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