mm/madvise.c

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