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 const struct mm_walk_ops swapin_walk_ops = {
 230         .pmd_entry              = swapin_walk_pmd_entry,
 231 };
 232
 233 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 234                 unsigned long start, unsigned long end,
 235                 struct address_space *mapping)
 236 {
 237         pgoff_t index;
 238         struct page *page;
 239         swp_entry_t swap;
 240
 241         for (; start < end; start += PAGE_SIZE) {
 242                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 243
 244                 page = find_get_entry(mapping, index);
 245                 if (!xa_is_value(page)) {
 246                         if (page)
 247                                 put_page(page);
 248                         continue;
 249                 }
 250                 swap = radix_to_swp_entry(page);
 251                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 252                                                         NULL, 0, false);
 253                 if (page)
 254                         put_page(page);
 255         }
 256
 257         lru_add_drain();        /* Push any new pages onto the LRU now */
 258 }
 259 #endif          /* CONFIG_SWAP */
 260
 261 /*
 262  * Schedule all required I/O operations.  Do not wait for completion.
 263  */
 264 static long madvise_willneed(struct vm_area_struct *vma,
 265                              struct vm_area_struct **prev,
 266                              unsigned long start, unsigned long end)
 267 {
 268         struct file *file = vma->vm_file;
 269
 270         *prev = vma;
 271 #ifdef CONFIG_SWAP
 272         if (!file) {
 273                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 274                 lru_add_drain(); /* Push any new pages onto the LRU now */
 275                 return 0;
 276         }
 277
 278         if (shmem_mapping(file->f_mapping)) {
 279                 force_shm_swapin_readahead(vma, start, end,
 280                                         file->f_mapping);
 281                 return 0;
 282         }
 283 #else
 284         if (!file)
 285                 return -EBADF;
 286 #endif
 287
 288         if (IS_DAX(file_inode(file))) {
 289                 /* no bad return value, but ignore advice */
 290                 return 0;
 291         }
 292
 293         start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 294         if (end > vma->vm_end)
 295                 end = vma->vm_end;
 296         end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 297
 298         force_page_cache_readahead(file->f_mapping, file, start, end - start);
 299         return 0;
 300 }
 301
 302 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 303                                 unsigned long end, struct mm_walk *walk)
 304
 305 {
 306         struct mmu_gather *tlb = walk->private;
 307         struct mm_struct *mm = tlb->mm;
 308         struct vm_area_struct *vma = walk->vma;
 309         spinlock_t *ptl;
 310         pte_t *orig_pte, *pte, ptent;
 311         struct page *page;
 312         int nr_swap = 0;
 313         unsigned long next;
 314
 315         next = pmd_addr_end(addr, end);
 316         if (pmd_trans_huge(*pmd))
 317                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 318                         goto next;
 319
 320         if (pmd_trans_unstable(pmd))
 321                 return 0;
 322
 323         tlb_change_page_size(tlb, PAGE_SIZE);
 324         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 325         flush_tlb_batched_pending(mm);
 326         arch_enter_lazy_mmu_mode();
 327         for (; addr != end; pte++, addr += PAGE_SIZE) {
 328                 ptent = *pte;
 329
 330                 if (pte_none(ptent))
 331                         continue;
 332                 /*
 333                  * If the pte has swp_entry, just clear page table to
 334                  * prevent swap-in which is more expensive rather than
 335                  * (page allocation + zeroing).
 336                  */
 337                 if (!pte_present(ptent)) {
 338                         swp_entry_t entry;
 339
 340                         entry = pte_to_swp_entry(ptent);
 341                         if (non_swap_entry(entry))
 342                                 continue;
 343                         nr_swap--;
 344                         free_swap_and_cache(entry);
 345                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 346                         continue;
 347                 }
 348
 349                 page = vm_normal_page(vma, addr, ptent);
 350                 if (!page)
 351                         continue;
 352
 353                 /*
 354                  * If pmd isn't transhuge but the page is THP and
 355                  * is owned by only this process, split it and
 356                  * deactivate all pages.
 357                  */
 358                 if (PageTransCompound(page)) {
 359                         if (page_mapcount(page) != 1)
 360                                 goto out;
 361                         get_page(page);
 362                         if (!trylock_page(page)) {
 363                                 put_page(page);
 364                                 goto out;
 365                         }
 366                         pte_unmap_unlock(orig_pte, ptl);
 367                         if (split_huge_page(page)) {
 368                                 unlock_page(page);
 369                                 put_page(page);
 370                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 371                                 goto out;
 372                         }
 373                         unlock_page(page);
 374                         put_page(page);
 375                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 376                         pte--;
 377                         addr -= PAGE_SIZE;
 378                         continue;
 379                 }
 380
 381                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 382
 383                 if (PageSwapCache(page) || PageDirty(page)) {
 384                         if (!trylock_page(page))
 385                                 continue;
 386                         /*
 387                          * If page is shared with others, we couldn't clear
 388                          * PG_dirty of the page.
 389                          */
 390                         if (page_mapcount(page) != 1) {
 391                                 unlock_page(page);
 392                                 continue;
 393                         }
 394
 395                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 396                                 unlock_page(page);
 397                                 continue;
 398                         }
 399
 400                         ClearPageDirty(page);
 401                         unlock_page(page);
 402                 }
 403
 404                 if (pte_young(ptent) || pte_dirty(ptent)) {
 405                         /*
 406                          * Some of architecture(ex, PPC) don't update TLB
 407                          * with set_pte_at and tlb_remove_tlb_entry so for
 408                          * the portability, remap the pte with old|clean
 409                          * after pte clearing.
 410                          */
 411                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 412                                                         tlb->fullmm);
 413
 414                         ptent = pte_mkold(ptent);
 415                         ptent = pte_mkclean(ptent);
 416                         set_pte_at(mm, addr, pte, ptent);
 417                         tlb_remove_tlb_entry(tlb, pte, addr);
 418                 }
 419                 mark_page_lazyfree(page);
 420         }
 421 out:
 422         if (nr_swap) {
 423                 if (current->mm == mm)
 424                         sync_mm_rss(mm);
 425
 426                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 427         }
 428         arch_leave_lazy_mmu_mode();
 429         pte_unmap_unlock(orig_pte, ptl);
 430         cond_resched();
 431 next:
 432         return 0;
 433 }
 434
 435 static const struct mm_walk_ops madvise_free_walk_ops = {
 436         .pmd_entry              = madvise_free_pte_range,
 437 };
 438
 439 static int madvise_free_single_vma(struct vm_area_struct *vma,
 440                         unsigned long start_addr, unsigned long end_addr)
 441 {
 442         struct mm_struct *mm = vma->vm_mm;
 443         struct mmu_notifier_range range;
 444         struct mmu_gather tlb;
 445
 446         /* MADV_FREE works for only anon vma at the moment */
 447         if (!vma_is_anonymous(vma))
 448                 return -EINVAL;
 449
 450         range.start = max(vma->vm_start, start_addr);
 451         if (range.start >= vma->vm_end)
 452                 return -EINVAL;
 453         range.end = min(vma->vm_end, end_addr);
 454         if (range.end <= vma->vm_start)
 455                 return -EINVAL;
 456         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 457                                 range.start, range.end);
 458
 459         lru_add_drain();
 460         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 461         update_hiwater_rss(mm);
 462
 463         mmu_notifier_invalidate_range_start(&range);
 464         tlb_start_vma(&tlb, vma);
 465         walk_page_range(vma->vm_mm, range.start, range.end,
 466                         &madvise_free_walk_ops, &tlb);
 467         tlb_end_vma(&tlb, vma);
 468         mmu_notifier_invalidate_range_end(&range);
 469         tlb_finish_mmu(&tlb, range.start, range.end);
 470
 471         return 0;
 472 }
 473
 474 /*
 475  * Application no longer needs these pages.  If the pages are dirty,
 476  * it's OK to just throw them away.  The app will be more careful about
 477  * data it wants to keep.  Be sure to free swap resources too.  The
 478  * zap_page_range call sets things up for shrink_active_list to actually free
 479  * these pages later if no one else has touched them in the meantime,
 480  * although we could add these pages to a global reuse list for
 481  * shrink_active_list to pick up before reclaiming other pages.
 482  *
 483  * NB: This interface discards data rather than pushes it out to swap,
 484  * as some implementations do.  This has performance implications for
 485  * applications like large transactional databases which want to discard
 486  * pages in anonymous maps after committing to backing store the data
 487  * that was kept in them.  There is no reason to write this data out to
 488  * the swap area if the application is discarding it.
 489  *
 490  * An interface that causes the system to free clean pages and flush
 491  * dirty pages is already available as msync(MS_INVALIDATE).
 492  */
 493 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 494                                         unsigned long start, unsigned long end)
 495 {
 496         zap_page_range(vma, start, end - start);
 497         return 0;
 498 }
 499
 500 static long madvise_dontneed_free(struct vm_area_struct *vma,
 501                                   struct vm_area_struct **prev,
 502                                   unsigned long start, unsigned long end,
 503                                   int behavior)
 504 {
 505         *prev = vma;
 506         if (!can_madv_dontneed_vma(vma))
 507                 return -EINVAL;
 508
 509         if (!userfaultfd_remove(vma, start, end)) {
 510                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 511
 512                 down_read(&current->mm->mmap_sem);
 513                 vma = find_vma(current->mm, start);
 514                 if (!vma)
 515                         return -ENOMEM;
 516                 if (start < vma->vm_start) {
 517                         /*
 518                          * This "vma" under revalidation is the one
 519                          * with the lowest vma->vm_start where start
 520                          * is also < vma->vm_end. If start <
 521                          * vma->vm_start it means an hole materialized
 522                          * in the user address space within the
 523                          * virtual range passed to MADV_DONTNEED
 524                          * or MADV_FREE.
 525                          */
 526                         return -ENOMEM;
 527                 }
 528                 if (!can_madv_dontneed_vma(vma))
 529                         return -EINVAL;
 530                 if (end > vma->vm_end) {
 531                         /*
 532                          * Don't fail if end > vma->vm_end. If the old
 533                          * vma was splitted while the mmap_sem was
 534                          * released the effect of the concurrent
 535                          * operation may not cause madvise() to
 536                          * have an undefined result. There may be an
 537                          * adjacent next vma that we'll walk
 538                          * next. userfaultfd_remove() will generate an
 539                          * UFFD_EVENT_REMOVE repetition on the
 540                          * end-vma->vm_end range, but the manager can
 541                          * handle a repetition fine.
 542                          */
 543                         end = vma->vm_end;
 544                 }
 545                 VM_WARN_ON(start >= end);
 546         }
 547
 548         if (behavior == MADV_DONTNEED)
 549                 return madvise_dontneed_single_vma(vma, start, end);
 550         else if (behavior == MADV_FREE)
 551                 return madvise_free_single_vma(vma, start, end);
 552         else
 553                 return -EINVAL;
 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         struct zone *zone;
 613         unsigned int order;
 614
 615         if (!capable(CAP_SYS_ADMIN))
 616                 return -EPERM;
 617
 618
 619         for (; start < end; start += PAGE_SIZE << order) {
 620                 unsigned long pfn;
 621                 int ret;
 622
 623                 ret = get_user_pages_fast(start, 1, 0, &page);
 624                 if (ret != 1)
 625                         return ret;
 626                 pfn = page_to_pfn(page);
 627
 628                 /*
 629                  * When soft offlining hugepages, after migrating the page
 630                  * we dissolve it, therefore in the second loop "page" will
 631                  * no longer be a compound page, and order will be 0.
 632                  */
 633                 order = compound_order(compound_head(page));
 634
 635                 if (PageHWPoison(page)) {
 636                         put_page(page);
 637                         continue;
 638                 }
 639
 640                 if (behavior == MADV_SOFT_OFFLINE) {
 641                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 642                                         pfn, start);
 643
 644                         ret = soft_offline_page(page, MF_COUNT_INCREASED);
 645                         if (ret)
 646                                 return ret;
 647                         continue;
 648                 }
 649
 650                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 651                                 pfn, start);
 652
 653                 /*
 654                  * Drop the page reference taken by get_user_pages_fast(). In
 655                  * the absence of MF_COUNT_INCREASED the memory_failure()
 656                  * routine is responsible for pinning the page to prevent it
 657                  * from being released back to the page allocator.
 658                  */
 659                 put_page(page);
 660                 ret = memory_failure(pfn, 0);
 661                 if (ret)
 662                         return ret;
 663         }
 664
 665         /* Ensure that all poisoned pages are removed from per-cpu lists */
 666         for_each_populated_zone(zone)
 667                 drain_all_pages(zone);
 668
 669         return 0;
 670 }
 671 #endif
 672
 673 static long
 674 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 675                 unsigned long start, unsigned long end, int behavior)
 676 {
 677         switch (behavior) {
 678         case MADV_REMOVE:
 679                 return madvise_remove(vma, prev, start, end);
 680         case MADV_WILLNEED:
 681                 return madvise_willneed(vma, prev, start, end);
 682         case MADV_FREE:
 683         case MADV_DONTNEED:
 684                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 685         default:
 686                 return madvise_behavior(vma, prev, start, end, behavior);
 687         }
 688 }
 689
 690 static bool
 691 madvise_behavior_valid(int behavior)
 692 {
 693         switch (behavior) {
 694         case MADV_DOFORK:
 695         case MADV_DONTFORK:
 696         case MADV_NORMAL:
 697         case MADV_SEQUENTIAL:
 698         case MADV_RANDOM:
 699         case MADV_REMOVE:
 700         case MADV_WILLNEED:
 701         case MADV_DONTNEED:
 702         case MADV_FREE:
 703 #ifdef CONFIG_KSM
 704         case MADV_MERGEABLE:
 705         case MADV_UNMERGEABLE:
 706 #endif
 707 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 708         case MADV_HUGEPAGE:
 709         case MADV_NOHUGEPAGE:
 710 #endif
 711         case MADV_DONTDUMP:
 712         case MADV_DODUMP:
 713         case MADV_WIPEONFORK:
 714         case MADV_KEEPONFORK:
 715 #ifdef CONFIG_MEMORY_FAILURE
 716         case MADV_SOFT_OFFLINE:
 717         case MADV_HWPOISON:
 718 #endif
 719                 return true;
 720
 721         default:
 722                 return false;
 723         }
 724 }
 725
 726 /*
 727  * The madvise(2) system call.
 728  *
 729  * Applications can use madvise() to advise the kernel how it should
 730  * handle paging I/O in this VM area.  The idea is to help the kernel
 731  * use appropriate read-ahead and caching techniques.  The information
 732  * provided is advisory only, and can be safely disregarded by the
 733  * kernel without affecting the correct operation of the application.
 734  *
 735  * behavior values:
 736  *  MADV_NORMAL - the default behavior is to read clusters.  This
 737  *              results in some read-ahead and read-behind.
 738  *  MADV_RANDOM - the system should read the minimum amount of data
 739  *              on any access, since it is unlikely that the appli-
 740  *              cation will need more than what it asks for.
 741  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 742  *              once, so they can be aggressively read ahead, and
 743  *              can be freed soon after they are accessed.
 744  *  MADV_WILLNEED - the application is notifying the system to read
 745  *              some pages ahead.
 746  *  MADV_DONTNEED - the application is finished with the given range,
 747  *              so the kernel can free resources associated with it.
 748  *  MADV_FREE - the application marks pages in the given range as lazy free,
 749  *              where actual purges are postponed until memory pressure happens.
 750  *  MADV_REMOVE - the application wants to free up the given range of
 751  *              pages and associated backing store.
 752  *  MADV_DONTFORK - omit this area from child's address space when forking:
 753  *              typically, to avoid COWing pages pinned by get_user_pages().
 754  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 755  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
 756  *              range after a fork.
 757  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
 758  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 759  *              were corrupted by unrecoverable hardware memory failure.
 760  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 761  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 762  *              this area with pages of identical content from other such areas.
 763  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 764  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 765  *              huge pages in the future. Existing pages might be coalesced and
 766  *              new pages might be allocated as THP.
 767  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 768  *              transparent huge pages so the existing pages will not be
 769  *              coalesced into THP and new pages will not be allocated as THP.
 770  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 771  *              from being included in its core dump.
 772  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 773  *
 774  * return values:
 775  *  zero    - success
 776  *  -EINVAL - start + len < 0, start is not page-aligned,
 777  *              "behavior" is not a valid value, or application
 778  *              is attempting to release locked or shared pages,
 779  *              or the specified address range includes file, Huge TLB,
 780  *              MAP_SHARED or VMPFNMAP range.
 781  *  -ENOMEM - addresses in the specified range are not currently
 782  *              mapped, or are outside the AS of the process.
 783  *  -EIO    - an I/O error occurred while paging in data.
 784  *  -EBADF  - map exists, but area maps something that isn't a file.
 785  *  -EAGAIN - a kernel resource was temporarily unavailable.
 786  */
 787 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 788 {
 789         unsigned long end, tmp;
 790         struct vm_area_struct *vma, *prev;
 791         int unmapped_error = 0;
 792         int error = -EINVAL;
 793         int write;
 794         size_t len;
 795         struct blk_plug plug;
 796
 797         if (!madvise_behavior_valid(behavior))
 798                 return error;
 799
 800         if (start & ~PAGE_MASK)
 801                 return error;
 802         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 803
 804         /* Check to see whether len was rounded up from small -ve to zero */
 805         if (len_in && !len)
 806                 return error;
 807
 808         end = start + len;
 809         if (end < start)
 810                 return error;
 811
 812         error = 0;
 813         if (end == start)
 814                 return error;
 815
 816 #ifdef CONFIG_MEMORY_FAILURE
 817         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 818                 return madvise_inject_error(behavior, start, start + len_in);
 819 #endif
 820
 821         write = madvise_need_mmap_write(behavior);
 822         if (write) {
 823                 if (down_write_killable(&current->mm->mmap_sem))
 824                         return -EINTR;
 825         } else {
 826                 down_read(&current->mm->mmap_sem);
 827         }
 828
 829         /*
 830          * If the interval [start,end) covers some unmapped address
 831          * ranges, just ignore them, but return -ENOMEM at the end.
 832          * - different from the way of handling in mlock etc.
 833          */
 834         vma = find_vma_prev(current->mm, start, &prev);
 835         if (vma && start > vma->vm_start)
 836                 prev = vma;
 837
 838         blk_start_plug(&plug);
 839         for (;;) {
 840                 /* Still start < end. */
 841                 error = -ENOMEM;
 842                 if (!vma)
 843                         goto out;
 844
 845                 /* Here start < (end|vma->vm_end). */
 846                 if (start < vma->vm_start) {
 847                         unmapped_error = -ENOMEM;
 848                         start = vma->vm_start;
 849                         if (start >= end)
 850                                 goto out;
 851                 }
 852
 853                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
 854                 tmp = vma->vm_end;
 855                 if (end < tmp)
 856                         tmp = end;
 857
 858                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
 859                 error = madvise_vma(vma, &prev, start, tmp, behavior);
 860                 if (error)
 861                         goto out;
 862                 start = tmp;
 863                 if (prev && start < prev->vm_end)
 864                         start = prev->vm_end;
 865                 error = unmapped_error;
 866                 if (start >= end)
 867                         goto out;
 868                 if (prev)
 869                         vma = prev->vm_next;
 870                 else    /* madvise_remove dropped mmap_sem */
 871                         vma = find_vma(current->mm, start);
 872         }
 873 out:
 874         blk_finish_plug(&plug);
 875         if (write)
 876                 up_write(&current->mm->mmap_sem);
 877         else
 878                 up_read(&current->mm->mmap_sem);
 879
 880         return error;
 881 }