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