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 bool is_transparent_hugepage(struct page *page);
 187
 188 bool can_split_huge_page(struct page *page, int *pextra_pins);
 189 int split_huge_page_to_list(struct page *page, struct list_head *list);
 190 static inline int split_huge_page(struct page *page)
 191 {
 192         return split_huge_page_to_list(page, NULL);
 193 }
 194 void deferred_split_huge_page(struct page *page);
 195
 196 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 197                 unsigned long address, bool freeze, struct page *page);
 198
 199 #define split_huge_pmd(__vma, __pmd, __address)                         \
 200         do {                                                            \
 201                 pmd_t *____pmd = (__pmd);                               \
 202                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 203                                         || pmd_devmap(*____pmd))        \
 204                         __split_huge_pmd(__vma, __pmd, __address,       \
 205                                                 false, NULL);           \
 206         }  while (0)
 207
 208
 209 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 210                 bool freeze, struct page *page);
 211
 212 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 213                 unsigned long address);
 214
 215 #define split_huge_pud(__vma, __pud, __address)                         \
 216         do {                                                            \
 217                 pud_t *____pud = (__pud);                               \
 218                 if (pud_trans_huge(*____pud)                            \
 219                                         || pud_devmap(*____pud))        \
 220                         __split_huge_pud(__vma, __pud, __address);      \
 221         }  while (0)
 222
 223 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 224                      int advice);
 225 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 226                            unsigned long end, long adjust_next);
 227 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 228 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 229
 230 static inline int is_swap_pmd(pmd_t pmd)
 231 {
 232         return !pmd_none(pmd) && !pmd_present(pmd);
 233 }
 234
 235 /* mmap_lock must be held on entry */
 236 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 237                 struct vm_area_struct *vma)
 238 {
 239         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 240                 return __pmd_trans_huge_lock(pmd, vma);
 241         else
 242                 return NULL;
 243 }
 244 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 245                 struct vm_area_struct *vma)
 246 {
 247         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 248                 return __pud_trans_huge_lock(pud, vma);
 249         else
 250                 return NULL;
 251 }
 252
 253 /**
 254  * thp_head - Head page of a transparent huge page.
 255  * @page: Any page (tail, head or regular) found in the page cache.
 256  */
 257 static inline struct page *thp_head(struct page *page)
 258 {
 259         return compound_head(page);
 260 }
 261
 262 /**
 263  * thp_order - Order of a transparent huge page.
 264  * @page: Head page of a transparent huge page.
 265  */
 266 static inline unsigned int thp_order(struct page *page)
 267 {
 268         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 269         if (PageHead(page))
 270                 return HPAGE_PMD_ORDER;
 271         return 0;
 272 }
 273
 274 /**
 275  * thp_nr_pages - The number of regular pages in this huge page.
 276  * @page: The head page of a huge page.
 277  */
 278 static inline int thp_nr_pages(struct page *page)
 279 {
 280         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 281         if (PageHead(page))
 282                 return HPAGE_PMD_NR;
 283         return 1;
 284 }
 285
 286 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 287                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 288 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 289                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 290
 291 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf);
 292
 293 extern struct page *huge_zero_page;
 294 extern unsigned long huge_zero_pfn;
 295
 296 static inline bool is_huge_zero_page(struct page *page)
 297 {
 298         return READ_ONCE(huge_zero_page) == page;
 299 }
 300
 301 static inline bool is_huge_zero_pmd(pmd_t pmd)
 302 {
 303         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 304 }
 305
 306 static inline bool is_huge_zero_pud(pud_t pud)
 307 {
 308         return false;
 309 }
 310
 311 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 312 void mm_put_huge_zero_page(struct mm_struct *mm);
 313
 314 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 315
 316 static inline bool thp_migration_supported(void)
 317 {
 318         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 319 }
 320
 321 static inline struct list_head *page_deferred_list(struct page *page)
 322 {
 323         /*
 324          * Global or memcg deferred list in the second tail pages is
 325          * occupied by compound_head.
 326          */
 327         return &page[2].deferred_list;
 328 }
 329
 330 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 331 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 332 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 333 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 334
 335 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 336 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 337 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 338
 339 static inline struct page *thp_head(struct page *page)
 340 {
 341         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 342         return page;
 343 }
 344
 345 static inline unsigned int thp_order(struct page *page)
 346 {
 347         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 348         return 0;
 349 }
 350
 351 static inline int thp_nr_pages(struct page *page)
 352 {
 353         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 354         return 1;
 355 }
 356
 357 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 358 {
 359         return false;
 360 }
 361
 362 static inline bool transparent_hugepage_active(struct vm_area_struct *vma)
 363 {
 364         return false;
 365 }
 366
 367 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 368                 unsigned long haddr)
 369 {
 370         return false;
 371 }
 372
 373 static inline bool transhuge_vma_enabled(struct vm_area_struct *vma,
 374                                           unsigned long vm_flags)
 375 {
 376         return false;
 377 }
 378
 379 static inline void prep_transhuge_page(struct page *page) {}
 380
 381 static inline bool is_transparent_hugepage(struct page *page)
 382 {
 383         return false;
 384 }
 385
 386 #define transparent_hugepage_flags 0UL
 387
 388 #define thp_get_unmapped_area   NULL
 389
 390 static inline bool
 391 can_split_huge_page(struct page *page, int *pextra_pins)
 392 {
 393         BUILD_BUG();
 394         return false;
 395 }
 396 static inline int
 397 split_huge_page_to_list(struct page *page, struct list_head *list)
 398 {
 399         return 0;
 400 }
 401 static inline int split_huge_page(struct page *page)
 402 {
 403         return 0;
 404 }
 405 static inline void deferred_split_huge_page(struct page *page) {}
 406 #define split_huge_pmd(__vma, __pmd, __address) \
 407         do { } while (0)
 408
 409 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 410                 unsigned long address, bool freeze, struct page *page) {}
 411 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 412                 unsigned long address, bool freeze, struct page *page) {}
 413
 414 #define split_huge_pud(__vma, __pmd, __address) \
 415         do { } while (0)
 416
 417 static inline int hugepage_madvise(struct vm_area_struct *vma,
 418                                    unsigned long *vm_flags, int advice)
 419 {
 420         BUG();
 421         return 0;
 422 }
 423 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 424                                          unsigned long start,
 425                                          unsigned long end,
 426                                          long adjust_next)
 427 {
 428 }
 429 static inline int is_swap_pmd(pmd_t pmd)
 430 {
 431         return 0;
 432 }
 433 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 434                 struct vm_area_struct *vma)
 435 {
 436         return NULL;
 437 }
 438 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 439                 struct vm_area_struct *vma)
 440 {
 441         return NULL;
 442 }
 443
 444 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 445 {
 446         return 0;
 447 }
 448
 449 static inline bool is_huge_zero_page(struct page *page)
 450 {
 451         return false;
 452 }
 453
 454 static inline bool is_huge_zero_pmd(pmd_t pmd)
 455 {
 456         return false;
 457 }
 458
 459 static inline bool is_huge_zero_pud(pud_t pud)
 460 {
 461         return false;
 462 }
 463
 464 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 465 {
 466         return;
 467 }
 468
 469 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 470         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 471 {
 472         return NULL;
 473 }
 474
 475 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 476         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 477 {
 478         return NULL;
 479 }
 480
 481 static inline bool thp_migration_supported(void)
 482 {
 483         return false;
 484 }
 485 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 486
 487 /**
 488  * thp_size - Size of a transparent huge page.
 489  * @page: Head page of a transparent huge page.
 490  *
 491  * Return: Number of bytes in this page.
 492  */
 493 static inline unsigned long thp_size(struct page *page)
 494 {
 495         return PAGE_SIZE << thp_order(page);
 496 }
 497
 498 #endif /* _LINUX_HUGE_MM_H */