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