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 /*
 120  * Do the below checks:
 121  *   - For file vma, check if the linear page offset of vma is
 122  *     HPAGE_PMD_NR aligned within the file.  The hugepage is
 123  *     guaranteed to be hugepage-aligned within the file, but we must
 124  *     check that the PMD-aligned addresses in the VMA map to
 125  *     PMD-aligned offsets within the file, else the hugepage will
 126  *     not be PMD-mappable.
 127  *   - For all vmas, check if the haddr is in an aligned HPAGE_PMD_SIZE
 128  *     area.
 129  */
 130 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 131                 unsigned long addr)
 132 {
 133         unsigned long haddr;
 134
 135         /* Don't have to check pgoff for anonymous vma */
 136         if (!vma_is_anonymous(vma)) {
 137                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 138                                 HPAGE_PMD_NR))
 139                         return false;
 140         }
 141
 142         haddr = addr & HPAGE_PMD_MASK;
 143
 144         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 145                 return false;
 146         return true;
 147 }
 148
 149 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 150                                           unsigned long vm_flags)
 151 {
 152         /* Explicitly disabled through madvise. */
 153         if ((vm_flags & VM_NOHUGEPAGE) ||
 154             test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 155                 return false;
 156         return true;
 157 }
 158
 159 /*
 160  * to be used on vmas which are known to support THP.
 161  * Use transparent_hugepage_active otherwise
 162  */
 163 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 164 {
 165
 166         /*
 167          * If the hardware/firmware marked hugepage support disabled.
 168          */
 169         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
 170                 return false;
 171
 172         if (!transhuge_vma_enabled(vma, vma->vm_flags))
 173                 return false;
 174
 175         if (vma_is_temporary_stack(vma))
 176                 return false;
 177
 178         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 179                 return true;
 180
 181         if (vma_is_dax(vma))
 182                 return true;
 183
 184         if (transparent_hugepage_flags &
 185                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 186                 return !!(vma->vm_flags & VM_HUGEPAGE);
 187
 188         return false;
 189 }
 190
 191 static inline bool file_thp_enabled(struct vm_area_struct *vma)
 192 {
 193         struct inode *inode;
 194
 195         if (!vma->vm_file)
 196                 return false;
 197
 198         inode = vma->vm_file->f_inode;
 199
 200         return (IS_ENABLED(CONFIG_READ_ONLY_THP_FOR_FS)) &&
 201                (vma->vm_flags & VM_EXEC) &&
 202                !inode_is_open_for_write(inode) && S_ISREG(inode->i_mode);
 203 }
 204
 205 bool hugepage_vma_check(struct vm_area_struct *vma,
 206                         unsigned long vm_flags,
 207                         bool smaps);
 208
 209 #define transparent_hugepage_use_zero_page()                            \
 210         (transparent_hugepage_flags &                                   \
 211          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 212
 213 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 214                 unsigned long len, unsigned long pgoff, unsigned long flags);
 215
 216 void prep_transhuge_page(struct page *page);
 217 void free_transhuge_page(struct page *page);
 218
 219 bool can_split_folio(struct folio *folio, int *pextra_pins);
 220 int split_huge_page_to_list(struct page *page, struct list_head *list);
 221 static inline int split_huge_page(struct page *page)
 222 {
 223         return split_huge_page_to_list(page, NULL);
 224 }
 225 void deferred_split_huge_page(struct page *page);
 226
 227 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 228                 unsigned long address, bool freeze, struct folio *folio);
 229
 230 #define split_huge_pmd(__vma, __pmd, __address)                         \
 231         do {                                                            \
 232                 pmd_t *____pmd = (__pmd);                               \
 233                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 234                                         || pmd_devmap(*____pmd))        \
 235                         __split_huge_pmd(__vma, __pmd, __address,       \
 236                                                 false, NULL);           \
 237         }  while (0)
 238
 239
 240 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 241                 bool freeze, struct folio *folio);
 242
 243 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 244                 unsigned long address);
 245
 246 #define split_huge_pud(__vma, __pud, __address)                         \
 247         do {                                                            \
 248                 pud_t *____pud = (__pud);                               \
 249                 if (pud_trans_huge(*____pud)                            \
 250                                         || pud_devmap(*____pud))        \
 251                         __split_huge_pud(__vma, __pud, __address);      \
 252         }  while (0)
 253
 254 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 255                      int advice);
 256 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 257                            unsigned long end, long adjust_next);
 258 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 259 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 260
 261 static inline int is_swap_pmd(pmd_t pmd)
 262 {
 263         return !pmd_none(pmd) && !pmd_present(pmd);
 264 }
 265
 266 /* mmap_lock must be held on entry */
 267 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 268                 struct vm_area_struct *vma)
 269 {
 270         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 271                 return __pmd_trans_huge_lock(pmd, vma);
 272         else
 273                 return NULL;
 274 }
 275 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 276                 struct vm_area_struct *vma)
 277 {
 278         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 279                 return __pud_trans_huge_lock(pud, vma);
 280         else
 281                 return NULL;
 282 }
 283
 284 /**
 285  * folio_test_pmd_mappable - Can we map this folio with a PMD?
 286  * @folio: The folio to test
 287  */
 288 static inline bool folio_test_pmd_mappable(struct folio *folio)
 289 {
 290         return folio_order(folio) >= HPAGE_PMD_ORDER;
 291 }
 292
 293 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 294                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 295 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 296                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 297
 298 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 299
 300 extern struct page *huge_zero_page;
 301 extern unsigned long huge_zero_pfn;
 302
 303 static inline bool is_huge_zero_page(struct page *page)
 304 {
 305         return READ_ONCE(huge_zero_page) == page;
 306 }
 307
 308 static inline bool is_huge_zero_pmd(pmd_t pmd)
 309 {
 310         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 311 }
 312
 313 static inline bool is_huge_zero_pud(pud_t pud)
 314 {
 315         return false;
 316 }
 317
 318 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 319 void mm_put_huge_zero_page(struct mm_struct *mm);
 320
 321 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 322
 323 static inline bool thp_migration_supported(void)
 324 {
 325         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 326 }
 327
 328 static inline struct list_head *page_deferred_list(struct page *page)
 329 {
 330         /*
 331          * Global or memcg deferred list in the second tail pages is
 332          * occupied by compound_head.
 333          */
 334         return &page[2].deferred_list;
 335 }
 336
 337 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 338 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 339 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 340 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 341
 342 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 343 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 344 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 345
 346 static inline bool folio_test_pmd_mappable(struct folio *folio)
 347 {
 348         return false;
 349 }
 350
 351 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 352 {
 353         return false;
 354 }
 355
 356 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 357                 unsigned long addr)
 358 {
 359         return false;
 360 }
 361
 362 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 363                                           unsigned long vm_flags)
 364 {
 365         return false;
 366 }
 367
 368 static inline bool hugepage_vma_check(struct vm_area_struct *vma,
 369                                        unsigned long vm_flags,
 370                                        bool smaps)
 371 {
 372         return false;
 373 }
 374
 375 static inline void prep_transhuge_page(struct page *page) {}
 376
 377 #define transparent_hugepage_flags 0UL
 378
 379 #define thp_get_unmapped_area   NULL
 380
 381 static inline bool
 382 can_split_folio(struct folio *folio, int *pextra_pins)
 383 {
 384         return false;
 385 }
 386 static inline int
 387 split_huge_page_to_list(struct page *page, struct list_head *list)
 388 {
 389         return 0;
 390 }
 391 static inline int split_huge_page(struct page *page)
 392 {
 393         return 0;
 394 }
 395 static inline void deferred_split_huge_page(struct page *page) {}
 396 #define split_huge_pmd(__vma, __pmd, __address) \
 397         do { } while (0)
 398
 399 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 400                 unsigned long address, bool freeze, struct folio *folio) {}
 401 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 402                 unsigned long address, bool freeze, struct folio *folio) {}
 403
 404 #define split_huge_pud(__vma, __pmd, __address) \
 405         do { } while (0)
 406
 407 static inline int hugepage_madvise(struct vm_area_struct *vma,
 408                                    unsigned long *vm_flags, int advice)
 409 {
 410         BUG();
 411         return 0;
 412 }
 413 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 414                                          unsigned long start,
 415                                          unsigned long end,
 416                                          long adjust_next)
 417 {
 418 }
 419 static inline int is_swap_pmd(pmd_t pmd)
 420 {
 421         return 0;
 422 }
 423 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 424                 struct vm_area_struct *vma)
 425 {
 426         return NULL;
 427 }
 428 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 429                 struct vm_area_struct *vma)
 430 {
 431         return NULL;
 432 }
 433
 434 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 435 {
 436         return 0;
 437 }
 438
 439 static inline bool is_huge_zero_page(struct page *page)
 440 {
 441         return false;
 442 }
 443
 444 static inline bool is_huge_zero_pmd(pmd_t pmd)
 445 {
 446         return false;
 447 }
 448
 449 static inline bool is_huge_zero_pud(pud_t pud)
 450 {
 451         return false;
 452 }
 453
 454 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 455 {
 456         return;
 457 }
 458
 459 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 460         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 461 {
 462         return NULL;
 463 }
 464
 465 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 466         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 467 {
 468         return NULL;
 469 }
 470
 471 static inline bool thp_migration_supported(void)
 472 {
 473         return false;
 474 }
 475 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 476
 477 static inline int split_folio_to_list(struct folio *folio,
 478                 struct list_head *list)
 479 {
 480         return split_huge_page_to_list(&folio->page, list);
 481 }
 482
 483 #endif /* _LINUX_HUGE_MM_H */