mm/madvise.c

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