mm/madvise.c

   1 /*
   2  *      linux/mm/madvise.c
   3  *
   4  * Copyright (C) 1999  Linus Torvalds
   5  * Copyright (C) 2002  Christoph Hellwig
   6  */
   7
   8 #include <linux/mman.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/syscalls.h>
  11 #include <linux/mempolicy.h>
  12 #include <linux/page-isolation.h>
  13 #include <linux/userfaultfd_k.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/falloc.h>
  16 #include <linux/sched.h>
  17 #include <linux/ksm.h>
  18 #include <linux/fs.h>
  19 #include <linux/file.h>
  20 #include <linux/blkdev.h>
  21 #include <linux/backing-dev.h>
  22 #include <linux/swap.h>
  23 #include <linux/swapops.h>
  24 #include <linux/mmu_notifier.h>
  25
  26 #include <asm/tlb.h>
  27
  28 /*
  29  * Any behaviour which results in changes to the vma->vm_flags needs to
  30  * take mmap_sem for writing. Others, which simply traverse vmas, need
  31  * to only take it for reading.
  32  */
  33 static int madvise_need_mmap_write(int behavior)
  34 {
  35         switch (behavior) {
  36         case MADV_REMOVE:
  37         case MADV_WILLNEED:
  38         case MADV_DONTNEED:
  39         case MADV_FREE:
  40                 return 0;
  41         default:
  42                 /* be safe, default to 1. list exceptions explicitly */
  43                 return 1;
  44         }
  45 }
  46
  47 /*
  48  * We can potentially split a vm area into separate
  49  * areas, each area with its own behavior.
  50  */
  51 static long madvise_behavior(struct vm_area_struct *vma,
  52                      struct vm_area_struct **prev,
  53                      unsigned long start, unsigned long end, int behavior)
  54 {
  55         struct mm_struct *mm = vma->vm_mm;
  56         int error = 0;
  57         pgoff_t pgoff;
  58         unsigned long new_flags = vma->vm_flags;
  59
  60         switch (behavior) {
  61         case MADV_NORMAL:
  62                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  63                 break;
  64         case MADV_SEQUENTIAL:
  65                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  66                 break;
  67         case MADV_RANDOM:
  68                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  69                 break;
  70         case MADV_DONTFORK:
  71                 new_flags |= VM_DONTCOPY;
  72                 break;
  73         case MADV_DOFORK:
  74                 if (vma->vm_flags & VM_IO) {
  75                         error = -EINVAL;
  76                         goto out;
  77                 }
  78                 new_flags &= ~VM_DONTCOPY;
  79                 break;
  80         case MADV_DONTDUMP:
  81                 new_flags |= VM_DONTDUMP;
  82                 break;
  83         case MADV_DODUMP:
  84                 if (new_flags & VM_SPECIAL) {
  85                         error = -EINVAL;
  86                         goto out;
  87                 }
  88                 new_flags &= ~VM_DONTDUMP;
  89                 break;
  90         case MADV_MERGEABLE:
  91         case MADV_UNMERGEABLE:
  92                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  93                 if (error)
  94                         goto out;
  95                 break;
  96         case MADV_HUGEPAGE:
  97         case MADV_NOHUGEPAGE:
  98                 error = hugepage_madvise(vma, &new_flags, behavior);
  99                 if (error)
 100                         goto out;
 101                 break;
 102         }
 103
 104         if (new_flags == vma->vm_flags) {
 105                 *prev = vma;
 106                 goto out;
 107         }
 108
 109         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 110         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 111                           vma->vm_file, pgoff, vma_policy(vma),
 112                           vma->vm_userfaultfd_ctx);
 113         if (*prev) {
 114                 vma = *prev;
 115                 goto success;
 116         }
 117
 118         *prev = vma;
 119
 120         if (start != vma->vm_start) {
 121                 error = split_vma(mm, vma, start, 1);
 122                 if (error)
 123                         goto out;
 124         }
 125
 126         if (end != vma->vm_end) {
 127                 error = split_vma(mm, vma, end, 0);
 128                 if (error)
 129                         goto out;
 130         }
 131
 132 success:
 133         /*
 134          * vm_flags is protected by the mmap_sem held in write mode.
 135          */
 136         vma->vm_flags = new_flags;
 137
 138 out:
 139         if (error == -ENOMEM)
 140                 error = -EAGAIN;
 141         return error;
 142 }
 143
 144 #ifdef CONFIG_SWAP
 145 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 146         unsigned long end, struct mm_walk *walk)
 147 {
 148         pte_t *orig_pte;
 149         struct vm_area_struct *vma = walk->private;
 150         unsigned long index;
 151
 152         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 153                 return 0;
 154
 155         for (index = start; index != end; index += PAGE_SIZE) {
 156                 pte_t pte;
 157                 swp_entry_t entry;
 158                 struct page *page;
 159                 spinlock_t *ptl;
 160
 161                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 162                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 163                 pte_unmap_unlock(orig_pte, ptl);
 164
 165                 if (pte_present(pte) || pte_none(pte))
 166                         continue;
 167                 entry = pte_to_swp_entry(pte);
 168                 if (unlikely(non_swap_entry(entry)))
 169                         continue;
 170
 171                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 172                                                                 vma, index);
 173                 if (page)
 174                         put_page(page);
 175         }
 176
 177         return 0;
 178 }
 179
 180 static void force_swapin_readahead(struct vm_area_struct *vma,
 181                 unsigned long start, unsigned long end)
 182 {
 183         struct mm_walk walk = {
 184                 .mm = vma->vm_mm,
 185                 .pmd_entry = swapin_walk_pmd_entry,
 186                 .private = vma,
 187         };
 188
 189         walk_page_range(start, end, &walk);
 190
 191         lru_add_drain();        /* Push any new pages onto the LRU now */
 192 }
 193
 194 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 195                 unsigned long start, unsigned long end,
 196                 struct address_space *mapping)
 197 {
 198         pgoff_t index;
 199         struct page *page;
 200         swp_entry_t swap;
 201
 202         for (; start < end; start += PAGE_SIZE) {
 203                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 204
 205                 page = find_get_entry(mapping, index);
 206                 if (!radix_tree_exceptional_entry(page)) {
 207                         if (page)
 208                                 put_page(page);
 209                         continue;
 210                 }
 211                 swap = radix_to_swp_entry(page);
 212                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 213                                                                 NULL, 0);
 214                 if (page)
 215                         put_page(page);
 216         }
 217
 218         lru_add_drain();        /* Push any new pages onto the LRU now */
 219 }
 220 #endif          /* CONFIG_SWAP */
 221
 222 /*
 223  * Schedule all required I/O operations.  Do not wait for completion.
 224  */
 225 static long madvise_willneed(struct vm_area_struct *vma,
 226                              struct vm_area_struct **prev,
 227                              unsigned long start, unsigned long end)
 228 {
 229         struct file *file = vma->vm_file;
 230
 231 #ifdef CONFIG_SWAP
 232         if (!file) {
 233                 *prev = vma;
 234                 force_swapin_readahead(vma, start, end);
 235                 return 0;
 236         }
 237
 238         if (shmem_mapping(file->f_mapping)) {
 239                 *prev = vma;
 240                 force_shm_swapin_readahead(vma, start, end,
 241                                         file->f_mapping);
 242                 return 0;
 243         }
 244 #else
 245         if (!file)
 246                 return -EBADF;
 247 #endif
 248
 249         if (IS_DAX(file_inode(file))) {
 250                 /* no bad return value, but ignore advice */
 251                 return 0;
 252         }
 253
 254         *prev = vma;
 255         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 256         if (end > vma->vm_end)
 257                 end = vma->vm_end;
 258         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 259
 260         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 261         return 0;
 262 }
 263
 264 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 265                                 unsigned long end, struct mm_walk *walk)
 266
 267 {
 268         struct mmu_gather *tlb = walk->private;
 269         struct mm_struct *mm = tlb->mm;
 270         struct vm_area_struct *vma = walk->vma;
 271         spinlock_t *ptl;
 272         pte_t *orig_pte, *pte, ptent;
 273         struct page *page;
 274         int nr_swap = 0;
 275         unsigned long next;
 276
 277         next = pmd_addr_end(addr, end);
 278         if (pmd_trans_huge(*pmd))
 279                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 280                         goto next;
 281
 282         if (pmd_trans_unstable(pmd))
 283                 return 0;
 284
 285         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
 286         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 287         arch_enter_lazy_mmu_mode();
 288         for (; addr != end; pte++, addr += PAGE_SIZE) {
 289                 ptent = *pte;
 290
 291                 if (pte_none(ptent))
 292                         continue;
 293                 /*
 294                  * If the pte has swp_entry, just clear page table to
 295                  * prevent swap-in which is more expensive rather than
 296                  * (page allocation + zeroing).
 297                  */
 298                 if (!pte_present(ptent)) {
 299                         swp_entry_t entry;
 300
 301                         entry = pte_to_swp_entry(ptent);
 302                         if (non_swap_entry(entry))
 303                                 continue;
 304                         nr_swap--;
 305                         free_swap_and_cache(entry);
 306                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 307                         continue;
 308                 }
 309
 310                 page = vm_normal_page(vma, addr, ptent);
 311                 if (!page)
 312                         continue;
 313
 314                 /*
 315                  * If pmd isn't transhuge but the page is THP and
 316                  * is owned by only this process, split it and
 317                  * deactivate all pages.
 318                  */
 319                 if (PageTransCompound(page)) {
 320                         if (page_mapcount(page) != 1)
 321                                 goto out;
 322                         get_page(page);
 323                         if (!trylock_page(page)) {
 324                                 put_page(page);
 325                                 goto out;
 326                         }
 327                         pte_unmap_unlock(orig_pte, ptl);
 328                         if (split_huge_page(page)) {
 329                                 unlock_page(page);
 330                                 put_page(page);
 331                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 332                                 goto out;
 333                         }
 334                         put_page(page);
 335                         unlock_page(page);
 336                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 337                         pte--;
 338                         addr -= PAGE_SIZE;
 339                         continue;
 340                 }
 341
 342                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 343
 344                 if (PageSwapCache(page) || PageDirty(page)) {
 345                         if (!trylock_page(page))
 346                                 continue;
 347                         /*
 348                          * If page is shared with others, we couldn't clear
 349                          * PG_dirty of the page.
 350                          */
 351                         if (page_mapcount(page) != 1) {
 352                                 unlock_page(page);
 353                                 continue;
 354                         }
 355
 356                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 357                                 unlock_page(page);
 358                                 continue;
 359                         }
 360
 361                         ClearPageDirty(page);
 362                         unlock_page(page);
 363                 }
 364
 365                 if (pte_young(ptent) || pte_dirty(ptent)) {
 366                         /*
 367                          * Some of architecture(ex, PPC) don't update TLB
 368                          * with set_pte_at and tlb_remove_tlb_entry so for
 369                          * the portability, remap the pte with old|clean
 370                          * after pte clearing.
 371                          */
 372                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 373                                                         tlb->fullmm);
 374
 375                         ptent = pte_mkold(ptent);
 376                         ptent = pte_mkclean(ptent);
 377                         set_pte_at(mm, addr, pte, ptent);
 378                         if (PageActive(page))
 379                                 deactivate_page(page);
 380                         tlb_remove_tlb_entry(tlb, pte, addr);
 381                 }
 382         }
 383 out:
 384         if (nr_swap) {
 385                 if (current->mm == mm)
 386                         sync_mm_rss(mm);
 387
 388                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 389         }
 390         arch_leave_lazy_mmu_mode();
 391         pte_unmap_unlock(orig_pte, ptl);
 392         cond_resched();
 393 next:
 394         return 0;
 395 }
 396
 397 static void madvise_free_page_range(struct mmu_gather *tlb,
 398                              struct vm_area_struct *vma,
 399                              unsigned long addr, unsigned long end)
 400 {
 401         struct mm_walk free_walk = {
 402                 .pmd_entry = madvise_free_pte_range,
 403                 .mm = vma->vm_mm,
 404                 .private = tlb,
 405         };
 406
 407         tlb_start_vma(tlb, vma);
 408         walk_page_range(addr, end, &free_walk);
 409         tlb_end_vma(tlb, vma);
 410 }
 411
 412 static int madvise_free_single_vma(struct vm_area_struct *vma,
 413                         unsigned long start_addr, unsigned long end_addr)
 414 {
 415         unsigned long start, end;
 416         struct mm_struct *mm = vma->vm_mm;
 417         struct mmu_gather tlb;
 418
 419         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 420                 return -EINVAL;
 421
 422         /* MADV_FREE works for only anon vma at the moment */
 423         if (!vma_is_anonymous(vma))
 424                 return -EINVAL;
 425
 426         start = max(vma->vm_start, start_addr);
 427         if (start >= vma->vm_end)
 428                 return -EINVAL;
 429         end = min(vma->vm_end, end_addr);
 430         if (end <= vma->vm_start)
 431                 return -EINVAL;
 432
 433         lru_add_drain();
 434         tlb_gather_mmu(&tlb, mm, start, end);
 435         update_hiwater_rss(mm);
 436
 437         mmu_notifier_invalidate_range_start(mm, start, end);
 438         madvise_free_page_range(&tlb, vma, start, end);
 439         mmu_notifier_invalidate_range_end(mm, start, end);
 440         tlb_finish_mmu(&tlb, start, end);
 441
 442         return 0;
 443 }
 444
 445 static long madvise_free(struct vm_area_struct *vma,
 446                              struct vm_area_struct **prev,
 447                              unsigned long start, unsigned long end)
 448 {
 449         *prev = vma;
 450         return madvise_free_single_vma(vma, start, end);
 451 }
 452
 453 /*
 454  * Application no longer needs these pages.  If the pages are dirty,
 455  * it's OK to just throw them away.  The app will be more careful about
 456  * data it wants to keep.  Be sure to free swap resources too.  The
 457  * zap_page_range call sets things up for shrink_active_list to actually free
 458  * these pages later if no one else has touched them in the meantime,
 459  * although we could add these pages to a global reuse list for
 460  * shrink_active_list to pick up before reclaiming other pages.
 461  *
 462  * NB: This interface discards data rather than pushes it out to swap,
 463  * as some implementations do.  This has performance implications for
 464  * applications like large transactional databases which want to discard
 465  * pages in anonymous maps after committing to backing store the data
 466  * that was kept in them.  There is no reason to write this data out to
 467  * the swap area if the application is discarding it.
 468  *
 469  * An interface that causes the system to free clean pages and flush
 470  * dirty pages is already available as msync(MS_INVALIDATE).
 471  */
 472 static long madvise_dontneed(struct vm_area_struct *vma,
 473                              struct vm_area_struct **prev,
 474                              unsigned long start, unsigned long end)
 475 {
 476         *prev = vma;
 477         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 478                 return -EINVAL;
 479
 480         madvise_userfault_dontneed(vma, prev, start, end);
 481         zap_page_range(vma, start, end - start, NULL);
 482         return 0;
 483 }
 484
 485 /*
 486  * Application wants to free up the pages and associated backing store.
 487  * This is effectively punching a hole into the middle of a file.
 488  */
 489 static long madvise_remove(struct vm_area_struct *vma,
 490                                 struct vm_area_struct **prev,
 491                                 unsigned long start, unsigned long end)
 492 {
 493         loff_t offset;
 494         int error;
 495         struct file *f;
 496
 497         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 498
 499         if (vma->vm_flags & VM_LOCKED)
 500                 return -EINVAL;
 501
 502         f = vma->vm_file;
 503
 504         if (!f || !f->f_mapping || !f->f_mapping->host) {
 505                         return -EINVAL;
 506         }
 507
 508         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 509                 return -EACCES;
 510
 511         offset = (loff_t)(start - vma->vm_start)
 512                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 513
 514         /*
 515          * Filesystem's fallocate may need to take i_mutex.  We need to
 516          * explicitly grab a reference because the vma (and hence the
 517          * vma's reference to the file) can go away as soon as we drop
 518          * mmap_sem.
 519          */
 520         get_file(f);
 521         up_read(&current->mm->mmap_sem);
 522         error = vfs_fallocate(f,
 523                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 524                                 offset, end - start);
 525         fput(f);
 526         down_read(&current->mm->mmap_sem);
 527         return error;
 528 }
 529
 530 #ifdef CONFIG_MEMORY_FAILURE
 531 /*
 532  * Error injection support for memory error handling.
 533  */
 534 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 535 {
 536         struct page *p;
 537         if (!capable(CAP_SYS_ADMIN))
 538                 return -EPERM;
 539         for (; start < end; start += PAGE_SIZE <<
 540                                 compound_order(compound_head(p))) {
 541                 int ret;
 542
 543                 ret = get_user_pages_fast(start, 1, 0, &p);
 544                 if (ret != 1)
 545                         return ret;
 546
 547                 if (PageHWPoison(p)) {
 548                         put_page(p);
 549                         continue;
 550                 }
 551                 if (bhv == MADV_SOFT_OFFLINE) {
 552                         pr_info("Soft offlining page %#lx at %#lx\n",
 553                                 page_to_pfn(p), start);
 554                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 555                         if (ret)
 556                                 return ret;
 557                         continue;
 558                 }
 559                 pr_info("Injecting memory failure for page %#lx at %#lx\n",
 560                        page_to_pfn(p), start);
 561                 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 562                 if (ret)
 563                         return ret;
 564         }
 565         return 0;
 566 }
 567 #endif
 568
 569 static long
 570 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 571                 unsigned long start, unsigned long end, int behavior)
 572 {
 573         switch (behavior) {
 574         case MADV_REMOVE:
 575                 return madvise_remove(vma, prev, start, end);
 576         case MADV_WILLNEED:
 577                 return madvise_willneed(vma, prev, start, end);
 578         case MADV_FREE:
 579                 /*
 580                  * XXX: In this implementation, MADV_FREE works like
 581                  * MADV_DONTNEED on swapless system or full swap.
 582                  */
 583                 if (get_nr_swap_pages() > 0)
 584                         return madvise_free(vma, prev, start, end);
 585                 /* passthrough */
 586         case MADV_DONTNEED:
 587                 return madvise_dontneed(vma, prev, start, end);
 588         default:
 589                 return madvise_behavior(vma, prev, start, end, behavior);
 590         }
 591 }
 592
 593 static bool
 594 madvise_behavior_valid(int behavior)
 595 {
 596         switch (behavior) {
 597         case MADV_DOFORK:
 598         case MADV_DONTFORK:
 599         case MADV_NORMAL:
 600         case MADV_SEQUENTIAL:
 601         case MADV_RANDOM:
 602         case MADV_REMOVE:
 603         case MADV_WILLNEED:
 604         case MADV_DONTNEED:
 605         case MADV_FREE:
 606 #ifdef CONFIG_KSM
 607         case MADV_MERGEABLE:
 608         case MADV_UNMERGEABLE:
 609 #endif
 610 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 611         case MADV_HUGEPAGE:
 612         case MADV_NOHUGEPAGE:
 613 #endif
 614         case MADV_DONTDUMP:
 615         case MADV_DODUMP:
 616                 return true;
 617
 618         default:
 619                 return false;
 620         }
 621 }
 622
 623 /*
 624  * The madvise(2) system call.
 625  *
 626  * Applications can use madvise() to advise the kernel how it should
 627  * handle paging I/O in this VM area.  The idea is to help the kernel
 628  * use appropriate read-ahead and caching techniques.  The information
 629  * provided is advisory only, and can be safely disregarded by the
 630  * kernel without affecting the correct operation of the application.
 631  *
 632  * behavior values:
 633  *  MADV_NORMAL - the default behavior is to read clusters.  This
 634  *              results in some read-ahead and read-behind.
 635  *  MADV_RANDOM - the system should read the minimum amount of data
 636  *              on any access, since it is unlikely that the appli-
 637  *              cation will need more than what it asks for.
 638  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 639  *              once, so they can be aggressively read ahead, and
 640  *              can be freed soon after they are accessed.
 641  *  MADV_WILLNEED - the application is notifying the system to read
 642  *              some pages ahead.
 643  *  MADV_DONTNEED - the application is finished with the given range,
 644  *              so the kernel can free resources associated with it.
 645  *  MADV_FREE - the application marks pages in the given range as lazy free,
 646  *              where actual purges are postponed until memory pressure happens.
 647  *  MADV_REMOVE - the application wants to free up the given range of
 648  *              pages and associated backing store.
 649  *  MADV_DONTFORK - omit this area from child's address space when forking:
 650  *              typically, to avoid COWing pages pinned by get_user_pages().
 651  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 652  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 653  *              were corrupted by unrecoverable hardware memory failure.
 654  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 655  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 656  *              this area with pages of identical content from other such areas.
 657  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 658  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 659  *              huge pages in the future. Existing pages might be coalesced and
 660  *              new pages might be allocated as THP.
 661  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 662  *              transparent huge pages so the existing pages will not be
 663  *              coalesced into THP and new pages will not be allocated as THP.
 664  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 665  *              from being included in its core dump.
 666  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 667  *
 668  * return values:
 669  *  zero    - success
 670  *  -EINVAL - start + len < 0, start is not page-aligned,
 671  *              "behavior" is not a valid value, or application
 672  *              is attempting to release locked or shared pages.
 673  *  -ENOMEM - addresses in the specified range are not currently
 674  *              mapped, or are outside the AS of the process.
 675  *  -EIO    - an I/O error occurred while paging in data.
 676  *  -EBADF  - map exists, but area maps something that isn't a file.
 677  *  -EAGAIN - a kernel resource was temporarily unavailable.
 678  */
 679 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 680 {
 681         unsigned long end, tmp;
 682         struct vm_area_struct *vma, *prev;
 683         int unmapped_error = 0;
 684         int error = -EINVAL;
 685         int write;
 686         size_t len;
 687         struct blk_plug plug;
 688
 689 #ifdef CONFIG_MEMORY_FAILURE
 690         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 691                 return madvise_hwpoison(behavior, start, start+len_in);
 692 #endif
 693         if (!madvise_behavior_valid(behavior))
 694                 return error;
 695
 696         if (start & ~PAGE_MASK)
 697                 return error;
 698         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 699
 700         /* Check to see whether len was rounded up from small -ve to zero */
 701         if (len_in && !len)
 702                 return error;
 703
 704         end = start + len;
 705         if (end < start)
 706                 return error;
 707
 708         error = 0;
 709         if (end == start)
 710                 return error;
 711
 712         write = madvise_need_mmap_write(behavior);
 713         if (write) {
 714                 if (down_write_killable(&current->mm->mmap_sem))
 715                         return -EINTR;
 716         } else {
 717                 down_read(&current->mm->mmap_sem);
 718         }
 719
 720         /*
 721          * If the interval [start,end) covers some unmapped address
 722          * ranges, just ignore them, but return -ENOMEM at the end.
 723          * - different from the way of handling in mlock etc.
 724          */
 725         vma = find_vma_prev(current->mm, start, &prev);
 726         if (vma && start > vma->vm_start)
 727                 prev = vma;
 728
 729         blk_start_plug(&plug);
 730         for (;;) {
 731                 /* Still start < end. */
 732                 error = -ENOMEM;
 733                 if (!vma)
 734                         goto out;
 735
 736                 /* Here start < (end|vma->vm_end). */
 737                 if (start < vma->vm_start) {
 738                         unmapped_error = -ENOMEM;
 739                         start = vma->vm_start;
 740                         if (start >= end)
 741                                 goto out;
 742                 }
 743
 744                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 745                 tmp = vma->vm_end;
 746                 if (end < tmp)
 747                         tmp = end;
 748
 749                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 750                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 751                 if (error)
 752                         goto out;
 753                 start = tmp;
 754                 if (prev && start < prev->vm_end)
 755                         start = prev->vm_end;
 756                 error = unmapped_error;
 757                 if (start >= end)
 758                         goto out;
 759                 if (prev)
 760                         vma = prev->vm_next;
 761                 else    /* madvise_remove dropped mmap_sem */
 762                         vma = find_vma(current->mm, start);
 763         }
 764 out:
 765         blk_finish_plug(&plug);
 766         if (write)
 767                 up_write(&current->mm->mmap_sem);
 768         else
 769                 up_read(&current->mm->mmap_sem);
 770
 771         return error;
 772 }