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 struct madvise_walk_private {
  36         struct mmu_gather *tlb;
  37         bool pageout;
  38 };
  39
  40 /*
  41  * Any behaviour which results in changes to the vma->vm_flags needs to
  42  * take mmap_sem for writing. Others, which simply traverse vmas, need
  43  * to only take it for reading.
  44  */
  45 static int madvise_need_mmap_write(int behavior)
  46 {
  47         switch (behavior) {
  48         case MADV_REMOVE:
  49         case MADV_WILLNEED:
  50         case MADV_DONTNEED:
  51         case MADV_COLD:
  52         case MADV_PAGEOUT:
  53         case MADV_FREE:
  54                 return 0;
  55         default:
  56                 /* be safe, default to 1. list exceptions explicitly */
  57                 return 1;
  58         }
  59 }
  60
  61 /*
  62  * We can potentially split a vm area into separate
  63  * areas, each area with its own behavior.
  64  */
  65 static long madvise_behavior(struct vm_area_struct *vma,
  66                      struct vm_area_struct **prev,
  67                      unsigned long start, unsigned long end, int behavior)
  68 {
  69         struct mm_struct *mm = vma->vm_mm;
  70         int error = 0;
  71         pgoff_t pgoff;
  72         unsigned long new_flags = vma->vm_flags;
  73
  74         switch (behavior) {
  75         case MADV_NORMAL:
  76                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  77                 break;
  78         case MADV_SEQUENTIAL:
  79                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  80                 break;
  81         case MADV_RANDOM:
  82                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  83                 break;
  84         case MADV_DONTFORK:
  85                 new_flags |= VM_DONTCOPY;
  86                 break;
  87         case MADV_DOFORK:
  88                 if (vma->vm_flags & VM_IO) {
  89                         error = -EINVAL;
  90                         goto out;
  91                 }
  92                 new_flags &= ~VM_DONTCOPY;
  93                 break;
  94         case MADV_WIPEONFORK:
  95                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  96                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  97                         error = -EINVAL;
  98                         goto out;
  99                 }
 100                 new_flags |= VM_WIPEONFORK;
 101                 break;
 102         case MADV_KEEPONFORK:
 103                 new_flags &= ~VM_WIPEONFORK;
 104                 break;
 105         case MADV_DONTDUMP:
 106                 new_flags |= VM_DONTDUMP;
 107                 break;
 108         case MADV_DODUMP:
 109                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 110                         error = -EINVAL;
 111                         goto out;
 112                 }
 113                 new_flags &= ~VM_DONTDUMP;
 114                 break;
 115         case MADV_MERGEABLE:
 116         case MADV_UNMERGEABLE:
 117                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 118                 if (error)
 119                         goto out_convert_errno;
 120                 break;
 121         case MADV_HUGEPAGE:
 122         case MADV_NOHUGEPAGE:
 123                 error = hugepage_madvise(vma, &new_flags, behavior);
 124                 if (error)
 125                         goto out_convert_errno;
 126                 break;
 127         }
 128
 129         if (new_flags == vma->vm_flags) {
 130                 *prev = vma;
 131                 goto out;
 132         }
 133
 134         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 135         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 136                           vma->vm_file, pgoff, vma_policy(vma),
 137                           vma->vm_userfaultfd_ctx);
 138         if (*prev) {
 139                 vma = *prev;
 140                 goto success;
 141         }
 142
 143         *prev = vma;
 144
 145         if (start != vma->vm_start) {
 146                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 147                         error = -ENOMEM;
 148                         goto out;
 149                 }
 150                 error = __split_vma(mm, vma, start, 1);
 151                 if (error)
 152                         goto out_convert_errno;
 153         }
 154
 155         if (end != vma->vm_end) {
 156                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 157                         error = -ENOMEM;
 158                         goto out;
 159                 }
 160                 error = __split_vma(mm, vma, end, 0);
 161                 if (error)
 162                         goto out_convert_errno;
 163         }
 164
 165 success:
 166         /*
 167          * vm_flags is protected by the mmap_sem held in write mode.
 168          */
 169         vma->vm_flags = new_flags;
 170
 171 out_convert_errno:
 172         /*
 173          * madvise() returns EAGAIN if kernel resources, such as
 174          * slab, are temporarily unavailable.
 175          */
 176         if (error == -ENOMEM)
 177                 error = -EAGAIN;
 178 out:
 179         return error;
 180 }
 181
 182 #ifdef CONFIG_SWAP
 183 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 184         unsigned long end, struct mm_walk *walk)
 185 {
 186         pte_t *orig_pte;
 187         struct vm_area_struct *vma = walk->private;
 188         unsigned long index;
 189
 190         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 191                 return 0;
 192
 193         for (index = start; index != end; index += PAGE_SIZE) {
 194                 pte_t pte;
 195                 swp_entry_t entry;
 196                 struct page *page;
 197                 spinlock_t *ptl;
 198
 199                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 200                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 201                 pte_unmap_unlock(orig_pte, ptl);
 202
 203                 if (pte_present(pte) || pte_none(pte))
 204                         continue;
 205                 entry = pte_to_swp_entry(pte);
 206                 if (unlikely(non_swap_entry(entry)))
 207                         continue;
 208
 209                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 210                                                         vma, index, false);
 211                 if (page)
 212                         put_page(page);
 213         }
 214
 215         return 0;
 216 }
 217
 218 static const struct mm_walk_ops swapin_walk_ops = {
 219         .pmd_entry              = swapin_walk_pmd_entry,
 220 };
 221
 222 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 223                 unsigned long start, unsigned long end,
 224                 struct address_space *mapping)
 225 {
 226         pgoff_t index;
 227         struct page *page;
 228         swp_entry_t swap;
 229
 230         for (; start < end; start += PAGE_SIZE) {
 231                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 232
 233                 page = find_get_entry(mapping, index);
 234                 if (!xa_is_value(page)) {
 235                         if (page)
 236                                 put_page(page);
 237                         continue;
 238                 }
 239                 swap = radix_to_swp_entry(page);
 240                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 241                                                         NULL, 0, false);
 242                 if (page)
 243                         put_page(page);
 244         }
 245
 246         lru_add_drain();        /* Push any new pages onto the LRU now */
 247 }
 248 #endif          /* CONFIG_SWAP */
 249
 250 /*
 251  * Schedule all required I/O operations.  Do not wait for completion.
 252  */
 253 static long madvise_willneed(struct vm_area_struct *vma,
 254                              struct vm_area_struct **prev,
 255                              unsigned long start, unsigned long end)
 256 {
 257         struct file *file = vma->vm_file;
 258         loff_t offset;
 259
 260         *prev = vma;
 261 #ifdef CONFIG_SWAP
 262         if (!file) {
 263                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 264                 lru_add_drain(); /* Push any new pages onto the LRU now */
 265                 return 0;
 266         }
 267
 268         if (shmem_mapping(file->f_mapping)) {
 269                 force_shm_swapin_readahead(vma, start, end,
 270                                         file->f_mapping);
 271                 return 0;
 272         }
 273 #else
 274         if (!file)
 275                 return -EBADF;
 276 #endif
 277
 278         if (IS_DAX(file_inode(file))) {
 279                 /* no bad return value, but ignore advice */
 280                 return 0;
 281         }
 282
 283         /*
 284          * Filesystem's fadvise may need to take various locks.  We need to
 285          * explicitly grab a reference because the vma (and hence the
 286          * vma's reference to the file) can go away as soon as we drop
 287          * mmap_sem.
 288          */
 289         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 290         get_file(file);
 291         up_read(&current->mm->mmap_sem);
 292         offset = (loff_t)(start - vma->vm_start)
 293                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 294         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 295         fput(file);
 296         down_read(&current->mm->mmap_sem);
 297         return 0;
 298 }
 299
 300 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 301                                 unsigned long addr, unsigned long end,
 302                                 struct mm_walk *walk)
 303 {
 304         struct madvise_walk_private *private = walk->private;
 305         struct mmu_gather *tlb = private->tlb;
 306         bool pageout = private->pageout;
 307         struct mm_struct *mm = tlb->mm;
 308         struct vm_area_struct *vma = walk->vma;
 309         pte_t *orig_pte, *pte, ptent;
 310         spinlock_t *ptl;
 311         struct page *page = NULL;
 312         LIST_HEAD(page_list);
 313
 314         if (fatal_signal_pending(current))
 315                 return -EINTR;
 316
 317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 318         if (pmd_trans_huge(*pmd)) {
 319                 pmd_t orig_pmd;
 320                 unsigned long next = pmd_addr_end(addr, end);
 321
 322                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 323                 ptl = pmd_trans_huge_lock(pmd, vma);
 324                 if (!ptl)
 325                         return 0;
 326
 327                 orig_pmd = *pmd;
 328                 if (is_huge_zero_pmd(orig_pmd))
 329                         goto huge_unlock;
 330
 331                 if (unlikely(!pmd_present(orig_pmd))) {
 332                         VM_BUG_ON(thp_migration_supported() &&
 333                                         !is_pmd_migration_entry(orig_pmd));
 334                         goto huge_unlock;
 335                 }
 336
 337                 page = pmd_page(orig_pmd);
 338                 if (next - addr != HPAGE_PMD_SIZE) {
 339                         int err;
 340
 341                         if (page_mapcount(page) != 1)
 342                                 goto huge_unlock;
 343
 344                         get_page(page);
 345                         spin_unlock(ptl);
 346                         lock_page(page);
 347                         err = split_huge_page(page);
 348                         unlock_page(page);
 349                         put_page(page);
 350                         if (!err)
 351                                 goto regular_page;
 352                         return 0;
 353                 }
 354
 355                 if (pmd_young(orig_pmd)) {
 356                         pmdp_invalidate(vma, addr, pmd);
 357                         orig_pmd = pmd_mkold(orig_pmd);
 358
 359                         set_pmd_at(mm, addr, pmd, orig_pmd);
 360                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 361                 }
 362
 363                 ClearPageReferenced(page);
 364                 test_and_clear_page_young(page);
 365                 if (pageout) {
 366                         if (!isolate_lru_page(page))
 367                                 list_add(&page->lru, &page_list);
 368                 } else
 369                         deactivate_page(page);
 370 huge_unlock:
 371                 spin_unlock(ptl);
 372                 if (pageout)
 373                         reclaim_pages(&page_list);
 374                 return 0;
 375         }
 376
 377         if (pmd_trans_unstable(pmd))
 378                 return 0;
 379 regular_page:
 380 #endif
 381         tlb_change_page_size(tlb, PAGE_SIZE);
 382         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 383         flush_tlb_batched_pending(mm);
 384         arch_enter_lazy_mmu_mode();
 385         for (; addr < end; pte++, addr += PAGE_SIZE) {
 386                 ptent = *pte;
 387
 388                 if (pte_none(ptent))
 389                         continue;
 390
 391                 if (!pte_present(ptent))
 392                         continue;
 393
 394                 page = vm_normal_page(vma, addr, ptent);
 395                 if (!page)
 396                         continue;
 397
 398                 /*
 399                  * Creating a THP page is expensive so split it only if we
 400                  * are sure it's worth. Split it if we are only owner.
 401                  */
 402                 if (PageTransCompound(page)) {
 403                         if (page_mapcount(page) != 1)
 404                                 break;
 405                         get_page(page);
 406                         if (!trylock_page(page)) {
 407                                 put_page(page);
 408                                 break;
 409                         }
 410                         pte_unmap_unlock(orig_pte, ptl);
 411                         if (split_huge_page(page)) {
 412                                 unlock_page(page);
 413                                 put_page(page);
 414                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 415                                 break;
 416                         }
 417                         unlock_page(page);
 418                         put_page(page);
 419                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 420                         pte--;
 421                         addr -= PAGE_SIZE;
 422                         continue;
 423                 }
 424
 425                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 426
 427                 if (pte_young(ptent)) {
 428                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 429                                                         tlb->fullmm);
 430                         ptent = pte_mkold(ptent);
 431                         set_pte_at(mm, addr, pte, ptent);
 432                         tlb_remove_tlb_entry(tlb, pte, addr);
 433                 }
 434
 435                 /*
 436                  * We are deactivating a page for accelerating reclaiming.
 437                  * VM couldn't reclaim the page unless we clear PG_young.
 438                  * As a side effect, it makes confuse idle-page tracking
 439                  * because they will miss recent referenced history.
 440                  */
 441                 ClearPageReferenced(page);
 442                 test_and_clear_page_young(page);
 443                 if (pageout) {
 444                         if (!isolate_lru_page(page))
 445                                 list_add(&page->lru, &page_list);
 446                 } else
 447                         deactivate_page(page);
 448         }
 449
 450         arch_leave_lazy_mmu_mode();
 451         pte_unmap_unlock(orig_pte, ptl);
 452         if (pageout)
 453                 reclaim_pages(&page_list);
 454         cond_resched();
 455
 456         return 0;
 457 }
 458
 459 static const struct mm_walk_ops cold_walk_ops = {
 460         .pmd_entry = madvise_cold_or_pageout_pte_range,
 461 };
 462
 463 static void madvise_cold_page_range(struct mmu_gather *tlb,
 464                              struct vm_area_struct *vma,
 465                              unsigned long addr, unsigned long end)
 466 {
 467         struct madvise_walk_private walk_private = {
 468                 .pageout = false,
 469                 .tlb = tlb,
 470         };
 471
 472         tlb_start_vma(tlb, vma);
 473         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 474         tlb_end_vma(tlb, vma);
 475 }
 476
 477 static long madvise_cold(struct vm_area_struct *vma,
 478                         struct vm_area_struct **prev,
 479                         unsigned long start_addr, unsigned long end_addr)
 480 {
 481         struct mm_struct *mm = vma->vm_mm;
 482         struct mmu_gather tlb;
 483
 484         *prev = vma;
 485         if (!can_madv_lru_vma(vma))
 486                 return -EINVAL;
 487
 488         lru_add_drain();
 489         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 490         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 491         tlb_finish_mmu(&tlb, start_addr, end_addr);
 492
 493         return 0;
 494 }
 495
 496 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 497                              struct vm_area_struct *vma,
 498                              unsigned long addr, unsigned long end)
 499 {
 500         struct madvise_walk_private walk_private = {
 501                 .pageout = true,
 502                 .tlb = tlb,
 503         };
 504
 505         tlb_start_vma(tlb, vma);
 506         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 507         tlb_end_vma(tlb, vma);
 508 }
 509
 510 static inline bool can_do_pageout(struct vm_area_struct *vma)
 511 {
 512         if (vma_is_anonymous(vma))
 513                 return true;
 514         if (!vma->vm_file)
 515                 return false;
 516         /*
 517          * paging out pagecache only for non-anonymous mappings that correspond
 518          * to the files the calling process could (if tried) open for writing;
 519          * otherwise we'd be including shared non-exclusive mappings, which
 520          * opens a side channel.
 521          */
 522         return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 523                 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 524 }
 525
 526 static long madvise_pageout(struct vm_area_struct *vma,
 527                         struct vm_area_struct **prev,
 528                         unsigned long start_addr, unsigned long end_addr)
 529 {
 530         struct mm_struct *mm = vma->vm_mm;
 531         struct mmu_gather tlb;
 532
 533         *prev = vma;
 534         if (!can_madv_lru_vma(vma))
 535                 return -EINVAL;
 536
 537         if (!can_do_pageout(vma))
 538                 return 0;
 539
 540         lru_add_drain();
 541         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 542         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 543         tlb_finish_mmu(&tlb, start_addr, end_addr);
 544
 545         return 0;
 546 }
 547
 548 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 549                                 unsigned long end, struct mm_walk *walk)
 550
 551 {
 552         struct mmu_gather *tlb = walk->private;
 553         struct mm_struct *mm = tlb->mm;
 554         struct vm_area_struct *vma = walk->vma;
 555         spinlock_t *ptl;
 556         pte_t *orig_pte, *pte, ptent;
 557         struct page *page;
 558         int nr_swap = 0;
 559         unsigned long next;
 560
 561         next = pmd_addr_end(addr, end);
 562         if (pmd_trans_huge(*pmd))
 563                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 564                         goto next;
 565
 566         if (pmd_trans_unstable(pmd))
 567                 return 0;
 568
 569         tlb_change_page_size(tlb, PAGE_SIZE);
 570         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 571         flush_tlb_batched_pending(mm);
 572         arch_enter_lazy_mmu_mode();
 573         for (; addr != end; pte++, addr += PAGE_SIZE) {
 574                 ptent = *pte;
 575
 576                 if (pte_none(ptent))
 577                         continue;
 578                 /*
 579                  * If the pte has swp_entry, just clear page table to
 580                  * prevent swap-in which is more expensive rather than
 581                  * (page allocation + zeroing).
 582                  */
 583                 if (!pte_present(ptent)) {
 584                         swp_entry_t entry;
 585
 586                         entry = pte_to_swp_entry(ptent);
 587                         if (non_swap_entry(entry))
 588                                 continue;
 589                         nr_swap--;
 590                         free_swap_and_cache(entry);
 591                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 592                         continue;
 593                 }
 594
 595                 page = vm_normal_page(vma, addr, ptent);
 596                 if (!page)
 597                         continue;
 598
 599                 /*
 600                  * If pmd isn't transhuge but the page is THP and
 601                  * is owned by only this process, split it and
 602                  * deactivate all pages.
 603                  */
 604                 if (PageTransCompound(page)) {
 605                         if (page_mapcount(page) != 1)
 606                                 goto out;
 607                         get_page(page);
 608                         if (!trylock_page(page)) {
 609                                 put_page(page);
 610                                 goto out;
 611                         }
 612                         pte_unmap_unlock(orig_pte, ptl);
 613                         if (split_huge_page(page)) {
 614                                 unlock_page(page);
 615                                 put_page(page);
 616                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 617                                 goto out;
 618                         }
 619                         unlock_page(page);
 620                         put_page(page);
 621                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 622                         pte--;
 623                         addr -= PAGE_SIZE;
 624                         continue;
 625                 }
 626
 627                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 628
 629                 if (PageSwapCache(page) || PageDirty(page)) {
 630                         if (!trylock_page(page))
 631                                 continue;
 632                         /*
 633                          * If page is shared with others, we couldn't clear
 634                          * PG_dirty of the page.
 635                          */
 636                         if (page_mapcount(page) != 1) {
 637                                 unlock_page(page);
 638                                 continue;
 639                         }
 640
 641                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 642                                 unlock_page(page);
 643                                 continue;
 644                         }
 645
 646                         ClearPageDirty(page);
 647                         unlock_page(page);
 648                 }
 649
 650                 if (pte_young(ptent) || pte_dirty(ptent)) {
 651                         /*
 652                          * Some of architecture(ex, PPC) don't update TLB
 653                          * with set_pte_at and tlb_remove_tlb_entry so for
 654                          * the portability, remap the pte with old|clean
 655                          * after pte clearing.
 656                          */
 657                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 658                                                         tlb->fullmm);
 659
 660                         ptent = pte_mkold(ptent);
 661                         ptent = pte_mkclean(ptent);
 662                         set_pte_at(mm, addr, pte, ptent);
 663                         tlb_remove_tlb_entry(tlb, pte, addr);
 664                 }
 665                 mark_page_lazyfree(page);
 666         }
 667 out:
 668         if (nr_swap) {
 669                 if (current->mm == mm)
 670                         sync_mm_rss(mm);
 671
 672                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 673         }
 674         arch_leave_lazy_mmu_mode();
 675         pte_unmap_unlock(orig_pte, ptl);
 676         cond_resched();
 677 next:
 678         return 0;
 679 }
 680
 681 static const struct mm_walk_ops madvise_free_walk_ops = {
 682         .pmd_entry              = madvise_free_pte_range,
 683 };
 684
 685 static int madvise_free_single_vma(struct vm_area_struct *vma,
 686                         unsigned long start_addr, unsigned long end_addr)
 687 {
 688         struct mm_struct *mm = vma->vm_mm;
 689         struct mmu_notifier_range range;
 690         struct mmu_gather tlb;
 691
 692         /* MADV_FREE works for only anon vma at the moment */
 693         if (!vma_is_anonymous(vma))
 694                 return -EINVAL;
 695
 696         range.start = max(vma->vm_start, start_addr);
 697         if (range.start >= vma->vm_end)
 698                 return -EINVAL;
 699         range.end = min(vma->vm_end, end_addr);
 700         if (range.end <= vma->vm_start)
 701                 return -EINVAL;
 702         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 703                                 range.start, range.end);
 704
 705         lru_add_drain();
 706         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 707         update_hiwater_rss(mm);
 708
 709         mmu_notifier_invalidate_range_start(&range);
 710         tlb_start_vma(&tlb, vma);
 711         walk_page_range(vma->vm_mm, range.start, range.end,
 712                         &madvise_free_walk_ops, &tlb);
 713         tlb_end_vma(&tlb, vma);
 714         mmu_notifier_invalidate_range_end(&range);
 715         tlb_finish_mmu(&tlb, range.start, range.end);
 716
 717         return 0;
 718 }
 719
 720 /*
 721  * Application no longer needs these pages.  If the pages are dirty,
 722  * it's OK to just throw them away.  The app will be more careful about
 723  * data it wants to keep.  Be sure to free swap resources too.  The
 724  * zap_page_range call sets things up for shrink_active_list to actually free
 725  * these pages later if no one else has touched them in the meantime,
 726  * although we could add these pages to a global reuse list for
 727  * shrink_active_list to pick up before reclaiming other pages.
 728  *
 729  * NB: This interface discards data rather than pushes it out to swap,
 730  * as some implementations do.  This has performance implications for
 731  * applications like large transactional databases which want to discard
 732  * pages in anonymous maps after committing to backing store the data
 733  * that was kept in them.  There is no reason to write this data out to
 734  * the swap area if the application is discarding it.
 735  *
 736  * An interface that causes the system to free clean pages and flush
 737  * dirty pages is already available as msync(MS_INVALIDATE).
 738  */
 739 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 740                                         unsigned long start, unsigned long end)
 741 {
 742         zap_page_range(vma, start, end - start);
 743         return 0;
 744 }
 745
 746 static long madvise_dontneed_free(struct vm_area_struct *vma,
 747                                   struct vm_area_struct **prev,
 748                                   unsigned long start, unsigned long end,
 749                                   int behavior)
 750 {
 751         *prev = vma;
 752         if (!can_madv_lru_vma(vma))
 753                 return -EINVAL;
 754
 755         if (!userfaultfd_remove(vma, start, end)) {
 756                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 757
 758                 down_read(&current->mm->mmap_sem);
 759                 vma = find_vma(current->mm, start);
 760                 if (!vma)
 761                         return -ENOMEM;
 762                 if (start < vma->vm_start) {
 763                         /*
 764                          * This "vma" under revalidation is the one
 765                          * with the lowest vma->vm_start where start
 766                          * is also < vma->vm_end. If start <
 767                          * vma->vm_start it means an hole materialized
 768                          * in the user address space within the
 769                          * virtual range passed to MADV_DONTNEED
 770                          * or MADV_FREE.
 771                          */
 772                         return -ENOMEM;
 773                 }
 774                 if (!can_madv_lru_vma(vma))
 775                         return -EINVAL;
 776                 if (end > vma->vm_end) {
 777                         /*
 778                          * Don't fail if end > vma->vm_end. If the old
 779                          * vma was splitted while the mmap_sem was
 780                          * released the effect of the concurrent
 781                          * operation may not cause madvise() to
 782                          * have an undefined result. There may be an
 783                          * adjacent next vma that we'll walk
 784                          * next. userfaultfd_remove() will generate an
 785                          * UFFD_EVENT_REMOVE repetition on the
 786                          * end-vma->vm_end range, but the manager can
 787                          * handle a repetition fine.
 788                          */
 789                         end = vma->vm_end;
 790                 }
 791                 VM_WARN_ON(start >= end);
 792         }
 793
 794         if (behavior == MADV_DONTNEED)
 795                 return madvise_dontneed_single_vma(vma, start, end);
 796         else if (behavior == MADV_FREE)
 797                 return madvise_free_single_vma(vma, start, end);
 798         else
 799                 return -EINVAL;
 800 }
 801
 802 /*
 803  * Application wants to free up the pages and associated backing store.
 804  * This is effectively punching a hole into the middle of a file.
 805  */
 806 static long madvise_remove(struct vm_area_struct *vma,
 807                                 struct vm_area_struct **prev,
 808                                 unsigned long start, unsigned long end)
 809 {
 810         loff_t offset;
 811         int error;
 812         struct file *f;
 813
 814         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 815
 816         if (vma->vm_flags & VM_LOCKED)
 817                 return -EINVAL;
 818
 819         f = vma->vm_file;
 820
 821         if (!f || !f->f_mapping || !f->f_mapping->host) {
 822                         return -EINVAL;
 823         }
 824
 825         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 826                 return -EACCES;
 827
 828         offset = (loff_t)(start - vma->vm_start)
 829                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 830
 831         /*
 832          * Filesystem's fallocate may need to take i_mutex.  We need to
 833          * explicitly grab a reference because the vma (and hence the
 834          * vma's reference to the file) can go away as soon as we drop
 835          * mmap_sem.
 836          */
 837         get_file(f);
 838         if (userfaultfd_remove(vma, start, end)) {
 839                 /* mmap_sem was not released by userfaultfd_remove() */
 840                 up_read(&current->mm->mmap_sem);
 841         }
 842         error = vfs_fallocate(f,
 843                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 844                                 offset, end - start);
 845         fput(f);
 846         down_read(&current->mm->mmap_sem);
 847         return error;
 848 }
 849
 850 #ifdef CONFIG_MEMORY_FAILURE
 851 /*
 852  * Error injection support for memory error handling.
 853  */
 854 static int madvise_inject_error(int behavior,
 855                 unsigned long start, unsigned long end)
 856 {
 857         struct page *page;
 858         struct zone *zone;
 859         unsigned int order;
 860
 861         if (!capable(CAP_SYS_ADMIN))
 862                 return -EPERM;
 863
 864
 865         for (; start < end; start += PAGE_SIZE << order) {
 866                 unsigned long pfn;
 867                 int ret;
 868
 869                 ret = get_user_pages_fast(start, 1, 0, &page);
 870                 if (ret != 1)
 871                         return ret;
 872                 pfn = page_to_pfn(page);
 873
 874                 /*
 875                  * When soft offlining hugepages, after migrating the page
 876                  * we dissolve it, therefore in the second loop "page" will
 877                  * no longer be a compound page, and order will be 0.
 878                  */
 879                 order = compound_order(compound_head(page));
 880
 881                 if (PageHWPoison(page)) {
 882                         put_page(page);
 883                         continue;
 884                 }
 885
 886                 if (behavior == MADV_SOFT_OFFLINE) {
 887                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 888                                         pfn, start);
 889
 890                         ret = soft_offline_page(page, MF_COUNT_INCREASED);
 891                         if (ret)
 892                                 return ret;
 893                         continue;
 894                 }
 895
 896                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 897                                 pfn, start);
 898
 899                 /*
 900                  * Drop the page reference taken by get_user_pages_fast(). In
 901                  * the absence of MF_COUNT_INCREASED the memory_failure()
 902                  * routine is responsible for pinning the page to prevent it
 903                  * from being released back to the page allocator.
 904                  */
 905                 put_page(page);
 906                 ret = memory_failure(pfn, 0);
 907                 if (ret)
 908                         return ret;
 909         }
 910
 911         /* Ensure that all poisoned pages are removed from per-cpu lists */
 912         for_each_populated_zone(zone)
 913                 drain_all_pages(zone);
 914
 915         return 0;
 916 }
 917 #endif
 918
 919 static long
 920 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 921                 unsigned long start, unsigned long end, int behavior)
 922 {
 923         switch (behavior) {
 924         case MADV_REMOVE:
 925                 return madvise_remove(vma, prev, start, end);
 926         case MADV_WILLNEED:
 927                 return madvise_willneed(vma, prev, start, end);
 928         case MADV_COLD:
 929                 return madvise_cold(vma, prev, start, end);
 930         case MADV_PAGEOUT:
 931                 return madvise_pageout(vma, prev, start, end);
 932         case MADV_FREE:
 933         case MADV_DONTNEED:
 934                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 935         default:
 936                 return madvise_behavior(vma, prev, start, end, behavior);
 937         }
 938 }
 939
 940 static bool
 941 madvise_behavior_valid(int behavior)
 942 {
 943         switch (behavior) {
 944         case MADV_DOFORK:
 945         case MADV_DONTFORK:
 946         case MADV_NORMAL:
 947         case MADV_SEQUENTIAL:
 948         case MADV_RANDOM:
 949         case MADV_REMOVE:
 950         case MADV_WILLNEED:
 951         case MADV_DONTNEED:
 952         case MADV_FREE:
 953         case MADV_COLD:
 954         case MADV_PAGEOUT:
 955 #ifdef CONFIG_KSM
 956         case MADV_MERGEABLE:
 957         case MADV_UNMERGEABLE:
 958 #endif
 959 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 960         case MADV_HUGEPAGE:
 961         case MADV_NOHUGEPAGE:
 962 #endif
 963         case MADV_DONTDUMP:
 964         case MADV_DODUMP:
 965         case MADV_WIPEONFORK:
 966         case MADV_KEEPONFORK:
 967 #ifdef CONFIG_MEMORY_FAILURE
 968         case MADV_SOFT_OFFLINE:
 969         case MADV_HWPOISON:
 970 #endif
 971                 return true;
 972
 973         default:
 974                 return false;
 975         }
 976 }
 977
 978 /*
 979  * The madvise(2) system call.
 980  *
 981  * Applications can use madvise() to advise the kernel how it should
 982  * handle paging I/O in this VM area.  The idea is to help the kernel
 983  * use appropriate read-ahead and caching techniques.  The information
 984  * provided is advisory only, and can be safely disregarded by the
 985  * kernel without affecting the correct operation of the application.
 986  *
 987  * behavior values:
 988  *  MADV_NORMAL - the default behavior is to read clusters.  This
 989  *              results in some read-ahead and read-behind.
 990  *  MADV_RANDOM - the system should read the minimum amount of data
 991  *              on any access, since it is unlikely that the appli-
 992  *              cation will need more than what it asks for.
 993  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
 994  *              once, so they can be aggressively read ahead, and
 995  *              can be freed soon after they are accessed.
 996  *  MADV_WILLNEED - the application is notifying the system to read
 997  *              some pages ahead.
 998  *  MADV_DONTNEED - the application is finished with the given range,
 999  *              so the kernel can free resources associated with it.
1000  *  MADV_FREE - the application marks pages in the given range as lazy free,
1001  *              where actual purges are postponed until memory pressure happens.
1002  *  MADV_REMOVE - the application wants to free up the given range of
1003  *              pages and associated backing store.
1004  *  MADV_DONTFORK - omit this area from child's address space when forking:
1005  *              typically, to avoid COWing pages pinned by get_user_pages().
1006  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1007  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1008  *              range after a fork.
1009  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1010  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1011  *              were corrupted by unrecoverable hardware memory failure.
1012  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1013  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1014  *              this area with pages of identical content from other such areas.
1015  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1016  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1017  *              huge pages in the future. Existing pages might be coalesced and
1018  *              new pages might be allocated as THP.
1019  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1020  *              transparent huge pages so the existing pages will not be
1021  *              coalesced into THP and new pages will not be allocated as THP.
1022  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1023  *              from being included in its core dump.
1024  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1025  *
1026  * return values:
1027  *  zero    - success
1028  *  -EINVAL - start + len < 0, start is not page-aligned,
1029  *              "behavior" is not a valid value, or application
1030  *              is attempting to release locked or shared pages,
1031  *              or the specified address range includes file, Huge TLB,
1032  *              MAP_SHARED or VMPFNMAP range.
1033  *  -ENOMEM - addresses in the specified range are not currently
1034  *              mapped, or are outside the AS of the process.
1035  *  -EIO    - an I/O error occurred while paging in data.
1036  *  -EBADF  - map exists, but area maps something that isn't a file.
1037  *  -EAGAIN - a kernel resource was temporarily unavailable.
1038  */
1039 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1040 {
1041         unsigned long end, tmp;
1042         struct vm_area_struct *vma, *prev;
1043         int unmapped_error = 0;
1044         int error = -EINVAL;
1045         int write;
1046         size_t len;
1047         struct blk_plug plug;
1048
1049         start = untagged_addr(start);
1050
1051         if (!madvise_behavior_valid(behavior))
1052                 return error;
1053
1054         if (start & ~PAGE_MASK)
1055                 return error;
1056         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1057
1058         /* Check to see whether len was rounded up from small -ve to zero */
1059         if (len_in && !len)
1060                 return error;
1061
1062         end = start + len;
1063         if (end < start)
1064                 return error;
1065
1066         error = 0;
1067         if (end == start)
1068                 return error;
1069
1070 #ifdef CONFIG_MEMORY_FAILURE
1071         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1072                 return madvise_inject_error(behavior, start, start + len_in);
1073 #endif
1074
1075         write = madvise_need_mmap_write(behavior);
1076         if (write) {
1077                 if (down_write_killable(&current->mm->mmap_sem))
1078                         return -EINTR;
1079         } else {
1080                 down_read(&current->mm->mmap_sem);
1081         }
1082
1083         /*
1084          * If the interval [start,end) covers some unmapped address
1085          * ranges, just ignore them, but return -ENOMEM at the end.
1086          * - different from the way of handling in mlock etc.
1087          */
1088         vma = find_vma_prev(current->mm, start, &prev);
1089         if (vma && start > vma->vm_start)
1090                 prev = vma;
1091
1092         blk_start_plug(&plug);
1093         for (;;) {
1094                 /* Still start < end. */
1095                 error = -ENOMEM;
1096                 if (!vma)
1097                         goto out;
1098
1099                 /* Here start < (end|vma->vm_end). */
1100                 if (start < vma->vm_start) {
1101                         unmapped_error = -ENOMEM;
1102                         start = vma->vm_start;
1103                         if (start >= end)
1104                                 goto out;
1105                 }
1106
1107                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1108                 tmp = vma->vm_end;
1109                 if (end < tmp)
1110                         tmp = end;
1111
1112                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1113                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1114                 if (error)
1115                         goto out;
1116                 start = tmp;
1117                 if (prev && start < prev->vm_end)
1118                         start = prev->vm_end;
1119                 error = unmapped_error;
1120                 if (start >= end)
1121                         goto out;
1122                 if (prev)
1123                         vma = prev->vm_next;
1124                 else    /* madvise_remove dropped mmap_sem */
1125                         vma = find_vma(current->mm, start);
1126         }
1127 out:
1128         blk_finish_plug(&plug);
1129         if (write)
1130                 up_write(&current->mm->mmap_sem);
1131         else
1132                 up_read(&current->mm->mmap_sem);
1133
1134         return error;
1135 }