet.

I mean zero-size array at the end of pg_data_t, no slab is needed.

> 
> I think getting the array compile-time is the best shot.
> 
> -- 
>  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 09/15] mm/hugetlb: Refactor code around vmemmap_walk

[Muchun Song]

On 2026/1/15 22:45, Kiryl Shutsemau wrote:
> To prepare for removing fake head pages, the vmemmap_walk code is being reworked.
>
> The reuse_page and reuse_addr variables are being eliminated. There will
> no longer be an expectation regarding the reuse address in relation to
> the operated range. Instead, the caller will provide head and tail
> vmemmap pages, along with the vmemmap_start address where the head page
> is located.
>
> Currently, vmemmap_head and vmemmap_tail are set to the same page, but
> this will change in the future.
>
> The only functional change is that __hugetlb_vmemmap_optimize_folio()
> will abandon optimization if memory allocation fails.
>
> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> ---
>   mm/hugetlb_vmemmap.c | 198 ++++++++++++++++++-------------------------
>   1 file changed, 83 insertions(+), 115 deletions(-)
>
> diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
> index ba0fb1b6a5a8..2b19c2205091 100644
> --- a/mm/hugetlb_vmemmap.c
> +++ b/mm/hugetlb_vmemmap.c
> @@ -24,8 +24,9 @@
>    *
>    * @remap_pte:		called for each lowest-level entry (PTE).
>    * @nr_walked:		the number of walked pte.
> - * @reuse_page:		the page which is reused for the tail vmemmap pages.
> - * @reuse_addr:		the virtual address of the @reuse_page page.
> + * @vmemmap_start:	the start of vmemmap range, where head page is located
> + * @vmemmap_head:	the page to be installed as first in the vmemmap range
> + * @vmemmap_tail:	the page to be installed as non-first in the vmemmap range
>    * @vmemmap_pages:	the list head of the vmemmap pages that can be freed
>    *			or is mapped from.
>    * @flags:		used to modify behavior in vmemmap page table walking
> @@ -34,11 +35,14 @@
>   struct vmemmap_remap_walk {
>   	void			(*remap_pte)(pte_t *pte, unsigned long addr,
>   					     struct vmemmap_remap_walk *walk);
> +
>   	unsigned long		nr_walked;
> -	struct page		*reuse_page;
> -	unsigned long		reuse_addr;
> +	unsigned long		vmemmap_start;
> +	struct page		*vmemmap_head;
> +	struct page		*vmemmap_tail;
>   	struct list_head	*vmemmap_pages;
>   
> +
>   /* Skip the TLB flush when we split the PMD */
>   #define VMEMMAP_SPLIT_NO_TLB_FLUSH	BIT(0)
>   /* Skip the TLB flush when we remap the PTE */
> @@ -140,14 +144,7 @@ static int vmemmap_pte_entry(pte_t *pte, unsigned long addr,
>   {
>   	struct vmemmap_remap_walk *vmemmap_walk = walk->private;
>   
> -	/*
> -	 * The reuse_page is found 'first' in page table walking before
> -	 * starting remapping.
> -	 */
> -	if (!vmemmap_walk->reuse_page)
> -		vmemmap_walk->reuse_page = pte_page(ptep_get(pte));
> -	else
> -		vmemmap_walk->remap_pte(pte, addr, vmemmap_walk);
> +	vmemmap_walk->remap_pte(pte, addr, vmemmap_walk);
>   	vmemmap_walk->nr_walked++;
>   
>   	return 0;
> @@ -207,18 +204,12 @@ static void free_vmemmap_page_list(struct list_head *list)
>   static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
>   			      struct vmemmap_remap_walk *walk)
>   {
> -	/*
> -	 * Remap the tail pages as read-only to catch illegal write operation
> -	 * to the tail pages.
> -	 */
> -	pgprot_t pgprot = PAGE_KERNEL_RO;
>   	struct page *page = pte_page(ptep_get(pte));
>   	pte_t entry;
>   
>   	/* Remapping the head page requires r/w */
> -	if (unlikely(addr == walk->reuse_addr)) {
> -		pgprot = PAGE_KERNEL;
> -		list_del(&walk->reuse_page->lru);
> +	if (unlikely(addr == walk->vmemmap_start)) {
> +		list_del(&walk->vmemmap_head->lru);
>   
>   		/*
>   		 * Makes sure that preceding stores to the page contents from
> @@ -226,9 +217,16 @@ static void vmemmap_remap_pte(pte_t *pte, unsigned long addr,
>   		 * write.
>   		 */
>   		smp_wmb();
> +
> +		entry = mk_pte(walk->vmemmap_head, PAGE_KERNEL);
> +	} else {
> +		/*
> +		 * Remap the tail pages as read-only to catch illegal write
> +		 * operation to the tail pages.
> +		 */
> +		entry = mk_pte(walk->vmemmap_tail, PAGE_KERNEL_RO);
>   	}
>   
> -	entry = mk_pte(walk->reuse_page, pgprot);
>   	list_add(&page->lru, walk->vmemmap_pages);
>   	set_pte_at(&init_mm, addr, pte, entry);
>   }
> @@ -255,16 +253,13 @@ static inline void reset_struct_pages(struct page *start)
>   static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
>   				struct vmemmap_remap_walk *walk)
>   {
> -	pgprot_t pgprot = PAGE_KERNEL;
>   	struct page *page;
>   	void *to;
>   
> -	BUG_ON(pte_page(ptep_get(pte)) != walk->reuse_page);
> -
>   	page = list_first_entry(walk->vmemmap_pages, struct page, lru);
>   	list_del(&page->lru);
>   	to = page_to_virt(page);
> -	copy_page(to, (void *)walk->reuse_addr);
> +	copy_page(to, (void *)walk->vmemmap_start);
>   	reset_struct_pages(to);
>   
>   	/*
> @@ -272,7 +267,7 @@ static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
>   	 * before the set_pte_at() write.
>   	 */
>   	smp_wmb();
> -	set_pte_at(&init_mm, addr, pte, mk_pte(page, pgprot));
> +	set_pte_at(&init_mm, addr, pte, mk_pte(page, PAGE_KERNEL));
>   }
>   
>   /**
> @@ -282,33 +277,29 @@ static void vmemmap_restore_pte(pte_t *pte, unsigned long addr,
>    *             to remap.
>    * @end:       end address of the vmemmap virtual address range that we want to
>    *             remap.
> - * @reuse:     reuse address.
> - *
>    * Return: %0 on success, negative error code otherwise.
>    */
> -static int vmemmap_remap_split(unsigned long start, unsigned long end,
> -			       unsigned long reuse)
> +static int vmemmap_remap_split(unsigned long start, unsigned long end)
>   {
>   	struct vmemmap_remap_walk walk = {
>   		.remap_pte	= NULL,
> +		.vmemmap_start	= start,
>   		.flags		= VMEMMAP_SPLIT_NO_TLB_FLUSH,
>   	};
>   
> -	/* See the comment in the vmemmap_remap_free(). */
> -	BUG_ON(start - reuse != PAGE_SIZE);
> -
> -	return vmemmap_remap_range(reuse, end, &walk);
> +	return vmemmap_remap_range(start, end, &walk);
>   }
>   
>   /**
>    * vmemmap_remap_free - remap the vmemmap virtual address range [@start, @end)
> - *			to the page which @reuse is mapped to, then free vmemmap
> - *			which the range are mapped to.
> + *			to use @vmemmap_head/tail, then free vmemmap which
> + *			the range are mapped to.
>    * @start:	start address of the vmemmap virtual address range that we want
>    *		to remap.
>    * @end:	end address of the vmemmap virtual address range that we want to
>    *		remap.
> - * @reuse:	reuse address.
> + * @vmemmap_head: the page to be installed as first in the vmemmap range
> + * @vmemmap_tail: the page to be installed as non-first in the vmemmap range
>    * @vmemmap_pages: list to deposit vmemmap pages to be freed.  It is callers
>    *		responsibility to free pages.
>    * @flags:	modifications to vmemmap_remap_walk flags
> @@ -316,69 +307,40 @@ static int vmemmap_remap_split(unsigned long start, unsigned long end,
>    * Return: %0 on success, negative error code otherwise.
>    */
>   static int vmemmap_remap_free(unsigned long start, unsigned long end,
> -			      unsigned long reuse,
> +			      struct page *vmemmap_head,
> +			      struct page *vmemmap_tail,
>   			      struct list_head *vmemmap_pages,
>   			      unsigned long flags)
>   {
>   	int ret;
>   	struct vmemmap_remap_walk walk = {
>   		.remap_pte	= vmemmap_remap_pte,
> -		.reuse_addr	= reuse,
> +		.vmemmap_start	= start,
> +		.vmemmap_head	= vmemmap_head,
> +		.vmemmap_tail	= vmemmap_tail,
>   		.vmemmap_pages	= vmemmap_pages,
>   		.flags		= flags,
>   	};
> -	int nid = page_to_nid((struct page *)reuse);
> -	gfp_t gfp_mask = GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
> +
> +	ret = vmemmap_remap_range(start, end, &walk);
> +	if (!ret || !walk.nr_walked)
> +		return ret;
> +
> +	end = start + walk.nr_walked * PAGE_SIZE;
>   
>   	/*
> -	 * Allocate a new head vmemmap page to avoid breaking a contiguous
> -	 * block of struct page memory when freeing it back to page allocator
> -	 * in free_vmemmap_page_list(). This will allow the likely contiguous
> -	 * struct page backing memory to be kept contiguous and allowing for
> -	 * more allocations of hugepages. Fallback to the currently
> -	 * mapped head page in case should it fail to allocate.
> +	 * vmemmap_pages contains pages from the previous vmemmap_remap_range()
> +	 * call which failed.  These are pages which were removed from
> +	 * the vmemmap. They will be restored in the following call.
>   	 */
> -	walk.reuse_page = alloc_pages_node(nid, gfp_mask, 0);
> -	if (walk.reuse_page) {
> -		copy_page(page_to_virt(walk.reuse_page),
> -			  (void *)walk.reuse_addr);
> -		list_add(&walk.reuse_page->lru, vmemmap_pages);
> -		memmap_pages_add(1);
> -	}
> +	walk = (struct vmemmap_remap_walk) {
> +		.remap_pte	= vmemmap_restore_pte,
> +		.vmemmap_start	= start,
> +		.vmemmap_pages	= vmemmap_pages,
> +		.flags		= 0,
> +	};
>   
> -	/*
> -	 * In order to make remapping routine most efficient for the huge pages,
> -	 * the routine of vmemmap page table walking has the following rules
> -	 * (see more details from the vmemmap_pte_range()):
> -	 *
> -	 * - The range [@start, @end) and the range [@reuse, @reuse + PAGE_SIZE)
> -	 *   should be continuous.
> -	 * - The @reuse address is part of the range [@reuse, @end) that we are
> -	 *   walking which is passed to vmemmap_remap_range().
> -	 * - The @reuse address is the first in the complete range.
> -	 *
> -	 * So we need to make sure that @start and @reuse meet the above rules.
> -	 */
> -	BUG_ON(start - reuse != PAGE_SIZE);
> -
> -	ret = vmemmap_remap_range(reuse, end, &walk);
> -	if (ret && walk.nr_walked) {
> -		end = reuse + walk.nr_walked * PAGE_SIZE;
> -		/*
> -		 * vmemmap_pages contains pages from the previous
> -		 * vmemmap_remap_range call which failed.  These
> -		 * are pages which were removed from the vmemmap.
> -		 * They will be restored in the following call.
> -		 */
> -		walk = (struct vmemmap_remap_walk) {
> -			.remap_pte	= vmemmap_restore_pte,
> -			.reuse_addr	= reuse,
> -			.vmemmap_pages	= vmemmap_pages,
> -			.flags		= 0,
> -		};
> -
> -		vmemmap_remap_range(reuse, end, &walk);
> -	}
> +	vmemmap_remap_range(start, end, &walk);
>   
>   	return ret;
>   }
> @@ -415,29 +377,27 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
>    *		to remap.
>    * @end:	end address of the vmemmap virtual address range that we want to
>    *		remap.
> - * @reuse:	reuse address.
>    * @flags:	modifications to vmemmap_remap_walk flags
>    *
>    * Return: %0 on success, negative error code otherwise.
>    */
>   static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
> -			       unsigned long reuse, unsigned long flags)
> +			       unsigned long flags)
>   {
>   	LIST_HEAD(vmemmap_pages);
>   	struct vmemmap_remap_walk walk = {
>   		.remap_pte	= vmemmap_restore_pte,
> -		.reuse_addr	= reuse,
> +		.vmemmap_start	= start,
>   		.vmemmap_pages	= &vmemmap_pages,
>   		.flags		= flags,
>   	};
>   
> -	/* See the comment in the vmemmap_remap_free(). */
> -	BUG_ON(start - reuse != PAGE_SIZE);
> +	start += HUGETLB_VMEMMAP_RESERVE_SIZE;

 From the @start comment above this function, we can see that the @start
address is the starting address for the remap operation we want to perform.
However, the modification here does not conform to the semantics here.
Therefore, I suggest letting the caller __hugetlb_vmemmap_restore_folio()
pass the correct parameters itself.

The reason we previously used the reuse address here is that the
initialization of ->reuse_page was inside vmemmap_pte_entry. Now,
->reuse_page has been removed in your patch.

>   
>   	if (alloc_vmemmap_page_list(start, end, &vmemmap_pages))
>   		return -ENOMEM;
>   
> -	return vmemmap_remap_range(reuse, end, &walk);
> +	return vmemmap_remap_range(start, end, &walk);
>   }
>   
>   DEFINE_STATIC_KEY_FALSE(hugetlb_optimize_vmemmap_key);
> @@ -454,8 +414,7 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
>   					   struct folio *folio, unsigned long flags)
>   {
>   	int ret;
> -	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
> -	unsigned long vmemmap_reuse;
> +	unsigned long vmemmap_start, vmemmap_end;
>   
>   	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio), folio);
>   	VM_WARN_ON_ONCE_FOLIO(folio_ref_count(folio), folio);
> @@ -466,18 +425,16 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
>   	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
>   		synchronize_rcu();
>   
> +	vmemmap_start	= (unsigned long)folio;

(unsigned long)&folio->page？The folio is the metadata for contiguous pages,
while the page is our optimization target. Although their values are the
same now, their semantics are different.

>   	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
> -	vmemmap_reuse	= vmemmap_start;
> -	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
>   
>   	/*
>   	 * The pages which the vmemmap virtual address range [@vmemmap_start,
> -	 * @vmemmap_end) are mapped to are freed to the buddy allocator, and
> -	 * the range is mapped to the page which @vmemmap_reuse is mapped to.
> +	 * @vmemmap_end) are mapped to are freed to the buddy allocator.
>   	 * When a HugeTLB page is freed to the buddy allocator, previously
>   	 * discarded vmemmap pages must be allocated and remapping.
>   	 */
> -	ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, vmemmap_reuse, flags);
> +	ret = vmemmap_remap_alloc(vmemmap_start, vmemmap_end, flags);
>   	if (!ret) {
>   		folio_clear_hugetlb_vmemmap_optimized(folio);
>   		static_branch_dec(&hugetlb_optimize_vmemmap_key);
> @@ -565,9 +522,9 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
>   					    struct list_head *vmemmap_pages,
>   					    unsigned long flags)
>   {
> -	int ret = 0;
> -	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
> -	unsigned long vmemmap_reuse;
> +	unsigned long vmemmap_start, vmemmap_end;
> +	struct page *vmemmap_head, *vmemmap_tail;
> +	int nid, ret = 0;
>   
>   	VM_WARN_ON_ONCE_FOLIO(!folio_test_hugetlb(folio), folio);
>   	VM_WARN_ON_ONCE_FOLIO(folio_ref_count(folio), folio);
> @@ -592,18 +549,31 @@ static int __hugetlb_vmemmap_optimize_folio(const struct hstate *h,
>   	 */
>   	folio_set_hugetlb_vmemmap_optimized(folio);
>   
> +	nid = folio_nid(folio);
> +	vmemmap_head = alloc_pages_node(nid, GFP_KERNEL, 0);
> +
> +	if (!vmemmap_head) {
> +		ret = -ENOMEM;
> +		goto out;
> +	}
> +
> +	copy_page(page_to_virt(vmemmap_head), folio);
> +	list_add(&vmemmap_head->lru, vmemmap_pages);
> +	memmap_pages_add(1);
> +
> +	vmemmap_tail	= vmemmap_head;
> +	vmemmap_start	= (unsigned long)folio;

Ditto.

>   	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
> -	vmemmap_reuse	= vmemmap_start;
> -	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
>   
>   	/*
> -	 * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end)
> -	 * to the page which @vmemmap_reuse is mapped to.  Add pages previously
> -	 * mapping the range to vmemmap_pages list so that they can be freed by
> -	 * the caller.
> +	 * Remap the vmemmap virtual address range [@vmemmap_start, @vmemmap_end).
> +	 * Add pages previously mapping the range to vmemmap_pages list so that
> +	 * they can be freed by the caller.
>   	 */
> -	ret = vmemmap_remap_free(vmemmap_start, vmemmap_end, vmemmap_reuse,
> +	ret = vmemmap_remap_free(vmemmap_start, vmemmap_end,
> +				 vmemmap_head, vmemmap_tail,
>   				 vmemmap_pages, flags);
> +out:
>   	if (ret) {
>   		static_branch_dec(&hugetlb_optimize_vmemmap_key);
>   		folio_clear_hugetlb_vmemmap_optimized(folio);
> @@ -632,21 +602,19 @@ void hugetlb_vmemmap_optimize_folio(const struct hstate *h, struct folio *folio)
>   
>   static int hugetlb_vmemmap_split_folio(const struct hstate *h, struct folio *folio)
>   {
> -	unsigned long vmemmap_start = (unsigned long)&folio->page, vmemmap_end;
> -	unsigned long vmemmap_reuse;
> +	unsigned long vmemmap_start, vmemmap_end;
>   
>   	if (!vmemmap_should_optimize_folio(h, folio))
>   		return 0;
>   
> +	vmemmap_start	= (unsigned long)folio;

Same here.

>   	vmemmap_end	= vmemmap_start + hugetlb_vmemmap_size(h);
> -	vmemmap_reuse	= vmemmap_start;
> -	vmemmap_start	+= HUGETLB_VMEMMAP_RESERVE_SIZE;
>   
>   	/*
>   	 * Split PMDs on the vmemmap virtual address range [@vmemmap_start,
>   	 * @vmemmap_end]
>   	 */
> -	return vmemmap_remap_split(vmemmap_start, vmemmap_end, vmemmap_reuse);
> +	return vmemmap_remap_split(vmemmap_start, vmemmap_end);
>   }
>   
>   static void __hugetlb_vmemmap_optimize_folios(struct hstate *h,

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Kiryl Shutsemau]

On Sat, Jan 17, 2026 at 10:38:48AM +0800, Muchun Song wrote:
> 
> 
> > On Jan 16, 2026, at 23:52, Kiryl Shutsemau <kas@kernel.org> wrote:
> > 
> > On Fri, Jan 16, 2026 at 10:38:02AM +0800, Muchun Song wrote:
> >> 
> >> 
> >>> On Jan 16, 2026, at 01:23, Kiryl Shutsemau <kas@kernel.org> wrote:
> >>> 
> >>> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
> >>>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
> >>>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
> >>>>> vmemmap pages for huge pages and remapping the freed range to a single
> >>>>> page containing the struct page metadata.
> >>>>> 
> >>>>> With the new mask-based compound_info encoding (for power-of-2 struct
> >>>>> page sizes), all tail pages of the same order are now identical
> >>>>> regardless of which compound page they belong to. This means the tail
> >>>>> pages can be truly shared without fake heads.
> >>>>> 
> >>>>> Allocate a single page of initialized tail struct pages per NUMA node
> >>>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
> >>>>> of that order on the node share this tail page, mapped read-only into
> >>>>> their vmemmap. The head page remains unique per huge page.
> >>>>> 
> >>>>> This eliminates fake heads while maintaining the same memory savings,
> >>>>> and simplifies compound_head() by removing fake head detection.
> >>>>> 
> >>>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
> >>>>> ---
> >>>>> include/linux/mmzone.h | 16 ++++++++++++++-
> >>>>> mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
> >>>>> mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
> >>>>> 3 files changed, 93 insertions(+), 11 deletions(-)
> >>>>> 
> >>>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
> >>>>> index 322ed4c42cfc..2ee3eb610291 100644
> >>>>> --- a/include/linux/mmzone.h
> >>>>> +++ b/include/linux/mmzone.h
> >>>>> @@ -82,7 +82,11 @@
> >>>>>  * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
> >>>>>  * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
> >>>>>  */
> >>>>> -#define MAX_FOLIO_ORDER get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
> >>>>> +#ifdef CONFIG_64BIT
> >>>>> +#define MAX_FOLIO_ORDER (34 - PAGE_SHIFT)
> >>>>> +#else
> >>>>> +#define MAX_FOLIO_ORDER (30 - PAGE_SHIFT)
> >>>>> +#endif
> >>>> 
> >>>> Where do these magic values stem from, and how do they related to the
> >>>> comment above that clearly spells out 16G vs. 1G ?
> >>> 
> >>> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
> >>> is 1G. Subtract PAGE_SHIFT to get the order.
> >>> 
> >>> The change allows the value to be used to define NR_VMEMMAP_TAILS which
> >>> is used specify size of vmemmap_tails array.
> >> 
> >> How about allocate ->vmemmap_tails array dynamically? If sizeof of struct
> >> page is not power of two, then we could optimize away this array. Besides,
> >> the original MAX_FOLIO_ORDER could work as well.
> > 
> > This is tricky.
> > 
> > We need vmemmap_tails array to be around early, in
> > hugetlb_vmemmap_init_early(). By the time, we don't have slab
> > functional yet.
> 
> I mean zero-size array at the end of pg_data_t, no slab is needed.

For !NUMA, the struct is in BSS. See contig_page_data.

Dynamic array won't fly there.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 09/15] mm/hugetlb: Refactor code around vmemmap_walk

[Kiryl Shutsemau]

On Mon, Jan 19, 2026 at 06:04:32PM +0800, Muchun Song wrote:
> > @@ -415,29 +377,27 @@ static int alloc_vmemmap_page_list(unsigned long start, unsigned long end,
> >    *		to remap.
> >    * @end:	end address of the vmemmap virtual address range that we want to
> >    *		remap.
> > - * @reuse:	reuse address.
> >    * @flags:	modifications to vmemmap_remap_walk flags
> >    *
> >    * Return: %0 on success, negative error code otherwise.
> >    */
> >   static int vmemmap_remap_alloc(unsigned long start, unsigned long end,
> > -			       unsigned long reuse, unsigned long flags)
> > +			       unsigned long flags)
> >   {
> >   	LIST_HEAD(vmemmap_pages);
> >   	struct vmemmap_remap_walk walk = {
> >   		.remap_pte	= vmemmap_restore_pte,
> > -		.reuse_addr	= reuse,
> > +		.vmemmap_start	= start,
> >   		.vmemmap_pages	= &vmemmap_pages,
> >   		.flags		= flags,
> >   	};
> > -	/* See the comment in the vmemmap_remap_free(). */
> > -	BUG_ON(start - reuse != PAGE_SIZE);
> > +	start += HUGETLB_VMEMMAP_RESERVE_SIZE;
> 
> From the @start comment above this function, we can see that the @start
> address is the starting address for the remap operation we want to perform.
> However, the modification here does not conform to the semantics here.
> Therefore, I suggest letting the caller __hugetlb_vmemmap_restore_folio()
> pass the correct parameters itself.
> 
> The reason we previously used the reuse address here is that the
> initialization of ->reuse_page was inside vmemmap_pte_entry. Now,
> ->reuse_page has been removed in your patch.

Fair enough.

> > @@ -466,18 +425,16 @@ static int __hugetlb_vmemmap_restore_folio(const struct hstate *h,
> >   	if (flags & VMEMMAP_SYNCHRONIZE_RCU)
> >   		synchronize_rcu();
> > +	vmemmap_start	= (unsigned long)folio;
> 
> (unsigned long)&folio->page？The folio is the metadata for contiguous pages,
> while the page is our optimization target. Although their values are the
> same now, their semantics are different.

Okay.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCHv3 10/15] mm/hugetlb: Remove fake head pages

[Muchun Song]

> On Jan 19, 2026, at 23:15, Kiryl Shutsemau <kas@kernel.org> wrote:
> 
> On Sat, Jan 17, 2026 at 10:38:48AM +0800, Muchun Song wrote:
>> 
>> 
>>> On Jan 16, 2026, at 23:52, Kiryl Shutsemau <kas@kernel.org> wrote:
>>> 
>>> On Fri, Jan 16, 2026 at 10:38:02AM +0800, Muchun Song wrote:
>>>> 
>>>> 
>>>>> On Jan 16, 2026, at 01:23, Kiryl Shutsemau <kas@kernel.org> wrote:
>>>>> 
>>>>> On Thu, Jan 15, 2026 at 05:49:43PM +0100, David Hildenbrand (Red Hat) wrote:
>>>>>> On 1/15/26 15:45, Kiryl Shutsemau wrote:
>>>>>>> HugeTLB Vmemmap Optimization (HVO) reduces memory usage by freeing most
>>>>>>> vmemmap pages for huge pages and remapping the freed range to a single
>>>>>>> page containing the struct page metadata.
>>>>>>> 
>>>>>>> With the new mask-based compound_info encoding (for power-of-2 struct
>>>>>>> page sizes), all tail pages of the same order are now identical
>>>>>>> regardless of which compound page they belong to. This means the tail
>>>>>>> pages can be truly shared without fake heads.
>>>>>>> 
>>>>>>> Allocate a single page of initialized tail struct pages per NUMA node
>>>>>>> per order in the vmemmap_tails[] array in pglist_data. All huge pages
>>>>>>> of that order on the node share this tail page, mapped read-only into
>>>>>>> their vmemmap. The head page remains unique per huge page.
>>>>>>> 
>>>>>>> This eliminates fake heads while maintaining the same memory savings,
>>>>>>> and simplifies compound_head() by removing fake head detection.
>>>>>>> 
>>>>>>> Signed-off-by: Kiryl Shutsemau <kas@kernel.org>
>>>>>>> ---
>>>>>>> include/linux/mmzone.h | 16 ++++++++++++++-
>>>>>>> mm/hugetlb_vmemmap.c   | 44 ++++++++++++++++++++++++++++++++++++++++--
>>>>>>> mm/sparse-vmemmap.c    | 44 ++++++++++++++++++++++++++++++++++--------
>>>>>>> 3 files changed, 93 insertions(+), 11 deletions(-)
>>>>>>> 
>>>>>>> diff --git a/include/linux/mmzone.h b/include/linux/mmzone.h
>>>>>>> index 322ed4c42cfc..2ee3eb610291 100644
>>>>>>> --- a/include/linux/mmzone.h
>>>>>>> +++ b/include/linux/mmzone.h
>>>>>>> @@ -82,7 +82,11 @@
>>>>>>> * currently expect (see CONFIG_HAVE_GIGANTIC_FOLIOS): with hugetlb, we expect
>>>>>>> * no folios larger than 16 GiB on 64bit and 1 GiB on 32bit.
>>>>>>> */
>>>>>>> -#define MAX_FOLIO_ORDER get_order(IS_ENABLED(CONFIG_64BIT) ? SZ_16G : SZ_1G)
>>>>>>> +#ifdef CONFIG_64BIT
>>>>>>> +#define MAX_FOLIO_ORDER (34 - PAGE_SHIFT)
>>>>>>> +#else
>>>>>>> +#define MAX_FOLIO_ORDER (30 - PAGE_SHIFT)
>>>>>>> +#endif
>>>>>> 
>>>>>> Where do these magic values stem from, and how do they related to the
>>>>>> comment above that clearly spells out 16G vs. 1G ?
>>>>> 
>>>>> This doesn't change the resulting value: 1UL << 34 is 16GiB, 1UL << 30
>>>>> is 1G. Subtract PAGE_SHIFT to get the order.
>>>>> 
>>>>> The change allows the value to be used to define NR_VMEMMAP_TAILS which
>>>>> is used specify size of vmemmap_tails array.
>>>> 
>>>> How about allocate ->vmemmap_tails array dynamically? If sizeof of struct
>>>> page is not power of two, then we could optimize away this array. Besides,
>>>> the original MAX_FOLIO_ORDER could work as well.
>>> 
>>> This is tricky.
>>> 
>>> We need vmemmap_tails array to be around early, in
>>> hugetlb_vmemmap_init_early(). By the time, we don't have slab
>>> functional yet.
>> 
>> I mean zero-size array at the end of pg_data_t, no slab is needed.
> 
> For !NUMA, the struct is in BSS. See contig_page_data.

Right. I missed that.

> 
> Dynamic array won't fly there.
> 
> -- 
>  Kiryl Shutsemau / Kirill A. Shutemov