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