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