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 extern vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf);
  11 extern 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 *vma);
  14 extern void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd);
  15 extern 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 extern 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 extern vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd);
  28 extern struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
  29                                           unsigned long addr,
  30                                           pmd_t *pmd,
  31                                           unsigned int flags);
  32 extern bool madvise_free_huge_pmd(struct mmu_gather *tlb,
  33                         struct vm_area_struct *vma,
  34                         pmd_t *pmd, unsigned long addr, unsigned long next);
  35 extern int zap_huge_pmd(struct mmu_gather *tlb,
  36                         struct vm_area_struct *vma,
  37                         pmd_t *pmd, unsigned long addr);
  38 extern int zap_huge_pud(struct mmu_gather *tlb,
  39                         struct vm_area_struct *vma,
  40                         pud_t *pud, unsigned long addr);
  41 extern int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
  42                         unsigned long addr, unsigned long end,
  43                         unsigned char *vec);
  44 extern bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
  45                          unsigned long new_addr,
  46                          pmd_t *old_pmd, pmd_t *new_pmd);
  47 extern int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
  48                         unsigned long addr, pgprot_t newprot,
  49                         unsigned long cp_flags);
  50 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
  51                                    pgprot_t pgprot, bool write);
  52
  53 /**
  54  * vmf_insert_pfn_pmd - insert a pmd size pfn
  55  * @vmf: Structure describing the fault
  56  * @pfn: pfn to insert
  57  * @pgprot: page protection to use
  58  * @write: whether it's a write fault
  59  *
  60  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
  61  *
  62  * Return: vm_fault_t value.
  63  */
  64 static inline vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn,
  65                                             bool write)
  66 {
  67         return vmf_insert_pfn_pmd_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  68 }
  69 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
  70                                    pgprot_t pgprot, bool write);
  71
  72 /**
  73  * vmf_insert_pfn_pud - insert a pud size pfn
  74  * @vmf: Structure describing the fault
  75  * @pfn: pfn to insert
  76  * @pgprot: page protection to use
  77  * @write: whether it's a write fault
  78  *
  79  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
  80  *
  81  * Return: vm_fault_t value.
  82  */
  83 static inline vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn,
  84                                             bool write)
  85 {
  86         return vmf_insert_pfn_pud_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  87 }
  88
  89 enum transparent_hugepage_flag {
  90         TRANSPARENT_HUGEPAGE_FLAG,
  91         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  92         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  93         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  94         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  95         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  96         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  97         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  98 #ifdef CONFIG_DEBUG_VM
  99         TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG,
 100 #endif
 101 };
 102
 103 struct kobject;
 104 struct kobj_attribute;
 105
 106 extern ssize_t single_hugepage_flag_store(struct kobject *kobj,
 107                                  struct kobj_attribute *attr,
 108                                  const char *buf, size_t count,
 109                                  enum transparent_hugepage_flag flag);
 110 extern ssize_t single_hugepage_flag_show(struct kobject *kobj,
 111                                 struct kobj_attribute *attr, char *buf,
 112                                 enum transparent_hugepage_flag flag);
 113 extern struct kobj_attribute shmem_enabled_attr;
 114
 115 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
 116 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
 117
 118 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 119 #define HPAGE_PMD_SHIFT PMD_SHIFT
 120 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
 121 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
 122
 123 #define HPAGE_PUD_SHIFT PUD_SHIFT
 124 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
 125 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
 126
 127 extern unsigned long transparent_hugepage_flags;
 128
 129 /*
 130  * to be used on vmas which are known to support THP.
 131  * Use transparent_hugepage_enabled otherwise
 132  */
 133 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 134 {
 135         if (vma->vm_flags & VM_NOHUGEPAGE)
 136                 return false;
 137
 138         if (vma_is_temporary_stack(vma))
 139                 return false;
 140
 141         if (test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 142                 return false;
 143
 144         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 145                 return true;
 146         /*
 147          * For dax vmas, try to always use hugepage mappings. If the kernel does
 148          * not support hugepages, fsdax mappings will fallback to PAGE_SIZE
 149          * mappings, and device-dax namespaces, that try to guarantee a given
 150          * mapping size, will fail to enable
 151          */
 152         if (vma_is_dax(vma))
 153                 return true;
 154
 155         if (transparent_hugepage_flags &
 156                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 157                 return !!(vma->vm_flags & VM_HUGEPAGE);
 158
 159         return false;
 160 }
 161
 162 bool transparent_hugepage_enabled(struct vm_area_struct *vma);
 163
 164 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
 165
 166 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 167                 unsigned long haddr)
 168 {
 169         /* Don't have to check pgoff for anonymous vma */
 170         if (!vma_is_anonymous(vma)) {
 171                 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
 172                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
 173                         return false;
 174         }
 175
 176         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 177                 return false;
 178         return true;
 179 }
 180
 181 #define transparent_hugepage_use_zero_page()                            \
 182         (transparent_hugepage_flags &                                   \
 183          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 184
 185 extern unsigned long thp_get_unmapped_area(struct file *filp,
 186                 unsigned long addr, unsigned long len, unsigned long pgoff,
 187                 unsigned long flags);
 188
 189 extern void prep_transhuge_page(struct page *page);
 190 extern void free_transhuge_page(struct page *page);
 191 bool is_transparent_hugepage(struct page *page);
 192
 193 bool can_split_huge_page(struct page *page, int *pextra_pins);
 194 int split_huge_page_to_list(struct page *page, struct list_head *list);
 195 static inline int split_huge_page(struct page *page)
 196 {
 197         return split_huge_page_to_list(page, NULL);
 198 }
 199 void deferred_split_huge_page(struct page *page);
 200
 201 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 202                 unsigned long address, bool freeze, struct page *page);
 203
 204 #define split_huge_pmd(__vma, __pmd, __address)                         \
 205         do {                                                            \
 206                 pmd_t *____pmd = (__pmd);                               \
 207                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 208                                         || pmd_devmap(*____pmd))        \
 209                         __split_huge_pmd(__vma, __pmd, __address,       \
 210                                                 false, NULL);           \
 211         }  while (0)
 212
 213
 214 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 215                 bool freeze, struct page *page);
 216
 217 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 218                 unsigned long address);
 219
 220 #define split_huge_pud(__vma, __pud, __address)                         \
 221         do {                                                            \
 222                 pud_t *____pud = (__pud);                               \
 223                 if (pud_trans_huge(*____pud)                            \
 224                                         || pud_devmap(*____pud))        \
 225                         __split_huge_pud(__vma, __pud, __address);      \
 226         }  while (0)
 227
 228 extern int hugepage_madvise(struct vm_area_struct *vma,
 229                             unsigned long *vm_flags, int advice);
 230 extern void vma_adjust_trans_huge(struct vm_area_struct *vma,
 231                                     unsigned long start,
 232                                     unsigned long end,
 233                                     long adjust_next);
 234 extern spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd,
 235                 struct vm_area_struct *vma);
 236 extern spinlock_t *__pud_trans_huge_lock(pud_t *pud,
 237                 struct vm_area_struct *vma);
 238
 239 static inline int is_swap_pmd(pmd_t pmd)
 240 {
 241         return !pmd_none(pmd) && !pmd_present(pmd);
 242 }
 243
 244 /* mmap_lock must be held on entry */
 245 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 246                 struct vm_area_struct *vma)
 247 {
 248         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 249                 return __pmd_trans_huge_lock(pmd, vma);
 250         else
 251                 return NULL;
 252 }
 253 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 254                 struct vm_area_struct *vma)
 255 {
 256         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 257                 return __pud_trans_huge_lock(pud, vma);
 258         else
 259                 return NULL;
 260 }
 261
 262 /**
 263  * thp_head - Head page of a transparent huge page.
 264  * @page: Any page (tail, head or regular) found in the page cache.
 265  */
 266 static inline struct page *thp_head(struct page *page)
 267 {
 268         return compound_head(page);
 269 }
 270
 271 /**
 272  * thp_order - Order of a transparent huge page.
 273  * @page: Head page of a transparent huge page.
 274  */
 275 static inline unsigned int thp_order(struct page *page)
 276 {
 277         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 278         if (PageHead(page))
 279                 return HPAGE_PMD_ORDER;
 280         return 0;
 281 }
 282
 283 /**
 284  * thp_nr_pages - The number of regular pages in this huge page.
 285  * @page: The head page of a huge page.
 286  */
 287 static inline int thp_nr_pages(struct page *page)
 288 {
 289         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 290         if (PageHead(page))
 291                 return HPAGE_PMD_NR;
 292         return 1;
 293 }
 294
 295 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 296                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 297 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 298                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 299
 300 extern vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 301
 302 extern struct page *huge_zero_page;
 303
 304 static inline bool is_huge_zero_page(struct page *page)
 305 {
 306         return READ_ONCE(huge_zero_page) == page;
 307 }
 308
 309 static inline bool is_huge_zero_pmd(pmd_t pmd)
 310 {
 311         return is_huge_zero_page(pmd_page(pmd));
 312 }
 313
 314 static inline bool is_huge_zero_pud(pud_t pud)
 315 {
 316         return false;
 317 }
 318
 319 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 320 void mm_put_huge_zero_page(struct mm_struct *mm);
 321
 322 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 323
 324 static inline bool thp_migration_supported(void)
 325 {
 326         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 327 }
 328
 329 static inline struct list_head *page_deferred_list(struct page *page)
 330 {
 331         /*
 332          * Global or memcg deferred list in the second tail pages is
 333          * occupied by compound_head.
 334          */
 335         return &page[2].deferred_list;
 336 }
 337
 338 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 339 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 340 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 341 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 342
 343 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 344 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 345 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 346
 347 static inline struct page *thp_head(struct page *page)
 348 {
 349         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 350         return page;
 351 }
 352
 353 static inline unsigned int thp_order(struct page *page)
 354 {
 355         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 356         return 0;
 357 }
 358
 359 static inline int thp_nr_pages(struct page *page)
 360 {
 361         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 362         return 1;
 363 }
 364
 365 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 366 {
 367         return false;
 368 }
 369
 370 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
 371 {
 372         return false;
 373 }
 374
 375 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 376                 unsigned long haddr)
 377 {
 378         return false;
 379 }
 380
 381 static inline void prep_transhuge_page(struct page *page) {}
 382
 383 static inline bool is_transparent_hugepage(struct page *page)
 384 {
 385         return false;
 386 }
 387
 388 #define transparent_hugepage_flags 0UL
 389
 390 #define thp_get_unmapped_area   NULL
 391
 392 static inline bool
 393 can_split_huge_page(struct page *page, int *pextra_pins)
 394 {
 395         BUILD_BUG();
 396         return false;
 397 }
 398 static inline int
 399 split_huge_page_to_list(struct page *page, struct list_head *list)
 400 {
 401         return 0;
 402 }
 403 static inline int split_huge_page(struct page *page)
 404 {
 405         return 0;
 406 }
 407 static inline void deferred_split_huge_page(struct page *page) {}
 408 #define split_huge_pmd(__vma, __pmd, __address) \
 409         do { } while (0)
 410
 411 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 412                 unsigned long address, bool freeze, struct page *page) {}
 413 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 414                 unsigned long address, bool freeze, struct page *page) {}
 415
 416 #define split_huge_pud(__vma, __pmd, __address) \
 417         do { } while (0)
 418
 419 static inline int hugepage_madvise(struct vm_area_struct *vma,
 420                                    unsigned long *vm_flags, int advice)
 421 {
 422         BUG();
 423         return 0;
 424 }
 425 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 426                                          unsigned long start,
 427                                          unsigned long end,
 428                                          long adjust_next)
 429 {
 430 }
 431 static inline int is_swap_pmd(pmd_t pmd)
 432 {
 433         return 0;
 434 }
 435 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 436                 struct vm_area_struct *vma)
 437 {
 438         return NULL;
 439 }
 440 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 441                 struct vm_area_struct *vma)
 442 {
 443         return NULL;
 444 }
 445
 446 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 447                 pmd_t orig_pmd)
 448 {
 449         return 0;
 450 }
 451
 452 static inline bool is_huge_zero_page(struct page *page)
 453 {
 454         return false;
 455 }
 456
 457 static inline bool is_huge_zero_pud(pud_t pud)
 458 {
 459         return false;
 460 }
 461
 462 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 463 {
 464         return;
 465 }
 466
 467 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 468         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 469 {
 470         return NULL;
 471 }
 472
 473 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 474         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 475 {
 476         return NULL;
 477 }
 478
 479 static inline bool thp_migration_supported(void)
 480 {
 481         return false;
 482 }
 483 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 484
 485 /**
 486  * thp_size - Size of a transparent huge page.
 487  * @page: Head page of a transparent huge page.
 488  *
 489  * Return: Number of bytes in this page.
 490  */
 491 static inline unsigned long thp_size(struct page *page)
 492 {
 493         return PAGE_SIZE << thp_order(page);
 494 }
 495
 496 #endif /* _LINUX_HUGE_MM_H */