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 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 120                 unsigned long addr)
 121 {
 122         unsigned long haddr;
 123
 124         /* Don't have to check pgoff for anonymous vma */
 125         if (!vma_is_anonymous(vma)) {
 126                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 127                                 HPAGE_PMD_NR))
 128                         return false;
 129         }
 130
 131         haddr = addr & HPAGE_PMD_MASK;
 132
 133         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 134                 return false;
 135         return true;
 136 }
 137
 138 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 139                                           unsigned long vm_flags)
 140 {
 141         /* Explicitly disabled through madvise. */
 142         if ((vm_flags & VM_NOHUGEPAGE) ||
 143             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 144                 return false;
 145         return true;
 146 }
 147
 148 /*
 149  * to be used on vmas which are known to support THP.
 150  * Use transparent_hugepage_active otherwise
 151  */
 152 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 153 {
 154
 155         /*
 156          * If the hardware/firmware marked hugepage support disabled.
 157          */
 158         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
 159                 return false;
 160
 161         if (!transhuge_vma_enabled(vma, vma->vm_flags))
 162                 return false;
 163
 164         if (vma_is_temporary_stack(vma))
 165                 return false;
 166
 167         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 168                 return true;
 169
 170         if (vma_is_dax(vma))
 171                 return true;
 172
 173         if (transparent_hugepage_flags &
 174                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 175                 return !!(vma->vm_flags & VM_HUGEPAGE);
 176
 177         return false;
 178 }
 179
 180 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 181 {
 182         struct inode *inode;
 183
 184         if (!vma->vm_file)
 185                 return false;
 186
 187         inode = vma->vm_file->f_inode;
 188
 189         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 190                (vma->vm_flags & VM_EXEC) &&
 191                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 192 }
 193
 194 bool transparent_hugepage_active(struct vm_area_struct *vma);
 195
 196 #define transparent_hugepage_use_zero_page()                            \
 197         (transparent_hugepage_flags &                                   \
 198          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 199
 200 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 201                 unsigned long len, unsigned long pgoff, unsigned long flags);
 202
 203 void prep_transhuge_page(struct page *page);
 204 void free_transhuge_page(struct page *page);
 205
 206 bool can_split_folio(struct folio *folio, int *pextra_pins);
 207 int split_huge_page_to_list(struct page *page, struct list_head *list);
 208 static inline int split_huge_page(struct page *page)
 209 {
 210         return split_huge_page_to_list(page, NULL);
 211 }
 212 void deferred_split_huge_page(struct page *page);
 213
 214 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 215                 unsigned long address, bool freeze, struct folio *folio);
 216
 217 #define split_huge_pmd(__vma, __pmd, __address)                         \
 218         do {                                                            \
 219                 pmd_t *____pmd = (__pmd);                               \
 220                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 221                                         || pmd_devmap(*____pmd))        \
 222                         __split_huge_pmd(__vma, __pmd, __address,       \
 223                                                 false, NULL);           \
 224         }  while (0)
 225
 226
 227 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 228                 bool freeze, struct folio *folio);
 229
 230 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 231                 unsigned long address);
 232
 233 #define split_huge_pud(__vma, __pud, __address)                         \
 234         do {                                                            \
 235                 pud_t *____pud = (__pud);                               \
 236                 if (pud_trans_huge(*____pud)                            \
 237                                         || pud_devmap(*____pud))        \
 238                         __split_huge_pud(__vma, __pud, __address);      \
 239         }  while (0)
 240
 241 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 242                      int advice);
 243 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 244                            unsigned long end, long adjust_next);
 245 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 246 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 247
 248 static inline int is_swap_pmd(pmd_t pmd)
 249 {
 250         return !pmd_none(pmd) && !pmd_present(pmd);
 251 }
 252
 253 /* mmap_lock must be held on entry */
 254 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 255                 struct vm_area_struct *vma)
 256 {
 257         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 258                 return __pmd_trans_huge_lock(pmd, vma);
 259         else
 260                 return NULL;
 261 }
 262 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 263                 struct vm_area_struct *vma)
 264 {
 265         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 266                 return __pud_trans_huge_lock(pud, vma);
 267         else
 268                 return NULL;
 269 }
 270
 271 /**
 272  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 273  * @folio: The folio to test
 274  */
 275 static inline bool folio_test_pmd_mappable(struct folio *folio)
 276 {
 277         return folio_order(folio) >= HPAGE_PMD_ORDER;
 278 }
 279
 280 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 281                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 282 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 283                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 284
 285 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 286
 287 extern struct page *huge_zero_page;
 288 extern unsigned long huge_zero_pfn;
 289
 290 static inline bool is_huge_zero_page(struct page *page)
 291 {
 292         return READ_ONCE(huge_zero_page) == page;
 293 }
 294
 295 static inline bool is_huge_zero_pmd(pmd_t pmd)
 296 {
 297         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 298 }
 299
 300 static inline bool is_huge_zero_pud(pud_t pud)
 301 {
 302         return false;
 303 }
 304
 305 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 306 void mm_put_huge_zero_page(struct mm_struct *mm);
 307
 308 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 309
 310 static inline bool thp_migration_supported(void)
 311 {
 312         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 313 }
 314
 315 static inline struct list_head *page_deferred_list(struct page *page)
 316 {
 317         /*
 318          * Global or memcg deferred list in the second tail pages is
 319          * occupied by compound_head.
 320          */
 321         return &page[2].deferred_list;
 322 }
 323
 324 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 325 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 326 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 327 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 328
 329 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 330 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 331 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 332
 333 static inline bool folio_test_pmd_mappable(struct folio *folio)
 334 {
 335         return false;
 336 }
 337
 338 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 339 {
 340         return false;
 341 }
 342
 343 static inline bool transparent_hugepage_active(struct vm_area_struct *vma)
 344 {
 345         return false;
 346 }
 347
 348 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 349                 unsigned long addr)
 350 {
 351         return false;
 352 }
 353
 354 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 355                                           unsigned long vm_flags)
 356 {
 357         return false;
 358 }
 359
 360 static inline void prep_transhuge_page(struct page *page) {}
 361
 362 #define transparent_hugepage_flags 0UL
 363
 364 #define thp_get_unmapped_area   NULL
 365
 366 static inline bool
 367 can_split_folio(struct folio *folio, int *pextra_pins)
 368 {
 369         return false;
 370 }
 371 static inline int
 372 split_huge_page_to_list(struct page *page, struct list_head *list)
 373 {
 374         return 0;
 375 }
 376 static inline int split_huge_page(struct page *page)
 377 {
 378         return 0;
 379 }
 380 static inline void deferred_split_huge_page(struct page *page) {}
 381 #define split_huge_pmd(__vma, __pmd, __address) \
 382         do { } while (0)
 383
 384 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 385                 unsigned long address, bool freeze, struct folio *folio) {}
 386 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 387                 unsigned long address, bool freeze, struct folio *folio) {}
 388
 389 #define split_huge_pud(__vma, __pmd, __address) \
 390         do { } while (0)
 391
 392 static inline int hugepage_madvise(struct vm_area_struct *vma,
 393                                    unsigned long *vm_flags, int advice)
 394 {
 395         BUG();
 396         return 0;
 397 }
 398 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 399                                          unsigned long start,
 400                                          unsigned long end,
 401                                          long adjust_next)
 402 {
 403 }
 404 static inline int is_swap_pmd(pmd_t pmd)
 405 {
 406         return 0;
 407 }
 408 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 409                 struct vm_area_struct *vma)
 410 {
 411         return NULL;
 412 }
 413 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 414                 struct vm_area_struct *vma)
 415 {
 416         return NULL;
 417 }
 418
 419 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 420 {
 421         return 0;
 422 }
 423
 424 static inline bool is_huge_zero_page(struct page *page)
 425 {
 426         return false;
 427 }
 428
 429 static inline bool is_huge_zero_pmd(pmd_t pmd)
 430 {
 431         return false;
 432 }
 433
 434 static inline bool is_huge_zero_pud(pud_t pud)
 435 {
 436         return false;
 437 }
 438
 439 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 440 {
 441         return;
 442 }
 443
 444 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 445         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 446 {
 447         return NULL;
 448 }
 449
 450 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 451         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 452 {
 453         return NULL;
 454 }
 455
 456 static inline bool thp_migration_supported(void)
 457 {
 458         return false;
 459 }
 460 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 461
 462 static inline int split_folio_to_list(struct folio *folio,
 463                 struct list_head *list)
 464 {
 465         return split_huge_page_to_list(&folio->page, list);
 466 }
 467
 468 #endif /* _LINUX_HUGE_MM_H */