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/shmem_fs.h>
  25 #include <linux/mmu_notifier.h>
  26
  27 #include <asm/tlb.h>
  28
  29 #include "internal.h"
  30
  31 /*
  32  * Any behaviour which results in changes to the vma->vm_flags needs to
  33  * take mmap_sem for writing. Others, which simply traverse vmas, need
  34  * to only take it for reading.
  35  */
  36 static int madvise_need_mmap_write(int behavior)
  37 {
  38         switch (behavior) {
  39         case MADV_REMOVE:
  40         case MADV_WILLNEED:
  41         case MADV_DONTNEED:
  42         case MADV_FREE:
  43                 return 0;
  44         default:
  45                 /* be safe, default to 1. list exceptions explicitly */
  46                 return 1;
  47         }
  48 }
  49
  50 /*
  51  * We can potentially split a vm area into separate
  52  * areas, each area with its own behavior.
  53  */
  54 static long madvise_behavior(struct vm_area_struct *vma,
  55                      struct vm_area_struct **prev,
  56                      unsigned long start, unsigned long end, int behavior)
  57 {
  58         struct mm_struct *mm = vma->vm_mm;
  59         int error = 0;
  60         pgoff_t pgoff;
  61         unsigned long new_flags = vma->vm_flags;
  62
  63         switch (behavior) {
  64         case MADV_NORMAL:
  65                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  66                 break;
  67         case MADV_SEQUENTIAL:
  68                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  69                 break;
  70         case MADV_RANDOM:
  71                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  72                 break;
  73         case MADV_DONTFORK:
  74                 new_flags |= VM_DONTCOPY;
  75                 break;
  76         case MADV_DOFORK:
  77                 if (vma->vm_flags & VM_IO) {
  78                         error = -EINVAL;
  79                         goto out;
  80                 }
  81                 new_flags &= ~VM_DONTCOPY;
  82                 break;
  83         case MADV_WIPEONFORK:
  84                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  85                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  86                         error = -EINVAL;
  87                         goto out;
  88                 }
  89                 new_flags |= VM_WIPEONFORK;
  90                 break;
  91         case MADV_KEEPONFORK:
  92                 new_flags &= ~VM_WIPEONFORK;
  93                 break;
  94         case MADV_DONTDUMP:
  95                 new_flags |= VM_DONTDUMP;
  96                 break;
  97         case MADV_DODUMP:
  98                 if (new_flags & VM_SPECIAL) {
  99                         error = -EINVAL;
 100                         goto out;
 101                 }
 102                 new_flags &= ~VM_DONTDUMP;
 103                 break;
 104         case MADV_MERGEABLE:
 105         case MADV_UNMERGEABLE:
 106                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 107                 if (error) {
 108                         /*
 109                          * madvise() returns EAGAIN if kernel resources, such as
 110                          * slab, are temporarily unavailable.
 111                          */
 112                         if (error == -ENOMEM)
 113                                 error = -EAGAIN;
 114                         goto out;
 115                 }
 116                 break;
 117         case MADV_HUGEPAGE:
 118         case MADV_NOHUGEPAGE:
 119                 error = hugepage_madvise(vma, &new_flags, behavior);
 120                 if (error) {
 121                         /*
 122                          * madvise() returns EAGAIN if kernel resources, such as
 123                          * slab, are temporarily unavailable.
 124                          */
 125                         if (error == -ENOMEM)
 126                                 error = -EAGAIN;
 127                         goto out;
 128                 }
 129                 break;
 130         }
 131
 132         if (new_flags == vma->vm_flags) {
 133                 *prev = vma;
 134                 goto out;
 135         }
 136
 137         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 138         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 139                           vma->vm_file, pgoff, vma_policy(vma),
 140                           vma->vm_userfaultfd_ctx);
 141         if (*prev) {
 142                 vma = *prev;
 143                 goto success;
 144         }
 145
 146         *prev = vma;
 147
 148         if (start != vma->vm_start) {
 149                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 150                         error = -ENOMEM;
 151                         goto out;
 152                 }
 153                 error = __split_vma(mm, vma, start, 1);
 154                 if (error) {
 155                         /*
 156                          * madvise() returns EAGAIN if kernel resources, such as
 157                          * slab, are temporarily unavailable.
 158                          */
 159                         if (error == -ENOMEM)
 160                                 error = -EAGAIN;
 161                         goto out;
 162                 }
 163         }
 164
 165         if (end != vma->vm_end) {
 166                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 167                         error = -ENOMEM;
 168                         goto out;
 169                 }
 170                 error = __split_vma(mm, vma, end, 0);
 171                 if (error) {
 172                         /*
 173                          * madvise() returns EAGAIN if kernel resources, such as
 174                          * slab, are temporarily unavailable.
 175                          */
 176                         if (error == -ENOMEM)
 177                                 error = -EAGAIN;
 178                         goto out;
 179                 }
 180         }
 181
 182 success:
 183         /*
 184          * vm_flags is protected by the mmap_sem held in write mode.
 185          */
 186         vma->vm_flags = new_flags;
 187 out:
 188         return error;
 189 }
 190
 191 #ifdef CONFIG_SWAP
 192 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 193         unsigned long end, struct mm_walk *walk)
 194 {
 195         pte_t *orig_pte;
 196         struct vm_area_struct *vma = walk->private;
 197         unsigned long index;
 198
 199         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 200                 return 0;
 201
 202         for (index = start; index != end; index += PAGE_SIZE) {
 203                 pte_t pte;
 204                 swp_entry_t entry;
 205                 struct page *page;
 206                 spinlock_t *ptl;
 207
 208                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 209                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 210                 pte_unmap_unlock(orig_pte, ptl);
 211
 212                 if (pte_present(pte) || pte_none(pte))
 213                         continue;
 214                 entry = pte_to_swp_entry(pte);
 215                 if (unlikely(non_swap_entry(entry)))
 216                         continue;
 217
 218                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 219                                                         vma, index, false);
 220                 if (page)
 221                         put_page(page);
 222         }
 223
 224         return 0;
 225 }
 226
 227 static void force_swapin_readahead(struct vm_area_struct *vma,
 228                 unsigned long start, unsigned long end)
 229 {
 230         struct mm_walk walk = {
 231                 .mm = vma->vm_mm,
 232                 .pmd_entry = swapin_walk_pmd_entry,
 233                 .private = vma,
 234         };
 235
 236         walk_page_range(start, end, &walk);
 237
 238         lru_add_drain();        /* Push any new pages onto the LRU now */
 239 }
 240
 241 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 242                 unsigned long start, unsigned long end,
 243                 struct address_space *mapping)
 244 {
 245         pgoff_t index;
 246         struct page *page;
 247         swp_entry_t swap;
 248
 249         for (; start < end; start += PAGE_SIZE) {
 250                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 251
 252                 page = find_get_entry(mapping, index);
 253                 if (!radix_tree_exceptional_entry(page)) {
 254                         if (page)
 255                                 put_page(page);
 256                         continue;
 257                 }
 258                 swap = radix_to_swp_entry(page);
 259                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 260                                                         NULL, 0, false);
 261                 if (page)
 262                         put_page(page);
 263         }
 264
 265         lru_add_drain();        /* Push any new pages onto the LRU now */
 266 }
 267 #endif          /* CONFIG_SWAP */
 268
 269 /*
 270  * Schedule all required I/O operations.  Do not wait for completion.
 271  */
 272 static long madvise_willneed(struct vm_area_struct *vma,
 273                              struct vm_area_struct **prev,
 274                              unsigned long start, unsigned long end)
 275 {
 276         struct file *file = vma->vm_file;
 277
 278 #ifdef CONFIG_SWAP
 279         if (!file) {
 280                 *prev = vma;
 281                 force_swapin_readahead(vma, start, end);
 282                 return 0;
 283         }
 284
 285         if (shmem_mapping(file->f_mapping)) {
 286                 *prev = vma;
 287                 force_shm_swapin_readahead(vma, start, end,
 288                                         file->f_mapping);
 289                 return 0;
 290         }
 291 #else
 292         if (!file)
 293                 return -EBADF;
 294 #endif
 295
 296         if (IS_DAX(file_inode(file))) {
 297                 /* no bad return value, but ignore advice */
 298                 return 0;
 299         }
 300
 301         *prev = vma;
 302         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 303         if (end > vma->vm_end)
 304                 end = vma->vm_end;
 305         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 306
 307         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 308         return 0;
 309 }
 310
 311 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 312                                 unsigned long end, struct mm_walk *walk)
 313
 314 {
 315         struct mmu_gather *tlb = walk->private;
 316         struct mm_struct *mm = tlb->mm;
 317         struct vm_area_struct *vma = walk->vma;
 318         spinlock_t *ptl;
 319         pte_t *orig_pte, *pte, ptent;
 320         struct page *page;
 321         int nr_swap = 0;
 322         unsigned long next;
 323
 324         next = pmd_addr_end(addr, end);
 325         if (pmd_trans_huge(*pmd))
 326                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 327                         goto next;
 328
 329         if (pmd_trans_unstable(pmd))
 330                 return 0;
 331
 332         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
 333         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 334         flush_tlb_batched_pending(mm);
 335         arch_enter_lazy_mmu_mode();
 336         for (; addr != end; pte++, addr += PAGE_SIZE) {
 337                 ptent = *pte;
 338
 339                 if (pte_none(ptent))
 340                         continue;
 341                 /*
 342                  * If the pte has swp_entry, just clear page table to
 343                  * prevent swap-in which is more expensive rather than
 344                  * (page allocation + zeroing).
 345                  */
 346                 if (!pte_present(ptent)) {
 347                         swp_entry_t entry;
 348
 349                         entry = pte_to_swp_entry(ptent);
 350                         if (non_swap_entry(entry))
 351                                 continue;
 352                         nr_swap--;
 353                         free_swap_and_cache(entry);
 354                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 355                         continue;
 356                 }
 357
 358                 page = _vm_normal_page(vma, addr, ptent, true);
 359                 if (!page)
 360                         continue;
 361
 362                 /*
 363                  * If pmd isn't transhuge but the page is THP and
 364                  * is owned by only this process, split it and
 365                  * deactivate all pages.
 366                  */
 367                 if (PageTransCompound(page)) {
 368                         if (page_mapcount(page) != 1)
 369                                 goto out;
 370                         get_page(page);
 371                         if (!trylock_page(page)) {
 372                                 put_page(page);
 373                                 goto out;
 374                         }
 375                         pte_unmap_unlock(orig_pte, ptl);
 376                         if (split_huge_page(page)) {
 377                                 unlock_page(page);
 378                                 put_page(page);
 379                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 380                                 goto out;
 381                         }
 382                         unlock_page(page);
 383                         put_page(page);
 384                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 385                         pte--;
 386                         addr -= PAGE_SIZE;
 387                         continue;
 388                 }
 389
 390                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 391
 392                 if (PageSwapCache(page) || PageDirty(page)) {
 393                         if (!trylock_page(page))
 394                                 continue;
 395                         /*
 396                          * If page is shared with others, we couldn't clear
 397                          * PG_dirty of the page.
 398                          */
 399                         if (page_mapcount(page) != 1) {
 400                                 unlock_page(page);
 401                                 continue;
 402                         }
 403
 404                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 405                                 unlock_page(page);
 406                                 continue;
 407                         }
 408
 409                         ClearPageDirty(page);
 410                         unlock_page(page);
 411                 }
 412
 413                 if (pte_young(ptent) || pte_dirty(ptent)) {
 414                         /*
 415                          * Some of architecture(ex, PPC) don't update TLB
 416                          * with set_pte_at and tlb_remove_tlb_entry so for
 417                          * the portability, remap the pte with old|clean
 418                          * after pte clearing.
 419                          */
 420                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 421                                                         tlb->fullmm);
 422
 423                         ptent = pte_mkold(ptent);
 424                         ptent = pte_mkclean(ptent);
 425                         set_pte_at(mm, addr, pte, ptent);
 426                         tlb_remove_tlb_entry(tlb, pte, addr);
 427                 }
 428                 mark_page_lazyfree(page);
 429         }
 430 out:
 431         if (nr_swap) {
 432                 if (current->mm == mm)
 433                         sync_mm_rss(mm);
 434
 435                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 436         }
 437         arch_leave_lazy_mmu_mode();
 438         pte_unmap_unlock(orig_pte, ptl);
 439         cond_resched();
 440 next:
 441         return 0;
 442 }
 443
 444 static void madvise_free_page_range(struct mmu_gather *tlb,
 445                              struct vm_area_struct *vma,
 446                              unsigned long addr, unsigned long end)
 447 {
 448         struct mm_walk free_walk = {
 449                 .pmd_entry = madvise_free_pte_range,
 450                 .mm = vma->vm_mm,
 451                 .private = tlb,
 452         };
 453
 454         tlb_start_vma(tlb, vma);
 455         walk_page_range(addr, end, &free_walk);
 456         tlb_end_vma(tlb, vma);
 457 }
 458
 459 static int madvise_free_single_vma(struct vm_area_struct *vma,
 460                         unsigned long start_addr, unsigned long end_addr)
 461 {
 462         unsigned long start, end;
 463         struct mm_struct *mm = vma->vm_mm;
 464         struct mmu_gather tlb;
 465
 466         /* MADV_FREE works for only anon vma at the moment */
 467         if (!vma_is_anonymous(vma))
 468                 return -EINVAL;
 469
 470         start = max(vma->vm_start, start_addr);
 471         if (start >= vma->vm_end)
 472                 return -EINVAL;
 473         end = min(vma->vm_end, end_addr);
 474         if (end <= vma->vm_start)
 475                 return -EINVAL;
 476
 477         lru_add_drain();
 478         tlb_gather_mmu(&tlb, mm, start, end);
 479         update_hiwater_rss(mm);
 480
 481         mmu_notifier_invalidate_range_start(mm, start, end);
 482         madvise_free_page_range(&tlb, vma, start, end);
 483         mmu_notifier_invalidate_range_end(mm, start, end);
 484         tlb_finish_mmu(&tlb, start, end);
 485
 486         return 0;
 487 }
 488
 489 /*
 490  * Application no longer needs these pages.  If the pages are dirty,
 491  * it's OK to just throw them away.  The app will be more careful about
 492  * data it wants to keep.  Be sure to free swap resources too.  The
 493  * zap_page_range call sets things up for shrink_active_list to actually free
 494  * these pages later if no one else has touched them in the meantime,
 495  * although we could add these pages to a global reuse list for
 496  * shrink_active_list to pick up before reclaiming other pages.
 497  *
 498  * NB: This interface discards data rather than pushes it out to swap,
 499  * as some implementations do.  This has performance implications for
 500  * applications like large transactional databases which want to discard
 501  * pages in anonymous maps after committing to backing store the data
 502  * that was kept in them.  There is no reason to write this data out to
 503  * the swap area if the application is discarding it.
 504  *
 505  * An interface that causes the system to free clean pages and flush
 506  * dirty pages is already available as msync(MS_INVALIDATE).
 507  */
 508 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 509                                         unsigned long start, unsigned long end)
 510 {
 511         zap_page_range(vma, start, end - start);
 512         return 0;
 513 }
 514
 515 static long madvise_dontneed_free(struct vm_area_struct *vma,
 516                                   struct vm_area_struct **prev,
 517                                   unsigned long start, unsigned long end,
 518                                   int behavior)
 519 {
 520         *prev = vma;
 521         if (!can_madv_dontneed_vma(vma))
 522                 return -EINVAL;
 523
 524         if (!userfaultfd_remove(vma, start, end)) {
 525                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 526
 527                 down_read(&current->mm->mmap_sem);
 528                 vma = find_vma(current->mm, start);
 529                 if (!vma)
 530                         return -ENOMEM;
 531                 if (start < vma->vm_start) {
 532                         /*
 533                          * This "vma" under revalidation is the one
 534                          * with the lowest vma->vm_start where start
 535                          * is also < vma->vm_end. If start <
 536                          * vma->vm_start it means an hole materialized
 537                          * in the user address space within the
 538                          * virtual range passed to MADV_DONTNEED
 539                          * or MADV_FREE.
 540                          */
 541                         return -ENOMEM;
 542                 }
 543                 if (!can_madv_dontneed_vma(vma))
 544                         return -EINVAL;
 545                 if (end > vma->vm_end) {
 546                         /*
 547                          * Don't fail if end > vma->vm_end. If the old
 548                          * vma was splitted while the mmap_sem was
 549                          * released the effect of the concurrent
 550                          * operation may not cause madvise() to
 551                          * have an undefined result. There may be an
 552                          * adjacent next vma that we'll walk
 553                          * next. userfaultfd_remove() will generate an
 554                          * UFFD_EVENT_REMOVE repetition on the
 555                          * end-vma->vm_end range, but the manager can
 556                          * handle a repetition fine.
 557                          */
 558                         end = vma->vm_end;
 559                 }
 560                 VM_WARN_ON(start >= end);
 561         }
 562
 563         if (behavior == MADV_DONTNEED)
 564                 return madvise_dontneed_single_vma(vma, start, end);
 565         else if (behavior == MADV_FREE)
 566                 return madvise_free_single_vma(vma, start, end);
 567         else
 568                 return -EINVAL;
 569 }
 570
 571 /*
 572  * Application wants to free up the pages and associated backing store.
 573  * This is effectively punching a hole into the middle of a file.
 574  */
 575 static long madvise_remove(struct vm_area_struct *vma,
 576                                 struct vm_area_struct **prev,
 577                                 unsigned long start, unsigned long end)
 578 {
 579         loff_t offset;
 580         int error;
 581         struct file *f;
 582
 583         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 584
 585         if (vma->vm_flags & VM_LOCKED)
 586                 return -EINVAL;
 587
 588         f = vma->vm_file;
 589
 590         if (!f || !f->f_mapping || !f->f_mapping->host) {
 591                         return -EINVAL;
 592         }
 593
 594         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 595                 return -EACCES;
 596
 597         offset = (loff_t)(start - vma->vm_start)
 598                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 599
 600         /*
 601          * Filesystem's fallocate may need to take i_mutex.  We need to
 602          * explicitly grab a reference because the vma (and hence the
 603          * vma's reference to the file) can go away as soon as we drop
 604          * mmap_sem.
 605          */
 606         get_file(f);
 607         if (userfaultfd_remove(vma, start, end)) {
 608                 /* mmap_sem was not released by userfaultfd_remove() */
 609                 up_read(&current->mm->mmap_sem);
 610         }
 611         error = vfs_fallocate(f,
 612                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 613                                 offset, end - start);
 614         fput(f);
 615         down_read(&current->mm->mmap_sem);
 616         return error;
 617 }
 618
 619 #ifdef CONFIG_MEMORY_FAILURE
 620 /*
 621  * Error injection support for memory error handling.
 622  */
 623 static int madvise_inject_error(int behavior,
 624                 unsigned long start, unsigned long end)
 625 {
 626         struct page *page;
 627         struct zone *zone;
 628         unsigned int order;
 629
 630         if (!capable(CAP_SYS_ADMIN))
 631                 return -EPERM;
 632
 633
 634         for (; start < end; start += PAGE_SIZE << order) {
 635                 int ret;
 636
 637                 ret = get_user_pages_fast(start, 1, 0, &page);
 638                 if (ret != 1)
 639                         return ret;
 640
 641                 /*
 642                  * When soft offlining hugepages, after migrating the page
 643                  * we dissolve it, therefore in the second loop "page" will
 644                  * no longer be a compound page, and order will be 0.
 645                  */
 646                 order = compound_order(compound_head(page));
 647
 648                 if (PageHWPoison(page)) {
 649                         put_page(page);
 650                         continue;
 651                 }
 652
 653                 if (behavior == MADV_SOFT_OFFLINE) {
 654                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 655                                                 page_to_pfn(page), start);
 656
 657                         ret = soft_offline_page(page, MF_COUNT_INCREASED);
 658                         if (ret)
 659                                 return ret;
 660                         continue;
 661                 }
 662                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 663                                                 page_to_pfn(page), start);
 664
 665                 ret = memory_failure(page_to_pfn(page), 0, MF_COUNT_INCREASED);
 666                 if (ret)
 667                         return ret;
 668         }
 669
 670         /* Ensure that all poisoned pages are removed from per-cpu lists */
 671         for_each_populated_zone(zone)
 672                 drain_all_pages(zone);
 673
 674         return 0;
 675 }
 676 #endif
 677
 678 static long
 679 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 680                 unsigned long start, unsigned long end, int behavior)
 681 {
 682         switch (behavior) {
 683         case MADV_REMOVE:
 684                 return madvise_remove(vma, prev, start, end);
 685         case MADV_WILLNEED:
 686                 return madvise_willneed(vma, prev, start, end);
 687         case MADV_FREE:
 688         case MADV_DONTNEED:
 689                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 690         default:
 691                 return madvise_behavior(vma, prev, start, end, behavior);
 692         }
 693 }
 694
 695 static bool
 696 madvise_behavior_valid(int behavior)
 697 {
 698         switch (behavior) {
 699         case MADV_DOFORK:
 700         case MADV_DONTFORK:
 701         case MADV_NORMAL:
 702         case MADV_SEQUENTIAL:
 703         case MADV_RANDOM:
 704         case MADV_REMOVE:
 705         case MADV_WILLNEED:
 706         case MADV_DONTNEED:
 707         case MADV_FREE:
 708 #ifdef CONFIG_KSM
 709         case MADV_MERGEABLE:
 710         case MADV_UNMERGEABLE:
 711 #endif
 712 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 713         case MADV_HUGEPAGE:
 714         case MADV_NOHUGEPAGE:
 715 #endif
 716         case MADV_DONTDUMP:
 717         case MADV_DODUMP:
 718         case MADV_WIPEONFORK:
 719         case MADV_KEEPONFORK:
 720 #ifdef CONFIG_MEMORY_FAILURE
 721         case MADV_SOFT_OFFLINE:
 722         case MADV_HWPOISON:
 723 #endif
 724                 return true;
 725
 726         default:
 727                 return false;
 728         }
 729 }
 730
 731 /*
 732  * The madvise(2) system call.
 733  *
 734  * Applications can use madvise() to advise the kernel how it should
 735  * handle paging I/O in this VM area.  The idea is to help the kernel
 736  * use appropriate read-ahead and caching techniques.  The information
 737  * provided is advisory only, and can be safely disregarded by the
 738  * kernel without affecting the correct operation of the application.
 739  *
 740  * behavior values:
 741  *  MADV_NORMAL - the default behavior is to read clusters.  This
 742  *              results in some read-ahead and read-behind.
 743  *  MADV_RANDOM - the system should read the minimum amount of data
 744  *              on any access, since it is unlikely that the appli-
 745  *              cation will need more than what it asks for.
 746  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 747  *              once, so they can be aggressively read ahead, and
 748  *              can be freed soon after they are accessed.
 749  *  MADV_WILLNEED - the application is notifying the system to read
 750  *              some pages ahead.
 751  *  MADV_DONTNEED - the application is finished with the given range,
 752  *              so the kernel can free resources associated with it.
 753  *  MADV_FREE - the application marks pages in the given range as lazy free,
 754  *              where actual purges are postponed until memory pressure happens.
 755  *  MADV_REMOVE - the application wants to free up the given range of
 756  *              pages and associated backing store.
 757  *  MADV_DONTFORK - omit this area from child's address space when forking:
 758  *              typically, to avoid COWing pages pinned by get_user_pages().
 759  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 760  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
 761  *              range after a fork.
 762  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
 763  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 764  *              were corrupted by unrecoverable hardware memory failure.
 765  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 766  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 767  *              this area with pages of identical content from other such areas.
 768  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 769  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 770  *              huge pages in the future. Existing pages might be coalesced and
 771  *              new pages might be allocated as THP.
 772  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 773  *              transparent huge pages so the existing pages will not be
 774  *              coalesced into THP and new pages will not be allocated as THP.
 775  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 776  *              from being included in its core dump.
 777  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 778  *
 779  * return values:
 780  *  zero    - success
 781  *  -EINVAL - start + len < 0, start is not page-aligned,
 782  *              "behavior" is not a valid value, or application
 783  *              is attempting to release locked or shared pages,
 784  *              or the specified address range includes file, Huge TLB,
 785  *              MAP_SHARED or VMPFNMAP range.
 786  *  -ENOMEM - addresses in the specified range are not currently
 787  *              mapped, or are outside the AS of the process.
 788  *  -EIO    - an I/O error occurred while paging in data.
 789  *  -EBADF  - map exists, but area maps something that isn't a file.
 790  *  -EAGAIN - a kernel resource was temporarily unavailable.
 791  */
 792 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 793 {
 794         unsigned long end, tmp;
 795         struct vm_area_struct *vma, *prev;
 796         int unmapped_error = 0;
 797         int error = -EINVAL;
 798         int write;
 799         size_t len;
 800         struct blk_plug plug;
 801
 802         if (!madvise_behavior_valid(behavior))
 803                 return error;
 804
 805         if (start & ~PAGE_MASK)
 806                 return error;
 807         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 808
 809         /* Check to see whether len was rounded up from small -ve to zero */
 810         if (len_in && !len)
 811                 return error;
 812
 813         end = start + len;
 814         if (end < start)
 815                 return error;
 816
 817         error = 0;
 818         if (end == start)
 819                 return error;
 820
 821 #ifdef CONFIG_MEMORY_FAILURE
 822         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 823                 return madvise_inject_error(behavior, start, start + len_in);
 824 #endif
 825
 826         write = madvise_need_mmap_write(behavior);
 827         if (write) {
 828                 if (down_write_killable(&current->mm->mmap_sem))
 829                         return -EINTR;
 830         } else {
 831                 down_read(&current->mm->mmap_sem);
 832         }
 833
 834         /*
 835          * If the interval [start,end) covers some unmapped address
 836          * ranges, just ignore them, but return -ENOMEM at the end.
 837          * - different from the way of handling in mlock etc.
 838          */
 839         vma = find_vma_prev(current->mm, start, &prev);
 840         if (vma && start > vma->vm_start)
 841                 prev = vma;
 842
 843         blk_start_plug(&plug);
 844         for (;;) {
 845                 /* Still start < end. */
 846                 error = -ENOMEM;
 847                 if (!vma)
 848                         goto out;
 849
 850                 /* Here start < (end|vma->vm_end). */
 851                 if (start < vma->vm_start) {
 852                         unmapped_error = -ENOMEM;
 853                         start = vma->vm_start;
 854                         if (start >= end)
 855                                 goto out;
 856                 }
 857
 858                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 859                 tmp = vma->vm_end;
 860                 if (end < tmp)
 861                         tmp = end;
 862
 863                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 864                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 865                 if (error)
 866                         goto out;
 867                 start = tmp;
 868                 if (prev && start < prev->vm_end)
 869                         start = prev->vm_end;
 870                 error = unmapped_error;
 871                 if (start >= end)
 872                         goto out;
 873                 if (prev)
 874                         vma = prev->vm_next;
 875                 else    /* madvise_remove dropped mmap_sem */
 876                         vma = find_vma(current->mm, start);
 877         }
 878 out:
 879         blk_finish_plug(&plug);
 880         if (write)
 881                 up_write(&current->mm->mmap_sem);
 882         else
 883                 up_read(&current->mm->mmap_sem);
 884
 885         return error;
 886 }