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/sched/mm.h>
  21 #include <linux/uio.h>
  22 #include <linux/ksm.h>
  23 #include <linux/fs.h>
  24 #include <linux/file.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/backing-dev.h>
  27 #include <linux/pagewalk.h>
  28 #include <linux/swap.h>
  29 #include <linux/swapops.h>
  30 #include <linux/shmem_fs.h>
  31 #include <linux/mmu_notifier.h>
  32
  33 #include <asm/tlb.h>
  34
  35 #include "internal.h"
  36
  37 struct madvise_walk_private {
  38         struct mmu_gather *tlb;
  39         bool pageout;
  40 };
  41
  42 /*
  43  * Any behaviour which results in changes to the vma->vm_flags needs to
  44  * take mmap_lock for writing. Others, which simply traverse vmas, need
  45  * to only take it for reading.
  46  */
  47 static int madvise_need_mmap_write(int behavior)
  48 {
  49         switch (behavior) {
  50         case MADV_REMOVE:
  51         case MADV_WILLNEED:
  52         case MADV_DONTNEED:
  53         case MADV_COLD:
  54         case MADV_PAGEOUT:
  55         case MADV_FREE:
  56                 return 0;
  57         default:
  58                 /* be safe, default to 1. list exceptions explicitly */
  59                 return 1;
  60         }
  61 }
  62
  63 /*
  64  * We can potentially split a vm area into separate
  65  * areas, each area with its own behavior.
  66  */
  67 static long madvise_behavior(struct vm_area_struct *vma,
  68                      struct vm_area_struct **prev,
  69                      unsigned long start, unsigned long end, int behavior)
  70 {
  71         struct mm_struct *mm = vma->vm_mm;
  72         int error = 0;
  73         pgoff_t pgoff;
  74         unsigned long new_flags = vma->vm_flags;
  75
  76         switch (behavior) {
  77         case MADV_NORMAL:
  78                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  79                 break;
  80         case MADV_SEQUENTIAL:
  81                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  82                 break;
  83         case MADV_RANDOM:
  84                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  85                 break;
  86         case MADV_DONTFORK:
  87                 new_flags |= VM_DONTCOPY;
  88                 break;
  89         case MADV_DOFORK:
  90                 if (vma->vm_flags & VM_IO) {
  91                         error = -EINVAL;
  92                         goto out;
  93                 }
  94                 new_flags &= ~VM_DONTCOPY;
  95                 break;
  96         case MADV_WIPEONFORK:
  97                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  98                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  99                         error = -EINVAL;
 100                         goto out;
 101                 }
 102                 new_flags |= VM_WIPEONFORK;
 103                 break;
 104         case MADV_KEEPONFORK:
 105                 new_flags &= ~VM_WIPEONFORK;
 106                 break;
 107         case MADV_DONTDUMP:
 108                 new_flags |= VM_DONTDUMP;
 109                 break;
 110         case MADV_DODUMP:
 111                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 112                         error = -EINVAL;
 113                         goto out;
 114                 }
 115                 new_flags &= ~VM_DONTDUMP;
 116                 break;
 117         case MADV_MERGEABLE:
 118         case MADV_UNMERGEABLE:
 119                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 120                 if (error)
 121                         goto out_convert_errno;
 122                 break;
 123         case MADV_HUGEPAGE:
 124         case MADV_NOHUGEPAGE:
 125                 error = hugepage_madvise(vma, &new_flags, behavior);
 126                 if (error)
 127                         goto out_convert_errno;
 128                 break;
 129         }
 130
 131         if (new_flags == vma->vm_flags) {
 132                 *prev = vma;
 133                 goto out;
 134         }
 135
 136         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 137         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 138                           vma->vm_file, pgoff, vma_policy(vma),
 139                           vma->vm_userfaultfd_ctx);
 140         if (*prev) {
 141                 vma = *prev;
 142                 goto success;
 143         }
 144
 145         *prev = vma;
 146
 147         if (start != vma->vm_start) {
 148                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 149                         error = -ENOMEM;
 150                         goto out;
 151                 }
 152                 error = __split_vma(mm, vma, start, 1);
 153                 if (error)
 154                         goto out_convert_errno;
 155         }
 156
 157         if (end != vma->vm_end) {
 158                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 159                         error = -ENOMEM;
 160                         goto out;
 161                 }
 162                 error = __split_vma(mm, vma, end, 0);
 163                 if (error)
 164                         goto out_convert_errno;
 165         }
 166
 167 success:
 168         /*
 169          * vm_flags is protected by the mmap_lock held in write mode.
 170          */
 171         vma->vm_flags = new_flags;
 172
 173 out_convert_errno:
 174         /*
 175          * madvise() returns EAGAIN if kernel resources, such as
 176          * slab, are temporarily unavailable.
 177          */
 178         if (error == -ENOMEM)
 179                 error = -EAGAIN;
 180 out:
 181         return error;
 182 }
 183
 184 #ifdef CONFIG_SWAP
 185 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 186         unsigned long end, struct mm_walk *walk)
 187 {
 188         pte_t *orig_pte;
 189         struct vm_area_struct *vma = walk->private;
 190         unsigned long index;
 191
 192         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 193                 return 0;
 194
 195         for (index = start; index != end; index += PAGE_SIZE) {
 196                 pte_t pte;
 197                 swp_entry_t entry;
 198                 struct page *page;
 199                 spinlock_t *ptl;
 200
 201                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 202                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 203                 pte_unmap_unlock(orig_pte, ptl);
 204
 205                 if (pte_present(pte) || pte_none(pte))
 206                         continue;
 207                 entry = pte_to_swp_entry(pte);
 208                 if (unlikely(non_swap_entry(entry)))
 209                         continue;
 210
 211                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 212                                                         vma, index, false);
 213                 if (page)
 214                         put_page(page);
 215         }
 216
 217         return 0;
 218 }
 219
 220 static const struct mm_walk_ops swapin_walk_ops = {
 221         .pmd_entry              = swapin_walk_pmd_entry,
 222 };
 223
 224 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 225                 unsigned long start, unsigned long end,
 226                 struct address_space *mapping)
 227 {
 228         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 229         pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
 230         struct page *page;
 231
 232         rcu_read_lock();
 233         xas_for_each(&xas, page, end_index) {
 234                 swp_entry_t swap;
 235
 236                 if (!xa_is_value(page))
 237                         continue;
 238                 xas_pause(&xas);
 239                 rcu_read_unlock();
 240
 241                 swap = radix_to_swp_entry(page);
 242                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 243                                                         NULL, 0, false);
 244                 if (page)
 245                         put_page(page);
 246
 247                 rcu_read_lock();
 248         }
 249         rcu_read_unlock();
 250
 251         lru_add_drain();        /* Push any new pages onto the LRU now */
 252 }
 253 #endif          /* CONFIG_SWAP */
 254
 255 /*
 256  * Schedule all required I/O operations.  Do not wait for completion.
 257  */
 258 static long madvise_willneed(struct vm_area_struct *vma,
 259                              struct vm_area_struct **prev,
 260                              unsigned long start, unsigned long end)
 261 {
 262         struct mm_struct *mm = vma->vm_mm;
 263         struct file *file = vma->vm_file;
 264         loff_t offset;
 265
 266         *prev = vma;
 267 #ifdef CONFIG_SWAP
 268         if (!file) {
 269                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 270                 lru_add_drain(); /* Push any new pages onto the LRU now */
 271                 return 0;
 272         }
 273
 274         if (shmem_mapping(file->f_mapping)) {
 275                 force_shm_swapin_readahead(vma, start, end,
 276                                         file->f_mapping);
 277                 return 0;
 278         }
 279 #else
 280         if (!file)
 281                 return -EBADF;
 282 #endif
 283
 284         if (IS_DAX(file_inode(file))) {
 285                 /* no bad return value, but ignore advice */
 286                 return 0;
 287         }
 288
 289         /*
 290          * Filesystem's fadvise may need to take various locks.  We need to
 291          * explicitly grab a reference because the vma (and hence the
 292          * vma's reference to the file) can go away as soon as we drop
 293          * mmap_lock.
 294          */
 295         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 296         get_file(file);
 297         offset = (loff_t)(start - vma->vm_start)
 298                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 299         mmap_read_unlock(mm);
 300         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 301         fput(file);
 302         mmap_read_lock(mm);
 303         return 0;
 304 }
 305
 306 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 307                                 unsigned long addr, unsigned long end,
 308                                 struct mm_walk *walk)
 309 {
 310         struct madvise_walk_private *private = walk->private;
 311         struct mmu_gather *tlb = private->tlb;
 312         bool pageout = private->pageout;
 313         struct mm_struct *mm = tlb->mm;
 314         struct vm_area_struct *vma = walk->vma;
 315         pte_t *orig_pte, *pte, ptent;
 316         spinlock_t *ptl;
 317         struct page *page = NULL;
 318         LIST_HEAD(page_list);
 319
 320         if (fatal_signal_pending(current))
 321                 return -EINTR;
 322
 323 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 324         if (pmd_trans_huge(*pmd)) {
 325                 pmd_t orig_pmd;
 326                 unsigned long next = pmd_addr_end(addr, end);
 327
 328                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 329                 ptl = pmd_trans_huge_lock(pmd, vma);
 330                 if (!ptl)
 331                         return 0;
 332
 333                 orig_pmd = *pmd;
 334                 if (is_huge_zero_pmd(orig_pmd))
 335                         goto huge_unlock;
 336
 337                 if (unlikely(!pmd_present(orig_pmd))) {
 338                         VM_BUG_ON(thp_migration_supported() &&
 339                                         !is_pmd_migration_entry(orig_pmd));
 340                         goto huge_unlock;
 341                 }
 342
 343                 page = pmd_page(orig_pmd);
 344
 345                 /* Do not interfere with other mappings of this page */
 346                 if (page_mapcount(page) != 1)
 347                         goto huge_unlock;
 348
 349                 if (next - addr != HPAGE_PMD_SIZE) {
 350                         int err;
 351
 352                         get_page(page);
 353                         spin_unlock(ptl);
 354                         lock_page(page);
 355                         err = split_huge_page(page);
 356                         unlock_page(page);
 357                         put_page(page);
 358                         if (!err)
 359                                 goto regular_page;
 360                         return 0;
 361                 }
 362
 363                 if (pmd_young(orig_pmd)) {
 364                         pmdp_invalidate(vma, addr, pmd);
 365                         orig_pmd = pmd_mkold(orig_pmd);
 366
 367                         set_pmd_at(mm, addr, pmd, orig_pmd);
 368                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 369                 }
 370
 371                 ClearPageReferenced(page);
 372                 test_and_clear_page_young(page);
 373                 if (pageout) {
 374                         if (!isolate_lru_page(page)) {
 375                                 if (PageUnevictable(page))
 376                                         putback_lru_page(page);
 377                                 else
 378                                         list_add(&page->lru, &page_list);
 379                         }
 380                 } else
 381                         deactivate_page(page);
 382 huge_unlock:
 383                 spin_unlock(ptl);
 384                 if (pageout)
 385                         reclaim_pages(&page_list);
 386                 return 0;
 387         }
 388
 389 regular_page:
 390         if (pmd_trans_unstable(pmd))
 391                 return 0;
 392 #endif
 393         tlb_change_page_size(tlb, PAGE_SIZE);
 394         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 395         flush_tlb_batched_pending(mm);
 396         arch_enter_lazy_mmu_mode();
 397         for (; addr < end; pte++, addr += PAGE_SIZE) {
 398                 ptent = *pte;
 399
 400                 if (pte_none(ptent))
 401                         continue;
 402
 403                 if (!pte_present(ptent))
 404                         continue;
 405
 406                 page = vm_normal_page(vma, addr, ptent);
 407                 if (!page)
 408                         continue;
 409
 410 #ifdef CONFIG_FINEGRAINED_THP
 411                 if (pte_cont(ptent))
 412                         split_huge_pte_address(vma, addr, false, NULL);
 413 #endif
 414
 415                 /*
 416                  * Creating a THP page is expensive so split it only if we
 417                  * are sure it's worth. Split it if we are only owner.
 418                  */
 419                 if (PageTransCompound(page)) {
 420                         if (page_mapcount(page) != 1)
 421                                 break;
 422                         get_page(page);
 423                         if (!trylock_page(page)) {
 424                                 put_page(page);
 425                                 break;
 426                         }
 427                         pte_unmap_unlock(orig_pte, ptl);
 428                         if (split_huge_page(page)) {
 429                                 unlock_page(page);
 430                                 put_page(page);
 431                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 432                                 break;
 433                         }
 434                         unlock_page(page);
 435                         put_page(page);
 436                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 437                         pte--;
 438                         addr -= PAGE_SIZE;
 439                         continue;
 440                 }
 441
 442                 /* Do not interfere with other mappings of this page */
 443                 if (page_mapcount(page) != 1)
 444                         continue;
 445
 446                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 447
 448                 if (pte_young(ptent)) {
 449                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 450                                                         tlb->fullmm);
 451                         ptent = pte_mkold(ptent);
 452                         set_pte_at(mm, addr, pte, ptent);
 453                         tlb_remove_tlb_entry(tlb, pte, addr);
 454                 }
 455
 456                 /*
 457                  * We are deactivating a page for accelerating reclaiming.
 458                  * VM couldn't reclaim the page unless we clear PG_young.
 459                  * As a side effect, it makes confuse idle-page tracking
 460                  * because they will miss recent referenced history.
 461                  */
 462                 ClearPageReferenced(page);
 463                 test_and_clear_page_young(page);
 464                 if (pageout) {
 465                         if (!isolate_lru_page(page)) {
 466                                 if (PageUnevictable(page))
 467                                         putback_lru_page(page);
 468                                 else
 469                                         list_add(&page->lru, &page_list);
 470                         }
 471                 } else
 472                         deactivate_page(page);
 473         }
 474
 475         arch_leave_lazy_mmu_mode();
 476         pte_unmap_unlock(orig_pte, ptl);
 477         if (pageout)
 478                 reclaim_pages(&page_list);
 479         cond_resched();
 480
 481         return 0;
 482 }
 483
 484 static const struct mm_walk_ops cold_walk_ops = {
 485         .pmd_entry = madvise_cold_or_pageout_pte_range,
 486 };
 487
 488 static void madvise_cold_page_range(struct mmu_gather *tlb,
 489                              struct vm_area_struct *vma,
 490                              unsigned long addr, unsigned long end)
 491 {
 492         struct madvise_walk_private walk_private = {
 493                 .pageout = false,
 494                 .tlb = tlb,
 495         };
 496
 497         tlb_start_vma(tlb, vma);
 498         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 499         tlb_end_vma(tlb, vma);
 500 }
 501
 502 static long madvise_cold(struct vm_area_struct *vma,
 503                         struct vm_area_struct **prev,
 504                         unsigned long start_addr, unsigned long end_addr)
 505 {
 506         struct mm_struct *mm = vma->vm_mm;
 507         struct mmu_gather tlb;
 508
 509         *prev = vma;
 510         if (!can_madv_lru_vma(vma))
 511                 return -EINVAL;
 512
 513         lru_add_drain();
 514         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 515         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 516         tlb_finish_mmu(&tlb, start_addr, end_addr);
 517
 518         return 0;
 519 }
 520
 521 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 522                              struct vm_area_struct *vma,
 523                              unsigned long addr, unsigned long end)
 524 {
 525         struct madvise_walk_private walk_private = {
 526                 .pageout = true,
 527                 .tlb = tlb,
 528         };
 529
 530         tlb_start_vma(tlb, vma);
 531         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 532         tlb_end_vma(tlb, vma);
 533 }
 534
 535 static inline bool can_do_pageout(struct vm_area_struct *vma)
 536 {
 537         if (vma_is_anonymous(vma))
 538                 return true;
 539         if (!vma->vm_file)
 540                 return false;
 541         /*
 542          * paging out pagecache only for non-anonymous mappings that correspond
 543          * to the files the calling process could (if tried) open for writing;
 544          * otherwise we'd be including shared non-exclusive mappings, which
 545          * opens a side channel.
 546          */
 547         return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 548                 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 549 }
 550
 551 static long madvise_pageout(struct vm_area_struct *vma,
 552                         struct vm_area_struct **prev,
 553                         unsigned long start_addr, unsigned long end_addr)
 554 {
 555         struct mm_struct *mm = vma->vm_mm;
 556         struct mmu_gather tlb;
 557
 558         *prev = vma;
 559         if (!can_madv_lru_vma(vma))
 560                 return -EINVAL;
 561
 562         if (!can_do_pageout(vma))
 563                 return 0;
 564
 565         lru_add_drain();
 566         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 567         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 568         tlb_finish_mmu(&tlb, start_addr, end_addr);
 569
 570         return 0;
 571 }
 572
 573 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 574                                 unsigned long end, struct mm_walk *walk)
 575
 576 {
 577         struct mmu_gather *tlb = walk->private;
 578         struct mm_struct *mm = tlb->mm;
 579         struct vm_area_struct *vma = walk->vma;
 580         spinlock_t *ptl;
 581         pte_t *orig_pte, *pte, ptent;
 582         struct page *page;
 583         int nr_swap = 0;
 584         unsigned long next;
 585
 586         next = pmd_addr_end(addr, end);
 587         if (pmd_trans_huge(*pmd))
 588                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 589                         goto next;
 590
 591         if (pmd_trans_unstable(pmd))
 592                 return 0;
 593
 594         tlb_change_page_size(tlb, PAGE_SIZE);
 595         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 596         flush_tlb_batched_pending(mm);
 597         arch_enter_lazy_mmu_mode();
 598         for (; addr != end; pte++, addr += PAGE_SIZE) {
 599                 ptent = *pte;
 600
 601                 if (pte_none(ptent))
 602                         continue;
 603                 /*
 604                  * If the pte has swp_entry, just clear page table to
 605                  * prevent swap-in which is more expensive rather than
 606                  * (page allocation + zeroing).
 607                  */
 608                 if (!pte_present(ptent)) {
 609                         swp_entry_t entry;
 610
 611                         entry = pte_to_swp_entry(ptent);
 612                         if (non_swap_entry(entry))
 613                                 continue;
 614                         nr_swap--;
 615                         free_swap_and_cache(entry);
 616                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 617                         continue;
 618                 }
 619
 620                 page = vm_normal_page(vma, addr, ptent);
 621                 if (!page)
 622                         continue;
 623
 624 #ifdef CONFIG_FINEGRAINED_THP
 625                 if (pte_cont(ptent))
 626                         split_huge_pte_address(vma, addr, false, NULL);
 627 #endif /* CONFIG_FINEGRAINED_THP */
 628
 629                 /*
 630                  * If pmd isn't transhuge but the page is THP and
 631                  * is owned by only this process, split it and
 632                  * deactivate all pages.
 633                  */
 634                 if (PageTransCompound(page)) {
 635                         if (page_mapcount(page) != 1)
 636                                 goto out;
 637                         get_page(page);
 638                         if (!trylock_page(page)) {
 639                                 put_page(page);
 640                                 goto out;
 641                         }
 642                         pte_unmap_unlock(orig_pte, ptl);
 643                         if (split_huge_page(page)) {
 644                                 unlock_page(page);
 645                                 put_page(page);
 646                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 647                                 goto out;
 648                         }
 649                         unlock_page(page);
 650                         put_page(page);
 651                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 652                         pte--;
 653                         addr -= PAGE_SIZE;
 654                         continue;
 655                 }
 656
 657                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 658
 659                 if (PageSwapCache(page) || PageDirty(page)) {
 660                         if (!trylock_page(page))
 661                                 continue;
 662                         /*
 663                          * If page is shared with others, we couldn't clear
 664                          * PG_dirty of the page.
 665                          */
 666                         if (page_mapcount(page) != 1) {
 667                                 unlock_page(page);
 668                                 continue;
 669                         }
 670
 671                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 672                                 unlock_page(page);
 673                                 continue;
 674                         }
 675
 676                         ClearPageDirty(page);
 677                         unlock_page(page);
 678                 }
 679
 680                 if (pte_young(ptent) || pte_dirty(ptent)) {
 681                         /*
 682                          * Some of architecture(ex, PPC) don't update TLB
 683                          * with set_pte_at and tlb_remove_tlb_entry so for
 684                          * the portability, remap the pte with old|clean
 685                          * after pte clearing.
 686                          */
 687                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 688                                                         tlb->fullmm);
 689
 690                         ptent = pte_mkold(ptent);
 691                         ptent = pte_mkclean(ptent);
 692                         set_pte_at(mm, addr, pte, ptent);
 693                         tlb_remove_tlb_entry(tlb, pte, addr);
 694                 }
 695                 mark_page_lazyfree(page);
 696         }
 697 out:
 698         if (nr_swap) {
 699                 if (current->mm == mm)
 700                         sync_mm_rss(mm);
 701
 702                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 703         }
 704         arch_leave_lazy_mmu_mode();
 705         pte_unmap_unlock(orig_pte, ptl);
 706         cond_resched();
 707 next:
 708         return 0;
 709 }
 710
 711 static const struct mm_walk_ops madvise_free_walk_ops = {
 712         .pmd_entry              = madvise_free_pte_range,
 713 };
 714
 715 static int madvise_free_single_vma(struct vm_area_struct *vma,
 716                         unsigned long start_addr, unsigned long end_addr)
 717 {
 718         struct mm_struct *mm = vma->vm_mm;
 719         struct mmu_notifier_range range;
 720         struct mmu_gather tlb;
 721
 722         /* MADV_FREE works for only anon vma at the moment */
 723         if (!vma_is_anonymous(vma))
 724                 return -EINVAL;
 725
 726         range.start = max(vma->vm_start, start_addr);
 727         if (range.start >= vma->vm_end)
 728                 return -EINVAL;
 729         range.end = min(vma->vm_end, end_addr);
 730         if (range.end <= vma->vm_start)
 731                 return -EINVAL;
 732         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 733                                 range.start, range.end);
 734
 735         lru_add_drain();
 736         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 737         update_hiwater_rss(mm);
 738
 739         mmu_notifier_invalidate_range_start(&range);
 740         tlb_start_vma(&tlb, vma);
 741         walk_page_range(vma->vm_mm, range.start, range.end,
 742                         &madvise_free_walk_ops, &tlb);
 743         tlb_end_vma(&tlb, vma);
 744         mmu_notifier_invalidate_range_end(&range);
 745         tlb_finish_mmu(&tlb, range.start, range.end);
 746
 747         return 0;
 748 }
 749
 750 /*
 751  * Application no longer needs these pages.  If the pages are dirty,
 752  * it's OK to just throw them away.  The app will be more careful about
 753  * data it wants to keep.  Be sure to free swap resources too.  The
 754  * zap_page_range call sets things up for shrink_active_list to actually free
 755  * these pages later if no one else has touched them in the meantime,
 756  * although we could add these pages to a global reuse list for
 757  * shrink_active_list to pick up before reclaiming other pages.
 758  *
 759  * NB: This interface discards data rather than pushes it out to swap,
 760  * as some implementations do.  This has performance implications for
 761  * applications like large transactional databases which want to discard
 762  * pages in anonymous maps after committing to backing store the data
 763  * that was kept in them.  There is no reason to write this data out to
 764  * the swap area if the application is discarding it.
 765  *
 766  * An interface that causes the system to free clean pages and flush
 767  * dirty pages is already available as msync(MS_INVALIDATE).
 768  */
 769 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 770                                         unsigned long start, unsigned long end)
 771 {
 772         zap_page_range(vma, start, end - start);
 773         return 0;
 774 }
 775
 776 static long madvise_dontneed_free(struct vm_area_struct *vma,
 777                                   struct vm_area_struct **prev,
 778                                   unsigned long start, unsigned long end,
 779                                   int behavior)
 780 {
 781         struct mm_struct *mm = vma->vm_mm;
 782
 783         *prev = vma;
 784         if (!can_madv_lru_vma(vma))
 785                 return -EINVAL;
 786
 787         if (!userfaultfd_remove(vma, start, end)) {
 788                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 789
 790                 mmap_read_lock(mm);
 791                 vma = find_vma(mm, start);
 792                 if (!vma)
 793                         return -ENOMEM;
 794                 if (start < vma->vm_start) {
 795                         /*
 796                          * This "vma" under revalidation is the one
 797                          * with the lowest vma->vm_start where start
 798                          * is also < vma->vm_end. If start <
 799                          * vma->vm_start it means an hole materialized
 800                          * in the user address space within the
 801                          * virtual range passed to MADV_DONTNEED
 802                          * or MADV_FREE.
 803                          */
 804                         return -ENOMEM;
 805                 }
 806                 if (!can_madv_lru_vma(vma))
 807                         return -EINVAL;
 808                 if (end > vma->vm_end) {
 809                         /*
 810                          * Don't fail if end > vma->vm_end. If the old
 811                          * vma was splitted while the mmap_lock was
 812                          * released the effect of the concurrent
 813                          * operation may not cause madvise() to
 814                          * have an undefined result. There may be an
 815                          * adjacent next vma that we'll walk
 816                          * next. userfaultfd_remove() will generate an
 817                          * UFFD_EVENT_REMOVE repetition on the
 818                          * end-vma->vm_end range, but the manager can
 819                          * handle a repetition fine.
 820                          */
 821                         end = vma->vm_end;
 822                 }
 823                 VM_WARN_ON(start >= end);
 824         }
 825
 826         if (behavior == MADV_DONTNEED)
 827                 return madvise_dontneed_single_vma(vma, start, end);
 828         else if (behavior == MADV_FREE)
 829                 return madvise_free_single_vma(vma, start, end);
 830         else
 831                 return -EINVAL;
 832 }
 833
 834 /*
 835  * Application wants to free up the pages and associated backing store.
 836  * This is effectively punching a hole into the middle of a file.
 837  */
 838 static long madvise_remove(struct vm_area_struct *vma,
 839                                 struct vm_area_struct **prev,
 840                                 unsigned long start, unsigned long end)
 841 {
 842         loff_t offset;
 843         int error;
 844         struct file *f;
 845         struct mm_struct *mm = vma->vm_mm;
 846
 847         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 848
 849         if (vma->vm_flags & VM_LOCKED)
 850                 return -EINVAL;
 851
 852         f = vma->vm_file;
 853
 854         if (!f || !f->f_mapping || !f->f_mapping->host) {
 855                         return -EINVAL;
 856         }
 857
 858         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 859                 return -EACCES;
 860
 861         offset = (loff_t)(start - vma->vm_start)
 862                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 863
 864         /*
 865          * Filesystem's fallocate may need to take i_mutex.  We need to
 866          * explicitly grab a reference because the vma (and hence the
 867          * vma's reference to the file) can go away as soon as we drop
 868          * mmap_lock.
 869          */
 870         get_file(f);
 871         if (userfaultfd_remove(vma, start, end)) {
 872                 /* mmap_lock was not released by userfaultfd_remove() */
 873                 mmap_read_unlock(mm);
 874         }
 875         error = vfs_fallocate(f,
 876                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 877                                 offset, end - start);
 878         fput(f);
 879         mmap_read_lock(mm);
 880         return error;
 881 }
 882
 883 #ifdef CONFIG_MEMORY_FAILURE
 884 /*
 885  * Error injection support for memory error handling.
 886  */
 887 static int madvise_inject_error(int behavior,
 888                 unsigned long start, unsigned long end)
 889 {
 890         struct zone *zone;
 891         unsigned long size;
 892
 893         if (!capable(CAP_SYS_ADMIN))
 894                 return -EPERM;
 895
 896
 897         for (; start < end; start += size) {
 898                 unsigned long pfn;
 899                 struct page *page;
 900                 int ret;
 901
 902                 ret = get_user_pages_fast(start, 1, 0, &page);
 903                 if (ret != 1)
 904                         return ret;
 905                 pfn = page_to_pfn(page);
 906
 907                 /*
 908                  * When soft offlining hugepages, after migrating the page
 909                  * we dissolve it, therefore in the second loop "page" will
 910                  * no longer be a compound page.
 911                  */
 912                 size = page_size(compound_head(page));
 913
 914                 if (behavior == MADV_SOFT_OFFLINE) {
 915                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 916                                  pfn, start);
 917                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
 918                 } else {
 919                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 920                                  pfn, start);
 921                         ret = memory_failure(pfn, MF_COUNT_INCREASED);
 922                 }
 923
 924                 if (ret)
 925                         return ret;
 926         }
 927
 928         /* Ensure that all poisoned pages are removed from per-cpu lists */
 929         for_each_populated_zone(zone)
 930                 drain_all_pages(zone);
 931
 932         return 0;
 933 }
 934 #endif
 935
 936 static long
 937 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 938                 unsigned long start, unsigned long end, int behavior)
 939 {
 940         switch (behavior) {
 941         case MADV_REMOVE:
 942                 return madvise_remove(vma, prev, start, end);
 943         case MADV_WILLNEED:
 944                 return madvise_willneed(vma, prev, start, end);
 945         case MADV_COLD:
 946                 return madvise_cold(vma, prev, start, end);
 947         case MADV_PAGEOUT:
 948                 return madvise_pageout(vma, prev, start, end);
 949         case MADV_FREE:
 950         case MADV_DONTNEED:
 951                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 952         default:
 953                 return madvise_behavior(vma, prev, start, end, behavior);
 954         }
 955 }
 956
 957 static bool
 958 madvise_behavior_valid(int behavior)
 959 {
 960         switch (behavior) {
 961         case MADV_DOFORK:
 962         case MADV_DONTFORK:
 963         case MADV_NORMAL:
 964         case MADV_SEQUENTIAL:
 965         case MADV_RANDOM:
 966         case MADV_REMOVE:
 967         case MADV_WILLNEED:
 968         case MADV_DONTNEED:
 969         case MADV_FREE:
 970         case MADV_COLD:
 971         case MADV_PAGEOUT:
 972 #ifdef CONFIG_KSM
 973         case MADV_MERGEABLE:
 974         case MADV_UNMERGEABLE:
 975 #endif
 976 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 977         case MADV_HUGEPAGE:
 978         case MADV_NOHUGEPAGE:
 979 #endif
 980         case MADV_DONTDUMP:
 981         case MADV_DODUMP:
 982         case MADV_WIPEONFORK:
 983         case MADV_KEEPONFORK:
 984 #ifdef CONFIG_MEMORY_FAILURE
 985         case MADV_SOFT_OFFLINE:
 986         case MADV_HWPOISON:
 987 #endif
 988                 return true;
 989
 990         default:
 991                 return false;
 992         }
 993 }
 994
 995 static bool
 996 process_madvise_behavior_valid(int behavior)
 997 {
 998         switch (behavior) {
 999         case MADV_COLD:
1000         case MADV_PAGEOUT:
1001                 return true;
1002         default:
1003                 return false;
1004         }
1005 }
1006
1007 /*
1008  * The madvise(2) system call.
1009  *
1010  * Applications can use madvise() to advise the kernel how it should
1011  * handle paging I/O in this VM area.  The idea is to help the kernel
1012  * use appropriate read-ahead and caching techniques.  The information
1013  * provided is advisory only, and can be safely disregarded by the
1014  * kernel without affecting the correct operation of the application.
1015  *
1016  * behavior values:
1017  *  MADV_NORMAL - the default behavior is to read clusters.  This
1018  *              results in some read-ahead and read-behind.
1019  *  MADV_RANDOM - the system should read the minimum amount of data
1020  *              on any access, since it is unlikely that the appli-
1021  *              cation will need more than what it asks for.
1022  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1023  *              once, so they can be aggressively read ahead, and
1024  *              can be freed soon after they are accessed.
1025  *  MADV_WILLNEED - the application is notifying the system to read
1026  *              some pages ahead.
1027  *  MADV_DONTNEED - the application is finished with the given range,
1028  *              so the kernel can free resources associated with it.
1029  *  MADV_FREE - the application marks pages in the given range as lazy free,
1030  *              where actual purges are postponed until memory pressure happens.
1031  *  MADV_REMOVE - the application wants to free up the given range of
1032  *              pages and associated backing store.
1033  *  MADV_DONTFORK - omit this area from child's address space when forking:
1034  *              typically, to avoid COWing pages pinned by get_user_pages().
1035  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1036  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1037  *              range after a fork.
1038  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1039  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1040  *              were corrupted by unrecoverable hardware memory failure.
1041  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1042  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1043  *              this area with pages of identical content from other such areas.
1044  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1045  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1046  *              huge pages in the future. Existing pages might be coalesced and
1047  *              new pages might be allocated as THP.
1048  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1049  *              transparent huge pages so the existing pages will not be
1050  *              coalesced into THP and new pages will not be allocated as THP.
1051  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1052  *              from being included in its core dump.
1053  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1054  *  MADV_COLD - the application is not expected to use this memory soon,
1055  *              deactivate pages in this range so that they can be reclaimed
1056  *              easily if memory pressure hanppens.
1057  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1058  *              page out the pages in this range immediately.
1059  *
1060  * return values:
1061  *  zero    - success
1062  *  -EINVAL - start + len < 0, start is not page-aligned,
1063  *              "behavior" is not a valid value, or application
1064  *              is attempting to release locked or shared pages,
1065  *              or the specified address range includes file, Huge TLB,
1066  *              MAP_SHARED or VMPFNMAP range.
1067  *  -ENOMEM - addresses in the specified range are not currently
1068  *              mapped, or are outside the AS of the process.
1069  *  -EIO    - an I/O error occurred while paging in data.
1070  *  -EBADF  - map exists, but area maps something that isn't a file.
1071  *  -EAGAIN - a kernel resource was temporarily unavailable.
1072  */
1073 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1074 {
1075         unsigned long end, tmp;
1076         struct vm_area_struct *vma, *prev;
1077         int unmapped_error = 0;
1078         int error = -EINVAL;
1079         int write;
1080         size_t len;
1081         struct blk_plug plug;
1082
1083         start = untagged_addr(start);
1084
1085         if (!madvise_behavior_valid(behavior))
1086                 return error;
1087
1088         if (!PAGE_ALIGNED(start))
1089                 return error;
1090         len = PAGE_ALIGN(len_in);
1091
1092         /* Check to see whether len was rounded up from small -ve to zero */
1093         if (len_in && !len)
1094                 return error;
1095
1096         end = start + len;
1097         if (end < start)
1098                 return error;
1099
1100         error = 0;
1101         if (end == start)
1102                 return error;
1103
1104 #ifdef CONFIG_MEMORY_FAILURE
1105         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1106                 return madvise_inject_error(behavior, start, start + len_in);
1107 #endif
1108
1109         write = madvise_need_mmap_write(behavior);
1110         if (write) {
1111                 if (mmap_write_lock_killable(mm))
1112                         return -EINTR;
1113         } else {
1114                 mmap_read_lock(mm);
1115         }
1116
1117         /*
1118          * If the interval [start,end) covers some unmapped address
1119          * ranges, just ignore them, but return -ENOMEM at the end.
1120          * - different from the way of handling in mlock etc.
1121          */
1122         vma = find_vma_prev(mm, start, &prev);
1123         if (vma && start > vma->vm_start)
1124                 prev = vma;
1125
1126         blk_start_plug(&plug);
1127         for (;;) {
1128                 /* Still start < end. */
1129                 error = -ENOMEM;
1130                 if (!vma)
1131                         goto out;
1132
1133                 /* Here start < (end|vma->vm_end). */
1134                 if (start < vma->vm_start) {
1135                         unmapped_error = -ENOMEM;
1136                         start = vma->vm_start;
1137                         if (start >= end)
1138                                 goto out;
1139                 }
1140
1141                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1142                 tmp = vma->vm_end;
1143                 if (end < tmp)
1144                         tmp = end;
1145
1146                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1147                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1148                 if (error)
1149                         goto out;
1150                 start = tmp;
1151                 if (prev && start < prev->vm_end)
1152                         start = prev->vm_end;
1153                 error = unmapped_error;
1154                 if (start >= end)
1155                         goto out;
1156                 if (prev)
1157                         vma = prev->vm_next;
1158                 else    /* madvise_remove dropped mmap_lock */
1159                         vma = find_vma(mm, start);
1160         }
1161 out:
1162         blk_finish_plug(&plug);
1163         if (write)
1164                 mmap_write_unlock(mm);
1165         else
1166                 mmap_read_unlock(mm);
1167
1168         return error;
1169 }
1170
1171 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1172 {
1173         return do_madvise(current->mm, start, len_in, behavior);
1174 }
1175
1176 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1177                 size_t, vlen, int, behavior, unsigned int, flags)
1178 {
1179         ssize_t ret;
1180         struct iovec iovstack[UIO_FASTIOV], iovec;
1181         struct iovec *iov = iovstack;
1182         struct iov_iter iter;
1183         struct pid *pid;
1184         struct task_struct *task;
1185         struct mm_struct *mm;
1186         size_t total_len;
1187         unsigned int f_flags;
1188
1189         if (flags != 0) {
1190                 ret = -EINVAL;
1191                 goto out;
1192         }
1193
1194         ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1195         if (ret < 0)
1196                 goto out;
1197
1198         pid = pidfd_get_pid(pidfd, &f_flags);
1199         if (IS_ERR(pid)) {
1200                 ret = PTR_ERR(pid);
1201                 goto free_iov;
1202         }
1203
1204         task = get_pid_task(pid, PIDTYPE_PID);
1205         if (!task) {
1206                 ret = -ESRCH;
1207                 goto put_pid;
1208         }
1209
1210         if (!process_madvise_behavior_valid(behavior)) {
1211                 ret = -EINVAL;
1212                 goto release_task;
1213         }
1214
1215         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1216         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1217         if (IS_ERR_OR_NULL(mm)) {
1218                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1219                 goto release_task;
1220         }
1221
1222         /*
1223          * Require CAP_SYS_NICE for influencing process performance. Note that
1224          * only non-destructive hints are currently supported.
1225          */
1226         if (!capable(CAP_SYS_NICE)) {
1227                 ret = -EPERM;
1228                 goto release_mm;
1229         }
1230
1231         total_len = iov_iter_count(&iter);
1232
1233         while (iov_iter_count(&iter)) {
1234                 iovec = iov_iter_iovec(&iter);
1235                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1236                                         iovec.iov_len, behavior);
1237                 if (ret < 0)
1238                         break;
1239                 iov_iter_advance(&iter, iovec.iov_len);
1240         }
1241
1242         if (ret == 0)
1243                 ret = total_len - iov_iter_count(&iter);
1244
1245 release_mm:
1246         mmput(mm);
1247 release_task:
1248         put_task_struct(task);
1249 put_pid:
1250         put_pid(pid);
1251 free_iov:
1252         kfree(iov);
1253 out:
1254         return ret;
1255 }