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