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_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 /*
 119  * to be used on vmas which are known to support THP.
 120  * Use transparent_hugepage_enabled otherwise
 121  */
 122 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 123 {
 124
 125         /*
 126          * If the hardware/firmware marked hugepage support disabled.
 127          */
 128         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
 129                 return false;
 130
 131         if (vma->vm_flags & VM_NOHUGEPAGE)
 132                 return false;
 133
 134         if (vma_is_temporary_stack(vma))
 135                 return false;
 136
 137         if (test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 138                 return false;
 139
 140         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 141                 return true;
 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 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 156                 unsigned long haddr)
 157 {
 158         /* Don't have to check pgoff for anonymous vma */
 159         if (!vma_is_anonymous(vma)) {
 160                 if (!IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 161                                 HPAGE_PMD_NR))
 162                         return false;
 163         }
 164
 165         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 166                 return false;
 167         return true;
 168 }
 169
 170 #define transparent_hugepage_use_zero_page()                            \
 171         (transparent_hugepage_flags &                                   \
 172          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 173
 174 unsigned long thp_get_unmapped_area(struct file *filp, unsigned long addr,
 175                 unsigned long len, unsigned long pgoff, unsigned long flags);
 176
 177 void prep_transhuge_page(struct page *page);
 178 void free_transhuge_page(struct page *page);
 179 bool is_transparent_hugepage(struct page *page);
 180
 181 bool can_split_huge_page(struct page *page, int *pextra_pins);
 182 int split_huge_page_to_list(struct page *page, struct list_head *list);
 183 static inline int split_huge_page(struct page *page)
 184 {
 185         return split_huge_page_to_list(page, NULL);
 186 }
 187 void deferred_split_huge_page(struct page *page);
 188
 189 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 190                 unsigned long address, bool freeze, struct page *page);
 191
 192 #define split_huge_pmd(__vma, __pmd, __address)                         \
 193         do {                                                            \
 194                 pmd_t *____pmd = (__pmd);                               \
 195                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 196                                         || pmd_devmap(*____pmd))        \
 197                         __split_huge_pmd(__vma, __pmd, __address,       \
 198                                                 false, NULL);           \
 199         }  while (0)
 200
 201
 202 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 203                 bool freeze, struct page *page);
 204
 205 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 206                 unsigned long address);
 207
 208 #define split_huge_pud(__vma, __pud, __address)                         \
 209         do {                                                            \
 210                 pud_t *____pud = (__pud);                               \
 211                 if (pud_trans_huge(*____pud)                            \
 212                                         || pud_devmap(*____pud))        \
 213                         __split_huge_pud(__vma, __pud, __address);      \
 214         }  while (0)
 215
 216 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 217                      int advice);
 218 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 219                            unsigned long end, long adjust_next);
 220 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 221 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 222
 223 static inline int is_swap_pmd(pmd_t pmd)
 224 {
 225         return !pmd_none(pmd) && !pmd_present(pmd);
 226 }
 227
 228 /* mmap_lock must be held on entry */
 229 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 230                 struct vm_area_struct *vma)
 231 {
 232         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 233                 return __pmd_trans_huge_lock(pmd, vma);
 234         else
 235                 return NULL;
 236 }
 237 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 238                 struct vm_area_struct *vma)
 239 {
 240         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 241                 return __pud_trans_huge_lock(pud, vma);
 242         else
 243                 return NULL;
 244 }
 245
 246 /**
 247  * thp_head - Head page of a transparent huge page.
 248  * @page: Any page (tail, head or regular) found in the page cache.
 249  */
 250 static inline struct page *thp_head(struct page *page)
 251 {
 252         return compound_head(page);
 253 }
 254
 255 /**
 256  * thp_order - Order of a transparent huge page.
 257  * @page: Head page of a transparent huge page.
 258  */
 259 static inline unsigned int thp_order(struct page *page)
 260 {
 261         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 262         if (PageHead(page))
 263                 return HPAGE_PMD_ORDER;
 264         return 0;
 265 }
 266
 267 /**
 268  * thp_nr_pages - The number of regular pages in this huge page.
 269  * @page: The head page of a huge page.
 270  */
 271 static inline int thp_nr_pages(struct page *page)
 272 {
 273         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 274         if (PageHead(page))
 275                 return HPAGE_PMD_NR;
 276         return 1;
 277 }
 278
 279 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 280                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 281 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 282                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 283
 284 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 285
 286 extern struct page *huge_zero_page;
 287 extern unsigned long huge_zero_pfn;
 288
 289 static inline bool is_huge_zero_page(struct page *page)
 290 {
 291         return READ_ONCE(huge_zero_page) == page;
 292 }
 293
 294 static inline bool is_huge_zero_pmd(pmd_t pmd)
 295 {
 296         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 297 }
 298
 299 static inline bool is_huge_zero_pud(pud_t pud)
 300 {
 301         return false;
 302 }
 303
 304 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 305 void mm_put_huge_zero_page(struct mm_struct *mm);
 306
 307 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 308
 309 static inline bool thp_migration_supported(void)
 310 {
 311         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 312 }
 313
 314 static inline struct list_head *page_deferred_list(struct page *page)
 315 {
 316         /*
 317          * Global or memcg deferred list in the second tail pages is
 318          * occupied by compound_head.
 319          */
 320         return &page[2].deferred_list;
 321 }
 322
 323 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 324 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 325 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 326 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 327
 328 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 329 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 330 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 331
 332 static inline struct page *thp_head(struct page *page)
 333 {
 334         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 335         return page;
 336 }
 337
 338 static inline unsigned int thp_order(struct page *page)
 339 {
 340         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 341         return 0;
 342 }
 343
 344 static inline int thp_nr_pages(struct page *page)
 345 {
 346         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 347         return 1;
 348 }
 349
 350 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 351 {
 352         return false;
 353 }
 354
 355 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
 356 {
 357         return false;
 358 }
 359
 360 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 361                 unsigned long haddr)
 362 {
 363         return false;
 364 }
 365
 366 static inline void prep_transhuge_page(struct page *page) {}
 367
 368 static inline bool is_transparent_hugepage(struct page *page)
 369 {
 370         return false;
 371 }
 372
 373 #define transparent_hugepage_flags 0UL
 374
 375 #define thp_get_unmapped_area   NULL
 376
 377 static inline bool
 378 can_split_huge_page(struct page *page, int *pextra_pins)
 379 {
 380         BUILD_BUG();
 381         return false;
 382 }
 383 static inline int
 384 split_huge_page_to_list(struct page *page, struct list_head *list)
 385 {
 386         return 0;
 387 }
 388 static inline int split_huge_page(struct page *page)
 389 {
 390         return 0;
 391 }
 392 static inline void deferred_split_huge_page(struct page *page) {}
 393 #define split_huge_pmd(__vma, __pmd, __address) \
 394         do { } while (0)
 395
 396 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 397                 unsigned long address, bool freeze, struct page *page) {}
 398 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 399                 unsigned long address, bool freeze, struct page *page) {}
 400
 401 #define split_huge_pud(__vma, __pmd, __address) \
 402         do { } while (0)
 403
 404 static inline int hugepage_madvise(struct vm_area_struct *vma,
 405                                    unsigned long *vm_flags, int advice)
 406 {
 407         BUG();
 408         return 0;
 409 }
 410 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 411                                          unsigned long start,
 412                                          unsigned long end,
 413                                          long adjust_next)
 414 {
 415 }
 416 static inline int is_swap_pmd(pmd_t pmd)
 417 {
 418         return 0;
 419 }
 420 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 421                 struct vm_area_struct *vma)
 422 {
 423         return NULL;
 424 }
 425 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 426                 struct vm_area_struct *vma)
 427 {
 428         return NULL;
 429 }
 430
 431 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 432                 pmd_t orig_pmd)
 433 {
 434         return 0;
 435 }
 436
 437 static inline bool is_huge_zero_page(struct page *page)
 438 {
 439         return false;
 440 }
 441
 442 static inline bool is_huge_zero_pmd(pmd_t pmd)
 443 {
 444         return false;
 445 }
 446
 447 static inline bool is_huge_zero_pud(pud_t pud)
 448 {
 449         return false;
 450 }
 451
 452 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 453 {
 454         return;
 455 }
 456
 457 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 458         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 459 {
 460         return NULL;
 461 }
 462
 463 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 464         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 465 {
 466         return NULL;
 467 }
 468
 469 static inline bool thp_migration_supported(void)
 470 {
 471         return false;
 472 }
 473 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 474
 475 /**
 476  * thp_size - Size of a transparent huge page.
 477  * @page: Head page of a transparent huge page.
 478  *
 479  * Return: Number of bytes in this page.
 480  */
 481 static inline unsigned long thp_size(struct page *page)
 482 {
 483         return PAGE_SIZE << thp_order(page);
 484 }
 485
 486 #endif /* _LINUX_HUGE_MM_H */