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 #ifdef CONFIG_DEBUG_VM
  91         TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG,
  92 #endif
  93 };
  94
  95 struct kobject;
  96 struct kobj_attribute;
  97
  98 ssize_t single_hugepage_flag_store(struct kobject *kobj,
  99                                    struct kobj_attribute *attr,
 100                                    const char *buf, size_t count,
 101                                    enum transparent_hugepage_flag flag);
 102 ssize_t single_hugepage_flag_show(struct kobject *kobj,
 103                                   struct kobj_attribute *attr, char *buf,
 104                                   enum transparent_hugepage_flag flag);
 105 extern struct kobj_attribute shmem_enabled_attr;
 106
 107 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
 108 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
 109
 110 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 111 #define HPAGE_PMD_SHIFT PMD_SHIFT
 112 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
 113 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
 114
 115 #define HPAGE_PUD_SHIFT PUD_SHIFT
 116 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
 117 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
 118
 119 extern unsigned long transparent_hugepage_flags;
 120
 121 /*
 122  * to be used on vmas which are known to support THP.
 123  * Use transparent_hugepage_enabled otherwise
 124  */
 125 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 126 {
 127
 128         /*
 129          * If the hardware/firmware marked hugepage support disabled.
 130          */
 131         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_NEVER_DAX))
 132                 return false;
 133
 134         if (vma->vm_flags & VM_NOHUGEPAGE)
 135                 return false;
 136
 137         if (vma_is_temporary_stack(vma))
 138                 return false;
 139
 140         if (test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 141                 return false;
 142
 143         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 144                 return true;
 145
 146         if (vma_is_dax(vma))
 147                 return true;
 148
 149         if (transparent_hugepage_flags &
 150                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 151                 return !!(vma->vm_flags & VM_HUGEPAGE);
 152
 153         return false;
 154 }
 155
 156 bool transparent_hugepage_enabled(struct vm_area_struct *vma);
 157
 158 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
 159
 160 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 161                 unsigned long haddr)
 162 {
 163         /* Don't have to check pgoff for anonymous vma */
 164         if (!vma_is_anonymous(vma)) {
 165                 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
 166                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
 167                         return false;
 168         }
 169
 170         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 171                 return false;
 172         return true;
 173 }
 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 bool is_transparent_hugepage(struct page *page);
 185
 186 bool can_split_huge_page(struct page *page, int *pextra_pins);
 187 int split_huge_page_to_list(struct page *page, struct list_head *list);
 188 static inline int split_huge_page(struct page *page)
 189 {
 190         return split_huge_page_to_list(page, NULL);
 191 }
 192 void deferred_split_huge_page(struct page *page);
 193
 194 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 195                 unsigned long address, bool freeze, struct page *page);
 196
 197 #define split_huge_pmd(__vma, __pmd, __address)                         \
 198         do {                                                            \
 199                 pmd_t *____pmd = (__pmd);                               \
 200                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 201                                         || pmd_devmap(*____pmd))        \
 202                         __split_huge_pmd(__vma, __pmd, __address,       \
 203                                                 false, NULL);           \
 204         }  while (0)
 205
 206
 207 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 208                 bool freeze, struct page *page);
 209
 210 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 211                 unsigned long address);
 212
 213 #define split_huge_pud(__vma, __pud, __address)                         \
 214         do {                                                            \
 215                 pud_t *____pud = (__pud);                               \
 216                 if (pud_trans_huge(*____pud)                            \
 217                                         || pud_devmap(*____pud))        \
 218                         __split_huge_pud(__vma, __pud, __address);      \
 219         }  while (0)
 220
 221 int hugepage_madvise(struct vm_area_struct *vma, unsigned long *vm_flags,
 222                      int advice);
 223 void vma_adjust_trans_huge(struct vm_area_struct *vma, unsigned long start,
 224                            unsigned long end, long adjust_next);
 225 spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd, struct vm_area_struct *vma);
 226 spinlock_t *__pud_trans_huge_lock(pud_t *pud, struct vm_area_struct *vma);
 227
 228 static inline int is_swap_pmd(pmd_t pmd)
 229 {
 230         return !pmd_none(pmd) && !pmd_present(pmd);
 231 }
 232
 233 /* mmap_lock must be held on entry */
 234 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 235                 struct vm_area_struct *vma)
 236 {
 237         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 238                 return __pmd_trans_huge_lock(pmd, vma);
 239         else
 240                 return NULL;
 241 }
 242 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 243                 struct vm_area_struct *vma)
 244 {
 245         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 246                 return __pud_trans_huge_lock(pud, vma);
 247         else
 248                 return NULL;
 249 }
 250
 251 /**
 252  * thp_head - Head page of a transparent huge page.
 253  * @page: Any page (tail, head or regular) found in the page cache.
 254  */
 255 static inline struct page *thp_head(struct page *page)
 256 {
 257         return compound_head(page);
 258 }
 259
 260 /**
 261  * thp_order - Order of a transparent huge page.
 262  * @page: Head page of a transparent huge page.
 263  */
 264 static inline unsigned int thp_order(struct page *page)
 265 {
 266         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 267         if (PageHead(page))
 268                 return HPAGE_PMD_ORDER;
 269         return 0;
 270 }
 271
 272 /**
 273  * thp_nr_pages - The number of regular pages in this huge page.
 274  * @page: The head page of a huge page.
 275  */
 276 static inline int thp_nr_pages(struct page *page)
 277 {
 278         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 279         if (PageHead(page))
 280                 return HPAGE_PMD_NR;
 281         return 1;
 282 }
 283
 284 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 285                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 286 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 287                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 288
 289 vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 290
 291 extern struct page *huge_zero_page;
 292
 293 static inline bool is_huge_zero_page(struct page *page)
 294 {
 295         return READ_ONCE(huge_zero_page) == page;
 296 }
 297
 298 static inline bool is_huge_zero_pmd(pmd_t pmd)
 299 {
 300         return is_huge_zero_page(pmd_page(pmd));
 301 }
 302
 303 static inline bool is_huge_zero_pud(pud_t pud)
 304 {
 305         return false;
 306 }
 307
 308 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 309 void mm_put_huge_zero_page(struct mm_struct *mm);
 310
 311 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 312
 313 static inline bool thp_migration_supported(void)
 314 {
 315         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 316 }
 317
 318 static inline struct list_head *page_deferred_list(struct page *page)
 319 {
 320         /*
 321          * Global or memcg deferred list in the second tail pages is
 322          * occupied by compound_head.
 323          */
 324         return &page[2].deferred_list;
 325 }
 326
 327 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 328 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 329 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 330 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 331
 332 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 333 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 334 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 335
 336 static inline struct page *thp_head(struct page *page)
 337 {
 338         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 339         return page;
 340 }
 341
 342 static inline unsigned int thp_order(struct page *page)
 343 {
 344         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 345         return 0;
 346 }
 347
 348 static inline int thp_nr_pages(struct page *page)
 349 {
 350         VM_BUG_ON_PGFLAGS(PageTail(page), page);
 351         return 1;
 352 }
 353
 354 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 355 {
 356         return false;
 357 }
 358
 359 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
 360 {
 361         return false;
 362 }
 363
 364 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 365                 unsigned long haddr)
 366 {
 367         return false;
 368 }
 369
 370 static inline void prep_transhuge_page(struct page *page) {}
 371
 372 static inline bool is_transparent_hugepage(struct page *page)
 373 {
 374         return false;
 375 }
 376
 377 #define transparent_hugepage_flags 0UL
 378
 379 #define thp_get_unmapped_area   NULL
 380
 381 static inline bool
 382 can_split_huge_page(struct page *page, int *pextra_pins)
 383 {
 384         BUILD_BUG();
 385         return false;
 386 }
 387 static inline int
 388 split_huge_page_to_list(struct page *page, struct list_head *list)
 389 {
 390         return 0;
 391 }
 392 static inline int split_huge_page(struct page *page)
 393 {
 394         return 0;
 395 }
 396 static inline void deferred_split_huge_page(struct page *page) {}
 397 #define split_huge_pmd(__vma, __pmd, __address) \
 398         do { } while (0)
 399
 400 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 401                 unsigned long address, bool freeze, struct page *page) {}
 402 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 403                 unsigned long address, bool freeze, struct page *page) {}
 404
 405 #define split_huge_pud(__vma, __pmd, __address) \
 406         do { } while (0)
 407
 408 static inline int hugepage_madvise(struct vm_area_struct *vma,
 409                                    unsigned long *vm_flags, int advice)
 410 {
 411         BUG();
 412         return 0;
 413 }
 414 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 415                                          unsigned long start,
 416                                          unsigned long end,
 417                                          long adjust_next)
 418 {
 419 }
 420 static inline int is_swap_pmd(pmd_t pmd)
 421 {
 422         return 0;
 423 }
 424 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 425                 struct vm_area_struct *vma)
 426 {
 427         return NULL;
 428 }
 429 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 430                 struct vm_area_struct *vma)
 431 {
 432         return NULL;
 433 }
 434
 435 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 436                 pmd_t orig_pmd)
 437 {
 438         return 0;
 439 }
 440
 441 static inline bool is_huge_zero_page(struct page *page)
 442 {
 443         return false;
 444 }
 445
 446 static inline bool is_huge_zero_pud(pud_t pud)
 447 {
 448         return false;
 449 }
 450
 451 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 452 {
 453         return;
 454 }
 455
 456 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 457         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 458 {
 459         return NULL;
 460 }
 461
 462 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 463         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 464 {
 465         return NULL;
 466 }
 467
 468 static inline bool thp_migration_supported(void)
 469 {
 470         return false;
 471 }
 472 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 473
 474 /**
 475  * thp_size - Size of a transparent huge page.
 476  * @page: Head page of a transparent huge page.
 477  *
 478  * Return: Number of bytes in this page.
 479  */
 480 static inline unsigned long thp_size(struct page *page)
 481 {
 482         return PAGE_SIZE << thp_order(page);
 483 }
 484
 485 #endif /* _LINUX_HUGE_MM_H */