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 bool is_vma_temporary_stack(struct vm_area_struct *vma);
 128
 129 extern unsigned long transparent_hugepage_flags;
 130
 131 /*
 132  * to be used on vmas which are known to support THP.
 133  * Use transparent_hugepage_enabled otherwise
 134  */
 135 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 136 {
 137         if (vma->vm_flags & VM_NOHUGEPAGE)
 138                 return false;
 139
 140         if (is_vma_temporary_stack(vma))
 141                 return false;
 142
 143         if (test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 144                 return false;
 145
 146         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 147                 return true;
 148         /*
 149          * For dax vmas, try to always use hugepage mappings. If the kernel does
 150          * not support hugepages, fsdax mappings will fallback to PAGE_SIZE
 151          * mappings, and device-dax namespaces, that try to guarantee a given
 152          * mapping size, will fail to enable
 153          */
 154         if (vma_is_dax(vma))
 155                 return true;
 156
 157         if (transparent_hugepage_flags &
 158                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 159                 return !!(vma->vm_flags & VM_HUGEPAGE);
 160
 161         return false;
 162 }
 163
 164 bool transparent_hugepage_enabled(struct vm_area_struct *vma);
 165
 166 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
 167
 168 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 169                 unsigned long haddr)
 170 {
 171         /* Don't have to check pgoff for anonymous vma */
 172         if (!vma_is_anonymous(vma)) {
 173                 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
 174                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
 175                         return false;
 176         }
 177
 178         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 179                 return false;
 180         return true;
 181 }
 182
 183 #define transparent_hugepage_use_zero_page()                            \
 184         (transparent_hugepage_flags &                                   \
 185          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 186 #ifdef CONFIG_DEBUG_VM
 187 #define transparent_hugepage_debug_cow()                                \
 188         (transparent_hugepage_flags &                                   \
 189          (1<<TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG))
 190 #else /* CONFIG_DEBUG_VM */
 191 #define transparent_hugepage_debug_cow() 0
 192 #endif /* CONFIG_DEBUG_VM */
 193
 194 extern unsigned long thp_get_unmapped_area(struct file *filp,
 195                 unsigned long addr, unsigned long len, unsigned long pgoff,
 196                 unsigned long flags);
 197
 198 extern void prep_transhuge_page(struct page *page);
 199 extern void free_transhuge_page(struct page *page);
 200 bool is_transparent_hugepage(struct page *page);
 201
 202 bool can_split_huge_page(struct page *page, 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 page *page);
 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 page *page);
 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 extern int hugepage_madvise(struct vm_area_struct *vma,
 238                             unsigned long *vm_flags, int advice);
 239 extern void vma_adjust_trans_huge(struct vm_area_struct *vma,
 240                                     unsigned long start,
 241                                     unsigned long end,
 242                                     long adjust_next);
 243 extern spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd,
 244                 struct vm_area_struct *vma);
 245 extern spinlock_t *__pud_trans_huge_lock(pud_t *pud,
 246                 struct vm_area_struct *vma);
 247
 248 static inline int is_swap_pmd(pmd_t pmd)
 249 {
 250         return !pmd_none(pmd) && !pmd_present(pmd);
 251 }
 252
 253 /* mmap_sem must be held on entry */
 254 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 255                 struct vm_area_struct *vma)
 256 {
 257         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 258                 return __pmd_trans_huge_lock(pmd, vma);
 259         else
 260                 return NULL;
 261 }
 262 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 263                 struct vm_area_struct *vma)
 264 {
 265         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 266                 return __pud_trans_huge_lock(pud, vma);
 267         else
 268                 return NULL;
 269 }
 270 static inline int hpage_nr_pages(struct page *page)
 271 {
 272         if (unlikely(PageTransHuge(page)))
 273                 return HPAGE_PMD_NR;
 274         return 1;
 275 }
 276
 277 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 278                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 279 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 280                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 281
 282 extern vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 283
 284 extern struct page *huge_zero_page;
 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 is_huge_zero_page(pmd_page(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 #define hpage_nr_pages(x) 1
 330
 331 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 332 {
 333         return false;
 334 }
 335
 336 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
 337 {
 338         return false;
 339 }
 340
 341 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 342                 unsigned long haddr)
 343 {
 344         return false;
 345 }
 346
 347 static inline void prep_transhuge_page(struct page *page) {}
 348
 349 static inline bool is_transparent_hugepage(struct page *page)
 350 {
 351         return false;
 352 }
 353
 354 #define transparent_hugepage_flags 0UL
 355
 356 #define thp_get_unmapped_area   NULL
 357
 358 static inline bool
 359 can_split_huge_page(struct page *page, int *pextra_pins)
 360 {
 361         BUILD_BUG();
 362         return false;
 363 }
 364 static inline int
 365 split_huge_page_to_list(struct page *page, struct list_head *list)
 366 {
 367         return 0;
 368 }
 369 static inline int split_huge_page(struct page *page)
 370 {
 371         return 0;
 372 }
 373 static inline void deferred_split_huge_page(struct page *page) {}
 374 #define split_huge_pmd(__vma, __pmd, __address) \
 375         do { } while (0)
 376
 377 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 378                 unsigned long address, bool freeze, struct page *page) {}
 379 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 380                 unsigned long address, bool freeze, struct page *page) {}
 381
 382 #define split_huge_pud(__vma, __pmd, __address) \
 383         do { } while (0)
 384
 385 static inline int hugepage_madvise(struct vm_area_struct *vma,
 386                                    unsigned long *vm_flags, int advice)
 387 {
 388         BUG();
 389         return 0;
 390 }
 391 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 392                                          unsigned long start,
 393                                          unsigned long end,
 394                                          long adjust_next)
 395 {
 396 }
 397 static inline int is_swap_pmd(pmd_t pmd)
 398 {
 399         return 0;
 400 }
 401 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 402                 struct vm_area_struct *vma)
 403 {
 404         return NULL;
 405 }
 406 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 407                 struct vm_area_struct *vma)
 408 {
 409         return NULL;
 410 }
 411
 412 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 413                 pmd_t orig_pmd)
 414 {
 415         return 0;
 416 }
 417
 418 static inline bool is_huge_zero_page(struct page *page)
 419 {
 420         return false;
 421 }
 422
 423 static inline bool is_huge_zero_pud(pud_t pud)
 424 {
 425         return false;
 426 }
 427
 428 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 429 {
 430         return;
 431 }
 432
 433 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 434         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 435 {
 436         return NULL;
 437 }
 438
 439 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 440         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 441 {
 442         return NULL;
 443 }
 444
 445 static inline bool thp_migration_supported(void)
 446 {
 447         return false;
 448 }
 449 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 450
 451 #endif /* _LINUX_HUGE_MM_H */