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/page_idle.h>
  15 #include <linux/userfaultfd_k.h>
  16 #include <linux/hugetlb.h>
  17 #include <linux/falloc.h>
  18 #include <linux/fadvise.h>
  19 #include <linux/sched.h>
  20 #include <linux/ksm.h>
  21 #include <linux/fs.h>
  22 #include <linux/file.h>
  23 #include <linux/blkdev.h>
  24 #include <linux/backing-dev.h>
  25 #include <linux/pagewalk.h>
  26 #include <linux/swap.h>
  27 #include <linux/swapops.h>
  28 #include <linux/shmem_fs.h>
  29 #include <linux/mmu_notifier.h>
  30
  31 #include <asm/tlb.h>
  32
  33 #include "internal.h"
  34
  35 /*
  36  * Any behaviour which results in changes to the vma->vm_flags needs to
  37  * take mmap_sem for writing. Others, which simply traverse vmas, need
  38  * to only take it for reading.
  39  */
  40 static int madvise_need_mmap_write(int behavior)
  41 {
  42         switch (behavior) {
  43         case MADV_REMOVE:
  44         case MADV_WILLNEED:
  45         case MADV_DONTNEED:
  46         case MADV_COLD:
  47         case MADV_FREE:
  48                 return 0;
  49         default:
  50                 /* be safe, default to 1. list exceptions explicitly */
  51                 return 1;
  52         }
  53 }
  54
  55 /*
  56  * We can potentially split a vm area into separate
  57  * areas, each area with its own behavior.
  58  */
  59 static long madvise_behavior(struct vm_area_struct *vma,
  60                      struct vm_area_struct **prev,
  61                      unsigned long start, unsigned long end, int behavior)
  62 {
  63         struct mm_struct *mm = vma->vm_mm;
  64         int error = 0;
  65         pgoff_t pgoff;
  66         unsigned long new_flags = vma->vm_flags;
  67
  68         switch (behavior) {
  69         case MADV_NORMAL:
  70                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  71                 break;
  72         case MADV_SEQUENTIAL:
  73                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  74                 break;
  75         case MADV_RANDOM:
  76                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  77                 break;
  78         case MADV_DONTFORK:
  79                 new_flags |= VM_DONTCOPY;
  80                 break;
  81         case MADV_DOFORK:
  82                 if (vma->vm_flags & VM_IO) {
  83                         error = -EINVAL;
  84                         goto out;
  85                 }
  86                 new_flags &= ~VM_DONTCOPY;
  87                 break;
  88         case MADV_WIPEONFORK:
  89                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  90                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  91                         error = -EINVAL;
  92                         goto out;
  93                 }
  94                 new_flags |= VM_WIPEONFORK;
  95                 break;
  96         case MADV_KEEPONFORK:
  97                 new_flags &= ~VM_WIPEONFORK;
  98                 break;
  99         case MADV_DONTDUMP:
 100                 new_flags |= VM_DONTDUMP;
 101                 break;
 102         case MADV_DODUMP:
 103                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 104                         error = -EINVAL;
 105                         goto out;
 106                 }
 107                 new_flags &= ~VM_DONTDUMP;
 108                 break;
 109         case MADV_MERGEABLE:
 110         case MADV_UNMERGEABLE:
 111                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 112                 if (error)
 113                         goto out_convert_errno;
 114                 break;
 115         case MADV_HUGEPAGE:
 116         case MADV_NOHUGEPAGE:
 117                 error = hugepage_madvise(vma, &new_flags, behavior);
 118                 if (error)
 119                         goto out_convert_errno;
 120                 break;
 121         }
 122
 123         if (new_flags == vma->vm_flags) {
 124                 *prev = vma;
 125                 goto out;
 126         }
 127
 128         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 129         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 130                           vma->vm_file, pgoff, vma_policy(vma),
 131                           vma->vm_userfaultfd_ctx);
 132         if (*prev) {
 133                 vma = *prev;
 134                 goto success;
 135         }
 136
 137         *prev = vma;
 138
 139         if (start != vma->vm_start) {
 140                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 141                         error = -ENOMEM;
 142                         goto out;
 143                 }
 144                 error = __split_vma(mm, vma, start, 1);
 145                 if (error)
 146                         goto out_convert_errno;
 147         }
 148
 149         if (end != vma->vm_end) {
 150                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 151                         error = -ENOMEM;
 152                         goto out;
 153                 }
 154                 error = __split_vma(mm, vma, end, 0);
 155                 if (error)
 156                         goto out_convert_errno;
 157         }
 158
 159 success:
 160         /*
 161          * vm_flags is protected by the mmap_sem held in write mode.
 162          */
 163         vma->vm_flags = new_flags;
 164
 165 out_convert_errno:
 166         /*
 167          * madvise() returns EAGAIN if kernel resources, such as
 168          * slab, are temporarily unavailable.
 169          */
 170         if (error == -ENOMEM)
 171                 error = -EAGAIN;
 172 out:
 173         return error;
 174 }
 175
 176 #ifdef CONFIG_SWAP
 177 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 178         unsigned long end, struct mm_walk *walk)
 179 {
 180         pte_t *orig_pte;
 181         struct vm_area_struct *vma = walk->private;
 182         unsigned long index;
 183
 184         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 185                 return 0;
 186
 187         for (index = start; index != end; index += PAGE_SIZE) {
 188                 pte_t pte;
 189                 swp_entry_t entry;
 190                 struct page *page;
 191                 spinlock_t *ptl;
 192
 193                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 194                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 195                 pte_unmap_unlock(orig_pte, ptl);
 196
 197                 if (pte_present(pte) || pte_none(pte))
 198                         continue;
 199                 entry = pte_to_swp_entry(pte);
 200                 if (unlikely(non_swap_entry(entry)))
 201                         continue;
 202
 203                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 204                                                         vma, index, false);
 205                 if (page)
 206                         put_page(page);
 207         }
 208
 209         return 0;
 210 }
 211
 212 static const struct mm_walk_ops swapin_walk_ops = {
 213         .pmd_entry              = swapin_walk_pmd_entry,
 214 };
 215
 216 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 217                 unsigned long start, unsigned long end,
 218                 struct address_space *mapping)
 219 {
 220         pgoff_t index;
 221         struct page *page;
 222         swp_entry_t swap;
 223
 224         for (; start < end; start += PAGE_SIZE) {
 225                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 226
 227                 page = find_get_entry(mapping, index);
 228                 if (!xa_is_value(page)) {
 229                         if (page)
 230                                 put_page(page);
 231                         continue;
 232                 }
 233                 swap = radix_to_swp_entry(page);
 234                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 235                                                         NULL, 0, false);
 236                 if (page)
 237                         put_page(page);
 238         }
 239
 240         lru_add_drain();        /* Push any new pages onto the LRU now */
 241 }
 242 #endif          /* CONFIG_SWAP */
 243
 244 /*
 245  * Schedule all required I/O operations.  Do not wait for completion.
 246  */
 247 static long madvise_willneed(struct vm_area_struct *vma,
 248                              struct vm_area_struct **prev,
 249                              unsigned long start, unsigned long end)
 250 {
 251         struct file *file = vma->vm_file;
 252         loff_t offset;
 253
 254         *prev = vma;
 255 #ifdef CONFIG_SWAP
 256         if (!file) {
 257                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 258                 lru_add_drain(); /* Push any new pages onto the LRU now */
 259                 return 0;
 260         }
 261
 262         if (shmem_mapping(file->f_mapping)) {
 263                 force_shm_swapin_readahead(vma, start, end,
 264                                         file->f_mapping);
 265                 return 0;
 266         }
 267 #else
 268         if (!file)
 269                 return -EBADF;
 270 #endif
 271
 272         if (IS_DAX(file_inode(file))) {
 273                 /* no bad return value, but ignore advice */
 274                 return 0;
 275         }
 276
 277         /*
 278          * Filesystem's fadvise may need to take various locks.  We need to
 279          * explicitly grab a reference because the vma (and hence the
 280          * vma's reference to the file) can go away as soon as we drop
 281          * mmap_sem.
 282          */
 283         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 284         get_file(file);
 285         up_read(&current->mm->mmap_sem);
 286         offset = (loff_t)(start - vma->vm_start)
 287                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 288         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 289         fput(file);
 290         down_read(&current->mm->mmap_sem);
 291         return 0;
 292 }
 293
 294 static int madvise_cold_pte_range(pmd_t *pmd, unsigned long addr,
 295                                 unsigned long end, struct mm_walk *walk)
 296 {
 297         struct mmu_gather *tlb = walk->private;
 298         struct mm_struct *mm = tlb->mm;
 299         struct vm_area_struct *vma = walk->vma;
 300         pte_t *orig_pte, *pte, ptent;
 301         spinlock_t *ptl;
 302         struct page *page;
 303
 304 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 305         if (pmd_trans_huge(*pmd)) {
 306                 pmd_t orig_pmd;
 307                 unsigned long next = pmd_addr_end(addr, end);
 308
 309                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 310                 ptl = pmd_trans_huge_lock(pmd, vma);
 311                 if (!ptl)
 312                         return 0;
 313
 314                 orig_pmd = *pmd;
 315                 if (is_huge_zero_pmd(orig_pmd))
 316                         goto huge_unlock;
 317
 318                 if (unlikely(!pmd_present(orig_pmd))) {
 319                         VM_BUG_ON(thp_migration_supported() &&
 320                                         !is_pmd_migration_entry(orig_pmd));
 321                         goto huge_unlock;
 322                 }
 323
 324                 page = pmd_page(orig_pmd);
 325                 if (next - addr != HPAGE_PMD_SIZE) {
 326                         int err;
 327
 328                         if (page_mapcount(page) != 1)
 329                                 goto huge_unlock;
 330
 331                         get_page(page);
 332                         spin_unlock(ptl);
 333                         lock_page(page);
 334                         err = split_huge_page(page);
 335                         unlock_page(page);
 336                         put_page(page);
 337                         if (!err)
 338                                 goto regular_page;
 339                         return 0;
 340                 }
 341
 342                 if (pmd_young(orig_pmd)) {
 343                         pmdp_invalidate(vma, addr, pmd);
 344                         orig_pmd = pmd_mkold(orig_pmd);
 345
 346                         set_pmd_at(mm, addr, pmd, orig_pmd);
 347                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 348                 }
 349
 350                 test_and_clear_page_young(page);
 351                 deactivate_page(page);
 352 huge_unlock:
 353                 spin_unlock(ptl);
 354                 return 0;
 355         }
 356
 357         if (pmd_trans_unstable(pmd))
 358                 return 0;
 359 regular_page:
 360 #endif
 361         tlb_change_page_size(tlb, PAGE_SIZE);
 362         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 363         flush_tlb_batched_pending(mm);
 364         arch_enter_lazy_mmu_mode();
 365         for (; addr < end; pte++, addr += PAGE_SIZE) {
 366                 ptent = *pte;
 367
 368                 if (pte_none(ptent))
 369                         continue;
 370
 371                 if (!pte_present(ptent))
 372                         continue;
 373
 374                 page = vm_normal_page(vma, addr, ptent);
 375                 if (!page)
 376                         continue;
 377
 378                 /*
 379                  * Creating a THP page is expensive so split it only if we
 380                  * are sure it's worth. Split it if we are only owner.
 381                  */
 382                 if (PageTransCompound(page)) {
 383                         if (page_mapcount(page) != 1)
 384                                 break;
 385                         get_page(page);
 386                         if (!trylock_page(page)) {
 387                                 put_page(page);
 388                                 break;
 389                         }
 390                         pte_unmap_unlock(orig_pte, ptl);
 391                         if (split_huge_page(page)) {
 392                                 unlock_page(page);
 393                                 put_page(page);
 394                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 395                                 break;
 396                         }
 397                         unlock_page(page);
 398                         put_page(page);
 399                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 400                         pte--;
 401                         addr -= PAGE_SIZE;
 402                         continue;
 403                 }
 404
 405                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 406
 407                 if (pte_young(ptent)) {
 408                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 409                                                         tlb->fullmm);
 410                         ptent = pte_mkold(ptent);
 411                         set_pte_at(mm, addr, pte, ptent);
 412                         tlb_remove_tlb_entry(tlb, pte, addr);
 413                 }
 414
 415                 /*
 416                  * We are deactivating a page for accelerating reclaiming.
 417                  * VM couldn't reclaim the page unless we clear PG_young.
 418                  * As a side effect, it makes confuse idle-page tracking
 419                  * because they will miss recent referenced history.
 420                  */
 421                 test_and_clear_page_young(page);
 422                 deactivate_page(page);
 423         }
 424
 425         arch_leave_lazy_mmu_mode();
 426         pte_unmap_unlock(orig_pte, ptl);
 427         cond_resched();
 428
 429         return 0;
 430 }
 431
 432 static const struct mm_walk_ops cold_walk_ops = {
 433         .pmd_entry = madvise_cold_pte_range,
 434 };
 435
 436 static void madvise_cold_page_range(struct mmu_gather *tlb,
 437                              struct vm_area_struct *vma,
 438                              unsigned long addr, unsigned long end)
 439 {
 440         tlb_start_vma(tlb, vma);
 441         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, NULL);
 442         tlb_end_vma(tlb, vma);
 443 }
 444
 445 static long madvise_cold(struct vm_area_struct *vma,
 446                         struct vm_area_struct **prev,
 447                         unsigned long start_addr, unsigned long end_addr)
 448 {
 449         struct mm_struct *mm = vma->vm_mm;
 450         struct mmu_gather tlb;
 451
 452         *prev = vma;
 453         if (!can_madv_lru_vma(vma))
 454                 return -EINVAL;
 455
 456         lru_add_drain();
 457         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 458         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 459         tlb_finish_mmu(&tlb, start_addr, end_addr);
 460
 461         return 0;
 462 }
 463
 464 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 465                                 unsigned long end, struct mm_walk *walk)
 466
 467 {
 468         struct mmu_gather *tlb = walk->private;
 469         struct mm_struct *mm = tlb->mm;
 470         struct vm_area_struct *vma = walk->vma;
 471         spinlock_t *ptl;
 472         pte_t *orig_pte, *pte, ptent;
 473         struct page *page;
 474         int nr_swap = 0;
 475         unsigned long next;
 476
 477         next = pmd_addr_end(addr, end);
 478         if (pmd_trans_huge(*pmd))
 479                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 480                         goto next;
 481
 482         if (pmd_trans_unstable(pmd))
 483                 return 0;
 484
 485         tlb_change_page_size(tlb, PAGE_SIZE);
 486         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 487         flush_tlb_batched_pending(mm);
 488         arch_enter_lazy_mmu_mode();
 489         for (; addr != end; pte++, addr += PAGE_SIZE) {
 490                 ptent = *pte;
 491
 492                 if (pte_none(ptent))
 493                         continue;
 494                 /*
 495                  * If the pte has swp_entry, just clear page table to
 496                  * prevent swap-in which is more expensive rather than
 497                  * (page allocation + zeroing).
 498                  */
 499                 if (!pte_present(ptent)) {
 500                         swp_entry_t entry;
 501
 502                         entry = pte_to_swp_entry(ptent);
 503                         if (non_swap_entry(entry))
 504                                 continue;
 505                         nr_swap--;
 506                         free_swap_and_cache(entry);
 507                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 508                         continue;
 509                 }
 510
 511                 page = vm_normal_page(vma, addr, ptent);
 512                 if (!page)
 513                         continue;
 514
 515                 /*
 516                  * If pmd isn't transhuge but the page is THP and
 517                  * is owned by only this process, split it and
 518                  * deactivate all pages.
 519                  */
 520                 if (PageTransCompound(page)) {
 521                         if (page_mapcount(page) != 1)
 522                                 goto out;
 523                         get_page(page);
 524                         if (!trylock_page(page)) {
 525                                 put_page(page);
 526                                 goto out;
 527                         }
 528                         pte_unmap_unlock(orig_pte, ptl);
 529                         if (split_huge_page(page)) {
 530                                 unlock_page(page);
 531                                 put_page(page);
 532                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 533                                 goto out;
 534                         }
 535                         unlock_page(page);
 536                         put_page(page);
 537                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 538                         pte--;
 539                         addr -= PAGE_SIZE;
 540                         continue;
 541                 }
 542
 543                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 544
 545                 if (PageSwapCache(page) || PageDirty(page)) {
 546                         if (!trylock_page(page))
 547                                 continue;
 548                         /*
 549                          * If page is shared with others, we couldn't clear
 550                          * PG_dirty of the page.
 551                          */
 552                         if (page_mapcount(page) != 1) {
 553                                 unlock_page(page);
 554                                 continue;
 555                         }
 556
 557                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 558                                 unlock_page(page);
 559                                 continue;
 560                         }
 561
 562                         ClearPageDirty(page);
 563                         unlock_page(page);
 564                 }
 565
 566                 if (pte_young(ptent) || pte_dirty(ptent)) {
 567                         /*
 568                          * Some of architecture(ex, PPC) don't update TLB
 569                          * with set_pte_at and tlb_remove_tlb_entry so for
 570                          * the portability, remap the pte with old|clean
 571                          * after pte clearing.
 572                          */
 573                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 574                                                         tlb->fullmm);
 575
 576                         ptent = pte_mkold(ptent);
 577                         ptent = pte_mkclean(ptent);
 578                         set_pte_at(mm, addr, pte, ptent);
 579                         tlb_remove_tlb_entry(tlb, pte, addr);
 580                 }
 581                 mark_page_lazyfree(page);
 582         }
 583 out:
 584         if (nr_swap) {
 585                 if (current->mm == mm)
 586                         sync_mm_rss(mm);
 587
 588                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 589         }
 590         arch_leave_lazy_mmu_mode();
 591         pte_unmap_unlock(orig_pte, ptl);
 592         cond_resched();
 593 next:
 594         return 0;
 595 }
 596
 597 static const struct mm_walk_ops madvise_free_walk_ops = {
 598         .pmd_entry              = madvise_free_pte_range,
 599 };
 600
 601 static int madvise_free_single_vma(struct vm_area_struct *vma,
 602                         unsigned long start_addr, unsigned long end_addr)
 603 {
 604         struct mm_struct *mm = vma->vm_mm;
 605         struct mmu_notifier_range range;
 606         struct mmu_gather tlb;
 607
 608         /* MADV_FREE works for only anon vma at the moment */
 609         if (!vma_is_anonymous(vma))
 610                 return -EINVAL;
 611
 612         range.start = max(vma->vm_start, start_addr);
 613         if (range.start >= vma->vm_end)
 614                 return -EINVAL;
 615         range.end = min(vma->vm_end, end_addr);
 616         if (range.end <= vma->vm_start)
 617                 return -EINVAL;
 618         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 619                                 range.start, range.end);
 620
 621         lru_add_drain();
 622         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 623         update_hiwater_rss(mm);
 624
 625         mmu_notifier_invalidate_range_start(&range);
 626         tlb_start_vma(&tlb, vma);
 627         walk_page_range(vma->vm_mm, range.start, range.end,
 628                         &madvise_free_walk_ops, &tlb);
 629         tlb_end_vma(&tlb, vma);
 630         mmu_notifier_invalidate_range_end(&range);
 631         tlb_finish_mmu(&tlb, range.start, range.end);
 632
 633         return 0;
 634 }
 635
 636 /*
 637  * Application no longer needs these pages.  If the pages are dirty,
 638  * it's OK to just throw them away.  The app will be more careful about
 639  * data it wants to keep.  Be sure to free swap resources too.  The
 640  * zap_page_range call sets things up for shrink_active_list to actually free
 641  * these pages later if no one else has touched them in the meantime,
 642  * although we could add these pages to a global reuse list for
 643  * shrink_active_list to pick up before reclaiming other pages.
 644  *
 645  * NB: This interface discards data rather than pushes it out to swap,
 646  * as some implementations do.  This has performance implications for
 647  * applications like large transactional databases which want to discard
 648  * pages in anonymous maps after committing to backing store the data
 649  * that was kept in them.  There is no reason to write this data out to
 650  * the swap area if the application is discarding it.
 651  *
 652  * An interface that causes the system to free clean pages and flush
 653  * dirty pages is already available as msync(MS_INVALIDATE).
 654  */
 655 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 656                                         unsigned long start, unsigned long end)
 657 {
 658         zap_page_range(vma, start, end - start);
 659         return 0;
 660 }
 661
 662 static long madvise_dontneed_free(struct vm_area_struct *vma,
 663                                   struct vm_area_struct **prev,
 664                                   unsigned long start, unsigned long end,
 665                                   int behavior)
 666 {
 667         *prev = vma;
 668         if (!can_madv_lru_vma(vma))
 669                 return -EINVAL;
 670
 671         if (!userfaultfd_remove(vma, start, end)) {
 672                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 673
 674                 down_read(&current->mm->mmap_sem);
 675                 vma = find_vma(current->mm, start);
 676                 if (!vma)
 677                         return -ENOMEM;
 678                 if (start < vma->vm_start) {
 679                         /*
 680                          * This "vma" under revalidation is the one
 681                          * with the lowest vma->vm_start where start
 682                          * is also < vma->vm_end. If start <
 683                          * vma->vm_start it means an hole materialized
 684                          * in the user address space within the
 685                          * virtual range passed to MADV_DONTNEED
 686                          * or MADV_FREE.
 687                          */
 688                         return -ENOMEM;
 689                 }
 690                 if (!can_madv_lru_vma(vma))
 691                         return -EINVAL;
 692                 if (end > vma->vm_end) {
 693                         /*
 694                          * Don't fail if end > vma->vm_end. If the old
 695                          * vma was splitted while the mmap_sem was
 696                          * released the effect of the concurrent
 697                          * operation may not cause madvise() to
 698                          * have an undefined result. There may be an
 699                          * adjacent next vma that we'll walk
 700                          * next. userfaultfd_remove() will generate an
 701                          * UFFD_EVENT_REMOVE repetition on the
 702                          * end-vma->vm_end range, but the manager can
 703                          * handle a repetition fine.
 704                          */
 705                         end = vma->vm_end;
 706                 }
 707                 VM_WARN_ON(start >= end);
 708         }
 709
 710         if (behavior == MADV_DONTNEED)
 711                 return madvise_dontneed_single_vma(vma, start, end);
 712         else if (behavior == MADV_FREE)
 713                 return madvise_free_single_vma(vma, start, end);
 714         else
 715                 return -EINVAL;
 716 }
 717
 718 /*
 719  * Application wants to free up the pages and associated backing store.
 720  * This is effectively punching a hole into the middle of a file.
 721  */
 722 static long madvise_remove(struct vm_area_struct *vma,
 723                                 struct vm_area_struct **prev,
 724                                 unsigned long start, unsigned long end)
 725 {
 726         loff_t offset;
 727         int error;
 728         struct file *f;
 729
 730         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 731
 732         if (vma->vm_flags & VM_LOCKED)
 733                 return -EINVAL;
 734
 735         f = vma->vm_file;
 736
 737         if (!f || !f->f_mapping || !f->f_mapping->host) {
 738                         return -EINVAL;
 739         }
 740
 741         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 742                 return -EACCES;
 743
 744         offset = (loff_t)(start - vma->vm_start)
 745                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 746
 747         /*
 748          * Filesystem's fallocate may need to take i_mutex.  We need to
 749          * explicitly grab a reference because the vma (and hence the
 750          * vma's reference to the file) can go away as soon as we drop
 751          * mmap_sem.
 752          */
 753         get_file(f);
 754         if (userfaultfd_remove(vma, start, end)) {
 755                 /* mmap_sem was not released by userfaultfd_remove() */
 756                 up_read(&current->mm->mmap_sem);
 757         }
 758         error = vfs_fallocate(f,
 759                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 760                                 offset, end - start);
 761         fput(f);
 762         down_read(&current->mm->mmap_sem);
 763         return error;
 764 }
 765
 766 #ifdef CONFIG_MEMORY_FAILURE
 767 /*
 768  * Error injection support for memory error handling.
 769  */
 770 static int madvise_inject_error(int behavior,
 771                 unsigned long start, unsigned long end)
 772 {
 773         struct page *page;
 774         struct zone *zone;
 775         unsigned int order;
 776
 777         if (!capable(CAP_SYS_ADMIN))
 778                 return -EPERM;
 779
 780
 781         for (; start < end; start += PAGE_SIZE << order) {
 782                 unsigned long pfn;
 783                 int ret;
 784
 785                 ret = get_user_pages_fast(start, 1, 0, &page);
 786                 if (ret != 1)
 787                         return ret;
 788                 pfn = page_to_pfn(page);
 789
 790                 /*
 791                  * When soft offlining hugepages, after migrating the page
 792                  * we dissolve it, therefore in the second loop "page" will
 793                  * no longer be a compound page, and order will be 0.
 794                  */
 795                 order = compound_order(compound_head(page));
 796
 797                 if (PageHWPoison(page)) {
 798                         put_page(page);
 799                         continue;
 800                 }
 801
 802                 if (behavior == MADV_SOFT_OFFLINE) {
 803                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 804                                         pfn, start);
 805
 806                         ret = soft_offline_page(page, MF_COUNT_INCREASED);
 807                         if (ret)
 808                                 return ret;
 809                         continue;
 810                 }
 811
 812                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 813                                 pfn, start);
 814
 815                 /*
 816                  * Drop the page reference taken by get_user_pages_fast(). In
 817                  * the absence of MF_COUNT_INCREASED the memory_failure()
 818                  * routine is responsible for pinning the page to prevent it
 819                  * from being released back to the page allocator.
 820                  */
 821                 put_page(page);
 822                 ret = memory_failure(pfn, 0);
 823                 if (ret)
 824                         return ret;
 825         }
 826
 827         /* Ensure that all poisoned pages are removed from per-cpu lists */
 828         for_each_populated_zone(zone)
 829                 drain_all_pages(zone);
 830
 831         return 0;
 832 }
 833 #endif
 834
 835 static long
 836 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 837                 unsigned long start, unsigned long end, int behavior)
 838 {
 839         switch (behavior) {
 840         case MADV_REMOVE:
 841                 return madvise_remove(vma, prev, start, end);
 842         case MADV_WILLNEED:
 843                 return madvise_willneed(vma, prev, start, end);
 844         case MADV_COLD:
 845                 return madvise_cold(vma, prev, start, end);
 846         case MADV_FREE:
 847         case MADV_DONTNEED:
 848                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 849         default:
 850                 return madvise_behavior(vma, prev, start, end, behavior);
 851         }
 852 }
 853
 854 static bool
 855 madvise_behavior_valid(int behavior)
 856 {
 857         switch (behavior) {
 858         case MADV_DOFORK:
 859         case MADV_DONTFORK:
 860         case MADV_NORMAL:
 861         case MADV_SEQUENTIAL:
 862         case MADV_RANDOM:
 863         case MADV_REMOVE:
 864         case MADV_WILLNEED:
 865         case MADV_DONTNEED:
 866         case MADV_FREE:
 867         case MADV_COLD:
 868 #ifdef CONFIG_KSM
 869         case MADV_MERGEABLE:
 870         case MADV_UNMERGEABLE:
 871 #endif
 872 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 873         case MADV_HUGEPAGE:
 874         case MADV_NOHUGEPAGE:
 875 #endif
 876         case MADV_DONTDUMP:
 877         case MADV_DODUMP:
 878         case MADV_WIPEONFORK:
 879         case MADV_KEEPONFORK:
 880 #ifdef CONFIG_MEMORY_FAILURE
 881         case MADV_SOFT_OFFLINE:
 882         case MADV_HWPOISON:
 883 #endif
 884                 return true;
 885
 886         default:
 887                 return false;
 888         }
 889 }
 890
 891 /*
 892  * The madvise(2) system call.
 893  *
 894  * Applications can use madvise() to advise the kernel how it should
 895  * handle paging I/O in this VM area.  The idea is to help the kernel
 896  * use appropriate read-ahead and caching techniques.  The information
 897  * provided is advisory only, and can be safely disregarded by the
 898  * kernel without affecting the correct operation of the application.
 899  *
 900  * behavior values:
 901  *  MADV_NORMAL - the default behavior is to read clusters.  This
 902  *              results in some read-ahead and read-behind.
 903  *  MADV_RANDOM - the system should read the minimum amount of data
 904  *              on any access, since it is unlikely that the appli-
 905  *              cation will need more than what it asks for.
 906  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 907  *              once, so they can be aggressively read ahead, and
 908  *              can be freed soon after they are accessed.
 909  *  MADV_WILLNEED - the application is notifying the system to read
 910  *              some pages ahead.
 911  *  MADV_DONTNEED - the application is finished with the given range,
 912  *              so the kernel can free resources associated with it.
 913  *  MADV_FREE - the application marks pages in the given range as lazy free,
 914  *              where actual purges are postponed until memory pressure happens.
 915  *  MADV_REMOVE - the application wants to free up the given range of
 916  *              pages and associated backing store.
 917  *  MADV_DONTFORK - omit this area from child's address space when forking:
 918  *              typically, to avoid COWing pages pinned by get_user_pages().
 919  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
 920  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
 921  *              range after a fork.
 922  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
 923  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
 924  *              were corrupted by unrecoverable hardware memory failure.
 925  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
 926  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
 927  *              this area with pages of identical content from other such areas.
 928  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
 929  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
 930  *              huge pages in the future. Existing pages might be coalesced and
 931  *              new pages might be allocated as THP.
 932  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
 933  *              transparent huge pages so the existing pages will not be
 934  *              coalesced into THP and new pages will not be allocated as THP.
 935  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
 936  *              from being included in its core dump.
 937  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
 938  *
 939  * return values:
 940  *  zero    - success
 941  *  -EINVAL - start + len < 0, start is not page-aligned,
 942  *              "behavior" is not a valid value, or application
 943  *              is attempting to release locked or shared pages,
 944  *              or the specified address range includes file, Huge TLB,
 945  *              MAP_SHARED or VMPFNMAP range.
 946  *  -ENOMEM - addresses in the specified range are not currently
 947  *              mapped, or are outside the AS of the process.
 948  *  -EIO    - an I/O error occurred while paging in data.
 949  *  -EBADF  - map exists, but area maps something that isn't a file.
 950  *  -EAGAIN - a kernel resource was temporarily unavailable.
 951  */
 952 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
 953 {
 954         unsigned long end, tmp;
 955         struct vm_area_struct *vma, *prev;
 956         int unmapped_error = 0;
 957         int error = -EINVAL;
 958         int write;
 959         size_t len;
 960         struct blk_plug plug;
 961
 962         start = untagged_addr(start);
 963
 964         if (!madvise_behavior_valid(behavior))
 965                 return error;
 966
 967         if (start & ~PAGE_MASK)
 968                 return error;
 969         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
 970
 971         /* Check to see whether len was rounded up from small -ve to zero */
 972         if (len_in && !len)
 973                 return error;
 974
 975         end = start + len;
 976         if (end < start)
 977                 return error;
 978
 979         error = 0;
 980         if (end == start)
 981                 return error;
 982
 983 #ifdef CONFIG_MEMORY_FAILURE
 984         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
 985                 return madvise_inject_error(behavior, start, start + len_in);
 986 #endif
 987
 988         write = madvise_need_mmap_write(behavior);
 989         if (write) {
 990                 if (down_write_killable(&current->mm->mmap_sem))
 991                         return -EINTR;
 992         } else {
 993                 down_read(&current->mm->mmap_sem);
 994         }
 995
 996         /*
 997          * If the interval [start,end) covers some unmapped address
 998          * ranges, just ignore them, but return -ENOMEM at the end.
 999          * - different from the way of handling in mlock etc.
1000          */
1001         vma = find_vma_prev(current->mm, start, &prev);
1002         if (vma && start > vma->vm_start)
1003                 prev = vma;
1004
1005         blk_start_plug(&plug);
1006         for (;;) {
1007                 /* Still start < end. */
1008                 error = -ENOMEM;
1009                 if (!vma)
1010                         goto out;
1011
1012                 /* Here start < (end|vma->vm_end). */
1013                 if (start < vma->vm_start) {
1014                         unmapped_error = -ENOMEM;
1015                         start = vma->vm_start;
1016                         if (start >= end)
1017                                 goto out;
1018                 }
1019
1020                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1021                 tmp = vma->vm_end;
1022                 if (end < tmp)
1023                         tmp = end;
1024
1025                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1026                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1027                 if (error)
1028                         goto out;
1029                 start = tmp;
1030                 if (prev && start < prev->vm_end)
1031                         start = prev->vm_end;
1032                 error = unmapped_error;
1033                 if (start >= end)
1034                         goto out;
1035                 if (prev)
1036                         vma = prev->vm_next;
1037                 else    /* madvise_remove dropped mmap_sem */
1038                         vma = find_vma(current->mm, start);
1039         }
1040 out:
1041         blk_finish_plug(&plug);
1042         if (write)
1043                 up_write(&current->mm->mmap_sem);
1044         else
1045                 up_read(&current->mm->mmap_sem);
1046
1047         return error;
1048 }