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