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                         if (PageActive(page))
 415                                 deactivate_page(page);
 416                         tlb_remove_tlb_entry(tlb, pte, addr);
 417                 }
 418         }
 419 out:
 420         if (nr_swap) {
 421                 if (current->mm == mm)
 422                         sync_mm_rss(mm);
 423
 424                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 425         }
 426         arch_leave_lazy_mmu_mode();
 427         pte_unmap_unlock(orig_pte, ptl);
 428         cond_resched();
 429 next:
 430         return 0;
 431 }
 432
 433 static void madvise_free_page_range(struct mmu_gather *tlb,
 434                              struct vm_area_struct *vma,
 435                              unsigned long addr, unsigned long end)
 436 {
 437         struct mm_walk free_walk = {
 438                 .pmd_entry = madvise_free_pte_range,
 439                 .mm = vma->vm_mm,
 440                 .private = tlb,
 441         };
 442
 443         tlb_start_vma(tlb, vma);
 444         walk_page_range(addr, end, &free_walk);
 445         tlb_end_vma(tlb, vma);
 446 }
 447
 448 static int madvise_free_single_vma(struct vm_area_struct *vma,
 449                         unsigned long start_addr, unsigned long end_addr)
 450 {
 451         unsigned long start, end;
 452         struct mm_struct *mm = vma->vm_mm;
 453         struct mmu_gather tlb;
 454
 455         if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
 456                 return -EINVAL;
 457
 458         /* MADV_FREE works for only anon vma at the moment */
 459         if (!vma_is_anonymous(vma))
 460                 return -EINVAL;
 461
 462         start = max(vma->vm_start, start_addr);
 463         if (start >= vma->vm_end)
 464                 return -EINVAL;
 465         end = min(vma->vm_end, end_addr);
 466         if (end <= vma->vm_start)
 467                 return -EINVAL;
 468
 469         lru_add_drain();
 470         tlb_gather_mmu(&tlb, mm, start, end);
 471         update_hiwater_rss(mm);
 472
 473         mmu_notifier_invalidate_range_start(mm, start, end);
 474         madvise_free_page_range(&tlb, vma, start, end);
 475         mmu_notifier_invalidate_range_end(mm, start, end);
 476         tlb_finish_mmu(&tlb, start, end);
 477
 478         return 0;
 479 }
 480
 481 static long madvise_free(struct vm_area_struct *vma,
 482                              struct vm_area_struct **prev,
 483                              unsigned long start, unsigned long end)
 484 {
 485         *prev = vma;
 486         return madvise_free_single_vma(vma, start, end);
 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(struct vm_area_struct *vma,
 509                              struct vm_area_struct **prev,
 510                              unsigned long start, unsigned long end)
 511 {
 512         *prev = vma;
 513         if (!can_madv_dontneed_vma(vma))
 514                 return -EINVAL;
 515
 516         if (!userfaultfd_remove(vma, start, end)) {
 517                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 518
 519                 down_read(&current->mm->mmap_sem);
 520                 vma = find_vma(current->mm, start);
 521                 if (!vma)
 522                         return -ENOMEM;
 523                 if (start < vma->vm_start) {
 524                         /*
 525                          * This "vma" under revalidation is the one
 526                          * with the lowest vma->vm_start where start
 527                          * is also < vma->vm_end. If start <
 528                          * vma->vm_start it means an hole materialized
 529                          * in the user address space within the
 530                          * virtual range passed to MADV_DONTNEED.
 531                          */
 532                         return -ENOMEM;
 533                 }
 534                 if (!can_madv_dontneed_vma(vma))
 535                         return -EINVAL;
 536                 if (end > vma->vm_end) {
 537                         /*
 538                          * Don't fail if end > vma->vm_end. If the old
 539                          * vma was splitted while the mmap_sem was
 540                          * released the effect of the concurrent
 541                          * operation may not cause MADV_DONTNEED to
 542                          * have an undefined result. There may be an
 543                          * adjacent next vma that we'll walk
 544                          * next. userfaultfd_remove() will generate an
 545                          * UFFD_EVENT_REMOVE repetition on the
 546                          * end-vma->vm_end range, but the manager can
 547                          * handle a repetition fine.
 548                          */
 549                         end = vma->vm_end;
 550                 }
 551                 VM_WARN_ON(start >= end);
 552         }
 553         zap_page_range(vma, start, end - start);
 554         return 0;
 555 }
 556
 557 /*
 558  * Application wants to free up the pages and associated backing store.
 559  * This is effectively punching a hole into the middle of a file.
 560  */
 561 static long madvise_remove(struct vm_area_struct *vma,
 562                                 struct vm_area_struct **prev,
 563                                 unsigned long start, unsigned long end)
 564 {
 565         loff_t offset;
 566         int error;
 567         struct file *f;
 568
 569         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 570
 571         if (vma->vm_flags & VM_LOCKED)
 572                 return -EINVAL;
 573
 574         f = vma->vm_file;
 575
 576         if (!f || !f->f_mapping || !f->f_mapping->host) {
 577                         return -EINVAL;
 578         }
 579
 580         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 581                 return -EACCES;
 582
 583         offset = (loff_t)(start - vma->vm_start)
 584                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 585
 586         /*
 587          * Filesystem's fallocate may need to take i_mutex.  We need to
 588          * explicitly grab a reference because the vma (and hence the
 589          * vma's reference to the file) can go away as soon as we drop
 590          * mmap_sem.
 591          */
 592         get_file(f);
 593         if (userfaultfd_remove(vma, start, end)) {
 594                 /* mmap_sem was not released by userfaultfd_remove() */
 595                 up_read(&current->mm->mmap_sem);
 596         }
 597         error = vfs_fallocate(f,
 598                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 599                                 offset, end - start);
 600         fput(f);
 601         down_read(&current->mm->mmap_sem);
 602         return error;
 603 }
 604
 605 #ifdef CONFIG_MEMORY_FAILURE
 606 /*
 607  * Error injection support for memory error handling.
 608  */
 609 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
 610 {
 611         struct page *p;
 612         if (!capable(CAP_SYS_ADMIN))
 613                 return -EPERM;
 614         for (; start < end; start += PAGE_SIZE <<
 615                                 compound_order(compound_head(p))) {
 616                 int ret;
 617
 618                 ret = get_user_pages_fast(start, 1, 0, &p);
 619                 if (ret != 1)
 620                         return ret;
 621
 622                 if (PageHWPoison(p)) {
 623                         put_page(p);
 624                         continue;
 625                 }
 626                 if (bhv == MADV_SOFT_OFFLINE) {
 627                         pr_info("Soft offlining page %#lx at %#lx\n",
 628                                 page_to_pfn(p), start);
 629                         ret = soft_offline_page(p, MF_COUNT_INCREASED);
 630                         if (ret)
 631                                 return ret;
 632                         continue;
 633                 }
 634                 pr_info("Injecting memory failure for page %#lx at %#lx\n",
 635                        page_to_pfn(p), start);
 636                 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
 637                 if (ret)
 638                         return ret;
 639         }
 640         return 0;
 641 }
 642 #endif
 643
 644 static long
 645 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 646                 unsigned long start, unsigned long end, int behavior)
 647 {
 648         switch (behavior) {
 649         case MADV_REMOVE:
 650                 return madvise_remove(vma, prev, start, end);
 651         case MADV_WILLNEED:
 652                 return madvise_willneed(vma, prev, start, end);
 653         case MADV_FREE:
 654                 /*
 655                  * XXX: In this implementation, MADV_FREE works like
 656                  * MADV_DONTNEED on swapless system or full swap.
 657                  */
 658                 if (get_nr_swap_pages() > 0)
 659                         return madvise_free(vma, prev, start, end);
 660                 /* passthrough */
 661         case MADV_DONTNEED:
 662                 return madvise_dontneed(vma, prev, start, end);
 663         default:
 664                 return madvise_behavior(vma, prev, start, end, behavior);
 665         }
 666 }
 667
 668 static bool
 669 madvise_behavior_valid(int behavior)
 670 {
 671         switch (behavior) {
 672         case MADV_DOFORK:
 673         case MADV_DONTFORK:
 674         case MADV_NORMAL:
 675         case MADV_SEQUENTIAL:
 676         case MADV_RANDOM:
 677         case MADV_REMOVE:
 678         case MADV_WILLNEED:
 679         case MADV_DONTNEED:
 680         case MADV_FREE:
 681 #ifdef CONFIG_KSM
 682         case MADV_MERGEABLE:
 683         case MADV_UNMERGEABLE:
 684 #endif
 685 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 686         case MADV_HUGEPAGE:
 687         case MADV_NOHUGEPAGE:
 688 #endif
 689         case MADV_DONTDUMP:
 690         case MADV_DODUMP:
 691                 return true;
 692
 693         default:
 694                 return false;
 695         }
 696 }
 697
 698 /*
 699  * The madvise(2) system call.
 700  *
 701  * Applications can use madvise() to advise the kernel how it should
 702  * handle paging I/O in this VM area.  The idea is to help the kernel
 703  * use appropriate read-ahead and caching techniques.  The information
 704  * provided is advisory only, and can be safely disregarded by the
 705  * kernel without affecting the correct operation of the application.
 706  *
 707  * behavior values:
 708  *  MADV_NORMAL - the default behavior is to read clusters.  This
 709  *              results in some read-ahead and read-behind.
 710  *  MADV_RANDOM - the system should read the minimum amount of data
 711  *              on any access, since it is unlikely that the appli-
 712  *              cation will need more than what it asks for.
 713  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 714  *              once, so they can be aggressively read ahead, and
 715  *              can be freed soon after they are accessed.
 716  *  MADV_WILLNEED - the application is notifying the system to read
 717  *              some pages ahead.
 718  *  MADV_DONTNEED - the application is finished with the given range,
 719  *              so the kernel can free resources associated with it.
 720  *  MADV_FREE - the application marks pages in the given range as lazy free,
 721  *              where actual purges are postponed until memory pressure happens.
 722  *  MADV_REMOVE - the application wants to free up the given range of
 723  *              pages and associated backing store.
 724  *  MADV_DONTFORK - omit this area from child's address space when forking:
 725  *              typically, to avoid COWing pages pinned by get_user_pages().
 726  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 727  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 728  *              were corrupted by unrecoverable hardware memory failure.
 729  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 730  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 731  *              this area with pages of identical content from other such areas.
 732  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 733  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 734  *              huge pages in the future. Existing pages might be coalesced and
 735  *              new pages might be allocated as THP.
 736  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 737  *              transparent huge pages so the existing pages will not be
 738  *              coalesced into THP and new pages will not be allocated as THP.
 739  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 740  *              from being included in its core dump.
 741  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 742  *
 743  * return values:
 744  *  zero    - success
 745  *  -EINVAL - start + len < 0, start is not page-aligned,
 746  *              "behavior" is not a valid value, or application
 747  *              is attempting to release locked or shared pages.
 748  *  -ENOMEM - addresses in the specified range are not currently
 749  *              mapped, or are outside the AS of the process.
 750  *  -EIO    - an I/O error occurred while paging in data.
 751  *  -EBADF  - map exists, but area maps something that isn't a file.
 752  *  -EAGAIN - a kernel resource was temporarily unavailable.
 753  */
 754 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 755 {
 756         unsigned long end, tmp;
 757         struct vm_area_struct *vma, *prev;
 758         int unmapped_error = 0;
 759         int error = -EINVAL;
 760         int write;
 761         size_t len;
 762         struct blk_plug plug;
 763
 764 #ifdef CONFIG_MEMORY_FAILURE
 765         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 766                 return madvise_hwpoison(behavior, start, start+len_in);
 767 #endif
 768         if (!madvise_behavior_valid(behavior))
 769                 return error;
 770
 771         if (start & ~PAGE_MASK)
 772                 return error;
 773         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 774
 775         /* Check to see whether len was rounded up from small -ve to zero */
 776         if (len_in && !len)
 777                 return error;
 778
 779         end = start + len;
 780         if (end < start)
 781                 return error;
 782
 783         error = 0;
 784         if (end == start)
 785                 return error;
 786
 787         write = madvise_need_mmap_write(behavior);
 788         if (write) {
 789                 if (down_write_killable(&current->mm->mmap_sem))
 790                         return -EINTR;
 791         } else {
 792                 down_read(&current->mm->mmap_sem);
 793         }
 794
 795         /*
 796          * If the interval [start,end) covers some unmapped address
 797          * ranges, just ignore them, but return -ENOMEM at the end.
 798          * - different from the way of handling in mlock etc.
 799          */
 800         vma = find_vma_prev(current->mm, start, &prev);
 801         if (vma && start > vma->vm_start)
 802                 prev = vma;
 803
 804         blk_start_plug(&plug);
 805         for (;;) {
 806                 /* Still start < end. */
 807                 error = -ENOMEM;
 808                 if (!vma)
 809                         goto out;
 810
 811                 /* Here start < (end|vma->vm_end). */
 812                 if (start < vma->vm_start) {
 813                         unmapped_error = -ENOMEM;
 814                         start = vma->vm_start;
 815                         if (start >= end)
 816                                 goto out;
 817                 }
 818
 819                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 820                 tmp = vma->vm_end;
 821                 if (end < tmp)
 822                         tmp = end;
 823
 824                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 825                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 826                 if (error)
 827                         goto out;
 828                 start = tmp;
 829                 if (prev && start < prev->vm_end)
 830                         start = prev->vm_end;
 831                 error = unmapped_error;
 832                 if (start >= end)
 833                         goto out;
 834                 if (prev)
 835                         vma = prev->vm_next;
 836                 else    /* madvise_remove dropped mmap_sem */
 837                         vma = find_vma(current->mm, start);
 838         }
 839 out:
 840         blk_finish_plug(&plug);
 841         if (write)
 842                 up_write(&current->mm->mmap_sem);
 843         else
 844                 up_read(&current->mm->mmap_sem);
 845
 846         return error;
 847 }