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/hugetlb.h>
  14 #include <linux/falloc.h>
  15 #include <linux/sched.h>
  16 #include <linux/ksm.h>
  17 #include <linux/fs.h>
  18 #include <linux/file.h>
  19 #include <linux/blkdev.h>
  20 #include <linux/backing-dev.h>
  21 #include <linux/swap.h>
  22 #include <linux/swapops.h>
  23 #include <linux/mmu_notifier.h>
  24 #include "internal.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 (!is_vm_hugetlb_page(vma) && 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         *prev = vma;
 232 #ifdef CONFIG_SWAP
 233         if (!file) {
 234                 force_swapin_readahead(vma, start, end);
 235                 return 0;
 236         }
 237
 238         if (shmem_mapping(file->f_mapping)) {
 239                 force_shm_swapin_readahead(vma, start, end,
 240                                         file->f_mapping);
 241                 return 0;
 242         }
 243 #else
 244         if (!file)
 245                 return -EBADF;
 246 #endif
 247
 248         if (IS_DAX(file_inode(file))) {
 249                 /* no bad return value, but ignore advice */
 250                 return 0;
 251         }
 252
 253         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 254         if (end > vma->vm_end)
 255                 end = vma->vm_end;
 256         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 257
 258         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 259         return 0;
 260 }
 261
 262 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 263                                 unsigned long end, struct mm_walk *walk)
 264
 265 {
 266         struct mmu_gather *tlb = walk->private;
 267         struct mm_struct *mm = tlb->mm;
 268         struct vm_area_struct *vma = walk->vma;
 269         spinlock_t *ptl;
 270         pte_t *orig_pte, *pte, ptent;
 271         struct page *page;
 272         int nr_swap = 0;
 273         unsigned long next;
 274
 275         next = pmd_addr_end(addr, end);
 276         if (pmd_trans_huge(*pmd))
 277                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 278                         goto next;
 279
 280         if (pmd_trans_unstable(pmd))
 281                 return 0;
 282
 283         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 284         flush_tlb_batched_pending(mm);
 285         arch_enter_lazy_mmu_mode();
 286         for (; addr != end; pte++, addr += PAGE_SIZE) {
 287                 ptent = *pte;
 288
 289                 if (pte_none(ptent))
 290                         continue;
 291                 /*
 292                  * If the pte has swp_entry, just clear page table to
 293                  * prevent swap-in which is more expensive rather than
 294                  * (page allocation + zeroing).
 295                  */
 296                 if (!pte_present(ptent)) {
 297                         swp_entry_t entry;
 298
 299                         entry = pte_to_swp_entry(ptent);
 300                         if (non_swap_entry(entry))
 301                                 continue;
 302                         nr_swap--;
 303                         free_swap_and_cache(entry);
 304                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 305                         continue;
 306                 }
 307
 308                 page = vm_normal_page(vma, addr, ptent);
 309                 if (!page)
 310                         continue;
 311
 312                 /*
 313                  * If pmd isn't transhuge but the page is THP and
 314                  * is owned by only this process, split it and
 315                  * deactivate all pages.
 316                  */
 317                 if (PageTransCompound(page)) {
 318                         if (page_mapcount(page) != 1)
 319                                 goto out;
 320                         get_page(page);
 321                         if (!trylock_page(page)) {
 322                                 put_page(page);
 323                                 goto out;
 324                         }
 325                         pte_unmap_unlock(orig_pte, ptl);
 326                         if (split_huge_page(page)) {
 327                                 unlock_page(page);
 328                                 put_page(page);
 329                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 330                                 goto out;
 331                         }
 332                         unlock_page(page);
 333                         put_page(page);
 334                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 335                         pte--;
 336                         addr -= PAGE_SIZE;
 337                         continue;
 338                 }
 339
 340                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 341
 342                 if (PageSwapCache(page) || PageDirty(page)) {
 343                         if (!trylock_page(page))
 344                                 continue;
 345                         /*
 346                          * If page is shared with others, we couldn't clear
 347                          * PG_dirty of the page.
 348                          */
 349                         if (page_mapcount(page) != 1) {
 350                                 unlock_page(page);
 351                                 continue;
 352                         }
 353
 354                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 355                                 unlock_page(page);
 356                                 continue;
 357                         }
 358
 359                         ClearPageDirty(page);
 360                         unlock_page(page);
 361                 }
 362
 363                 if (pte_young(ptent) || pte_dirty(ptent)) {
 364                         /*
 365                          * Some of architecture(ex, PPC) don't update TLB
 366                          * with set_pte_at and tlb_remove_tlb_entry so for
 367                          * the portability, remap the pte with old|clean
 368                          * after pte clearing.
 369                          */
 370                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 371                                                         tlb->fullmm);
 372
 373                         ptent = pte_mkold(ptent);
 374                         ptent = pte_mkclean(ptent);
 375                         set_pte_at(mm, addr, pte, ptent);
 376                         if (PageActive(page))
 377                                 deactivate_page(page);
 378                         tlb_remove_tlb_entry(tlb, pte, addr);
 379                 }
 380         }
 381 out:
 382         if (nr_swap) {
 383                 if (current->mm == mm)
 384                         sync_mm_rss(mm);
 385
 386                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 387         }
 388         arch_leave_lazy_mmu_mode();
 389         pte_unmap_unlock(orig_pte, ptl);
 390         cond_resched();
 391 next:
 392         return 0;
 393 }
 394
 395 static void madvise_free_page_range(struct mmu_gather *tlb,
 396                              struct vm_area_struct *vma,
 397                              unsigned long addr, unsigned long end)
 398 {
 399         struct mm_walk free_walk = {
 400                 .pmd_entry = madvise_free_pte_range,
 401                 .mm = vma->vm_mm,
 402                 .private = tlb,
 403         };
 404
 405         tlb_start_vma(tlb, vma);
 406         walk_page_range(addr, end, &free_walk);
 407         tlb_end_vma(tlb, vma);
 408 }
 409
 410 static int madvise_free_single_vma(struct vm_area_struct *vma,
 411                         unsigned long start_addr, unsigned long end_addr)
 412 {
 413         unsigned long start, end;
 414         struct mm_struct *mm = vma->vm_mm;
 415         struct mmu_gather tlb;
 416
 417         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 418                 return -EINVAL;
 419
 420         /* MADV_FREE works for only anon vma at the moment */
 421         if (!vma_is_anonymous(vma))
 422                 return -EINVAL;
 423
 424         start = max(vma->vm_start, start_addr);
 425         if (start >= vma->vm_end)
 426                 return -EINVAL;
 427         end = min(vma->vm_end, end_addr);
 428         if (end <= vma->vm_start)
 429                 return -EINVAL;
 430
 431         lru_add_drain();
 432         tlb_gather_mmu(&tlb, mm, start, end);
 433         update_hiwater_rss(mm);
 434
 435         mmu_notifier_invalidate_range_start(mm, start, end);
 436         madvise_free_page_range(&tlb, vma, start, end);
 437         mmu_notifier_invalidate_range_end(mm, start, end);
 438         tlb_finish_mmu(&tlb, start, end);
 439
 440         return 0;
 441 }
 442
 443 static long madvise_free(struct vm_area_struct *vma,
 444                              struct vm_area_struct **prev,
 445                              unsigned long start, unsigned long end)
 446 {
 447         *prev = vma;
 448         return madvise_free_single_vma(vma, start, end);
 449 }
 450
 451 /*
 452  * Application no longer needs these pages.  If the pages are dirty,
 453  * it's OK to just throw them away.  The app will be more careful about
 454  * data it wants to keep.  Be sure to free swap resources too.  The
 455  * zap_page_range call sets things up for shrink_active_list to actually free
 456  * these pages later if no one else has touched them in the meantime,
 457  * although we could add these pages to a global reuse list for
 458  * shrink_active_list to pick up before reclaiming other pages.
 459  *
 460  * NB: This interface discards data rather than pushes it out to swap,
 461  * as some implementations do.  This has performance implications for
 462  * applications like large transactional databases which want to discard
 463  * pages in anonymous maps after committing to backing store the data
 464  * that was kept in them.  There is no reason to write this data out to
 465  * the swap area if the application is discarding it.
 466  *
 467  * An interface that causes the system to free clean pages and flush
 468  * dirty pages is already available as msync(MS_INVALIDATE).
 469  */
 470 static long madvise_dontneed(struct vm_area_struct *vma,
 471                              struct vm_area_struct **prev,
 472                              unsigned long start, unsigned long end)
 473 {
 474         *prev = vma;
 475         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 476                 return -EINVAL;
 477
 478         zap_page_range(vma, start, end - start, NULL);
 479         return 0;
 480 }
 481
 482 /*
 483  * Application wants to free up the pages and associated backing store.
 484  * This is effectively punching a hole into the middle of a file.
 485  */
 486 static long madvise_remove(struct vm_area_struct *vma,
 487                                 struct vm_area_struct **prev,
 488                                 unsigned long start, unsigned long end)
 489 {
 490         loff_t offset;
 491         int error;
 492         struct file *f;
 493
 494         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 495
 496         if (vma->vm_flags & VM_LOCKED)
 497                 return -EINVAL;
 498
 499         f = vma->vm_file;
 500
 501         if (!f || !f->f_mapping || !f->f_mapping->host) {
 502                         return -EINVAL;
 503         }
 504
 505         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 506                 return -EACCES;
 507
 508         offset = (loff_t)(start - vma->vm_start)
 509                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 510
 511         /*
 512          * Filesystem's fallocate may need to take i_mutex.  We need to
 513          * explicitly grab a reference because the vma (and hence the
 514          * vma's reference to the file) can go away as soon as we drop
 515          * mmap_sem.
 516          */
 517         get_file(f);
 518         up_read(&current->mm->mmap_sem);
 519         error = vfs_fallocate(f,
 520                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 521                                 offset, end - start);
 522         fput(f);
 523         down_read(&current->mm->mmap_sem);
 524         return error;
 525 }
 526
 527 #ifdef CONFIG_MEMORY_FAILURE
 528 /*
 529  * Error injection support for memory error handling.
 530  */
 531 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 532 {
 533         struct page *p;
 534         struct zone *zone;
 535
 536         if (!capable(CAP_SYS_ADMIN))
 537                 return -EPERM;
 538         for (; start < end; start += PAGE_SIZE <<
 539                                 compound_order(compound_head(p))) {
 540                 int ret;
 541
 542                 ret = get_user_pages_fast(start, 1, 0, &p);
 543                 if (ret != 1)
 544                         return ret;
 545
 546                 if (PageHWPoison(p)) {
 547                         put_page(p);
 548                         continue;
 549                 }
 550                 if (bhv == MADV_SOFT_OFFLINE) {
 551                         pr_info("Soft offlining page %#lx at %#lx\n",
 552                                 page_to_pfn(p), start);
 553                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 554                         if (ret)
 555                                 return ret;
 556                         continue;
 557                 }
 558                 pr_info("Injecting memory failure for page %#lx at %#lx\n",
 559                        page_to_pfn(p), start);
 560                 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 561                 if (ret)
 562                         return ret;
 563         }
 564
 565         /* Ensure that all poisoned pages are removed from per-cpu lists */
 566         for_each_populated_zone(zone)
 567                 drain_all_pages(zone);
 568
 569         return 0;
 570 }
 571 #endif
 572
 573 static long
 574 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 575                 unsigned long start, unsigned long end, int behavior)
 576 {
 577         switch (behavior) {
 578         case MADV_REMOVE:
 579                 return madvise_remove(vma, prev, start, end);
 580         case MADV_WILLNEED:
 581                 return madvise_willneed(vma, prev, start, end);
 582         case MADV_FREE:
 583                 /*
 584                  * XXX: In this implementation, MADV_FREE works like
 585                  * MADV_DONTNEED on swapless system or full swap.
 586                  */
 587                 if (get_nr_swap_pages() > 0)
 588                         return madvise_free(vma, prev, start, end);
 589                 /* passthrough */
 590         case MADV_DONTNEED:
 591                 return madvise_dontneed(vma, prev, start, end);
 592         default:
 593                 return madvise_behavior(vma, prev, start, end, behavior);
 594         }
 595 }
 596
 597 static bool
 598 madvise_behavior_valid(int behavior)
 599 {
 600         switch (behavior) {
 601         case MADV_DOFORK:
 602         case MADV_DONTFORK:
 603         case MADV_NORMAL:
 604         case MADV_SEQUENTIAL:
 605         case MADV_RANDOM:
 606         case MADV_REMOVE:
 607         case MADV_WILLNEED:
 608         case MADV_DONTNEED:
 609         case MADV_FREE:
 610 #ifdef CONFIG_KSM
 611         case MADV_MERGEABLE:
 612         case MADV_UNMERGEABLE:
 613 #endif
 614 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 615         case MADV_HUGEPAGE:
 616         case MADV_NOHUGEPAGE:
 617 #endif
 618         case MADV_DONTDUMP:
 619         case MADV_DODUMP:
 620                 return true;
 621
 622         default:
 623                 return false;
 624         }
 625 }
 626
 627 /*
 628  * The madvise(2) system call.
 629  *
 630  * Applications can use madvise() to advise the kernel how it should
 631  * handle paging I/O in this VM area.  The idea is to help the kernel
 632  * use appropriate read-ahead and caching techniques.  The information
 633  * provided is advisory only, and can be safely disregarded by the
 634  * kernel without affecting the correct operation of the application.
 635  *
 636  * behavior values:
 637  *  MADV_NORMAL - the default behavior is to read clusters.  This
 638  *              results in some read-ahead and read-behind.
 639  *  MADV_RANDOM - the system should read the minimum amount of data
 640  *              on any access, since it is unlikely that the appli-
 641  *              cation will need more than what it asks for.
 642  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 643  *              once, so they can be aggressively read ahead, and
 644  *              can be freed soon after they are accessed.
 645  *  MADV_WILLNEED - the application is notifying the system to read
 646  *              some pages ahead.
 647  *  MADV_DONTNEED - the application is finished with the given range,
 648  *              so the kernel can free resources associated with it.
 649  *  MADV_FREE - the application marks pages in the given range as lazy free,
 650  *              where actual purges are postponed until memory pressure happens.
 651  *  MADV_REMOVE - the application wants to free up the given range of
 652  *              pages and associated backing store.
 653  *  MADV_DONTFORK - omit this area from child's address space when forking:
 654  *              typically, to avoid COWing pages pinned by get_user_pages().
 655  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 656  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 657  *              were corrupted by unrecoverable hardware memory failure.
 658  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 659  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 660  *              this area with pages of identical content from other such areas.
 661  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 662  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 663  *              huge pages in the future. Existing pages might be coalesced and
 664  *              new pages might be allocated as THP.
 665  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 666  *              transparent huge pages so the existing pages will not be
 667  *              coalesced into THP and new pages will not be allocated as THP.
 668  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 669  *              from being included in its core dump.
 670  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 671  *
 672  * return values:
 673  *  zero    - success
 674  *  -EINVAL - start + len < 0, start is not page-aligned,
 675  *              "behavior" is not a valid value, or application
 676  *              is attempting to release locked or shared pages.
 677  *  -ENOMEM - addresses in the specified range are not currently
 678  *              mapped, or are outside the AS of the process.
 679  *  -EIO    - an I/O error occurred while paging in data.
 680  *  -EBADF  - map exists, but area maps something that isn't a file.
 681  *  -EAGAIN - a kernel resource was temporarily unavailable.
 682  */
 683 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 684 {
 685         unsigned long end, tmp;
 686         struct vm_area_struct *vma, *prev;
 687         int unmapped_error = 0;
 688         int error = -EINVAL;
 689         int write;
 690         size_t len;
 691         struct blk_plug plug;
 692
 693 #ifdef CONFIG_MEMORY_FAILURE
 694         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 695                 return madvise_hwpoison(behavior, start, start+len_in);
 696 #endif
 697         if (!madvise_behavior_valid(behavior))
 698                 return error;
 699
 700         if (start & ~PAGE_MASK)
 701                 return error;
 702         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 703
 704         /* Check to see whether len was rounded up from small -ve to zero */
 705         if (len_in && !len)
 706                 return error;
 707
 708         end = start + len;
 709         if (end < start)
 710                 return error;
 711
 712         error = 0;
 713         if (end == start)
 714                 return error;
 715
 716         write = madvise_need_mmap_write(behavior);
 717         if (write) {
 718                 if (down_write_killable(&current->mm->mmap_sem))
 719                         return -EINTR;
 720         } else {
 721                 down_read(&current->mm->mmap_sem);
 722         }
 723
 724         /*
 725          * If the interval [start,end) covers some unmapped address
 726          * ranges, just ignore them, but return -ENOMEM at the end.
 727          * - different from the way of handling in mlock etc.
 728          */
 729         vma = find_vma_prev(current->mm, start, &prev);
 730         if (vma && start > vma->vm_start)
 731                 prev = vma;
 732
 733         blk_start_plug(&plug);
 734         for (;;) {
 735                 /* Still start < end. */
 736                 error = -ENOMEM;
 737                 if (!vma)
 738                         goto out;
 739
 740                 /* Here start < (end|vma->vm_end). */
 741                 if (start < vma->vm_start) {
 742                         unmapped_error = -ENOMEM;
 743                         start = vma->vm_start;
 744                         if (start >= end)
 745                                 goto out;
 746                 }
 747
 748                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 749                 tmp = vma->vm_end;
 750                 if (end < tmp)
 751                         tmp = end;
 752
 753                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 754                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 755                 if (error)
 756                         goto out;
 757                 start = tmp;
 758                 if (prev && start < prev->vm_end)
 759                         start = prev->vm_end;
 760                 error = unmapped_error;
 761                 if (start >= end)
 762                         goto out;
 763                 if (prev)
 764                         vma = prev->vm_next;
 765                 else    /* madvise_remove dropped mmap_sem */
 766                         vma = find_vma(current->mm, start);
 767         }
 768 out:
 769         blk_finish_plug(&plug);
 770         if (write)
 771                 up_write(&current->mm->mmap_sem);
 772         else
 773                 up_read(&current->mm->mmap_sem);
 774
 775         return error;
 776 }