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