mm/madvise.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *      linux/mm/madvise.c
   4  *
   5  * Copyright (C) 1999  Linus Torvalds
   6  * Copyright (C) 2002  Christoph Hellwig
   7  */
   8
   9 #include <linux/mman.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/syscalls.h>
  12 #include <linux/mempolicy.h>
  13 #include <linux/page-isolation.h>
  14 #include <linux/page_idle.h>
  15 #include <linux/userfaultfd_k.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/falloc.h>
  18 #include <linux/fadvise.h>
  19 #include <linux/sched.h>
  20 #include <linux/sched/mm.h>
  21 #include <linux/mm_inline.h>
  22 #include <linux/string.h>
  23 #include <linux/uio.h>
  24 #include <linux/ksm.h>
  25 #include <linux/fs.h>
  26 #include <linux/file.h>
  27 #include <linux/blkdev.h>
  28 #include <linux/backing-dev.h>
  29 #include <linux/pagewalk.h>
  30 #include <linux/swap.h>
  31 #include <linux/swapops.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/mmu_notifier.h>
  34
  35 #include <asm/tlb.h>
  36
  37 #include "internal.h"
  38 #include "swap.h"
  39
  40 struct madvise_walk_private {
  41         struct mmu_gather *tlb;
  42         bool pageout;
  43 };
  44
  45 /*
  46  * Any behaviour which results in changes to the vma->vm_flags needs to
  47  * take mmap_lock for writing. Others, which simply traverse vmas, need
  48  * to only take it for reading.
  49  */
  50 static int madvise_need_mmap_write(int behavior)
  51 {
  52         switch (behavior) {
  53         case MADV_REMOVE:
  54         case MADV_WILLNEED:
  55         case MADV_DONTNEED:
  56         case MADV_DONTNEED_LOCKED:
  57         case MADV_COLD:
  58         case MADV_PAGEOUT:
  59         case MADV_FREE:
  60         case MADV_POPULATE_READ:
  61         case MADV_POPULATE_WRITE:
  62         case MADV_COLLAPSE:
  63                 return 0;
  64         default:
  65                 /* be safe, default to 1. list exceptions explicitly */
  66                 return 1;
  67         }
  68 }
  69
  70 #ifdef CONFIG_ANON_VMA_NAME
  71 struct anon_vma_name *anon_vma_name_alloc(const char *name)
  72 {
  73         struct anon_vma_name *anon_name;
  74         size_t count;
  75
  76         /* Add 1 for NUL terminator at the end of the anon_name->name */
  77         count = strlen(name) + 1;
  78         anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
  79         if (anon_name) {
  80                 kref_init(&anon_name->kref);
  81                 memcpy(anon_name->name, name, count);
  82         }
  83
  84         return anon_name;
  85 }
  86
  87 void anon_vma_name_free(struct kref *kref)
  88 {
  89         struct anon_vma_name *anon_name =
  90                         container_of(kref, struct anon_vma_name, kref);
  91         kfree(anon_name);
  92 }
  93
  94 struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
  95 {
  96         mmap_assert_locked(vma->vm_mm);
  97
  98         return vma->anon_name;
  99 }
 100
 101 /* mmap_lock should be write-locked */
 102 static int replace_anon_vma_name(struct vm_area_struct *vma,
 103                                  struct anon_vma_name *anon_name)
 104 {
 105         struct anon_vma_name *orig_name = anon_vma_name(vma);
 106
 107         if (!anon_name) {
 108                 vma->anon_name = NULL;
 109                 anon_vma_name_put(orig_name);
 110                 return 0;
 111         }
 112
 113         if (anon_vma_name_eq(orig_name, anon_name))
 114                 return 0;
 115
 116         vma->anon_name = anon_vma_name_reuse(anon_name);
 117         anon_vma_name_put(orig_name);
 118
 119         return 0;
 120 }
 121 #else /* CONFIG_ANON_VMA_NAME */
 122 static int replace_anon_vma_name(struct vm_area_struct *vma,
 123                                  struct anon_vma_name *anon_name)
 124 {
 125         if (anon_name)
 126                 return -EINVAL;
 127
 128         return 0;
 129 }
 130 #endif /* CONFIG_ANON_VMA_NAME */
 131 /*
 132  * Update the vm_flags on region of a vma, splitting it or merging it as
 133  * necessary.  Must be called with mmap_lock held for writing;
 134  * Caller should ensure anon_name stability by raising its refcount even when
 135  * anon_name belongs to a valid vma because this function might free that vma.
 136  */
 137 static int madvise_update_vma(struct vm_area_struct *vma,
 138                               struct vm_area_struct **prev, unsigned long start,
 139                               unsigned long end, unsigned long new_flags,
 140                               struct anon_vma_name *anon_name)
 141 {
 142         struct mm_struct *mm = vma->vm_mm;
 143         int error;
 144         pgoff_t pgoff;
 145         VMA_ITERATOR(vmi, mm, start);
 146
 147         if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
 148                 *prev = vma;
 149                 return 0;
 150         }
 151
 152         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 153         *prev = vma_merge(&vmi, mm, *prev, start, end, new_flags,
 154                           vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
 155                           vma->vm_userfaultfd_ctx, anon_name);
 156         if (*prev) {
 157                 vma = *prev;
 158                 goto success;
 159         }
 160
 161         *prev = vma;
 162
 163         if (start != vma->vm_start) {
 164                 error = split_vma(&vmi, vma, start, 1);
 165                 if (error)
 166                         return error;
 167         }
 168
 169         if (end != vma->vm_end) {
 170                 error = split_vma(&vmi, vma, end, 0);
 171                 if (error)
 172                         return error;
 173         }
 174
 175 success:
 176         /*
 177          * vm_flags is protected by the mmap_lock held in write mode.
 178          */
 179         vm_flags_reset(vma, new_flags);
 180         if (!vma->vm_file || vma_is_anon_shmem(vma)) {
 181                 error = replace_anon_vma_name(vma, anon_name);
 182                 if (error)
 183                         return error;
 184         }
 185
 186         return 0;
 187 }
 188
 189 #ifdef CONFIG_SWAP
 190 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 191                 unsigned long end, struct mm_walk *walk)
 192 {
 193         struct vm_area_struct *vma = walk->private;
 194         struct swap_iocb *splug = NULL;
 195         pte_t *ptep = NULL;
 196         spinlock_t *ptl;
 197         unsigned long addr;
 198
 199         for (addr = start; addr < end; addr += PAGE_SIZE) {
 200                 pte_t pte;
 201                 swp_entry_t entry;
 202                 struct page *page;
 203
 204                 if (!ptep++) {
 205                         ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 206                         if (!ptep)
 207                                 break;
 208                 }
 209
 210                 pte = ptep_get(ptep);
 211                 if (!is_swap_pte(pte))
 212                         continue;
 213                 entry = pte_to_swp_entry(pte);
 214                 if (unlikely(non_swap_entry(entry)))
 215                         continue;
 216
 217                 pte_unmap_unlock(ptep, ptl);
 218                 ptep = NULL;
 219
 220                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 221                                              vma, addr, false, &splug);
 222                 if (page)
 223                         put_page(page);
 224         }
 225
 226         if (ptep)
 227                 pte_unmap_unlock(ptep, ptl);
 228         swap_read_unplug(splug);
 229         cond_resched();
 230
 231         return 0;
 232 }
 233
 234 static const struct mm_walk_ops swapin_walk_ops = {
 235         .pmd_entry              = swapin_walk_pmd_entry,
 236 };
 237
 238 static void shmem_swapin_range(struct vm_area_struct *vma,
 239                 unsigned long start, unsigned long end,
 240                 struct address_space *mapping)
 241 {
 242         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 243         pgoff_t end_index = linear_page_index(vma, end) - 1;
 244         struct page *page;
 245         struct swap_iocb *splug = NULL;
 246
 247         rcu_read_lock();
 248         xas_for_each(&xas, page, end_index) {
 249                 unsigned long addr;
 250                 swp_entry_t entry;
 251
 252                 if (!xa_is_value(page))
 253                         continue;
 254                 entry = radix_to_swp_entry(page);
 255                 /* There might be swapin error entries in shmem mapping. */
 256                 if (non_swap_entry(entry))
 257                         continue;
 258
 259                 addr = vma->vm_start +
 260                         ((xas.xa_index - vma->vm_pgoff) << PAGE_SHIFT);
 261                 xas_pause(&xas);
 262                 rcu_read_unlock();
 263
 264                 page = read_swap_cache_async(entry, mapping_gfp_mask(mapping),
 265                                              vma, addr, false, &splug);
 266                 if (page)
 267                         put_page(page);
 268
 269                 rcu_read_lock();
 270         }
 271         rcu_read_unlock();
 272         swap_read_unplug(splug);
 273 }
 274 #endif          /* CONFIG_SWAP */
 275
 276 /*
 277  * Schedule all required I/O operations.  Do not wait for completion.
 278  */
 279 static long madvise_willneed(struct vm_area_struct *vma,
 280                              struct vm_area_struct **prev,
 281                              unsigned long start, unsigned long end)
 282 {
 283         struct mm_struct *mm = vma->vm_mm;
 284         struct file *file = vma->vm_file;
 285         loff_t offset;
 286
 287         *prev = vma;
 288 #ifdef CONFIG_SWAP
 289         if (!file) {
 290                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 291                 lru_add_drain(); /* Push any new pages onto the LRU now */
 292                 return 0;
 293         }
 294
 295         if (shmem_mapping(file->f_mapping)) {
 296                 shmem_swapin_range(vma, start, end, file->f_mapping);
 297                 lru_add_drain(); /* Push any new pages onto the LRU now */
 298                 return 0;
 299         }
 300 #else
 301         if (!file)
 302                 return -EBADF;
 303 #endif
 304
 305         if (IS_DAX(file_inode(file))) {
 306                 /* no bad return value, but ignore advice */
 307                 return 0;
 308         }
 309
 310         /*
 311          * Filesystem's fadvise may need to take various locks.  We need to
 312          * explicitly grab a reference because the vma (and hence the
 313          * vma's reference to the file) can go away as soon as we drop
 314          * mmap_lock.
 315          */
 316         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 317         get_file(file);
 318         offset = (loff_t)(start - vma->vm_start)
 319                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 320         mmap_read_unlock(mm);
 321         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 322         fput(file);
 323         mmap_read_lock(mm);
 324         return 0;
 325 }
 326
 327 static inline bool can_do_file_pageout(struct vm_area_struct *vma)
 328 {
 329         if (!vma->vm_file)
 330                 return false;
 331         /*
 332          * paging out pagecache only for non-anonymous mappings that correspond
 333          * to the files the calling process could (if tried) open for writing;
 334          * otherwise we'd be including shared non-exclusive mappings, which
 335          * opens a side channel.
 336          */
 337         return inode_owner_or_capable(&nop_mnt_idmap,
 338                                       file_inode(vma->vm_file)) ||
 339                file_permission(vma->vm_file, MAY_WRITE) == 0;
 340 }
 341
 342 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 343                                 unsigned long addr, unsigned long end,
 344                                 struct mm_walk *walk)
 345 {
 346         struct madvise_walk_private *private = walk->private;
 347         struct mmu_gather *tlb = private->tlb;
 348         bool pageout = private->pageout;
 349         struct mm_struct *mm = tlb->mm;
 350         struct vm_area_struct *vma = walk->vma;
 351         pte_t *start_pte, *pte, ptent;
 352         spinlock_t *ptl;
 353         struct folio *folio = NULL;
 354         LIST_HEAD(folio_list);
 355         bool pageout_anon_only_filter;
 356
 357         if (fatal_signal_pending(current))
 358                 return -EINTR;
 359
 360         pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
 361                                         !can_do_file_pageout(vma);
 362
 363 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 364         if (pmd_trans_huge(*pmd)) {
 365                 pmd_t orig_pmd;
 366                 unsigned long next = pmd_addr_end(addr, end);
 367
 368                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 369                 ptl = pmd_trans_huge_lock(pmd, vma);
 370                 if (!ptl)
 371                         return 0;
 372
 373                 orig_pmd = *pmd;
 374                 if (is_huge_zero_pmd(orig_pmd))
 375                         goto huge_unlock;
 376
 377                 if (unlikely(!pmd_present(orig_pmd))) {
 378                         VM_BUG_ON(thp_migration_supported() &&
 379                                         !is_pmd_migration_entry(orig_pmd));
 380                         goto huge_unlock;
 381                 }
 382
 383                 folio = pfn_folio(pmd_pfn(orig_pmd));
 384
 385                 /* Do not interfere with other mappings of this folio */
 386                 if (folio_mapcount(folio) != 1)
 387                         goto huge_unlock;
 388
 389                 if (pageout_anon_only_filter && !folio_test_anon(folio))
 390                         goto huge_unlock;
 391
 392                 if (next - addr != HPAGE_PMD_SIZE) {
 393                         int err;
 394
 395                         folio_get(folio);
 396                         spin_unlock(ptl);
 397                         folio_lock(folio);
 398                         err = split_folio(folio);
 399                         folio_unlock(folio);
 400                         folio_put(folio);
 401                         if (!err)
 402                                 goto regular_folio;
 403                         return 0;
 404                 }
 405
 406                 if (pmd_young(orig_pmd)) {
 407                         pmdp_invalidate(vma, addr, pmd);
 408                         orig_pmd = pmd_mkold(orig_pmd);
 409
 410                         set_pmd_at(mm, addr, pmd, orig_pmd);
 411                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 412                 }
 413
 414                 folio_clear_referenced(folio);
 415                 folio_test_clear_young(folio);
 416                 if (folio_test_active(folio))
 417                         folio_set_workingset(folio);
 418                 if (pageout) {
 419                         if (folio_isolate_lru(folio)) {
 420                                 if (folio_test_unevictable(folio))
 421                                         folio_putback_lru(folio);
 422                                 else
 423                                         list_add(&folio->lru, &folio_list);
 424                         }
 425                 } else
 426                         folio_deactivate(folio);
 427 huge_unlock:
 428                 spin_unlock(ptl);
 429                 if (pageout)
 430                         reclaim_pages(&folio_list);
 431                 return 0;
 432         }
 433
 434 regular_folio:
 435 #endif
 436         tlb_change_page_size(tlb, PAGE_SIZE);
 437         start_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 438         if (!start_pte)
 439                 return 0;
 440         flush_tlb_batched_pending(mm);
 441         arch_enter_lazy_mmu_mode();
 442         for (; addr < end; pte++, addr += PAGE_SIZE) {
 443                 ptent = ptep_get(pte);
 444
 445                 if (pte_none(ptent))
 446                         continue;
 447
 448                 if (!pte_present(ptent))
 449                         continue;
 450
 451                 folio = vm_normal_folio(vma, addr, ptent);
 452                 if (!folio || folio_is_zone_device(folio))
 453                         continue;
 454
 455                 /*
 456                  * Creating a THP page is expensive so split it only if we
 457                  * are sure it's worth. Split it if we are only owner.
 458                  */
 459                 if (folio_test_large(folio)) {
 460                         int err;
 461
 462                         if (folio_mapcount(folio) != 1)
 463                                 break;
 464                         if (pageout_anon_only_filter && !folio_test_anon(folio))
 465                                 break;
 466                         if (!folio_trylock(folio))
 467                                 break;
 468                         folio_get(folio);
 469                         arch_leave_lazy_mmu_mode();
 470                         pte_unmap_unlock(start_pte, ptl);
 471                         start_pte = NULL;
 472                         err = split_folio(folio);
 473                         folio_unlock(folio);
 474                         folio_put(folio);
 475                         if (err)
 476                                 break;
 477                         start_pte = pte =
 478                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 479                         if (!start_pte)
 480                                 break;
 481                         arch_enter_lazy_mmu_mode();
 482                         pte--;
 483                         addr -= PAGE_SIZE;
 484                         continue;
 485                 }
 486
 487                 /*
 488                  * Do not interfere with other mappings of this folio and
 489                  * non-LRU folio.
 490                  */
 491                 if (!folio_test_lru(folio) || folio_mapcount(folio) != 1)
 492                         continue;
 493
 494                 if (pageout_anon_only_filter && !folio_test_anon(folio))
 495                         continue;
 496
 497                 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
 498
 499                 if (pte_young(ptent)) {
 500                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 501                                                         tlb->fullmm);
 502                         ptent = pte_mkold(ptent);
 503                         set_pte_at(mm, addr, pte, ptent);
 504                         tlb_remove_tlb_entry(tlb, pte, addr);
 505                 }
 506
 507                 /*
 508                  * We are deactivating a folio for accelerating reclaiming.
 509                  * VM couldn't reclaim the folio unless we clear PG_young.
 510                  * As a side effect, it makes confuse idle-page tracking
 511                  * because they will miss recent referenced history.
 512                  */
 513                 folio_clear_referenced(folio);
 514                 folio_test_clear_young(folio);
 515                 if (folio_test_active(folio))
 516                         folio_set_workingset(folio);
 517                 if (pageout) {
 518                         if (folio_isolate_lru(folio)) {
 519                                 if (folio_test_unevictable(folio))
 520                                         folio_putback_lru(folio);
 521                                 else
 522                                         list_add(&folio->lru, &folio_list);
 523                         }
 524                 } else
 525                         folio_deactivate(folio);
 526         }
 527
 528         if (start_pte) {
 529                 arch_leave_lazy_mmu_mode();
 530                 pte_unmap_unlock(start_pte, ptl);
 531         }
 532         if (pageout)
 533                 reclaim_pages(&folio_list);
 534         cond_resched();
 535
 536         return 0;
 537 }
 538
 539 static const struct mm_walk_ops cold_walk_ops = {
 540         .pmd_entry = madvise_cold_or_pageout_pte_range,
 541 };
 542
 543 static void madvise_cold_page_range(struct mmu_gather *tlb,
 544                              struct vm_area_struct *vma,
 545                              unsigned long addr, unsigned long end)
 546 {
 547         struct madvise_walk_private walk_private = {
 548                 .pageout = false,
 549                 .tlb = tlb,
 550         };
 551
 552         tlb_start_vma(tlb, vma);
 553         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 554         tlb_end_vma(tlb, vma);
 555 }
 556
 557 static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
 558 {
 559         return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
 560 }
 561
 562 static long madvise_cold(struct vm_area_struct *vma,
 563                         struct vm_area_struct **prev,
 564                         unsigned long start_addr, unsigned long end_addr)
 565 {
 566         struct mm_struct *mm = vma->vm_mm;
 567         struct mmu_gather tlb;
 568
 569         *prev = vma;
 570         if (!can_madv_lru_vma(vma))
 571                 return -EINVAL;
 572
 573         lru_add_drain();
 574         tlb_gather_mmu(&tlb, mm);
 575         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 576         tlb_finish_mmu(&tlb);
 577
 578         return 0;
 579 }
 580
 581 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 582                              struct vm_area_struct *vma,
 583                              unsigned long addr, unsigned long end)
 584 {
 585         struct madvise_walk_private walk_private = {
 586                 .pageout = true,
 587                 .tlb = tlb,
 588         };
 589
 590         tlb_start_vma(tlb, vma);
 591         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 592         tlb_end_vma(tlb, vma);
 593 }
 594
 595 static long madvise_pageout(struct vm_area_struct *vma,
 596                         struct vm_area_struct **prev,
 597                         unsigned long start_addr, unsigned long end_addr)
 598 {
 599         struct mm_struct *mm = vma->vm_mm;
 600         struct mmu_gather tlb;
 601
 602         *prev = vma;
 603         if (!can_madv_lru_vma(vma))
 604                 return -EINVAL;
 605
 606         /*
 607          * If the VMA belongs to a private file mapping, there can be private
 608          * dirty pages which can be paged out if even this process is neither
 609          * owner nor write capable of the file. We allow private file mappings
 610          * further to pageout dirty anon pages.
 611          */
 612         if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
 613                                 (vma->vm_flags & VM_MAYSHARE)))
 614                 return 0;
 615
 616         lru_add_drain();
 617         tlb_gather_mmu(&tlb, mm);
 618         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 619         tlb_finish_mmu(&tlb);
 620
 621         return 0;
 622 }
 623
 624 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 625                                 unsigned long end, struct mm_walk *walk)
 626
 627 {
 628         struct mmu_gather *tlb = walk->private;
 629         struct mm_struct *mm = tlb->mm;
 630         struct vm_area_struct *vma = walk->vma;
 631         spinlock_t *ptl;
 632         pte_t *start_pte, *pte, ptent;
 633         struct folio *folio;
 634         int nr_swap = 0;
 635         unsigned long next;
 636
 637         next = pmd_addr_end(addr, end);
 638         if (pmd_trans_huge(*pmd))
 639                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 640                         return 0;
 641
 642         tlb_change_page_size(tlb, PAGE_SIZE);
 643         start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 644         if (!start_pte)
 645                 return 0;
 646         flush_tlb_batched_pending(mm);
 647         arch_enter_lazy_mmu_mode();
 648         for (; addr != end; pte++, addr += PAGE_SIZE) {
 649                 ptent = ptep_get(pte);
 650
 651                 if (pte_none(ptent))
 652                         continue;
 653                 /*
 654                  * If the pte has swp_entry, just clear page table to
 655                  * prevent swap-in which is more expensive rather than
 656                  * (page allocation + zeroing).
 657                  */
 658                 if (!pte_present(ptent)) {
 659                         swp_entry_t entry;
 660
 661                         entry = pte_to_swp_entry(ptent);
 662                         if (!non_swap_entry(entry)) {
 663                                 nr_swap--;
 664                                 free_swap_and_cache(entry);
 665                                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 666                         } else if (is_hwpoison_entry(entry) ||
 667                                    is_poisoned_swp_entry(entry)) {
 668                                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 669                         }
 670                         continue;
 671                 }
 672
 673                 folio = vm_normal_folio(vma, addr, ptent);
 674                 if (!folio || folio_is_zone_device(folio))
 675                         continue;
 676
 677                 /*
 678                  * If pmd isn't transhuge but the folio is large and
 679                  * is owned by only this process, split it and
 680                  * deactivate all pages.
 681                  */
 682                 if (folio_test_large(folio)) {
 683                         int err;
 684
 685                         if (folio_mapcount(folio) != 1)
 686                                 break;
 687                         if (!folio_trylock(folio))
 688                                 break;
 689                         folio_get(folio);
 690                         arch_leave_lazy_mmu_mode();
 691                         pte_unmap_unlock(start_pte, ptl);
 692                         start_pte = NULL;
 693                         err = split_folio(folio);
 694                         folio_unlock(folio);
 695                         folio_put(folio);
 696                         if (err)
 697                                 break;
 698                         start_pte = pte =
 699                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 700                         if (!start_pte)
 701                                 break;
 702                         arch_enter_lazy_mmu_mode();
 703                         pte--;
 704                         addr -= PAGE_SIZE;
 705                         continue;
 706                 }
 707
 708                 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
 709                         if (!folio_trylock(folio))
 710                                 continue;
 711                         /*
 712                          * If folio is shared with others, we mustn't clear
 713                          * the folio's dirty flag.
 714                          */
 715                         if (folio_mapcount(folio) != 1) {
 716                                 folio_unlock(folio);
 717                                 continue;
 718                         }
 719
 720                         if (folio_test_swapcache(folio) &&
 721                             !folio_free_swap(folio)) {
 722                                 folio_unlock(folio);
 723                                 continue;
 724                         }
 725
 726                         folio_clear_dirty(folio);
 727                         folio_unlock(folio);
 728                 }
 729
 730                 if (pte_young(ptent) || pte_dirty(ptent)) {
 731                         /*
 732                          * Some of architecture(ex, PPC) don't update TLB
 733                          * with set_pte_at and tlb_remove_tlb_entry so for
 734                          * the portability, remap the pte with old|clean
 735                          * after pte clearing.
 736                          */
 737                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 738                                                         tlb->fullmm);
 739
 740                         ptent = pte_mkold(ptent);
 741                         ptent = pte_mkclean(ptent);
 742                         set_pte_at(mm, addr, pte, ptent);
 743                         tlb_remove_tlb_entry(tlb, pte, addr);
 744                 }
 745                 folio_mark_lazyfree(folio);
 746         }
 747
 748         if (nr_swap) {
 749                 if (current->mm == mm)
 750                         sync_mm_rss(mm);
 751                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 752         }
 753         if (start_pte) {
 754                 arch_leave_lazy_mmu_mode();
 755                 pte_unmap_unlock(start_pte, ptl);
 756         }
 757         cond_resched();
 758
 759         return 0;
 760 }
 761
 762 static const struct mm_walk_ops madvise_free_walk_ops = {
 763         .pmd_entry              = madvise_free_pte_range,
 764 };
 765
 766 static int madvise_free_single_vma(struct vm_area_struct *vma,
 767                         unsigned long start_addr, unsigned long end_addr)
 768 {
 769         struct mm_struct *mm = vma->vm_mm;
 770         struct mmu_notifier_range range;
 771         struct mmu_gather tlb;
 772
 773         /* MADV_FREE works for only anon vma at the moment */
 774         if (!vma_is_anonymous(vma))
 775                 return -EINVAL;
 776
 777         range.start = max(vma->vm_start, start_addr);
 778         if (range.start >= vma->vm_end)
 779                 return -EINVAL;
 780         range.end = min(vma->vm_end, end_addr);
 781         if (range.end <= vma->vm_start)
 782                 return -EINVAL;
 783         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
 784                                 range.start, range.end);
 785
 786         lru_add_drain();
 787         tlb_gather_mmu(&tlb, mm);
 788         update_hiwater_rss(mm);
 789
 790         mmu_notifier_invalidate_range_start(&range);
 791         tlb_start_vma(&tlb, vma);
 792         walk_page_range(vma->vm_mm, range.start, range.end,
 793                         &madvise_free_walk_ops, &tlb);
 794         tlb_end_vma(&tlb, vma);
 795         mmu_notifier_invalidate_range_end(&range);
 796         tlb_finish_mmu(&tlb);
 797
 798         return 0;
 799 }
 800
 801 /*
 802  * Application no longer needs these pages.  If the pages are dirty,
 803  * it's OK to just throw them away.  The app will be more careful about
 804  * data it wants to keep.  Be sure to free swap resources too.  The
 805  * zap_page_range_single call sets things up for shrink_active_list to actually
 806  * free these pages later if no one else has touched them in the meantime,
 807  * although we could add these pages to a global reuse list for
 808  * shrink_active_list to pick up before reclaiming other pages.
 809  *
 810  * NB: This interface discards data rather than pushes it out to swap,
 811  * as some implementations do.  This has performance implications for
 812  * applications like large transactional databases which want to discard
 813  * pages in anonymous maps after committing to backing store the data
 814  * that was kept in them.  There is no reason to write this data out to
 815  * the swap area if the application is discarding it.
 816  *
 817  * An interface that causes the system to free clean pages and flush
 818  * dirty pages is already available as msync(MS_INVALIDATE).
 819  */
 820 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 821                                         unsigned long start, unsigned long end)
 822 {
 823         zap_page_range_single(vma, start, end - start, NULL);
 824         return 0;
 825 }
 826
 827 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
 828                                             unsigned long start,
 829                                             unsigned long *end,
 830                                             int behavior)
 831 {
 832         if (!is_vm_hugetlb_page(vma)) {
 833                 unsigned int forbidden = VM_PFNMAP;
 834
 835                 if (behavior != MADV_DONTNEED_LOCKED)
 836                         forbidden |= VM_LOCKED;
 837
 838                 return !(vma->vm_flags & forbidden);
 839         }
 840
 841         if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
 842                 return false;
 843         if (start & ~huge_page_mask(hstate_vma(vma)))
 844                 return false;
 845
 846         /*
 847          * Madvise callers expect the length to be rounded up to PAGE_SIZE
 848          * boundaries, and may be unaware that this VMA uses huge pages.
 849          * Avoid unexpected data loss by rounding down the number of
 850          * huge pages freed.
 851          */
 852         *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
 853
 854         return true;
 855 }
 856
 857 static long madvise_dontneed_free(struct vm_area_struct *vma,
 858                                   struct vm_area_struct **prev,
 859                                   unsigned long start, unsigned long end,
 860                                   int behavior)
 861 {
 862         struct mm_struct *mm = vma->vm_mm;
 863
 864         *prev = vma;
 865         if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
 866                 return -EINVAL;
 867
 868         if (start == end)
 869                 return 0;
 870
 871         if (!userfaultfd_remove(vma, start, end)) {
 872                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 873
 874                 mmap_read_lock(mm);
 875                 vma = vma_lookup(mm, start);
 876                 if (!vma)
 877                         return -ENOMEM;
 878                 /*
 879                  * Potential end adjustment for hugetlb vma is OK as
 880                  * the check below keeps end within vma.
 881                  */
 882                 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
 883                                                      behavior))
 884                         return -EINVAL;
 885                 if (end > vma->vm_end) {
 886                         /*
 887                          * Don't fail if end > vma->vm_end. If the old
 888                          * vma was split while the mmap_lock was
 889                          * released the effect of the concurrent
 890                          * operation may not cause madvise() to
 891                          * have an undefined result. There may be an
 892                          * adjacent next vma that we'll walk
 893                          * next. userfaultfd_remove() will generate an
 894                          * UFFD_EVENT_REMOVE repetition on the
 895                          * end-vma->vm_end range, but the manager can
 896                          * handle a repetition fine.
 897                          */
 898                         end = vma->vm_end;
 899                 }
 900                 VM_WARN_ON(start >= end);
 901         }
 902
 903         if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
 904                 return madvise_dontneed_single_vma(vma, start, end);
 905         else if (behavior == MADV_FREE)
 906                 return madvise_free_single_vma(vma, start, end);
 907         else
 908                 return -EINVAL;
 909 }
 910
 911 static long madvise_populate(struct vm_area_struct *vma,
 912                              struct vm_area_struct **prev,
 913                              unsigned long start, unsigned long end,
 914                              int behavior)
 915 {
 916         const bool write = behavior == MADV_POPULATE_WRITE;
 917         struct mm_struct *mm = vma->vm_mm;
 918         unsigned long tmp_end;
 919         int locked = 1;
 920         long pages;
 921
 922         *prev = vma;
 923
 924         while (start < end) {
 925                 /*
 926                  * We might have temporarily dropped the lock. For example,
 927                  * our VMA might have been split.
 928                  */
 929                 if (!vma || start >= vma->vm_end) {
 930                         vma = vma_lookup(mm, start);
 931                         if (!vma)
 932                                 return -ENOMEM;
 933                 }
 934
 935                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 936                 /* Populate (prefault) page tables readable/writable. */
 937                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 938                                                &locked);
 939                 if (!locked) {
 940                         mmap_read_lock(mm);
 941                         locked = 1;
 942                         *prev = NULL;
 943                         vma = NULL;
 944                 }
 945                 if (pages < 0) {
 946                         switch (pages) {
 947                         case -EINTR:
 948                                 return -EINTR;
 949                         case -EINVAL: /* Incompatible mappings / permissions. */
 950                                 return -EINVAL;
 951                         case -EHWPOISON:
 952                                 return -EHWPOISON;
 953                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 954                                 return -EFAULT;
 955                         default:
 956                                 pr_warn_once("%s: unhandled return value: %ld\n",
 957                                              __func__, pages);
 958                                 fallthrough;
 959                         case -ENOMEM:
 960                                 return -ENOMEM;
 961                         }
 962                 }
 963                 start += pages * PAGE_SIZE;
 964         }
 965         return 0;
 966 }
 967
 968 /*
 969  * Application wants to free up the pages and associated backing store.
 970  * This is effectively punching a hole into the middle of a file.
 971  */
 972 static long madvise_remove(struct vm_area_struct *vma,
 973                                 struct vm_area_struct **prev,
 974                                 unsigned long start, unsigned long end)
 975 {
 976         loff_t offset;
 977         int error;
 978         struct file *f;
 979         struct mm_struct *mm = vma->vm_mm;
 980
 981         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 982
 983         if (vma->vm_flags & VM_LOCKED)
 984                 return -EINVAL;
 985
 986         f = vma->vm_file;
 987
 988         if (!f || !f->f_mapping || !f->f_mapping->host) {
 989                         return -EINVAL;
 990         }
 991
 992         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 993                 return -EACCES;
 994
 995         offset = (loff_t)(start - vma->vm_start)
 996                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 997
 998         /*
 999          * Filesystem's fallocate may need to take i_rwsem.  We need to
1000          * explicitly grab a reference because the vma (and hence the
1001          * vma's reference to the file) can go away as soon as we drop
1002          * mmap_lock.
1003          */
1004         get_file(f);
1005         if (userfaultfd_remove(vma, start, end)) {
1006                 /* mmap_lock was not released by userfaultfd_remove() */
1007                 mmap_read_unlock(mm);
1008         }
1009         error = vfs_fallocate(f,
1010                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1011                                 offset, end - start);
1012         fput(f);
1013         mmap_read_lock(mm);
1014         return error;
1015 }
1016
1017 /*
1018  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
1019  * will handle splitting a vm area into separate areas, each area with its own
1020  * behavior.
1021  */
1022 static int madvise_vma_behavior(struct vm_area_struct *vma,
1023                                 struct vm_area_struct **prev,
1024                                 unsigned long start, unsigned long end,
1025                                 unsigned long behavior)
1026 {
1027         int error;
1028         struct anon_vma_name *anon_name;
1029         unsigned long new_flags = vma->vm_flags;
1030
1031         switch (behavior) {
1032         case MADV_REMOVE:
1033                 return madvise_remove(vma, prev, start, end);
1034         case MADV_WILLNEED:
1035                 return madvise_willneed(vma, prev, start, end);
1036         case MADV_COLD:
1037                 return madvise_cold(vma, prev, start, end);
1038         case MADV_PAGEOUT:
1039                 return madvise_pageout(vma, prev, start, end);
1040         case MADV_FREE:
1041         case MADV_DONTNEED:
1042         case MADV_DONTNEED_LOCKED:
1043                 return madvise_dontneed_free(vma, prev, start, end, behavior);
1044         case MADV_POPULATE_READ:
1045         case MADV_POPULATE_WRITE:
1046                 return madvise_populate(vma, prev, start, end, behavior);
1047         case MADV_NORMAL:
1048                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1049                 break;
1050         case MADV_SEQUENTIAL:
1051                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1052                 break;
1053         case MADV_RANDOM:
1054                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1055                 break;
1056         case MADV_DONTFORK:
1057                 new_flags |= VM_DONTCOPY;
1058                 break;
1059         case MADV_DOFORK:
1060                 if (vma->vm_flags & VM_IO)
1061                         return -EINVAL;
1062                 new_flags &= ~VM_DONTCOPY;
1063                 break;
1064         case MADV_WIPEONFORK:
1065                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1066                 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1067                         return -EINVAL;
1068                 new_flags |= VM_WIPEONFORK;
1069                 break;
1070         case MADV_KEEPONFORK:
1071                 new_flags &= ~VM_WIPEONFORK;
1072                 break;
1073         case MADV_DONTDUMP:
1074                 new_flags |= VM_DONTDUMP;
1075                 break;
1076         case MADV_DODUMP:
1077                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1078                         return -EINVAL;
1079                 new_flags &= ~VM_DONTDUMP;
1080                 break;
1081         case MADV_MERGEABLE:
1082         case MADV_UNMERGEABLE:
1083                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1084                 if (error)
1085                         goto out;
1086                 break;
1087         case MADV_HUGEPAGE:
1088         case MADV_NOHUGEPAGE:
1089                 error = hugepage_madvise(vma, &new_flags, behavior);
1090                 if (error)
1091                         goto out;
1092                 break;
1093         case MADV_COLLAPSE:
1094                 return madvise_collapse(vma, prev, start, end);
1095         }
1096
1097         anon_name = anon_vma_name(vma);
1098         anon_vma_name_get(anon_name);
1099         error = madvise_update_vma(vma, prev, start, end, new_flags,
1100                                    anon_name);
1101         anon_vma_name_put(anon_name);
1102
1103 out:
1104         /*
1105          * madvise() returns EAGAIN if kernel resources, such as
1106          * slab, are temporarily unavailable.
1107          */
1108         if (error == -ENOMEM)
1109                 error = -EAGAIN;
1110         return error;
1111 }
1112
1113 #ifdef CONFIG_MEMORY_FAILURE
1114 /*
1115  * Error injection support for memory error handling.
1116  */
1117 static int madvise_inject_error(int behavior,
1118                 unsigned long start, unsigned long end)
1119 {
1120         unsigned long size;
1121
1122         if (!capable(CAP_SYS_ADMIN))
1123                 return -EPERM;
1124
1125
1126         for (; start < end; start += size) {
1127                 unsigned long pfn;
1128                 struct page *page;
1129                 int ret;
1130
1131                 ret = get_user_pages_fast(start, 1, 0, &page);
1132                 if (ret != 1)
1133                         return ret;
1134                 pfn = page_to_pfn(page);
1135
1136                 /*
1137                  * When soft offlining hugepages, after migrating the page
1138                  * we dissolve it, therefore in the second loop "page" will
1139                  * no longer be a compound page.
1140                  */
1141                 size = page_size(compound_head(page));
1142
1143                 if (behavior == MADV_SOFT_OFFLINE) {
1144                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1145                                  pfn, start);
1146                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1147                 } else {
1148                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1149                                  pfn, start);
1150                         ret = memory_failure(pfn, MF_COUNT_INCREASED | MF_SW_SIMULATED);
1151                         if (ret == -EOPNOTSUPP)
1152                                 ret = 0;
1153                 }
1154
1155                 if (ret)
1156                         return ret;
1157         }
1158
1159         return 0;
1160 }
1161 #endif
1162
1163 static bool
1164 madvise_behavior_valid(int behavior)
1165 {
1166         switch (behavior) {
1167         case MADV_DOFORK:
1168         case MADV_DONTFORK:
1169         case MADV_NORMAL:
1170         case MADV_SEQUENTIAL:
1171         case MADV_RANDOM:
1172         case MADV_REMOVE:
1173         case MADV_WILLNEED:
1174         case MADV_DONTNEED:
1175         case MADV_DONTNEED_LOCKED:
1176         case MADV_FREE:
1177         case MADV_COLD:
1178         case MADV_PAGEOUT:
1179         case MADV_POPULATE_READ:
1180         case MADV_POPULATE_WRITE:
1181 #ifdef CONFIG_KSM
1182         case MADV_MERGEABLE:
1183         case MADV_UNMERGEABLE:
1184 #endif
1185 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1186         case MADV_HUGEPAGE:
1187         case MADV_NOHUGEPAGE:
1188         case MADV_COLLAPSE:
1189 #endif
1190         case MADV_DONTDUMP:
1191         case MADV_DODUMP:
1192         case MADV_WIPEONFORK:
1193         case MADV_KEEPONFORK:
1194 #ifdef CONFIG_MEMORY_FAILURE
1195         case MADV_SOFT_OFFLINE:
1196         case MADV_HWPOISON:
1197 #endif
1198                 return true;
1199
1200         default:
1201                 return false;
1202         }
1203 }
1204
1205 static bool process_madvise_behavior_valid(int behavior)
1206 {
1207         switch (behavior) {
1208         case MADV_COLD:
1209         case MADV_PAGEOUT:
1210         case MADV_WILLNEED:
1211         case MADV_COLLAPSE:
1212                 return true;
1213         default:
1214                 return false;
1215         }
1216 }
1217
1218 /*
1219  * Walk the vmas in range [start,end), and call the visit function on each one.
1220  * The visit function will get start and end parameters that cover the overlap
1221  * between the current vma and the original range.  Any unmapped regions in the
1222  * original range will result in this function returning -ENOMEM while still
1223  * calling the visit function on all of the existing vmas in the range.
1224  * Must be called with the mmap_lock held for reading or writing.
1225  */
1226 static
1227 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1228                       unsigned long end, unsigned long arg,
1229                       int (*visit)(struct vm_area_struct *vma,
1230                                    struct vm_area_struct **prev, unsigned long start,
1231                                    unsigned long end, unsigned long arg))
1232 {
1233         struct vm_area_struct *vma;
1234         struct vm_area_struct *prev;
1235         unsigned long tmp;
1236         int unmapped_error = 0;
1237
1238         /*
1239          * If the interval [start,end) covers some unmapped address
1240          * ranges, just ignore them, but return -ENOMEM at the end.
1241          * - different from the way of handling in mlock etc.
1242          */
1243         vma = find_vma_prev(mm, start, &prev);
1244         if (vma && start > vma->vm_start)
1245                 prev = vma;
1246
1247         for (;;) {
1248                 int error;
1249
1250                 /* Still start < end. */
1251                 if (!vma)
1252                         return -ENOMEM;
1253
1254                 /* Here start < (end|vma->vm_end). */
1255                 if (start < vma->vm_start) {
1256                         unmapped_error = -ENOMEM;
1257                         start = vma->vm_start;
1258                         if (start >= end)
1259                                 break;
1260                 }
1261
1262                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1263                 tmp = vma->vm_end;
1264                 if (end < tmp)
1265                         tmp = end;
1266
1267                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1268                 error = visit(vma, &prev, start, tmp, arg);
1269                 if (error)
1270                         return error;
1271                 start = tmp;
1272                 if (prev && start < prev->vm_end)
1273                         start = prev->vm_end;
1274                 if (start >= end)
1275                         break;
1276                 if (prev)
1277                         vma = find_vma(mm, prev->vm_end);
1278                 else    /* madvise_remove dropped mmap_lock */
1279                         vma = find_vma(mm, start);
1280         }
1281
1282         return unmapped_error;
1283 }
1284
1285 #ifdef CONFIG_ANON_VMA_NAME
1286 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1287                                  struct vm_area_struct **prev,
1288                                  unsigned long start, unsigned long end,
1289                                  unsigned long anon_name)
1290 {
1291         int error;
1292
1293         /* Only anonymous mappings can be named */
1294         if (vma->vm_file && !vma_is_anon_shmem(vma))
1295                 return -EBADF;
1296
1297         error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1298                                    (struct anon_vma_name *)anon_name);
1299
1300         /*
1301          * madvise() returns EAGAIN if kernel resources, such as
1302          * slab, are temporarily unavailable.
1303          */
1304         if (error == -ENOMEM)
1305                 error = -EAGAIN;
1306         return error;
1307 }
1308
1309 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1310                           unsigned long len_in, struct anon_vma_name *anon_name)
1311 {
1312         unsigned long end;
1313         unsigned long len;
1314
1315         if (start & ~PAGE_MASK)
1316                 return -EINVAL;
1317         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1318
1319         /* Check to see whether len was rounded up from small -ve to zero */
1320         if (len_in && !len)
1321                 return -EINVAL;
1322
1323         end = start + len;
1324         if (end < start)
1325                 return -EINVAL;
1326
1327         if (end == start)
1328                 return 0;
1329
1330         return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1331                                  madvise_vma_anon_name);
1332 }
1333 #endif /* CONFIG_ANON_VMA_NAME */
1334 /*
1335  * The madvise(2) system call.
1336  *
1337  * Applications can use madvise() to advise the kernel how it should
1338  * handle paging I/O in this VM area.  The idea is to help the kernel
1339  * use appropriate read-ahead and caching techniques.  The information
1340  * provided is advisory only, and can be safely disregarded by the
1341  * kernel without affecting the correct operation of the application.
1342  *
1343  * behavior values:
1344  *  MADV_NORMAL - the default behavior is to read clusters.  This
1345  *              results in some read-ahead and read-behind.
1346  *  MADV_RANDOM - the system should read the minimum amount of data
1347  *              on any access, since it is unlikely that the appli-
1348  *              cation will need more than what it asks for.
1349  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1350  *              once, so they can be aggressively read ahead, and
1351  *              can be freed soon after they are accessed.
1352  *  MADV_WILLNEED - the application is notifying the system to read
1353  *              some pages ahead.
1354  *  MADV_DONTNEED - the application is finished with the given range,
1355  *              so the kernel can free resources associated with it.
1356  *  MADV_FREE - the application marks pages in the given range as lazy free,
1357  *              where actual purges are postponed until memory pressure happens.
1358  *  MADV_REMOVE - the application wants to free up the given range of
1359  *              pages and associated backing store.
1360  *  MADV_DONTFORK - omit this area from child's address space when forking:
1361  *              typically, to avoid COWing pages pinned by get_user_pages().
1362  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1363  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1364  *              range after a fork.
1365  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1366  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1367  *              were corrupted by unrecoverable hardware memory failure.
1368  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1369  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1370  *              this area with pages of identical content from other such areas.
1371  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1372  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1373  *              huge pages in the future. Existing pages might be coalesced and
1374  *              new pages might be allocated as THP.
1375  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1376  *              transparent huge pages so the existing pages will not be
1377  *              coalesced into THP and new pages will not be allocated as THP.
1378  *  MADV_COLLAPSE - synchronously coalesce pages into new THP.
1379  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1380  *              from being included in its core dump.
1381  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1382  *  MADV_COLD - the application is not expected to use this memory soon,
1383  *              deactivate pages in this range so that they can be reclaimed
1384  *              easily if memory pressure happens.
1385  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1386  *              page out the pages in this range immediately.
1387  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1388  *              triggering read faults if required
1389  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1390  *              triggering write faults if required
1391  *
1392  * return values:
1393  *  zero    - success
1394  *  -EINVAL - start + len < 0, start is not page-aligned,
1395  *              "behavior" is not a valid value, or application
1396  *              is attempting to release locked or shared pages,
1397  *              or the specified address range includes file, Huge TLB,
1398  *              MAP_SHARED or VMPFNMAP range.
1399  *  -ENOMEM - addresses in the specified range are not currently
1400  *              mapped, or are outside the AS of the process.
1401  *  -EIO    - an I/O error occurred while paging in data.
1402  *  -EBADF  - map exists, but area maps something that isn't a file.
1403  *  -EAGAIN - a kernel resource was temporarily unavailable.
1404  */
1405 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1406 {
1407         unsigned long end;
1408         int error;
1409         int write;
1410         size_t len;
1411         struct blk_plug plug;
1412
1413         if (!madvise_behavior_valid(behavior))
1414                 return -EINVAL;
1415
1416         if (!PAGE_ALIGNED(start))
1417                 return -EINVAL;
1418         len = PAGE_ALIGN(len_in);
1419
1420         /* Check to see whether len was rounded up from small -ve to zero */
1421         if (len_in && !len)
1422                 return -EINVAL;
1423
1424         end = start + len;
1425         if (end < start)
1426                 return -EINVAL;
1427
1428         if (end == start)
1429                 return 0;
1430
1431 #ifdef CONFIG_MEMORY_FAILURE
1432         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1433                 return madvise_inject_error(behavior, start, start + len_in);
1434 #endif
1435
1436         write = madvise_need_mmap_write(behavior);
1437         if (write) {
1438                 if (mmap_write_lock_killable(mm))
1439                         return -EINTR;
1440         } else {
1441                 mmap_read_lock(mm);
1442         }
1443
1444         start = untagged_addr_remote(mm, start);
1445         end = start + len;
1446
1447         blk_start_plug(&plug);
1448         error = madvise_walk_vmas(mm, start, end, behavior,
1449                         madvise_vma_behavior);
1450         blk_finish_plug(&plug);
1451         if (write)
1452                 mmap_write_unlock(mm);
1453         else
1454                 mmap_read_unlock(mm);
1455
1456         return error;
1457 }
1458
1459 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1460 {
1461         return do_madvise(current->mm, start, len_in, behavior);
1462 }
1463
1464 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1465                 size_t, vlen, int, behavior, unsigned int, flags)
1466 {
1467         ssize_t ret;
1468         struct iovec iovstack[UIO_FASTIOV];
1469         struct iovec *iov = iovstack;
1470         struct iov_iter iter;
1471         struct task_struct *task;
1472         struct mm_struct *mm;
1473         size_t total_len;
1474         unsigned int f_flags;
1475
1476         if (flags != 0) {
1477                 ret = -EINVAL;
1478                 goto out;
1479         }
1480
1481         ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1482         if (ret < 0)
1483                 goto out;
1484
1485         task = pidfd_get_task(pidfd, &f_flags);
1486         if (IS_ERR(task)) {
1487                 ret = PTR_ERR(task);
1488                 goto free_iov;
1489         }
1490
1491         if (!process_madvise_behavior_valid(behavior)) {
1492                 ret = -EINVAL;
1493                 goto release_task;
1494         }
1495
1496         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1497         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1498         if (IS_ERR_OR_NULL(mm)) {
1499                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1500                 goto release_task;
1501         }
1502
1503         /*
1504          * Require CAP_SYS_NICE for influencing process performance. Note that
1505          * only non-destructive hints are currently supported.
1506          */
1507         if (!capable(CAP_SYS_NICE)) {
1508                 ret = -EPERM;
1509                 goto release_mm;
1510         }
1511
1512         total_len = iov_iter_count(&iter);
1513
1514         while (iov_iter_count(&iter)) {
1515                 ret = do_madvise(mm, (unsigned long)iter_iov_addr(&iter),
1516                                         iter_iov_len(&iter), behavior);
1517                 if (ret < 0)
1518                         break;
1519                 iov_iter_advance(&iter, iter_iov_len(&iter));
1520         }
1521
1522         ret = (total_len - iov_iter_count(&iter)) ? : ret;
1523
1524 release_mm:
1525         mmput(mm);
1526 release_task:
1527         put_task_struct(task);
1528 free_iov:
1529         kfree(iov);
1530 out:
1531         return ret;
1532 }