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