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