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