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 haddr)
 121 {
 122         /* Don't have to check pgoff for anonymous vma */
 123         if (!vma_is_anonymous(vma)) {
 124                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 125                                 HPAGE_PMD_NR))
 126                         return false;
 127         }
 128
 129         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 130                 return false;
 131         return true;
 132 }
 133
 134 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 135                                           unsigned long vm_flags)
 136 {
 137         /* Explicitly disabled through madvise. */
 138         if ((vm_flags & VM_NOHUGEPAGE) ||
 139             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 140                 return false;
 141         return true;
 142 }
 143
 144 /*
 145  * to be used on vmas which are known to support THP.
 146  * Use transparent_hugepage_active otherwise
 147  */
 148 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 149 {
 150
 151         /*
 152          * If the hardware/firmware marked hugepage support disabled.
 153          */
 154         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
 155                 return false;
 156
 157         if (!transhuge_vma_enabled(vma, vma->vm_flags))
 158                 return false;
 159
 160         if (vma_is_temporary_stack(vma))
 161                 return false;
 162
 163         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 164                 return true;
 165
 166         if (vma_is_dax(vma))
 167                 return true;
 168
 169         if (transparent_hugepage_flags &
 170                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 171                 return !!(vma->vm_flags & VM_HUGEPAGE);
 172
 173         return false;
 174 }
 175
 176 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 177 {
 178         struct inode *inode;
 179
 180         if (!vma->vm_file)
 181                 return false;
 182
 183         inode = vma->vm_file->f_inode;
 184
 185         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 186                (vma->vm_flags & VM_EXEC) &&
 187                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 188 }
 189
 190 bool transparent_hugepage_active(struct vm_area_struct *vma);
 191
 192 #define transparent_hugepage_use_zero_page()                            \
 193         (transparent_hugepage_flags &                                   \
 194          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 195
 196 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 197                 unsigned long len, unsigned long pgoff, unsigned long flags);
 198
 199 void prep_transhuge_page(struct page *page);
 200 void free_transhuge_page(struct page *page);
 201
 202 bool can_split_folio(struct folio *folio, int *pextra_pins);
 203 int split_huge_page_to_list(struct page *page, struct list_head *list);
 204 static inline int split_huge_page(struct page *page)
 205 {
 206         return split_huge_page_to_list(page, NULL);
 207 }
 208 void deferred_split_huge_page(struct page *page);
 209
 210 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 211                 unsigned long address, bool freeze, struct folio *folio);
 212
 213 #define split_huge_pmd(__vma, __pmd, __address)                         \
 214         do {                                                            \
 215                 pmd_t *____pmd = (__pmd);                               \
 216                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 217                                         || pmd_devmap(*____pmd))        \
 218                         __split_huge_pmd(__vma, __pmd, __address,       \
 219                                                 false, NULL);           \
 220         }  while (0)
 221
 222
 223 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 224                 bool freeze, struct folio *folio);
 225
 226 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 227                 unsigned long address);
 228
 229 #define split_huge_pud(__vma, __pud, __address)                         \
 230         do {                                                            \
 231                 pud_t *____pud = (__pud);                               \
 232                 if (pud_trans_huge(*____pud)                            \
 233                                         || pud_devmap(*____pud))        \
 234                         __split_huge_pud(__vma, __pud, __address);      \
 235         }  while (0)
 236
 237 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 238                      int advice);
 239 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 240                            unsigned long end, long adjust_next);
 241 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 242 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 243
 244 static inline int is_swap_pmd(pmd_t pmd)
 245 {
 246         return !pmd_none(pmd) && !pmd_present(pmd);
 247 }
 248
 249 /* mmap_lock must be held on entry */
 250 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 251                 struct vm_area_struct *vma)
 252 {
 253         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 254                 return __pmd_trans_huge_lock(pmd, vma);
 255         else
 256                 return NULL;
 257 }
 258 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 259                 struct vm_area_struct *vma)
 260 {
 261         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 262                 return __pud_trans_huge_lock(pud, vma);
 263         else
 264                 return NULL;
 265 }
 266
 267 /**
 268  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 269  * @folio: The folio to test
 270  */
 271 static inline bool folio_test_pmd_mappable(struct folio *folio)
 272 {
 273         return folio_order(folio) >= HPAGE_PMD_ORDER;
 274 }
 275
 276 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 277                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 278 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 279                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 280
 281 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 282
 283 extern struct page *huge_zero_page;
 284 extern unsigned long huge_zero_pfn;
 285
 286 static inline bool is_huge_zero_page(struct page *page)
 287 {
 288         return READ_ONCE(huge_zero_page) == page;
 289 }
 290
 291 static inline bool is_huge_zero_pmd(pmd_t pmd)
 292 {
 293         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 294 }
 295
 296 static inline bool is_huge_zero_pud(pud_t pud)
 297 {
 298         return false;
 299 }
 300
 301 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 302 void mm_put_huge_zero_page(struct mm_struct *mm);
 303
 304 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 305
 306 static inline bool thp_migration_supported(void)
 307 {
 308         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 309 }
 310
 311 static inline struct list_head *page_deferred_list(struct page *page)
 312 {
 313         /*
 314          * Global or memcg deferred list in the second tail pages is
 315          * occupied by compound_head.
 316          */
 317         return &page[2].deferred_list;
 318 }
 319
 320 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 321 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 322 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 323 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 324
 325 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 326 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 327 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 328
 329 static inline bool folio_test_pmd_mappable(struct folio *folio)
 330 {
 331         return false;
 332 }
 333
 334 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 335 {
 336         return false;
 337 }
 338
 339 static inline bool transparent_hugepage_active(struct vm_area_struct *vma)
 340 {
 341         return false;
 342 }
 343
 344 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 345                 unsigned long haddr)
 346 {
 347         return false;
 348 }
 349
 350 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 351                                           unsigned long vm_flags)
 352 {
 353         return false;
 354 }
 355
 356 static inline void prep_transhuge_page(struct page *page) {}
 357
 358 #define transparent_hugepage_flags 0UL
 359
 360 #define thp_get_unmapped_area   NULL
 361
 362 static inline bool
 363 can_split_folio(struct folio *folio, int *pextra_pins)
 364 {
 365         return false;
 366 }
 367 static inline int
 368 split_huge_page_to_list(struct page *page, struct list_head *list)
 369 {
 370         return 0;
 371 }
 372 static inline int split_huge_page(struct page *page)
 373 {
 374         return 0;
 375 }
 376 static inline void deferred_split_huge_page(struct page *page) {}
 377 #define split_huge_pmd(__vma, __pmd, __address) \
 378         do { } while (0)
 379
 380 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 381                 unsigned long address, bool freeze, struct folio *folio) {}
 382 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 383                 unsigned long address, bool freeze, struct folio *folio) {}
 384
 385 #define split_huge_pud(__vma, __pmd, __address) \
 386         do { } while (0)
 387
 388 static inline int hugepage_madvise(struct vm_area_struct *vma,
 389                                    unsigned long *vm_flags, int advice)
 390 {
 391         BUG();
 392         return 0;
 393 }
 394 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 395                                          unsigned long start,
 396                                          unsigned long end,
 397                                          long adjust_next)
 398 {
 399 }
 400 static inline int is_swap_pmd(pmd_t pmd)
 401 {
 402         return 0;
 403 }
 404 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 405                 struct vm_area_struct *vma)
 406 {
 407         return NULL;
 408 }
 409 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 410                 struct vm_area_struct *vma)
 411 {
 412         return NULL;
 413 }
 414
 415 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 416 {
 417         return 0;
 418 }
 419
 420 static inline bool is_huge_zero_page(struct page *page)
 421 {
 422         return false;
 423 }
 424
 425 static inline bool is_huge_zero_pmd(pmd_t pmd)
 426 {
 427         return false;
 428 }
 429
 430 static inline bool is_huge_zero_pud(pud_t pud)
 431 {
 432         return false;
 433 }
 434
 435 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 436 {
 437         return;
 438 }
 439
 440 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 441         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 442 {
 443         return NULL;
 444 }
 445
 446 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 447         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 448 {
 449         return NULL;
 450 }
 451
 452 static inline bool thp_migration_supported(void)
 453 {
 454         return false;
 455 }
 456 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 457
 458 static inline int split_folio_to_list(struct folio *folio,
 459                 struct list_head *list)
 460 {
 461         return split_huge_page_to_list(&folio->page, list);
 462 }
 463
 464 /*
 465  * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
 466  * limitations in the implementation like arm64 MTE can override this to
 467  * false
 468  */
 469 #ifndef arch_thp_swp_supported
 470 static inline bool arch_thp_swp_supported(void)
 471 {
 472         return true;
 473 }
 474 #endif
 475
 476 #endif /* _LINUX_HUGE_MM_H */