xref: /linux-6.15/include/linux/huge_mm.h (revision da1fda28)
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 
10 extern vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf);
11 extern 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 *vma);
14 extern void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd);
15 extern 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);
18 
19 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
20 extern void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud);
21 #else
22 static inline void huge_pud_set_accessed(struct vm_fault *vmf, pud_t orig_pud)
23 {
24 }
25 #endif
26 
27 extern vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd);
28 extern struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
29 					  unsigned long addr,
30 					  pmd_t *pmd,
31 					  unsigned int flags);
32 extern bool madvise_free_huge_pmd(struct mmu_gather *tlb,
33 			struct vm_area_struct *vma,
34 			pmd_t *pmd, unsigned long addr, unsigned long next);
35 extern int zap_huge_pmd(struct mmu_gather *tlb,
36 			struct vm_area_struct *vma,
37 			pmd_t *pmd, unsigned long addr);
38 extern int zap_huge_pud(struct mmu_gather *tlb,
39 			struct vm_area_struct *vma,
40 			pud_t *pud, unsigned long addr);
41 extern int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
42 			unsigned long addr, unsigned long end,
43 			unsigned char *vec);
44 extern bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
45 			 unsigned long new_addr, unsigned long old_end,
46 			 pmd_t *old_pmd, pmd_t *new_pmd);
47 extern int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
48 			unsigned long addr, pgprot_t newprot,
49 			int prot_numa);
50 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write);
51 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write);
52 enum transparent_hugepage_flag {
53 	TRANSPARENT_HUGEPAGE_FLAG,
54 	TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
55 	TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
56 	TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
57 	TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
58 	TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
59 	TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
60 	TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
61 #ifdef CONFIG_DEBUG_VM
62 	TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG,
63 #endif
64 };
65 
66 struct kobject;
67 struct kobj_attribute;
68 
69 extern ssize_t single_hugepage_flag_store(struct kobject *kobj,
70 				 struct kobj_attribute *attr,
71 				 const char *buf, size_t count,
72 				 enum transparent_hugepage_flag flag);
73 extern ssize_t single_hugepage_flag_show(struct kobject *kobj,
74 				struct kobj_attribute *attr, char *buf,
75 				enum transparent_hugepage_flag flag);
76 extern struct kobj_attribute shmem_enabled_attr;
77 
78 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
79 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
80 
81 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
82 #define HPAGE_PMD_SHIFT PMD_SHIFT
83 #define HPAGE_PMD_SIZE	((1UL) << HPAGE_PMD_SHIFT)
84 #define HPAGE_PMD_MASK	(~(HPAGE_PMD_SIZE - 1))
85 
86 #define HPAGE_PUD_SHIFT PUD_SHIFT
87 #define HPAGE_PUD_SIZE	((1UL) << HPAGE_PUD_SHIFT)
88 #define HPAGE_PUD_MASK	(~(HPAGE_PUD_SIZE - 1))
89 
90 extern unsigned long transparent_hugepage_flags;
91 
92 /*
93  * to be used on vmas which are known to support THP.
94  * Use transparent_hugepage_enabled otherwise
95  */
96 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
97 {
98 	if (vma->vm_flags & VM_NOHUGEPAGE)
99 		return false;
100 
101 	if (vma_is_temporary_stack(vma))
102 		return false;
103 
104 	if (test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
105 		return false;
106 
107 	if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
108 		return true;
109 	/*
110 	 * For dax vmas, try to always use hugepage mappings. If the kernel does
111 	 * not support hugepages, fsdax mappings will fallback to PAGE_SIZE
112 	 * mappings, and device-dax namespaces, that try to guarantee a given
113 	 * mapping size, will fail to enable
114 	 */
115 	if (vma_is_dax(vma))
116 		return true;
117 
118 	if (transparent_hugepage_flags &
119 				(1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
120 		return !!(vma->vm_flags & VM_HUGEPAGE);
121 
122 	return false;
123 }
124 
125 bool transparent_hugepage_enabled(struct vm_area_struct *vma);
126 
127 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
128 
129 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
130 		unsigned long haddr)
131 {
132 	/* Don't have to check pgoff for anonymous vma */
133 	if (!vma_is_anonymous(vma)) {
134 		if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
135 			(vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
136 			return false;
137 	}
138 
139 	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
140 		return false;
141 	return true;
142 }
143 
144 #define transparent_hugepage_use_zero_page()				\
145 	(transparent_hugepage_flags &					\
146 	 (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
147 #ifdef CONFIG_DEBUG_VM
148 #define transparent_hugepage_debug_cow()				\
149 	(transparent_hugepage_flags &					\
150 	 (1<<TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG))
151 #else /* CONFIG_DEBUG_VM */
152 #define transparent_hugepage_debug_cow() 0
153 #endif /* CONFIG_DEBUG_VM */
154 
155 extern unsigned long thp_get_unmapped_area(struct file *filp,
156 		unsigned long addr, unsigned long len, unsigned long pgoff,
157 		unsigned long flags);
158 
159 extern void prep_transhuge_page(struct page *page);
160 extern void free_transhuge_page(struct page *page);
161 bool is_transparent_hugepage(struct page *page);
162 
163 bool can_split_huge_page(struct page *page, int *pextra_pins);
164 int split_huge_page_to_list(struct page *page, struct list_head *list);
165 static inline int split_huge_page(struct page *page)
166 {
167 	return split_huge_page_to_list(page, NULL);
168 }
169 void deferred_split_huge_page(struct page *page);
170 
171 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
172 		unsigned long address, bool freeze, struct page *page);
173 
174 #define split_huge_pmd(__vma, __pmd, __address)				\
175 	do {								\
176 		pmd_t *____pmd = (__pmd);				\
177 		if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)	\
178 					|| pmd_devmap(*____pmd))	\
179 			__split_huge_pmd(__vma, __pmd, __address,	\
180 						false, NULL);		\
181 	}  while (0)
182 
183 
184 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
185 		bool freeze, struct page *page);
186 
187 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
188 		unsigned long address);
189 
190 #define split_huge_pud(__vma, __pud, __address)				\
191 	do {								\
192 		pud_t *____pud = (__pud);				\
193 		if (pud_trans_huge(*____pud)				\
194 					|| pud_devmap(*____pud))	\
195 			__split_huge_pud(__vma, __pud, __address);	\
196 	}  while (0)
197 
198 extern int hugepage_madvise(struct vm_area_struct *vma,
199 			    unsigned long *vm_flags, int advice);
200 extern void vma_adjust_trans_huge(struct vm_area_struct *vma,
201 				    unsigned long start,
202 				    unsigned long end,
203 				    long adjust_next);
204 extern spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd,
205 		struct vm_area_struct *vma);
206 extern spinlock_t *__pud_trans_huge_lock(pud_t *pud,
207 		struct vm_area_struct *vma);
208 
209 static inline int is_swap_pmd(pmd_t pmd)
210 {
211 	return !pmd_none(pmd) && !pmd_present(pmd);
212 }
213 
214 /* mmap_sem must be held on entry */
215 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
216 		struct vm_area_struct *vma)
217 {
218 	if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
219 		return __pmd_trans_huge_lock(pmd, vma);
220 	else
221 		return NULL;
222 }
223 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
224 		struct vm_area_struct *vma)
225 {
226 	if (pud_trans_huge(*pud) || pud_devmap(*pud))
227 		return __pud_trans_huge_lock(pud, vma);
228 	else
229 		return NULL;
230 }
231 static inline int hpage_nr_pages(struct page *page)
232 {
233 	if (unlikely(PageTransHuge(page)))
234 		return HPAGE_PMD_NR;
235 	return 1;
236 }
237 
238 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
239 		pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
240 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
241 		pud_t *pud, int flags, struct dev_pagemap **pgmap);
242 
243 extern vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
244 
245 extern struct page *huge_zero_page;
246 
247 static inline bool is_huge_zero_page(struct page *page)
248 {
249 	return READ_ONCE(huge_zero_page) == page;
250 }
251 
252 static inline bool is_huge_zero_pmd(pmd_t pmd)
253 {
254 	return is_huge_zero_page(pmd_page(pmd));
255 }
256 
257 static inline bool is_huge_zero_pud(pud_t pud)
258 {
259 	return false;
260 }
261 
262 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
263 void mm_put_huge_zero_page(struct mm_struct *mm);
264 
265 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
266 
267 static inline bool thp_migration_supported(void)
268 {
269 	return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
270 }
271 
272 static inline struct list_head *page_deferred_list(struct page *page)
273 {
274 	/*
275 	 * Global or memcg deferred list in the second tail pages is
276 	 * occupied by compound_head.
277 	 */
278 	return &page[2].deferred_list;
279 }
280 
281 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
282 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
283 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
284 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
285 
286 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
287 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
288 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
289 
290 static inline int hpage_nr_pages(struct page *page)
291 {
292 	VM_BUG_ON_PAGE(PageTail(page), page);
293 	return 1;
294 }
295 
296 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
297 {
298 	return false;
299 }
300 
301 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
302 {
303 	return false;
304 }
305 
306 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
307 		unsigned long haddr)
308 {
309 	return false;
310 }
311 
312 static inline void prep_transhuge_page(struct page *page) {}
313 
314 static inline bool is_transparent_hugepage(struct page *page)
315 {
316 	return false;
317 }
318 
319 #define transparent_hugepage_flags 0UL
320 
321 #define thp_get_unmapped_area	NULL
322 
323 static inline bool
324 can_split_huge_page(struct page *page, int *pextra_pins)
325 {
326 	BUILD_BUG();
327 	return false;
328 }
329 static inline int
330 split_huge_page_to_list(struct page *page, struct list_head *list)
331 {
332 	return 0;
333 }
334 static inline int split_huge_page(struct page *page)
335 {
336 	return 0;
337 }
338 static inline void deferred_split_huge_page(struct page *page) {}
339 #define split_huge_pmd(__vma, __pmd, __address)	\
340 	do { } while (0)
341 
342 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
343 		unsigned long address, bool freeze, struct page *page) {}
344 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
345 		unsigned long address, bool freeze, struct page *page) {}
346 
347 #define split_huge_pud(__vma, __pmd, __address)	\
348 	do { } while (0)
349 
350 static inline int hugepage_madvise(struct vm_area_struct *vma,
351 				   unsigned long *vm_flags, int advice)
352 {
353 	BUG();
354 	return 0;
355 }
356 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
357 					 unsigned long start,
358 					 unsigned long end,
359 					 long adjust_next)
360 {
361 }
362 static inline int is_swap_pmd(pmd_t pmd)
363 {
364 	return 0;
365 }
366 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
367 		struct vm_area_struct *vma)
368 {
369 	return NULL;
370 }
371 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
372 		struct vm_area_struct *vma)
373 {
374 	return NULL;
375 }
376 
377 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
378 		pmd_t orig_pmd)
379 {
380 	return 0;
381 }
382 
383 static inline bool is_huge_zero_page(struct page *page)
384 {
385 	return false;
386 }
387 
388 static inline bool is_huge_zero_pud(pud_t pud)
389 {
390 	return false;
391 }
392 
393 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
394 {
395 	return;
396 }
397 
398 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
399 	unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
400 {
401 	return NULL;
402 }
403 
404 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
405 	unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
406 {
407 	return NULL;
408 }
409 
410 static inline bool thp_migration_supported(void)
411 {
412 	return false;
413 }
414 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
415 
416 #endif /* _LINUX_HUGE_MM_H */
417