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
 164 bool can_split_huge_page(struct page *page, int *pextra_pins);
 165 int split_huge_page_to_list(struct page *page, struct list_head *list);
 166 static inline int split_huge_page(struct page *page)
 167 {
 168         return split_huge_page_to_list(page, NULL);
 169 }
 170 void deferred_split_huge_page(struct page *page);
 171
 172 void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 173                 unsigned long address, bool freeze, struct page *page);
 174
 175 #define split_huge_pmd(__vma, __pmd, __address)                         \
 176         do {                                                            \
 177                 pmd_t *____pmd = (__pmd);                               \
 178                 if (is_swap_pmd(*____pmd) || pmd_trans_huge(*____pmd)   \
 179                                         || pmd_devmap(*____pmd))        \
 180                         __split_huge_pmd(__vma, __pmd, __address,       \
 181                                                 false, NULL);           \
 182         }  while (0)
 183
 184
 185 void split_huge_pmd_address(struct vm_area_struct *vma, unsigned long address,
 186                 bool freeze, struct page *page);
 187
 188 void __split_huge_pud(struct vm_area_struct *vma, pud_t *pud,
 189                 unsigned long address);
 190
 191 #define split_huge_pud(__vma, __pud, __address)                         \
 192         do {                                                            \
 193                 pud_t *____pud = (__pud);                               \
 194                 if (pud_trans_huge(*____pud)                            \
 195                                         || pud_devmap(*____pud))        \
 196                         __split_huge_pud(__vma, __pud, __address);      \
 197         }  while (0)
 198
 199 extern int hugepage_madvise(struct vm_area_struct *vma,
 200                             unsigned long *vm_flags, int advice);
 201 extern void vma_adjust_trans_huge(struct vm_area_struct *vma,
 202                                     unsigned long start,
 203                                     unsigned long end,
 204                                     long adjust_next);
 205 extern spinlock_t *__pmd_trans_huge_lock(pmd_t *pmd,
 206                 struct vm_area_struct *vma);
 207 extern spinlock_t *__pud_trans_huge_lock(pud_t *pud,
 208                 struct vm_area_struct *vma);
 209
 210 static inline int is_swap_pmd(pmd_t pmd)
 211 {
 212         return !pmd_none(pmd) && !pmd_present(pmd);
 213 }
 214
 215 /* mmap_sem must be held on entry */
 216 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 217                 struct vm_area_struct *vma)
 218 {
 219         VM_BUG_ON_VMA(!rwsem_is_locked(&vma->vm_mm->mmap_sem), vma);
 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         VM_BUG_ON_VMA(!rwsem_is_locked(&vma->vm_mm->mmap_sem), vma);
 229         if (pud_trans_huge(*pud) || pud_devmap(*pud))
 230                 return __pud_trans_huge_lock(pud, vma);
 231         else
 232                 return NULL;
 233 }
 234 static inline int hpage_nr_pages(struct page *page)
 235 {
 236         if (unlikely(PageTransHuge(page)))
 237                 return HPAGE_PMD_NR;
 238         return 1;
 239 }
 240
 241 struct page *follow_devmap_pmd(struct vm_area_struct *vma, unsigned long addr,
 242                 pmd_t *pmd, int flags, struct dev_pagemap **pgmap);
 243 struct page *follow_devmap_pud(struct vm_area_struct *vma, unsigned long addr,
 244                 pud_t *pud, int flags, struct dev_pagemap **pgmap);
 245
 246 extern vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf, pmd_t orig_pmd);
 247
 248 extern struct page *huge_zero_page;
 249 extern unsigned long huge_zero_pfn;
 250
 251 static inline bool is_huge_zero_page(struct page *page)
 252 {
 253         return READ_ONCE(huge_zero_page) == page;
 254 }
 255
 256 static inline bool is_huge_zero_pmd(pmd_t pmd)
 257 {
 258         return READ_ONCE(huge_zero_pfn) == pmd_pfn(pmd) && pmd_present(pmd);
 259 }
 260
 261 static inline bool is_huge_zero_pud(pud_t pud)
 262 {
 263         return false;
 264 }
 265
 266 struct page *mm_get_huge_zero_page(struct mm_struct *mm);
 267 void mm_put_huge_zero_page(struct mm_struct *mm);
 268
 269 #define mk_huge_pmd(page, prot) pmd_mkhuge(mk_pmd(page, prot))
 270
 271 static inline bool thp_migration_supported(void)
 272 {
 273         return IS_ENABLED(CONFIG_ARCH_ENABLE_THP_MIGRATION);
 274 }
 275
 276 static inline struct list_head *page_deferred_list(struct page *page)
 277 {
 278         /*
 279          * Global or memcg deferred list in the second tail pages is
 280          * occupied by compound_head.
 281          */
 282         return &page[2].deferred_list;
 283 }
 284
 285 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
 286 #define HPAGE_PMD_SHIFT ({ BUILD_BUG(); 0; })
 287 #define HPAGE_PMD_MASK ({ BUILD_BUG(); 0; })
 288 #define HPAGE_PMD_SIZE ({ BUILD_BUG(); 0; })
 289
 290 #define HPAGE_PUD_SHIFT ({ BUILD_BUG(); 0; })
 291 #define HPAGE_PUD_MASK ({ BUILD_BUG(); 0; })
 292 #define HPAGE_PUD_SIZE ({ BUILD_BUG(); 0; })
 293
 294 #define hpage_nr_pages(x) 1
 295
 296 static inline bool __transparent_hugepage_enabled(struct vm_area_struct *vma)
 297 {
 298         return false;
 299 }
 300
 301 static inline bool transparent_hugepage_enabled(struct vm_area_struct *vma)
 302 {
 303         return false;
 304 }
 305
 306 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 307                 unsigned long haddr)
 308 {
 309         return false;
 310 }
 311
 312 static inline void prep_transhuge_page(struct page *page) {}
 313
 314 #define transparent_hugepage_flags 0UL
 315
 316 #define thp_get_unmapped_area   NULL
 317
 318 static inline bool
 319 can_split_huge_page(struct page *page, int *pextra_pins)
 320 {
 321         BUILD_BUG();
 322         return false;
 323 }
 324 static inline int
 325 split_huge_page_to_list(struct page *page, struct list_head *list)
 326 {
 327         return 0;
 328 }
 329 static inline int split_huge_page(struct page *page)
 330 {
 331         return 0;
 332 }
 333 static inline void deferred_split_huge_page(struct page *page) {}
 334 #define split_huge_pmd(__vma, __pmd, __address) \
 335         do { } while (0)
 336
 337 static inline void __split_huge_pmd(struct vm_area_struct *vma, pmd_t *pmd,
 338                 unsigned long address, bool freeze, struct page *page) {}
 339 static inline void split_huge_pmd_address(struct vm_area_struct *vma,
 340                 unsigned long address, bool freeze, struct page *page) {}
 341
 342 #define split_huge_pud(__vma, __pmd, __address) \
 343         do { } while (0)
 344
 345 static inline int hugepage_madvise(struct vm_area_struct *vma,
 346                                    unsigned long *vm_flags, int advice)
 347 {
 348         BUG();
 349         return 0;
 350 }
 351 static inline void vma_adjust_trans_huge(struct vm_area_struct *vma,
 352                                          unsigned long start,
 353                                          unsigned long end,
 354                                          long adjust_next)
 355 {
 356 }
 357 static inline int is_swap_pmd(pmd_t pmd)
 358 {
 359         return 0;
 360 }
 361 static inline spinlock_t *pmd_trans_huge_lock(pmd_t *pmd,
 362                 struct vm_area_struct *vma)
 363 {
 364         return NULL;
 365 }
 366 static inline spinlock_t *pud_trans_huge_lock(pud_t *pud,
 367                 struct vm_area_struct *vma)
 368 {
 369         return NULL;
 370 }
 371
 372 static inline vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf,
 373                 pmd_t orig_pmd)
 374 {
 375         return 0;
 376 }
 377
 378 static inline bool is_huge_zero_page(struct page *page)
 379 {
 380         return false;
 381 }
 382
 383 static inline bool is_huge_zero_pmd(pmd_t pmd)
 384 {
 385         return false;
 386 }
 387
 388 static inline bool is_huge_zero_pud(pud_t pud)
 389 {
 390         return false;
 391 }
 392
 393 static inline void mm_put_huge_zero_page(struct mm_struct *mm)
 394 {
 395         return;
 396 }
 397
 398 static inline struct page *follow_devmap_pmd(struct vm_area_struct *vma,
 399         unsigned long addr, pmd_t *pmd, int flags, struct dev_pagemap **pgmap)
 400 {
 401         return NULL;
 402 }
 403
 404 static inline struct page *follow_devmap_pud(struct vm_area_struct *vma,
 405         unsigned long addr, pud_t *pud, int flags, struct dev_pagemap **pgmap)
 406 {
 407         return NULL;
 408 }
 409
 410 static inline bool thp_migration_supported(void)
 411 {
 412         return false;
 413 }
 414 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
 415
 416 #endif /* _LINUX_HUGE_MM_H */