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         int nr_swap = 0;
 621         unsigned long next;
 622
 623         next = pmd_addr_end(addr, end);
 624         if (pmd_trans_huge(*pmd))
 625                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 626                         goto next;
 627
 628         if (pmd_trans_unstable(pmd))
 629                 return 0;
 630
 631         tlb_change_page_size(tlb, PAGE_SIZE);
 632         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 633         flush_tlb_batched_pending(mm);
 634         arch_enter_lazy_mmu_mode();
 635         for (; addr != end; pte++, addr += PAGE_SIZE) {
 636                 ptent = *pte;
 637
 638                 if (pte_none(ptent))
 639                         continue;
 640                 /*
 641                  * If the pte has swp_entry, just clear page table to
 642                  * prevent swap-in which is more expensive rather than
 643                  * (page allocation + zeroing).
 644                  */
 645                 if (!pte_present(ptent)) {
 646                         swp_entry_t entry;
 647
 648                         entry = pte_to_swp_entry(ptent);
 649                         if (!non_swap_entry(entry)) {
 650                                 nr_swap--;
 651                                 free_swap_and_cache(entry);
 652                                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 653                         } else if (is_hwpoison_entry(entry) ||
 654                                    is_swapin_error_entry(entry)) {
 655                                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 656                         }
 657                         continue;
 658                 }
 659
 660                 folio = vm_normal_folio(vma, addr, ptent);
 661                 if (!folio || folio_is_zone_device(folio))
 662                         continue;
 663
 664                 /*
 665                  * If pmd isn't transhuge but the folio is large and
 666                  * is owned by only this process, split it and
 667                  * deactivate all pages.
 668                  */
 669                 if (folio_test_large(folio)) {
 670                         if (folio_mapcount(folio) != 1)
 671                                 goto out;
 672                         folio_get(folio);
 673                         if (!folio_trylock(folio)) {
 674                                 folio_put(folio);
 675                                 goto out;
 676                         }
 677                         pte_unmap_unlock(orig_pte, ptl);
 678                         if (split_folio(folio)) {
 679                                 folio_unlock(folio);
 680                                 folio_put(folio);
 681                                 orig_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 682                                 goto out;
 683                         }
 684                         folio_unlock(folio);
 685                         folio_put(folio);
 686                         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 687                         pte--;
 688                         addr -= PAGE_SIZE;
 689                         continue;
 690                 }
 691
 692                 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
 693                         if (!folio_trylock(folio))
 694                                 continue;
 695                         /*
 696                          * If folio is shared with others, we mustn't clear
 697                          * the folio's dirty flag.
 698                          */
 699                         if (folio_mapcount(folio) != 1) {
 700                                 folio_unlock(folio);
 701                                 continue;
 702                         }
 703
 704                         if (folio_test_swapcache(folio) &&
 705                             !folio_free_swap(folio)) {
 706                                 folio_unlock(folio);
 707                                 continue;
 708                         }
 709
 710                         folio_clear_dirty(folio);
 711                         folio_unlock(folio);
 712                 }
 713
 714                 if (pte_young(ptent) || pte_dirty(ptent)) {
 715                         /*
 716                          * Some of architecture(ex, PPC) don't update TLB
 717                          * with set_pte_at and tlb_remove_tlb_entry so for
 718                          * the portability, remap the pte with old|clean
 719                          * after pte clearing.
 720                          */
 721                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 722                                                         tlb->fullmm);
 723
 724                         ptent = pte_mkold(ptent);
 725                         ptent = pte_mkclean(ptent);
 726                         set_pte_at(mm, addr, pte, ptent);
 727                         tlb_remove_tlb_entry(tlb, pte, addr);
 728                 }
 729                 folio_mark_lazyfree(folio);
 730         }
 731 out:
 732         if (nr_swap) {
 733                 if (current->mm == mm)
 734                         sync_mm_rss(mm);
 735
 736                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 737         }
 738         arch_leave_lazy_mmu_mode();
 739         pte_unmap_unlock(orig_pte, ptl);
 740         cond_resched();
 741 next:
 742         return 0;
 743 }
 744
 745 static const struct mm_walk_ops madvise_free_walk_ops = {
 746         .pmd_entry              = madvise_free_pte_range,
 747 };
 748
 749 static int madvise_free_single_vma(struct vm_area_struct *vma,
 750                         unsigned long start_addr, unsigned long end_addr)
 751 {
 752         struct mm_struct *mm = vma->vm_mm;
 753         struct mmu_notifier_range range;
 754         struct mmu_gather tlb;
 755
 756         /* MADV_FREE works for only anon vma at the moment */
 757         if (!vma_is_anonymous(vma))
 758                 return -EINVAL;
 759
 760         range.start = max(vma->vm_start, start_addr);
 761         if (range.start >= vma->vm_end)
 762                 return -EINVAL;
 763         range.end = min(vma->vm_end, end_addr);
 764         if (range.end <= vma->vm_start)
 765                 return -EINVAL;
 766         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
 767                                 range.start, range.end);
 768
 769         lru_add_drain();
 770         tlb_gather_mmu(&tlb, mm);
 771         update_hiwater_rss(mm);
 772
 773         mmu_notifier_invalidate_range_start(&range);
 774         tlb_start_vma(&tlb, vma);
 775         walk_page_range(vma->vm_mm, range.start, range.end,
 776                         &madvise_free_walk_ops, &tlb);
 777         tlb_end_vma(&tlb, vma);
 778         mmu_notifier_invalidate_range_end(&range);
 779         tlb_finish_mmu(&tlb);
 780
 781         return 0;
 782 }
 783
 784 /*
 785  * Application no longer needs these pages.  If the pages are dirty,
 786  * it's OK to just throw them away.  The app will be more careful about
 787  * data it wants to keep.  Be sure to free swap resources too.  The
 788  * zap_page_range_single call sets things up for shrink_active_list to actually
 789  * free these pages later if no one else has touched them in the meantime,
 790  * although we could add these pages to a global reuse list for
 791  * shrink_active_list to pick up before reclaiming other pages.
 792  *
 793  * NB: This interface discards data rather than pushes it out to swap,
 794  * as some implementations do.  This has performance implications for
 795  * applications like large transactional databases which want to discard
 796  * pages in anonymous maps after committing to backing store the data
 797  * that was kept in them.  There is no reason to write this data out to
 798  * the swap area if the application is discarding it.
 799  *
 800  * An interface that causes the system to free clean pages and flush
 801  * dirty pages is already available as msync(MS_INVALIDATE).
 802  */
 803 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 804                                         unsigned long start, unsigned long end)
 805 {
 806         zap_page_range_single(vma, start, end - start, NULL);
 807         return 0;
 808 }
 809
 810 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
 811                                             unsigned long start,
 812                                             unsigned long *end,
 813                                             int behavior)
 814 {
 815         if (!is_vm_hugetlb_page(vma)) {
 816                 unsigned int forbidden = VM_PFNMAP;
 817
 818                 if (behavior != MADV_DONTNEED_LOCKED)
 819                         forbidden |= VM_LOCKED;
 820
 821                 return !(vma->vm_flags & forbidden);
 822         }
 823
 824         if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
 825                 return false;
 826         if (start & ~huge_page_mask(hstate_vma(vma)))
 827                 return false;
 828
 829         /*
 830          * Madvise callers expect the length to be rounded up to PAGE_SIZE
 831          * boundaries, and may be unaware that this VMA uses huge pages.
 832          * Avoid unexpected data loss by rounding down the number of
 833          * huge pages freed.
 834          */
 835         *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
 836
 837         return true;
 838 }
 839
 840 static long madvise_dontneed_free(struct vm_area_struct *vma,
 841                                   struct vm_area_struct **prev,
 842                                   unsigned long start, unsigned long end,
 843                                   int behavior)
 844 {
 845         struct mm_struct *mm = vma->vm_mm;
 846
 847         *prev = vma;
 848         if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
 849                 return -EINVAL;
 850
 851         if (start == end)
 852                 return 0;
 853
 854         if (!userfaultfd_remove(vma, start, end)) {
 855                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 856
 857                 mmap_read_lock(mm);
 858                 vma = find_vma(mm, start);
 859                 if (!vma)
 860                         return -ENOMEM;
 861                 if (start < vma->vm_start) {
 862                         /*
 863                          * This "vma" under revalidation is the one
 864                          * with the lowest vma->vm_start where start
 865                          * is also < vma->vm_end. If start <
 866                          * vma->vm_start it means an hole materialized
 867                          * in the user address space within the
 868                          * virtual range passed to MADV_DONTNEED
 869                          * or MADV_FREE.
 870                          */
 871                         return -ENOMEM;
 872                 }
 873                 /*
 874                  * Potential end adjustment for hugetlb vma is OK as
 875                  * the check below keeps end within vma.
 876                  */
 877                 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
 878                                                      behavior))
 879                         return -EINVAL;
 880                 if (end > vma->vm_end) {
 881                         /*
 882                          * Don't fail if end > vma->vm_end. If the old
 883                          * vma was split while the mmap_lock was
 884                          * released the effect of the concurrent
 885                          * operation may not cause madvise() to
 886                          * have an undefined result. There may be an
 887                          * adjacent next vma that we'll walk
 888                          * next. userfaultfd_remove() will generate an
 889                          * UFFD_EVENT_REMOVE repetition on the
 890                          * end-vma->vm_end range, but the manager can
 891                          * handle a repetition fine.
 892                          */
 893                         end = vma->vm_end;
 894                 }
 895                 VM_WARN_ON(start >= end);
 896         }
 897
 898         if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
 899                 return madvise_dontneed_single_vma(vma, start, end);
 900         else if (behavior == MADV_FREE)
 901                 return madvise_free_single_vma(vma, start, end);
 902         else
 903                 return -EINVAL;
 904 }
 905
 906 static long madvise_populate(struct vm_area_struct *vma,
 907                              struct vm_area_struct **prev,
 908                              unsigned long start, unsigned long end,
 909                              int behavior)
 910 {
 911         const bool write = behavior == MADV_POPULATE_WRITE;
 912         struct mm_struct *mm = vma->vm_mm;
 913         unsigned long tmp_end;
 914         int locked = 1;
 915         long pages;
 916
 917         *prev = vma;
 918
 919         while (start < end) {
 920                 /*
 921                  * We might have temporarily dropped the lock. For example,
 922                  * our VMA might have been split.
 923                  */
 924                 if (!vma || start >= vma->vm_end) {
 925                         vma = vma_lookup(mm, start);
 926                         if (!vma)
 927                                 return -ENOMEM;
 928                 }
 929
 930                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 931                 /* Populate (prefault) page tables readable/writable. */
 932                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 933                                                &locked);
 934                 if (!locked) {
 935                         mmap_read_lock(mm);
 936                         locked = 1;
 937                         *prev = NULL;
 938                         vma = NULL;
 939                 }
 940                 if (pages < 0) {
 941                         switch (pages) {
 942                         case -EINTR:
 943                                 return -EINTR;
 944                         case -EINVAL: /* Incompatible mappings / permissions. */
 945                                 return -EINVAL;
 946                         case -EHWPOISON:
 947                                 return -EHWPOISON;
 948                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 949                                 return -EFAULT;
 950                         default:
 951                                 pr_warn_once("%s: unhandled return value: %ld\n",
 952                                              __func__, pages);
 953                                 fallthrough;
 954                         case -ENOMEM:
 955                                 return -ENOMEM;
 956                         }
 957                 }
 958                 start += pages * PAGE_SIZE;
 959         }
 960         return 0;
 961 }
 962
 963 /*
 964  * Application wants to free up the pages and associated backing store.
 965  * This is effectively punching a hole into the middle of a file.
 966  */
 967 static long madvise_remove(struct vm_area_struct *vma,
 968                                 struct vm_area_struct **prev,
 969                                 unsigned long start, unsigned long end)
 970 {
 971         loff_t offset;
 972         int error;
 973         struct file *f;
 974         struct mm_struct *mm = vma->vm_mm;
 975
 976         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 977
 978         if (vma->vm_flags & VM_LOCKED)
 979                 return -EINVAL;
 980
 981         f = vma->vm_file;
 982
 983         if (!f || !f->f_mapping || !f->f_mapping->host) {
 984                         return -EINVAL;
 985         }
 986
 987         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 988                 return -EACCES;
 989
 990         offset = (loff_t)(start - vma->vm_start)
 991                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 992
 993         /*
 994          * Filesystem's fallocate may need to take i_rwsem.  We need to
 995          * explicitly grab a reference because the vma (and hence the
 996          * vma's reference to the file) can go away as soon as we drop
 997          * mmap_lock.
 998          */
 999         get_file(f);
1000         if (userfaultfd_remove(vma, start, end)) {
1001                 /* mmap_lock was not released by userfaultfd_remove() */
1002                 mmap_read_unlock(mm);
1003         }
1004         error = vfs_fallocate(f,
1005                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1006                                 offset, end - start);
1007         fput(f);
1008         mmap_read_lock(mm);
1009         return error;
1010 }
1011
1012 /*
1013  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
1014  * will handle splitting a vm area into separate areas, each area with its own
1015  * behavior.
1016  */
1017 static int madvise_vma_behavior(struct vm_area_struct *vma,
1018                                 struct vm_area_struct **prev,
1019                                 unsigned long start, unsigned long end,
1020                                 unsigned long behavior)
1021 {
1022         int error;
1023         struct anon_vma_name *anon_name;
1024         unsigned long new_flags = vma->vm_flags;
1025
1026         switch (behavior) {
1027         case MADV_REMOVE:
1028                 return madvise_remove(vma, prev, start, end);
1029         case MADV_WILLNEED:
1030                 return madvise_willneed(vma, prev, start, end);
1031         case MADV_COLD:
1032                 return madvise_cold(vma, prev, start, end);
1033         case MADV_PAGEOUT:
1034                 return madvise_pageout(vma, prev, start, end);
1035         case MADV_FREE:
1036         case MADV_DONTNEED:
1037         case MADV_DONTNEED_LOCKED:
1038                 return madvise_dontneed_free(vma, prev, start, end, behavior);
1039         case MADV_POPULATE_READ:
1040         case MADV_POPULATE_WRITE:
1041                 return madvise_populate(vma, prev, start, end, behavior);
1042         case MADV_NORMAL:
1043                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1044                 break;
1045         case MADV_SEQUENTIAL:
1046                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1047                 break;
1048         case MADV_RANDOM:
1049                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1050                 break;
1051         case MADV_DONTFORK:
1052                 new_flags |= VM_DONTCOPY;
1053                 break;
1054         case MADV_DOFORK:
1055                 if (vma->vm_flags & VM_IO)
1056                         return -EINVAL;
1057                 new_flags &= ~VM_DONTCOPY;
1058                 break;
1059         case MADV_WIPEONFORK:
1060                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1061                 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1062                         return -EINVAL;
1063                 new_flags |= VM_WIPEONFORK;
1064                 break;
1065         case MADV_KEEPONFORK:
1066                 new_flags &= ~VM_WIPEONFORK;
1067                 break;
1068         case MADV_DONTDUMP:
1069                 new_flags |= VM_DONTDUMP;
1070                 break;
1071         case MADV_DODUMP:
1072                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1073                         return -EINVAL;
1074                 new_flags &= ~VM_DONTDUMP;
1075                 break;
1076         case MADV_MERGEABLE:
1077         case MADV_UNMERGEABLE:
1078                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1079                 if (error)
1080                         goto out;
1081                 break;
1082         case MADV_HUGEPAGE:
1083         case MADV_NOHUGEPAGE:
1084                 error = hugepage_madvise(vma, &new_flags, behavior);
1085                 if (error)
1086                         goto out;
1087                 break;
1088         case MADV_COLLAPSE:
1089                 return madvise_collapse(vma, prev, start, end);
1090         }
1091
1092         anon_name = anon_vma_name(vma);
1093         anon_vma_name_get(anon_name);
1094         error = madvise_update_vma(vma, prev, start, end, new_flags,
1095                                    anon_name);
1096         anon_vma_name_put(anon_name);
1097
1098 out:
1099         /*
1100          * madvise() returns EAGAIN if kernel resources, such as
1101          * slab, are temporarily unavailable.
1102          */
1103         if (error == -ENOMEM)
1104                 error = -EAGAIN;
1105         return error;
1106 }
1107
1108 #ifdef CONFIG_MEMORY_FAILURE
1109 /*
1110  * Error injection support for memory error handling.
1111  */
1112 static int madvise_inject_error(int behavior,
1113                 unsigned long start, unsigned long end)
1114 {
1115         unsigned long size;
1116
1117         if (!capable(CAP_SYS_ADMIN))
1118                 return -EPERM;
1119
1120
1121         for (; start < end; start += size) {
1122                 unsigned long pfn;
1123                 struct page *page;
1124                 int ret;
1125
1126                 ret = get_user_pages_fast(start, 1, 0, &page);
1127                 if (ret != 1)
1128                         return ret;
1129                 pfn = page_to_pfn(page);
1130
1131                 /*
1132                  * When soft offlining hugepages, after migrating the page
1133                  * we dissolve it, therefore in the second loop "page" will
1134                  * no longer be a compound page.
1135                  */
1136                 size = page_size(compound_head(page));
1137
1138                 if (behavior == MADV_SOFT_OFFLINE) {
1139                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1140                                  pfn, start);
1141                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1142                 } else {
1143                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1144                                  pfn, start);
1145                         ret = memory_failure(pfn, MF_COUNT_INCREASED | MF_SW_SIMULATED);
1146                         if (ret == -EOPNOTSUPP)
1147                                 ret = 0;
1148                 }
1149
1150                 if (ret)
1151                         return ret;
1152         }
1153
1154         return 0;
1155 }
1156 #endif
1157
1158 static bool
1159 madvise_behavior_valid(int behavior)
1160 {
1161         switch (behavior) {
1162         case MADV_DOFORK:
1163         case MADV_DONTFORK:
1164         case MADV_NORMAL:
1165         case MADV_SEQUENTIAL:
1166         case MADV_RANDOM:
1167         case MADV_REMOVE:
1168         case MADV_WILLNEED:
1169         case MADV_DONTNEED:
1170         case MADV_DONTNEED_LOCKED:
1171         case MADV_FREE:
1172         case MADV_COLD:
1173         case MADV_PAGEOUT:
1174         case MADV_POPULATE_READ:
1175         case MADV_POPULATE_WRITE:
1176 #ifdef CONFIG_KSM
1177         case MADV_MERGEABLE:
1178         case MADV_UNMERGEABLE:
1179 #endif
1180 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1181         case MADV_HUGEPAGE:
1182         case MADV_NOHUGEPAGE:
1183         case MADV_COLLAPSE:
1184 #endif
1185         case MADV_DONTDUMP:
1186         case MADV_DODUMP:
1187         case MADV_WIPEONFORK:
1188         case MADV_KEEPONFORK:
1189 #ifdef CONFIG_MEMORY_FAILURE
1190         case MADV_SOFT_OFFLINE:
1191         case MADV_HWPOISON:
1192 #endif
1193                 return true;
1194
1195         default:
1196                 return false;
1197         }
1198 }
1199
1200 static bool process_madvise_behavior_valid(int behavior)
1201 {
1202         switch (behavior) {
1203         case MADV_COLD:
1204         case MADV_PAGEOUT:
1205         case MADV_WILLNEED:
1206         case MADV_COLLAPSE:
1207                 return true;
1208         default:
1209                 return false;
1210         }
1211 }
1212
1213 /*
1214  * Walk the vmas in range [start,end), and call the visit function on each one.
1215  * The visit function will get start and end parameters that cover the overlap
1216  * between the current vma and the original range.  Any unmapped regions in the
1217  * original range will result in this function returning -ENOMEM while still
1218  * calling the visit function on all of the existing vmas in the range.
1219  * Must be called with the mmap_lock held for reading or writing.
1220  */
1221 static
1222 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1223                       unsigned long end, unsigned long arg,
1224                       int (*visit)(struct vm_area_struct *vma,
1225                                    struct vm_area_struct **prev, unsigned long start,
1226                                    unsigned long end, unsigned long arg))
1227 {
1228         struct vm_area_struct *vma;
1229         struct vm_area_struct *prev;
1230         unsigned long tmp;
1231         int unmapped_error = 0;
1232
1233         /*
1234          * If the interval [start,end) covers some unmapped address
1235          * ranges, just ignore them, but return -ENOMEM at the end.
1236          * - different from the way of handling in mlock etc.
1237          */
1238         vma = find_vma_prev(mm, start, &prev);
1239         if (vma && start > vma->vm_start)
1240                 prev = vma;
1241
1242         for (;;) {
1243                 int error;
1244
1245                 /* Still start < end. */
1246                 if (!vma)
1247                         return -ENOMEM;
1248
1249                 /* Here start < (end|vma->vm_end). */
1250                 if (start < vma->vm_start) {
1251                         unmapped_error = -ENOMEM;
1252                         start = vma->vm_start;
1253                         if (start >= end)
1254                                 break;
1255                 }
1256
1257                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1258                 tmp = vma->vm_end;
1259                 if (end < tmp)
1260                         tmp = end;
1261
1262                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1263                 error = visit(vma, &prev, start, tmp, arg);
1264                 if (error)
1265                         return error;
1266                 start = tmp;
1267                 if (prev && start < prev->vm_end)
1268                         start = prev->vm_end;
1269                 if (start >= end)
1270                         break;
1271                 if (prev)
1272                         vma = find_vma(mm, prev->vm_end);
1273                 else    /* madvise_remove dropped mmap_lock */
1274                         vma = find_vma(mm, start);
1275         }
1276
1277         return unmapped_error;
1278 }
1279
1280 #ifdef CONFIG_ANON_VMA_NAME
1281 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1282                                  struct vm_area_struct **prev,
1283                                  unsigned long start, unsigned long end,
1284                                  unsigned long anon_name)
1285 {
1286         int error;
1287
1288         /* Only anonymous mappings can be named */
1289         if (vma->vm_file && !vma_is_anon_shmem(vma))
1290                 return -EBADF;
1291
1292         error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1293                                    (struct anon_vma_name *)anon_name);
1294
1295         /*
1296          * madvise() returns EAGAIN if kernel resources, such as
1297          * slab, are temporarily unavailable.
1298          */
1299         if (error == -ENOMEM)
1300                 error = -EAGAIN;
1301         return error;
1302 }
1303
1304 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1305                           unsigned long len_in, struct anon_vma_name *anon_name)
1306 {
1307         unsigned long end;
1308         unsigned long len;
1309
1310         if (start & ~PAGE_MASK)
1311                 return -EINVAL;
1312         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1313
1314         /* Check to see whether len was rounded up from small -ve to zero */
1315         if (len_in && !len)
1316                 return -EINVAL;
1317
1318         end = start + len;
1319         if (end < start)
1320                 return -EINVAL;
1321
1322         if (end == start)
1323                 return 0;
1324
1325         return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1326                                  madvise_vma_anon_name);
1327 }
1328 #endif /* CONFIG_ANON_VMA_NAME */
1329 /*
1330  * The madvise(2) system call.
1331  *
1332  * Applications can use madvise() to advise the kernel how it should
1333  * handle paging I/O in this VM area.  The idea is to help the kernel
1334  * use appropriate read-ahead and caching techniques.  The information
1335  * provided is advisory only, and can be safely disregarded by the
1336  * kernel without affecting the correct operation of the application.
1337  *
1338  * behavior values:
1339  *  MADV_NORMAL - the default behavior is to read clusters.  This
1340  *              results in some read-ahead and read-behind.
1341  *  MADV_RANDOM - the system should read the minimum amount of data
1342  *              on any access, since it is unlikely that the appli-
1343  *              cation will need more than what it asks for.
1344  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1345  *              once, so they can be aggressively read ahead, and
1346  *              can be freed soon after they are accessed.
1347  *  MADV_WILLNEED - the application is notifying the system to read
1348  *              some pages ahead.
1349  *  MADV_DONTNEED - the application is finished with the given range,
1350  *              so the kernel can free resources associated with it.
1351  *  MADV_FREE - the application marks pages in the given range as lazy free,
1352  *              where actual purges are postponed until memory pressure happens.
1353  *  MADV_REMOVE - the application wants to free up the given range of
1354  *              pages and associated backing store.
1355  *  MADV_DONTFORK - omit this area from child's address space when forking:
1356  *              typically, to avoid COWing pages pinned by get_user_pages().
1357  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1358  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1359  *              range after a fork.
1360  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1361  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1362  *              were corrupted by unrecoverable hardware memory failure.
1363  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1364  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1365  *              this area with pages of identical content from other such areas.
1366  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1367  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1368  *              huge pages in the future. Existing pages might be coalesced and
1369  *              new pages might be allocated as THP.
1370  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1371  *              transparent huge pages so the existing pages will not be
1372  *              coalesced into THP and new pages will not be allocated as THP.
1373  *  MADV_COLLAPSE - synchronously coalesce pages into new THP.
1374  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1375  *              from being included in its core dump.
1376  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1377  *  MADV_COLD - the application is not expected to use this memory soon,
1378  *              deactivate pages in this range so that they can be reclaimed
1379  *              easily if memory pressure happens.
1380  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1381  *              page out the pages in this range immediately.
1382  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1383  *              triggering read faults if required
1384  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1385  *              triggering write faults if required
1386  *
1387  * return values:
1388  *  zero    - success
1389  *  -EINVAL - start + len < 0, start is not page-aligned,
1390  *              "behavior" is not a valid value, or application
1391  *              is attempting to release locked or shared pages,
1392  *              or the specified address range includes file, Huge TLB,
1393  *              MAP_SHARED or VMPFNMAP range.
1394  *  -ENOMEM - addresses in the specified range are not currently
1395  *              mapped, or are outside the AS of the process.
1396  *  -EIO    - an I/O error occurred while paging in data.
1397  *  -EBADF  - map exists, but area maps something that isn't a file.
1398  *  -EAGAIN - a kernel resource was temporarily unavailable.
1399  */
1400 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1401 {
1402         unsigned long end;
1403         int error;
1404         int write;
1405         size_t len;
1406         struct blk_plug plug;
1407
1408         start = untagged_addr(start);
1409
1410         if (!madvise_behavior_valid(behavior))
1411                 return -EINVAL;
1412
1413         if (!PAGE_ALIGNED(start))
1414                 return -EINVAL;
1415         len = PAGE_ALIGN(len_in);
1416
1417         /* Check to see whether len was rounded up from small -ve to zero */
1418         if (len_in && !len)
1419                 return -EINVAL;
1420
1421         end = start + len;
1422         if (end < start)
1423                 return -EINVAL;
1424
1425         if (end == start)
1426                 return 0;
1427
1428 #ifdef CONFIG_MEMORY_FAILURE
1429         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1430                 return madvise_inject_error(behavior, start, start + len_in);
1431 #endif
1432
1433         write = madvise_need_mmap_write(behavior);
1434         if (write) {
1435                 if (mmap_write_lock_killable(mm))
1436                         return -EINTR;
1437         } else {
1438                 mmap_read_lock(mm);
1439         }
1440
1441         blk_start_plug(&plug);
1442         error = madvise_walk_vmas(mm, start, end, behavior,
1443                         madvise_vma_behavior);
1444         blk_finish_plug(&plug);
1445         if (write)
1446                 mmap_write_unlock(mm);
1447         else
1448                 mmap_read_unlock(mm);
1449
1450         return error;
1451 }
1452
1453 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1454 {
1455         return do_madvise(current->mm, start, len_in, behavior);
1456 }
1457
1458 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1459                 size_t, vlen, int, behavior, unsigned int, flags)
1460 {
1461         ssize_t ret;
1462         struct iovec iovstack[UIO_FASTIOV], iovec;
1463         struct iovec *iov = iovstack;
1464         struct iov_iter iter;
1465         struct task_struct *task;
1466         struct mm_struct *mm;
1467         size_t total_len;
1468         unsigned int f_flags;
1469
1470         if (flags != 0) {
1471                 ret = -EINVAL;
1472                 goto out;
1473         }
1474
1475         ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1476         if (ret < 0)
1477                 goto out;
1478
1479         task = pidfd_get_task(pidfd, &f_flags);
1480         if (IS_ERR(task)) {
1481                 ret = PTR_ERR(task);
1482                 goto free_iov;
1483         }
1484
1485         if (!process_madvise_behavior_valid(behavior)) {
1486                 ret = -EINVAL;
1487                 goto release_task;
1488         }
1489
1490         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1491         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1492         if (IS_ERR_OR_NULL(mm)) {
1493                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1494                 goto release_task;
1495         }
1496
1497         /*
1498          * Require CAP_SYS_NICE for influencing process performance. Note that
1499          * only non-destructive hints are currently supported.
1500          */
1501         if (!capable(CAP_SYS_NICE)) {
1502                 ret = -EPERM;
1503                 goto release_mm;
1504         }
1505
1506         total_len = iov_iter_count(&iter);
1507
1508         while (iov_iter_count(&iter)) {
1509                 iovec = iov_iter_iovec(&iter);
1510                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1511                                         iovec.iov_len, behavior);
1512                 if (ret < 0)
1513                         break;
1514                 iov_iter_advance(&iter, iovec.iov_len);
1515         }
1516
1517         ret = (total_len - iov_iter_count(&iter)) ? : ret;
1518
1519 release_mm:
1520         mmput(mm);
1521 release_task:
1522         put_task_struct(task);
1523 free_iov:
1524         kfree(iov);
1525 out:
1526         return ret;
1527 }