1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _LINUX_HUGE_MM_H
3 #define _LINUX_HUGE_MM_H
5 #include <linux/mm_types.h>
7 #include <linux/fs.h> /* only for vma_is_dax() */
8 #include <linux/kobject.h>
10 vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf);
11 int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
12 pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
13 struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma);
14 void huge_pmd_set_accessed(struct vm_fault *vmf);
15 int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
16 pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
17 struct vm_area_struct *vma);
19 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
20 void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud);
22 static inline void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
27 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf);
28 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
29 pmd_t *pmd, unsigned long addr, unsigned long next);
30 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, pmd_t *pmd,
32 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, pud_t *pud,
34 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
35 unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd);
36 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
37 pmd_t *pmd, unsigned long addr, pgprot_t newprot,
38 unsigned long cp_flags);
40 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write);
41 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write);
42 vm_fault_t vmf_insert_folio_pmd(struct vm_fault *vmf, struct folio *folio,
44 vm_fault_t vmf_insert_folio_pud(struct vm_fault *vmf, struct folio *folio,
47 enum transparent_hugepage_flag {
48 TRANSPARENT_HUGEPAGE_UNSUPPORTED,
49 TRANSPARENT_HUGEPAGE_FLAG,
50 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
51 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
52 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
53 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
54 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
55 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
56 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
60 struct kobj_attribute;
62 ssize_t single_hugepage_flag_store(struct kobject *kobj,
63 struct kobj_attribute *attr,
64 const char *buf, size_t count,
65 enum transparent_hugepage_flag flag);
66 ssize_t single_hugepage_flag_show(struct kobject *kobj,
67 struct kobj_attribute *attr, char *buf,
68 enum transparent_hugepage_flag flag);
69 extern struct kobj_attribute shmem_enabled_attr;
70 extern struct kobj_attribute thpsize_shmem_enabled_attr;
73 * Mask of all large folio orders supported for anonymous THP; all orders up to
74 * and including PMD_ORDER, except order-0 (which is not "huge") and order-1
75 * (which is a limitation of the THP implementation).
77 #define THP_ORDERS_ALL_ANON ((BIT(PMD_ORDER + 1) - 1) & ~(BIT(0) | BIT(1)))
80 * Mask of all large folio orders supported for file THP. Folios in a DAX
81 * file is never split and the MAX_PAGECACHE_ORDER limit does not apply to
82 * it. Same to PFNMAPs where there's neither page* nor pagecache.
84 #define THP_ORDERS_ALL_SPECIAL \
85 (BIT(PMD_ORDER) | BIT(PUD_ORDER))
86 #define THP_ORDERS_ALL_FILE_DEFAULT \
87 ((BIT(MAX_PAGECACHE_ORDER + 1) - 1) & ~BIT(0))
90 * Mask of all large folio orders supported for THP.
92 #define THP_ORDERS_ALL \
93 (THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_SPECIAL | THP_ORDERS_ALL_FILE_DEFAULT)
95 #define TVA_SMAPS (1 << 0) /* Will be used for procfs */
96 #define TVA_IN_PF (1 << 1) /* Page fault handler */
97 #define TVA_ENFORCE_SYSFS (1 << 2) /* Obey sysfs configuration */
99 #define thp_vma_allowable_order(vma, vm_flags, tva_flags, order) \
100 (!!thp_vma_allowable_orders(vma, vm_flags, tva_flags, BIT(order)))
102 #define split_folio(f) split_folio_to_list(f, NULL)
104 #ifdef CONFIG_PGTABLE_HAS_HUGE_LEAVES
105 #define HPAGE_PMD_SHIFT PMD_SHIFT
106 #define HPAGE_PUD_SHIFT PUD_SHIFT
108 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
109 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
112 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
113 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
114 #define HPAGE_PMD_MASK (~(HPAGE_PMD_SIZE - 1))
115 #define HPAGE_PMD_SIZE ((1UL) << HPAGE_PMD_SHIFT)
117 #define HPAGE_PUD_ORDER (HPAGE_PUD_SHIFT-PAGE_SHIFT)
118 #define HPAGE_PUD_NR (1<<HPAGE_PUD_ORDER)
119 #define HPAGE_PUD_MASK (~(HPAGE_PUD_SIZE - 1))
120 #define HPAGE_PUD_SIZE ((1UL) << HPAGE_PUD_SHIFT)
122 enum mthp_stat_item {
123 MTHP_STAT_ANON_FAULT_ALLOC,
124 MTHP_STAT_ANON_FAULT_FALLBACK,
125 MTHP_STAT_ANON_FAULT_FALLBACK_CHARGE,
128 MTHP_STAT_SWPIN_FALLBACK,
129 MTHP_STAT_SWPIN_FALLBACK_CHARGE,
131 MTHP_STAT_SWPOUT_FALLBACK,
132 MTHP_STAT_SHMEM_ALLOC,
133 MTHP_STAT_SHMEM_FALLBACK,
134 MTHP_STAT_SHMEM_FALLBACK_CHARGE,
136 MTHP_STAT_SPLIT_FAILED,
137 MTHP_STAT_SPLIT_DEFERRED,
139 MTHP_STAT_NR_ANON_PARTIALLY_MAPPED,
143 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) && defined(CONFIG_SYSFS)
145 unsigned long stats[ilog2(MAX_PTRS_PER_PTE) + 1][__MTHP_STAT_COUNT];
148 DECLARE_PER_CPU(struct mthp_stat, mthp_stats);
150 static inline void mod_mthp_stat(int order, enum mthp_stat_item item, int delta)
152 if (order <= 0 || order > PMD_ORDER)
155 this_cpu_add(mthp_stats.stats[order][item], delta);
158 static inline void count_mthp_stat(int order, enum mthp_stat_item item)
160 mod_mthp_stat(order, item, 1);
164 static inline void mod_mthp_stat(int order, enum mthp_stat_item item, int delta)
168 static inline void count_mthp_stat(int order, enum mthp_stat_item item)
173 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
175 extern unsigned long transparent_hugepage_flags;
176 extern unsigned long huge_anon_orders_always;
177 extern unsigned long huge_anon_orders_madvise;
178 extern unsigned long huge_anon_orders_inherit;
180 static inline bool hugepage_global_enabled(void)
182 return transparent_hugepage_flags &
183 ((1<<TRANSPARENT_HUGEPAGE_FLAG) |
184 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG));
187 static inline bool hugepage_global_always(void)
189 return transparent_hugepage_flags &
190 (1<<TRANSPARENT_HUGEPAGE_FLAG);
193 static inline int highest_order(unsigned long orders)
195 return fls_long(orders) - 1;
198 static inline int next_order(unsigned long *orders, int prev)
200 *orders &= ~BIT(prev);
201 return highest_order(*orders);
205 * Do the below checks:
206 * - For file vma, check if the linear page offset of vma is
207 * order-aligned within the file. The hugepage is
208 * guaranteed to be order-aligned within the file, but we must
209 * check that the order-aligned addresses in the VMA map to
210 * order-aligned offsets within the file, else the hugepage will
212 * - For all vmas, check if the haddr is in an aligned hugepage
215 static inline bool thp_vma_suitable_order(struct vm_area_struct *vma,
216 unsigned long addr, int order)
218 unsigned long hpage_size = PAGE_SIZE << order;
221 /* Don't have to check pgoff for anonymous vma */
222 if (!vma_is_anonymous(vma)) {
223 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
224 hpage_size >> PAGE_SHIFT))
228 haddr = ALIGN_DOWN(addr, hpage_size);
230 if (haddr < vma->vm_start || haddr + hpage_size > vma->vm_end)
236 * Filter the bitfield of input orders to the ones suitable for use in the vma.
237 * See thp_vma_suitable_order().
238 * All orders that pass the checks are returned as a bitfield.
240 static inline unsigned long thp_vma_suitable_orders(struct vm_area_struct *vma,
241 unsigned long addr, unsigned long orders)
246 * Iterate over orders, highest to lowest, removing orders that don't
247 * meet alignment requirements from the set. Exit loop at first order
248 * that meets requirements, since all lower orders must also meet
252 order = highest_order(orders);
255 if (thp_vma_suitable_order(vma, addr, order))
257 order = next_order(&orders, order);
263 unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
264 unsigned long vm_flags,
265 unsigned long tva_flags,
266 unsigned long orders);
269 * thp_vma_allowable_orders - determine hugepage orders that are allowed for vma
270 * @vma: the vm area to check
271 * @vm_flags: use these vm_flags instead of vma->vm_flags
272 * @tva_flags: Which TVA flags to honour
273 * @orders: bitfield of all orders to consider
275 * Calculates the intersection of the requested hugepage orders and the allowed
276 * hugepage orders for the provided vma. Permitted orders are encoded as a set
277 * bit at the corresponding bit position (bit-2 corresponds to order-2, bit-3
278 * corresponds to order-3, etc). Order-0 is never considered a hugepage order.
280 * Return: bitfield of orders allowed for hugepage in the vma. 0 if no hugepage
281 * orders are allowed.
284 unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
285 unsigned long vm_flags,
286 unsigned long tva_flags,
287 unsigned long orders)
289 /* Optimization to check if required orders are enabled early. */
290 if ((tva_flags & TVA_ENFORCE_SYSFS) && vma_is_anonymous(vma)) {
291 unsigned long mask = READ_ONCE(huge_anon_orders_always);
293 if (vm_flags & VM_HUGEPAGE)
294 mask |= READ_ONCE(huge_anon_orders_madvise);
295 if (hugepage_global_always() ||
296 ((vm_flags & VM_HUGEPAGE) && hugepage_global_enabled()))
297 mask |= READ_ONCE(huge_anon_orders_inherit);
304 return __thp_vma_allowable_orders(vma, vm_flags, tva_flags, orders);
309 struct list_head node;
313 #define to_thpsize(kobj) container_of(kobj, struct thpsize, kobj)
315 #define transparent_hugepage_use_zero_page() \
316 (transparent_hugepage_flags & \
317 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
319 static inline bool vma_thp_disabled(struct vm_area_struct *vma,
320 unsigned long vm_flags)
323 * Explicitly disabled through madvise or prctl, or some
324 * architectures may disable THP for some mappings, for
327 return (vm_flags & VM_NOHUGEPAGE) ||
328 test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags);
331 static inline bool thp_disabled_by_hw(void)
333 /* If the hardware/firmware marked hugepage support disabled. */
334 return transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_UNSUPPORTED);
337 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
338 unsigned long len, unsigned long pgoff, unsigned long flags);
339 unsigned long thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr,
340 unsigned long len, unsigned long pgoff, unsigned long flags,
341 vm_flags_t vm_flags);
343 bool can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins);
344 int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
345 unsigned int new_order);
346 int min_order_for_split(struct folio *folio);
347 int split_folio_to_list(struct folio *folio, struct list_head *list);
348 bool uniform_split_supported(struct folio *folio, unsigned int new_order,
350 bool non_uniform_split_supported(struct folio *folio, unsigned int new_order,
352 int folio_split(struct folio *folio, unsigned int new_order, struct page *page,
353 struct list_head *list);
355 * try_folio_split - try to split a @folio at @page using non uniform split.
356 * @folio: folio to be split
357 * @page: split to order-0 at the given page
358 * @list: store the after-split folios
360 * Try to split a @folio at @page using non uniform split to order-0, if
361 * non uniform split is not supported, fall back to uniform split.
363 * Return: 0: split is successful, otherwise split failed.
365 static inline int try_folio_split(struct folio *folio, struct page *page,
366 struct list_head *list)
368 int ret = min_order_for_split(folio);
373 if (!non_uniform_split_supported(folio, 0, false))
374 return split_huge_page_to_list_to_order(&folio->page, list,
376 return folio_split(folio, ret, page, list);
378 static inline int split_huge_page(struct page *page)
380 struct folio *folio = page_folio(page);
381 int ret = min_order_for_split(folio);
387 * split_huge_page() locks the page before splitting and
388 * expects the same page that has been split to be locked when
389 * returned. split_folio(page_folio(page)) cannot be used here
390 * because it converts the page to folio and passes the head
393 return split_huge_page_to_list_to_order(page, NULL, ret);
395 void deferred_split_folio(struct folio *folio, bool partially_mapped);
397 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
398 unsigned long address, bool freeze);
400 #define split_huge_pmd(__vma, __pmd, __address) \
402 pmd_t *____pmd = (__pmd); \
403 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd) \
404 || pmd_devmap(*____pmd)) \
405 __split_huge_pmd(__vma, __pmd, __address, \
409 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
412 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
413 unsigned long address);
415 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
416 int change_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
417 pud_t *pudp, unsigned long addr, pgprot_t newprot,
418 unsigned long cp_flags);
421 change_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma,
422 pud_t *pudp, unsigned long addr, pgprot_t newprot,
423 unsigned long cp_flags) { return 0; }
426 #define split_huge_pud(__vma, __pud, __address) \
428 pud_t *____pud = (__pud); \
429 if (pud_trans_huge(*____pud) \
430 || pud_devmap(*____pud)) \
431 __split_huge_pud(__vma, __pud, __address); \
434 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
436 int madvise_collapse(struct vm_area_struct *vma,
437 struct vm_area_struct **prev,
438 unsigned long start, unsigned long end);
439 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
440 unsigned long end, struct vm_area_struct *next);
441 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
442 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
444 static inline int is_swap_pmd(pmd_t pmd)
446 return !pmd_none(pmd) && !pmd_present(pmd);
449 /* mmap_lock must be held on entry */
450 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
451 struct vm_area_struct *vma)
453 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
454 return __pmd_trans_huge_lock(pmd, vma);
458 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
459 struct vm_area_struct *vma)
461 if (pud_trans_huge(*pud) || pud_devmap(*pud))
462 return __pud_trans_huge_lock(pud, vma);
468 * folio_test_pmd_mappable - Can we map this folio with a PMD?
469 * @folio: The folio to test
471 static inline bool folio_test_pmd_mappable(struct folio *folio)
473 return folio_order(folio) >= HPAGE_PMD_ORDER;
476 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
477 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
479 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
481 extern struct folio *huge_zero_folio;
482 extern unsigned long huge_zero_pfn;
484 static inline bool is_huge_zero_folio(const struct folio *folio)
486 return READ_ONCE(huge_zero_folio) == folio;
489 static inline bool is_huge_zero_pmd(pmd_t pmd)
491 return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
494 struct folio *mm_get_huge_zero_folio(struct mm_struct *mm);
495 void mm_put_huge_zero_folio(struct mm_struct *mm);
497 static inline bool thp_migration_supported(void)
499 return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
502 void split_huge_pmd_locked(struct vm_area_struct *vma, unsigned long address,
503 pmd_t *pmd, bool freeze);
504 bool unmap_huge_pmd_locked(struct vm_area_struct *vma, unsigned long addr,
505 pmd_t *pmdp, struct folio *folio);
507 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
509 static inline bool folio_test_pmd_mappable(struct folio *folio)
514 static inline bool thp_vma_suitable_order(struct vm_area_struct *vma,
515 unsigned long addr, int order)
520 static inline unsigned long thp_vma_suitable_orders(struct vm_area_struct *vma,
521 unsigned long addr, unsigned long orders)
526 static inline unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
527 unsigned long vm_flags,
528 unsigned long tva_flags,
529 unsigned long orders)
534 #define transparent_hugepage_flags 0UL
536 #define thp_get_unmapped_area NULL
538 static inline unsigned long
539 thp_get_unmapped_area_vmflags(struct file *filp, unsigned long addr,
540 unsigned long len, unsigned long pgoff,
541 unsigned long flags, vm_flags_t vm_flags)
547 can_split_folio(struct folio *folio, int caller_pins, int *pextra_pins)
552 split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
553 unsigned int new_order)
557 static inline int split_huge_page(struct page *page)
562 static inline int split_folio_to_list(struct folio *folio, struct list_head *list)
567 static inline int try_folio_split(struct folio *folio, struct page *page,
568 struct list_head *list)
573 static inline void deferred_split_folio(struct folio *folio, bool partially_mapped) {}
574 #define split_huge_pmd(__vma, __pmd, __address) \
577 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
578 unsigned long address, bool freeze) {}
579 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
580 unsigned long address, bool freeze) {}
581 static inline void split_huge_pmd_locked(struct vm_area_struct *vma,
582 unsigned long address, pmd_t *pmd,
585 static inline bool unmap_huge_pmd_locked(struct vm_area_struct *vma,
586 unsigned long addr, pmd_t *pmdp,
592 #define split_huge_pud(__vma, __pmd, __address) \
595 static inline int hugepage_madvise(struct vm_area_struct *vma,
596 unsigned long *vm_flags, int advice)
601 static inline int madvise_collapse(struct vm_area_struct *vma,
602 struct vm_area_struct **prev,
603 unsigned long start, unsigned long end)
608 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
611 struct vm_area_struct *next)
614 static inline int is_swap_pmd(pmd_t pmd)
618 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
619 struct vm_area_struct *vma)
623 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
624 struct vm_area_struct *vma)
629 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
634 static inline bool is_huge_zero_folio(const struct folio *folio)
639 static inline bool is_huge_zero_pmd(pmd_t pmd)
644 static inline void mm_put_huge_zero_folio(struct mm_struct *mm)
649 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
650 unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
655 static inline bool thp_migration_supported(void)
660 static inline int highest_order(unsigned long orders)
665 static inline int next_order(unsigned long *orders, int prev)
670 static inline void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
671 unsigned long address)
675 static inline int change_huge_pud(struct mmu_gather *tlb,
676 struct vm_area_struct *vma, pud_t *pudp,
677 unsigned long addr, pgprot_t newprot,
678 unsigned long cp_flags)
682 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
684 static inline int split_folio_to_list_to_order(struct folio *folio,
685 struct list_head *list, int new_order)
687 return split_huge_page_to_list_to_order(&folio->page, list, new_order);
690 static inline int split_folio_to_order(struct folio *folio, int new_order)
692 return split_folio_to_list_to_order(folio, NULL, new_order);
695 #endif /* _LINUX_HUGE_MM_H */