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 prep_transhuge_page(struct page *page);
 144 void free_transhuge_page(struct page *page);
 145
 146 bool can_split_folio(struct folio *folio, int *pextra_pins);
 147 int split_huge_page_to_list(struct page *page, struct list_head *list);
 148 static inline int split_huge_page(struct page *page)
 149 {
 150         return split_huge_page_to_list(page, NULL);
 151 }
 152 void deferred_split_folio(struct folio *folio);
 153
 154 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 155                 unsigned long address, bool freeze, struct folio *folio);
 156
 157 #define split_huge_pmd(__vma, __pmd, __address)                         \
 158         do {                                                            \
 159                 pmd_t *____pmd = (__pmd);                               \
 160                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 161                                         || pmd_devmap(*____pmd))        \
 162                         __split_huge_pmd(__vma, __pmd, __address,       \
 163                                                 false, NULL);           \
 164         }  while (0)
 165
 166
 167 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 168                 bool freeze, struct folio *folio);
 169
 170 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 171                 unsigned long address);
 172
 173 #define split_huge_pud(__vma, __pud, __address)                         \
 174         do {                                                            \
 175                 pud_t *____pud = (__pud);                               \
 176                 if (pud_trans_huge(*____pud)                            \
 177                                         || pud_devmap(*____pud))        \
 178                         __split_huge_pud(__vma, __pud, __address);      \
 179         }  while (0)
 180
 181 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 182                      int advice);
 183 int madvise_collapse(struct vm_area_struct *vma,
 184                      struct vm_area_struct **prev,
 185                      unsigned long start, unsigned long end);
 186 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 187                            unsigned long end, long adjust_next);
 188 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 189 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 190
 191 static inline int is_swap_pmd(pmd_t pmd)
 192 {
 193         return !pmd_none(pmd) && !pmd_present(pmd);
 194 }
 195
 196 /* mmap_lock must be held on entry */
 197 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 198                 struct vm_area_struct *vma)
 199 {
 200         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 201                 return __pmd_trans_huge_lock(pmd, vma);
 202         else
 203                 return NULL;
 204 }
 205 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 206                 struct vm_area_struct *vma)
 207 {
 208         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 209                 return __pud_trans_huge_lock(pud, vma);
 210         else
 211                 return NULL;
 212 }
 213
 214 /**
 215  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 216  * @folio: The folio to test
 217  */
 218 static inline bool folio_test_pmd_mappable(struct folio *folio)
 219 {
 220         return folio_order(folio) >= HPAGE_PMD_ORDER;
 221 }
 222
 223 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 224                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 225 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 226                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 227
 228 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 229
 230 extern struct page *huge_zero_page;
 231 extern unsigned long huge_zero_pfn;
 232
 233 static inline bool is_huge_zero_page(struct page *page)
 234 {
 235         return READ_ONCE(huge_zero_page) == page;
 236 }
 237
 238 static inline bool is_huge_zero_pmd(pmd_t pmd)
 239 {
 240         return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
 241 }
 242
 243 static inline bool is_huge_zero_pud(pud_t pud)
 244 {
 245         return false;
 246 }
 247
 248 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 249 void mm_put_huge_zero_page(struct mm_struct *mm);
 250
 251 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 252
 253 static inline bool thp_migration_supported(void)
 254 {
 255         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 256 }
 257
 258 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 259 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 260 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 261 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 262
 263 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 264 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 265 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 266
 267 static inline bool folio_test_pmd_mappable(struct folio *folio)
 268 {
 269         return false;
 270 }
 271
 272 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 273                 unsigned long addr)
 274 {
 275         return false;
 276 }
 277
 278 static inline bool hugepage_vma_check(struct vm_area_struct *vma,
 279                                       unsigned long vm_flags, bool smaps,
 280                                       bool in_pf, bool enforce_sysfs)
 281 {
 282         return false;
 283 }
 284
 285 static inline void prep_transhuge_page(struct page *page) {}
 286
 287 #define transparent_hugepage_flags 0UL
 288
 289 #define thp_get_unmapped_area   NULL
 290
 291 static inline bool
 292 can_split_folio(struct folio *folio, int *pextra_pins)
 293 {
 294         return false;
 295 }
 296 static inline int
 297 split_huge_page_to_list(struct page *page, struct list_head *list)
 298 {
 299         return 0;
 300 }
 301 static inline int split_huge_page(struct page *page)
 302 {
 303         return 0;
 304 }
 305 static inline void deferred_split_folio(struct folio *folio) {}
 306 #define split_huge_pmd(__vma, __pmd, __address) \
 307         do { } while (0)
 308
 309 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 310                 unsigned long address, bool freeze, struct folio *folio) {}
 311 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 312                 unsigned long address, bool freeze, struct folio *folio) {}
 313
 314 #define split_huge_pud(__vma, __pmd, __address) \
 315         do { } while (0)
 316
 317 static inline int hugepage_madvise(struct vm_area_struct *vma,
 318                                    unsigned long *vm_flags, int advice)
 319 {
 320         return -EINVAL;
 321 }
 322
 323 static inline int madvise_collapse(struct vm_area_struct *vma,
 324                                    struct vm_area_struct **prev,
 325                                    unsigned long start, unsigned long end)
 326 {
 327         return -EINVAL;
 328 }
 329
 330 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 331                                          unsigned long start,
 332                                          unsigned long end,
 333                                          long adjust_next)
 334 {
 335 }
 336 static inline int is_swap_pmd(pmd_t pmd)
 337 {
 338         return 0;
 339 }
 340 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 341                 struct vm_area_struct *vma)
 342 {
 343         return NULL;
 344 }
 345 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 346                 struct vm_area_struct *vma)
 347 {
 348         return NULL;
 349 }
 350
 351 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 352 {
 353         return 0;
 354 }
 355
 356 static inline bool is_huge_zero_page(struct page *page)
 357 {
 358         return false;
 359 }
 360
 361 static inline bool is_huge_zero_pmd(pmd_t pmd)
 362 {
 363         return false;
 364 }
 365
 366 static inline bool is_huge_zero_pud(pud_t pud)
 367 {
 368         return false;
 369 }
 370
 371 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 372 {
 373         return;
 374 }
 375
 376 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 377         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 378 {
 379         return NULL;
 380 }
 381
 382 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 383         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 384 {
 385         return NULL;
 386 }
 387
 388 static inline bool thp_migration_supported(void)
 389 {
 390         return false;
 391 }
 392 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 393
 394 static inline int split_folio_to_list(struct folio *folio,
 395                 struct list_head *list)
 396 {
 397         return split_huge_page_to_list(&folio->page, list);
 398 }
 399
 400 static inline int split_folio(struct folio *folio)
 401 {
 402         return split_folio_to_list(folio, NULL);
 403 }
 404
 405 /*
 406  * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
 407  * limitations in the implementation like arm64 MTE can override this to
 408  * false
 409  */
 410 #ifndef arch_thp_swp_supported
 411 static inline bool arch_thp_swp_supported(void)
 412 {
 413         return true;
 414 }
 415 #endif
 416
 417 #endif /* _LINUX_HUGE_MM_H */