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         case MADV_POPULATE_READ:
  57         case MADV_POPULATE_WRITE:
  58                 return 0;
  59         default:
  60                 /* be safe, default to 1. list exceptions explicitly */
  61                 return 1;
  62         }
  63 }
  64
  65 /*
  66  * We can potentially split a vm area into separate
  67  * areas, each area with its own behavior.
  68  */
  69 static long madvise_behavior(struct vm_area_struct *vma,
  70                      struct vm_area_struct **prev,
  71                      unsigned long start, unsigned long end, int behavior)
  72 {
  73         struct mm_struct *mm = vma->vm_mm;
  74         int error = 0;
  75         pgoff_t pgoff;
  76         unsigned long new_flags = vma->vm_flags;
  77
  78         switch (behavior) {
  79         case MADV_NORMAL:
  80                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  81                 break;
  82         case MADV_SEQUENTIAL:
  83                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  84                 break;
  85         case MADV_RANDOM:
  86                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  87                 break;
  88         case MADV_DONTFORK:
  89                 new_flags |= VM_DONTCOPY;
  90                 break;
  91         case MADV_DOFORK:
  92                 if (vma->vm_flags & VM_IO) {
  93                         error = -EINVAL;
  94                         goto out;
  95                 }
  96                 new_flags &= ~VM_DONTCOPY;
  97                 break;
  98         case MADV_WIPEONFORK:
  99                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
 100                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
 101                         error = -EINVAL;
 102                         goto out;
 103                 }
 104                 new_flags |= VM_WIPEONFORK;
 105                 break;
 106         case MADV_KEEPONFORK:
 107                 new_flags &= ~VM_WIPEONFORK;
 108                 break;
 109         case MADV_DONTDUMP:
 110                 new_flags |= VM_DONTDUMP;
 111                 break;
 112         case MADV_DODUMP:
 113                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 114                         error = -EINVAL;
 115                         goto out;
 116                 }
 117                 new_flags &= ~VM_DONTDUMP;
 118                 break;
 119         case MADV_MERGEABLE:
 120         case MADV_UNMERGEABLE:
 121                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 122                 if (error)
 123                         goto out_convert_errno;
 124                 break;
 125         case MADV_HUGEPAGE:
 126         case MADV_NOHUGEPAGE:
 127                 error = hugepage_madvise(vma, &new_flags, behavior);
 128                 if (error)
 129                         goto out_convert_errno;
 130                 break;
 131         }
 132
 133         if (new_flags == vma->vm_flags) {
 134                 *prev = vma;
 135                 goto out;
 136         }
 137
 138         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 139         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 140                           vma->vm_file, pgoff, vma_policy(vma),
 141                           vma->vm_userfaultfd_ctx);
 142         if (*prev) {
 143                 vma = *prev;
 144                 goto success;
 145         }
 146
 147         *prev = vma;
 148
 149         if (start != vma->vm_start) {
 150                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 151                         error = -ENOMEM;
 152                         goto out;
 153                 }
 154                 error = __split_vma(mm, vma, start, 1);
 155                 if (error)
 156                         goto out_convert_errno;
 157         }
 158
 159         if (end != vma->vm_end) {
 160                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 161                         error = -ENOMEM;
 162                         goto out;
 163                 }
 164                 error = __split_vma(mm, vma, end, 0);
 165                 if (error)
 166                         goto out_convert_errno;
 167         }
 168
 169 success:
 170         /*
 171          * vm_flags is protected by the mmap_lock held in write mode.
 172          */
 173         vma->vm_flags = new_flags;
 174
 175 out_convert_errno:
 176         /*
 177          * madvise() returns EAGAIN if kernel resources, such as
 178          * slab, are temporarily unavailable.
 179          */
 180         if (error == -ENOMEM)
 181                 error = -EAGAIN;
 182 out:
 183         return error;
 184 }
 185
 186 #ifdef CONFIG_SWAP
 187 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 188         unsigned long end, struct mm_walk *walk)
 189 {
 190         pte_t *orig_pte;
 191         struct vm_area_struct *vma = walk->private;
 192         unsigned long index;
 193
 194         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 195                 return 0;
 196
 197         for (index = start; index != end; index += PAGE_SIZE) {
 198                 pte_t pte;
 199                 swp_entry_t entry;
 200                 struct page *page;
 201                 spinlock_t *ptl;
 202
 203                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 204                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 205                 pte_unmap_unlock(orig_pte, ptl);
 206
 207                 if (pte_present(pte) || pte_none(pte))
 208                         continue;
 209                 entry = pte_to_swp_entry(pte);
 210                 if (unlikely(non_swap_entry(entry)))
 211                         continue;
 212
 213                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 214                                                         vma, index, false);
 215                 if (page)
 216                         put_page(page);
 217         }
 218
 219         return 0;
 220 }
 221
 222 static const struct mm_walk_ops swapin_walk_ops = {
 223         .pmd_entry              = swapin_walk_pmd_entry,
 224 };
 225
 226 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 227                 unsigned long start, unsigned long end,
 228                 struct address_space *mapping)
 229 {
 230         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 231         pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
 232         struct page *page;
 233
 234         rcu_read_lock();
 235         xas_for_each(&xas, page, end_index) {
 236                 swp_entry_t swap;
 237
 238                 if (!xa_is_value(page))
 239                         continue;
 240                 xas_pause(&xas);
 241                 rcu_read_unlock();
 242
 243                 swap = radix_to_swp_entry(page);
 244                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 245                                                         NULL, 0, false);
 246                 if (page)
 247                         put_page(page);
 248
 249                 rcu_read_lock();
 250         }
 251         rcu_read_unlock();
 252
 253         lru_add_drain();        /* Push any new pages onto the LRU now */
 254 }
 255 #endif          /* CONFIG_SWAP */
 256
 257 /*
 258  * Schedule all required I/O operations.  Do not wait for completion.
 259  */
 260 static long madvise_willneed(struct vm_area_struct *vma,
 261                              struct vm_area_struct **prev,
 262                              unsigned long start, unsigned long end)
 263 {
 264         struct mm_struct *mm = vma->vm_mm;
 265         struct file *file = vma->vm_file;
 266         loff_t offset;
 267
 268         *prev = vma;
 269 #ifdef CONFIG_SWAP
 270         if (!file) {
 271                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 272                 lru_add_drain(); /* Push any new pages onto the LRU now */
 273                 return 0;
 274         }
 275
 276         if (shmem_mapping(file->f_mapping)) {
 277                 force_shm_swapin_readahead(vma, start, end,
 278                                         file->f_mapping);
 279                 return 0;
 280         }
 281 #else
 282         if (!file)
 283                 return -EBADF;
 284 #endif
 285
 286         if (IS_DAX(file_inode(file))) {
 287                 /* no bad return value, but ignore advice */
 288                 return 0;
 289         }
 290
 291         /*
 292          * Filesystem's fadvise may need to take various locks.  We need to
 293          * explicitly grab a reference because the vma (and hence the
 294          * vma's reference to the file) can go away as soon as we drop
 295          * mmap_lock.
 296          */
 297         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 298         get_file(file);
 299         offset = (loff_t)(start - vma->vm_start)
 300                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 301         mmap_read_unlock(mm);
 302         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 303         fput(file);
 304         mmap_read_lock(mm);
 305         return 0;
 306 }
 307
 308 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 309                                 unsigned long addr, unsigned long end,
 310                                 struct mm_walk *walk)
 311 {
 312         struct madvise_walk_private *private = walk->private;
 313         struct mmu_gather *tlb = private->tlb;
 314         bool pageout = private->pageout;
 315         struct mm_struct *mm = tlb->mm;
 316         struct vm_area_struct *vma = walk->vma;
 317         pte_t *orig_pte, *pte, ptent;
 318         spinlock_t *ptl;
 319         struct page *page = NULL;
 320         LIST_HEAD(page_list);
 321
 322         if (fatal_signal_pending(current))
 323                 return -EINTR;
 324
 325 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 326         if (pmd_trans_huge(*pmd)) {
 327                 pmd_t orig_pmd;
 328                 unsigned long next = pmd_addr_end(addr, end);
 329
 330                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 331                 ptl = pmd_trans_huge_lock(pmd, vma);
 332                 if (!ptl)
 333                         return 0;
 334
 335                 orig_pmd = *pmd;
 336                 if (is_huge_zero_pmd(orig_pmd))
 337                         goto huge_unlock;
 338
 339                 if (unlikely(!pmd_present(orig_pmd))) {
 340                         VM_BUG_ON(thp_migration_supported() &&
 341                                         !is_pmd_migration_entry(orig_pmd));
 342                         goto huge_unlock;
 343                 }
 344
 345                 page = pmd_page(orig_pmd);
 346
 347                 /* Do not interfere with other mappings of this page */
 348                 if (page_mapcount(page) != 1)
 349                         goto huge_unlock;
 350
 351                 if (next - addr != HPAGE_PMD_SIZE) {
 352                         int err;
 353
 354                         get_page(page);
 355                         spin_unlock(ptl);
 356                         lock_page(page);
 357                         err = split_huge_page(page);
 358                         unlock_page(page);
 359                         put_page(page);
 360                         if (!err)
 361                                 goto regular_page;
 362                         return 0;
 363                 }
 364
 365                 if (pmd_young(orig_pmd)) {
 366                         pmdp_invalidate(vma, addr, pmd);
 367                         orig_pmd = pmd_mkold(orig_pmd);
 368
 369                         set_pmd_at(mm, addr, pmd, orig_pmd);
 370                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 371                 }
 372
 373                 ClearPageReferenced(page);
 374                 test_and_clear_page_young(page);
 375                 if (pageout) {
 376                         if (!isolate_lru_page(page)) {
 377                                 if (PageUnevictable(page))
 378                                         putback_lru_page(page);
 379                                 else
 380                                         list_add(&page->lru, &page_list);
 381                         }
 382                 } else
 383                         deactivate_page(page);
 384 huge_unlock:
 385                 spin_unlock(ptl);
 386                 if (pageout)
 387                         reclaim_pages(&page_list);
 388                 return 0;
 389         }
 390
 391 regular_page:
 392         if (pmd_trans_unstable(pmd))
 393                 return 0;
 394 #endif
 395         tlb_change_page_size(tlb, PAGE_SIZE);
 396         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 397         flush_tlb_batched_pending(mm);
 398         arch_enter_lazy_mmu_mode();
 399         for (; addr < end; pte++, addr += PAGE_SIZE) {
 400                 ptent = *pte;
 401
 402                 if (pte_none(ptent))
 403                         continue;
 404
 405                 if (!pte_present(ptent))
 406                         continue;
 407
 408                 page = vm_normal_page(vma, addr, ptent);
 409                 if (!page)
 410                         continue;
 411
 412                 /*
 413                  * Creating a THP page is expensive so split it only if we
 414                  * are sure it's worth. Split it if we are only owner.
 415                  */
 416                 if (PageTransCompound(page)) {
 417                         if (page_mapcount(page) != 1)
 418                                 break;
 419                         get_page(page);
 420                         if (!trylock_page(page)) {
 421                                 put_page(page);
 422                                 break;
 423                         }
 424                         pte_unmap_unlock(orig_pte, ptl);
 425                         if (split_huge_page(page)) {
 426                                 unlock_page(page);
 427                                 put_page(page);
 428                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 429                                 break;
 430                         }
 431                         unlock_page(page);
 432                         put_page(page);
 433                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 434                         pte--;
 435                         addr -= PAGE_SIZE;
 436                         continue;
 437                 }
 438
 439                 /* Do not interfere with other mappings of this page */
 440                 if (page_mapcount(page) != 1)
 441                         continue;
 442
 443                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 444
 445                 if (pte_young(ptent)) {
 446                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 447                                                         tlb->fullmm);
 448                         ptent = pte_mkold(ptent);
 449                         set_pte_at(mm, addr, pte, ptent);
 450                         tlb_remove_tlb_entry(tlb, pte, addr);
 451                 }
 452
 453                 /*
 454                  * We are deactivating a page for accelerating reclaiming.
 455                  * VM couldn't reclaim the page unless we clear PG_young.
 456                  * As a side effect, it makes confuse idle-page tracking
 457                  * because they will miss recent referenced history.
 458                  */
 459                 ClearPageReferenced(page);
 460                 test_and_clear_page_young(page);
 461                 if (pageout) {
 462                         if (!isolate_lru_page(page)) {
 463                                 if (PageUnevictable(page))
 464                                         putback_lru_page(page);
 465                                 else
 466                                         list_add(&page->lru, &page_list);
 467                         }
 468                 } else
 469                         deactivate_page(page);
 470         }
 471
 472         arch_leave_lazy_mmu_mode();
 473         pte_unmap_unlock(orig_pte, ptl);
 474         if (pageout)
 475                 reclaim_pages(&page_list);
 476         cond_resched();
 477
 478         return 0;
 479 }
 480
 481 static const struct mm_walk_ops cold_walk_ops = {
 482         .pmd_entry = madvise_cold_or_pageout_pte_range,
 483 };
 484
 485 static void madvise_cold_page_range(struct mmu_gather *tlb,
 486                              struct vm_area_struct *vma,
 487                              unsigned long addr, unsigned long end)
 488 {
 489         struct madvise_walk_private walk_private = {
 490                 .pageout = false,
 491                 .tlb = tlb,
 492         };
 493
 494         tlb_start_vma(tlb, vma);
 495         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 496         tlb_end_vma(tlb, vma);
 497 }
 498
 499 static long madvise_cold(struct vm_area_struct *vma,
 500                         struct vm_area_struct **prev,
 501                         unsigned long start_addr, unsigned long end_addr)
 502 {
 503         struct mm_struct *mm = vma->vm_mm;
 504         struct mmu_gather tlb;
 505
 506         *prev = vma;
 507         if (!can_madv_lru_vma(vma))
 508                 return -EINVAL;
 509
 510         lru_add_drain();
 511         tlb_gather_mmu(&tlb, mm);
 512         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 513         tlb_finish_mmu(&tlb);
 514
 515         return 0;
 516 }
 517
 518 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 519                              struct vm_area_struct *vma,
 520                              unsigned long addr, unsigned long end)
 521 {
 522         struct madvise_walk_private walk_private = {
 523                 .pageout = true,
 524                 .tlb = tlb,
 525         };
 526
 527         tlb_start_vma(tlb, vma);
 528         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 529         tlb_end_vma(tlb, vma);
 530 }
 531
 532 static inline bool can_do_pageout(struct vm_area_struct *vma)
 533 {
 534         if (vma_is_anonymous(vma))
 535                 return true;
 536         if (!vma->vm_file)
 537                 return false;
 538         /*
 539          * paging out pagecache only for non-anonymous mappings that correspond
 540          * to the files the calling process could (if tried) open for writing;
 541          * otherwise we'd be including shared non-exclusive mappings, which
 542          * opens a side channel.
 543          */
 544         return inode_owner_or_capable(&init_user_ns,
 545                                       file_inode(vma->vm_file)) ||
 546                file_permission(vma->vm_file, MAY_WRITE) == 0;
 547 }
 548
 549 static long madvise_pageout(struct vm_area_struct *vma,
 550                         struct vm_area_struct **prev,
 551                         unsigned long start_addr, unsigned long end_addr)
 552 {
 553         struct mm_struct *mm = vma->vm_mm;
 554         struct mmu_gather tlb;
 555
 556         *prev = vma;
 557         if (!can_madv_lru_vma(vma))
 558                 return -EINVAL;
 559
 560         if (!can_do_pageout(vma))
 561                 return 0;
 562
 563         lru_add_drain();
 564         tlb_gather_mmu(&tlb, mm);
 565         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 566         tlb_finish_mmu(&tlb);
 567
 568         return 0;
 569 }
 570
 571 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 572                                 unsigned long end, struct mm_walk *walk)
 573
 574 {
 575         struct mmu_gather *tlb = walk->private;
 576         struct mm_struct *mm = tlb->mm;
 577         struct vm_area_struct *vma = walk->vma;
 578         spinlock_t *ptl;
 579         pte_t *orig_pte, *pte, ptent;
 580         struct page *page;
 581         int nr_swap = 0;
 582         unsigned long next;
 583
 584         next = pmd_addr_end(addr, end);
 585         if (pmd_trans_huge(*pmd))
 586                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 587                         goto next;
 588
 589         if (pmd_trans_unstable(pmd))
 590                 return 0;
 591
 592         tlb_change_page_size(tlb, PAGE_SIZE);
 593         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 594         flush_tlb_batched_pending(mm);
 595         arch_enter_lazy_mmu_mode();
 596         for (; addr != end; pte++, addr += PAGE_SIZE) {
 597                 ptent = *pte;
 598
 599                 if (pte_none(ptent))
 600                         continue;
 601                 /*
 602                  * If the pte has swp_entry, just clear page table to
 603                  * prevent swap-in which is more expensive rather than
 604                  * (page allocation + zeroing).
 605                  */
 606                 if (!pte_present(ptent)) {
 607                         swp_entry_t entry;
 608
 609                         entry = pte_to_swp_entry(ptent);
 610                         if (non_swap_entry(entry))
 611                                 continue;
 612                         nr_swap--;
 613                         free_swap_and_cache(entry);
 614                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 615                         continue;
 616                 }
 617
 618                 page = vm_normal_page(vma, addr, ptent);
 619                 if (!page)
 620                         continue;
 621
 622                 /*
 623                  * If pmd isn't transhuge but the page is THP and
 624                  * is owned by only this process, split it and
 625                  * deactivate all pages.
 626                  */
 627                 if (PageTransCompound(page)) {
 628                         if (page_mapcount(page) != 1)
 629                                 goto out;
 630                         get_page(page);
 631                         if (!trylock_page(page)) {
 632                                 put_page(page);
 633                                 goto out;
 634                         }
 635                         pte_unmap_unlock(orig_pte, ptl);
 636                         if (split_huge_page(page)) {
 637                                 unlock_page(page);
 638                                 put_page(page);
 639                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 640                                 goto out;
 641                         }
 642                         unlock_page(page);
 643                         put_page(page);
 644                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 645                         pte--;
 646                         addr -= PAGE_SIZE;
 647                         continue;
 648                 }
 649
 650                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 651
 652                 if (PageSwapCache(page) || PageDirty(page)) {
 653                         if (!trylock_page(page))
 654                                 continue;
 655                         /*
 656                          * If page is shared with others, we couldn't clear
 657                          * PG_dirty of the page.
 658                          */
 659                         if (page_mapcount(page) != 1) {
 660                                 unlock_page(page);
 661                                 continue;
 662                         }
 663
 664                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 665                                 unlock_page(page);
 666                                 continue;
 667                         }
 668
 669                         ClearPageDirty(page);
 670                         unlock_page(page);
 671                 }
 672
 673                 if (pte_young(ptent) || pte_dirty(ptent)) {
 674                         /*
 675                          * Some of architecture(ex, PPC) don't update TLB
 676                          * with set_pte_at and tlb_remove_tlb_entry so for
 677                          * the portability, remap the pte with old|clean
 678                          * after pte clearing.
 679                          */
 680                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 681                                                         tlb->fullmm);
 682
 683                         ptent = pte_mkold(ptent);
 684                         ptent = pte_mkclean(ptent);
 685                         set_pte_at(mm, addr, pte, ptent);
 686                         tlb_remove_tlb_entry(tlb, pte, addr);
 687                 }
 688                 mark_page_lazyfree(page);
 689         }
 690 out:
 691         if (nr_swap) {
 692                 if (current->mm == mm)
 693                         sync_mm_rss(mm);
 694
 695                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 696         }
 697         arch_leave_lazy_mmu_mode();
 698         pte_unmap_unlock(orig_pte, ptl);
 699         cond_resched();
 700 next:
 701         return 0;
 702 }
 703
 704 static const struct mm_walk_ops madvise_free_walk_ops = {
 705         .pmd_entry              = madvise_free_pte_range,
 706 };
 707
 708 static int madvise_free_single_vma(struct vm_area_struct *vma,
 709                         unsigned long start_addr, unsigned long end_addr)
 710 {
 711         struct mm_struct *mm = vma->vm_mm;
 712         struct mmu_notifier_range range;
 713         struct mmu_gather tlb;
 714
 715         /* MADV_FREE works for only anon vma at the moment */
 716         if (!vma_is_anonymous(vma))
 717                 return -EINVAL;
 718
 719         range.start = max(vma->vm_start, start_addr);
 720         if (range.start >= vma->vm_end)
 721                 return -EINVAL;
 722         range.end = min(vma->vm_end, end_addr);
 723         if (range.end <= vma->vm_start)
 724                 return -EINVAL;
 725         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 726                                 range.start, range.end);
 727
 728         lru_add_drain();
 729         tlb_gather_mmu(&tlb, mm);
 730         update_hiwater_rss(mm);
 731
 732         mmu_notifier_invalidate_range_start(&range);
 733         tlb_start_vma(&tlb, vma);
 734         walk_page_range(vma->vm_mm, range.start, range.end,
 735                         &madvise_free_walk_ops, &tlb);
 736         tlb_end_vma(&tlb, vma);
 737         mmu_notifier_invalidate_range_end(&range);
 738         tlb_finish_mmu(&tlb);
 739
 740         return 0;
 741 }
 742
 743 /*
 744  * Application no longer needs these pages.  If the pages are dirty,
 745  * it's OK to just throw them away.  The app will be more careful about
 746  * data it wants to keep.  Be sure to free swap resources too.  The
 747  * zap_page_range call sets things up for shrink_active_list to actually free
 748  * these pages later if no one else has touched them in the meantime,
 749  * although we could add these pages to a global reuse list for
 750  * shrink_active_list to pick up before reclaiming other pages.
 751  *
 752  * NB: This interface discards data rather than pushes it out to swap,
 753  * as some implementations do.  This has performance implications for
 754  * applications like large transactional databases which want to discard
 755  * pages in anonymous maps after committing to backing store the data
 756  * that was kept in them.  There is no reason to write this data out to
 757  * the swap area if the application is discarding it.
 758  *
 759  * An interface that causes the system to free clean pages and flush
 760  * dirty pages is already available as msync(MS_INVALIDATE).
 761  */
 762 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 763                                         unsigned long start, unsigned long end)
 764 {
 765         zap_page_range(vma, start, end - start);
 766         return 0;
 767 }
 768
 769 static long madvise_dontneed_free(struct vm_area_struct *vma,
 770                                   struct vm_area_struct **prev,
 771                                   unsigned long start, unsigned long end,
 772                                   int behavior)
 773 {
 774         struct mm_struct *mm = vma->vm_mm;
 775
 776         *prev = vma;
 777         if (!can_madv_lru_vma(vma))
 778                 return -EINVAL;
 779
 780         if (!userfaultfd_remove(vma, start, end)) {
 781                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 782
 783                 mmap_read_lock(mm);
 784                 vma = find_vma(mm, start);
 785                 if (!vma)
 786                         return -ENOMEM;
 787                 if (start < vma->vm_start) {
 788                         /*
 789                          * This "vma" under revalidation is the one
 790                          * with the lowest vma->vm_start where start
 791                          * is also < vma->vm_end. If start <
 792                          * vma->vm_start it means an hole materialized
 793                          * in the user address space within the
 794                          * virtual range passed to MADV_DONTNEED
 795                          * or MADV_FREE.
 796                          */
 797                         return -ENOMEM;
 798                 }
 799                 if (!can_madv_lru_vma(vma))
 800                         return -EINVAL;
 801                 if (end > vma->vm_end) {
 802                         /*
 803                          * Don't fail if end > vma->vm_end. If the old
 804                          * vma was split while the mmap_lock was
 805                          * released the effect of the concurrent
 806                          * operation may not cause madvise() to
 807                          * have an undefined result. There may be an
 808                          * adjacent next vma that we'll walk
 809                          * next. userfaultfd_remove() will generate an
 810                          * UFFD_EVENT_REMOVE repetition on the
 811                          * end-vma->vm_end range, but the manager can
 812                          * handle a repetition fine.
 813                          */
 814                         end = vma->vm_end;
 815                 }
 816                 VM_WARN_ON(start >= end);
 817         }
 818
 819         if (behavior == MADV_DONTNEED)
 820                 return madvise_dontneed_single_vma(vma, start, end);
 821         else if (behavior == MADV_FREE)
 822                 return madvise_free_single_vma(vma, start, end);
 823         else
 824                 return -EINVAL;
 825 }
 826
 827 static long madvise_populate(struct vm_area_struct *vma,
 828                              struct vm_area_struct **prev,
 829                              unsigned long start, unsigned long end,
 830                              int behavior)
 831 {
 832         const bool write = behavior == MADV_POPULATE_WRITE;
 833         struct mm_struct *mm = vma->vm_mm;
 834         unsigned long tmp_end;
 835         int locked = 1;
 836         long pages;
 837
 838         *prev = vma;
 839
 840         while (start < end) {
 841                 /*
 842                  * We might have temporarily dropped the lock. For example,
 843                  * our VMA might have been split.
 844                  */
 845                 if (!vma || start >= vma->vm_end) {
 846                         vma = find_vma(mm, start);
 847                         if (!vma || start < vma->vm_start)
 848                                 return -ENOMEM;
 849                 }
 850
 851                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 852                 /* Populate (prefault) page tables readable/writable. */
 853                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 854                                                &locked);
 855                 if (!locked) {
 856                         mmap_read_lock(mm);
 857                         locked = 1;
 858                         *prev = NULL;
 859                         vma = NULL;
 860                 }
 861                 if (pages < 0) {
 862                         switch (pages) {
 863                         case -EINTR:
 864                                 return -EINTR;
 865                         case -EINVAL: /* Incompatible mappings / permissions. */
 866                                 return -EINVAL;
 867                         case -EHWPOISON:
 868                                 return -EHWPOISON;
 869                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 870                                 return -EFAULT;
 871                         default:
 872                                 pr_warn_once("%s: unhandled return value: %ld\n",
 873                                              __func__, pages);
 874                                 fallthrough;
 875                         case -ENOMEM:
 876                                 return -ENOMEM;
 877                         }
 878                 }
 879                 start += pages * PAGE_SIZE;
 880         }
 881         return 0;
 882 }
 883
 884 /*
 885  * Application wants to free up the pages and associated backing store.
 886  * This is effectively punching a hole into the middle of a file.
 887  */
 888 static long madvise_remove(struct vm_area_struct *vma,
 889                                 struct vm_area_struct **prev,
 890                                 unsigned long start, unsigned long end)
 891 {
 892         loff_t offset;
 893         int error;
 894         struct file *f;
 895         struct mm_struct *mm = vma->vm_mm;
 896
 897         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 898
 899         if (vma->vm_flags & VM_LOCKED)
 900                 return -EINVAL;
 901
 902         f = vma->vm_file;
 903
 904         if (!f || !f->f_mapping || !f->f_mapping->host) {
 905                         return -EINVAL;
 906         }
 907
 908         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 909                 return -EACCES;
 910
 911         offset = (loff_t)(start - vma->vm_start)
 912                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 913
 914         /*
 915          * Filesystem's fallocate may need to take i_rwsem.  We need to
 916          * explicitly grab a reference because the vma (and hence the
 917          * vma's reference to the file) can go away as soon as we drop
 918          * mmap_lock.
 919          */
 920         get_file(f);
 921         if (userfaultfd_remove(vma, start, end)) {
 922                 /* mmap_lock was not released by userfaultfd_remove() */
 923                 mmap_read_unlock(mm);
 924         }
 925         error = vfs_fallocate(f,
 926                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 927                                 offset, end - start);
 928         fput(f);
 929         mmap_read_lock(mm);
 930         return error;
 931 }
 932
 933 #ifdef CONFIG_MEMORY_FAILURE
 934 /*
 935  * Error injection support for memory error handling.
 936  */
 937 static int madvise_inject_error(int behavior,
 938                 unsigned long start, unsigned long end)
 939 {
 940         unsigned long size;
 941
 942         if (!capable(CAP_SYS_ADMIN))
 943                 return -EPERM;
 944
 945
 946         for (; start < end; start += size) {
 947                 unsigned long pfn;
 948                 struct page *page;
 949                 int ret;
 950
 951                 ret = get_user_pages_fast(start, 1, 0, &page);
 952                 if (ret != 1)
 953                         return ret;
 954                 pfn = page_to_pfn(page);
 955
 956                 /*
 957                  * When soft offlining hugepages, after migrating the page
 958                  * we dissolve it, therefore in the second loop "page" will
 959                  * no longer be a compound page.
 960                  */
 961                 size = page_size(compound_head(page));
 962
 963                 if (behavior == MADV_SOFT_OFFLINE) {
 964                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 965                                  pfn, start);
 966                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
 967                 } else {
 968                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 969                                  pfn, start);
 970                         ret = memory_failure(pfn, MF_COUNT_INCREASED);
 971                 }
 972
 973                 if (ret)
 974                         return ret;
 975         }
 976
 977         return 0;
 978 }
 979 #endif
 980
 981 static long
 982 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 983                 unsigned long start, unsigned long end, int behavior)
 984 {
 985         switch (behavior) {
 986         case MADV_REMOVE:
 987                 return madvise_remove(vma, prev, start, end);
 988         case MADV_WILLNEED:
 989                 return madvise_willneed(vma, prev, start, end);
 990         case MADV_COLD:
 991                 return madvise_cold(vma, prev, start, end);
 992         case MADV_PAGEOUT:
 993                 return madvise_pageout(vma, prev, start, end);
 994         case MADV_FREE:
 995         case MADV_DONTNEED:
 996                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 997         case MADV_POPULATE_READ:
 998         case MADV_POPULATE_WRITE:
 999                 return madvise_populate(vma, prev, start, end, behavior);
1000         default:
1001                 return madvise_behavior(vma, prev, start, end, behavior);
1002         }
1003 }
1004
1005 static bool
1006 madvise_behavior_valid(int behavior)
1007 {
1008         switch (behavior) {
1009         case MADV_DOFORK:
1010         case MADV_DONTFORK:
1011         case MADV_NORMAL:
1012         case MADV_SEQUENTIAL:
1013         case MADV_RANDOM:
1014         case MADV_REMOVE:
1015         case MADV_WILLNEED:
1016         case MADV_DONTNEED:
1017         case MADV_FREE:
1018         case MADV_COLD:
1019         case MADV_PAGEOUT:
1020         case MADV_POPULATE_READ:
1021         case MADV_POPULATE_WRITE:
1022 #ifdef CONFIG_KSM
1023         case MADV_MERGEABLE:
1024         case MADV_UNMERGEABLE:
1025 #endif
1026 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1027         case MADV_HUGEPAGE:
1028         case MADV_NOHUGEPAGE:
1029 #endif
1030         case MADV_DONTDUMP:
1031         case MADV_DODUMP:
1032         case MADV_WIPEONFORK:
1033         case MADV_KEEPONFORK:
1034 #ifdef CONFIG_MEMORY_FAILURE
1035         case MADV_SOFT_OFFLINE:
1036         case MADV_HWPOISON:
1037 #endif
1038                 return true;
1039
1040         default:
1041                 return false;
1042         }
1043 }
1044
1045 static bool
1046 process_madvise_behavior_valid(int behavior)
1047 {
1048         switch (behavior) {
1049         case MADV_COLD:
1050         case MADV_PAGEOUT:
1051         case MADV_WILLNEED:
1052                 return true;
1053         default:
1054                 return false;
1055         }
1056 }
1057
1058 /*
1059  * The madvise(2) system call.
1060  *
1061  * Applications can use madvise() to advise the kernel how it should
1062  * handle paging I/O in this VM area.  The idea is to help the kernel
1063  * use appropriate read-ahead and caching techniques.  The information
1064  * provided is advisory only, and can be safely disregarded by the
1065  * kernel without affecting the correct operation of the application.
1066  *
1067  * behavior values:
1068  *  MADV_NORMAL - the default behavior is to read clusters.  This
1069  *              results in some read-ahead and read-behind.
1070  *  MADV_RANDOM - the system should read the minimum amount of data
1071  *              on any access, since it is unlikely that the appli-
1072  *              cation will need more than what it asks for.
1073  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1074  *              once, so they can be aggressively read ahead, and
1075  *              can be freed soon after they are accessed.
1076  *  MADV_WILLNEED - the application is notifying the system to read
1077  *              some pages ahead.
1078  *  MADV_DONTNEED - the application is finished with the given range,
1079  *              so the kernel can free resources associated with it.
1080  *  MADV_FREE - the application marks pages in the given range as lazy free,
1081  *              where actual purges are postponed until memory pressure happens.
1082  *  MADV_REMOVE - the application wants to free up the given range of
1083  *              pages and associated backing store.
1084  *  MADV_DONTFORK - omit this area from child's address space when forking:
1085  *              typically, to avoid COWing pages pinned by get_user_pages().
1086  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1087  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1088  *              range after a fork.
1089  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1090  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1091  *              were corrupted by unrecoverable hardware memory failure.
1092  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1093  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1094  *              this area with pages of identical content from other such areas.
1095  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1096  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1097  *              huge pages in the future. Existing pages might be coalesced and
1098  *              new pages might be allocated as THP.
1099  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1100  *              transparent huge pages so the existing pages will not be
1101  *              coalesced into THP and new pages will not be allocated as THP.
1102  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1103  *              from being included in its core dump.
1104  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1105  *  MADV_COLD - the application is not expected to use this memory soon,
1106  *              deactivate pages in this range so that they can be reclaimed
1107  *              easily if memory pressure happens.
1108  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1109  *              page out the pages in this range immediately.
1110  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1111  *              triggering read faults if required
1112  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1113  *              triggering write faults if required
1114  *
1115  * return values:
1116  *  zero    - success
1117  *  -EINVAL - start + len < 0, start is not page-aligned,
1118  *              "behavior" is not a valid value, or application
1119  *              is attempting to release locked or shared pages,
1120  *              or the specified address range includes file, Huge TLB,
1121  *              MAP_SHARED or VMPFNMAP range.
1122  *  -ENOMEM - addresses in the specified range are not currently
1123  *              mapped, or are outside the AS of the process.
1124  *  -EIO    - an I/O error occurred while paging in data.
1125  *  -EBADF  - map exists, but area maps something that isn't a file.
1126  *  -EAGAIN - a kernel resource was temporarily unavailable.
1127  */
1128 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1129 {
1130         unsigned long end, tmp;
1131         struct vm_area_struct *vma, *prev;
1132         int unmapped_error = 0;
1133         int error = -EINVAL;
1134         int write;
1135         size_t len;
1136         struct blk_plug plug;
1137
1138         start = untagged_addr(start);
1139
1140         if (!madvise_behavior_valid(behavior))
1141                 return error;
1142
1143         if (!PAGE_ALIGNED(start))
1144                 return error;
1145         len = PAGE_ALIGN(len_in);
1146
1147         /* Check to see whether len was rounded up from small -ve to zero */
1148         if (len_in && !len)
1149                 return error;
1150
1151         end = start + len;
1152         if (end < start)
1153                 return error;
1154
1155         error = 0;
1156         if (end == start)
1157                 return error;
1158
1159 #ifdef CONFIG_MEMORY_FAILURE
1160         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1161                 return madvise_inject_error(behavior, start, start + len_in);
1162 #endif
1163
1164         write = madvise_need_mmap_write(behavior);
1165         if (write) {
1166                 if (mmap_write_lock_killable(mm))
1167                         return -EINTR;
1168         } else {
1169                 mmap_read_lock(mm);
1170         }
1171
1172         /*
1173          * If the interval [start,end) covers some unmapped address
1174          * ranges, just ignore them, but return -ENOMEM at the end.
1175          * - different from the way of handling in mlock etc.
1176          */
1177         vma = find_vma_prev(mm, start, &prev);
1178         if (vma && start > vma->vm_start)
1179                 prev = vma;
1180
1181         blk_start_plug(&plug);
1182         for (;;) {
1183                 /* Still start < end. */
1184                 error = -ENOMEM;
1185                 if (!vma)
1186                         goto out;
1187
1188                 /* Here start < (end|vma->vm_end). */
1189                 if (start < vma->vm_start) {
1190                         unmapped_error = -ENOMEM;
1191                         start = vma->vm_start;
1192                         if (start >= end)
1193                                 goto out;
1194                 }
1195
1196                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1197                 tmp = vma->vm_end;
1198                 if (end < tmp)
1199                         tmp = end;
1200
1201                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1202                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1203                 if (error)
1204                         goto out;
1205                 start = tmp;
1206                 if (prev && start < prev->vm_end)
1207                         start = prev->vm_end;
1208                 error = unmapped_error;
1209                 if (start >= end)
1210                         goto out;
1211                 if (prev)
1212                         vma = prev->vm_next;
1213                 else    /* madvise_remove dropped mmap_lock */
1214                         vma = find_vma(mm, start);
1215         }
1216 out:
1217         blk_finish_plug(&plug);
1218         if (write)
1219                 mmap_write_unlock(mm);
1220         else
1221                 mmap_read_unlock(mm);
1222
1223         return error;
1224 }
1225
1226 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1227 {
1228         return do_madvise(current->mm, start, len_in, behavior);
1229 }
1230
1231 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1232                 size_t, vlen, int, behavior, unsigned int, flags)
1233 {
1234         ssize_t ret;
1235         struct iovec iovstack[UIO_FASTIOV], iovec;
1236         struct iovec *iov = iovstack;
1237         struct iov_iter iter;
1238         struct pid *pid;
1239         struct task_struct *task;
1240         struct mm_struct *mm;
1241         size_t total_len;
1242         unsigned int f_flags;
1243
1244         if (flags != 0) {
1245                 ret = -EINVAL;
1246                 goto out;
1247         }
1248
1249         ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1250         if (ret < 0)
1251                 goto out;
1252
1253         pid = pidfd_get_pid(pidfd, &f_flags);
1254         if (IS_ERR(pid)) {
1255                 ret = PTR_ERR(pid);
1256                 goto free_iov;
1257         }
1258
1259         task = get_pid_task(pid, PIDTYPE_PID);
1260         if (!task) {
1261                 ret = -ESRCH;
1262                 goto put_pid;
1263         }
1264
1265         if (!process_madvise_behavior_valid(behavior)) {
1266                 ret = -EINVAL;
1267                 goto release_task;
1268         }
1269
1270         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1271         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1272         if (IS_ERR_OR_NULL(mm)) {
1273                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1274                 goto release_task;
1275         }
1276
1277         /*
1278          * Require CAP_SYS_NICE for influencing process performance. Note that
1279          * only non-destructive hints are currently supported.
1280          */
1281         if (!capable(CAP_SYS_NICE)) {
1282                 ret = -EPERM;
1283                 goto release_mm;
1284         }
1285
1286         total_len = iov_iter_count(&iter);
1287
1288         while (iov_iter_count(&iter)) {
1289                 iovec = iov_iter_iovec(&iter);
1290                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1291                                         iovec.iov_len, behavior);
1292                 if (ret < 0)
1293                         break;
1294                 iov_iter_advance(&iter, iovec.iov_len);
1295         }
1296
1297         if (ret == 0)
1298                 ret = total_len - iov_iter_count(&iter);
1299
1300 release_mm:
1301         mmput(mm);
1302 release_task:
1303         put_task_struct(task);
1304 put_pid:
1305         put_pid(pid);
1306 free_iov:
1307         kfree(iov);
1308 out:
1309         return ret;
1310 }