mm/madvise.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *      linux/mm/madvise.c
   4  *
   5  * Copyright (C) 1999  Linus Torvalds
   6  * Copyright (C) 2002  Christoph Hellwig
   7  */
   8
   9 #include <linux/mman.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/syscalls.h>
  12 #include <linux/mempolicy.h>
  13 #include <linux/page-isolation.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/falloc.h>
  17 #include <linux/sched.h>
  18 #include <linux/ksm.h>
  19 #include <linux/fs.h>
  20 #include <linux/file.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/backing-dev.h>
  23 #include <linux/pagewalk.h>
  24 #include <linux/swap.h>
  25 #include <linux/swapops.h>
  26 #include <linux/shmem_fs.h>
  27 #include <linux/mmu_notifier.h>
  28
  29 #include <asm/tlb.h>
  30
  31 #include "internal.h"
  32
  33 /*
  34  * Any behaviour which results in changes to the vma->vm_flags needs to
  35  * take mmap_sem for writing. Others, which simply traverse vmas, need
  36  * to only take it for reading.
  37  */
  38 static int madvise_need_mmap_write(int behavior)
  39 {
  40         switch (behavior) {
  41         case MADV_REMOVE:
  42         case MADV_WILLNEED:
  43         case MADV_DONTNEED:
  44         case MADV_FREE:
  45                 return 0;
  46         default:
  47                 /* be safe, default to 1. list exceptions explicitly */
  48                 return 1;
  49         }
  50 }
  51
  52 /*
  53  * We can potentially split a vm area into separate
  54  * areas, each area with its own behavior.
  55  */
  56 static long madvise_behavior(struct vm_area_struct *vma,
  57                      struct vm_area_struct **prev,
  58                      unsigned long start, unsigned long end, int behavior)
  59 {
  60         struct mm_struct *mm = vma->vm_mm;
  61         int error = 0;
  62         pgoff_t pgoff;
  63         unsigned long new_flags = vma->vm_flags;
  64
  65         switch (behavior) {
  66         case MADV_NORMAL:
  67                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  68                 break;
  69         case MADV_SEQUENTIAL:
  70                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  71                 break;
  72         case MADV_RANDOM:
  73                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  74                 break;
  75         case MADV_DONTFORK:
  76                 new_flags |= VM_DONTCOPY;
  77                 break;
  78         case MADV_DOFORK:
  79                 if (vma->vm_flags & VM_IO) {
  80                         error = -EINVAL;
  81                         goto out;
  82                 }
  83                 new_flags &= ~VM_DONTCOPY;
  84                 break;
  85         case MADV_WIPEONFORK:
  86                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  87                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  88                         error = -EINVAL;
  89                         goto out;
  90                 }
  91                 new_flags |= VM_WIPEONFORK;
  92                 break;
  93         case MADV_KEEPONFORK:
  94                 new_flags &= ~VM_WIPEONFORK;
  95                 break;
  96         case MADV_DONTDUMP:
  97                 new_flags |= VM_DONTDUMP;
  98                 break;
  99         case MADV_DODUMP:
 100                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 101                         error = -EINVAL;
 102                         goto out;
 103                 }
 104                 new_flags &= ~VM_DONTDUMP;
 105                 break;
 106         case MADV_MERGEABLE:
 107         case MADV_UNMERGEABLE:
 108                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 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         case MADV_HUGEPAGE:
 120         case MADV_NOHUGEPAGE:
 121                 error = hugepage_madvise(vma, &new_flags, behavior);
 122                 if (error) {
 123                         /*
 124                          * madvise() returns EAGAIN if kernel resources, such as
 125                          * slab, are temporarily unavailable.
 126                          */
 127                         if (error == -ENOMEM)
 128                                 error = -EAGAIN;
 129                         goto out;
 130                 }
 131                 break;
 132         }
 133
 134         if (new_flags == vma->vm_flags) {
 135                 *prev = vma;
 136                 goto out;
 137         }
 138
 139         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 140         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 141                           vma->vm_file, pgoff, vma_policy(vma),
 142                           vma->vm_userfaultfd_ctx);
 143         if (*prev) {
 144                 vma = *prev;
 145                 goto success;
 146         }
 147
 148         *prev = vma;
 149
 150         if (start != vma->vm_start) {
 151                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 152                         error = -ENOMEM;
 153                         goto out;
 154                 }
 155                 error = __split_vma(mm, vma, start, 1);
 156                 if (error) {
 157                         /*
 158                          * madvise() returns EAGAIN if kernel resources, such as
 159                          * slab, are temporarily unavailable.
 160                          */
 161                         if (error == -ENOMEM)
 162                                 error = -EAGAIN;
 163                         goto out;
 164                 }
 165         }
 166
 167         if (end != vma->vm_end) {
 168                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 169                         error = -ENOMEM;
 170                         goto out;
 171                 }
 172                 error = __split_vma(mm, vma, end, 0);
 173                 if (error) {
 174                         /*
 175                          * madvise() returns EAGAIN if kernel resources, such as
 176                          * slab, are temporarily unavailable.
 177                          */
 178                         if (error == -ENOMEM)
 179                                 error = -EAGAIN;
 180                         goto out;
 181                 }
 182         }
 183
 184 success:
 185         /*
 186          * vm_flags is protected by the mmap_sem held in write mode.
 187          */
 188         vma->vm_flags = new_flags;
 189 out:
 190         return error;
 191 }
 192
 193 #ifdef CONFIG_SWAP
 194 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 195         unsigned long end, struct mm_walk *walk)
 196 {
 197         pte_t *orig_pte;
 198         struct vm_area_struct *vma = walk->private;
 199         unsigned long index;
 200
 201         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 202                 return 0;
 203
 204         for (index = start; index != end; index += PAGE_SIZE) {
 205                 pte_t pte;
 206                 swp_entry_t entry;
 207                 struct page *page;
 208                 spinlock_t *ptl;
 209
 210                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 211                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 212                 pte_unmap_unlock(orig_pte, ptl);
 213
 214                 if (pte_present(pte) || pte_none(pte))
 215                         continue;
 216                 entry = pte_to_swp_entry(pte);
 217                 if (unlikely(non_swap_entry(entry)))
 218                         continue;
 219
 220                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 221                                                         vma, index, false);
 222                 if (page)
 223                         put_page(page);
 224         }
 225
 226         return 0;
 227 }
 228
 229 static void force_swapin_readahead(struct vm_area_struct *vma,
 230                 unsigned long start, unsigned long end)
 231 {
 232         struct mm_walk walk = {
 233                 .mm = vma->vm_mm,
 234                 .pmd_entry = swapin_walk_pmd_entry,
 235                 .private = vma,
 236         };
 237
 238         walk_page_range(start, end, &walk);
 239
 240         lru_add_drain();        /* Push any new pages onto the LRU now */
 241 }
 242
 243 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 244                 unsigned long start, unsigned long end,
 245                 struct address_space *mapping)
 246 {
 247         pgoff_t index;
 248         struct page *page;
 249         swp_entry_t swap;
 250
 251         for (; start < end; start += PAGE_SIZE) {
 252                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 253
 254                 page = find_get_entry(mapping, index);
 255                 if (!xa_is_value(page)) {
 256                         if (page)
 257                                 put_page(page);
 258                         continue;
 259                 }
 260                 swap = radix_to_swp_entry(page);
 261                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 262                                                         NULL, 0, false);
 263                 if (page)
 264                         put_page(page);
 265         }
 266
 267         lru_add_drain();        /* Push any new pages onto the LRU now */
 268 }
 269 #endif          /* CONFIG_SWAP */
 270
 271 /*
 272  * Schedule all required I/O operations.  Do not wait for completion.
 273  */
 274 static long madvise_willneed(struct vm_area_struct *vma,
 275                              struct vm_area_struct **prev,
 276                              unsigned long start, unsigned long end)
 277 {
 278         struct file *file = vma->vm_file;
 279
 280         *prev = vma;
 281 #ifdef CONFIG_SWAP
 282         if (!file) {
 283                 force_swapin_readahead(vma, start, end);
 284                 return 0;
 285         }
 286
 287         if (shmem_mapping(file->f_mapping)) {
 288                 force_shm_swapin_readahead(vma, start, end,
 289                                         file->f_mapping);
 290                 return 0;
 291         }
 292 #else
 293         if (!file)
 294                 return -EBADF;
 295 #endif
 296
 297         if (IS_DAX(file_inode(file))) {
 298                 /* no bad return value, but ignore advice */
 299                 return 0;
 300         }
 301
 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_change_page_size(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);
 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         struct mm_struct *mm = vma->vm_mm;
 463         struct mmu_notifier_range range;
 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         range.start = max(vma->vm_start, start_addr);
 471         if (range.start >= vma->vm_end)
 472                 return -EINVAL;
 473         range.end = min(vma->vm_end, end_addr);
 474         if (range.end <= vma->vm_start)
 475                 return -EINVAL;
 476         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 477                                 range.start, range.end);
 478
 479         lru_add_drain();
 480         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 481         update_hiwater_rss(mm);
 482
 483         mmu_notifier_invalidate_range_start(&range);
 484         madvise_free_page_range(&tlb, vma, range.start, range.end);
 485         mmu_notifier_invalidate_range_end(&range);
 486         tlb_finish_mmu(&tlb, range.start, range.end);
 487
 488         return 0;
 489 }
 490
 491 /*
 492  * Application no longer needs these pages.  If the pages are dirty,
 493  * it's OK to just throw them away.  The app will be more careful about
 494  * data it wants to keep.  Be sure to free swap resources too.  The
 495  * zap_page_range call sets things up for shrink_active_list to actually free
 496  * these pages later if no one else has touched them in the meantime,
 497  * although we could add these pages to a global reuse list for
 498  * shrink_active_list to pick up before reclaiming other pages.
 499  *
 500  * NB: This interface discards data rather than pushes it out to swap,
 501  * as some implementations do.  This has performance implications for
 502  * applications like large transactional databases which want to discard
 503  * pages in anonymous maps after committing to backing store the data
 504  * that was kept in them.  There is no reason to write this data out to
 505  * the swap area if the application is discarding it.
 506  *
 507  * An interface that causes the system to free clean pages and flush
 508  * dirty pages is already available as msync(MS_INVALIDATE).
 509  */
 510 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 511                                         unsigned long start, unsigned long end)
 512 {
 513         zap_page_range(vma, start, end - start);
 514         return 0;
 515 }
 516
 517 static long madvise_dontneed_free(struct vm_area_struct *vma,
 518                                   struct vm_area_struct **prev,
 519                                   unsigned long start, unsigned long end,
 520                                   int behavior)
 521 {
 522         *prev = vma;
 523         if (!can_madv_dontneed_vma(vma))
 524                 return -EINVAL;
 525
 526         if (!userfaultfd_remove(vma, start, end)) {
 527                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 528
 529                 down_read(&current->mm->mmap_sem);
 530                 vma = find_vma(current->mm, start);
 531                 if (!vma)
 532                         return -ENOMEM;
 533                 if (start < vma->vm_start) {
 534                         /*
 535                          * This "vma" under revalidation is the one
 536                          * with the lowest vma->vm_start where start
 537                          * is also < vma->vm_end. If start <
 538                          * vma->vm_start it means an hole materialized
 539                          * in the user address space within the
 540                          * virtual range passed to MADV_DONTNEED
 541                          * or MADV_FREE.
 542                          */
 543                         return -ENOMEM;
 544                 }
 545                 if (!can_madv_dontneed_vma(vma))
 546                         return -EINVAL;
 547                 if (end > vma->vm_end) {
 548                         /*
 549                          * Don't fail if end > vma->vm_end. If the old
 550                          * vma was splitted while the mmap_sem was
 551                          * released the effect of the concurrent
 552                          * operation may not cause madvise() to
 553                          * have an undefined result. There may be an
 554                          * adjacent next vma that we'll walk
 555                          * next. userfaultfd_remove() will generate an
 556                          * UFFD_EVENT_REMOVE repetition on the
 557                          * end-vma->vm_end range, but the manager can
 558                          * handle a repetition fine.
 559                          */
 560                         end = vma->vm_end;
 561                 }
 562                 VM_WARN_ON(start >= end);
 563         }
 564
 565         if (behavior == MADV_DONTNEED)
 566                 return madvise_dontneed_single_vma(vma, start, end);
 567         else if (behavior == MADV_FREE)
 568                 return madvise_free_single_vma(vma, start, end);
 569         else
 570                 return -EINVAL;
 571 }
 572
 573 /*
 574  * Application wants to free up the pages and associated backing store.
 575  * This is effectively punching a hole into the middle of a file.
 576  */
 577 static long madvise_remove(struct vm_area_struct *vma,
 578                                 struct vm_area_struct **prev,
 579                                 unsigned long start, unsigned long end)
 580 {
 581         loff_t offset;
 582         int error;
 583         struct file *f;
 584
 585         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 586
 587         if (vma->vm_flags & VM_LOCKED)
 588                 return -EINVAL;
 589
 590         f = vma->vm_file;
 591
 592         if (!f || !f->f_mapping || !f->f_mapping->host) {
 593                         return -EINVAL;
 594         }
 595
 596         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 597                 return -EACCES;
 598
 599         offset = (loff_t)(start - vma->vm_start)
 600                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 601
 602         /*
 603          * Filesystem's fallocate may need to take i_mutex.  We need to
 604          * explicitly grab a reference because the vma (and hence the
 605          * vma's reference to the file) can go away as soon as we drop
 606          * mmap_sem.
 607          */
 608         get_file(f);
 609         if (userfaultfd_remove(vma, start, end)) {
 610                 /* mmap_sem was not released by userfaultfd_remove() */
 611                 up_read(&current->mm->mmap_sem);
 612         }
 613         error = vfs_fallocate(f,
 614                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 615                                 offset, end - start);
 616         fput(f);
 617         down_read(&current->mm->mmap_sem);
 618         return error;
 619 }
 620
 621 #ifdef CONFIG_MEMORY_FAILURE
 622 /*
 623  * Error injection support for memory error handling.
 624  */
 625 static int madvise_inject_error(int behavior,
 626                 unsigned long start, unsigned long end)
 627 {
 628         struct page *page;
 629         struct zone *zone;
 630         unsigned int order;
 631
 632         if (!capable(CAP_SYS_ADMIN))
 633                 return -EPERM;
 634
 635
 636         for (; start < end; start += PAGE_SIZE << order) {
 637                 unsigned long pfn;
 638                 int ret;
 639
 640                 ret = get_user_pages_fast(start, 1, 0, &page);
 641                 if (ret != 1)
 642                         return ret;
 643                 pfn = page_to_pfn(page);
 644
 645                 /*
 646                  * When soft offlining hugepages, after migrating the page
 647                  * we dissolve it, therefore in the second loop "page" will
 648                  * no longer be a compound page, and order will be 0.
 649                  */
 650                 order = compound_order(compound_head(page));
 651
 652                 if (PageHWPoison(page)) {
 653                         put_page(page);
 654                         continue;
 655                 }
 656
 657                 if (behavior == MADV_SOFT_OFFLINE) {
 658                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 659                                         pfn, start);
 660
 661                         ret = soft_offline_page(page, MF_COUNT_INCREASED);
 662                         if (ret)
 663                                 return ret;
 664                         continue;
 665                 }
 666
 667                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 668                                 pfn, start);
 669
 670                 /*
 671                  * Drop the page reference taken by get_user_pages_fast(). In
 672                  * the absence of MF_COUNT_INCREASED the memory_failure()
 673                  * routine is responsible for pinning the page to prevent it
 674                  * from being released back to the page allocator.
 675                  */
 676                 put_page(page);
 677                 ret = memory_failure(pfn, 0);
 678                 if (ret)
 679                         return ret;
 680         }
 681
 682         /* Ensure that all poisoned pages are removed from per-cpu lists */
 683         for_each_populated_zone(zone)
 684                 drain_all_pages(zone);
 685
 686         return 0;
 687 }
 688 #endif
 689
 690 static long
 691 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 692                 unsigned long start, unsigned long end, int behavior)
 693 {
 694         switch (behavior) {
 695         case MADV_REMOVE:
 696                 return madvise_remove(vma, prev, start, end);
 697         case MADV_WILLNEED:
 698                 return madvise_willneed(vma, prev, start, end);
 699         case MADV_FREE:
 700         case MADV_DONTNEED:
 701                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 702         default:
 703                 return madvise_behavior(vma, prev, start, end, behavior);
 704         }
 705 }
 706
 707 static bool
 708 madvise_behavior_valid(int behavior)
 709 {
 710         switch (behavior) {
 711         case MADV_DOFORK:
 712         case MADV_DONTFORK:
 713         case MADV_NORMAL:
 714         case MADV_SEQUENTIAL:
 715         case MADV_RANDOM:
 716         case MADV_REMOVE:
 717         case MADV_WILLNEED:
 718         case MADV_DONTNEED:
 719         case MADV_FREE:
 720 #ifdef CONFIG_KSM
 721         case MADV_MERGEABLE:
 722         case MADV_UNMERGEABLE:
 723 #endif
 724 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 725         case MADV_HUGEPAGE:
 726         case MADV_NOHUGEPAGE:
 727 #endif
 728         case MADV_DONTDUMP:
 729         case MADV_DODUMP:
 730         case MADV_WIPEONFORK:
 731         case MADV_KEEPONFORK:
 732 #ifdef CONFIG_MEMORY_FAILURE
 733         case MADV_SOFT_OFFLINE:
 734         case MADV_HWPOISON:
 735 #endif
 736                 return true;
 737
 738         default:
 739                 return false;
 740         }
 741 }
 742
 743 /*
 744  * The madvise(2) system call.
 745  *
 746  * Applications can use madvise() to advise the kernel how it should
 747  * handle paging I/O in this VM area.  The idea is to help the kernel
 748  * use appropriate read-ahead and caching techniques.  The information
 749  * provided is advisory only, and can be safely disregarded by the
 750  * kernel without affecting the correct operation of the application.
 751  *
 752  * behavior values:
 753  *  MADV_NORMAL - the default behavior is to read clusters.  This
 754  *              results in some read-ahead and read-behind.
 755  *  MADV_RANDOM - the system should read the minimum amount of data
 756  *              on any access, since it is unlikely that the appli-
 757  *              cation will need more than what it asks for.
 758  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 759  *              once, so they can be aggressively read ahead, and
 760  *              can be freed soon after they are accessed.
 761  *  MADV_WILLNEED - the application is notifying the system to read
 762  *              some pages ahead.
 763  *  MADV_DONTNEED - the application is finished with the given range,
 764  *              so the kernel can free resources associated with it.
 765  *  MADV_FREE - the application marks pages in the given range as lazy free,
 766  *              where actual purges are postponed until memory pressure happens.
 767  *  MADV_REMOVE - the application wants to free up the given range of
 768  *              pages and associated backing store.
 769  *  MADV_DONTFORK - omit this area from child's address space when forking:
 770  *              typically, to avoid COWing pages pinned by get_user_pages().
 771  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 772  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
 773  *              range after a fork.
 774  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
 775  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 776  *              were corrupted by unrecoverable hardware memory failure.
 777  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 778  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 779  *              this area with pages of identical content from other such areas.
 780  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 781  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 782  *              huge pages in the future. Existing pages might be coalesced and
 783  *              new pages might be allocated as THP.
 784  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 785  *              transparent huge pages so the existing pages will not be
 786  *              coalesced into THP and new pages will not be allocated as THP.
 787  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 788  *              from being included in its core dump.
 789  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 790  *
 791  * return values:
 792  *  zero    - success
 793  *  -EINVAL - start + len < 0, start is not page-aligned,
 794  *              "behavior" is not a valid value, or application
 795  *              is attempting to release locked or shared pages,
 796  *              or the specified address range includes file, Huge TLB,
 797  *              MAP_SHARED or VMPFNMAP range.
 798  *  -ENOMEM - addresses in the specified range are not currently
 799  *              mapped, or are outside the AS of the process.
 800  *  -EIO    - an I/O error occurred while paging in data.
 801  *  -EBADF  - map exists, but area maps something that isn't a file.
 802  *  -EAGAIN - a kernel resource was temporarily unavailable.
 803  */
 804 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 805 {
 806         unsigned long end, tmp;
 807         struct vm_area_struct *vma, *prev;
 808         int unmapped_error = 0;
 809         int error = -EINVAL;
 810         int write;
 811         size_t len;
 812         struct blk_plug plug;
 813
 814         if (!madvise_behavior_valid(behavior))
 815                 return error;
 816
 817         if (start & ~PAGE_MASK)
 818                 return error;
 819         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 820
 821         /* Check to see whether len was rounded up from small -ve to zero */
 822         if (len_in && !len)
 823                 return error;
 824
 825         end = start + len;
 826         if (end < start)
 827                 return error;
 828
 829         error = 0;
 830         if (end == start)
 831                 return error;
 832
 833 #ifdef CONFIG_MEMORY_FAILURE
 834         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 835                 return madvise_inject_error(behavior, start, start + len_in);
 836 #endif
 837
 838         write = madvise_need_mmap_write(behavior);
 839         if (write) {
 840                 if (down_write_killable(&current->mm->mmap_sem))
 841                         return -EINTR;
 842         } else {
 843                 down_read(&current->mm->mmap_sem);
 844         }
 845
 846         /*
 847          * If the interval [start,end) covers some unmapped address
 848          * ranges, just ignore them, but return -ENOMEM at the end.
 849          * - different from the way of handling in mlock etc.
 850          */
 851         vma = find_vma_prev(current->mm, start, &prev);
 852         if (vma && start > vma->vm_start)
 853                 prev = vma;
 854
 855         blk_start_plug(&plug);
 856         for (;;) {
 857                 /* Still start < end. */
 858                 error = -ENOMEM;
 859                 if (!vma)
 860                         goto out;
 861
 862                 /* Here start < (end|vma->vm_end). */
 863                 if (start < vma->vm_start) {
 864                         unmapped_error = -ENOMEM;
 865                         start = vma->vm_start;
 866                         if (start >= end)
 867                                 goto out;
 868                 }
 869
 870                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 871                 tmp = vma->vm_end;
 872                 if (end < tmp)
 873                         tmp = end;
 874
 875                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 876                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 877                 if (error)
 878                         goto out;
 879                 start = tmp;
 880                 if (prev && start < prev->vm_end)
 881                         start = prev->vm_end;
 882                 error = unmapped_error;
 883                 if (start >= end)
 884                         goto out;
 885                 if (prev)
 886                         vma = prev->vm_next;
 887                 else    /* madvise_remove dropped mmap_sem */
 888                         vma = find_vma(current->mm, start);
 889         }
 890 out:
 891         blk_finish_plug(&plug);
 892         if (write)
 893                 up_write(&current->mm->mmap_sem);
 894         else
 895                 up_read(&current->mm->mmap_sem);
 896
 897         return error;
 898 }