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                                 arch_leave_lazy_mmu_mode();
 433                                 pte_unmap_unlock(start_pte, ptl);
 434                                 cond_resched();
 435                                 goto restart;
 436                         }
 437                 }
 438
 439                 if (pte_none(ptent))
 440                         continue;
 441
 442                 if (!pte_present(ptent))
 443                         continue;
 444
 445                 folio = vm_normal_folio(vma, addr, ptent);
 446                 if (!folio || folio_is_zone_device(folio))
 447                         continue;
 448
 449                 /*
 450                  * Creating a THP page is expensive so split it only if we
 451                  * are sure it's worth. Split it if we are only owner.
 452                  */
 453                 if (folio_test_large(folio)) {
 454                         int err;
 455
 456                         if (folio_estimated_sharers(folio) != 1)
 457                                 break;
 458                         if (pageout_anon_only_filter && !folio_test_anon(folio))
 459                                 break;
 460                         if (!folio_trylock(folio))
 461                                 break;
 462                         folio_get(folio);
 463                         arch_leave_lazy_mmu_mode();
 464                         pte_unmap_unlock(start_pte, ptl);
 465                         start_pte = NULL;
 466                         err = split_folio(folio);
 467                         folio_unlock(folio);
 468                         folio_put(folio);
 469                         if (err)
 470                                 break;
 471                         start_pte = pte =
 472                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 473                         if (!start_pte)
 474                                 break;
 475                         arch_enter_lazy_mmu_mode();
 476                         pte--;
 477                         addr -= PAGE_SIZE;
 478                         continue;
 479                 }
 480
 481                 /*
 482                  * Do not interfere with other mappings of this folio and
 483                  * non-LRU folio.
 484                  */
 485                 if (!folio_test_lru(folio) || folio_mapcount(folio) != 1)
 486                         continue;
 487
 488                 if (pageout_anon_only_filter && !folio_test_anon(folio))
 489                         continue;
 490
 491                 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
 492
 493                 if (pte_young(ptent)) {
 494                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 495                                                         tlb->fullmm);
 496                         ptent = pte_mkold(ptent);
 497                         set_pte_at(mm, addr, pte, ptent);
 498                         tlb_remove_tlb_entry(tlb, pte, addr);
 499                 }
 500
 501                 /*
 502                  * We are deactivating a folio for accelerating reclaiming.
 503                  * VM couldn't reclaim the folio unless we clear PG_young.
 504                  * As a side effect, it makes confuse idle-page tracking
 505                  * because they will miss recent referenced history.
 506                  */
 507                 folio_clear_referenced(folio);
 508                 folio_test_clear_young(folio);
 509                 if (folio_test_active(folio))
 510                         folio_set_workingset(folio);
 511                 if (pageout) {
 512                         if (folio_isolate_lru(folio)) {
 513                                 if (folio_test_unevictable(folio))
 514                                         folio_putback_lru(folio);
 515                                 else
 516                                         list_add(&folio->lru, &folio_list);
 517                         }
 518                 } else
 519                         folio_deactivate(folio);
 520         }
 521
 522         if (start_pte) {
 523                 arch_leave_lazy_mmu_mode();
 524                 pte_unmap_unlock(start_pte, ptl);
 525         }
 526         if (pageout)
 527                 reclaim_pages(&folio_list);
 528         cond_resched();
 529
 530         return 0;
 531 }
 532
 533 static const struct mm_walk_ops cold_walk_ops = {
 534         .pmd_entry = madvise_cold_or_pageout_pte_range,
 535         .walk_lock = PGWALK_RDLOCK,
 536 };
 537
 538 static void madvise_cold_page_range(struct mmu_gather *tlb,
 539                              struct vm_area_struct *vma,
 540                              unsigned long addr, unsigned long end)
 541 {
 542         struct madvise_walk_private walk_private = {
 543                 .pageout = false,
 544                 .tlb = tlb,
 545         };
 546
 547         tlb_start_vma(tlb, vma);
 548         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 549         tlb_end_vma(tlb, vma);
 550 }
 551
 552 static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
 553 {
 554         return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
 555 }
 556
 557 static long madvise_cold(struct vm_area_struct *vma,
 558                         struct vm_area_struct **prev,
 559                         unsigned long start_addr, unsigned long end_addr)
 560 {
 561         struct mm_struct *mm = vma->vm_mm;
 562         struct mmu_gather tlb;
 563
 564         *prev = vma;
 565         if (!can_madv_lru_vma(vma))
 566                 return -EINVAL;
 567
 568         lru_add_drain();
 569         tlb_gather_mmu(&tlb, mm);
 570         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 571         tlb_finish_mmu(&tlb);
 572
 573         return 0;
 574 }
 575
 576 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 577                              struct vm_area_struct *vma,
 578                              unsigned long addr, unsigned long end)
 579 {
 580         struct madvise_walk_private walk_private = {
 581                 .pageout = true,
 582                 .tlb = tlb,
 583         };
 584
 585         tlb_start_vma(tlb, vma);
 586         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 587         tlb_end_vma(tlb, vma);
 588 }
 589
 590 static long madvise_pageout(struct vm_area_struct *vma,
 591                         struct vm_area_struct **prev,
 592                         unsigned long start_addr, unsigned long end_addr)
 593 {
 594         struct mm_struct *mm = vma->vm_mm;
 595         struct mmu_gather tlb;
 596
 597         *prev = vma;
 598         if (!can_madv_lru_vma(vma))
 599                 return -EINVAL;
 600
 601         /*
 602          * If the VMA belongs to a private file mapping, there can be private
 603          * dirty pages which can be paged out if even this process is neither
 604          * owner nor write capable of the file. We allow private file mappings
 605          * further to pageout dirty anon pages.
 606          */
 607         if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
 608                                 (vma->vm_flags & VM_MAYSHARE)))
 609                 return 0;
 610
 611         lru_add_drain();
 612         tlb_gather_mmu(&tlb, mm);
 613         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 614         tlb_finish_mmu(&tlb);
 615
 616         return 0;
 617 }
 618
 619 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 620                                 unsigned long end, struct mm_walk *walk)
 621
 622 {
 623         struct mmu_gather *tlb = walk->private;
 624         struct mm_struct *mm = tlb->mm;
 625         struct vm_area_struct *vma = walk->vma;
 626         spinlock_t *ptl;
 627         pte_t *start_pte, *pte, ptent;
 628         struct folio *folio;
 629         int nr_swap = 0;
 630         unsigned long next;
 631
 632         next = pmd_addr_end(addr, end);
 633         if (pmd_trans_huge(*pmd))
 634                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 635                         return 0;
 636
 637         tlb_change_page_size(tlb, PAGE_SIZE);
 638         start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 639         if (!start_pte)
 640                 return 0;
 641         flush_tlb_batched_pending(mm);
 642         arch_enter_lazy_mmu_mode();
 643         for (; addr != end; pte++, addr += PAGE_SIZE) {
 644                 ptent = ptep_get(pte);
 645
 646                 if (pte_none(ptent))
 647                         continue;
 648                 /*
 649                  * If the pte has swp_entry, just clear page table to
 650                  * prevent swap-in which is more expensive rather than
 651                  * (page allocation + zeroing).
 652                  */
 653                 if (!pte_present(ptent)) {
 654                         swp_entry_t entry;
 655
 656                         entry = pte_to_swp_entry(ptent);
 657                         if (!non_swap_entry(entry)) {
 658                                 nr_swap--;
 659                                 free_swap_and_cache(entry);
 660                                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 661                         } else if (is_hwpoison_entry(entry) ||
 662                                    is_poisoned_swp_entry(entry)) {
 663                                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 664                         }
 665                         continue;
 666                 }
 667
 668                 folio = vm_normal_folio(vma, addr, ptent);
 669                 if (!folio || folio_is_zone_device(folio))
 670                         continue;
 671
 672                 /*
 673                  * If pmd isn't transhuge but the folio is large and
 674                  * is owned by only this process, split it and
 675                  * deactivate all pages.
 676                  */
 677                 if (folio_test_large(folio)) {
 678                         int err;
 679
 680                         if (folio_estimated_sharers(folio) != 1)
 681                                 break;
 682                         if (!folio_trylock(folio))
 683                                 break;
 684                         folio_get(folio);
 685                         arch_leave_lazy_mmu_mode();
 686                         pte_unmap_unlock(start_pte, ptl);
 687                         start_pte = NULL;
 688                         err = split_folio(folio);
 689                         folio_unlock(folio);
 690                         folio_put(folio);
 691                         if (err)
 692                                 break;
 693                         start_pte = pte =
 694                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 695                         if (!start_pte)
 696                                 break;
 697                         arch_enter_lazy_mmu_mode();
 698                         pte--;
 699                         addr -= PAGE_SIZE;
 700                         continue;
 701                 }
 702
 703                 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
 704                         if (!folio_trylock(folio))
 705                                 continue;
 706                         /*
 707                          * If folio is shared with others, we mustn't clear
 708                          * the folio's dirty flag.
 709                          */
 710                         if (folio_mapcount(folio) != 1) {
 711                                 folio_unlock(folio);
 712                                 continue;
 713                         }
 714
 715                         if (folio_test_swapcache(folio) &&
 716                             !folio_free_swap(folio)) {
 717                                 folio_unlock(folio);
 718                                 continue;
 719                         }
 720
 721                         folio_clear_dirty(folio);
 722                         folio_unlock(folio);
 723                 }
 724
 725                 if (pte_young(ptent) || pte_dirty(ptent)) {
 726                         /*
 727                          * Some of architecture(ex, PPC) don't update TLB
 728                          * with set_pte_at and tlb_remove_tlb_entry so for
 729                          * the portability, remap the pte with old|clean
 730                          * after pte clearing.
 731                          */
 732                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 733                                                         tlb->fullmm);
 734
 735                         ptent = pte_mkold(ptent);
 736                         ptent = pte_mkclean(ptent);
 737                         set_pte_at(mm, addr, pte, ptent);
 738                         tlb_remove_tlb_entry(tlb, pte, addr);
 739                 }
 740                 folio_mark_lazyfree(folio);
 741         }
 742
 743         if (nr_swap)
 744                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 745         if (start_pte) {
 746                 arch_leave_lazy_mmu_mode();
 747                 pte_unmap_unlock(start_pte, ptl);
 748         }
 749         cond_resched();
 750
 751         return 0;
 752 }
 753
 754 static const struct mm_walk_ops madvise_free_walk_ops = {
 755         .pmd_entry              = madvise_free_pte_range,
 756         .walk_lock              = PGWALK_RDLOCK,
 757 };
 758
 759 static int madvise_free_single_vma(struct vm_area_struct *vma,
 760                         unsigned long start_addr, unsigned long end_addr)
 761 {
 762         struct mm_struct *mm = vma->vm_mm;
 763         struct mmu_notifier_range range;
 764         struct mmu_gather tlb;
 765
 766         /* MADV_FREE works for only anon vma at the moment */
 767         if (!vma_is_anonymous(vma))
 768                 return -EINVAL;
 769
 770         range.start = max(vma->vm_start, start_addr);
 771         if (range.start >= vma->vm_end)
 772                 return -EINVAL;
 773         range.end = min(vma->vm_end, end_addr);
 774         if (range.end <= vma->vm_start)
 775                 return -EINVAL;
 776         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
 777                                 range.start, range.end);
 778
 779         lru_add_drain();
 780         tlb_gather_mmu(&tlb, mm);
 781         update_hiwater_rss(mm);
 782
 783         mmu_notifier_invalidate_range_start(&range);
 784         tlb_start_vma(&tlb, vma);
 785         walk_page_range(vma->vm_mm, range.start, range.end,
 786                         &madvise_free_walk_ops, &tlb);
 787         tlb_end_vma(&tlb, vma);
 788         mmu_notifier_invalidate_range_end(&range);
 789         tlb_finish_mmu(&tlb);
 790
 791         return 0;
 792 }
 793
 794 /*
 795  * Application no longer needs these pages.  If the pages are dirty,
 796  * it's OK to just throw them away.  The app will be more careful about
 797  * data it wants to keep.  Be sure to free swap resources too.  The
 798  * zap_page_range_single call sets things up for shrink_active_list to actually
 799  * free these pages later if no one else has touched them in the meantime,
 800  * although we could add these pages to a global reuse list for
 801  * shrink_active_list to pick up before reclaiming other pages.
 802  *
 803  * NB: This interface discards data rather than pushes it out to swap,
 804  * as some implementations do.  This has performance implications for
 805  * applications like large transactional databases which want to discard
 806  * pages in anonymous maps after committing to backing store the data
 807  * that was kept in them.  There is no reason to write this data out to
 808  * the swap area if the application is discarding it.
 809  *
 810  * An interface that causes the system to free clean pages and flush
 811  * dirty pages is already available as msync(MS_INVALIDATE).
 812  */
 813 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 814                                         unsigned long start, unsigned long end)
 815 {
 816         zap_page_range_single(vma, start, end - start, NULL);
 817         return 0;
 818 }
 819
 820 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
 821                                             unsigned long start,
 822                                             unsigned long *end,
 823                                             int behavior)
 824 {
 825         if (!is_vm_hugetlb_page(vma)) {
 826                 unsigned int forbidden = VM_PFNMAP;
 827
 828                 if (behavior != MADV_DONTNEED_LOCKED)
 829                         forbidden |= VM_LOCKED;
 830
 831                 return !(vma->vm_flags & forbidden);
 832         }
 833
 834         if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
 835                 return false;
 836         if (start & ~huge_page_mask(hstate_vma(vma)))
 837                 return false;
 838
 839         /*
 840          * Madvise callers expect the length to be rounded up to PAGE_SIZE
 841          * boundaries, and may be unaware that this VMA uses huge pages.
 842          * Avoid unexpected data loss by rounding down the number of
 843          * huge pages freed.
 844          */
 845         *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
 846
 847         return true;
 848 }
 849
 850 static long madvise_dontneed_free(struct vm_area_struct *vma,
 851                                   struct vm_area_struct **prev,
 852                                   unsigned long start, unsigned long end,
 853                                   int behavior)
 854 {
 855         struct mm_struct *mm = vma->vm_mm;
 856
 857         *prev = vma;
 858         if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
 859                 return -EINVAL;
 860
 861         if (start == end)
 862                 return 0;
 863
 864         if (!userfaultfd_remove(vma, start, end)) {
 865                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 866
 867                 mmap_read_lock(mm);
 868                 vma = vma_lookup(mm, start);
 869                 if (!vma)
 870                         return -ENOMEM;
 871                 /*
 872                  * Potential end adjustment for hugetlb vma is OK as
 873                  * the check below keeps end within vma.
 874                  */
 875                 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
 876                                                      behavior))
 877                         return -EINVAL;
 878                 if (end > vma->vm_end) {
 879                         /*
 880                          * Don't fail if end > vma->vm_end. If the old
 881                          * vma was split while the mmap_lock was
 882                          * released the effect of the concurrent
 883                          * operation may not cause madvise() to
 884                          * have an undefined result. There may be an
 885                          * adjacent next vma that we'll walk
 886                          * next. userfaultfd_remove() will generate an
 887                          * UFFD_EVENT_REMOVE repetition on the
 888                          * end-vma->vm_end range, but the manager can
 889                          * handle a repetition fine.
 890                          */
 891                         end = vma->vm_end;
 892                 }
 893                 VM_WARN_ON(start >= end);
 894         }
 895
 896         if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
 897                 return madvise_dontneed_single_vma(vma, start, end);
 898         else if (behavior == MADV_FREE)
 899                 return madvise_free_single_vma(vma, start, end);
 900         else
 901                 return -EINVAL;
 902 }
 903
 904 static long madvise_populate(struct vm_area_struct *vma,
 905                              struct vm_area_struct **prev,
 906                              unsigned long start, unsigned long end,
 907                              int behavior)
 908 {
 909         const bool write = behavior == MADV_POPULATE_WRITE;
 910         struct mm_struct *mm = vma->vm_mm;
 911         unsigned long tmp_end;
 912         int locked = 1;
 913         long pages;
 914
 915         *prev = vma;
 916
 917         while (start < end) {
 918                 /*
 919                  * We might have temporarily dropped the lock. For example,
 920                  * our VMA might have been split.
 921                  */
 922                 if (!vma || start >= vma->vm_end) {
 923                         vma = vma_lookup(mm, start);
 924                         if (!vma)
 925                                 return -ENOMEM;
 926                 }
 927
 928                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 929                 /* Populate (prefault) page tables readable/writable. */
 930                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 931                                                &locked);
 932                 if (!locked) {
 933                         mmap_read_lock(mm);
 934                         locked = 1;
 935                         *prev = NULL;
 936                         vma = NULL;
 937                 }
 938                 if (pages < 0) {
 939                         switch (pages) {
 940                         case -EINTR:
 941                                 return -EINTR;
 942                         case -EINVAL: /* Incompatible mappings / permissions. */
 943                                 return -EINVAL;
 944                         case -EHWPOISON:
 945                                 return -EHWPOISON;
 946                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 947                                 return -EFAULT;
 948                         default:
 949                                 pr_warn_once("%s: unhandled return value: %ld\n",
 950                                              __func__, pages);
 951                                 fallthrough;
 952                         case -ENOMEM:
 953                                 return -ENOMEM;
 954                         }
 955                 }
 956                 start += pages * PAGE_SIZE;
 957         }
 958         return 0;
 959 }
 960
 961 /*
 962  * Application wants to free up the pages and associated backing store.
 963  * This is effectively punching a hole into the middle of a file.
 964  */
 965 static long madvise_remove(struct vm_area_struct *vma,
 966                                 struct vm_area_struct **prev,
 967                                 unsigned long start, unsigned long end)
 968 {
 969         loff_t offset;
 970         int error;
 971         struct file *f;
 972         struct mm_struct *mm = vma->vm_mm;
 973
 974         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 975
 976         if (vma->vm_flags & VM_LOCKED)
 977                 return -EINVAL;
 978
 979         f = vma->vm_file;
 980
 981         if (!f || !f->f_mapping || !f->f_mapping->host) {
 982                         return -EINVAL;
 983         }
 984
 985         if (!vma_is_shared_maywrite(vma))
 986                 return -EACCES;
 987
 988         offset = (loff_t)(start - vma->vm_start)
 989                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 990
 991         /*
 992          * Filesystem's fallocate may need to take i_rwsem.  We need to
 993          * explicitly grab a reference because the vma (and hence the
 994          * vma's reference to the file) can go away as soon as we drop
 995          * mmap_lock.
 996          */
 997         get_file(f);
 998         if (userfaultfd_remove(vma, start, end)) {
 999                 /* mmap_lock was not released by userfaultfd_remove() */
1000                 mmap_read_unlock(mm);
1001         }
1002         error = vfs_fallocate(f,
1003                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1004                                 offset, end - start);
1005         fput(f);
1006         mmap_read_lock(mm);
1007         return error;
1008 }
1009
1010 /*
1011  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
1012  * will handle splitting a vm area into separate areas, each area with its own
1013  * behavior.
1014  */
1015 static int madvise_vma_behavior(struct vm_area_struct *vma,
1016                                 struct vm_area_struct **prev,
1017                                 unsigned long start, unsigned long end,
1018                                 unsigned long behavior)
1019 {
1020         int error;
1021         struct anon_vma_name *anon_name;
1022         unsigned long new_flags = vma->vm_flags;
1023
1024         switch (behavior) {
1025         case MADV_REMOVE:
1026                 return madvise_remove(vma, prev, start, end);
1027         case MADV_WILLNEED:
1028                 return madvise_willneed(vma, prev, start, end);
1029         case MADV_COLD:
1030                 return madvise_cold(vma, prev, start, end);
1031         case MADV_PAGEOUT:
1032                 return madvise_pageout(vma, prev, start, end);
1033         case MADV_FREE:
1034         case MADV_DONTNEED:
1035         case MADV_DONTNEED_LOCKED:
1036                 return madvise_dontneed_free(vma, prev, start, end, behavior);
1037         case MADV_POPULATE_READ:
1038         case MADV_POPULATE_WRITE:
1039                 return madvise_populate(vma, prev, start, end, behavior);
1040         case MADV_NORMAL:
1041                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1042                 break;
1043         case MADV_SEQUENTIAL:
1044                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1045                 break;
1046         case MADV_RANDOM:
1047                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1048                 break;
1049         case MADV_DONTFORK:
1050                 new_flags |= VM_DONTCOPY;
1051                 break;
1052         case MADV_DOFORK:
1053                 if (vma->vm_flags & VM_IO)
1054                         return -EINVAL;
1055                 new_flags &= ~VM_DONTCOPY;
1056                 break;
1057         case MADV_WIPEONFORK:
1058                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1059                 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1060                         return -EINVAL;
1061                 new_flags |= VM_WIPEONFORK;
1062                 break;
1063         case MADV_KEEPONFORK:
1064                 new_flags &= ~VM_WIPEONFORK;
1065                 break;
1066         case MADV_DONTDUMP:
1067                 new_flags |= VM_DONTDUMP;
1068                 break;
1069         case MADV_DODUMP:
1070                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1071                         return -EINVAL;
1072                 new_flags &= ~VM_DONTDUMP;
1073                 break;
1074         case MADV_MERGEABLE:
1075         case MADV_UNMERGEABLE:
1076                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1077                 if (error)
1078                         goto out;
1079                 break;
1080         case MADV_HUGEPAGE:
1081         case MADV_NOHUGEPAGE:
1082                 error = hugepage_madvise(vma, &new_flags, behavior);
1083                 if (error)
1084                         goto out;
1085                 break;
1086         case MADV_COLLAPSE:
1087                 return madvise_collapse(vma, prev, start, end);
1088         }
1089
1090         anon_name = anon_vma_name(vma);
1091         anon_vma_name_get(anon_name);
1092         error = madvise_update_vma(vma, prev, start, end, new_flags,
1093                                    anon_name);
1094         anon_vma_name_put(anon_name);
1095
1096 out:
1097         /*
1098          * madvise() returns EAGAIN if kernel resources, such as
1099          * slab, are temporarily unavailable.
1100          */
1101         if (error == -ENOMEM)
1102                 error = -EAGAIN;
1103         return error;
1104 }
1105
1106 #ifdef CONFIG_MEMORY_FAILURE
1107 /*
1108  * Error injection support for memory error handling.
1109  */
1110 static int madvise_inject_error(int behavior,
1111                 unsigned long start, unsigned long end)
1112 {
1113         unsigned long size;
1114
1115         if (!capable(CAP_SYS_ADMIN))
1116                 return -EPERM;
1117
1118
1119         for (; start < end; start += size) {
1120                 unsigned long pfn;
1121                 struct page *page;
1122                 int ret;
1123
1124                 ret = get_user_pages_fast(start, 1, 0, &page);
1125                 if (ret != 1)
1126                         return ret;
1127                 pfn = page_to_pfn(page);
1128
1129                 /*
1130                  * When soft offlining hugepages, after migrating the page
1131                  * we dissolve it, therefore in the second loop "page" will
1132                  * no longer be a compound page.
1133                  */
1134                 size = page_size(compound_head(page));
1135
1136                 if (behavior == MADV_SOFT_OFFLINE) {
1137                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1138                                  pfn, start);
1139                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1140                 } else {
1141                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1142                                  pfn, start);
1143                         ret = memory_failure(pfn, MF_COUNT_INCREASED | MF_SW_SIMULATED);
1144                         if (ret == -EOPNOTSUPP)
1145                                 ret = 0;
1146                 }
1147
1148                 if (ret)
1149                         return ret;
1150         }
1151
1152         return 0;
1153 }
1154 #endif
1155
1156 static bool
1157 madvise_behavior_valid(int behavior)
1158 {
1159         switch (behavior) {
1160         case MADV_DOFORK:
1161         case MADV_DONTFORK:
1162         case MADV_NORMAL:
1163         case MADV_SEQUENTIAL:
1164         case MADV_RANDOM:
1165         case MADV_REMOVE:
1166         case MADV_WILLNEED:
1167         case MADV_DONTNEED:
1168         case MADV_DONTNEED_LOCKED:
1169         case MADV_FREE:
1170         case MADV_COLD:
1171         case MADV_PAGEOUT:
1172         case MADV_POPULATE_READ:
1173         case MADV_POPULATE_WRITE:
1174 #ifdef CONFIG_KSM
1175         case MADV_MERGEABLE:
1176         case MADV_UNMERGEABLE:
1177 #endif
1178 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1179         case MADV_HUGEPAGE:
1180         case MADV_NOHUGEPAGE:
1181         case MADV_COLLAPSE:
1182 #endif
1183         case MADV_DONTDUMP:
1184         case MADV_DODUMP:
1185         case MADV_WIPEONFORK:
1186         case MADV_KEEPONFORK:
1187 #ifdef CONFIG_MEMORY_FAILURE
1188         case MADV_SOFT_OFFLINE:
1189         case MADV_HWPOISON:
1190 #endif
1191                 return true;
1192
1193         default:
1194                 return false;
1195         }
1196 }
1197
1198 static bool process_madvise_behavior_valid(int behavior)
1199 {
1200         switch (behavior) {
1201         case MADV_COLD:
1202         case MADV_PAGEOUT:
1203         case MADV_WILLNEED:
1204         case MADV_COLLAPSE:
1205                 return true;
1206         default:
1207                 return false;
1208         }
1209 }
1210
1211 /*
1212  * Walk the vmas in range [start,end), and call the visit function on each one.
1213  * The visit function will get start and end parameters that cover the overlap
1214  * between the current vma and the original range.  Any unmapped regions in the
1215  * original range will result in this function returning -ENOMEM while still
1216  * calling the visit function on all of the existing vmas in the range.
1217  * Must be called with the mmap_lock held for reading or writing.
1218  */
1219 static
1220 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1221                       unsigned long end, unsigned long arg,
1222                       int (*visit)(struct vm_area_struct *vma,
1223                                    struct vm_area_struct **prev, unsigned long start,
1224                                    unsigned long end, unsigned long arg))
1225 {
1226         struct vm_area_struct *vma;
1227         struct vm_area_struct *prev;
1228         unsigned long tmp;
1229         int unmapped_error = 0;
1230
1231         /*
1232          * If the interval [start,end) covers some unmapped address
1233          * ranges, just ignore them, but return -ENOMEM at the end.
1234          * - different from the way of handling in mlock etc.
1235          */
1236         vma = find_vma_prev(mm, start, &prev);
1237         if (vma && start > vma->vm_start)
1238                 prev = vma;
1239
1240         for (;;) {
1241                 int error;
1242
1243                 /* Still start < end. */
1244                 if (!vma)
1245                         return -ENOMEM;
1246
1247                 /* Here start < (end|vma->vm_end). */
1248                 if (start < vma->vm_start) {
1249                         unmapped_error = -ENOMEM;
1250                         start = vma->vm_start;
1251                         if (start >= end)
1252                                 break;
1253                 }
1254
1255                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1256                 tmp = vma->vm_end;
1257                 if (end < tmp)
1258                         tmp = end;
1259
1260                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1261                 error = visit(vma, &prev, start, tmp, arg);
1262                 if (error)
1263                         return error;
1264                 start = tmp;
1265                 if (prev && start < prev->vm_end)
1266                         start = prev->vm_end;
1267                 if (start >= end)
1268                         break;
1269                 if (prev)
1270                         vma = find_vma(mm, prev->vm_end);
1271                 else    /* madvise_remove dropped mmap_lock */
1272                         vma = find_vma(mm, start);
1273         }
1274
1275         return unmapped_error;
1276 }
1277
1278 #ifdef CONFIG_ANON_VMA_NAME
1279 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1280                                  struct vm_area_struct **prev,
1281                                  unsigned long start, unsigned long end,
1282                                  unsigned long anon_name)
1283 {
1284         int error;
1285
1286         /* Only anonymous mappings can be named */
1287         if (vma->vm_file && !vma_is_anon_shmem(vma))
1288                 return -EBADF;
1289
1290         error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1291                                    (struct anon_vma_name *)anon_name);
1292
1293         /*
1294          * madvise() returns EAGAIN if kernel resources, such as
1295          * slab, are temporarily unavailable.
1296          */
1297         if (error == -ENOMEM)
1298                 error = -EAGAIN;
1299         return error;
1300 }
1301
1302 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1303                           unsigned long len_in, struct anon_vma_name *anon_name)
1304 {
1305         unsigned long end;
1306         unsigned long len;
1307
1308         if (start & ~PAGE_MASK)
1309                 return -EINVAL;
1310         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1311
1312         /* Check to see whether len was rounded up from small -ve to zero */
1313         if (len_in && !len)
1314                 return -EINVAL;
1315
1316         end = start + len;
1317         if (end < start)
1318                 return -EINVAL;
1319
1320         if (end == start)
1321                 return 0;
1322
1323         return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1324                                  madvise_vma_anon_name);
1325 }
1326 #endif /* CONFIG_ANON_VMA_NAME */
1327 /*
1328  * The madvise(2) system call.
1329  *
1330  * Applications can use madvise() to advise the kernel how it should
1331  * handle paging I/O in this VM area.  The idea is to help the kernel
1332  * use appropriate read-ahead and caching techniques.  The information
1333  * provided is advisory only, and can be safely disregarded by the
1334  * kernel without affecting the correct operation of the application.
1335  *
1336  * behavior values:
1337  *  MADV_NORMAL - the default behavior is to read clusters.  This
1338  *              results in some read-ahead and read-behind.
1339  *  MADV_RANDOM - the system should read the minimum amount of data
1340  *              on any access, since it is unlikely that the appli-
1341  *              cation will need more than what it asks for.
1342  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1343  *              once, so they can be aggressively read ahead, and
1344  *              can be freed soon after they are accessed.
1345  *  MADV_WILLNEED - the application is notifying the system to read
1346  *              some pages ahead.
1347  *  MADV_DONTNEED - the application is finished with the given range,
1348  *              so the kernel can free resources associated with it.
1349  *  MADV_FREE - the application marks pages in the given range as lazy free,
1350  *              where actual purges are postponed until memory pressure happens.
1351  *  MADV_REMOVE - the application wants to free up the given range of
1352  *              pages and associated backing store.
1353  *  MADV_DONTFORK - omit this area from child's address space when forking:
1354  *              typically, to avoid COWing pages pinned by get_user_pages().
1355  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1356  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1357  *              range after a fork.
1358  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1359  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1360  *              were corrupted by unrecoverable hardware memory failure.
1361  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1362  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1363  *              this area with pages of identical content from other such areas.
1364  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1365  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1366  *              huge pages in the future. Existing pages might be coalesced and
1367  *              new pages might be allocated as THP.
1368  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1369  *              transparent huge pages so the existing pages will not be
1370  *              coalesced into THP and new pages will not be allocated as THP.
1371  *  MADV_COLLAPSE - synchronously coalesce pages into new THP.
1372  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1373  *              from being included in its core dump.
1374  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1375  *  MADV_COLD - the application is not expected to use this memory soon,
1376  *              deactivate pages in this range so that they can be reclaimed
1377  *              easily if memory pressure happens.
1378  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1379  *              page out the pages in this range immediately.
1380  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1381  *              triggering read faults if required
1382  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1383  *              triggering write faults if required
1384  *
1385  * return values:
1386  *  zero    - success
1387  *  -EINVAL - start + len < 0, start is not page-aligned,
1388  *              "behavior" is not a valid value, or application
1389  *              is attempting to release locked or shared pages,
1390  *              or the specified address range includes file, Huge TLB,
1391  *              MAP_SHARED or VMPFNMAP range.
1392  *  -ENOMEM - addresses in the specified range are not currently
1393  *              mapped, or are outside the AS of the process.
1394  *  -EIO    - an I/O error occurred while paging in data.
1395  *  -EBADF  - map exists, but area maps something that isn't a file.
1396  *  -EAGAIN - a kernel resource was temporarily unavailable.
1397  */
1398 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1399 {
1400         unsigned long end;
1401         int error;
1402         int write;
1403         size_t len;
1404         struct blk_plug plug;
1405
1406         if (!madvise_behavior_valid(behavior))
1407                 return -EINVAL;
1408
1409         if (!PAGE_ALIGNED(start))
1410                 return -EINVAL;
1411         len = PAGE_ALIGN(len_in);
1412
1413         /* Check to see whether len was rounded up from small -ve to zero */
1414         if (len_in && !len)
1415                 return -EINVAL;
1416
1417         end = start + len;
1418         if (end < start)
1419                 return -EINVAL;
1420
1421         if (end == start)
1422                 return 0;
1423
1424 #ifdef CONFIG_MEMORY_FAILURE
1425         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1426                 return madvise_inject_error(behavior, start, start + len_in);
1427 #endif
1428
1429         write = madvise_need_mmap_write(behavior);
1430         if (write) {
1431                 if (mmap_write_lock_killable(mm))
1432                         return -EINTR;
1433         } else {
1434                 mmap_read_lock(mm);
1435         }
1436
1437         start = untagged_addr_remote(mm, start);
1438         end = start + len;
1439
1440         blk_start_plug(&plug);
1441         error = madvise_walk_vmas(mm, start, end, behavior,
1442                         madvise_vma_behavior);
1443         blk_finish_plug(&plug);
1444         if (write)
1445                 mmap_write_unlock(mm);
1446         else
1447                 mmap_read_unlock(mm);
1448
1449         return error;
1450 }
1451
1452 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1453 {
1454         return do_madvise(current->mm, start, len_in, behavior);
1455 }
1456
1457 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1458                 size_t, vlen, int, behavior, unsigned int, flags)
1459 {
1460         ssize_t ret;
1461         struct iovec iovstack[UIO_FASTIOV];
1462         struct iovec *iov = iovstack;
1463         struct iov_iter iter;
1464         struct task_struct *task;
1465         struct mm_struct *mm;
1466         size_t total_len;
1467         unsigned int f_flags;
1468
1469         if (flags != 0) {
1470                 ret = -EINVAL;
1471                 goto out;
1472         }
1473
1474         ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1475         if (ret < 0)
1476                 goto out;
1477
1478         task = pidfd_get_task(pidfd, &f_flags);
1479         if (IS_ERR(task)) {
1480                 ret = PTR_ERR(task);
1481                 goto free_iov;
1482         }
1483
1484         if (!process_madvise_behavior_valid(behavior)) {
1485                 ret = -EINVAL;
1486                 goto release_task;
1487         }
1488
1489         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1490         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1491         if (IS_ERR_OR_NULL(mm)) {
1492                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1493                 goto release_task;
1494         }
1495
1496         /*
1497          * Require CAP_SYS_NICE for influencing process performance. Note that
1498          * only non-destructive hints are currently supported.
1499          */
1500         if (!capable(CAP_SYS_NICE)) {
1501                 ret = -EPERM;
1502                 goto release_mm;
1503         }
1504
1505         total_len = iov_iter_count(&iter);
1506
1507         while (iov_iter_count(&iter)) {
1508                 ret = do_madvise(mm, (unsigned long)iter_iov_addr(&iter),
1509                                         iter_iov_len(&iter), behavior);
1510                 if (ret < 0)
1511                         break;
1512                 iov_iter_advance(&iter, iter_iov_len(&iter));
1513         }
1514
1515         ret = (total_len - iov_iter_count(&iter)) ? : ret;
1516
1517 release_mm:
1518         mmput(mm);
1519 release_task:
1520         put_task_struct(task);
1521 free_iov:
1522         kfree(iov);
1523 out:
1524         return ret;
1525 }