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, unsigned long old_end,
  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                         int prot_numa);
  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 #ifdef CONFIG_DEBUG_VM
 185 #define transparent_hugepage_debug_cow()                                \
 186         (transparent_hugepage_flags &                                   \
 187          (1<<TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG))
 188 #else /* CONFIG_DEBUG_VM */
 189 #define transparent_hugepage_debug_cow() 0
 190 #endif /* CONFIG_DEBUG_VM */
 191
 192 extern unsigned long thp_get_unmapped_area(struct file *filp,
 193                 unsigned long addr, unsigned long len, unsigned long pgoff,
 194                 unsigned long flags);
 195
 196 extern void prep_transhuge_page(struct page *page);
 197 extern void free_transhuge_page(struct page *page);
 198 bool is_transparent_hugepage(struct page *page);
 199
 200 bool can_split_huge_page(struct page *page, int *pextra_pins);
 201 int split_huge_page_to_list(struct page *page, struct list_head *list);
 202 static inline int split_huge_page(struct page *page)
 203 {
 204         return split_huge_page_to_list(page, NULL);
 205 }
 206 void deferred_split_huge_page(struct page *page);
 207
 208 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 209                 unsigned long address, bool freeze, struct page *page);
 210
 211 #define split_huge_pmd(__vma, __pmd, __address)                         \
 212         do {                                                            \
 213                 pmd_t *____pmd = (__pmd);                               \
 214                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 215                                         || pmd_devmap(*____pmd))        \
 216                         __split_huge_pmd(__vma, __pmd, __address,       \
 217                                                 false, NULL);           \
 218         }  while (0)
 219
 220
 221 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 222                 bool freeze, struct page *page);
 223
 224 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 225                 unsigned long address);
 226
 227 #define split_huge_pud(__vma, __pud, __address)                         \
 228         do {                                                            \
 229                 pud_t *____pud = (__pud);                               \
 230                 if (pud_trans_huge(*____pud)                            \
 231                                         || pud_devmap(*____pud))        \
 232                         __split_huge_pud(__vma, __pud, __address);      \
 233         }  while (0)
 234
 235 extern int hugepage_madvise(struct vm_area_struct *vma,
 236                             unsigned long *vm_flags, int advice);
 237 extern void vma_adjust_trans_huge(struct vm_area_struct *vma,
 238                                     unsigned long start,
 239                                     unsigned long end,
 240                                     long adjust_next);
 241 extern spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd,
 242                 struct vm_area_struct *vma);
 243 extern spinlock_t *__pud_trans_huge_lock(pud_t *pud,
 244                 struct vm_area_struct *vma);
 245
 246 static inline int is_swap_pmd(pmd_t pmd)
 247 {
 248         return !pmd_none(pmd) && !pmd_present(pmd);
 249 }
 250
 251 /* mmap_sem must be held on entry */
 252 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 253                 struct vm_area_struct *vma)
 254 {
 255         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 256                 return __pmd_trans_huge_lock(pmd, vma);
 257         else
 258                 return NULL;
 259 }
 260 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 261                 struct vm_area_struct *vma)
 262 {
 263         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 264                 return __pud_trans_huge_lock(pud, vma);
 265         else
 266                 return NULL;
 267 }
 268 static inline int hpage_nr_pages(struct page *page)
 269 {
 270         if (unlikely(PageTransHuge(page)))
 271                 return HPAGE_PMD_NR;
 272         return 1;
 273 }
 274
 275 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 276                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 277 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 278                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 279
 280 extern vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 281
 282 extern struct page *huge_zero_page;
 283
 284 static inline bool is_huge_zero_page(struct page *page)
 285 {
 286         return READ_ONCE(huge_zero_page) == page;
 287 }
 288
 289 static inline bool is_huge_zero_pmd(pmd_t pmd)
 290 {
 291         return is_huge_zero_page(pmd_page(pmd));
 292 }
 293
 294 static inline bool is_huge_zero_pud(pud_t pud)
 295 {
 296         return false;
 297 }
 298
 299 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 300 void mm_put_huge_zero_page(struct mm_struct *mm);
 301
 302 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 303
 304 static inline bool thp_migration_supported(void)
 305 {
 306         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 307 }
 308
 309 static inline struct list_head *page_deferred_list(struct page *page)
 310 {
 311         /*
 312          * Global or memcg deferred list in the second tail pages is
 313          * occupied by compound_head.
 314          */
 315         return &page[2].deferred_list;
 316 }
 317
 318 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 319 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 320 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 321 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 322
 323 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 324 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 325 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 326
 327 static inline int hpage_nr_pages(struct page *page)
 328 {
 329         VM_BUG_ON_PAGE(PageTail(page), page);
 330         return 1;
 331 }
 332
 333 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 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 transhuge_vma_suitable(struct vm_area_struct *vma,
 344                 unsigned long haddr)
 345 {
 346         return false;
 347 }
 348
 349 static inline void prep_transhuge_page(struct page *page) {}
 350
 351 static inline bool is_transparent_hugepage(struct page *page)
 352 {
 353         return false;
 354 }
 355
 356 #define transparent_hugepage_flags 0UL
 357
 358 #define thp_get_unmapped_area   NULL
 359
 360 static inline bool
 361 can_split_huge_page(struct page *page, int *pextra_pins)
 362 {
 363         BUILD_BUG();
 364         return false;
 365 }
 366 static inline int
 367 split_huge_page_to_list(struct page *page, struct list_head *list)
 368 {
 369         return 0;
 370 }
 371 static inline int split_huge_page(struct page *page)
 372 {
 373         return 0;
 374 }
 375 static inline void deferred_split_huge_page(struct page *page) {}
 376 #define split_huge_pmd(__vma, __pmd, __address) \
 377         do { } while (0)
 378
 379 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 380                 unsigned long address, bool freeze, struct page *page) {}
 381 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 382                 unsigned long address, bool freeze, struct page *page) {}
 383
 384 #define split_huge_pud(__vma, __pmd, __address) \
 385         do { } while (0)
 386
 387 static inline int hugepage_madvise(struct vm_area_struct *vma,
 388                                    unsigned long *vm_flags, int advice)
 389 {
 390         BUG();
 391         return 0;
 392 }
 393 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 394                                          unsigned long start,
 395                                          unsigned long end,
 396                                          long adjust_next)
 397 {
 398 }
 399 static inline int is_swap_pmd(pmd_t pmd)
 400 {
 401         return 0;
 402 }
 403 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 404                 struct vm_area_struct *vma)
 405 {
 406         return NULL;
 407 }
 408 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 409                 struct vm_area_struct *vma)
 410 {
 411         return NULL;
 412 }
 413
 414 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 415                 pmd_t orig_pmd)
 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_pud(pud_t pud)
 426 {
 427         return false;
 428 }
 429
 430 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 431 {
 432         return;
 433 }
 434
 435 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 436         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 437 {
 438         return NULL;
 439 }
 440
 441 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 442         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 443 {
 444         return NULL;
 445 }
 446
 447 static inline bool thp_migration_supported(void)
 448 {
 449         return false;
 450 }
 451 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 452
 453 #endif /* _LINUX_HUGE_MM_H */