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 extern vm_fault_t do_huge_pmd_anonymous_page(struct vm_fault *vmf);
  11 extern 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 extern void huge_pmd_set_accessed(struct vm_fault *vmf, pmd_t orig_pmd);
  15 extern 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 extern 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 extern vm_fault_t do_huge_pmd_wp_page(struct vm_fault *vmf, pmd_t orig_pmd);
  28 extern struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
  29                                           unsigned long addr,
  30                                           pmd_t *pmd,
  31                                           unsigned int flags);
  32 extern bool madvise_free_huge_pmd(struct mmu_gather *tlb,
  33                         struct vm_area_struct *vma,
  34                         pmd_t *pmd, unsigned long addr, unsigned long next);
  35 extern int zap_huge_pmd(struct mmu_gather *tlb,
  36                         struct vm_area_struct *vma,
  37                         pmd_t *pmd, unsigned long addr);
  38 extern int zap_huge_pud(struct mmu_gather *tlb,
  39                         struct vm_area_struct *vma,
  40                         pud_t *pud, unsigned long addr);
  41 extern int mincore_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
  42                         unsigned long addr, unsigned long end,
  43                         unsigned char *vec);
  44 extern bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr,
  45                          unsigned long new_addr, unsigned long old_end,
  46                          pmd_t *old_pmd, pmd_t *new_pmd);
  47 extern int change_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
  48                         unsigned long addr, pgprot_t newprot,
  49                         int prot_numa);
  50 vm_fault_t vmf_insert_pfn_pmd(struct vm_fault *vmf, pfn_t pfn, bool write);
  51 vm_fault_t vmf_insert_pfn_pud(struct vm_fault *vmf, pfn_t pfn, bool write);
  52 enum transparent_hugepage_flag {
  53         TRANSPARENT_HUGEPAGE_FLAG,
  54         TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG,
  55         TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
  56         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
  57         TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
  58         TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
  59         TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG,
  60         TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG,
  61 #ifdef CONFIG_DEBUG_VM
  62         TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG,
  63 #endif
  64 };
  65
  66 struct kobject;
  67 struct kobj_attribute;
  68
  69 extern ssize_t single_hugepage_flag_store(struct kobject *kobj,
  70                                  struct kobj_attribute *attr,
  71                                  const char *buf, size_t count,
  72                                  enum transparent_hugepage_flag flag);
  73 extern ssize_t single_hugepage_flag_show(struct kobject *kobj,
  74                                 struct kobj_attribute *attr, char *buf,
  75                                 enum transparent_hugepage_flag flag);
  76 extern struct kobj_attribute shmem_enabled_attr;
  77
  78 #define HPAGE_PMD_ORDER (HPAGE_PMD_SHIFT-PAGE_SHIFT)
  79 #define HPAGE_PMD_NR (1<<HPAGE_PMD_ORDER)
  80
  81 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
  82 #define HPAGE_PMD_SHIFT PMD_SHIFT
  83 #define HPAGE_PMD_SIZE  ((1UL) << HPAGE_PMD_SHIFT)
  84 #define HPAGE_PMD_MASK  (~(HPAGE_PMD_SIZE - 1))
  85
  86 #define HPAGE_PUD_SHIFT PUD_SHIFT
  87 #define HPAGE_PUD_SIZE  ((1UL) << HPAGE_PUD_SHIFT)
  88 #define HPAGE_PUD_MASK  (~(HPAGE_PUD_SIZE - 1))
  89
  90 extern bool is_vma_temporary_stack(struct vm_area_struct *vma);
  91
  92 extern unsigned long transparent_hugepage_flags;
  93
  94 /*
  95  * to be used on vmas which are known to support THP.
  96  * Use transparent_hugepage_enabled otherwise
  97  */
  98 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
  99 {
 100         if (vma->vm_flags & VM_NOHUGEPAGE)
 101                 return false;
 102
 103         if (is_vma_temporary_stack(vma))
 104                 return false;
 105
 106         if (test_bit(MMF_DISABLE_THP, &vma->vm_mm->flags))
 107                 return false;
 108
 109         if (transparent_hugepage_flags & (1 << TRANSPARENT_HUGEPAGE_FLAG))
 110                 return true;
 111         /*
 112          * For dax vmas, try to always use hugepage mappings. If the kernel does
 113          * not support hugepages, fsdax mappings will fallback to PAGE_SIZE
 114          * mappings, and device-dax namespaces, that try to guarantee a given
 115          * mapping size, will fail to enable
 116          */
 117         if (vma_is_dax(vma))
 118                 return true;
 119
 120         if (transparent_hugepage_flags &
 121                                 (1 << TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG))
 122                 return !!(vma->vm_flags & VM_HUGEPAGE);
 123
 124         return false;
 125 }
 126
 127 bool transparent_hugepage_enabled(struct vm_area_struct *vma);
 128
 129 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
 130
 131 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 132                 unsigned long haddr)
 133 {
 134         /* Don't have to check pgoff for anonymous vma */
 135         if (!vma_is_anonymous(vma)) {
 136                 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
 137                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
 138                         return false;
 139         }
 140
 141         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 142                 return false;
 143         return true;
 144 }
 145
 146 #define transparent_hugepage_use_zero_page()                            \
 147         (transparent_hugepage_flags &                                   \
 148          (1<<TRANSPARENT_HUGEPAGE_USE_ZERO_PAGE_FLAG))
 149 #ifdef CONFIG_DEBUG_VM
 150 #define transparent_hugepage_debug_cow()                                \
 151         (transparent_hugepage_flags &                                   \
 152          (1<<TRANSPARENT_HUGEPAGE_DEBUG_COW_FLAG))
 153 #else /* CONFIG_DEBUG_VM */
 154 #define transparent_hugepage_debug_cow() 0
 155 #endif /* CONFIG_DEBUG_VM */
 156
 157 extern unsigned long thp_get_unmapped_area(struct file *filp,
 158                 unsigned long addr, unsigned long len, unsigned long pgoff,
 159                 unsigned long flags);
 160
 161 extern void prep_transhuge_page(struct page *page);
 162 extern void free_transhuge_page(struct page *page);
 163 bool is_transparent_hugepage(struct page *page);
 164
 165 bool can_split_huge_page(struct page *page, int *pextra_pins);
 166 int split_huge_page_to_list(struct page *page, struct list_head *list);
 167 static inline int split_huge_page(struct page *page)
 168 {
 169         return split_huge_page_to_list(page, NULL);
 170 }
 171 void deferred_split_huge_page(struct page *page);
 172
 173 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 174                 unsigned long address, bool freeze, struct page *page);
 175
 176 #define split_huge_pmd(__vma, __pmd, __address)                         \
 177         do {                                                            \
 178                 pmd_t *____pmd = (__pmd);                               \
 179                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 180                                         || pmd_devmap(*____pmd))        \
 181                         __split_huge_pmd(__vma, __pmd, __address,       \
 182                                                 false, NULL);           \
 183         }  while (0)
 184
 185
 186 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 187                 bool freeze, struct page *page);
 188
 189 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 190                 unsigned long address);
 191
 192 #define split_huge_pud(__vma, __pud, __address)                         \
 193         do {                                                            \
 194                 pud_t *____pud = (__pud);                               \
 195                 if (pud_trans_huge(*____pud)                            \
 196                                         || pud_devmap(*____pud))        \
 197                         __split_huge_pud(__vma, __pud, __address);      \
 198         }  while (0)
 199
 200 extern int hugepage_madvise(struct vm_area_struct *vma,
 201                             unsigned long *vm_flags, int advice);
 202 extern void vma_adjust_trans_huge(struct vm_area_struct *vma,
 203                                     unsigned long start,
 204                                     unsigned long end,
 205                                     long adjust_next);
 206 extern spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd,
 207                 struct vm_area_struct *vma);
 208 extern spinlock_t *__pud_trans_huge_lock(pud_t *pud,
 209                 struct vm_area_struct *vma);
 210
 211 static inline int is_swap_pmd(pmd_t pmd)
 212 {
 213         return !pmd_none(pmd) && !pmd_present(pmd);
 214 }
 215
 216 /* mmap_sem must be held on entry */
 217 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 218                 struct vm_area_struct *vma)
 219 {
 220         if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd))
 221                 return __pmd_trans_huge_lock(pmd, vma);
 222         else
 223                 return NULL;
 224 }
 225 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 226                 struct vm_area_struct *vma)
 227 {
 228         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 229                 return __pud_trans_huge_lock(pud, vma);
 230         else
 231                 return NULL;
 232 }
 233 static inline int hpage_nr_pages(struct page *page)
 234 {
 235         if (unlikely(PageTransHuge(page)))
 236                 return HPAGE_PMD_NR;
 237         return 1;
 238 }
 239
 240 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 241                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 242 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 243                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 244
 245 extern vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 246
 247 extern struct page *huge_zero_page;
 248
 249 static inline bool is_huge_zero_page(struct page *page)
 250 {
 251         return READ_ONCE(huge_zero_page) == page;
 252 }
 253
 254 static inline bool is_huge_zero_pmd(pmd_t pmd)
 255 {
 256         return is_huge_zero_page(pmd_page(pmd));
 257 }
 258
 259 static inline bool is_huge_zero_pud(pud_t pud)
 260 {
 261         return false;
 262 }
 263
 264 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 265 void mm_put_huge_zero_page(struct mm_struct *mm);
 266
 267 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 268
 269 static inline bool thp_migration_supported(void)
 270 {
 271         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 272 }
 273
 274 static inline struct list_head *page_deferred_list(struct page *page)
 275 {
 276         /*
 277          * Global or memcg deferred list in the second tail pages is
 278          * occupied by compound_head.
 279          */
 280         return &page[2].deferred_list;
 281 }
 282
 283 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 284 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 285 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 286 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 287
 288 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 289 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 290 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 291
 292 #define hpage_nr_pages(x) 1
 293
 294 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 295 {
 296         return false;
 297 }
 298
 299 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
 300 {
 301         return false;
 302 }
 303
 304 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 305                 unsigned long haddr)
 306 {
 307         return false;
 308 }
 309
 310 static inline void prep_transhuge_page(struct page *page) {}
 311
 312 static inline bool is_transparent_hugepage(struct page *page)
 313 {
 314         return false;
 315 }
 316
 317 #define transparent_hugepage_flags 0UL
 318
 319 #define thp_get_unmapped_area   NULL
 320
 321 static inline bool
 322 can_split_huge_page(struct page *page, int *pextra_pins)
 323 {
 324         BUILD_BUG();
 325         return false;
 326 }
 327 static inline int
 328 split_huge_page_to_list(struct page *page, struct list_head *list)
 329 {
 330         return 0;
 331 }
 332 static inline int split_huge_page(struct page *page)
 333 {
 334         return 0;
 335 }
 336 static inline void deferred_split_huge_page(struct page *page) {}
 337 #define split_huge_pmd(__vma, __pmd, __address) \
 338         do { } while (0)
 339
 340 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 341                 unsigned long address, bool freeze, struct page *page) {}
 342 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 343                 unsigned long address, bool freeze, struct page *page) {}
 344
 345 #define split_huge_pud(__vma, __pmd, __address) \
 346         do { } while (0)
 347
 348 static inline int hugepage_madvise(struct vm_area_struct *vma,
 349                                    unsigned long *vm_flags, int advice)
 350 {
 351         BUG();
 352         return 0;
 353 }
 354 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 355                                          unsigned long start,
 356                                          unsigned long end,
 357                                          long adjust_next)
 358 {
 359 }
 360 static inline int is_swap_pmd(pmd_t pmd)
 361 {
 362         return 0;
 363 }
 364 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 365                 struct vm_area_struct *vma)
 366 {
 367         return NULL;
 368 }
 369 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 370                 struct vm_area_struct *vma)
 371 {
 372         return NULL;
 373 }
 374
 375 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 376                 pmd_t orig_pmd)
 377 {
 378         return 0;
 379 }
 380
 381 static inline bool is_huge_zero_page(struct page *page)
 382 {
 383         return false;
 384 }
 385
 386 static inline bool is_huge_zero_pud(pud_t pud)
 387 {
 388         return false;
 389 }
 390
 391 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 392 {
 393         return;
 394 }
 395
 396 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 397         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 398 {
 399         return NULL;
 400 }
 401
 402 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 403         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 404 {
 405         return NULL;
 406 }
 407
 408 static inline bool thp_migration_supported(void)
 409 {
 410         return false;
 411 }
 412 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 413
 414 #endif /* _LINUX_HUGE_MM_H */