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                 /*
 411                  * Creating a THP page is expensive so split it only if we
 412                  * are sure it's worth. Split it if we are only owner.
 413                  */
 414                 if (PageTransCompound(page)) {
 415                         if (page_mapcount(page) != 1)
 416                                 break;
 417                         get_page(page);
 418                         if (!trylock_page(page)) {
 419                                 put_page(page);
 420                                 break;
 421                         }
 422                         pte_unmap_unlock(orig_pte, ptl);
 423                         if (split_huge_page(page)) {
 424                                 unlock_page(page);
 425                                 put_page(page);
 426                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 427                                 break;
 428                         }
 429                         unlock_page(page);
 430                         put_page(page);
 431                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 432                         pte--;
 433                         addr -= PAGE_SIZE;
 434                         continue;
 435                 }
 436
 437                 /* Do not interfere with other mappings of this page */
 438                 if (page_mapcount(page) != 1)
 439                         continue;
 440
 441                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 442
 443                 if (pte_young(ptent)) {
 444                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 445                                                         tlb->fullmm);
 446                         ptent = pte_mkold(ptent);
 447                         set_pte_at(mm, addr, pte, ptent);
 448                         tlb_remove_tlb_entry(tlb, pte, addr);
 449                 }
 450
 451                 /*
 452                  * We are deactivating a page for accelerating reclaiming.
 453                  * VM couldn't reclaim the page unless we clear PG_young.
 454                  * As a side effect, it makes confuse idle-page tracking
 455                  * because they will miss recent referenced history.
 456                  */
 457                 ClearPageReferenced(page);
 458                 test_and_clear_page_young(page);
 459                 if (pageout) {
 460                         if (!isolate_lru_page(page)) {
 461                                 if (PageUnevictable(page))
 462                                         putback_lru_page(page);
 463                                 else
 464                                         list_add(&page->lru, &page_list);
 465                         }
 466                 } else
 467                         deactivate_page(page);
 468         }
 469
 470         arch_leave_lazy_mmu_mode();
 471         pte_unmap_unlock(orig_pte, ptl);
 472         if (pageout)
 473                 reclaim_pages(&page_list);
 474         cond_resched();
 475
 476         return 0;
 477 }
 478
 479 static const struct mm_walk_ops cold_walk_ops = {
 480         .pmd_entry = madvise_cold_or_pageout_pte_range,
 481 };
 482
 483 static void madvise_cold_page_range(struct mmu_gather *tlb,
 484                              struct vm_area_struct *vma,
 485                              unsigned long addr, unsigned long end)
 486 {
 487         struct madvise_walk_private walk_private = {
 488                 .pageout = false,
 489                 .tlb = tlb,
 490         };
 491
 492         tlb_start_vma(tlb, vma);
 493         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 494         tlb_end_vma(tlb, vma);
 495 }
 496
 497 static long madvise_cold(struct vm_area_struct *vma,
 498                         struct vm_area_struct **prev,
 499                         unsigned long start_addr, unsigned long end_addr)
 500 {
 501         struct mm_struct *mm = vma->vm_mm;
 502         struct mmu_gather tlb;
 503
 504         *prev = vma;
 505         if (!can_madv_lru_vma(vma))
 506                 return -EINVAL;
 507
 508         lru_add_drain();
 509         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 510         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 511         tlb_finish_mmu(&tlb, start_addr, end_addr);
 512
 513         return 0;
 514 }
 515
 516 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 517                              struct vm_area_struct *vma,
 518                              unsigned long addr, unsigned long end)
 519 {
 520         struct madvise_walk_private walk_private = {
 521                 .pageout = true,
 522                 .tlb = tlb,
 523         };
 524
 525         tlb_start_vma(tlb, vma);
 526         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 527         tlb_end_vma(tlb, vma);
 528 }
 529
 530 static inline bool can_do_pageout(struct vm_area_struct *vma)
 531 {
 532         if (vma_is_anonymous(vma))
 533                 return true;
 534         if (!vma->vm_file)
 535                 return false;
 536         /*
 537          * paging out pagecache only for non-anonymous mappings that correspond
 538          * to the files the calling process could (if tried) open for writing;
 539          * otherwise we'd be including shared non-exclusive mappings, which
 540          * opens a side channel.
 541          */
 542         return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 543                 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 544 }
 545
 546 static long madvise_pageout(struct vm_area_struct *vma,
 547                         struct vm_area_struct **prev,
 548                         unsigned long start_addr, unsigned long end_addr)
 549 {
 550         struct mm_struct *mm = vma->vm_mm;
 551         struct mmu_gather tlb;
 552
 553         *prev = vma;
 554         if (!can_madv_lru_vma(vma))
 555                 return -EINVAL;
 556
 557         if (!can_do_pageout(vma))
 558                 return 0;
 559
 560         lru_add_drain();
 561         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 562         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 563         tlb_finish_mmu(&tlb, start_addr, end_addr);
 564
 565         return 0;
 566 }
 567
 568 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 569                                 unsigned long end, struct mm_walk *walk)
 570
 571 {
 572         struct mmu_gather *tlb = walk->private;
 573         struct mm_struct *mm = tlb->mm;
 574         struct vm_area_struct *vma = walk->vma;
 575         spinlock_t *ptl;
 576         pte_t *orig_pte, *pte, ptent;
 577         struct page *page;
 578         int nr_swap = 0;
 579         unsigned long next;
 580
 581         next = pmd_addr_end(addr, end);
 582         if (pmd_trans_huge(*pmd))
 583                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 584                         goto next;
 585
 586         if (pmd_trans_unstable(pmd))
 587                 return 0;
 588
 589         tlb_change_page_size(tlb, PAGE_SIZE);
 590         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 591         flush_tlb_batched_pending(mm);
 592         arch_enter_lazy_mmu_mode();
 593         for (; addr != end; pte++, addr += PAGE_SIZE) {
 594                 ptent = *pte;
 595
 596                 if (pte_none(ptent))
 597                         continue;
 598                 /*
 599                  * If the pte has swp_entry, just clear page table to
 600                  * prevent swap-in which is more expensive rather than
 601                  * (page allocation + zeroing).
 602                  */
 603                 if (!pte_present(ptent)) {
 604                         swp_entry_t entry;
 605
 606                         entry = pte_to_swp_entry(ptent);
 607                         if (non_swap_entry(entry))
 608                                 continue;
 609                         nr_swap--;
 610                         free_swap_and_cache(entry);
 611                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 612                         continue;
 613                 }
 614
 615                 page = vm_normal_page(vma, addr, ptent);
 616                 if (!page)
 617                         continue;
 618
 619                 /*
 620                  * If pmd isn't transhuge but the page is THP and
 621                  * is owned by only this process, split it and
 622                  * deactivate all pages.
 623                  */
 624                 if (PageTransCompound(page)) {
 625                         if (page_mapcount(page) != 1)
 626                                 goto out;
 627                         get_page(page);
 628                         if (!trylock_page(page)) {
 629                                 put_page(page);
 630                                 goto out;
 631                         }
 632                         pte_unmap_unlock(orig_pte, ptl);
 633                         if (split_huge_page(page)) {
 634                                 unlock_page(page);
 635                                 put_page(page);
 636                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 637                                 goto out;
 638                         }
 639                         unlock_page(page);
 640                         put_page(page);
 641                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 642                         pte--;
 643                         addr -= PAGE_SIZE;
 644                         continue;
 645                 }
 646
 647                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 648
 649                 if (PageSwapCache(page) || PageDirty(page)) {
 650                         if (!trylock_page(page))
 651                                 continue;
 652                         /*
 653                          * If page is shared with others, we couldn't clear
 654                          * PG_dirty of the page.
 655                          */
 656                         if (page_mapcount(page) != 1) {
 657                                 unlock_page(page);
 658                                 continue;
 659                         }
 660
 661                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 662                                 unlock_page(page);
 663                                 continue;
 664                         }
 665
 666                         ClearPageDirty(page);
 667                         unlock_page(page);
 668                 }
 669
 670                 if (pte_young(ptent) || pte_dirty(ptent)) {
 671                         /*
 672                          * Some of architecture(ex, PPC) don't update TLB
 673                          * with set_pte_at and tlb_remove_tlb_entry so for
 674                          * the portability, remap the pte with old|clean
 675                          * after pte clearing.
 676                          */
 677                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 678                                                         tlb->fullmm);
 679
 680                         ptent = pte_mkold(ptent);
 681                         ptent = pte_mkclean(ptent);
 682                         set_pte_at(mm, addr, pte, ptent);
 683                         tlb_remove_tlb_entry(tlb, pte, addr);
 684                 }
 685                 mark_page_lazyfree(page);
 686         }
 687 out:
 688         if (nr_swap) {
 689                 if (current->mm == mm)
 690                         sync_mm_rss(mm);
 691
 692                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 693         }
 694         arch_leave_lazy_mmu_mode();
 695         pte_unmap_unlock(orig_pte, ptl);
 696         cond_resched();
 697 next:
 698         return 0;
 699 }
 700
 701 static const struct mm_walk_ops madvise_free_walk_ops = {
 702         .pmd_entry              = madvise_free_pte_range,
 703 };
 704
 705 static int madvise_free_single_vma(struct vm_area_struct *vma,
 706                         unsigned long start_addr, unsigned long end_addr)
 707 {
 708         struct mm_struct *mm = vma->vm_mm;
 709         struct mmu_notifier_range range;
 710         struct mmu_gather tlb;
 711
 712         /* MADV_FREE works for only anon vma at the moment */
 713         if (!vma_is_anonymous(vma))
 714                 return -EINVAL;
 715
 716         range.start = max(vma->vm_start, start_addr);
 717         if (range.start >= vma->vm_end)
 718                 return -EINVAL;
 719         range.end = min(vma->vm_end, end_addr);
 720         if (range.end <= vma->vm_start)
 721                 return -EINVAL;
 722         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 723                                 range.start, range.end);
 724
 725         lru_add_drain();
 726         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 727         update_hiwater_rss(mm);
 728
 729         mmu_notifier_invalidate_range_start(&range);
 730         tlb_start_vma(&tlb, vma);
 731         walk_page_range(vma->vm_mm, range.start, range.end,
 732                         &madvise_free_walk_ops, &tlb);
 733         tlb_end_vma(&tlb, vma);
 734         mmu_notifier_invalidate_range_end(&range);
 735         tlb_finish_mmu(&tlb, range.start, range.end);
 736
 737         return 0;
 738 }
 739
 740 /*
 741  * Application no longer needs these pages.  If the pages are dirty,
 742  * it's OK to just throw them away.  The app will be more careful about
 743  * data it wants to keep.  Be sure to free swap resources too.  The
 744  * zap_page_range call sets things up for shrink_active_list to actually free
 745  * these pages later if no one else has touched them in the meantime,
 746  * although we could add these pages to a global reuse list for
 747  * shrink_active_list to pick up before reclaiming other pages.
 748  *
 749  * NB: This interface discards data rather than pushes it out to swap,
 750  * as some implementations do.  This has performance implications for
 751  * applications like large transactional databases which want to discard
 752  * pages in anonymous maps after committing to backing store the data
 753  * that was kept in them.  There is no reason to write this data out to
 754  * the swap area if the application is discarding it.
 755  *
 756  * An interface that causes the system to free clean pages and flush
 757  * dirty pages is already available as msync(MS_INVALIDATE).
 758  */
 759 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 760                                         unsigned long start, unsigned long end)
 761 {
 762         zap_page_range(vma, start, end - start);
 763         return 0;
 764 }
 765
 766 static long madvise_dontneed_free(struct vm_area_struct *vma,
 767                                   struct vm_area_struct **prev,
 768                                   unsigned long start, unsigned long end,
 769                                   int behavior)
 770 {
 771         struct mm_struct *mm = vma->vm_mm;
 772
 773         *prev = vma;
 774         if (!can_madv_lru_vma(vma))
 775                 return -EINVAL;
 776
 777         if (!userfaultfd_remove(vma, start, end)) {
 778                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 779
 780                 mmap_read_lock(mm);
 781                 vma = find_vma(mm, start);
 782                 if (!vma)
 783                         return -ENOMEM;
 784                 if (start < vma->vm_start) {
 785                         /*
 786                          * This "vma" under revalidation is the one
 787                          * with the lowest vma->vm_start where start
 788                          * is also < vma->vm_end. If start <
 789                          * vma->vm_start it means an hole materialized
 790                          * in the user address space within the
 791                          * virtual range passed to MADV_DONTNEED
 792                          * or MADV_FREE.
 793                          */
 794                         return -ENOMEM;
 795                 }
 796                 if (!can_madv_lru_vma(vma))
 797                         return -EINVAL;
 798                 if (end > vma->vm_end) {
 799                         /*
 800                          * Don't fail if end > vma->vm_end. If the old
 801                          * vma was splitted while the mmap_lock was
 802                          * released the effect of the concurrent
 803                          * operation may not cause madvise() to
 804                          * have an undefined result. There may be an
 805                          * adjacent next vma that we'll walk
 806                          * next. userfaultfd_remove() will generate an
 807                          * UFFD_EVENT_REMOVE repetition on the
 808                          * end-vma->vm_end range, but the manager can
 809                          * handle a repetition fine.
 810                          */
 811                         end = vma->vm_end;
 812                 }
 813                 VM_WARN_ON(start >= end);
 814         }
 815
 816         if (behavior == MADV_DONTNEED)
 817                 return madvise_dontneed_single_vma(vma, start, end);
 818         else if (behavior == MADV_FREE)
 819                 return madvise_free_single_vma(vma, start, end);
 820         else
 821                 return -EINVAL;
 822 }
 823
 824 /*
 825  * Application wants to free up the pages and associated backing store.
 826  * This is effectively punching a hole into the middle of a file.
 827  */
 828 static long madvise_remove(struct vm_area_struct *vma,
 829                                 struct vm_area_struct **prev,
 830                                 unsigned long start, unsigned long end)
 831 {
 832         loff_t offset;
 833         int error;
 834         struct file *f;
 835         struct mm_struct *mm = vma->vm_mm;
 836
 837         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 838
 839         if (vma->vm_flags & VM_LOCKED)
 840                 return -EINVAL;
 841
 842         f = vma->vm_file;
 843
 844         if (!f || !f->f_mapping || !f->f_mapping->host) {
 845                         return -EINVAL;
 846         }
 847
 848         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 849                 return -EACCES;
 850
 851         offset = (loff_t)(start - vma->vm_start)
 852                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 853
 854         /*
 855          * Filesystem's fallocate may need to take i_mutex.  We need to
 856          * explicitly grab a reference because the vma (and hence the
 857          * vma's reference to the file) can go away as soon as we drop
 858          * mmap_lock.
 859          */
 860         get_file(f);
 861         if (userfaultfd_remove(vma, start, end)) {
 862                 /* mmap_lock was not released by userfaultfd_remove() */
 863                 mmap_read_unlock(mm);
 864         }
 865         error = vfs_fallocate(f,
 866                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 867                                 offset, end - start);
 868         fput(f);
 869         mmap_read_lock(mm);
 870         return error;
 871 }
 872
 873 #ifdef CONFIG_MEMORY_FAILURE
 874 /*
 875  * Error injection support for memory error handling.
 876  */
 877 static int madvise_inject_error(int behavior,
 878                 unsigned long start, unsigned long end)
 879 {
 880         struct zone *zone;
 881         unsigned long size;
 882
 883         if (!capable(CAP_SYS_ADMIN))
 884                 return -EPERM;
 885
 886
 887         for (; start < end; start += size) {
 888                 unsigned long pfn;
 889                 struct page *page;
 890                 int ret;
 891
 892                 ret = get_user_pages_fast(start, 1, 0, &page);
 893                 if (ret != 1)
 894                         return ret;
 895                 pfn = page_to_pfn(page);
 896
 897                 /*
 898                  * When soft offlining hugepages, after migrating the page
 899                  * we dissolve it, therefore in the second loop "page" will
 900                  * no longer be a compound page.
 901                  */
 902                 size = page_size(compound_head(page));
 903
 904                 if (behavior == MADV_SOFT_OFFLINE) {
 905                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 906                                  pfn, start);
 907                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
 908                 } else {
 909                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 910                                  pfn, start);
 911                         ret = memory_failure(pfn, MF_COUNT_INCREASED);
 912                 }
 913
 914                 if (ret)
 915                         return ret;
 916         }
 917
 918         /* Ensure that all poisoned pages are removed from per-cpu lists */
 919         for_each_populated_zone(zone)
 920                 drain_all_pages(zone);
 921
 922         return 0;
 923 }
 924 #endif
 925
 926 static long
 927 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 928                 unsigned long start, unsigned long end, int behavior)
 929 {
 930         switch (behavior) {
 931         case MADV_REMOVE:
 932                 return madvise_remove(vma, prev, start, end);
 933         case MADV_WILLNEED:
 934                 return madvise_willneed(vma, prev, start, end);
 935         case MADV_COLD:
 936                 return madvise_cold(vma, prev, start, end);
 937         case MADV_PAGEOUT:
 938                 return madvise_pageout(vma, prev, start, end);
 939         case MADV_FREE:
 940         case MADV_DONTNEED:
 941                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 942         default:
 943                 return madvise_behavior(vma, prev, start, end, behavior);
 944         }
 945 }
 946
 947 static bool
 948 madvise_behavior_valid(int behavior)
 949 {
 950         switch (behavior) {
 951         case MADV_DOFORK:
 952         case MADV_DONTFORK:
 953         case MADV_NORMAL:
 954         case MADV_SEQUENTIAL:
 955         case MADV_RANDOM:
 956         case MADV_REMOVE:
 957         case MADV_WILLNEED:
 958         case MADV_DONTNEED:
 959         case MADV_FREE:
 960         case MADV_COLD:
 961         case MADV_PAGEOUT:
 962 #ifdef CONFIG_KSM
 963         case MADV_MERGEABLE:
 964         case MADV_UNMERGEABLE:
 965 #endif
 966 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 967         case MADV_HUGEPAGE:
 968         case MADV_NOHUGEPAGE:
 969 #endif
 970         case MADV_DONTDUMP:
 971         case MADV_DODUMP:
 972         case MADV_WIPEONFORK:
 973         case MADV_KEEPONFORK:
 974 #ifdef CONFIG_MEMORY_FAILURE
 975         case MADV_SOFT_OFFLINE:
 976         case MADV_HWPOISON:
 977 #endif
 978                 return true;
 979
 980         default:
 981                 return false;
 982         }
 983 }
 984
 985 static bool
 986 process_madvise_behavior_valid(int behavior)
 987 {
 988         switch (behavior) {
 989         case MADV_COLD:
 990         case MADV_PAGEOUT:
 991                 return true;
 992         default:
 993                 return false;
 994         }
 995 }
 996
 997 /*
 998  * The madvise(2) system call.
 999  *
1000  * Applications can use madvise() to advise the kernel how it should
1001  * handle paging I/O in this VM area.  The idea is to help the kernel
1002  * use appropriate read-ahead and caching techniques.  The information
1003  * provided is advisory only, and can be safely disregarded by the
1004  * kernel without affecting the correct operation of the application.
1005  *
1006  * behavior values:
1007  *  MADV_NORMAL - the default behavior is to read clusters.  This
1008  *              results in some read-ahead and read-behind.
1009  *  MADV_RANDOM - the system should read the minimum amount of data
1010  *              on any access, since it is unlikely that the appli-
1011  *              cation will need more than what it asks for.
1012  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1013  *              once, so they can be aggressively read ahead, and
1014  *              can be freed soon after they are accessed.
1015  *  MADV_WILLNEED - the application is notifying the system to read
1016  *              some pages ahead.
1017  *  MADV_DONTNEED - the application is finished with the given range,
1018  *              so the kernel can free resources associated with it.
1019  *  MADV_FREE - the application marks pages in the given range as lazy free,
1020  *              where actual purges are postponed until memory pressure happens.
1021  *  MADV_REMOVE - the application wants to free up the given range of
1022  *              pages and associated backing store.
1023  *  MADV_DONTFORK - omit this area from child's address space when forking:
1024  *              typically, to avoid COWing pages pinned by get_user_pages().
1025  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1026  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1027  *              range after a fork.
1028  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1029  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1030  *              were corrupted by unrecoverable hardware memory failure.
1031  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1032  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1033  *              this area with pages of identical content from other such areas.
1034  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1035  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1036  *              huge pages in the future. Existing pages might be coalesced and
1037  *              new pages might be allocated as THP.
1038  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1039  *              transparent huge pages so the existing pages will not be
1040  *              coalesced into THP and new pages will not be allocated as THP.
1041  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1042  *              from being included in its core dump.
1043  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1044  *  MADV_COLD - the application is not expected to use this memory soon,
1045  *              deactivate pages in this range so that they can be reclaimed
1046  *              easily if memory pressure hanppens.
1047  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1048  *              page out the pages in this range immediately.
1049  *
1050  * return values:
1051  *  zero    - success
1052  *  -EINVAL - start + len < 0, start is not page-aligned,
1053  *              "behavior" is not a valid value, or application
1054  *              is attempting to release locked or shared pages,
1055  *              or the specified address range includes file, Huge TLB,
1056  *              MAP_SHARED or VMPFNMAP range.
1057  *  -ENOMEM - addresses in the specified range are not currently
1058  *              mapped, or are outside the AS of the process.
1059  *  -EIO    - an I/O error occurred while paging in data.
1060  *  -EBADF  - map exists, but area maps something that isn't a file.
1061  *  -EAGAIN - a kernel resource was temporarily unavailable.
1062  */
1063 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1064 {
1065         unsigned long end, tmp;
1066         struct vm_area_struct *vma, *prev;
1067         int unmapped_error = 0;
1068         int error = -EINVAL;
1069         int write;
1070         size_t len;
1071         struct blk_plug plug;
1072
1073         start = untagged_addr(start);
1074
1075         if (!madvise_behavior_valid(behavior))
1076                 return error;
1077
1078         if (!PAGE_ALIGNED(start))
1079                 return error;
1080         len = PAGE_ALIGN(len_in);
1081
1082         /* Check to see whether len was rounded up from small -ve to zero */
1083         if (len_in && !len)
1084                 return error;
1085
1086         end = start + len;
1087         if (end < start)
1088                 return error;
1089
1090         error = 0;
1091         if (end == start)
1092                 return error;
1093
1094 #ifdef CONFIG_MEMORY_FAILURE
1095         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1096                 return madvise_inject_error(behavior, start, start + len_in);
1097 #endif
1098
1099         write = madvise_need_mmap_write(behavior);
1100         if (write) {
1101                 if (mmap_write_lock_killable(mm))
1102                         return -EINTR;
1103         } else {
1104                 mmap_read_lock(mm);
1105         }
1106
1107         /*
1108          * If the interval [start,end) covers some unmapped address
1109          * ranges, just ignore them, but return -ENOMEM at the end.
1110          * - different from the way of handling in mlock etc.
1111          */
1112         vma = find_vma_prev(mm, start, &prev);
1113         if (vma && start > vma->vm_start)
1114                 prev = vma;
1115
1116         blk_start_plug(&plug);
1117         for (;;) {
1118                 /* Still start < end. */
1119                 error = -ENOMEM;
1120                 if (!vma)
1121                         goto out;
1122
1123                 /* Here start < (end|vma->vm_end). */
1124                 if (start < vma->vm_start) {
1125                         unmapped_error = -ENOMEM;
1126                         start = vma->vm_start;
1127                         if (start >= end)
1128                                 goto out;
1129                 }
1130
1131                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1132                 tmp = vma->vm_end;
1133                 if (end < tmp)
1134                         tmp = end;
1135
1136                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1137                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1138                 if (error)
1139                         goto out;
1140                 start = tmp;
1141                 if (prev && start < prev->vm_end)
1142                         start = prev->vm_end;
1143                 error = unmapped_error;
1144                 if (start >= end)
1145                         goto out;
1146                 if (prev)
1147                         vma = prev->vm_next;
1148                 else    /* madvise_remove dropped mmap_lock */
1149                         vma = find_vma(mm, start);
1150         }
1151 out:
1152         blk_finish_plug(&plug);
1153         if (write)
1154                 mmap_write_unlock(mm);
1155         else
1156                 mmap_read_unlock(mm);
1157
1158         return error;
1159 }
1160
1161 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1162 {
1163         return do_madvise(current->mm, start, len_in, behavior);
1164 }
1165
1166 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1167                 size_t, vlen, int, behavior, unsigned int, flags)
1168 {
1169         ssize_t ret;
1170         struct iovec iovstack[UIO_FASTIOV], iovec;
1171         struct iovec *iov = iovstack;
1172         struct iov_iter iter;
1173         struct pid *pid;
1174         struct task_struct *task;
1175         struct mm_struct *mm;
1176         size_t total_len;
1177         unsigned int f_flags;
1178
1179         if (flags != 0) {
1180                 ret = -EINVAL;
1181                 goto out;
1182         }
1183
1184         ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1185         if (ret < 0)
1186                 goto out;
1187
1188         pid = pidfd_get_pid(pidfd, &f_flags);
1189         if (IS_ERR(pid)) {
1190                 ret = PTR_ERR(pid);
1191                 goto free_iov;
1192         }
1193
1194         task = get_pid_task(pid, PIDTYPE_PID);
1195         if (!task) {
1196                 ret = -ESRCH;
1197                 goto put_pid;
1198         }
1199
1200         if (!process_madvise_behavior_valid(behavior)) {
1201                 ret = -EINVAL;
1202                 goto release_task;
1203         }
1204
1205         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1206         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1207         if (IS_ERR_OR_NULL(mm)) {
1208                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1209                 goto release_task;
1210         }
1211
1212         /*
1213          * Require CAP_SYS_NICE for influencing process performance. Note that
1214          * only non-destructive hints are currently supported.
1215          */
1216         if (!capable(CAP_SYS_NICE)) {
1217                 ret = -EPERM;
1218                 goto release_mm;
1219         }
1220
1221         total_len = iov_iter_count(&iter);
1222
1223         while (iov_iter_count(&iter)) {
1224                 iovec = iov_iter_iovec(&iter);
1225                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1226                                         iovec.iov_len, behavior);
1227                 if (ret < 0)
1228                         break;
1229                 iov_iter_advance(&iter, iovec.iov_len);
1230         }
1231
1232         if (ret == 0)
1233                 ret = total_len - iov_iter_count(&iter);
1234
1235 release_mm:
1236         mmput(mm);
1237 release_task:
1238         put_task_struct(task);
1239 put_pid:
1240         put_pid(pid);
1241 free_iov:
1242         kfree(iov);
1243 out:
1244         return ret;
1245 }