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_DONTNEED_LOCKED:
  56         case MADV_COLD:
  57         case MADV_PAGEOUT:
  58         case MADV_FREE:
  59         case MADV_POPULATE_READ:
  60         case MADV_POPULATE_WRITE:
  61                 return 0;
  62         default:
  63                 /* be safe, default to 1. list exceptions explicitly */
  64                 return 1;
  65         }
  66 }
  67
  68 #ifdef CONFIG_ANON_VMA_NAME
  69 struct anon_vma_name *anon_vma_name_alloc(const char *name)
  70 {
  71         struct anon_vma_name *anon_name;
  72         size_t count;
  73
  74         /* Add 1 for NUL terminator at the end of the anon_name->name */
  75         count = strlen(name) + 1;
  76         anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
  77         if (anon_name) {
  78                 kref_init(&anon_name->kref);
  79                 memcpy(anon_name->name, name, count);
  80         }
  81
  82         return anon_name;
  83 }
  84
  85 void anon_vma_name_free(struct kref *kref)
  86 {
  87         struct anon_vma_name *anon_name =
  88                         container_of(kref, struct anon_vma_name, kref);
  89         kfree(anon_name);
  90 }
  91
  92 struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
  93 {
  94         mmap_assert_locked(vma->vm_mm);
  95
  96         if (vma->vm_file)
  97                 return NULL;
  98
  99         return vma->anon_name;
 100 }
 101
 102 /* mmap_lock should be write-locked */
 103 static int replace_anon_vma_name(struct vm_area_struct *vma,
 104                                  struct anon_vma_name *anon_name)
 105 {
 106         struct anon_vma_name *orig_name = anon_vma_name(vma);
 107
 108         if (!anon_name) {
 109                 vma->anon_name = NULL;
 110                 anon_vma_name_put(orig_name);
 111                 return 0;
 112         }
 113
 114         if (anon_vma_name_eq(orig_name, anon_name))
 115                 return 0;
 116
 117         vma->anon_name = anon_vma_name_reuse(anon_name);
 118         anon_vma_name_put(orig_name);
 119
 120         return 0;
 121 }
 122 #else /* CONFIG_ANON_VMA_NAME */
 123 static int replace_anon_vma_name(struct vm_area_struct *vma,
 124                                  struct anon_vma_name *anon_name)
 125 {
 126         if (anon_name)
 127                 return -EINVAL;
 128
 129         return 0;
 130 }
 131 #endif /* CONFIG_ANON_VMA_NAME */
 132 /*
 133  * Update the vm_flags on region of a vma, splitting it or merging it as
 134  * necessary.  Must be called with mmap_sem held for writing;
 135  * Caller should ensure anon_name stability by raising its refcount even when
 136  * anon_name belongs to a valid vma because this function might free that vma.
 137  */
 138 static int madvise_update_vma(struct vm_area_struct *vma,
 139                               struct vm_area_struct **prev, unsigned long start,
 140                               unsigned long end, unsigned long new_flags,
 141                               struct anon_vma_name *anon_name)
 142 {
 143         struct mm_struct *mm = vma->vm_mm;
 144         int error;
 145         pgoff_t pgoff;
 146
 147         if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
 148                 *prev = vma;
 149                 return 0;
 150         }
 151
 152         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 153         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 154                           vma->vm_file, pgoff, vma_policy(vma),
 155                           vma->vm_userfaultfd_ctx, anon_name);
 156         if (*prev) {
 157                 vma = *prev;
 158                 goto success;
 159         }
 160
 161         *prev = vma;
 162
 163         if (start != vma->vm_start) {
 164                 if (unlikely(mm->map_count >= sysctl_max_map_count))
 165                         return -ENOMEM;
 166                 error = __split_vma(mm, vma, start, 1);
 167                 if (error)
 168                         return error;
 169         }
 170
 171         if (end != vma->vm_end) {
 172                 if (unlikely(mm->map_count >= sysctl_max_map_count))
 173                         return -ENOMEM;
 174                 error = __split_vma(mm, vma, end, 0);
 175                 if (error)
 176                         return error;
 177         }
 178
 179 success:
 180         /*
 181          * vm_flags is protected by the mmap_lock held in write mode.
 182          */
 183         vma->vm_flags = new_flags;
 184         if (!vma->vm_file) {
 185                 error = replace_anon_vma_name(vma, anon_name);
 186                 if (error)
 187                         return error;
 188         }
 189
 190         return 0;
 191 }
 192
 193 #ifdef CONFIG_SWAP
 194 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 195         unsigned long end, struct mm_walk *walk)
 196 {
 197         pte_t *orig_pte;
 198         struct vm_area_struct *vma = walk->private;
 199         unsigned long index;
 200
 201         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 202                 return 0;
 203
 204         for (index = start; index != end; index += PAGE_SIZE) {
 205                 pte_t pte;
 206                 swp_entry_t entry;
 207                 struct page *page;
 208                 spinlock_t *ptl;
 209
 210                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 211                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 212                 pte_unmap_unlock(orig_pte, ptl);
 213
 214                 if (pte_present(pte) || pte_none(pte))
 215                         continue;
 216                 entry = pte_to_swp_entry(pte);
 217                 if (unlikely(non_swap_entry(entry)))
 218                         continue;
 219
 220                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 221                                                         vma, index, false);
 222                 if (page)
 223                         put_page(page);
 224         }
 225
 226         return 0;
 227 }
 228
 229 static const struct mm_walk_ops swapin_walk_ops = {
 230         .pmd_entry              = swapin_walk_pmd_entry,
 231 };
 232
 233 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 234                 unsigned long start, unsigned long end,
 235                 struct address_space *mapping)
 236 {
 237         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 238         pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
 239         struct page *page;
 240
 241         rcu_read_lock();
 242         xas_for_each(&xas, page, end_index) {
 243                 swp_entry_t swap;
 244
 245                 if (!xa_is_value(page))
 246                         continue;
 247                 xas_pause(&xas);
 248                 rcu_read_unlock();
 249
 250                 swap = radix_to_swp_entry(page);
 251                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 252                                                         NULL, 0, false);
 253                 if (page)
 254                         put_page(page);
 255
 256                 rcu_read_lock();
 257         }
 258         rcu_read_unlock();
 259
 260         lru_add_drain();        /* Push any new pages onto the LRU now */
 261 }
 262 #endif          /* CONFIG_SWAP */
 263
 264 /*
 265  * Schedule all required I/O operations.  Do not wait for completion.
 266  */
 267 static long madvise_willneed(struct vm_area_struct *vma,
 268                              struct vm_area_struct **prev,
 269                              unsigned long start, unsigned long end)
 270 {
 271         struct mm_struct *mm = vma->vm_mm;
 272         struct file *file = vma->vm_file;
 273         loff_t offset;
 274
 275         *prev = vma;
 276 #ifdef CONFIG_SWAP
 277         if (!file) {
 278                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 279                 lru_add_drain(); /* Push any new pages onto the LRU now */
 280                 return 0;
 281         }
 282
 283         if (shmem_mapping(file->f_mapping)) {
 284                 force_shm_swapin_readahead(vma, start, end,
 285                                         file->f_mapping);
 286                 return 0;
 287         }
 288 #else
 289         if (!file)
 290                 return -EBADF;
 291 #endif
 292
 293         if (IS_DAX(file_inode(file))) {
 294                 /* no bad return value, but ignore advice */
 295                 return 0;
 296         }
 297
 298         /*
 299          * Filesystem's fadvise may need to take various locks.  We need to
 300          * explicitly grab a reference because the vma (and hence the
 301          * vma's reference to the file) can go away as soon as we drop
 302          * mmap_lock.
 303          */
 304         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 305         get_file(file);
 306         offset = (loff_t)(start - vma->vm_start)
 307                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 308         mmap_read_unlock(mm);
 309         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 310         fput(file);
 311         mmap_read_lock(mm);
 312         return 0;
 313 }
 314
 315 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 316                                 unsigned long addr, unsigned long end,
 317                                 struct mm_walk *walk)
 318 {
 319         struct madvise_walk_private *private = walk->private;
 320         struct mmu_gather *tlb = private->tlb;
 321         bool pageout = private->pageout;
 322         struct mm_struct *mm = tlb->mm;
 323         struct vm_area_struct *vma = walk->vma;
 324         pte_t *orig_pte, *pte, ptent;
 325         spinlock_t *ptl;
 326         struct page *page = NULL;
 327         LIST_HEAD(page_list);
 328
 329         if (fatal_signal_pending(current))
 330                 return -EINTR;
 331
 332 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 333         if (pmd_trans_huge(*pmd)) {
 334                 pmd_t orig_pmd;
 335                 unsigned long next = pmd_addr_end(addr, end);
 336
 337                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 338                 ptl = pmd_trans_huge_lock(pmd, vma);
 339                 if (!ptl)
 340                         return 0;
 341
 342                 orig_pmd = *pmd;
 343                 if (is_huge_zero_pmd(orig_pmd))
 344                         goto huge_unlock;
 345
 346                 if (unlikely(!pmd_present(orig_pmd))) {
 347                         VM_BUG_ON(thp_migration_supported() &&
 348                                         !is_pmd_migration_entry(orig_pmd));
 349                         goto huge_unlock;
 350                 }
 351
 352                 page = pmd_page(orig_pmd);
 353
 354                 /* Do not interfere with other mappings of this page */
 355                 if (page_mapcount(page) != 1)
 356                         goto huge_unlock;
 357
 358                 if (next - addr != HPAGE_PMD_SIZE) {
 359                         int err;
 360
 361                         get_page(page);
 362                         spin_unlock(ptl);
 363                         lock_page(page);
 364                         err = split_huge_page(page);
 365                         unlock_page(page);
 366                         put_page(page);
 367                         if (!err)
 368                                 goto regular_page;
 369                         return 0;
 370                 }
 371
 372                 if (pmd_young(orig_pmd)) {
 373                         pmdp_invalidate(vma, addr, pmd);
 374                         orig_pmd = pmd_mkold(orig_pmd);
 375
 376                         set_pmd_at(mm, addr, pmd, orig_pmd);
 377                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 378                 }
 379
 380                 ClearPageReferenced(page);
 381                 test_and_clear_page_young(page);
 382                 if (pageout) {
 383                         if (!isolate_lru_page(page)) {
 384                                 if (PageUnevictable(page))
 385                                         putback_lru_page(page);
 386                                 else
 387                                         list_add(&page->lru, &page_list);
 388                         }
 389                 } else
 390                         deactivate_page(page);
 391 huge_unlock:
 392                 spin_unlock(ptl);
 393                 if (pageout)
 394                         reclaim_pages(&page_list);
 395                 return 0;
 396         }
 397
 398 regular_page:
 399         if (pmd_trans_unstable(pmd))
 400                 return 0;
 401 #endif
 402         tlb_change_page_size(tlb, PAGE_SIZE);
 403         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 404         flush_tlb_batched_pending(mm);
 405         arch_enter_lazy_mmu_mode();
 406         for (; addr < end; pte++, addr += PAGE_SIZE) {
 407                 ptent = *pte;
 408
 409                 if (pte_none(ptent))
 410                         continue;
 411
 412                 if (!pte_present(ptent))
 413                         continue;
 414
 415                 page = vm_normal_page(vma, addr, ptent);
 416                 if (!page)
 417                         continue;
 418
 419                 /*
 420                  * Creating a THP page is expensive so split it only if we
 421                  * are sure it's worth. Split it if we are only owner.
 422                  */
 423                 if (PageTransCompound(page)) {
 424                         if (page_mapcount(page) != 1)
 425                                 break;
 426                         get_page(page);
 427                         if (!trylock_page(page)) {
 428                                 put_page(page);
 429                                 break;
 430                         }
 431                         pte_unmap_unlock(orig_pte, ptl);
 432                         if (split_huge_page(page)) {
 433                                 unlock_page(page);
 434                                 put_page(page);
 435                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 436                                 break;
 437                         }
 438                         unlock_page(page);
 439                         put_page(page);
 440                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 441                         pte--;
 442                         addr -= PAGE_SIZE;
 443                         continue;
 444                 }
 445
 446                 /* Do not interfere with other mappings of this page */
 447                 if (page_mapcount(page) != 1)
 448                         continue;
 449
 450                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 451
 452                 if (pte_young(ptent)) {
 453                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 454                                                         tlb->fullmm);
 455                         ptent = pte_mkold(ptent);
 456                         set_pte_at(mm, addr, pte, ptent);
 457                         tlb_remove_tlb_entry(tlb, pte, addr);
 458                 }
 459
 460                 /*
 461                  * We are deactivating a page for accelerating reclaiming.
 462                  * VM couldn't reclaim the page unless we clear PG_young.
 463                  * As a side effect, it makes confuse idle-page tracking
 464                  * because they will miss recent referenced history.
 465                  */
 466                 ClearPageReferenced(page);
 467                 test_and_clear_page_young(page);
 468                 if (pageout) {
 469                         if (!isolate_lru_page(page)) {
 470                                 if (PageUnevictable(page))
 471                                         putback_lru_page(page);
 472                                 else
 473                                         list_add(&page->lru, &page_list);
 474                         }
 475                 } else
 476                         deactivate_page(page);
 477         }
 478
 479         arch_leave_lazy_mmu_mode();
 480         pte_unmap_unlock(orig_pte, ptl);
 481         if (pageout)
 482                 reclaim_pages(&page_list);
 483         cond_resched();
 484
 485         return 0;
 486 }
 487
 488 static const struct mm_walk_ops cold_walk_ops = {
 489         .pmd_entry = madvise_cold_or_pageout_pte_range,
 490 };
 491
 492 static void madvise_cold_page_range(struct mmu_gather *tlb,
 493                              struct vm_area_struct *vma,
 494                              unsigned long addr, unsigned long end)
 495 {
 496         struct madvise_walk_private walk_private = {
 497                 .pageout = false,
 498                 .tlb = tlb,
 499         };
 500
 501         tlb_start_vma(tlb, vma);
 502         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 503         tlb_end_vma(tlb, vma);
 504 }
 505
 506 static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
 507 {
 508         return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
 509 }
 510
 511 static long madvise_cold(struct vm_area_struct *vma,
 512                         struct vm_area_struct **prev,
 513                         unsigned long start_addr, unsigned long end_addr)
 514 {
 515         struct mm_struct *mm = vma->vm_mm;
 516         struct mmu_gather tlb;
 517
 518         *prev = vma;
 519         if (!can_madv_lru_vma(vma))
 520                 return -EINVAL;
 521
 522         lru_add_drain();
 523         tlb_gather_mmu(&tlb, mm);
 524         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 525         tlb_finish_mmu(&tlb);
 526
 527         return 0;
 528 }
 529
 530 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 531                              struct vm_area_struct *vma,
 532                              unsigned long addr, unsigned long end)
 533 {
 534         struct madvise_walk_private walk_private = {
 535                 .pageout = true,
 536                 .tlb = tlb,
 537         };
 538
 539         tlb_start_vma(tlb, vma);
 540         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 541         tlb_end_vma(tlb, vma);
 542 }
 543
 544 static inline bool can_do_pageout(struct vm_area_struct *vma)
 545 {
 546         if (vma_is_anonymous(vma))
 547                 return true;
 548         if (!vma->vm_file)
 549                 return false;
 550         /*
 551          * paging out pagecache only for non-anonymous mappings that correspond
 552          * to the files the calling process could (if tried) open for writing;
 553          * otherwise we'd be including shared non-exclusive mappings, which
 554          * opens a side channel.
 555          */
 556         return inode_owner_or_capable(&init_user_ns,
 557                                       file_inode(vma->vm_file)) ||
 558                file_permission(vma->vm_file, MAY_WRITE) == 0;
 559 }
 560
 561 static long madvise_pageout(struct vm_area_struct *vma,
 562                         struct vm_area_struct **prev,
 563                         unsigned long start_addr, unsigned long end_addr)
 564 {
 565         struct mm_struct *mm = vma->vm_mm;
 566         struct mmu_gather tlb;
 567
 568         *prev = vma;
 569         if (!can_madv_lru_vma(vma))
 570                 return -EINVAL;
 571
 572         if (!can_do_pageout(vma))
 573                 return 0;
 574
 575         lru_add_drain();
 576         tlb_gather_mmu(&tlb, mm);
 577         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 578         tlb_finish_mmu(&tlb);
 579
 580         return 0;
 581 }
 582
 583 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 584                                 unsigned long end, struct mm_walk *walk)
 585
 586 {
 587         struct mmu_gather *tlb = walk->private;
 588         struct mm_struct *mm = tlb->mm;
 589         struct vm_area_struct *vma = walk->vma;
 590         spinlock_t *ptl;
 591         pte_t *orig_pte, *pte, ptent;
 592         struct page *page;
 593         int nr_swap = 0;
 594         unsigned long next;
 595
 596         next = pmd_addr_end(addr, end);
 597         if (pmd_trans_huge(*pmd))
 598                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 599                         goto next;
 600
 601         if (pmd_trans_unstable(pmd))
 602                 return 0;
 603
 604         tlb_change_page_size(tlb, PAGE_SIZE);
 605         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 606         flush_tlb_batched_pending(mm);
 607         arch_enter_lazy_mmu_mode();
 608         for (; addr != end; pte++, addr += PAGE_SIZE) {
 609                 ptent = *pte;
 610
 611                 if (pte_none(ptent))
 612                         continue;
 613                 /*
 614                  * If the pte has swp_entry, just clear page table to
 615                  * prevent swap-in which is more expensive rather than
 616                  * (page allocation + zeroing).
 617                  */
 618                 if (!pte_present(ptent)) {
 619                         swp_entry_t entry;
 620
 621                         entry = pte_to_swp_entry(ptent);
 622                         if (non_swap_entry(entry))
 623                                 continue;
 624                         nr_swap--;
 625                         free_swap_and_cache(entry);
 626                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 627                         continue;
 628                 }
 629
 630                 page = vm_normal_page(vma, addr, ptent);
 631                 if (!page)
 632                         continue;
 633
 634                 /*
 635                  * If pmd isn't transhuge but the page is THP and
 636                  * is owned by only this process, split it and
 637                  * deactivate all pages.
 638                  */
 639                 if (PageTransCompound(page)) {
 640                         if (page_mapcount(page) != 1)
 641                                 goto out;
 642                         get_page(page);
 643                         if (!trylock_page(page)) {
 644                                 put_page(page);
 645                                 goto out;
 646                         }
 647                         pte_unmap_unlock(orig_pte, ptl);
 648                         if (split_huge_page(page)) {
 649                                 unlock_page(page);
 650                                 put_page(page);
 651                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 652                                 goto out;
 653                         }
 654                         unlock_page(page);
 655                         put_page(page);
 656                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 657                         pte--;
 658                         addr -= PAGE_SIZE;
 659                         continue;
 660                 }
 661
 662                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 663
 664                 if (PageSwapCache(page) || PageDirty(page)) {
 665                         if (!trylock_page(page))
 666                                 continue;
 667                         /*
 668                          * If page is shared with others, we couldn't clear
 669                          * PG_dirty of the page.
 670                          */
 671                         if (page_mapcount(page) != 1) {
 672                                 unlock_page(page);
 673                                 continue;
 674                         }
 675
 676                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 677                                 unlock_page(page);
 678                                 continue;
 679                         }
 680
 681                         ClearPageDirty(page);
 682                         unlock_page(page);
 683                 }
 684
 685                 if (pte_young(ptent) || pte_dirty(ptent)) {
 686                         /*
 687                          * Some of architecture(ex, PPC) don't update TLB
 688                          * with set_pte_at and tlb_remove_tlb_entry so for
 689                          * the portability, remap the pte with old|clean
 690                          * after pte clearing.
 691                          */
 692                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 693                                                         tlb->fullmm);
 694
 695                         ptent = pte_mkold(ptent);
 696                         ptent = pte_mkclean(ptent);
 697                         set_pte_at(mm, addr, pte, ptent);
 698                         tlb_remove_tlb_entry(tlb, pte, addr);
 699                 }
 700                 mark_page_lazyfree(page);
 701         }
 702 out:
 703         if (nr_swap) {
 704                 if (current->mm == mm)
 705                         sync_mm_rss(mm);
 706
 707                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 708         }
 709         arch_leave_lazy_mmu_mode();
 710         pte_unmap_unlock(orig_pte, ptl);
 711         cond_resched();
 712 next:
 713         return 0;
 714 }
 715
 716 static const struct mm_walk_ops madvise_free_walk_ops = {
 717         .pmd_entry              = madvise_free_pte_range,
 718 };
 719
 720 static int madvise_free_single_vma(struct vm_area_struct *vma,
 721                         unsigned long start_addr, unsigned long end_addr)
 722 {
 723         struct mm_struct *mm = vma->vm_mm;
 724         struct mmu_notifier_range range;
 725         struct mmu_gather tlb;
 726
 727         /* MADV_FREE works for only anon vma at the moment */
 728         if (!vma_is_anonymous(vma))
 729                 return -EINVAL;
 730
 731         range.start = max(vma->vm_start, start_addr);
 732         if (range.start >= vma->vm_end)
 733                 return -EINVAL;
 734         range.end = min(vma->vm_end, end_addr);
 735         if (range.end <= vma->vm_start)
 736                 return -EINVAL;
 737         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 738                                 range.start, range.end);
 739
 740         lru_add_drain();
 741         tlb_gather_mmu(&tlb, mm);
 742         update_hiwater_rss(mm);
 743
 744         mmu_notifier_invalidate_range_start(&range);
 745         tlb_start_vma(&tlb, vma);
 746         walk_page_range(vma->vm_mm, range.start, range.end,
 747                         &madvise_free_walk_ops, &tlb);
 748         tlb_end_vma(&tlb, vma);
 749         mmu_notifier_invalidate_range_end(&range);
 750         tlb_finish_mmu(&tlb);
 751
 752         return 0;
 753 }
 754
 755 /*
 756  * Application no longer needs these pages.  If the pages are dirty,
 757  * it's OK to just throw them away.  The app will be more careful about
 758  * data it wants to keep.  Be sure to free swap resources too.  The
 759  * zap_page_range call sets things up for shrink_active_list to actually free
 760  * these pages later if no one else has touched them in the meantime,
 761  * although we could add these pages to a global reuse list for
 762  * shrink_active_list to pick up before reclaiming other pages.
 763  *
 764  * NB: This interface discards data rather than pushes it out to swap,
 765  * as some implementations do.  This has performance implications for
 766  * applications like large transactional databases which want to discard
 767  * pages in anonymous maps after committing to backing store the data
 768  * that was kept in them.  There is no reason to write this data out to
 769  * the swap area if the application is discarding it.
 770  *
 771  * An interface that causes the system to free clean pages and flush
 772  * dirty pages is already available as msync(MS_INVALIDATE).
 773  */
 774 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 775                                         unsigned long start, unsigned long end)
 776 {
 777         zap_page_range(vma, start, end - start);
 778         return 0;
 779 }
 780
 781 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
 782                                             unsigned long start,
 783                                             unsigned long *end,
 784                                             int behavior)
 785 {
 786         if (!is_vm_hugetlb_page(vma)) {
 787                 unsigned int forbidden = VM_PFNMAP;
 788
 789                 if (behavior != MADV_DONTNEED_LOCKED)
 790                         forbidden |= VM_LOCKED;
 791
 792                 return !(vma->vm_flags & forbidden);
 793         }
 794
 795         if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
 796                 return false;
 797         if (start & ~huge_page_mask(hstate_vma(vma)))
 798                 return false;
 799
 800         *end = ALIGN(*end, huge_page_size(hstate_vma(vma)));
 801         return true;
 802 }
 803
 804 static long madvise_dontneed_free(struct vm_area_struct *vma,
 805                                   struct vm_area_struct **prev,
 806                                   unsigned long start, unsigned long end,
 807                                   int behavior)
 808 {
 809         struct mm_struct *mm = vma->vm_mm;
 810
 811         *prev = vma;
 812         if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
 813                 return -EINVAL;
 814
 815         if (!userfaultfd_remove(vma, start, end)) {
 816                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 817
 818                 mmap_read_lock(mm);
 819                 vma = find_vma(mm, start);
 820                 if (!vma)
 821                         return -ENOMEM;
 822                 if (start < vma->vm_start) {
 823                         /*
 824                          * This "vma" under revalidation is the one
 825                          * with the lowest vma->vm_start where start
 826                          * is also < vma->vm_end. If start <
 827                          * vma->vm_start it means an hole materialized
 828                          * in the user address space within the
 829                          * virtual range passed to MADV_DONTNEED
 830                          * or MADV_FREE.
 831                          */
 832                         return -ENOMEM;
 833                 }
 834                 /*
 835                  * Potential end adjustment for hugetlb vma is OK as
 836                  * the check below keeps end within vma.
 837                  */
 838                 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
 839                                                      behavior))
 840                         return -EINVAL;
 841                 if (end > vma->vm_end) {
 842                         /*
 843                          * Don't fail if end > vma->vm_end. If the old
 844                          * vma was split while the mmap_lock was
 845                          * released the effect of the concurrent
 846                          * operation may not cause madvise() to
 847                          * have an undefined result. There may be an
 848                          * adjacent next vma that we'll walk
 849                          * next. userfaultfd_remove() will generate an
 850                          * UFFD_EVENT_REMOVE repetition on the
 851                          * end-vma->vm_end range, but the manager can
 852                          * handle a repetition fine.
 853                          */
 854                         end = vma->vm_end;
 855                 }
 856                 VM_WARN_ON(start >= end);
 857         }
 858
 859         if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
 860                 return madvise_dontneed_single_vma(vma, start, end);
 861         else if (behavior == MADV_FREE)
 862                 return madvise_free_single_vma(vma, start, end);
 863         else
 864                 return -EINVAL;
 865 }
 866
 867 static long madvise_populate(struct vm_area_struct *vma,
 868                              struct vm_area_struct **prev,
 869                              unsigned long start, unsigned long end,
 870                              int behavior)
 871 {
 872         const bool write = behavior == MADV_POPULATE_WRITE;
 873         struct mm_struct *mm = vma->vm_mm;
 874         unsigned long tmp_end;
 875         int locked = 1;
 876         long pages;
 877
 878         *prev = vma;
 879
 880         while (start < end) {
 881                 /*
 882                  * We might have temporarily dropped the lock. For example,
 883                  * our VMA might have been split.
 884                  */
 885                 if (!vma || start >= vma->vm_end) {
 886                         vma = vma_lookup(mm, start);
 887                         if (!vma)
 888                                 return -ENOMEM;
 889                 }
 890
 891                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 892                 /* Populate (prefault) page tables readable/writable. */
 893                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 894                                                &locked);
 895                 if (!locked) {
 896                         mmap_read_lock(mm);
 897                         locked = 1;
 898                         *prev = NULL;
 899                         vma = NULL;
 900                 }
 901                 if (pages < 0) {
 902                         switch (pages) {
 903                         case -EINTR:
 904                                 return -EINTR;
 905                         case -EINVAL: /* Incompatible mappings / permissions. */
 906                                 return -EINVAL;
 907                         case -EHWPOISON:
 908                                 return -EHWPOISON;
 909                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 910                                 return -EFAULT;
 911                         default:
 912                                 pr_warn_once("%s: unhandled return value: %ld\n",
 913                                              __func__, pages);
 914                                 fallthrough;
 915                         case -ENOMEM:
 916                                 return -ENOMEM;
 917                         }
 918                 }
 919                 start += pages * PAGE_SIZE;
 920         }
 921         return 0;
 922 }
 923
 924 /*
 925  * Application wants to free up the pages and associated backing store.
 926  * This is effectively punching a hole into the middle of a file.
 927  */
 928 static long madvise_remove(struct vm_area_struct *vma,
 929                                 struct vm_area_struct **prev,
 930                                 unsigned long start, unsigned long end)
 931 {
 932         loff_t offset;
 933         int error;
 934         struct file *f;
 935         struct mm_struct *mm = vma->vm_mm;
 936
 937         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 938
 939         if (vma->vm_flags & VM_LOCKED)
 940                 return -EINVAL;
 941
 942         f = vma->vm_file;
 943
 944         if (!f || !f->f_mapping || !f->f_mapping->host) {
 945                         return -EINVAL;
 946         }
 947
 948         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 949                 return -EACCES;
 950
 951         offset = (loff_t)(start - vma->vm_start)
 952                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 953
 954         /*
 955          * Filesystem's fallocate may need to take i_rwsem.  We need to
 956          * explicitly grab a reference because the vma (and hence the
 957          * vma's reference to the file) can go away as soon as we drop
 958          * mmap_lock.
 959          */
 960         get_file(f);
 961         if (userfaultfd_remove(vma, start, end)) {
 962                 /* mmap_lock was not released by userfaultfd_remove() */
 963                 mmap_read_unlock(mm);
 964         }
 965         error = vfs_fallocate(f,
 966                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 967                                 offset, end - start);
 968         fput(f);
 969         mmap_read_lock(mm);
 970         return error;
 971 }
 972
 973 /*
 974  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
 975  * will handle splitting a vm area into separate areas, each area with its own
 976  * behavior.
 977  */
 978 static int madvise_vma_behavior(struct vm_area_struct *vma,
 979                                 struct vm_area_struct **prev,
 980                                 unsigned long start, unsigned long end,
 981                                 unsigned long behavior)
 982 {
 983         int error;
 984         struct anon_vma_name *anon_name;
 985         unsigned long new_flags = vma->vm_flags;
 986
 987         switch (behavior) {
 988         case MADV_REMOVE:
 989                 return madvise_remove(vma, prev, start, end);
 990         case MADV_WILLNEED:
 991                 return madvise_willneed(vma, prev, start, end);
 992         case MADV_COLD:
 993                 return madvise_cold(vma, prev, start, end);
 994         case MADV_PAGEOUT:
 995                 return madvise_pageout(vma, prev, start, end);
 996         case MADV_FREE:
 997         case MADV_DONTNEED:
 998         case MADV_DONTNEED_LOCKED:
 999                 return madvise_dontneed_free(vma, prev, start, end, behavior);
1000         case MADV_POPULATE_READ:
1001         case MADV_POPULATE_WRITE:
1002                 return madvise_populate(vma, prev, start, end, behavior);
1003         case MADV_NORMAL:
1004                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1005                 break;
1006         case MADV_SEQUENTIAL:
1007                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1008                 break;
1009         case MADV_RANDOM:
1010                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1011                 break;
1012         case MADV_DONTFORK:
1013                 new_flags |= VM_DONTCOPY;
1014                 break;
1015         case MADV_DOFORK:
1016                 if (vma->vm_flags & VM_IO)
1017                         return -EINVAL;
1018                 new_flags &= ~VM_DONTCOPY;
1019                 break;
1020         case MADV_WIPEONFORK:
1021                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1022                 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1023                         return -EINVAL;
1024                 new_flags |= VM_WIPEONFORK;
1025                 break;
1026         case MADV_KEEPONFORK:
1027                 new_flags &= ~VM_WIPEONFORK;
1028                 break;
1029         case MADV_DONTDUMP:
1030                 new_flags |= VM_DONTDUMP;
1031                 break;
1032         case MADV_DODUMP:
1033                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1034                         return -EINVAL;
1035                 new_flags &= ~VM_DONTDUMP;
1036                 break;
1037         case MADV_MERGEABLE:
1038         case MADV_UNMERGEABLE:
1039                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1040                 if (error)
1041                         goto out;
1042                 break;
1043         case MADV_HUGEPAGE:
1044         case MADV_NOHUGEPAGE:
1045                 error = hugepage_madvise(vma, &new_flags, behavior);
1046                 if (error)
1047                         goto out;
1048                 break;
1049         }
1050
1051         anon_name = anon_vma_name(vma);
1052         anon_vma_name_get(anon_name);
1053         error = madvise_update_vma(vma, prev, start, end, new_flags,
1054                                    anon_name);
1055         anon_vma_name_put(anon_name);
1056
1057 out:
1058         /*
1059          * madvise() returns EAGAIN if kernel resources, such as
1060          * slab, are temporarily unavailable.
1061          */
1062         if (error == -ENOMEM)
1063                 error = -EAGAIN;
1064         return error;
1065 }
1066
1067 #ifdef CONFIG_MEMORY_FAILURE
1068 /*
1069  * Error injection support for memory error handling.
1070  */
1071 static int madvise_inject_error(int behavior,
1072                 unsigned long start, unsigned long end)
1073 {
1074         unsigned long size;
1075
1076         if (!capable(CAP_SYS_ADMIN))
1077                 return -EPERM;
1078
1079
1080         for (; start < end; start += size) {
1081                 unsigned long pfn;
1082                 struct page *page;
1083                 int ret;
1084
1085                 ret = get_user_pages_fast(start, 1, 0, &page);
1086                 if (ret != 1)
1087                         return ret;
1088                 pfn = page_to_pfn(page);
1089
1090                 /*
1091                  * When soft offlining hugepages, after migrating the page
1092                  * we dissolve it, therefore in the second loop "page" will
1093                  * no longer be a compound page.
1094                  */
1095                 size = page_size(compound_head(page));
1096
1097                 if (behavior == MADV_SOFT_OFFLINE) {
1098                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1099                                  pfn, start);
1100                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1101                 } else {
1102                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1103                                  pfn, start);
1104                         ret = memory_failure(pfn, MF_COUNT_INCREASED);
1105                         if (ret == -EOPNOTSUPP)
1106                                 ret = 0;
1107                 }
1108
1109                 if (ret)
1110                         return ret;
1111         }
1112
1113         return 0;
1114 }
1115 #endif
1116
1117 static bool
1118 madvise_behavior_valid(int behavior)
1119 {
1120         switch (behavior) {
1121         case MADV_DOFORK:
1122         case MADV_DONTFORK:
1123         case MADV_NORMAL:
1124         case MADV_SEQUENTIAL:
1125         case MADV_RANDOM:
1126         case MADV_REMOVE:
1127         case MADV_WILLNEED:
1128         case MADV_DONTNEED:
1129         case MADV_DONTNEED_LOCKED:
1130         case MADV_FREE:
1131         case MADV_COLD:
1132         case MADV_PAGEOUT:
1133         case MADV_POPULATE_READ:
1134         case MADV_POPULATE_WRITE:
1135 #ifdef CONFIG_KSM
1136         case MADV_MERGEABLE:
1137         case MADV_UNMERGEABLE:
1138 #endif
1139 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1140         case MADV_HUGEPAGE:
1141         case MADV_NOHUGEPAGE:
1142 #endif
1143         case MADV_DONTDUMP:
1144         case MADV_DODUMP:
1145         case MADV_WIPEONFORK:
1146         case MADV_KEEPONFORK:
1147 #ifdef CONFIG_MEMORY_FAILURE
1148         case MADV_SOFT_OFFLINE:
1149         case MADV_HWPOISON:
1150 #endif
1151                 return true;
1152
1153         default:
1154                 return false;
1155         }
1156 }
1157
1158 static bool
1159 process_madvise_behavior_valid(int behavior)
1160 {
1161         switch (behavior) {
1162         case MADV_COLD:
1163         case MADV_PAGEOUT:
1164         case MADV_WILLNEED:
1165                 return true;
1166         default:
1167                 return false;
1168         }
1169 }
1170
1171 /*
1172  * Walk the vmas in range [start,end), and call the visit function on each one.
1173  * The visit function will get start and end parameters that cover the overlap
1174  * between the current vma and the original range.  Any unmapped regions in the
1175  * original range will result in this function returning -ENOMEM while still
1176  * calling the visit function on all of the existing vmas in the range.
1177  * Must be called with the mmap_lock held for reading or writing.
1178  */
1179 static
1180 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1181                       unsigned long end, unsigned long arg,
1182                       int (*visit)(struct vm_area_struct *vma,
1183                                    struct vm_area_struct **prev, unsigned long start,
1184                                    unsigned long end, unsigned long arg))
1185 {
1186         struct vm_area_struct *vma;
1187         struct vm_area_struct *prev;
1188         unsigned long tmp;
1189         int unmapped_error = 0;
1190
1191         /*
1192          * If the interval [start,end) covers some unmapped address
1193          * ranges, just ignore them, but return -ENOMEM at the end.
1194          * - different from the way of handling in mlock etc.
1195          */
1196         vma = find_vma_prev(mm, start, &prev);
1197         if (vma && start > vma->vm_start)
1198                 prev = vma;
1199
1200         for (;;) {
1201                 int error;
1202
1203                 /* Still start < end. */
1204                 if (!vma)
1205                         return -ENOMEM;
1206
1207                 /* Here start < (end|vma->vm_end). */
1208                 if (start < vma->vm_start) {
1209                         unmapped_error = -ENOMEM;
1210                         start = vma->vm_start;
1211                         if (start >= end)
1212                                 break;
1213                 }
1214
1215                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1216                 tmp = vma->vm_end;
1217                 if (end < tmp)
1218                         tmp = end;
1219
1220                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1221                 error = visit(vma, &prev, start, tmp, arg);
1222                 if (error)
1223                         return error;
1224                 start = tmp;
1225                 if (prev && start < prev->vm_end)
1226                         start = prev->vm_end;
1227                 if (start >= end)
1228                         break;
1229                 if (prev)
1230                         vma = prev->vm_next;
1231                 else    /* madvise_remove dropped mmap_lock */
1232                         vma = find_vma(mm, start);
1233         }
1234
1235         return unmapped_error;
1236 }
1237
1238 #ifdef CONFIG_ANON_VMA_NAME
1239 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1240                                  struct vm_area_struct **prev,
1241                                  unsigned long start, unsigned long end,
1242                                  unsigned long anon_name)
1243 {
1244         int error;
1245
1246         /* Only anonymous mappings can be named */
1247         if (vma->vm_file)
1248                 return -EBADF;
1249
1250         error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1251                                    (struct anon_vma_name *)anon_name);
1252
1253         /*
1254          * madvise() returns EAGAIN if kernel resources, such as
1255          * slab, are temporarily unavailable.
1256          */
1257         if (error == -ENOMEM)
1258                 error = -EAGAIN;
1259         return error;
1260 }
1261
1262 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1263                           unsigned long len_in, struct anon_vma_name *anon_name)
1264 {
1265         unsigned long end;
1266         unsigned long len;
1267
1268         if (start & ~PAGE_MASK)
1269                 return -EINVAL;
1270         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1271
1272         /* Check to see whether len was rounded up from small -ve to zero */
1273         if (len_in && !len)
1274                 return -EINVAL;
1275
1276         end = start + len;
1277         if (end < start)
1278                 return -EINVAL;
1279
1280         if (end == start)
1281                 return 0;
1282
1283         return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1284                                  madvise_vma_anon_name);
1285 }
1286 #endif /* CONFIG_ANON_VMA_NAME */
1287 /*
1288  * The madvise(2) system call.
1289  *
1290  * Applications can use madvise() to advise the kernel how it should
1291  * handle paging I/O in this VM area.  The idea is to help the kernel
1292  * use appropriate read-ahead and caching techniques.  The information
1293  * provided is advisory only, and can be safely disregarded by the
1294  * kernel without affecting the correct operation of the application.
1295  *
1296  * behavior values:
1297  *  MADV_NORMAL - the default behavior is to read clusters.  This
1298  *              results in some read-ahead and read-behind.
1299  *  MADV_RANDOM - the system should read the minimum amount of data
1300  *              on any access, since it is unlikely that the appli-
1301  *              cation will need more than what it asks for.
1302  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1303  *              once, so they can be aggressively read ahead, and
1304  *              can be freed soon after they are accessed.
1305  *  MADV_WILLNEED - the application is notifying the system to read
1306  *              some pages ahead.
1307  *  MADV_DONTNEED - the application is finished with the given range,
1308  *              so the kernel can free resources associated with it.
1309  *  MADV_FREE - the application marks pages in the given range as lazy free,
1310  *              where actual purges are postponed until memory pressure happens.
1311  *  MADV_REMOVE - the application wants to free up the given range of
1312  *              pages and associated backing store.
1313  *  MADV_DONTFORK - omit this area from child's address space when forking:
1314  *              typically, to avoid COWing pages pinned by get_user_pages().
1315  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1316  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1317  *              range after a fork.
1318  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1319  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1320  *              were corrupted by unrecoverable hardware memory failure.
1321  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1322  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1323  *              this area with pages of identical content from other such areas.
1324  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1325  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1326  *              huge pages in the future. Existing pages might be coalesced and
1327  *              new pages might be allocated as THP.
1328  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1329  *              transparent huge pages so the existing pages will not be
1330  *              coalesced into THP and new pages will not be allocated as THP.
1331  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1332  *              from being included in its core dump.
1333  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1334  *  MADV_COLD - the application is not expected to use this memory soon,
1335  *              deactivate pages in this range so that they can be reclaimed
1336  *              easily if memory pressure happens.
1337  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1338  *              page out the pages in this range immediately.
1339  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1340  *              triggering read faults if required
1341  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1342  *              triggering write faults if required
1343  *
1344  * return values:
1345  *  zero    - success
1346  *  -EINVAL - start + len < 0, start is not page-aligned,
1347  *              "behavior" is not a valid value, or application
1348  *              is attempting to release locked or shared pages,
1349  *              or the specified address range includes file, Huge TLB,
1350  *              MAP_SHARED or VMPFNMAP range.
1351  *  -ENOMEM - addresses in the specified range are not currently
1352  *              mapped, or are outside the AS of the process.
1353  *  -EIO    - an I/O error occurred while paging in data.
1354  *  -EBADF  - map exists, but area maps something that isn't a file.
1355  *  -EAGAIN - a kernel resource was temporarily unavailable.
1356  */
1357 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1358 {
1359         unsigned long end;
1360         int error;
1361         int write;
1362         size_t len;
1363         struct blk_plug plug;
1364
1365         start = untagged_addr(start);
1366
1367         if (!madvise_behavior_valid(behavior))
1368                 return -EINVAL;
1369
1370         if (!PAGE_ALIGNED(start))
1371                 return -EINVAL;
1372         len = PAGE_ALIGN(len_in);
1373
1374         /* Check to see whether len was rounded up from small -ve to zero */
1375         if (len_in && !len)
1376                 return -EINVAL;
1377
1378         end = start + len;
1379         if (end < start)
1380                 return -EINVAL;
1381
1382         if (end == start)
1383                 return 0;
1384
1385 #ifdef CONFIG_MEMORY_FAILURE
1386         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1387                 return madvise_inject_error(behavior, start, start + len_in);
1388 #endif
1389
1390         write = madvise_need_mmap_write(behavior);
1391         if (write) {
1392                 if (mmap_write_lock_killable(mm))
1393                         return -EINTR;
1394         } else {
1395                 mmap_read_lock(mm);
1396         }
1397
1398         blk_start_plug(&plug);
1399         error = madvise_walk_vmas(mm, start, end, behavior,
1400                         madvise_vma_behavior);
1401         blk_finish_plug(&plug);
1402         if (write)
1403                 mmap_write_unlock(mm);
1404         else
1405                 mmap_read_unlock(mm);
1406
1407         return error;
1408 }
1409
1410 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1411 {
1412         return do_madvise(current->mm, start, len_in, behavior);
1413 }
1414
1415 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1416                 size_t, vlen, int, behavior, unsigned int, flags)
1417 {
1418         ssize_t ret;
1419         struct iovec iovstack[UIO_FASTIOV], iovec;
1420         struct iovec *iov = iovstack;
1421         struct iov_iter iter;
1422         struct task_struct *task;
1423         struct mm_struct *mm;
1424         size_t total_len;
1425         unsigned int f_flags;
1426
1427         if (flags != 0) {
1428                 ret = -EINVAL;
1429                 goto out;
1430         }
1431
1432         ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1433         if (ret < 0)
1434                 goto out;
1435
1436         task = pidfd_get_task(pidfd, &f_flags);
1437         if (IS_ERR(task)) {
1438                 ret = PTR_ERR(task);
1439                 goto free_iov;
1440         }
1441
1442         if (!process_madvise_behavior_valid(behavior)) {
1443                 ret = -EINVAL;
1444                 goto release_task;
1445         }
1446
1447         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1448         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1449         if (IS_ERR_OR_NULL(mm)) {
1450                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1451                 goto release_task;
1452         }
1453
1454         /*
1455          * Require CAP_SYS_NICE for influencing process performance. Note that
1456          * only non-destructive hints are currently supported.
1457          */
1458         if (!capable(CAP_SYS_NICE)) {
1459                 ret = -EPERM;
1460                 goto release_mm;
1461         }
1462
1463         total_len = iov_iter_count(&iter);
1464
1465         while (iov_iter_count(&iter)) {
1466                 iovec = iov_iter_iovec(&iter);
1467                 /*
1468                  * do_madvise returns ENOMEM if unmapped holes are present
1469                  * in the passed VMA. process_madvise() is expected to skip
1470                  * unmapped holes passed to it in the 'struct iovec' list
1471                  * and not fail because of them. Thus treat -ENOMEM return
1472                  * from do_madvise as valid and continue processing.
1473                  */
1474                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1475                                         iovec.iov_len, behavior);
1476                 if (ret < 0 && ret != -ENOMEM)
1477                         break;
1478                 iov_iter_advance(&iter, iovec.iov_len);
1479         }
1480
1481         ret = (total_len - iov_iter_count(&iter)) ? : ret;
1482
1483 release_mm:
1484         mmput(mm);
1485 release_task:
1486         put_task_struct(task);
1487 free_iov:
1488         kfree(iov);
1489 out:
1490         return ret;
1491 }