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);
  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);
  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 mmu_gather *tlb, struct vm_area_struct *vma,
  40                     pmd_t *pmd, unsigned long addr, pgprot_t newprot,
  41                     unsigned long cp_flags);
  42 vm_fault_t vmf_insert_pfn_pmd_prot(struct vm_fault *vmf, pfn_t pfn,
  43                                    pgprot_t pgprot, bool write);
  44
  45 /**
  46  * vmf_insert_pfn_pmd - insert a pmd size pfn
  47  * @vmf: Structure describing the fault
  48  * @pfn: pfn to insert
  49  * @pgprot: page protection to use
  50  * @write: whether it's a write fault
  51  *
  52  * Insert a pmd size pfn. See vmf_insert_pfn() for additional info.
  53  *
  54  * Return: vm_fault_t value.
  55  */
  56 static inline vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn,
  57                                             bool write)
  58 {
  59         return vmf_insert_pfn_pmd_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  60 }
  61 vm_fault_t vmf_insert_pfn_pud_prot(struct vm_fault *vmf, pfn_t pfn,
  62                                    pgprot_t pgprot, bool write);
  63
  64 /**
  65  * vmf_insert_pfn_pud - insert a pud size pfn
  66  * @vmf: Structure describing the fault
  67  * @pfn: pfn to insert
  68  * @pgprot: page protection to use
  69  * @write: whether it's a write fault
  70  *
  71  * Insert a pud size pfn. See vmf_insert_pfn() for additional info.
  72  *
  73  * Return: vm_fault_t value.
  74  */
  75 static inline vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn,
  76                                             bool write)
  77 {
  78         return vmf_insert_pfn_pud_prot(vmf, pfn, vmf->vma->vm_page_prot, write);
  79 }
  80
  81 enum transparent_hugepage_flag {
  82         TRANSPARENT_HUGEPAGE_NEVER_DAX,
  83         TRANSPARENT_HUGEPAGE_FLAG,
  84         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  85         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  86         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  87         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  88         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  89         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  90         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  91 };
  92
  93 struct kobject;
  94 struct kobj_attribute;
  95
  96 ssize_t single_hugepage_flag_store(struct kobject *kobj,
  97                                    struct kobj_attribute *attr,
  98                                    const char *buf, size_t count,
  99                                    enum transparent_hugepage_flag flag);
 100 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 101                                   struct kobj_attribute *attr, char *buf,
 102                                   enum transparent_hugepage_flag flag);
 103 extern struct kobj_attribute shmem_enabled_attr;
 104
 105 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
 106 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
 107
 108 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 109 #define HPAGE_PMD_SHIFT PMD_SHIFT
 110 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
 111 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
 112
 113 #define HPAGE_PUD_SHIFT PUD_SHIFT
 114 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
 115 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
 116
 117 extern unsigned long transparent_hugepage_flags;
 118
 119 #define hugepage_flags_enabled()                                               \
 120         (transparent_hugepage_flags &                                  \
 121          ((1<<TRANSPARENT_HUGEPAGE_FLAG) |                     \
 122           (1<<TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG)))
 123 #define hugepage_flags_always()                         \
 124         (transparent_hugepage_flags &                   \
 125          (1<<TRANSPARENT_HUGEPAGE_FLAG))
 126
 127 /*
 128  * Do the below checks:
 129  *   - For file vma, check if the linear page offset of vma is
 130  *     HPAGE_PMD_NR aligned within the file.  The hugepage is
 131  *     guaranteed to be hugepage-aligned within the file, but we must
 132  *     check that the PMD-aligned addresses in the VMA map to
 133  *     PMD-aligned offsets within the file, else the hugepage will
 134  *     not be PMD-mappable.
 135  *   - For all vmas, check if the haddr is in an aligned HPAGE_PMD_SIZE
 136  *     area.
 137  */
 138 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 139                 unsigned long addr)
 140 {
 141         unsigned long haddr;
 142
 143         /* Don't have to check pgoff for anonymous vma */
 144         if (!vma_is_anonymous(vma)) {
 145                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 146                                 HPAGE_PMD_NR))
 147                         return false;
 148         }
 149
 150         haddr = addr & HPAGE_PMD_MASK;
 151
 152         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 153                 return false;
 154         return true;
 155 }
 156
 157 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 158 {
 159         struct inode *inode;
 160
 161         if (!vma->vm_file)
 162                 return false;
 163
 164         inode = vma->vm_file->f_inode;
 165
 166         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 167                (vma->vm_flags & VM_EXEC) &&
 168                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 169 }
 170
 171 bool hugepage_vma_check(struct vm_area_struct *vma,
 172                         unsigned long vm_flags,
 173                         bool smaps, bool in_pf);
 174
 175 #define transparent_hugepage_use_zero_page()                            \
 176         (transparent_hugepage_flags &                                   \
 177          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 178
 179 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 180                 unsigned long len, unsigned long pgoff, unsigned long flags);
 181
 182 void prep_transhuge_page(struct page *page);
 183 void free_transhuge_page(struct page *page);
 184
 185 bool can_split_folio(struct folio *folio, int *pextra_pins);
 186 int split_huge_page_to_list(struct page *page, struct list_head *list);
 187 static inline int split_huge_page(struct page *page)
 188 {
 189         return split_huge_page_to_list(page, NULL);
 190 }
 191 void deferred_split_huge_page(struct page *page);
 192
 193 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 194                 unsigned long address, bool freeze, struct folio *folio);
 195
 196 #define split_huge_pmd(__vma, __pmd, __address)                         \
 197         do {                                                            \
 198                 pmd_t *____pmd = (__pmd);                               \
 199                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 200                                         || pmd_devmap(*____pmd))        \
 201                         __split_huge_pmd(__vma, __pmd, __address,       \
 202                                                 false, NULL);           \
 203         }  while (0)
 204
 205
 206 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 207                 bool freeze, struct folio *folio);
 208
 209 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 210                 unsigned long address);
 211
 212 #define split_huge_pud(__vma, __pud, __address)                         \
 213         do {                                                            \
 214                 pud_t *____pud = (__pud);                               \
 215                 if (pud_trans_huge(*____pud)                            \
 216                                         || pud_devmap(*____pud))        \
 217                         __split_huge_pud(__vma, __pud, __address);      \
 218         }  while (0)
 219
 220 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 221                      int advice);
 222 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 223                            unsigned long end, long adjust_next);
 224 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 225 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 226
 227 static inline int is_swap_pmd(pmd_t pmd)
 228 {
 229         return !pmd_none(pmd) && !pmd_present(pmd);
 230 }
 231
 232 /* mmap_lock must be held on entry */
 233 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 234                 struct vm_area_struct *vma)
 235 {
 236         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 237                 return __pmd_trans_huge_lock(pmd, vma);
 238         else
 239                 return NULL;
 240 }
 241 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 242                 struct vm_area_struct *vma)
 243 {
 244         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 245                 return __pud_trans_huge_lock(pud, vma);
 246         else
 247                 return NULL;
 248 }
 249
 250 /**
 251  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 252  * @folio: The folio to test
 253  */
 254 static inline bool folio_test_pmd_mappable(struct folio *folio)
 255 {
 256         return folio_order(folio) >= HPAGE_PMD_ORDER;
 257 }
 258
 259 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 260                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 261 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 262                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 263
 264 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 265
 266 extern struct page *huge_zero_page;
 267 extern unsigned long huge_zero_pfn;
 268
 269 static inline bool is_huge_zero_page(struct page *page)
 270 {
 271         return READ_ONCE(huge_zero_page) == page;
 272 }
 273
 274 static inline bool is_huge_zero_pmd(pmd_t pmd)
 275 {
 276         return pmd_present(pmd) && READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd);
 277 }
 278
 279 static inline bool is_huge_zero_pud(pud_t pud)
 280 {
 281         return false;
 282 }
 283
 284 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 285 void mm_put_huge_zero_page(struct mm_struct *mm);
 286
 287 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 288
 289 static inline bool thp_migration_supported(void)
 290 {
 291         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 292 }
 293
 294 static inline struct list_head *page_deferred_list(struct page *page)
 295 {
 296         /*
 297          * See organization of tail pages of compound page in
 298          * "struct page" definition.
 299          */
 300         return &page[2].deferred_list;
 301 }
 302
 303 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 304 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 305 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 306 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 307
 308 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 309 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 310 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 311
 312 static inline bool folio_test_pmd_mappable(struct folio *folio)
 313 {
 314         return false;
 315 }
 316
 317 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 318                 unsigned long addr)
 319 {
 320         return false;
 321 }
 322
 323 static inline bool hugepage_vma_check(struct vm_area_struct *vma,
 324                                        unsigned long vm_flags,
 325                                        bool smaps, bool in_pf)
 326 {
 327         return false;
 328 }
 329
 330 static inline void prep_transhuge_page(struct page *page) {}
 331
 332 #define transparent_hugepage_flags 0UL
 333
 334 #define thp_get_unmapped_area   NULL
 335
 336 static inline bool
 337 can_split_folio(struct folio *folio, int *pextra_pins)
 338 {
 339         return false;
 340 }
 341 static inline int
 342 split_huge_page_to_list(struct page *page, struct list_head *list)
 343 {
 344         return 0;
 345 }
 346 static inline int split_huge_page(struct page *page)
 347 {
 348         return 0;
 349 }
 350 static inline void deferred_split_huge_page(struct page *page) {}
 351 #define split_huge_pmd(__vma, __pmd, __address) \
 352         do { } while (0)
 353
 354 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 355                 unsigned long address, bool freeze, struct folio *folio) {}
 356 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 357                 unsigned long address, bool freeze, struct folio *folio) {}
 358
 359 #define split_huge_pud(__vma, __pmd, __address) \
 360         do { } while (0)
 361
 362 static inline int hugepage_madvise(struct vm_area_struct *vma,
 363                                    unsigned long *vm_flags, int advice)
 364 {
 365         BUG();
 366         return 0;
 367 }
 368 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 369                                          unsigned long start,
 370                                          unsigned long end,
 371                                          long adjust_next)
 372 {
 373 }
 374 static inline int is_swap_pmd(pmd_t pmd)
 375 {
 376         return 0;
 377 }
 378 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 379                 struct vm_area_struct *vma)
 380 {
 381         return NULL;
 382 }
 383 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 384                 struct vm_area_struct *vma)
 385 {
 386         return NULL;
 387 }
 388
 389 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 390 {
 391         return 0;
 392 }
 393
 394 static inline bool is_huge_zero_page(struct page *page)
 395 {
 396         return false;
 397 }
 398
 399 static inline bool is_huge_zero_pmd(pmd_t pmd)
 400 {
 401         return false;
 402 }
 403
 404 static inline bool is_huge_zero_pud(pud_t pud)
 405 {
 406         return false;
 407 }
 408
 409 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 410 {
 411         return;
 412 }
 413
 414 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 415         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 416 {
 417         return NULL;
 418 }
 419
 420 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 421         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 422 {
 423         return NULL;
 424 }
 425
 426 static inline bool thp_migration_supported(void)
 427 {
 428         return false;
 429 }
 430 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 431
 432 static inline int split_folio_to_list(struct folio *folio,
 433                 struct list_head *list)
 434 {
 435         return split_huge_page_to_list(&folio->page, list);
 436 }
 437
 438 /*
 439  * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
 440  * limitations in the implementation like arm64 MTE can override this to
 441  * false
 442  */
 443 #ifndef arch_thp_swp_supported
 444 static inline bool arch_thp_swp_supported(void)
 445 {
 446         return true;
 447 }
 448 #endif
 449
 450 #endif /* _LINUX_HUGE_MM_H */