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 #include <linux/sched/mm.h>
  31
  32 #include <asm/tlb.h>
  33
  34 #include "internal.h"
  35
  36 struct madvise_walk_private {
  37         struct mmu_gather *tlb;
  38         bool pageout;
  39 };
  40
  41 /*
  42  * Any behaviour which results in changes to the vma->vm_flags needs to
  43  * take mmap_lock for writing. Others, which simply traverse vmas, need
  44  * to only take it for reading.
  45  */
  46 static int madvise_need_mmap_write(int behavior)
  47 {
  48         switch (behavior) {
  49         case MADV_REMOVE:
  50         case MADV_WILLNEED:
  51         case MADV_DONTNEED:
  52         case MADV_COLD:
  53         case MADV_PAGEOUT:
  54         case MADV_FREE:
  55                 return 0;
  56         default:
  57                 /* be safe, default to 1. list exceptions explicitly */
  58                 return 1;
  59         }
  60 }
  61
  62 /*
  63  * We can potentially split a vm area into separate
  64  * areas, each area with its own behavior.
  65  */
  66 static long madvise_behavior(struct vm_area_struct *vma,
  67                      struct vm_area_struct **prev,
  68                      unsigned long start, unsigned long end, int behavior)
  69 {
  70         struct mm_struct *mm = vma->vm_mm;
  71         int error = 0;
  72         pgoff_t pgoff;
  73         unsigned long new_flags = vma->vm_flags;
  74
  75         switch (behavior) {
  76         case MADV_NORMAL:
  77                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  78                 break;
  79         case MADV_SEQUENTIAL:
  80                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  81                 break;
  82         case MADV_RANDOM:
  83                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  84                 break;
  85         case MADV_DONTFORK:
  86                 new_flags |= VM_DONTCOPY;
  87                 break;
  88         case MADV_DOFORK:
  89                 if (vma->vm_flags & VM_IO) {
  90                         error = -EINVAL;
  91                         goto out;
  92                 }
  93                 new_flags &= ~VM_DONTCOPY;
  94                 break;
  95         case MADV_WIPEONFORK:
  96                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  97                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  98                         error = -EINVAL;
  99                         goto out;
 100                 }
 101                 new_flags |= VM_WIPEONFORK;
 102                 break;
 103         case MADV_KEEPONFORK:
 104                 new_flags &= ~VM_WIPEONFORK;
 105                 break;
 106         case MADV_DONTDUMP:
 107                 new_flags |= VM_DONTDUMP;
 108                 break;
 109         case MADV_DODUMP:
 110                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 111                         error = -EINVAL;
 112                         goto out;
 113                 }
 114                 new_flags &= ~VM_DONTDUMP;
 115                 break;
 116         case MADV_MERGEABLE:
 117         case MADV_UNMERGEABLE:
 118                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 119                 if (error)
 120                         goto out_convert_errno;
 121                 break;
 122         case MADV_HUGEPAGE:
 123         case MADV_NOHUGEPAGE:
 124                 error = hugepage_madvise(vma, &new_flags, behavior);
 125                 if (error)
 126                         goto out_convert_errno;
 127                 break;
 128         }
 129
 130         if (new_flags == vma->vm_flags) {
 131                 *prev = vma;
 132                 goto out;
 133         }
 134
 135         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 136         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 137                           vma->vm_file, pgoff, vma_policy(vma),
 138                           vma->vm_userfaultfd_ctx);
 139         if (*prev) {
 140                 vma = *prev;
 141                 goto success;
 142         }
 143
 144         *prev = vma;
 145
 146         if (start != vma->vm_start) {
 147                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 148                         error = -ENOMEM;
 149                         goto out;
 150                 }
 151                 error = __split_vma(mm, vma, start, 1);
 152                 if (error)
 153                         goto out_convert_errno;
 154         }
 155
 156         if (end != vma->vm_end) {
 157                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 158                         error = -ENOMEM;
 159                         goto out;
 160                 }
 161                 error = __split_vma(mm, vma, end, 0);
 162                 if (error)
 163                         goto out_convert_errno;
 164         }
 165
 166 success:
 167         /*
 168          * vm_flags is protected by the mmap_lock held in write mode.
 169          */
 170         vma->vm_flags = new_flags;
 171
 172 out_convert_errno:
 173         /*
 174          * madvise() returns EAGAIN if kernel resources, such as
 175          * slab, are temporarily unavailable.
 176          */
 177         if (error == -ENOMEM)
 178                 error = -EAGAIN;
 179 out:
 180         return error;
 181 }
 182
 183 #ifdef CONFIG_SWAP
 184 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 185         unsigned long end, struct mm_walk *walk)
 186 {
 187         pte_t *orig_pte;
 188         struct vm_area_struct *vma = walk->private;
 189         unsigned long index;
 190
 191         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 192                 return 0;
 193
 194         for (index = start; index != end; index += PAGE_SIZE) {
 195                 pte_t pte;
 196                 swp_entry_t entry;
 197                 struct page *page;
 198                 spinlock_t *ptl;
 199
 200                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 201                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 202                 pte_unmap_unlock(orig_pte, ptl);
 203
 204                 if (pte_present(pte) || pte_none(pte))
 205                         continue;
 206                 entry = pte_to_swp_entry(pte);
 207                 if (unlikely(non_swap_entry(entry)))
 208                         continue;
 209
 210                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 211                                                         vma, index, false);
 212                 if (page)
 213                         put_page(page);
 214         }
 215
 216         return 0;
 217 }
 218
 219 static const struct mm_walk_ops swapin_walk_ops = {
 220         .pmd_entry              = swapin_walk_pmd_entry,
 221 };
 222
 223 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 224                 unsigned long start, unsigned long end,
 225                 struct address_space *mapping)
 226 {
 227         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 228         pgoff_t end_index = end / PAGE_SIZE;
 229         struct page *page;
 230
 231         rcu_read_lock();
 232         xas_for_each(&xas, page, end_index) {
 233                 swp_entry_t swap;
 234
 235                 if (!xa_is_value(page))
 236                         continue;
 237                 xas_pause(&xas);
 238                 rcu_read_unlock();
 239
 240                 swap = radix_to_swp_entry(page);
 241                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 242                                                         NULL, 0, false);
 243                 if (page)
 244                         put_page(page);
 245
 246                 rcu_read_lock();
 247         }
 248         rcu_read_unlock();
 249
 250         lru_add_drain();        /* Push any new pages onto the LRU now */
 251 }
 252 #endif          /* CONFIG_SWAP */
 253
 254 /*
 255  * Schedule all required I/O operations.  Do not wait for completion.
 256  */
 257 static long madvise_willneed(struct vm_area_struct *vma,
 258                              struct vm_area_struct **prev,
 259                              unsigned long start, unsigned long end)
 260 {
 261         struct file *file = vma->vm_file;
 262         loff_t offset;
 263
 264         *prev = vma;
 265 #ifdef CONFIG_SWAP
 266         if (!file) {
 267                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 268                 lru_add_drain(); /* Push any new pages onto the LRU now */
 269                 return 0;
 270         }
 271
 272         if (shmem_mapping(file->f_mapping)) {
 273                 force_shm_swapin_readahead(vma, start, end,
 274                                         file->f_mapping);
 275                 return 0;
 276         }
 277 #else
 278         if (!file)
 279                 return -EBADF;
 280 #endif
 281
 282         if (IS_DAX(file_inode(file))) {
 283                 /* no bad return value, but ignore advice */
 284                 return 0;
 285         }
 286
 287         /*
 288          * Filesystem's fadvise may need to take various locks.  We need to
 289          * explicitly grab a reference because the vma (and hence the
 290          * vma's reference to the file) can go away as soon as we drop
 291          * mmap_lock.
 292          */
 293         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 294         get_file(file);
 295         offset = (loff_t)(start - vma->vm_start)
 296                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 297         mmap_read_unlock(current->mm);
 298         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 299         fput(file);
 300         mmap_read_lock(current->mm);
 301         return 0;
 302 }
 303
 304 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 305                                 unsigned long addr, unsigned long end,
 306                                 struct mm_walk *walk)
 307 {
 308         struct madvise_walk_private *private = walk->private;
 309         struct mmu_gather *tlb = private->tlb;
 310         bool pageout = private->pageout;
 311         struct mm_struct *mm = tlb->mm;
 312         struct vm_area_struct *vma = walk->vma;
 313         pte_t *orig_pte, *pte, ptent;
 314         spinlock_t *ptl;
 315         struct page *page = NULL;
 316         LIST_HEAD(page_list);
 317
 318         if (fatal_signal_pending(current))
 319                 return -EINTR;
 320
 321 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 322         if (pmd_trans_huge(*pmd)) {
 323                 pmd_t orig_pmd;
 324                 unsigned long next = pmd_addr_end(addr, end);
 325
 326                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 327                 ptl = pmd_trans_huge_lock(pmd, vma);
 328                 if (!ptl)
 329                         return 0;
 330
 331                 orig_pmd = *pmd;
 332                 if (is_huge_zero_pmd(orig_pmd))
 333                         goto huge_unlock;
 334
 335                 if (unlikely(!pmd_present(orig_pmd))) {
 336                         VM_BUG_ON(thp_migration_supported() &&
 337                                         !is_pmd_migration_entry(orig_pmd));
 338                         goto huge_unlock;
 339                 }
 340
 341                 page = pmd_page(orig_pmd);
 342
 343                 /* Do not interfere with other mappings of this page */
 344                 if (page_mapcount(page) != 1)
 345                         goto huge_unlock;
 346
 347                 if (next - addr != HPAGE_PMD_SIZE) {
 348                         int err;
 349
 350                         get_page(page);
 351                         spin_unlock(ptl);
 352                         lock_page(page);
 353                         err = split_huge_page(page);
 354                         unlock_page(page);
 355                         put_page(page);
 356                         if (!err)
 357                                 goto regular_page;
 358                         return 0;
 359                 }
 360
 361                 if (pmd_young(orig_pmd)) {
 362                         pmdp_invalidate(vma, addr, pmd);
 363                         orig_pmd = pmd_mkold(orig_pmd);
 364
 365                         set_pmd_at(mm, addr, pmd, orig_pmd);
 366                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 367                 }
 368
 369                 ClearPageReferenced(page);
 370                 test_and_clear_page_young(page);
 371                 if (pageout) {
 372                         if (!isolate_lru_page(page)) {
 373                                 if (PageUnevictable(page))
 374                                         putback_lru_page(page);
 375                                 else
 376                                         list_add(&page->lru, &page_list);
 377                         }
 378                 } else
 379                         deactivate_page(page);
 380 huge_unlock:
 381                 spin_unlock(ptl);
 382                 if (pageout)
 383                         reclaim_pages(&page_list);
 384                 return 0;
 385         }
 386
 387 regular_page:
 388         if (pmd_trans_unstable(pmd))
 389                 return 0;
 390 #endif
 391         tlb_change_page_size(tlb, PAGE_SIZE);
 392         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 393         flush_tlb_batched_pending(mm);
 394         arch_enter_lazy_mmu_mode();
 395         for (; addr < end; pte++, addr += PAGE_SIZE) {
 396                 ptent = *pte;
 397
 398                 if (pte_none(ptent))
 399                         continue;
 400
 401                 if (!pte_present(ptent))
 402                         continue;
 403
 404                 page = vm_normal_page(vma, addr, ptent);
 405                 if (!page)
 406                         continue;
 407
 408                 /*
 409                  * Creating a THP page is expensive so split it only if we
 410                  * are sure it's worth. Split it if we are only owner.
 411                  */
 412                 if (PageTransCompound(page)) {
 413                         if (page_mapcount(page) != 1)
 414                                 break;
 415                         get_page(page);
 416                         if (!trylock_page(page)) {
 417                                 put_page(page);
 418                                 break;
 419                         }
 420                         pte_unmap_unlock(orig_pte, ptl);
 421                         if (split_huge_page(page)) {
 422                                 unlock_page(page);
 423                                 put_page(page);
 424                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 425                                 break;
 426                         }
 427                         unlock_page(page);
 428                         put_page(page);
 429                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 430                         pte--;
 431                         addr -= PAGE_SIZE;
 432                         continue;
 433                 }
 434
 435                 /* Do not interfere with other mappings of this page */
 436                 if (page_mapcount(page) != 1)
 437                         continue;
 438
 439                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 440
 441                 if (pte_young(ptent)) {
 442                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 443                                                         tlb->fullmm);
 444                         ptent = pte_mkold(ptent);
 445                         set_pte_at(mm, addr, pte, ptent);
 446                         tlb_remove_tlb_entry(tlb, pte, addr);
 447                 }
 448
 449                 /*
 450                  * We are deactivating a page for accelerating reclaiming.
 451                  * VM couldn't reclaim the page unless we clear PG_young.
 452                  * As a side effect, it makes confuse idle-page tracking
 453                  * because they will miss recent referenced history.
 454                  */
 455                 ClearPageReferenced(page);
 456                 test_and_clear_page_young(page);
 457                 if (pageout) {
 458                         if (!isolate_lru_page(page)) {
 459                                 if (PageUnevictable(page))
 460                                         putback_lru_page(page);
 461                                 else
 462                                         list_add(&page->lru, &page_list);
 463                         }
 464                 } else
 465                         deactivate_page(page);
 466         }
 467
 468         arch_leave_lazy_mmu_mode();
 469         pte_unmap_unlock(orig_pte, ptl);
 470         if (pageout)
 471                 reclaim_pages(&page_list);
 472         cond_resched();
 473
 474         return 0;
 475 }
 476
 477 static const struct mm_walk_ops cold_walk_ops = {
 478         .pmd_entry = madvise_cold_or_pageout_pte_range,
 479 };
 480
 481 static void madvise_cold_page_range(struct mmu_gather *tlb,
 482                              struct vm_area_struct *vma,
 483                              unsigned long addr, unsigned long end)
 484 {
 485         struct madvise_walk_private walk_private = {
 486                 .pageout = false,
 487                 .tlb = tlb,
 488         };
 489
 490         tlb_start_vma(tlb, vma);
 491         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 492         tlb_end_vma(tlb, vma);
 493 }
 494
 495 static long madvise_cold(struct vm_area_struct *vma,
 496                         struct vm_area_struct **prev,
 497                         unsigned long start_addr, unsigned long end_addr)
 498 {
 499         struct mm_struct *mm = vma->vm_mm;
 500         struct mmu_gather tlb;
 501
 502         *prev = vma;
 503         if (!can_madv_lru_vma(vma))
 504                 return -EINVAL;
 505
 506         lru_add_drain();
 507         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 508         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 509         tlb_finish_mmu(&tlb, start_addr, end_addr);
 510
 511         return 0;
 512 }
 513
 514 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 515                              struct vm_area_struct *vma,
 516                              unsigned long addr, unsigned long end)
 517 {
 518         struct madvise_walk_private walk_private = {
 519                 .pageout = true,
 520                 .tlb = tlb,
 521         };
 522
 523         tlb_start_vma(tlb, vma);
 524         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 525         tlb_end_vma(tlb, vma);
 526 }
 527
 528 static inline bool can_do_pageout(struct vm_area_struct *vma)
 529 {
 530         if (vma_is_anonymous(vma))
 531                 return true;
 532         if (!vma->vm_file)
 533                 return false;
 534         /*
 535          * paging out pagecache only for non-anonymous mappings that correspond
 536          * to the files the calling process could (if tried) open for writing;
 537          * otherwise we'd be including shared non-exclusive mappings, which
 538          * opens a side channel.
 539          */
 540         return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 541                 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 542 }
 543
 544 static long madvise_pageout(struct vm_area_struct *vma,
 545                         struct vm_area_struct **prev,
 546                         unsigned long start_addr, unsigned long end_addr)
 547 {
 548         struct mm_struct *mm = vma->vm_mm;
 549         struct mmu_gather tlb;
 550
 551         *prev = vma;
 552         if (!can_madv_lru_vma(vma))
 553                 return -EINVAL;
 554
 555         if (!can_do_pageout(vma))
 556                 return 0;
 557
 558         lru_add_drain();
 559         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 560         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 561         tlb_finish_mmu(&tlb, start_addr, end_addr);
 562
 563         return 0;
 564 }
 565
 566 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 567                                 unsigned long end, struct mm_walk *walk)
 568
 569 {
 570         struct mmu_gather *tlb = walk->private;
 571         struct mm_struct *mm = tlb->mm;
 572         struct vm_area_struct *vma = walk->vma;
 573         spinlock_t *ptl;
 574         pte_t *orig_pte, *pte, ptent;
 575         struct page *page;
 576         int nr_swap = 0;
 577         unsigned long next;
 578
 579         next = pmd_addr_end(addr, end);
 580         if (pmd_trans_huge(*pmd))
 581                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 582                         goto next;
 583
 584         if (pmd_trans_unstable(pmd))
 585                 return 0;
 586
 587         tlb_change_page_size(tlb, PAGE_SIZE);
 588         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 589         flush_tlb_batched_pending(mm);
 590         arch_enter_lazy_mmu_mode();
 591         for (; addr != end; pte++, addr += PAGE_SIZE) {
 592                 ptent = *pte;
 593
 594                 if (pte_none(ptent))
 595                         continue;
 596                 /*
 597                  * If the pte has swp_entry, just clear page table to
 598                  * prevent swap-in which is more expensive rather than
 599                  * (page allocation + zeroing).
 600                  */
 601                 if (!pte_present(ptent)) {
 602                         swp_entry_t entry;
 603
 604                         entry = pte_to_swp_entry(ptent);
 605                         if (non_swap_entry(entry))
 606                                 continue;
 607                         nr_swap--;
 608                         free_swap_and_cache(entry);
 609                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 610                         continue;
 611                 }
 612
 613                 page = vm_normal_page(vma, addr, ptent);
 614                 if (!page)
 615                         continue;
 616
 617                 /*
 618                  * If pmd isn't transhuge but the page is THP and
 619                  * is owned by only this process, split it and
 620                  * deactivate all pages.
 621                  */
 622                 if (PageTransCompound(page)) {
 623                         if (page_mapcount(page) != 1)
 624                                 goto out;
 625                         get_page(page);
 626                         if (!trylock_page(page)) {
 627                                 put_page(page);
 628                                 goto out;
 629                         }
 630                         pte_unmap_unlock(orig_pte, ptl);
 631                         if (split_huge_page(page)) {
 632                                 unlock_page(page);
 633                                 put_page(page);
 634                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 635                                 goto out;
 636                         }
 637                         unlock_page(page);
 638                         put_page(page);
 639                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 640                         pte--;
 641                         addr -= PAGE_SIZE;
 642                         continue;
 643                 }
 644
 645                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 646
 647                 if (PageSwapCache(page) || PageDirty(page)) {
 648                         if (!trylock_page(page))
 649                                 continue;
 650                         /*
 651                          * If page is shared with others, we couldn't clear
 652                          * PG_dirty of the page.
 653                          */
 654                         if (page_mapcount(page) != 1) {
 655                                 unlock_page(page);
 656                                 continue;
 657                         }
 658
 659                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 660                                 unlock_page(page);
 661                                 continue;
 662                         }
 663
 664                         ClearPageDirty(page);
 665                         unlock_page(page);
 666                 }
 667
 668                 if (pte_young(ptent) || pte_dirty(ptent)) {
 669                         /*
 670                          * Some of architecture(ex, PPC) don't update TLB
 671                          * with set_pte_at and tlb_remove_tlb_entry so for
 672                          * the portability, remap the pte with old|clean
 673                          * after pte clearing.
 674                          */
 675                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 676                                                         tlb->fullmm);
 677
 678                         ptent = pte_mkold(ptent);
 679                         ptent = pte_mkclean(ptent);
 680                         set_pte_at(mm, addr, pte, ptent);
 681                         tlb_remove_tlb_entry(tlb, pte, addr);
 682                 }
 683                 mark_page_lazyfree(page);
 684         }
 685 out:
 686         if (nr_swap) {
 687                 if (current->mm == mm)
 688                         sync_mm_rss(mm);
 689
 690                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 691         }
 692         arch_leave_lazy_mmu_mode();
 693         pte_unmap_unlock(orig_pte, ptl);
 694         cond_resched();
 695 next:
 696         return 0;
 697 }
 698
 699 static const struct mm_walk_ops madvise_free_walk_ops = {
 700         .pmd_entry              = madvise_free_pte_range,
 701 };
 702
 703 static int madvise_free_single_vma(struct vm_area_struct *vma,
 704                         unsigned long start_addr, unsigned long end_addr)
 705 {
 706         struct mm_struct *mm = vma->vm_mm;
 707         struct mmu_notifier_range range;
 708         struct mmu_gather tlb;
 709
 710         /* MADV_FREE works for only anon vma at the moment */
 711         if (!vma_is_anonymous(vma))
 712                 return -EINVAL;
 713
 714         range.start = max(vma->vm_start, start_addr);
 715         if (range.start >= vma->vm_end)
 716                 return -EINVAL;
 717         range.end = min(vma->vm_end, end_addr);
 718         if (range.end <= vma->vm_start)
 719                 return -EINVAL;
 720         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 721                                 range.start, range.end);
 722
 723         lru_add_drain();
 724         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 725         update_hiwater_rss(mm);
 726
 727         mmu_notifier_invalidate_range_start(&range);
 728         tlb_start_vma(&tlb, vma);
 729         walk_page_range(vma->vm_mm, range.start, range.end,
 730                         &madvise_free_walk_ops, &tlb);
 731         tlb_end_vma(&tlb, vma);
 732         mmu_notifier_invalidate_range_end(&range);
 733         tlb_finish_mmu(&tlb, range.start, range.end);
 734
 735         return 0;
 736 }
 737
 738 /*
 739  * Application no longer needs these pages.  If the pages are dirty,
 740  * it's OK to just throw them away.  The app will be more careful about
 741  * data it wants to keep.  Be sure to free swap resources too.  The
 742  * zap_page_range call sets things up for shrink_active_list to actually free
 743  * these pages later if no one else has touched them in the meantime,
 744  * although we could add these pages to a global reuse list for
 745  * shrink_active_list to pick up before reclaiming other pages.
 746  *
 747  * NB: This interface discards data rather than pushes it out to swap,
 748  * as some implementations do.  This has performance implications for
 749  * applications like large transactional databases which want to discard
 750  * pages in anonymous maps after committing to backing store the data
 751  * that was kept in them.  There is no reason to write this data out to
 752  * the swap area if the application is discarding it.
 753  *
 754  * An interface that causes the system to free clean pages and flush
 755  * dirty pages is already available as msync(MS_INVALIDATE).
 756  */
 757 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 758                                         unsigned long start, unsigned long end)
 759 {
 760         zap_page_range(vma, start, end - start);
 761         return 0;
 762 }
 763
 764 static long madvise_dontneed_free(struct vm_area_struct *vma,
 765                                   struct vm_area_struct **prev,
 766                                   unsigned long start, unsigned long end,
 767                                   int behavior)
 768 {
 769         *prev = vma;
 770         if (!can_madv_lru_vma(vma))
 771                 return -EINVAL;
 772
 773         if (!userfaultfd_remove(vma, start, end)) {
 774                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 775
 776                 mmap_read_lock(current->mm);
 777                 vma = find_vma(current->mm, start);
 778                 if (!vma)
 779                         return -ENOMEM;
 780                 if (start < vma->vm_start) {
 781                         /*
 782                          * This "vma" under revalidation is the one
 783                          * with the lowest vma->vm_start where start
 784                          * is also < vma->vm_end. If start <
 785                          * vma->vm_start it means an hole materialized
 786                          * in the user address space within the
 787                          * virtual range passed to MADV_DONTNEED
 788                          * or MADV_FREE.
 789                          */
 790                         return -ENOMEM;
 791                 }
 792                 if (!can_madv_lru_vma(vma))
 793                         return -EINVAL;
 794                 if (end > vma->vm_end) {
 795                         /*
 796                          * Don't fail if end > vma->vm_end. If the old
 797                          * vma was splitted while the mmap_lock was
 798                          * released the effect of the concurrent
 799                          * operation may not cause madvise() to
 800                          * have an undefined result. There may be an
 801                          * adjacent next vma that we'll walk
 802                          * next. userfaultfd_remove() will generate an
 803                          * UFFD_EVENT_REMOVE repetition on the
 804                          * end-vma->vm_end range, but the manager can
 805                          * handle a repetition fine.
 806                          */
 807                         end = vma->vm_end;
 808                 }
 809                 VM_WARN_ON(start >= end);
 810         }
 811
 812         if (behavior == MADV_DONTNEED)
 813                 return madvise_dontneed_single_vma(vma, start, end);
 814         else if (behavior == MADV_FREE)
 815                 return madvise_free_single_vma(vma, start, end);
 816         else
 817                 return -EINVAL;
 818 }
 819
 820 /*
 821  * Application wants to free up the pages and associated backing store.
 822  * This is effectively punching a hole into the middle of a file.
 823  */
 824 static long madvise_remove(struct vm_area_struct *vma,
 825                                 struct vm_area_struct **prev,
 826                                 unsigned long start, unsigned long end)
 827 {
 828         loff_t offset;
 829         int error;
 830         struct file *f;
 831
 832         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 833
 834         if (vma->vm_flags & VM_LOCKED)
 835                 return -EINVAL;
 836
 837         f = vma->vm_file;
 838
 839         if (!f || !f->f_mapping || !f->f_mapping->host) {
 840                         return -EINVAL;
 841         }
 842
 843         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 844                 return -EACCES;
 845
 846         offset = (loff_t)(start - vma->vm_start)
 847                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 848
 849         /*
 850          * Filesystem's fallocate may need to take i_mutex.  We need to
 851          * explicitly grab a reference because the vma (and hence the
 852          * vma's reference to the file) can go away as soon as we drop
 853          * mmap_lock.
 854          */
 855         get_file(f);
 856         if (userfaultfd_remove(vma, start, end)) {
 857                 /* mmap_lock was not released by userfaultfd_remove() */
 858                 mmap_read_unlock(current->mm);
 859         }
 860         error = vfs_fallocate(f,
 861                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 862                                 offset, end - start);
 863         fput(f);
 864         mmap_read_lock(current->mm);
 865         return error;
 866 }
 867
 868 #ifdef CONFIG_MEMORY_FAILURE
 869 /*
 870  * Error injection support for memory error handling.
 871  */
 872 static int madvise_inject_error(int behavior,
 873                 unsigned long start, unsigned long end)
 874 {
 875         struct page *page;
 876         struct zone *zone;
 877         unsigned long size;
 878
 879         if (!capable(CAP_SYS_ADMIN))
 880                 return -EPERM;
 881
 882
 883         for (; start < end; start += size) {
 884                 unsigned long pfn;
 885                 int ret;
 886
 887                 ret = get_user_pages_fast(start, 1, 0, &page);
 888                 if (ret != 1)
 889                         return ret;
 890                 pfn = page_to_pfn(page);
 891
 892                 /*
 893                  * When soft offlining hugepages, after migrating the page
 894                  * we dissolve it, therefore in the second loop "page" will
 895                  * no longer be a compound page.
 896                  */
 897                 size = page_size(compound_head(page));
 898
 899                 if (PageHWPoison(page)) {
 900                         put_page(page);
 901                         continue;
 902                 }
 903
 904                 if (behavior == MADV_SOFT_OFFLINE) {
 905                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 906                                         pfn, start);
 907
 908                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
 909                         if (ret)
 910                                 return ret;
 911                         continue;
 912                 }
 913
 914                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 915                                 pfn, start);
 916
 917                 /*
 918                  * Drop the page reference taken by get_user_pages_fast(). In
 919                  * the absence of MF_COUNT_INCREASED the memory_failure()
 920                  * routine is responsible for pinning the page to prevent it
 921                  * from being released back to the page allocator.
 922                  */
 923                 put_page(page);
 924                 ret = memory_failure(pfn, 0);
 925                 if (ret)
 926                         return ret;
 927         }
 928
 929         /* Ensure that all poisoned pages are removed from per-cpu lists */
 930         for_each_populated_zone(zone)
 931                 drain_all_pages(zone);
 932
 933         return 0;
 934 }
 935 #endif
 936
 937 static long
 938 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 939                 unsigned long start, unsigned long end, int behavior)
 940 {
 941         switch (behavior) {
 942         case MADV_REMOVE:
 943                 return madvise_remove(vma, prev, start, end);
 944         case MADV_WILLNEED:
 945                 return madvise_willneed(vma, prev, start, end);
 946         case MADV_COLD:
 947                 return madvise_cold(vma, prev, start, end);
 948         case MADV_PAGEOUT:
 949                 return madvise_pageout(vma, prev, start, end);
 950         case MADV_FREE:
 951         case MADV_DONTNEED:
 952                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 953         default:
 954                 return madvise_behavior(vma, prev, start, end, behavior);
 955         }
 956 }
 957
 958 static bool
 959 madvise_behavior_valid(int behavior)
 960 {
 961         switch (behavior) {
 962         case MADV_DOFORK:
 963         case MADV_DONTFORK:
 964         case MADV_NORMAL:
 965         case MADV_SEQUENTIAL:
 966         case MADV_RANDOM:
 967         case MADV_REMOVE:
 968         case MADV_WILLNEED:
 969         case MADV_DONTNEED:
 970         case MADV_FREE:
 971         case MADV_COLD:
 972         case MADV_PAGEOUT:
 973 #ifdef CONFIG_KSM
 974         case MADV_MERGEABLE:
 975         case MADV_UNMERGEABLE:
 976 #endif
 977 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 978         case MADV_HUGEPAGE:
 979         case MADV_NOHUGEPAGE:
 980 #endif
 981         case MADV_DONTDUMP:
 982         case MADV_DODUMP:
 983         case MADV_WIPEONFORK:
 984         case MADV_KEEPONFORK:
 985 #ifdef CONFIG_MEMORY_FAILURE
 986         case MADV_SOFT_OFFLINE:
 987         case MADV_HWPOISON:
 988 #endif
 989                 return true;
 990
 991         default:
 992                 return false;
 993         }
 994 }
 995
 996 /*
 997  * The madvise(2) system call.
 998  *
 999  * Applications can use madvise() to advise the kernel how it should
1000  * handle paging I/O in this VM area.  The idea is to help the kernel
1001  * use appropriate read-ahead and caching techniques.  The information
1002  * provided is advisory only, and can be safely disregarded by the
1003  * kernel without affecting the correct operation of the application.
1004  *
1005  * behavior values:
1006  *  MADV_NORMAL - the default behavior is to read clusters.  This
1007  *              results in some read-ahead and read-behind.
1008  *  MADV_RANDOM - the system should read the minimum amount of data
1009  *              on any access, since it is unlikely that the appli-
1010  *              cation will need more than what it asks for.
1011  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1012  *              once, so they can be aggressively read ahead, and
1013  *              can be freed soon after they are accessed.
1014  *  MADV_WILLNEED - the application is notifying the system to read
1015  *              some pages ahead.
1016  *  MADV_DONTNEED - the application is finished with the given range,
1017  *              so the kernel can free resources associated with it.
1018  *  MADV_FREE - the application marks pages in the given range as lazy free,
1019  *              where actual purges are postponed until memory pressure happens.
1020  *  MADV_REMOVE - the application wants to free up the given range of
1021  *              pages and associated backing store.
1022  *  MADV_DONTFORK - omit this area from child's address space when forking:
1023  *              typically, to avoid COWing pages pinned by get_user_pages().
1024  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1025  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1026  *              range after a fork.
1027  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1028  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1029  *              were corrupted by unrecoverable hardware memory failure.
1030  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1031  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1032  *              this area with pages of identical content from other such areas.
1033  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1034  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1035  *              huge pages in the future. Existing pages might be coalesced and
1036  *              new pages might be allocated as THP.
1037  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1038  *              transparent huge pages so the existing pages will not be
1039  *              coalesced into THP and new pages will not be allocated as THP.
1040  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1041  *              from being included in its core dump.
1042  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1043  *
1044  * return values:
1045  *  zero    - success
1046  *  -EINVAL - start + len < 0, start is not page-aligned,
1047  *              "behavior" is not a valid value, or application
1048  *              is attempting to release locked or shared pages,
1049  *              or the specified address range includes file, Huge TLB,
1050  *              MAP_SHARED or VMPFNMAP range.
1051  *  -ENOMEM - addresses in the specified range are not currently
1052  *              mapped, or are outside the AS of the process.
1053  *  -EIO    - an I/O error occurred while paging in data.
1054  *  -EBADF  - map exists, but area maps something that isn't a file.
1055  *  -EAGAIN - a kernel resource was temporarily unavailable.
1056  */
1057 int do_madvise(unsigned long start, size_t len_in, int behavior)
1058 {
1059         unsigned long end, tmp;
1060         struct vm_area_struct *vma, *prev;
1061         int unmapped_error = 0;
1062         int error = -EINVAL;
1063         int write;
1064         size_t len;
1065         struct blk_plug plug;
1066
1067         start = untagged_addr(start);
1068
1069         if (!madvise_behavior_valid(behavior))
1070                 return error;
1071
1072         if (!PAGE_ALIGNED(start))
1073                 return error;
1074         len = PAGE_ALIGN(len_in);
1075
1076         /* Check to see whether len was rounded up from small -ve to zero */
1077         if (len_in && !len)
1078                 return error;
1079
1080         end = start + len;
1081         if (end < start)
1082                 return error;
1083
1084         error = 0;
1085         if (end == start)
1086                 return error;
1087
1088 #ifdef CONFIG_MEMORY_FAILURE
1089         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1090                 return madvise_inject_error(behavior, start, start + len_in);
1091 #endif
1092
1093         write = madvise_need_mmap_write(behavior);
1094         if (write) {
1095                 if (mmap_write_lock_killable(current->mm))
1096                         return -EINTR;
1097
1098                 /*
1099                  * We may have stolen the mm from another process
1100                  * that is undergoing core dumping.
1101                  *
1102                  * Right now that's io_ring, in the future it may
1103                  * be remote process management and not "current"
1104                  * at all.
1105                  *
1106                  * We need to fix core dumping to not do this,
1107                  * but for now we have the mmget_still_valid()
1108                  * model.
1109                  */
1110                 if (!mmget_still_valid(current->mm)) {
1111                         mmap_write_unlock(current->mm);
1112                         return -EINTR;
1113                 }
1114         } else {
1115                 mmap_read_lock(current->mm);
1116         }
1117
1118         /*
1119          * If the interval [start,end) covers some unmapped address
1120          * ranges, just ignore them, but return -ENOMEM at the end.
1121          * - different from the way of handling in mlock etc.
1122          */
1123         vma = find_vma_prev(current->mm, start, &prev);
1124         if (vma && start > vma->vm_start)
1125                 prev = vma;
1126
1127         blk_start_plug(&plug);
1128         for (;;) {
1129                 /* Still start < end. */
1130                 error = -ENOMEM;
1131                 if (!vma)
1132                         goto out;
1133
1134                 /* Here start < (end|vma->vm_end). */
1135                 if (start < vma->vm_start) {
1136                         unmapped_error = -ENOMEM;
1137                         start = vma->vm_start;
1138                         if (start >= end)
1139                                 goto out;
1140                 }
1141
1142                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1143                 tmp = vma->vm_end;
1144                 if (end < tmp)
1145                         tmp = end;
1146
1147                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1148                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1149                 if (error)
1150                         goto out;
1151                 start = tmp;
1152                 if (prev && start < prev->vm_end)
1153                         start = prev->vm_end;
1154                 error = unmapped_error;
1155                 if (start >= end)
1156                         goto out;
1157                 if (prev)
1158                         vma = prev->vm_next;
1159                 else    /* madvise_remove dropped mmap_lock */
1160                         vma = find_vma(current->mm, start);
1161         }
1162 out:
1163         blk_finish_plug(&plug);
1164         if (write)
1165                 mmap_write_unlock(current->mm);
1166         else
1167                 mmap_read_unlock(current->mm);
1168
1169         return error;
1170 }
1171
1172 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1173 {
1174         return do_madvise(start, len_in, behavior);
1175 }