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