include/linux/huge_mm.h

   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 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);
  18
  19 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
  20 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 vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf);
  28 struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
  29                                    unsigned long addr, pmd_t *pmd,
  30                                    unsigned int flags);
  31 bool madvise_free_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
  32                            pmd_t *pmd, unsigned long addr, unsigned long next);
  33 int zap_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma, pmd_t *pmd,
  34                  unsigned long addr);
  35 int zap_huge_pud(struct mmu_gather *tlb, struct vm_area_struct *vma, pud_t *pud,
  36                  unsigned long addr);
  37 bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
  38                    unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd);
  39 int change_huge_pmd(struct mmu_gather *tlb, struct vm_area_struct *vma,
  40                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
  41                     unsigned long cp_flags);
  42
  43 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write);
  44 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write);
  45
  46 enum transparent_hugepage_flag {
  47         TRANSPARENT_HUGEPAGE_UNSUPPORTED,
  48         TRANSPARENT_HUGEPAGE_FLAG,
  49         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  50         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  51         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  52         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  53         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  54         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  55         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  56 };
  57
  58 struct kobject;
  59 struct kobj_attribute;
  60
  61 ssize_t single_hugepage_flag_store(struct kobject *kobj,
  62                                    struct kobj_attribute *attr,
  63                                    const char *buf, size_t count,
  64                                    enum transparent_hugepage_flag flag);
  65 ssize_t single_hugepage_flag_show(struct kobject *kobj,
  66                                   struct kobj_attribute *attr, char *buf,
  67                                   enum transparent_hugepage_flag flag);
  68 extern struct kobj_attribute shmem_enabled_attr;
  69
  70 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
  71 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
  72
  73 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  74 #define HPAGE_PMD_SHIFT PMD_SHIFT
  75 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
  76 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
  77
  78 #define HPAGE_PUD_SHIFT PUD_SHIFT
  79 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
  80 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
  81
  82 extern unsigned long transparent_hugepage_flags;
  83
  84 #define hugepage_flags_enabled()                                               \
  85         (transparent_hugepage_flags &                                  \
  86          ((1<<TRANSPARENT_HUGEPAGE_FLAG) |                     \
  87           (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)))
  88 #define hugepage_flags_always()                         \
  89         (transparent_hugepage_flags &                   \
  90          (1<<TRANSPARENT_HUGEPAGE_FLAG))
  91
  92 /*
  93  * Do the below checks:
  94  *   - For file vma, check if the linear page offset of vma is
  95  *     HPAGE_PMD_NR aligned within the file.  The hugepage is
  96  *     guaranteed to be hugepage-aligned within the file, but we must
  97  *     check that the PMD-aligned addresses in the VMA map to
  98  *     PMD-aligned offsets within the file, else the hugepage will
  99  *     not be PMD-mappable.
 100  *   - For all vmas, check if the haddr is in an aligned HPAGE_PMD_SIZE
 101  *     area.
 102  */
 103 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 104                 unsigned long addr)
 105 {
 106         unsigned long haddr;
 107
 108         /* Don't have to check pgoff for anonymous vma */
 109         if (!vma_is_anonymous(vma)) {
 110                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 111                                 HPAGE_PMD_NR))
 112                         return false;
 113         }
 114
 115         haddr = addr & HPAGE_PMD_MASK;
 116
 117         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 118                 return false;
 119         return true;
 120 }
 121
 122 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 123 {
 124         struct inode *inode;
 125
 126         if (!vma->vm_file)
 127                 return false;
 128
 129         inode = vma->vm_file->f_inode;
 130
 131         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 132                (vma->vm_flags & VM_EXEC) &&
 133                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 134 }
 135
 136 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
 137                         bool smaps, bool in_pf, bool enforce_sysfs);
 138
 139 #define transparent_hugepage_use_zero_page()                            \
 140         (transparent_hugepage_flags &                                   \
 141          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 142
 143 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 144                 unsigned long len, unsigned long pgoff, unsigned long flags);
 145
 146 void prep_transhuge_page(struct page *page);
 147 void free_transhuge_page(struct page *page);
 148
 149 bool can_split_folio(struct folio *folio, int *pextra_pins);
 150 int split_huge_page_to_list(struct page *page, struct list_head *list);
 151 static inline int split_huge_page(struct page *page)
 152 {
 153         return split_huge_page_to_list(page, NULL);
 154 }
 155 void deferred_split_folio(struct folio *folio);
 156
 157 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 158                 unsigned long address, bool freeze, struct folio *folio);
 159
 160 #define split_huge_pmd(__vma, __pmd, __address)                         \
 161         do {                                                            \
 162                 pmd_t *____pmd = (__pmd);                               \
 163                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 164                                         || pmd_devmap(*____pmd))        \
 165                         __split_huge_pmd(__vma, __pmd, __address,       \
 166                                                 false, NULL);           \
 167         }  while (0)
 168
 169
 170 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 171                 bool freeze, struct folio *folio);
 172
 173 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 174                 unsigned long address);
 175
 176 #define split_huge_pud(__vma, __pud, __address)                         \
 177         do {                                                            \
 178                 pud_t *____pud = (__pud);                               \
 179                 if (pud_trans_huge(*____pud)                            \
 180                                         || pud_devmap(*____pud))        \
 181                         __split_huge_pud(__vma, __pud, __address);      \
 182         }  while (0)
 183
 184 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 185                      int advice);
 186 int madvise_collapse(struct vm_area_struct *vma,
 187                      struct vm_area_struct **prev,
 188                      unsigned long start, unsigned long end);
 189 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 190                            unsigned long end, long adjust_next);
 191 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 192 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 193
 194 static inline int is_swap_pmd(pmd_t pmd)
 195 {
 196         return !pmd_none(pmd) && !pmd_present(pmd);
 197 }
 198
 199 /* mmap_lock must be held on entry */
 200 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 201                 struct vm_area_struct *vma)
 202 {
 203         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 204                 return __pmd_trans_huge_lock(pmd, vma);
 205         else
 206                 return NULL;
 207 }
 208 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 209                 struct vm_area_struct *vma)
 210 {
 211         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 212                 return __pud_trans_huge_lock(pud, vma);
 213         else
 214                 return NULL;
 215 }
 216
 217 /**
 218  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 219  * @folio: The folio to test
 220  */
 221 static inline bool folio_test_pmd_mappable(struct folio *folio)
 222 {
 223         return folio_order(folio) >= HPAGE_PMD_ORDER;
 224 }
 225
 226 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 227                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 228 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 229                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 230
 231 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 232
 233 extern struct page *huge_zero_page;
 234 extern unsigned long huge_zero_pfn;
 235
 236 static inline bool is_huge_zero_page(struct page *page)
 237 {
 238         return READ_ONCE(huge_zero_page) == page;
 239 }
 240
 241 static inline bool is_huge_zero_pmd(pmd_t pmd)
 242 {
 243         return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
 244 }
 245
 246 static inline bool is_huge_zero_pud(pud_t pud)
 247 {
 248         return false;
 249 }
 250
 251 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 252 void mm_put_huge_zero_page(struct mm_struct *mm);
 253
 254 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 255
 256 static inline bool thp_migration_supported(void)
 257 {
 258         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 259 }
 260
 261 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 262 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 263 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 264 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 265
 266 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 267 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 268 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 269
 270 static inline bool folio_test_pmd_mappable(struct folio *folio)
 271 {
 272         return false;
 273 }
 274
 275 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 276                 unsigned long addr)
 277 {
 278         return false;
 279 }
 280
 281 static inline bool hugepage_vma_check(struct vm_area_struct *vma,
 282                                       unsigned long vm_flags, bool smaps,
 283                                       bool in_pf, bool enforce_sysfs)
 284 {
 285         return false;
 286 }
 287
 288 static inline void prep_transhuge_page(struct page *page) {}
 289
 290 #define transparent_hugepage_flags 0UL
 291
 292 #define thp_get_unmapped_area   NULL
 293
 294 static inline bool
 295 can_split_folio(struct folio *folio, int *pextra_pins)
 296 {
 297         return false;
 298 }
 299 static inline int
 300 split_huge_page_to_list(struct page *page, struct list_head *list)
 301 {
 302         return 0;
 303 }
 304 static inline int split_huge_page(struct page *page)
 305 {
 306         return 0;
 307 }
 308 static inline void deferred_split_folio(struct folio *folio) {}
 309 #define split_huge_pmd(__vma, __pmd, __address) \
 310         do { } while (0)
 311
 312 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 313                 unsigned long address, bool freeze, struct folio *folio) {}
 314 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 315                 unsigned long address, bool freeze, struct folio *folio) {}
 316
 317 #define split_huge_pud(__vma, __pmd, __address) \
 318         do { } while (0)
 319
 320 static inline int hugepage_madvise(struct vm_area_struct *vma,
 321                                    unsigned long *vm_flags, int advice)
 322 {
 323         return -EINVAL;
 324 }
 325
 326 static inline int madvise_collapse(struct vm_area_struct *vma,
 327                                    struct vm_area_struct **prev,
 328                                    unsigned long start, unsigned long end)
 329 {
 330         return -EINVAL;
 331 }
 332
 333 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 334                                          unsigned long start,
 335                                          unsigned long end,
 336                                          long adjust_next)
 337 {
 338 }
 339 static inline int is_swap_pmd(pmd_t pmd)
 340 {
 341         return 0;
 342 }
 343 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 344                 struct vm_area_struct *vma)
 345 {
 346         return NULL;
 347 }
 348 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 349                 struct vm_area_struct *vma)
 350 {
 351         return NULL;
 352 }
 353
 354 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 355 {
 356         return 0;
 357 }
 358
 359 static inline bool is_huge_zero_page(struct page *page)
 360 {
 361         return false;
 362 }
 363
 364 static inline bool is_huge_zero_pmd(pmd_t pmd)
 365 {
 366         return false;
 367 }
 368
 369 static inline bool is_huge_zero_pud(pud_t pud)
 370 {
 371         return false;
 372 }
 373
 374 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 375 {
 376         return;
 377 }
 378
 379 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 380         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 381 {
 382         return NULL;
 383 }
 384
 385 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 386         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 387 {
 388         return NULL;
 389 }
 390
 391 static inline bool thp_migration_supported(void)
 392 {
 393         return false;
 394 }
 395 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 396
 397 static inline int split_folio_to_list(struct folio *folio,
 398                 struct list_head *list)
 399 {
 400         return split_huge_page_to_list(&folio->page, list);
 401 }
 402
 403 static inline int split_folio(struct folio *folio)
 404 {
 405         return split_folio_to_list(folio, NULL);
 406 }
 407
 408 /*
 409  * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
 410  * limitations in the implementation like arm64 MTE can override this to
 411  * false
 412  */
 413 #ifndef arch_thp_swp_supported
 414 static inline bool arch_thp_swp_supported(void)
 415 {
 416         return true;
 417 }
 418 #endif
 419
 420 #endif /* _LINUX_HUGE_MM_H */