1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _LINUX_HUGE_MM_H
3#define _LINUX_HUGE_MM_H
4
5#include <linux/sched/coredump.h>
6#include <linux/mm_types.h>
7
8#include <linux/fs.h> /* only for vma_is_dax() */
9
10vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf);
11int 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);
14void huge_pmd_set_accessed(struct vm_fault *vmf);
15int 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);
18
19#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
20void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud);
21#else
22static inline void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
23{
24}
25#endif
26
27vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf);
28bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
29 pmd_t *pmd, unsigned long addr, unsigned long next);
30int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, pmd_t *pmd,
31 unsigned long addr);
32int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, pud_t *pud,
33 unsigned long addr);
34bool 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);
36int 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);
39
40vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write);
41vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write);
42
43enum transparent_hugepage_flag {
44 TRANSPARENT_HUGEPAGE_UNSUPPORTED,
45 TRANSPARENT_HUGEPAGE_FLAG,
46 TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
47 TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
48 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
49 TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
50 TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
51 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
52 TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
53};
54
55struct kobject;
56struct kobj_attribute;
57
58ssize_t single_hugepage_flag_store(struct kobject *kobj,
59 struct kobj_attribute *attr,
60 const char *buf, size_t count,
61 enum transparent_hugepage_flag flag);
62ssize_t single_hugepage_flag_show(struct kobject *kobj,
63 struct kobj_attribute *attr, char *buf,
64 enum transparent_hugepage_flag flag);
65extern struct kobj_attribute shmem_enabled_attr;
66
67#define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
68#define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
69
70/*
71 * Mask of all large folio orders supported for anonymous THP; all orders up to
72 * and including PMD_ORDER, except order-0 (which is not "huge") and order-1
73 * (which is a limitation of the THP implementation).
74 */
75#define THP_ORDERS_ALL_ANON ((BIT(PMD_ORDER + 1) - 1) & ~(BIT(0) | BIT(1)))
76
77/*
78 * Mask of all large folio orders supported for file THP.
79 */
80#define THP_ORDERS_ALL_FILE (BIT(PMD_ORDER) | BIT(PUD_ORDER))
81
82/*
83 * Mask of all large folio orders supported for THP.
84 */
85#define THP_ORDERS_ALL (THP_ORDERS_ALL_ANON | THP_ORDERS_ALL_FILE)
86
87#define thp_vma_allowable_order(vma, vm_flags, smaps, in_pf, enforce_sysfs, order) \
88 (!!thp_vma_allowable_orders(vma, vm_flags, smaps, in_pf, enforce_sysfs, BIT(order)))
89
90#ifdef CONFIG_TRANSPARENT_HUGEPAGE
91#define HPAGE_PMD_SHIFT PMD_SHIFT
92#define HPAGE_PMD_SIZE ((1UL) << HPAGE_PMD_SHIFT)
93#define HPAGE_PMD_MASK (~(HPAGE_PMD_SIZE - 1))
94
95#define HPAGE_PUD_SHIFT PUD_SHIFT
96#define HPAGE_PUD_SIZE ((1UL) << HPAGE_PUD_SHIFT)
97#define HPAGE_PUD_MASK (~(HPAGE_PUD_SIZE - 1))
98
99extern unsigned long transparent_hugepage_flags;
100extern unsigned long huge_anon_orders_always;
101extern unsigned long huge_anon_orders_madvise;
102extern unsigned long huge_anon_orders_inherit;
103
104static inline bool hugepage_global_enabled(void)
105{
106 return transparent_hugepage_flags &
107 ((1<<TRANSPARENT_HUGEPAGE_FLAG) |
108 (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG));
109}
110
111static inline bool hugepage_global_always(void)
112{
113 return transparent_hugepage_flags &
114 (1<<TRANSPARENT_HUGEPAGE_FLAG);
115}
116
117static inline bool hugepage_flags_enabled(void)
118{
119 /*
120 * We cover both the anon and the file-backed case here; we must return
121 * true if globally enabled, even when all anon sizes are set to never.
122 * So we don't need to look at huge_anon_orders_inherit.
123 */
124 return hugepage_global_enabled() ||
125 huge_anon_orders_always ||
126 huge_anon_orders_madvise;
127}
128
129static inline int highest_order(unsigned long orders)
130{
131 return fls_long(l: orders) - 1;
132}
133
134static inline int next_order(unsigned long *orders, int prev)
135{
136 *orders &= ~BIT(prev);
137 return highest_order(orders: *orders);
138}
139
140/*
141 * Do the below checks:
142 * - For file vma, check if the linear page offset of vma is
143 * order-aligned within the file. The hugepage is
144 * guaranteed to be order-aligned within the file, but we must
145 * check that the order-aligned addresses in the VMA map to
146 * order-aligned offsets within the file, else the hugepage will
147 * not be mappable.
148 * - For all vmas, check if the haddr is in an aligned hugepage
149 * area.
150 */
151static inline bool thp_vma_suitable_order(struct vm_area_struct *vma,
152 unsigned long addr, int order)
153{
154 unsigned long hpage_size = PAGE_SIZE << order;
155 unsigned long haddr;
156
157 /* Don't have to check pgoff for anonymous vma */
158 if (!vma_is_anonymous(vma)) {
159 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
160 hpage_size >> PAGE_SHIFT))
161 return false;
162 }
163
164 haddr = ALIGN_DOWN(addr, hpage_size);
165
166 if (haddr < vma->vm_start || haddr + hpage_size > vma->vm_end)
167 return false;
168 return true;
169}
170
171/*
172 * Filter the bitfield of input orders to the ones suitable for use in the vma.
173 * See thp_vma_suitable_order().
174 * All orders that pass the checks are returned as a bitfield.
175 */
176static inline unsigned long thp_vma_suitable_orders(struct vm_area_struct *vma,
177 unsigned long addr, unsigned long orders)
178{
179 int order;
180
181 /*
182 * Iterate over orders, highest to lowest, removing orders that don't
183 * meet alignment requirements from the set. Exit loop at first order
184 * that meets requirements, since all lower orders must also meet
185 * requirements.
186 */
187
188 order = highest_order(orders);
189
190 while (orders) {
191 if (thp_vma_suitable_order(vma, addr, order))
192 break;
193 order = next_order(orders: &orders, prev: order);
194 }
195
196 return orders;
197}
198
199static inline bool file_thp_enabled(struct vm_area_struct *vma)
200{
201 struct inode *inode;
202
203 if (!vma->vm_file)
204 return false;
205
206 inode = vma->vm_file->f_inode;
207
208 return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
209 !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
210}
211
212unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
213 unsigned long vm_flags, bool smaps,
214 bool in_pf, bool enforce_sysfs,
215 unsigned long orders);
216
217/**
218 * thp_vma_allowable_orders - determine hugepage orders that are allowed for vma
219 * @vma: the vm area to check
220 * @vm_flags: use these vm_flags instead of vma->vm_flags
221 * @smaps: whether answer will be used for smaps file
222 * @in_pf: whether answer will be used by page fault handler
223 * @enforce_sysfs: whether sysfs config should be taken into account
224 * @orders: bitfield of all orders to consider
225 *
226 * Calculates the intersection of the requested hugepage orders and the allowed
227 * hugepage orders for the provided vma. Permitted orders are encoded as a set
228 * bit at the corresponding bit position (bit-2 corresponds to order-2, bit-3
229 * corresponds to order-3, etc). Order-0 is never considered a hugepage order.
230 *
231 * Return: bitfield of orders allowed for hugepage in the vma. 0 if no hugepage
232 * orders are allowed.
233 */
234static inline
235unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
236 unsigned long vm_flags, bool smaps,
237 bool in_pf, bool enforce_sysfs,
238 unsigned long orders)
239{
240 /* Optimization to check if required orders are enabled early. */
241 if (enforce_sysfs && vma_is_anonymous(vma)) {
242 unsigned long mask = READ_ONCE(huge_anon_orders_always);
243
244 if (vm_flags & VM_HUGEPAGE)
245 mask |= READ_ONCE(huge_anon_orders_madvise);
246 if (hugepage_global_always() ||
247 ((vm_flags & VM_HUGEPAGE) && hugepage_global_enabled()))
248 mask |= READ_ONCE(huge_anon_orders_inherit);
249
250 orders &= mask;
251 if (!orders)
252 return 0;
253 }
254
255 return __thp_vma_allowable_orders(vma, vm_flags, smaps, in_pf,
256 enforce_sysfs, orders);
257}
258
259#define transparent_hugepage_use_zero_page() \
260 (transparent_hugepage_flags & \
261 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
262
263unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
264 unsigned long len, unsigned long pgoff, unsigned long flags);
265
266void folio_prep_large_rmappable(struct folio *folio);
267bool can_split_folio(struct folio *folio, int *pextra_pins);
268int split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
269 unsigned int new_order);
270static inline int split_huge_page(struct page *page)
271{
272 return split_huge_page_to_list_to_order(page, NULL, new_order: 0);
273}
274void deferred_split_folio(struct folio *folio);
275
276void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
277 unsigned long address, bool freeze, struct folio *folio);
278
279#define split_huge_pmd(__vma, __pmd, __address) \
280 do { \
281 pmd_t *____pmd = (__pmd); \
282 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd) \
283 || pmd_devmap(*____pmd)) \
284 __split_huge_pmd(__vma, __pmd, __address, \
285 false, NULL); \
286 } while (0)
287
288
289void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
290 bool freeze, struct folio *folio);
291
292void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
293 unsigned long address);
294
295#define split_huge_pud(__vma, __pud, __address) \
296 do { \
297 pud_t *____pud = (__pud); \
298 if (pud_trans_huge(*____pud) \
299 || pud_devmap(*____pud)) \
300 __split_huge_pud(__vma, __pud, __address); \
301 } while (0)
302
303int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
304 int advice);
305int madvise_collapse(struct vm_area_struct *vma,
306 struct vm_area_struct **prev,
307 unsigned long start, unsigned long end);
308void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
309 unsigned long end, long adjust_next);
310spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
311spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
312
313static inline int is_swap_pmd(pmd_t pmd)
314{
315 return !pmd_none(pmd) && !pmd_present(pmd);
316}
317
318/* mmap_lock must be held on entry */
319static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
320 struct vm_area_struct *vma)
321{
322 if (is_swap_pmd(pmd: *pmd) || pmd_trans_huge(pmd: *pmd) || pmd_devmap(pmd: *pmd))
323 return __pmd_trans_huge_lock(pmd, vma);
324 else
325 return NULL;
326}
327static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
328 struct vm_area_struct *vma)
329{
330 if (pud_trans_huge(pud: *pud) || pud_devmap(pud: *pud))
331 return __pud_trans_huge_lock(pud, vma);
332 else
333 return NULL;
334}
335
336/**
337 * folio_test_pmd_mappable - Can we map this folio with a PMD?
338 * @folio: The folio to test
339 */
340static inline bool folio_test_pmd_mappable(struct folio *folio)
341{
342 return folio_order(folio) >= HPAGE_PMD_ORDER;
343}
344
345struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
346 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
347struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
348 pud_t *pud, int flags, struct dev_pagemap **pgmap);
349
350vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
351
352extern struct page *huge_zero_page;
353extern unsigned long huge_zero_pfn;
354
355static inline bool is_huge_zero_page(struct page *page)
356{
357 return READ_ONCE(huge_zero_page) == page;
358}
359
360static inline bool is_huge_zero_pmd(pmd_t pmd)
361{
362 return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
363}
364
365static inline bool is_huge_zero_pud(pud_t pud)
366{
367 return false;
368}
369
370struct page *mm_get_huge_zero_page(struct mm_struct *mm);
371void mm_put_huge_zero_page(struct mm_struct *mm);
372
373#define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
374
375static inline bool thp_migration_supported(void)
376{
377 return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
378}
379
380#else /* CONFIG_TRANSPARENT_HUGEPAGE */
381#define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
382#define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
383#define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
384
385#define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
386#define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
387#define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
388
389static inline bool folio_test_pmd_mappable(struct folio *folio)
390{
391 return false;
392}
393
394static inline bool thp_vma_suitable_order(struct vm_area_struct *vma,
395 unsigned long addr, int order)
396{
397 return false;
398}
399
400static inline unsigned long thp_vma_suitable_orders(struct vm_area_struct *vma,
401 unsigned long addr, unsigned long orders)
402{
403 return 0;
404}
405
406static inline unsigned long thp_vma_allowable_orders(struct vm_area_struct *vma,
407 unsigned long vm_flags, bool smaps,
408 bool in_pf, bool enforce_sysfs,
409 unsigned long orders)
410{
411 return 0;
412}
413
414static inline void folio_prep_large_rmappable(struct folio *folio) {}
415
416#define transparent_hugepage_flags 0UL
417
418#define thp_get_unmapped_area NULL
419
420static inline bool
421can_split_folio(struct folio *folio, int *pextra_pins)
422{
423 return false;
424}
425static inline int
426split_huge_page_to_list_to_order(struct page *page, struct list_head *list,
427 unsigned int new_order)
428{
429 return 0;
430}
431static inline int split_huge_page(struct page *page)
432{
433 return 0;
434}
435static inline void deferred_split_folio(struct folio *folio) {}
436#define split_huge_pmd(__vma, __pmd, __address) \
437 do { } while (0)
438
439static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
440 unsigned long address, bool freeze, struct folio *folio) {}
441static inline void split_huge_pmd_address(struct vm_area_struct *vma,
442 unsigned long address, bool freeze, struct folio *folio) {}
443
444#define split_huge_pud(__vma, __pmd, __address) \
445 do { } while (0)
446
447static inline int hugepage_madvise(struct vm_area_struct *vma,
448 unsigned long *vm_flags, int advice)
449{
450 return -EINVAL;
451}
452
453static inline int madvise_collapse(struct vm_area_struct *vma,
454 struct vm_area_struct **prev,
455 unsigned long start, unsigned long end)
456{
457 return -EINVAL;
458}
459
460static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
461 unsigned long start,
462 unsigned long end,
463 long adjust_next)
464{
465}
466static inline int is_swap_pmd(pmd_t pmd)
467{
468 return 0;
469}
470static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
471 struct vm_area_struct *vma)
472{
473 return NULL;
474}
475static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
476 struct vm_area_struct *vma)
477{
478 return NULL;
479}
480
481static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
482{
483 return 0;
484}
485
486static inline bool is_huge_zero_page(struct page *page)
487{
488 return false;
489}
490
491static inline bool is_huge_zero_pmd(pmd_t pmd)
492{
493 return false;
494}
495
496static inline bool is_huge_zero_pud(pud_t pud)
497{
498 return false;
499}
500
501static inline void mm_put_huge_zero_page(struct mm_struct *mm)
502{
503 return;
504}
505
506static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
507 unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
508{
509 return NULL;
510}
511
512static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
513 unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
514{
515 return NULL;
516}
517
518static inline bool thp_migration_supported(void)
519{
520 return false;
521}
522#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
523
524static inline int split_folio_to_list_to_order(struct folio *folio,
525 struct list_head *list, int new_order)
526{
527 return split_huge_page_to_list_to_order(page: &folio->page, list, new_order);
528}
529
530static inline int split_folio_to_order(struct folio *folio, int new_order)
531{
532 return split_folio_to_list_to_order(folio, NULL, new_order);
533}
534
535#define split_folio_to_list(f, l) split_folio_to_list_to_order(f, l, 0)
536#define split_folio(f) split_folio_to_order(f, 0)
537
538/*
539 * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
540 * limitations in the implementation like arm64 MTE can override this to
541 * false
542 */
543#ifndef arch_thp_swp_supported
544static inline bool arch_thp_swp_supported(void)
545{
546 return true;
547}
548#endif
549
550#endif /* _LINUX_HUGE_MM_H */
551

source code of linux/include/linux/huge_mm.h