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, pmd_t orig_pmd);
  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, pmd_t orig_pmd);
  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 vm_area_struct *vma, pmd_t *pmd, unsigned long addr,
  40                     pgprot_t newprot, unsigned long cp_flags);
  41 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
  42                                    pgprot_t pgprot, bool write);
  43
  44 /**
  45  * vmf_insert_pfn_pmd - insert a pmd size pfn
  46  * @vmf: Structure describing the fault
  47  * @pfn: pfn to insert
  48  * @pgprot: page protection to use
  49  * @write: whether it's a write fault
  50  *
  51  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
  52  *
  53  * Return: vm_fault_t value.
  54  */
  55 static inline vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn,
  56                                             bool write)
  57 {
  58         return vmf_insert_pfn_pmd_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  59 }
  60 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
  61                                    pgprot_t pgprot, bool write);
  62
  63 /**
  64  * vmf_insert_pfn_pud - insert a pud size pfn
  65  * @vmf: Structure describing the fault
  66  * @pfn: pfn to insert
  67  * @pgprot: page protection to use
  68  * @write: whether it's a write fault
  69  *
  70  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
  71  *
  72  * Return: vm_fault_t value.
  73  */
  74 static inline vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn,
  75                                             bool write)
  76 {
  77         return vmf_insert_pfn_pud_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  78 }
  79
  80 enum transparent_hugepage_flag {
  81         TRANSPARENT_HUGEPAGE_NEVER_DAX,
  82         TRANSPARENT_HUGEPAGE_FLAG,
  83         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  84         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  85         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  86         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  87         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  88         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  89         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  90 #ifdef CONFIG_DEBUG_VM
  91         TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG,
  92 #endif
  93 };
  94
  95 struct kobject;
  96 struct kobj_attribute;
  97
  98 ssize_t single_hugepage_flag_store(struct kobject *kobj,
  99                                    struct kobj_attribute *attr,
 100                                    const char *buf, size_t count,
 101                                    enum transparent_hugepage_flag flag);
 102 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 103                                   struct kobj_attribute *attr, char *buf,
 104                                   enum transparent_hugepage_flag flag);
 105 extern struct kobj_attribute shmem_enabled_attr;
 106
 107 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
 108 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
 109
 110 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 111 #define HPAGE_PMD_SHIFT PMD_SHIFT
 112 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
 113 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
 114
 115 #define HPAGE_PUD_SHIFT PUD_SHIFT
 116 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
 117 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
 118
 119 extern unsigned long transparent_hugepage_flags;
 120
 121 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 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         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 132                 return false;
 133         return true;
 134 }
 135
 136 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 137                                           unsigned long vm_flags)
 138 {
 139         /* Explicitly disabled through madvise. */
 140         if ((vm_flags & VM_NOHUGEPAGE) ||
 141             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 142                 return false;
 143         return true;
 144 }
 145
 146 /*
 147  * to be used on vmas which are known to support THP.
 148  * Use transparent_hugepage_active otherwise
 149  */
 150 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 151 {
 152
 153         /*
 154          * If the hardware/firmware marked hugepage support disabled.
 155          */
 156         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
 157                 return false;
 158
 159         if (!transhuge_vma_enabled(vma, vma->vm_flags))
 160                 return false;
 161
 162         if (vma_is_temporary_stack(vma))
 163                 return false;
 164
 165         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 166                 return true;
 167
 168         if (vma_is_dax(vma))
 169                 return true;
 170
 171         if (transparent_hugepage_flags &
 172                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 173                 return !!(vma->vm_flags & VM_HUGEPAGE);
 174
 175         return false;
 176 }
 177
 178 bool transparent_hugepage_active(struct vm_area_struct *vma);
 179
 180 #define transparent_hugepage_use_zero_page()                            \
 181         (transparent_hugepage_flags &                                   \
 182          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 183
 184 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 185                 unsigned long len, unsigned long pgoff, unsigned long flags);
 186
 187 void prep_transhuge_page(struct page *page);
 188 void free_transhuge_page(struct page *page);
 189 bool is_transparent_hugepage(struct page *page);
 190
 191 bool can_split_huge_page(struct page *page, int *pextra_pins);
 192 int split_huge_page_to_list(struct page *page, struct list_head *list);
 193 static inline int split_huge_page(struct page *page)
 194 {
 195         return split_huge_page_to_list(page, NULL);
 196 }
 197 void deferred_split_huge_page(struct page *page);
 198
 199 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 200                 unsigned long address, bool freeze, struct page *page);
 201
 202 #define split_huge_pmd(__vma, __pmd, __address)                         \
 203         do {                                                            \
 204                 pmd_t *____pmd = (__pmd);                               \
 205                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 206                                         || pmd_devmap(*____pmd))        \
 207                         __split_huge_pmd(__vma, __pmd, __address,       \
 208                                                 false, NULL);           \
 209         }  while (0)
 210
 211
 212 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 213                 bool freeze, struct page *page);
 214
 215 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 216                 unsigned long address);
 217
 218 #define split_huge_pud(__vma, __pud, __address)                         \
 219         do {                                                            \
 220                 pud_t *____pud = (__pud);                               \
 221                 if (pud_trans_huge(*____pud)                            \
 222                                         || pud_devmap(*____pud))        \
 223                         __split_huge_pud(__vma, __pud, __address);      \
 224         }  while (0)
 225
 226 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 227                      int advice);
 228 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 229                            unsigned long end, long adjust_next);
 230 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 231 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 232
 233 static inline int is_swap_pmd(pmd_t pmd)
 234 {
 235         return !pmd_none(pmd) && !pmd_present(pmd);
 236 }
 237
 238 /* mmap_lock must be held on entry */
 239 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 240                 struct vm_area_struct *vma)
 241 {
 242         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 243                 return __pmd_trans_huge_lock(pmd, vma);
 244         else
 245                 return NULL;
 246 }
 247 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 248                 struct vm_area_struct *vma)
 249 {
 250         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 251                 return __pud_trans_huge_lock(pud, vma);
 252         else
 253                 return NULL;
 254 }
 255
 256 /**
 257  * thp_head - Head page of a transparent huge page.
 258  * @page: Any page (tail, head or regular) found in the page cache.
 259  */
 260 static inline struct page *thp_head(struct page *page)
 261 {
 262         return compound_head(page);
 263 }
 264
 265 /**
 266  * thp_order - Order of a transparent huge page.
 267  * @page: Head page of a transparent huge page.
 268  */
 269 static inline unsigned int thp_order(struct page *page)
 270 {
 271         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 272         if (PageHead(page))
 273                 return HPAGE_PMD_ORDER;
 274         return 0;
 275 }
 276
 277 /**
 278  * thp_nr_pages - The number of regular pages in this huge page.
 279  * @page: The head page of a huge page.
 280  */
 281 static inline int thp_nr_pages(struct page *page)
 282 {
 283         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 284         if (PageHead(page))
 285                 return HPAGE_PMD_NR;
 286         return 1;
 287 }
 288
 289 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 290                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 291 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 292                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 293
 294 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 295
 296 extern struct page *huge_zero_page;
 297 extern unsigned long huge_zero_pfn;
 298
 299 static inline bool is_huge_zero_page(struct page *page)
 300 {
 301         return READ_ONCE(huge_zero_page) == page;
 302 }
 303
 304 static inline bool is_huge_zero_pmd(pmd_t pmd)
 305 {
 306         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 307 }
 308
 309 static inline bool is_huge_zero_pud(pud_t pud)
 310 {
 311         return false;
 312 }
 313
 314 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 315 void mm_put_huge_zero_page(struct mm_struct *mm);
 316
 317 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 318
 319 static inline bool thp_migration_supported(void)
 320 {
 321         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 322 }
 323
 324 static inline struct list_head *page_deferred_list(struct page *page)
 325 {
 326         /*
 327          * Global or memcg deferred list in the second tail pages is
 328          * occupied by compound_head.
 329          */
 330         return &page[2].deferred_list;
 331 }
 332
 333 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 334 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 335 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 336 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 337
 338 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 339 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 340 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 341
 342 static inline struct page *thp_head(struct page *page)
 343 {
 344         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 345         return page;
 346 }
 347
 348 static inline unsigned int thp_order(struct page *page)
 349 {
 350         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 351         return 0;
 352 }
 353
 354 static inline int thp_nr_pages(struct page *page)
 355 {
 356         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 357         return 1;
 358 }
 359
 360 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 361 {
 362         return false;
 363 }
 364
 365 static inline bool transparent_hugepage_active(struct vm_area_struct *vma)
 366 {
 367         return false;
 368 }
 369
 370 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 371                 unsigned long haddr)
 372 {
 373         return false;
 374 }
 375
 376 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 377                                           unsigned long vm_flags)
 378 {
 379         return false;
 380 }
 381
 382 static inline void prep_transhuge_page(struct page *page) {}
 383
 384 static inline bool is_transparent_hugepage(struct page *page)
 385 {
 386         return false;
 387 }
 388
 389 #define transparent_hugepage_flags 0UL
 390
 391 #define thp_get_unmapped_area   NULL
 392
 393 static inline bool
 394 can_split_huge_page(struct page *page, int *pextra_pins)
 395 {
 396         BUILD_BUG();
 397         return false;
 398 }
 399 static inline int
 400 split_huge_page_to_list(struct page *page, struct list_head *list)
 401 {
 402         return 0;
 403 }
 404 static inline int split_huge_page(struct page *page)
 405 {
 406         return 0;
 407 }
 408 static inline void deferred_split_huge_page(struct page *page) {}
 409 #define split_huge_pmd(__vma, __pmd, __address) \
 410         do { } while (0)
 411
 412 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 413                 unsigned long address, bool freeze, struct page *page) {}
 414 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 415                 unsigned long address, bool freeze, struct page *page) {}
 416
 417 #define split_huge_pud(__vma, __pmd, __address) \
 418         do { } while (0)
 419
 420 static inline int hugepage_madvise(struct vm_area_struct *vma,
 421                                    unsigned long *vm_flags, int advice)
 422 {
 423         BUG();
 424         return 0;
 425 }
 426 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 427                                          unsigned long start,
 428                                          unsigned long end,
 429                                          long adjust_next)
 430 {
 431 }
 432 static inline int is_swap_pmd(pmd_t pmd)
 433 {
 434         return 0;
 435 }
 436 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 437                 struct vm_area_struct *vma)
 438 {
 439         return NULL;
 440 }
 441 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 442                 struct vm_area_struct *vma)
 443 {
 444         return NULL;
 445 }
 446
 447 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 448                 pmd_t orig_pmd)
 449 {
 450         return 0;
 451 }
 452
 453 static inline bool is_huge_zero_page(struct page *page)
 454 {
 455         return false;
 456 }
 457
 458 static inline bool is_huge_zero_pmd(pmd_t pmd)
 459 {
 460         return false;
 461 }
 462
 463 static inline bool is_huge_zero_pud(pud_t pud)
 464 {
 465         return false;
 466 }
 467
 468 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 469 {
 470         return;
 471 }
 472
 473 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 474         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 475 {
 476         return NULL;
 477 }
 478
 479 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 480         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 481 {
 482         return NULL;
 483 }
 484
 485 static inline bool thp_migration_supported(void)
 486 {
 487         return false;
 488 }
 489 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 490
 491 /**
 492  * thp_size - Size of a transparent huge page.
 493  * @page: Head page of a transparent huge page.
 494  *
 495  * Return: Number of bytes in this page.
 496  */
 497 static inline unsigned long thp_size(struct page *page)
 498 {
 499         return PAGE_SIZE << thp_order(page);
 500 }
 501
 502 #endif /* _LINUX_HUGE_MM_H */