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 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
  43                                    pgprot_t pgprot, bool write);
  44
  45 /**
  46  * vmf_insert_pfn_pmd - insert a pmd size pfn
  47  * @vmf: Structure describing the fault
  48  * @pfn: pfn to insert
  49  * @pgprot: page protection to use
  50  * @write: whether it's a write fault
  51  *
  52  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
  53  *
  54  * Return: vm_fault_t value.
  55  */
  56 static inline vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn,
  57                                             bool write)
  58 {
  59         return vmf_insert_pfn_pmd_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  60 }
  61 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
  62                                    pgprot_t pgprot, bool write);
  63
  64 /**
  65  * vmf_insert_pfn_pud - insert a pud size pfn
  66  * @vmf: Structure describing the fault
  67  * @pfn: pfn to insert
  68  * @pgprot: page protection to use
  69  * @write: whether it's a write fault
  70  *
  71  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
  72  *
  73  * Return: vm_fault_t value.
  74  */
  75 static inline vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn,
  76                                             bool write)
  77 {
  78         return vmf_insert_pfn_pud_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  79 }
  80
  81 enum transparent_hugepage_flag {
  82         TRANSPARENT_HUGEPAGE_NEVER_DAX,
  83         TRANSPARENT_HUGEPAGE_FLAG,
  84         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  85         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  86         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  87         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  88         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  89         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  90         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  91 };
  92
  93 struct kobject;
  94 struct kobj_attribute;
  95
  96 ssize_t single_hugepage_flag_store(struct kobject *kobj,
  97                                    struct kobj_attribute *attr,
  98                                    const char *buf, size_t count,
  99                                    enum transparent_hugepage_flag flag);
 100 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 101                                   struct kobj_attribute *attr, char *buf,
 102                                   enum transparent_hugepage_flag flag);
 103 extern struct kobj_attribute shmem_enabled_attr;
 104
 105 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
 106 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
 107
 108 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 109 #define HPAGE_PMD_SHIFT PMD_SHIFT
 110 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
 111 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
 112
 113 #define HPAGE_PUD_SHIFT PUD_SHIFT
 114 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
 115 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
 116
 117 extern unsigned long transparent_hugepage_flags;
 118
 119 #define hugepage_flags_enabled()                                               \
 120         (transparent_hugepage_flags &                                  \
 121          ((1<<TRANSPARENT_HUGEPAGE_FLAG) |                     \
 122           (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)))
 123 #define hugepage_flags_always()                         \
 124         (transparent_hugepage_flags &                   \
 125          (1<<TRANSPARENT_HUGEPAGE_FLAG))
 126
 127 /*
 128  * Do the below checks:
 129  *   - For file vma, check if the linear page offset of vma is
 130  *     HPAGE_PMD_NR aligned within the file.  The hugepage is
 131  *     guaranteed to be hugepage-aligned within the file, but we must
 132  *     check that the PMD-aligned addresses in the VMA map to
 133  *     PMD-aligned offsets within the file, else the hugepage will
 134  *     not be PMD-mappable.
 135  *   - For all vmas, check if the haddr is in an aligned HPAGE_PMD_SIZE
 136  *     area.
 137  */
 138 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 139                 unsigned long addr)
 140 {
 141         unsigned long haddr;
 142
 143         /* Don't have to check pgoff for anonymous vma */
 144         if (!vma_is_anonymous(vma)) {
 145                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 146                                 HPAGE_PMD_NR))
 147                         return false;
 148         }
 149
 150         haddr = addr & HPAGE_PMD_MASK;
 151
 152         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 153                 return false;
 154         return true;
 155 }
 156
 157 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 158 {
 159         struct inode *inode;
 160
 161         if (!vma->vm_file)
 162                 return false;
 163
 164         inode = vma->vm_file->f_inode;
 165
 166         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 167                (vma->vm_flags & VM_EXEC) &&
 168                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 169 }
 170
 171 bool hugepage_vma_check(struct vm_area_struct *vma, unsigned long vm_flags,
 172                         bool smaps, bool in_pf, bool enforce_sysfs);
 173
 174 #define transparent_hugepage_use_zero_page()                            \
 175         (transparent_hugepage_flags &                                   \
 176          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 177
 178 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 179                 unsigned long len, unsigned long pgoff, unsigned long flags);
 180
 181 void prep_transhuge_page(struct page *page);
 182 void free_transhuge_page(struct page *page);
 183
 184 bool can_split_folio(struct folio *folio, int *pextra_pins);
 185 int split_huge_page_to_list(struct page *page, struct list_head *list);
 186 static inline int split_huge_page(struct page *page)
 187 {
 188         return split_huge_page_to_list(page, NULL);
 189 }
 190 void deferred_split_huge_page(struct page *page);
 191
 192 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 193                 unsigned long address, bool freeze, struct folio *folio);
 194
 195 #define split_huge_pmd(__vma, __pmd, __address)                         \
 196         do {                                                            \
 197                 pmd_t *____pmd = (__pmd);                               \
 198                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 199                                         || pmd_devmap(*____pmd))        \
 200                         __split_huge_pmd(__vma, __pmd, __address,       \
 201                                                 false, NULL);           \
 202         }  while (0)
 203
 204
 205 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 206                 bool freeze, struct folio *folio);
 207
 208 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 209                 unsigned long address);
 210
 211 #define split_huge_pud(__vma, __pud, __address)                         \
 212         do {                                                            \
 213                 pud_t *____pud = (__pud);                               \
 214                 if (pud_trans_huge(*____pud)                            \
 215                                         || pud_devmap(*____pud))        \
 216                         __split_huge_pud(__vma, __pud, __address);      \
 217         }  while (0)
 218
 219 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 220                      int advice);
 221 int madvise_collapse(struct vm_area_struct *vma,
 222                      struct vm_area_struct **prev,
 223                      unsigned long start, unsigned long end);
 224 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 225                            unsigned long end, long adjust_next);
 226 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 227 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 228
 229 static inline int is_swap_pmd(pmd_t pmd)
 230 {
 231         return !pmd_none(pmd) && !pmd_present(pmd);
 232 }
 233
 234 /* mmap_lock must be held on entry */
 235 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 236                 struct vm_area_struct *vma)
 237 {
 238         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 239                 return __pmd_trans_huge_lock(pmd, vma);
 240         else
 241                 return NULL;
 242 }
 243 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 244                 struct vm_area_struct *vma)
 245 {
 246         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 247                 return __pud_trans_huge_lock(pud, vma);
 248         else
 249                 return NULL;
 250 }
 251
 252 /**
 253  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 254  * @folio: The folio to test
 255  */
 256 static inline bool folio_test_pmd_mappable(struct folio *folio)
 257 {
 258         return folio_order(folio) >= HPAGE_PMD_ORDER;
 259 }
 260
 261 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 262                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 263 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 264                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 265
 266 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 267
 268 extern struct page *huge_zero_page;
 269 extern unsigned long huge_zero_pfn;
 270
 271 static inline bool is_huge_zero_page(struct page *page)
 272 {
 273         return READ_ONCE(huge_zero_page) == page;
 274 }
 275
 276 static inline bool is_huge_zero_pmd(pmd_t pmd)
 277 {
 278         return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
 279 }
 280
 281 static inline bool is_huge_zero_pud(pud_t pud)
 282 {
 283         return false;
 284 }
 285
 286 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 287 void mm_put_huge_zero_page(struct mm_struct *mm);
 288
 289 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 290
 291 static inline bool thp_migration_supported(void)
 292 {
 293         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 294 }
 295
 296 static inline struct list_head *page_deferred_list(struct page *page)
 297 {
 298         /*
 299          * See organization of tail pages of compound page in
 300          * "struct page" definition.
 301          */
 302         return &page[2].deferred_list;
 303 }
 304
 305 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 306 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 307 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 308 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 309
 310 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 311 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 312 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 313
 314 static inline bool folio_test_pmd_mappable(struct folio *folio)
 315 {
 316         return false;
 317 }
 318
 319 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 320                 unsigned long addr)
 321 {
 322         return false;
 323 }
 324
 325 static inline bool hugepage_vma_check(struct vm_area_struct *vma,
 326                                       unsigned long vm_flags, bool smaps,
 327                                       bool in_pf, bool enforce_sysfs)
 328 {
 329         return false;
 330 }
 331
 332 static inline void prep_transhuge_page(struct page *page) {}
 333
 334 #define transparent_hugepage_flags 0UL
 335
 336 #define thp_get_unmapped_area   NULL
 337
 338 static inline bool
 339 can_split_folio(struct folio *folio, int *pextra_pins)
 340 {
 341         return false;
 342 }
 343 static inline int
 344 split_huge_page_to_list(struct page *page, struct list_head *list)
 345 {
 346         return 0;
 347 }
 348 static inline int split_huge_page(struct page *page)
 349 {
 350         return 0;
 351 }
 352 static inline void deferred_split_huge_page(struct page *page) {}
 353 #define split_huge_pmd(__vma, __pmd, __address) \
 354         do { } while (0)
 355
 356 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 357                 unsigned long address, bool freeze, struct folio *folio) {}
 358 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 359                 unsigned long address, bool freeze, struct folio *folio) {}
 360
 361 #define split_huge_pud(__vma, __pmd, __address) \
 362         do { } while (0)
 363
 364 static inline int hugepage_madvise(struct vm_area_struct *vma,
 365                                    unsigned long *vm_flags, int advice)
 366 {
 367         return -EINVAL;
 368 }
 369
 370 static inline int madvise_collapse(struct vm_area_struct *vma,
 371                                    struct vm_area_struct **prev,
 372                                    unsigned long start, unsigned long end)
 373 {
 374         return -EINVAL;
 375 }
 376
 377 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 378                                          unsigned long start,
 379                                          unsigned long end,
 380                                          long adjust_next)
 381 {
 382 }
 383 static inline int is_swap_pmd(pmd_t pmd)
 384 {
 385         return 0;
 386 }
 387 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 388                 struct vm_area_struct *vma)
 389 {
 390         return NULL;
 391 }
 392 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 393                 struct vm_area_struct *vma)
 394 {
 395         return NULL;
 396 }
 397
 398 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 399 {
 400         return 0;
 401 }
 402
 403 static inline bool is_huge_zero_page(struct page *page)
 404 {
 405         return false;
 406 }
 407
 408 static inline bool is_huge_zero_pmd(pmd_t pmd)
 409 {
 410         return false;
 411 }
 412
 413 static inline bool is_huge_zero_pud(pud_t pud)
 414 {
 415         return false;
 416 }
 417
 418 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 419 {
 420         return;
 421 }
 422
 423 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 424         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 425 {
 426         return NULL;
 427 }
 428
 429 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 430         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 431 {
 432         return NULL;
 433 }
 434
 435 static inline bool thp_migration_supported(void)
 436 {
 437         return false;
 438 }
 439 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 440
 441 static inline int split_folio_to_list(struct folio *folio,
 442                 struct list_head *list)
 443 {
 444         return split_huge_page_to_list(&folio->page, list);
 445 }
 446
 447 static inline int split_folio(struct folio *folio)
 448 {
 449         return split_folio_to_list(folio, NULL);
 450 }
 451
 452 /*
 453  * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
 454  * limitations in the implementation like arm64 MTE can override this to
 455  * false
 456  */
 457 #ifndef arch_thp_swp_supported
 458 static inline bool arch_thp_swp_supported(void)
 459 {
 460         return true;
 461 }
 462 #endif
 463
 464 #endif /* _LINUX_HUGE_MM_H */