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