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 #include "internal.h"
  29
  30 /*
  31  * Any behaviour which results in changes to the vma->vm_flags needs to
  32  * take mmap_sem for writing. Others, which simply traverse vmas, need
  33  * to only take it for reading.
  34  */
  35 static int madvise_need_mmap_write(int behavior)
  36 {
  37         switch (behavior) {
  38         case MADV_REMOVE:
  39         case MADV_WILLNEED:
  40         case MADV_DONTNEED:
  41         case MADV_FREE:
  42                 return 0;
  43         default:
  44                 /* be safe, default to 1. list exceptions explicitly */
  45                 return 1;
  46         }
  47 }
  48
  49 /*
  50  * We can potentially split a vm area into separate
  51  * areas, each area with its own behavior.
  52  */
  53 static long madvise_behavior(struct vm_area_struct *vma,
  54                      struct vm_area_struct **prev,
  55                      unsigned long start, unsigned long end, int behavior)
  56 {
  57         struct mm_struct *mm = vma->vm_mm;
  58         int error = 0;
  59         pgoff_t pgoff;
  60         unsigned long new_flags = vma->vm_flags;
  61
  62         switch (behavior) {
  63         case MADV_NORMAL:
  64                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  65                 break;
  66         case MADV_SEQUENTIAL:
  67                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  68                 break;
  69         case MADV_RANDOM:
  70                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  71                 break;
  72         case MADV_DONTFORK:
  73                 new_flags |= VM_DONTCOPY;
  74                 break;
  75         case MADV_DOFORK:
  76                 if (vma->vm_flags & VM_IO) {
  77                         error = -EINVAL;
  78                         goto out;
  79                 }
  80                 new_flags &= ~VM_DONTCOPY;
  81                 break;
  82         case MADV_DONTDUMP:
  83                 new_flags |= VM_DONTDUMP;
  84                 break;
  85         case MADV_DODUMP:
  86                 if (new_flags & VM_SPECIAL) {
  87                         error = -EINVAL;
  88                         goto out;
  89                 }
  90                 new_flags &= ~VM_DONTDUMP;
  91                 break;
  92         case MADV_MERGEABLE:
  93         case MADV_UNMERGEABLE:
  94                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
  95                 if (error)
  96                         goto out;
  97                 break;
  98         case MADV_HUGEPAGE:
  99         case MADV_NOHUGEPAGE:
 100                 error = hugepage_madvise(vma, &new_flags, behavior);
 101                 if (error)
 102                         goto out;
 103                 break;
 104         }
 105
 106         if (new_flags == vma->vm_flags) {
 107                 *prev = vma;
 108                 goto out;
 109         }
 110
 111         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 112         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 113                           vma->vm_file, pgoff, vma_policy(vma),
 114                           vma->vm_userfaultfd_ctx);
 115         if (*prev) {
 116                 vma = *prev;
 117                 goto success;
 118         }
 119
 120         *prev = vma;
 121
 122         if (start != vma->vm_start) {
 123                 error = split_vma(mm, vma, start, 1);
 124                 if (error)
 125                         goto out;
 126         }
 127
 128         if (end != vma->vm_end) {
 129                 error = split_vma(mm, vma, end, 0);
 130                 if (error)
 131                         goto out;
 132         }
 133
 134 success:
 135         /*
 136          * vm_flags is protected by the mmap_sem held in write mode.
 137          */
 138         vma->vm_flags = new_flags;
 139
 140 out:
 141         if (error == -ENOMEM)
 142                 error = -EAGAIN;
 143         return error;
 144 }
 145
 146 #ifdef CONFIG_SWAP
 147 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 148         unsigned long end, struct mm_walk *walk)
 149 {
 150         pte_t *orig_pte;
 151         struct vm_area_struct *vma = walk->private;
 152         unsigned long index;
 153
 154         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 155                 return 0;
 156
 157         for (index = start; index != end; index += PAGE_SIZE) {
 158                 pte_t pte;
 159                 swp_entry_t entry;
 160                 struct page *page;
 161                 spinlock_t *ptl;
 162
 163                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 164                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 165                 pte_unmap_unlock(orig_pte, ptl);
 166
 167                 if (pte_present(pte) || pte_none(pte))
 168                         continue;
 169                 entry = pte_to_swp_entry(pte);
 170                 if (unlikely(non_swap_entry(entry)))
 171                         continue;
 172
 173                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 174                                                                 vma, index);
 175                 if (page)
 176                         put_page(page);
 177         }
 178
 179         return 0;
 180 }
 181
 182 static void force_swapin_readahead(struct vm_area_struct *vma,
 183                 unsigned long start, unsigned long end)
 184 {
 185         struct mm_walk walk = {
 186                 .mm = vma->vm_mm,
 187                 .pmd_entry = swapin_walk_pmd_entry,
 188                 .private = vma,
 189         };
 190
 191         walk_page_range(start, end, &walk);
 192
 193         lru_add_drain();        /* Push any new pages onto the LRU now */
 194 }
 195
 196 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 197                 unsigned long start, unsigned long end,
 198                 struct address_space *mapping)
 199 {
 200         pgoff_t index;
 201         struct page *page;
 202         swp_entry_t swap;
 203
 204         for (; start < end; start += PAGE_SIZE) {
 205                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 206
 207                 page = find_get_entry(mapping, index);
 208                 if (!radix_tree_exceptional_entry(page)) {
 209                         if (page)
 210                                 put_page(page);
 211                         continue;
 212                 }
 213                 swap = radix_to_swp_entry(page);
 214                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 215                                                                 NULL, 0);
 216                 if (page)
 217                         put_page(page);
 218         }
 219
 220         lru_add_drain();        /* Push any new pages onto the LRU now */
 221 }
 222 #endif          /* CONFIG_SWAP */
 223
 224 /*
 225  * Schedule all required I/O operations.  Do not wait for completion.
 226  */
 227 static long madvise_willneed(struct vm_area_struct *vma,
 228                              struct vm_area_struct **prev,
 229                              unsigned long start, unsigned long end)
 230 {
 231         struct file *file = vma->vm_file;
 232
 233 #ifdef CONFIG_SWAP
 234         if (!file) {
 235                 *prev = vma;
 236                 force_swapin_readahead(vma, start, end);
 237                 return 0;
 238         }
 239
 240         if (shmem_mapping(file->f_mapping)) {
 241                 *prev = vma;
 242                 force_shm_swapin_readahead(vma, start, end,
 243                                         file->f_mapping);
 244                 return 0;
 245         }
 246 #else
 247         if (!file)
 248                 return -EBADF;
 249 #endif
 250
 251         if (IS_DAX(file_inode(file))) {
 252                 /* no bad return value, but ignore advice */
 253                 return 0;
 254         }
 255
 256         *prev = vma;
 257         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 258         if (end > vma->vm_end)
 259                 end = vma->vm_end;
 260         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 261
 262         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 263         return 0;
 264 }
 265
 266 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 267                                 unsigned long end, struct mm_walk *walk)
 268
 269 {
 270         struct mmu_gather *tlb = walk->private;
 271         struct mm_struct *mm = tlb->mm;
 272         struct vm_area_struct *vma = walk->vma;
 273         spinlock_t *ptl;
 274         pte_t *orig_pte, *pte, ptent;
 275         struct page *page;
 276         int nr_swap = 0;
 277         unsigned long next;
 278
 279         next = pmd_addr_end(addr, end);
 280         if (pmd_trans_huge(*pmd))
 281                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 282                         goto next;
 283
 284         if (pmd_trans_unstable(pmd))
 285                 return 0;
 286
 287         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
 288         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 289         arch_enter_lazy_mmu_mode();
 290         for (; addr != end; pte++, addr += PAGE_SIZE) {
 291                 ptent = *pte;
 292
 293                 if (pte_none(ptent))
 294                         continue;
 295                 /*
 296                  * If the pte has swp_entry, just clear page table to
 297                  * prevent swap-in which is more expensive rather than
 298                  * (page allocation + zeroing).
 299                  */
 300                 if (!pte_present(ptent)) {
 301                         swp_entry_t entry;
 302
 303                         entry = pte_to_swp_entry(ptent);
 304                         if (non_swap_entry(entry))
 305                                 continue;
 306                         nr_swap--;
 307                         free_swap_and_cache(entry);
 308                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 309                         continue;
 310                 }
 311
 312                 page = vm_normal_page(vma, addr, ptent);
 313                 if (!page)
 314                         continue;
 315
 316                 /*
 317                  * If pmd isn't transhuge but the page is THP and
 318                  * is owned by only this process, split it and
 319                  * deactivate all pages.
 320                  */
 321                 if (PageTransCompound(page)) {
 322                         if (page_mapcount(page) != 1)
 323                                 goto out;
 324                         get_page(page);
 325                         if (!trylock_page(page)) {
 326                                 put_page(page);
 327                                 goto out;
 328                         }
 329                         pte_unmap_unlock(orig_pte, ptl);
 330                         if (split_huge_page(page)) {
 331                                 unlock_page(page);
 332                                 put_page(page);
 333                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 334                                 goto out;
 335                         }
 336                         put_page(page);
 337                         unlock_page(page);
 338                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 339                         pte--;
 340                         addr -= PAGE_SIZE;
 341                         continue;
 342                 }
 343
 344                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 345
 346                 if (PageSwapCache(page) || PageDirty(page)) {
 347                         if (!trylock_page(page))
 348                                 continue;
 349                         /*
 350                          * If page is shared with others, we couldn't clear
 351                          * PG_dirty of the page.
 352                          */
 353                         if (page_mapcount(page) != 1) {
 354                                 unlock_page(page);
 355                                 continue;
 356                         }
 357
 358                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 359                                 unlock_page(page);
 360                                 continue;
 361                         }
 362
 363                         ClearPageDirty(page);
 364                         unlock_page(page);
 365                 }
 366
 367                 if (pte_young(ptent) || pte_dirty(ptent)) {
 368                         /*
 369                          * Some of architecture(ex, PPC) don't update TLB
 370                          * with set_pte_at and tlb_remove_tlb_entry so for
 371                          * the portability, remap the pte with old|clean
 372                          * after pte clearing.
 373                          */
 374                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 375                                                         tlb->fullmm);
 376
 377                         ptent = pte_mkold(ptent);
 378                         ptent = pte_mkclean(ptent);
 379                         set_pte_at(mm, addr, pte, ptent);
 380                         if (PageActive(page))
 381                                 deactivate_page(page);
 382                         tlb_remove_tlb_entry(tlb, pte, addr);
 383                 }
 384         }
 385 out:
 386         if (nr_swap) {
 387                 if (current->mm == mm)
 388                         sync_mm_rss(mm);
 389
 390                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 391         }
 392         arch_leave_lazy_mmu_mode();
 393         pte_unmap_unlock(orig_pte, ptl);
 394         cond_resched();
 395 next:
 396         return 0;
 397 }
 398
 399 static void madvise_free_page_range(struct mmu_gather *tlb,
 400                              struct vm_area_struct *vma,
 401                              unsigned long addr, unsigned long end)
 402 {
 403         struct mm_walk free_walk = {
 404                 .pmd_entry = madvise_free_pte_range,
 405                 .mm = vma->vm_mm,
 406                 .private = tlb,
 407         };
 408
 409         tlb_start_vma(tlb, vma);
 410         walk_page_range(addr, end, &free_walk);
 411         tlb_end_vma(tlb, vma);
 412 }
 413
 414 static int madvise_free_single_vma(struct vm_area_struct *vma,
 415                         unsigned long start_addr, unsigned long end_addr)
 416 {
 417         unsigned long start, end;
 418         struct mm_struct *mm = vma->vm_mm;
 419         struct mmu_gather tlb;
 420
 421         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 422                 return -EINVAL;
 423
 424         /* MADV_FREE works for only anon vma at the moment */
 425         if (!vma_is_anonymous(vma))
 426                 return -EINVAL;
 427
 428         start = max(vma->vm_start, start_addr);
 429         if (start >= vma->vm_end)
 430                 return -EINVAL;
 431         end = min(vma->vm_end, end_addr);
 432         if (end <= vma->vm_start)
 433                 return -EINVAL;
 434
 435         lru_add_drain();
 436         tlb_gather_mmu(&tlb, mm, start, end);
 437         update_hiwater_rss(mm);
 438
 439         mmu_notifier_invalidate_range_start(mm, start, end);
 440         madvise_free_page_range(&tlb, vma, start, end);
 441         mmu_notifier_invalidate_range_end(mm, start, end);
 442         tlb_finish_mmu(&tlb, start, end);
 443
 444         return 0;
 445 }
 446
 447 static long madvise_free(struct vm_area_struct *vma,
 448                              struct vm_area_struct **prev,
 449                              unsigned long start, unsigned long end)
 450 {
 451         *prev = vma;
 452         return madvise_free_single_vma(vma, start, end);
 453 }
 454
 455 /*
 456  * Application no longer needs these pages.  If the pages are dirty,
 457  * it's OK to just throw them away.  The app will be more careful about
 458  * data it wants to keep.  Be sure to free swap resources too.  The
 459  * zap_page_range call sets things up for shrink_active_list to actually free
 460  * these pages later if no one else has touched them in the meantime,
 461  * although we could add these pages to a global reuse list for
 462  * shrink_active_list to pick up before reclaiming other pages.
 463  *
 464  * NB: This interface discards data rather than pushes it out to swap,
 465  * as some implementations do.  This has performance implications for
 466  * applications like large transactional databases which want to discard
 467  * pages in anonymous maps after committing to backing store the data
 468  * that was kept in them.  There is no reason to write this data out to
 469  * the swap area if the application is discarding it.
 470  *
 471  * An interface that causes the system to free clean pages and flush
 472  * dirty pages is already available as msync(MS_INVALIDATE).
 473  */
 474 static long madvise_dontneed(struct vm_area_struct *vma,
 475                              struct vm_area_struct **prev,
 476                              unsigned long start, unsigned long end)
 477 {
 478         *prev = vma;
 479         if (!can_madv_dontneed_vma(vma))
 480                 return -EINVAL;
 481
 482         userfaultfd_remove(vma, prev, start, end);
 483         zap_page_range(vma, start, end - start);
 484         return 0;
 485 }
 486
 487 /*
 488  * Application wants to free up the pages and associated backing store.
 489  * This is effectively punching a hole into the middle of a file.
 490  */
 491 static long madvise_remove(struct vm_area_struct *vma,
 492                                 struct vm_area_struct **prev,
 493                                 unsigned long start, unsigned long end)
 494 {
 495         loff_t offset;
 496         int error;
 497         struct file *f;
 498
 499         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 500
 501         if (vma->vm_flags & VM_LOCKED)
 502                 return -EINVAL;
 503
 504         f = vma->vm_file;
 505
 506         if (!f || !f->f_mapping || !f->f_mapping->host) {
 507                         return -EINVAL;
 508         }
 509
 510         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 511                 return -EACCES;
 512
 513         offset = (loff_t)(start - vma->vm_start)
 514                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 515
 516         /*
 517          * Filesystem's fallocate may need to take i_mutex.  We need to
 518          * explicitly grab a reference because the vma (and hence the
 519          * vma's reference to the file) can go away as soon as we drop
 520          * mmap_sem.
 521          */
 522         get_file(f);
 523         userfaultfd_remove(vma, prev, start, end);
 524         up_read(&current->mm->mmap_sem);
 525         error = vfs_fallocate(f,
 526                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 527                                 offset, end - start);
 528         fput(f);
 529         down_read(&current->mm->mmap_sem);
 530         return error;
 531 }
 532
 533 #ifdef CONFIG_MEMORY_FAILURE
 534 /*
 535  * Error injection support for memory error handling.
 536  */
 537 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 538 {
 539         struct page *p;
 540         if (!capable(CAP_SYS_ADMIN))
 541                 return -EPERM;
 542         for (; start < end; start += PAGE_SIZE <<
 543                                 compound_order(compound_head(p))) {
 544                 int ret;
 545
 546                 ret = get_user_pages_fast(start, 1, 0, &p);
 547                 if (ret != 1)
 548                         return ret;
 549
 550                 if (PageHWPoison(p)) {
 551                         put_page(p);
 552                         continue;
 553                 }
 554                 if (bhv == MADV_SOFT_OFFLINE) {
 555                         pr_info("Soft offlining page %#lx at %#lx\n",
 556                                 page_to_pfn(p), start);
 557                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 558                         if (ret)
 559                                 return ret;
 560                         continue;
 561                 }
 562                 pr_info("Injecting memory failure for page %#lx at %#lx\n",
 563                        page_to_pfn(p), start);
 564                 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 565                 if (ret)
 566                         return ret;
 567         }
 568         return 0;
 569 }
 570 #endif
 571
 572 static long
 573 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 574                 unsigned long start, unsigned long end, int behavior)
 575 {
 576         switch (behavior) {
 577         case MADV_REMOVE:
 578                 return madvise_remove(vma, prev, start, end);
 579         case MADV_WILLNEED:
 580                 return madvise_willneed(vma, prev, start, end);
 581         case MADV_FREE:
 582                 /*
 583                  * XXX: In this implementation, MADV_FREE works like
 584                  * MADV_DONTNEED on swapless system or full swap.
 585                  */
 586                 if (get_nr_swap_pages() > 0)
 587                         return madvise_free(vma, prev, start, end);
 588                 /* passthrough */
 589         case MADV_DONTNEED:
 590                 return madvise_dontneed(vma, prev, start, end);
 591         default:
 592                 return madvise_behavior(vma, prev, start, end, behavior);
 593         }
 594 }
 595
 596 static bool
 597 madvise_behavior_valid(int behavior)
 598 {
 599         switch (behavior) {
 600         case MADV_DOFORK:
 601         case MADV_DONTFORK:
 602         case MADV_NORMAL:
 603         case MADV_SEQUENTIAL:
 604         case MADV_RANDOM:
 605         case MADV_REMOVE:
 606         case MADV_WILLNEED:
 607         case MADV_DONTNEED:
 608         case MADV_FREE:
 609 #ifdef CONFIG_KSM
 610         case MADV_MERGEABLE:
 611         case MADV_UNMERGEABLE:
 612 #endif
 613 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 614         case MADV_HUGEPAGE:
 615         case MADV_NOHUGEPAGE:
 616 #endif
 617         case MADV_DONTDUMP:
 618         case MADV_DODUMP:
 619                 return true;
 620
 621         default:
 622                 return false;
 623         }
 624 }
 625
 626 /*
 627  * The madvise(2) system call.
 628  *
 629  * Applications can use madvise() to advise the kernel how it should
 630  * handle paging I/O in this VM area.  The idea is to help the kernel
 631  * use appropriate read-ahead and caching techniques.  The information
 632  * provided is advisory only, and can be safely disregarded by the
 633  * kernel without affecting the correct operation of the application.
 634  *
 635  * behavior values:
 636  *  MADV_NORMAL - the default behavior is to read clusters.  This
 637  *              results in some read-ahead and read-behind.
 638  *  MADV_RANDOM - the system should read the minimum amount of data
 639  *              on any access, since it is unlikely that the appli-
 640  *              cation will need more than what it asks for.
 641  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 642  *              once, so they can be aggressively read ahead, and
 643  *              can be freed soon after they are accessed.
 644  *  MADV_WILLNEED - the application is notifying the system to read
 645  *              some pages ahead.
 646  *  MADV_DONTNEED - the application is finished with the given range,
 647  *              so the kernel can free resources associated with it.
 648  *  MADV_FREE - the application marks pages in the given range as lazy free,
 649  *              where actual purges are postponed until memory pressure happens.
 650  *  MADV_REMOVE - the application wants to free up the given range of
 651  *              pages and associated backing store.
 652  *  MADV_DONTFORK - omit this area from child's address space when forking:
 653  *              typically, to avoid COWing pages pinned by get_user_pages().
 654  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 655  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 656  *              were corrupted by unrecoverable hardware memory failure.
 657  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 658  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 659  *              this area with pages of identical content from other such areas.
 660  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 661  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 662  *              huge pages in the future. Existing pages might be coalesced and
 663  *              new pages might be allocated as THP.
 664  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 665  *              transparent huge pages so the existing pages will not be
 666  *              coalesced into THP and new pages will not be allocated as THP.
 667  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 668  *              from being included in its core dump.
 669  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 670  *
 671  * return values:
 672  *  zero    - success
 673  *  -EINVAL - start + len < 0, start is not page-aligned,
 674  *              "behavior" is not a valid value, or application
 675  *              is attempting to release locked or shared pages.
 676  *  -ENOMEM - addresses in the specified range are not currently
 677  *              mapped, or are outside the AS of the process.
 678  *  -EIO    - an I/O error occurred while paging in data.
 679  *  -EBADF  - map exists, but area maps something that isn't a file.
 680  *  -EAGAIN - a kernel resource was temporarily unavailable.
 681  */
 682 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 683 {
 684         unsigned long end, tmp;
 685         struct vm_area_struct *vma, *prev;
 686         int unmapped_error = 0;
 687         int error = -EINVAL;
 688         int write;
 689         size_t len;
 690         struct blk_plug plug;
 691
 692 #ifdef CONFIG_MEMORY_FAILURE
 693         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 694                 return madvise_hwpoison(behavior, start, start+len_in);
 695 #endif
 696         if (!madvise_behavior_valid(behavior))
 697                 return error;
 698
 699         if (start & ~PAGE_MASK)
 700                 return error;
 701         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 702
 703         /* Check to see whether len was rounded up from small -ve to zero */
 704         if (len_in && !len)
 705                 return error;
 706
 707         end = start + len;
 708         if (end < start)
 709                 return error;
 710
 711         error = 0;
 712         if (end == start)
 713                 return error;
 714
 715         write = madvise_need_mmap_write(behavior);
 716         if (write) {
 717                 if (down_write_killable(&current->mm->mmap_sem))
 718                         return -EINTR;
 719         } else {
 720                 down_read(&current->mm->mmap_sem);
 721         }
 722
 723         /*
 724          * If the interval [start,end) covers some unmapped address
 725          * ranges, just ignore them, but return -ENOMEM at the end.
 726          * - different from the way of handling in mlock etc.
 727          */
 728         vma = find_vma_prev(current->mm, start, &prev);
 729         if (vma && start > vma->vm_start)
 730                 prev = vma;
 731
 732         blk_start_plug(&plug);
 733         for (;;) {
 734                 /* Still start < end. */
 735                 error = -ENOMEM;
 736                 if (!vma)
 737                         goto out;
 738
 739                 /* Here start < (end|vma->vm_end). */
 740                 if (start < vma->vm_start) {
 741                         unmapped_error = -ENOMEM;
 742                         start = vma->vm_start;
 743                         if (start >= end)
 744                                 goto out;
 745                 }
 746
 747                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 748                 tmp = vma->vm_end;
 749                 if (end < tmp)
 750                         tmp = end;
 751
 752                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 753                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 754                 if (error)
 755                         goto out;
 756                 start = tmp;
 757                 if (prev && start < prev->vm_end)
 758                         start = prev->vm_end;
 759                 error = unmapped_error;
 760                 if (start >= end)
 761                         goto out;
 762                 if (prev)
 763                         vma = prev->vm_next;
 764                 else    /* madvise_remove dropped mmap_sem */
 765                         vma = find_vma(current->mm, start);
 766         }
 767 out:
 768         blk_finish_plug(&plug);
 769         if (write)
 770                 up_write(&current->mm->mmap_sem);
 771         else
 772                 up_read(&current->mm->mmap_sem);
 773
 774         return error;
 775 }