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/mm_inline.h>
  22 #include <linux/string.h>
  23 #include <linux/uio.h>
  24 #include <linux/ksm.h>
  25 #include <linux/fs.h>
  26 #include <linux/file.h>
  27 #include <linux/blkdev.h>
  28 #include <linux/backing-dev.h>
  29 #include <linux/pagewalk.h>
  30 #include <linux/swap.h>
  31 #include <linux/swapops.h>
  32 #include <linux/shmem_fs.h>
  33 #include <linux/mmu_notifier.h>
  34
  35 #include <asm/tlb.h>
  36
  37 #include "internal.h"
  38
  39 struct madvise_walk_private {
  40         struct mmu_gather *tlb;
  41         bool pageout;
  42 };
  43
  44 /*
  45  * Any behaviour which results in changes to the vma->vm_flags needs to
  46  * take mmap_lock for writing. Others, which simply traverse vmas, need
  47  * to only take it for reading.
  48  */
  49 static int madvise_need_mmap_write(int behavior)
  50 {
  51         switch (behavior) {
  52         case MADV_REMOVE:
  53         case MADV_WILLNEED:
  54         case MADV_DONTNEED:
  55         case MADV_COLD:
  56         case MADV_PAGEOUT:
  57         case MADV_FREE:
  58         case MADV_POPULATE_READ:
  59         case MADV_POPULATE_WRITE:
  60                 return 0;
  61         default:
  62                 /* be safe, default to 1. list exceptions explicitly */
  63                 return 1;
  64         }
  65 }
  66
  67 #ifdef CONFIG_ANON_VMA_NAME
  68 struct anon_vma_name *anon_vma_name_alloc(const char *name)
  69 {
  70         struct anon_vma_name *anon_name;
  71         size_t count;
  72
  73         /* Add 1 for NUL terminator at the end of the anon_name->name */
  74         count = strlen(name) + 1;
  75         anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
  76         if (anon_name) {
  77                 kref_init(&anon_name->kref);
  78                 memcpy(anon_name->name, name, count);
  79         }
  80
  81         return anon_name;
  82 }
  83
  84 void anon_vma_name_free(struct kref *kref)
  85 {
  86         struct anon_vma_name *anon_name =
  87                         container_of(kref, struct anon_vma_name, kref);
  88         kfree(anon_name);
  89 }
  90
  91 struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
  92 {
  93         mmap_assert_locked(vma->vm_mm);
  94
  95         if (vma->vm_file)
  96                 return NULL;
  97
  98         return vma->anon_name;
  99 }
 100
 101 /* mmap_lock should be write-locked */
 102 static int replace_anon_vma_name(struct vm_area_struct *vma,
 103                                  struct anon_vma_name *anon_name)
 104 {
 105         struct anon_vma_name *orig_name = anon_vma_name(vma);
 106
 107         if (!anon_name) {
 108                 vma->anon_name = NULL;
 109                 anon_vma_name_put(orig_name);
 110                 return 0;
 111         }
 112
 113         if (anon_vma_name_eq(orig_name, anon_name))
 114                 return 0;
 115
 116         anon_vma_name_get(anon_name);
 117         vma->anon_name = anon_name;
 118         anon_vma_name_put(orig_name);
 119
 120         return 0;
 121 }
 122 #else /* CONFIG_ANON_VMA_NAME */
 123 static int replace_anon_vma_name(struct vm_area_struct *vma,
 124                                  struct anon_vma_name *anon_name)
 125 {
 126         if (anon_name)
 127                 return -EINVAL;
 128
 129         return 0;
 130 }
 131 #endif /* CONFIG_ANON_VMA_NAME */
 132 /*
 133  * Update the vm_flags on region of a vma, splitting it or merging it as
 134  * necessary.  Must be called with mmap_sem held for writing;
 135  */
 136 static int madvise_update_vma(struct vm_area_struct *vma,
 137                               struct vm_area_struct **prev, unsigned long start,
 138                               unsigned long end, unsigned long new_flags,
 139                               struct anon_vma_name *anon_name)
 140 {
 141         struct mm_struct *mm = vma->vm_mm;
 142         int error;
 143         pgoff_t pgoff;
 144
 145         if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
 146                 *prev = vma;
 147                 return 0;
 148         }
 149
 150         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 151         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 152                           vma->vm_file, pgoff, vma_policy(vma),
 153                           vma->vm_userfaultfd_ctx, anon_name);
 154         if (*prev) {
 155                 vma = *prev;
 156                 goto success;
 157         }
 158
 159         *prev = vma;
 160
 161         if (start != vma->vm_start) {
 162                 if (unlikely(mm->map_count >= sysctl_max_map_count))
 163                         return -ENOMEM;
 164                 error = __split_vma(mm, vma, start, 1);
 165                 if (error)
 166                         return error;
 167         }
 168
 169         if (end != vma->vm_end) {
 170                 if (unlikely(mm->map_count >= sysctl_max_map_count))
 171                         return -ENOMEM;
 172                 error = __split_vma(mm, vma, end, 0);
 173                 if (error)
 174                         return error;
 175         }
 176
 177 success:
 178         /*
 179          * vm_flags is protected by the mmap_lock held in write mode.
 180          */
 181         vma->vm_flags = new_flags;
 182         if (!vma->vm_file) {
 183                 error = replace_anon_vma_name(vma, anon_name);
 184                 if (error)
 185                         return error;
 186         }
 187
 188         return 0;
 189 }
 190
 191 #ifdef CONFIG_SWAP
 192 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 193         unsigned long end, struct mm_walk *walk)
 194 {
 195         pte_t *orig_pte;
 196         struct vm_area_struct *vma = walk->private;
 197         unsigned long index;
 198
 199         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 200                 return 0;
 201
 202         for (index = start; index != end; index += PAGE_SIZE) {
 203                 pte_t pte;
 204                 swp_entry_t entry;
 205                 struct page *page;
 206                 spinlock_t *ptl;
 207
 208                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 209                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 210                 pte_unmap_unlock(orig_pte, ptl);
 211
 212                 if (pte_present(pte) || pte_none(pte))
 213                         continue;
 214                 entry = pte_to_swp_entry(pte);
 215                 if (unlikely(non_swap_entry(entry)))
 216                         continue;
 217
 218                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 219                                                         vma, index, false);
 220                 if (page)
 221                         put_page(page);
 222         }
 223
 224         return 0;
 225 }
 226
 227 static const struct mm_walk_ops swapin_walk_ops = {
 228         .pmd_entry              = swapin_walk_pmd_entry,
 229 };
 230
 231 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 232                 unsigned long start, unsigned long end,
 233                 struct address_space *mapping)
 234 {
 235         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 236         pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
 237         struct page *page;
 238
 239         rcu_read_lock();
 240         xas_for_each(&xas, page, end_index) {
 241                 swp_entry_t swap;
 242
 243                 if (!xa_is_value(page))
 244                         continue;
 245                 xas_pause(&xas);
 246                 rcu_read_unlock();
 247
 248                 swap = radix_to_swp_entry(page);
 249                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 250                                                         NULL, 0, false);
 251                 if (page)
 252                         put_page(page);
 253
 254                 rcu_read_lock();
 255         }
 256         rcu_read_unlock();
 257
 258         lru_add_drain();        /* Push any new pages onto the LRU now */
 259 }
 260 #endif          /* CONFIG_SWAP */
 261
 262 /*
 263  * Schedule all required I/O operations.  Do not wait for completion.
 264  */
 265 static long madvise_willneed(struct vm_area_struct *vma,
 266                              struct vm_area_struct **prev,
 267                              unsigned long start, unsigned long end)
 268 {
 269         struct mm_struct *mm = vma->vm_mm;
 270         struct file *file = vma->vm_file;
 271         loff_t offset;
 272
 273         *prev = vma;
 274 #ifdef CONFIG_SWAP
 275         if (!file) {
 276                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 277                 lru_add_drain(); /* Push any new pages onto the LRU now */
 278                 return 0;
 279         }
 280
 281         if (shmem_mapping(file->f_mapping)) {
 282                 force_shm_swapin_readahead(vma, start, end,
 283                                         file->f_mapping);
 284                 return 0;
 285         }
 286 #else
 287         if (!file)
 288                 return -EBADF;
 289 #endif
 290
 291         if (IS_DAX(file_inode(file))) {
 292                 /* no bad return value, but ignore advice */
 293                 return 0;
 294         }
 295
 296         /*
 297          * Filesystem's fadvise may need to take various locks.  We need to
 298          * explicitly grab a reference because the vma (and hence the
 299          * vma's reference to the file) can go away as soon as we drop
 300          * mmap_lock.
 301          */
 302         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 303         get_file(file);
 304         offset = (loff_t)(start - vma->vm_start)
 305                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 306         mmap_read_unlock(mm);
 307         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 308         fput(file);
 309         mmap_read_lock(mm);
 310         return 0;
 311 }
 312
 313 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 314                                 unsigned long addr, unsigned long end,
 315                                 struct mm_walk *walk)
 316 {
 317         struct madvise_walk_private *private = walk->private;
 318         struct mmu_gather *tlb = private->tlb;
 319         bool pageout = private->pageout;
 320         struct mm_struct *mm = tlb->mm;
 321         struct vm_area_struct *vma = walk->vma;
 322         pte_t *orig_pte, *pte, ptent;
 323         spinlock_t *ptl;
 324         struct page *page = NULL;
 325         LIST_HEAD(page_list);
 326
 327         if (fatal_signal_pending(current))
 328                 return -EINTR;
 329
 330 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 331         if (pmd_trans_huge(*pmd)) {
 332                 pmd_t orig_pmd;
 333                 unsigned long next = pmd_addr_end(addr, end);
 334
 335                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 336                 ptl = pmd_trans_huge_lock(pmd, vma);
 337                 if (!ptl)
 338                         return 0;
 339
 340                 orig_pmd = *pmd;
 341                 if (is_huge_zero_pmd(orig_pmd))
 342                         goto huge_unlock;
 343
 344                 if (unlikely(!pmd_present(orig_pmd))) {
 345                         VM_BUG_ON(thp_migration_supported() &&
 346                                         !is_pmd_migration_entry(orig_pmd));
 347                         goto huge_unlock;
 348                 }
 349
 350                 page = pmd_page(orig_pmd);
 351
 352                 /* Do not interfere with other mappings of this page */
 353                 if (page_mapcount(page) != 1)
 354                         goto huge_unlock;
 355
 356                 if (next - addr != HPAGE_PMD_SIZE) {
 357                         int err;
 358
 359                         get_page(page);
 360                         spin_unlock(ptl);
 361                         lock_page(page);
 362                         err = split_huge_page(page);
 363                         unlock_page(page);
 364                         put_page(page);
 365                         if (!err)
 366                                 goto regular_page;
 367                         return 0;
 368                 }
 369
 370                 if (pmd_young(orig_pmd)) {
 371                         pmdp_invalidate(vma, addr, pmd);
 372                         orig_pmd = pmd_mkold(orig_pmd);
 373
 374                         set_pmd_at(mm, addr, pmd, orig_pmd);
 375                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 376                 }
 377
 378                 ClearPageReferenced(page);
 379                 test_and_clear_page_young(page);
 380                 if (pageout) {
 381                         if (!isolate_lru_page(page)) {
 382                                 if (PageUnevictable(page))
 383                                         putback_lru_page(page);
 384                                 else
 385                                         list_add(&page->lru, &page_list);
 386                         }
 387                 } else
 388                         deactivate_page(page);
 389 huge_unlock:
 390                 spin_unlock(ptl);
 391                 if (pageout)
 392                         reclaim_pages(&page_list);
 393                 return 0;
 394         }
 395
 396 regular_page:
 397         if (pmd_trans_unstable(pmd))
 398                 return 0;
 399 #endif
 400         tlb_change_page_size(tlb, PAGE_SIZE);
 401         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 402         flush_tlb_batched_pending(mm);
 403         arch_enter_lazy_mmu_mode();
 404         for (; addr < end; pte++, addr += PAGE_SIZE) {
 405                 ptent = *pte;
 406
 407                 if (pte_none(ptent))
 408                         continue;
 409
 410                 if (!pte_present(ptent))
 411                         continue;
 412
 413                 page = vm_normal_page(vma, addr, ptent);
 414                 if (!page)
 415                         continue;
 416
 417                 /*
 418                  * Creating a THP page is expensive so split it only if we
 419                  * are sure it's worth. Split it if we are only owner.
 420                  */
 421                 if (PageTransCompound(page)) {
 422                         if (page_mapcount(page) != 1)
 423                                 break;
 424                         get_page(page);
 425                         if (!trylock_page(page)) {
 426                                 put_page(page);
 427                                 break;
 428                         }
 429                         pte_unmap_unlock(orig_pte, ptl);
 430                         if (split_huge_page(page)) {
 431                                 unlock_page(page);
 432                                 put_page(page);
 433                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 434                                 break;
 435                         }
 436                         unlock_page(page);
 437                         put_page(page);
 438                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 439                         pte--;
 440                         addr -= PAGE_SIZE;
 441                         continue;
 442                 }
 443
 444                 /* Do not interfere with other mappings of this page */
 445                 if (page_mapcount(page) != 1)
 446                         continue;
 447
 448                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 449
 450                 if (pte_young(ptent)) {
 451                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 452                                                         tlb->fullmm);
 453                         ptent = pte_mkold(ptent);
 454                         set_pte_at(mm, addr, pte, ptent);
 455                         tlb_remove_tlb_entry(tlb, pte, addr);
 456                 }
 457
 458                 /*
 459                  * We are deactivating a page for accelerating reclaiming.
 460                  * VM couldn't reclaim the page unless we clear PG_young.
 461                  * As a side effect, it makes confuse idle-page tracking
 462                  * because they will miss recent referenced history.
 463                  */
 464                 ClearPageReferenced(page);
 465                 test_and_clear_page_young(page);
 466                 if (pageout) {
 467                         if (!isolate_lru_page(page)) {
 468                                 if (PageUnevictable(page))
 469                                         putback_lru_page(page);
 470                                 else
 471                                         list_add(&page->lru, &page_list);
 472                         }
 473                 } else
 474                         deactivate_page(page);
 475         }
 476
 477         arch_leave_lazy_mmu_mode();
 478         pte_unmap_unlock(orig_pte, ptl);
 479         if (pageout)
 480                 reclaim_pages(&page_list);
 481         cond_resched();
 482
 483         return 0;
 484 }
 485
 486 static const struct mm_walk_ops cold_walk_ops = {
 487         .pmd_entry = madvise_cold_or_pageout_pte_range,
 488 };
 489
 490 static void madvise_cold_page_range(struct mmu_gather *tlb,
 491                              struct vm_area_struct *vma,
 492                              unsigned long addr, unsigned long end)
 493 {
 494         struct madvise_walk_private walk_private = {
 495                 .pageout = false,
 496                 .tlb = tlb,
 497         };
 498
 499         tlb_start_vma(tlb, vma);
 500         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 501         tlb_end_vma(tlb, vma);
 502 }
 503
 504 static long madvise_cold(struct vm_area_struct *vma,
 505                         struct vm_area_struct **prev,
 506                         unsigned long start_addr, unsigned long end_addr)
 507 {
 508         struct mm_struct *mm = vma->vm_mm;
 509         struct mmu_gather tlb;
 510
 511         *prev = vma;
 512         if (!can_madv_lru_vma(vma))
 513                 return -EINVAL;
 514
 515         lru_add_drain();
 516         tlb_gather_mmu(&tlb, mm);
 517         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 518         tlb_finish_mmu(&tlb);
 519
 520         return 0;
 521 }
 522
 523 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 524                              struct vm_area_struct *vma,
 525                              unsigned long addr, unsigned long end)
 526 {
 527         struct madvise_walk_private walk_private = {
 528                 .pageout = true,
 529                 .tlb = tlb,
 530         };
 531
 532         tlb_start_vma(tlb, vma);
 533         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 534         tlb_end_vma(tlb, vma);
 535 }
 536
 537 static inline bool can_do_pageout(struct vm_area_struct *vma)
 538 {
 539         if (vma_is_anonymous(vma))
 540                 return true;
 541         if (!vma->vm_file)
 542                 return false;
 543         /*
 544          * paging out pagecache only for non-anonymous mappings that correspond
 545          * to the files the calling process could (if tried) open for writing;
 546          * otherwise we'd be including shared non-exclusive mappings, which
 547          * opens a side channel.
 548          */
 549         return inode_owner_or_capable(&init_user_ns,
 550                                       file_inode(vma->vm_file)) ||
 551                file_permission(vma->vm_file, MAY_WRITE) == 0;
 552 }
 553
 554 static long madvise_pageout(struct vm_area_struct *vma,
 555                         struct vm_area_struct **prev,
 556                         unsigned long start_addr, unsigned long end_addr)
 557 {
 558         struct mm_struct *mm = vma->vm_mm;
 559         struct mmu_gather tlb;
 560
 561         *prev = vma;
 562         if (!can_madv_lru_vma(vma))
 563                 return -EINVAL;
 564
 565         if (!can_do_pageout(vma))
 566                 return 0;
 567
 568         lru_add_drain();
 569         tlb_gather_mmu(&tlb, mm);
 570         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 571         tlb_finish_mmu(&tlb);
 572
 573         return 0;
 574 }
 575
 576 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 577                                 unsigned long end, struct mm_walk *walk)
 578
 579 {
 580         struct mmu_gather *tlb = walk->private;
 581         struct mm_struct *mm = tlb->mm;
 582         struct vm_area_struct *vma = walk->vma;
 583         spinlock_t *ptl;
 584         pte_t *orig_pte, *pte, ptent;
 585         struct page *page;
 586         int nr_swap = 0;
 587         unsigned long next;
 588
 589         next = pmd_addr_end(addr, end);
 590         if (pmd_trans_huge(*pmd))
 591                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 592                         goto next;
 593
 594         if (pmd_trans_unstable(pmd))
 595                 return 0;
 596
 597         tlb_change_page_size(tlb, PAGE_SIZE);
 598         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 599         flush_tlb_batched_pending(mm);
 600         arch_enter_lazy_mmu_mode();
 601         for (; addr != end; pte++, addr += PAGE_SIZE) {
 602                 ptent = *pte;
 603
 604                 if (pte_none(ptent))
 605                         continue;
 606                 /*
 607                  * If the pte has swp_entry, just clear page table to
 608                  * prevent swap-in which is more expensive rather than
 609                  * (page allocation + zeroing).
 610                  */
 611                 if (!pte_present(ptent)) {
 612                         swp_entry_t entry;
 613
 614                         entry = pte_to_swp_entry(ptent);
 615                         if (non_swap_entry(entry))
 616                                 continue;
 617                         nr_swap--;
 618                         free_swap_and_cache(entry);
 619                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 620                         continue;
 621                 }
 622
 623                 page = vm_normal_page(vma, addr, ptent);
 624                 if (!page)
 625                         continue;
 626
 627                 /*
 628                  * If pmd isn't transhuge but the page is THP and
 629                  * is owned by only this process, split it and
 630                  * deactivate all pages.
 631                  */
 632                 if (PageTransCompound(page)) {
 633                         if (page_mapcount(page) != 1)
 634                                 goto out;
 635                         get_page(page);
 636                         if (!trylock_page(page)) {
 637                                 put_page(page);
 638                                 goto out;
 639                         }
 640                         pte_unmap_unlock(orig_pte, ptl);
 641                         if (split_huge_page(page)) {
 642                                 unlock_page(page);
 643                                 put_page(page);
 644                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 645                                 goto out;
 646                         }
 647                         unlock_page(page);
 648                         put_page(page);
 649                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 650                         pte--;
 651                         addr -= PAGE_SIZE;
 652                         continue;
 653                 }
 654
 655                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 656
 657                 if (PageSwapCache(page) || PageDirty(page)) {
 658                         if (!trylock_page(page))
 659                                 continue;
 660                         /*
 661                          * If page is shared with others, we couldn't clear
 662                          * PG_dirty of the page.
 663                          */
 664                         if (page_mapcount(page) != 1) {
 665                                 unlock_page(page);
 666                                 continue;
 667                         }
 668
 669                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 670                                 unlock_page(page);
 671                                 continue;
 672                         }
 673
 674                         ClearPageDirty(page);
 675                         unlock_page(page);
 676                 }
 677
 678                 if (pte_young(ptent) || pte_dirty(ptent)) {
 679                         /*
 680                          * Some of architecture(ex, PPC) don't update TLB
 681                          * with set_pte_at and tlb_remove_tlb_entry so for
 682                          * the portability, remap the pte with old|clean
 683                          * after pte clearing.
 684                          */
 685                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 686                                                         tlb->fullmm);
 687
 688                         ptent = pte_mkold(ptent);
 689                         ptent = pte_mkclean(ptent);
 690                         set_pte_at(mm, addr, pte, ptent);
 691                         tlb_remove_tlb_entry(tlb, pte, addr);
 692                 }
 693                 mark_page_lazyfree(page);
 694         }
 695 out:
 696         if (nr_swap) {
 697                 if (current->mm == mm)
 698                         sync_mm_rss(mm);
 699
 700                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 701         }
 702         arch_leave_lazy_mmu_mode();
 703         pte_unmap_unlock(orig_pte, ptl);
 704         cond_resched();
 705 next:
 706         return 0;
 707 }
 708
 709 static const struct mm_walk_ops madvise_free_walk_ops = {
 710         .pmd_entry              = madvise_free_pte_range,
 711 };
 712
 713 static int madvise_free_single_vma(struct vm_area_struct *vma,
 714                         unsigned long start_addr, unsigned long end_addr)
 715 {
 716         struct mm_struct *mm = vma->vm_mm;
 717         struct mmu_notifier_range range;
 718         struct mmu_gather tlb;
 719
 720         /* MADV_FREE works for only anon vma at the moment */
 721         if (!vma_is_anonymous(vma))
 722                 return -EINVAL;
 723
 724         range.start = max(vma->vm_start, start_addr);
 725         if (range.start >= vma->vm_end)
 726                 return -EINVAL;
 727         range.end = min(vma->vm_end, end_addr);
 728         if (range.end <= vma->vm_start)
 729                 return -EINVAL;
 730         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 731                                 range.start, range.end);
 732
 733         lru_add_drain();
 734         tlb_gather_mmu(&tlb, mm);
 735         update_hiwater_rss(mm);
 736
 737         mmu_notifier_invalidate_range_start(&range);
 738         tlb_start_vma(&tlb, vma);
 739         walk_page_range(vma->vm_mm, range.start, range.end,
 740                         &madvise_free_walk_ops, &tlb);
 741         tlb_end_vma(&tlb, vma);
 742         mmu_notifier_invalidate_range_end(&range);
 743         tlb_finish_mmu(&tlb);
 744
 745         return 0;
 746 }
 747
 748 /*
 749  * Application no longer needs these pages.  If the pages are dirty,
 750  * it's OK to just throw them away.  The app will be more careful about
 751  * data it wants to keep.  Be sure to free swap resources too.  The
 752  * zap_page_range call sets things up for shrink_active_list to actually free
 753  * these pages later if no one else has touched them in the meantime,
 754  * although we could add these pages to a global reuse list for
 755  * shrink_active_list to pick up before reclaiming other pages.
 756  *
 757  * NB: This interface discards data rather than pushes it out to swap,
 758  * as some implementations do.  This has performance implications for
 759  * applications like large transactional databases which want to discard
 760  * pages in anonymous maps after committing to backing store the data
 761  * that was kept in them.  There is no reason to write this data out to
 762  * the swap area if the application is discarding it.
 763  *
 764  * An interface that causes the system to free clean pages and flush
 765  * dirty pages is already available as msync(MS_INVALIDATE).
 766  */
 767 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 768                                         unsigned long start, unsigned long end)
 769 {
 770         zap_page_range(vma, start, end - start);
 771         return 0;
 772 }
 773
 774 static long madvise_dontneed_free(struct vm_area_struct *vma,
 775                                   struct vm_area_struct **prev,
 776                                   unsigned long start, unsigned long end,
 777                                   int behavior)
 778 {
 779         struct mm_struct *mm = vma->vm_mm;
 780
 781         *prev = vma;
 782         if (!can_madv_lru_vma(vma))
 783                 return -EINVAL;
 784
 785         if (!userfaultfd_remove(vma, start, end)) {
 786                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 787
 788                 mmap_read_lock(mm);
 789                 vma = find_vma(mm, start);
 790                 if (!vma)
 791                         return -ENOMEM;
 792                 if (start < vma->vm_start) {
 793                         /*
 794                          * This "vma" under revalidation is the one
 795                          * with the lowest vma->vm_start where start
 796                          * is also < vma->vm_end. If start <
 797                          * vma->vm_start it means an hole materialized
 798                          * in the user address space within the
 799                          * virtual range passed to MADV_DONTNEED
 800                          * or MADV_FREE.
 801                          */
 802                         return -ENOMEM;
 803                 }
 804                 if (!can_madv_lru_vma(vma))
 805                         return -EINVAL;
 806                 if (end > vma->vm_end) {
 807                         /*
 808                          * Don't fail if end > vma->vm_end. If the old
 809                          * vma was split while the mmap_lock was
 810                          * released the effect of the concurrent
 811                          * operation may not cause madvise() to
 812                          * have an undefined result. There may be an
 813                          * adjacent next vma that we'll walk
 814                          * next. userfaultfd_remove() will generate an
 815                          * UFFD_EVENT_REMOVE repetition on the
 816                          * end-vma->vm_end range, but the manager can
 817                          * handle a repetition fine.
 818                          */
 819                         end = vma->vm_end;
 820                 }
 821                 VM_WARN_ON(start >= end);
 822         }
 823
 824         if (behavior == MADV_DONTNEED)
 825                 return madvise_dontneed_single_vma(vma, start, end);
 826         else if (behavior == MADV_FREE)
 827                 return madvise_free_single_vma(vma, start, end);
 828         else
 829                 return -EINVAL;
 830 }
 831
 832 static long madvise_populate(struct vm_area_struct *vma,
 833                              struct vm_area_struct **prev,
 834                              unsigned long start, unsigned long end,
 835                              int behavior)
 836 {
 837         const bool write = behavior == MADV_POPULATE_WRITE;
 838         struct mm_struct *mm = vma->vm_mm;
 839         unsigned long tmp_end;
 840         int locked = 1;
 841         long pages;
 842
 843         *prev = vma;
 844
 845         while (start < end) {
 846                 /*
 847                  * We might have temporarily dropped the lock. For example,
 848                  * our VMA might have been split.
 849                  */
 850                 if (!vma || start >= vma->vm_end) {
 851                         vma = find_vma(mm, start);
 852                         if (!vma || start < vma->vm_start)
 853                                 return -ENOMEM;
 854                 }
 855
 856                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 857                 /* Populate (prefault) page tables readable/writable. */
 858                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 859                                                &locked);
 860                 if (!locked) {
 861                         mmap_read_lock(mm);
 862                         locked = 1;
 863                         *prev = NULL;
 864                         vma = NULL;
 865                 }
 866                 if (pages < 0) {
 867                         switch (pages) {
 868                         case -EINTR:
 869                                 return -EINTR;
 870                         case -EINVAL: /* Incompatible mappings / permissions. */
 871                                 return -EINVAL;
 872                         case -EHWPOISON:
 873                                 return -EHWPOISON;
 874                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 875                                 return -EFAULT;
 876                         default:
 877                                 pr_warn_once("%s: unhandled return value: %ld\n",
 878                                              __func__, pages);
 879                                 fallthrough;
 880                         case -ENOMEM:
 881                                 return -ENOMEM;
 882                         }
 883                 }
 884                 start += pages * PAGE_SIZE;
 885         }
 886         return 0;
 887 }
 888
 889 /*
 890  * Application wants to free up the pages and associated backing store.
 891  * This is effectively punching a hole into the middle of a file.
 892  */
 893 static long madvise_remove(struct vm_area_struct *vma,
 894                                 struct vm_area_struct **prev,
 895                                 unsigned long start, unsigned long end)
 896 {
 897         loff_t offset;
 898         int error;
 899         struct file *f;
 900         struct mm_struct *mm = vma->vm_mm;
 901
 902         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 903
 904         if (vma->vm_flags & VM_LOCKED)
 905                 return -EINVAL;
 906
 907         f = vma->vm_file;
 908
 909         if (!f || !f->f_mapping || !f->f_mapping->host) {
 910                         return -EINVAL;
 911         }
 912
 913         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 914                 return -EACCES;
 915
 916         offset = (loff_t)(start - vma->vm_start)
 917                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 918
 919         /*
 920          * Filesystem's fallocate may need to take i_rwsem.  We need to
 921          * explicitly grab a reference because the vma (and hence the
 922          * vma's reference to the file) can go away as soon as we drop
 923          * mmap_lock.
 924          */
 925         get_file(f);
 926         if (userfaultfd_remove(vma, start, end)) {
 927                 /* mmap_lock was not released by userfaultfd_remove() */
 928                 mmap_read_unlock(mm);
 929         }
 930         error = vfs_fallocate(f,
 931                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 932                                 offset, end - start);
 933         fput(f);
 934         mmap_read_lock(mm);
 935         return error;
 936 }
 937
 938 /*
 939  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
 940  * will handle splitting a vm area into separate areas, each area with its own
 941  * behavior.
 942  */
 943 static int madvise_vma_behavior(struct vm_area_struct *vma,
 944                                 struct vm_area_struct **prev,
 945                                 unsigned long start, unsigned long end,
 946                                 unsigned long behavior)
 947 {
 948         int error;
 949         unsigned long new_flags = vma->vm_flags;
 950
 951         switch (behavior) {
 952         case MADV_REMOVE:
 953                 return madvise_remove(vma, prev, start, end);
 954         case MADV_WILLNEED:
 955                 return madvise_willneed(vma, prev, start, end);
 956         case MADV_COLD:
 957                 return madvise_cold(vma, prev, start, end);
 958         case MADV_PAGEOUT:
 959                 return madvise_pageout(vma, prev, start, end);
 960         case MADV_FREE:
 961         case MADV_DONTNEED:
 962                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 963         case MADV_POPULATE_READ:
 964         case MADV_POPULATE_WRITE:
 965                 return madvise_populate(vma, prev, start, end, behavior);
 966         case MADV_NORMAL:
 967                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
 968                 break;
 969         case MADV_SEQUENTIAL:
 970                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
 971                 break;
 972         case MADV_RANDOM:
 973                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
 974                 break;
 975         case MADV_DONTFORK:
 976                 new_flags |= VM_DONTCOPY;
 977                 break;
 978         case MADV_DOFORK:
 979                 if (vma->vm_flags & VM_IO)
 980                         return -EINVAL;
 981                 new_flags &= ~VM_DONTCOPY;
 982                 break;
 983         case MADV_WIPEONFORK:
 984                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
 985                 if (vma->vm_file || vma->vm_flags & VM_SHARED)
 986                         return -EINVAL;
 987                 new_flags |= VM_WIPEONFORK;
 988                 break;
 989         case MADV_KEEPONFORK:
 990                 new_flags &= ~VM_WIPEONFORK;
 991                 break;
 992         case MADV_DONTDUMP:
 993                 new_flags |= VM_DONTDUMP;
 994                 break;
 995         case MADV_DODUMP:
 996                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
 997                         return -EINVAL;
 998                 new_flags &= ~VM_DONTDUMP;
 999                 break;
1000         case MADV_MERGEABLE:
1001         case MADV_UNMERGEABLE:
1002                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1003                 if (error)
1004                         goto out;
1005                 break;
1006         case MADV_HUGEPAGE:
1007         case MADV_NOHUGEPAGE:
1008                 error = hugepage_madvise(vma, &new_flags, behavior);
1009                 if (error)
1010                         goto out;
1011                 break;
1012         }
1013
1014         error = madvise_update_vma(vma, prev, start, end, new_flags,
1015                                    anon_vma_name(vma));
1016
1017 out:
1018         /*
1019          * madvise() returns EAGAIN if kernel resources, such as
1020          * slab, are temporarily unavailable.
1021          */
1022         if (error == -ENOMEM)
1023                 error = -EAGAIN;
1024         return error;
1025 }
1026
1027 #ifdef CONFIG_MEMORY_FAILURE
1028 /*
1029  * Error injection support for memory error handling.
1030  */
1031 static int madvise_inject_error(int behavior,
1032                 unsigned long start, unsigned long end)
1033 {
1034         unsigned long size;
1035
1036         if (!capable(CAP_SYS_ADMIN))
1037                 return -EPERM;
1038
1039
1040         for (; start < end; start += size) {
1041                 unsigned long pfn;
1042                 struct page *page;
1043                 int ret;
1044
1045                 ret = get_user_pages_fast(start, 1, 0, &page);
1046                 if (ret != 1)
1047                         return ret;
1048                 pfn = page_to_pfn(page);
1049
1050                 /*
1051                  * When soft offlining hugepages, after migrating the page
1052                  * we dissolve it, therefore in the second loop "page" will
1053                  * no longer be a compound page.
1054                  */
1055                 size = page_size(compound_head(page));
1056
1057                 if (behavior == MADV_SOFT_OFFLINE) {
1058                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1059                                  pfn, start);
1060                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1061                 } else {
1062                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1063                                  pfn, start);
1064                         ret = memory_failure(pfn, MF_COUNT_INCREASED);
1065                 }
1066
1067                 if (ret)
1068                         return ret;
1069         }
1070
1071         return 0;
1072 }
1073 #endif
1074
1075 static bool
1076 madvise_behavior_valid(int behavior)
1077 {
1078         switch (behavior) {
1079         case MADV_DOFORK:
1080         case MADV_DONTFORK:
1081         case MADV_NORMAL:
1082         case MADV_SEQUENTIAL:
1083         case MADV_RANDOM:
1084         case MADV_REMOVE:
1085         case MADV_WILLNEED:
1086         case MADV_DONTNEED:
1087         case MADV_FREE:
1088         case MADV_COLD:
1089         case MADV_PAGEOUT:
1090         case MADV_POPULATE_READ:
1091         case MADV_POPULATE_WRITE:
1092 #ifdef CONFIG_KSM
1093         case MADV_MERGEABLE:
1094         case MADV_UNMERGEABLE:
1095 #endif
1096 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1097         case MADV_HUGEPAGE:
1098         case MADV_NOHUGEPAGE:
1099 #endif
1100         case MADV_DONTDUMP:
1101         case MADV_DODUMP:
1102         case MADV_WIPEONFORK:
1103         case MADV_KEEPONFORK:
1104 #ifdef CONFIG_MEMORY_FAILURE
1105         case MADV_SOFT_OFFLINE:
1106         case MADV_HWPOISON:
1107 #endif
1108                 return true;
1109
1110         default:
1111                 return false;
1112         }
1113 }
1114
1115 static bool
1116 process_madvise_behavior_valid(int behavior)
1117 {
1118         switch (behavior) {
1119         case MADV_COLD:
1120         case MADV_PAGEOUT:
1121         case MADV_WILLNEED:
1122                 return true;
1123         default:
1124                 return false;
1125         }
1126 }
1127
1128 /*
1129  * Walk the vmas in range [start,end), and call the visit function on each one.
1130  * The visit function will get start and end parameters that cover the overlap
1131  * between the current vma and the original range.  Any unmapped regions in the
1132  * original range will result in this function returning -ENOMEM while still
1133  * calling the visit function on all of the existing vmas in the range.
1134  * Must be called with the mmap_lock held for reading or writing.
1135  */
1136 static
1137 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1138                       unsigned long end, unsigned long arg,
1139                       int (*visit)(struct vm_area_struct *vma,
1140                                    struct vm_area_struct **prev, unsigned long start,
1141                                    unsigned long end, unsigned long arg))
1142 {
1143         struct vm_area_struct *vma;
1144         struct vm_area_struct *prev;
1145         unsigned long tmp;
1146         int unmapped_error = 0;
1147
1148         /*
1149          * If the interval [start,end) covers some unmapped address
1150          * ranges, just ignore them, but return -ENOMEM at the end.
1151          * - different from the way of handling in mlock etc.
1152          */
1153         vma = find_vma_prev(mm, start, &prev);
1154         if (vma && start > vma->vm_start)
1155                 prev = vma;
1156
1157         for (;;) {
1158                 int error;
1159
1160                 /* Still start < end. */
1161                 if (!vma)
1162                         return -ENOMEM;
1163
1164                 /* Here start < (end|vma->vm_end). */
1165                 if (start < vma->vm_start) {
1166                         unmapped_error = -ENOMEM;
1167                         start = vma->vm_start;
1168                         if (start >= end)
1169                                 break;
1170                 }
1171
1172                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1173                 tmp = vma->vm_end;
1174                 if (end < tmp)
1175                         tmp = end;
1176
1177                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1178                 error = visit(vma, &prev, start, tmp, arg);
1179                 if (error)
1180                         return error;
1181                 start = tmp;
1182                 if (prev && start < prev->vm_end)
1183                         start = prev->vm_end;
1184                 if (start >= end)
1185                         break;
1186                 if (prev)
1187                         vma = prev->vm_next;
1188                 else    /* madvise_remove dropped mmap_lock */
1189                         vma = find_vma(mm, start);
1190         }
1191
1192         return unmapped_error;
1193 }
1194
1195 #ifdef CONFIG_ANON_VMA_NAME
1196 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1197                                  struct vm_area_struct **prev,
1198                                  unsigned long start, unsigned long end,
1199                                  unsigned long anon_name)
1200 {
1201         int error;
1202
1203         /* Only anonymous mappings can be named */
1204         if (vma->vm_file)
1205                 return -EBADF;
1206
1207         error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1208                                    (struct anon_vma_name *)anon_name);
1209
1210         /*
1211          * madvise() returns EAGAIN if kernel resources, such as
1212          * slab, are temporarily unavailable.
1213          */
1214         if (error == -ENOMEM)
1215                 error = -EAGAIN;
1216         return error;
1217 }
1218
1219 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1220                           unsigned long len_in, struct anon_vma_name *anon_name)
1221 {
1222         unsigned long end;
1223         unsigned long len;
1224
1225         if (start & ~PAGE_MASK)
1226                 return -EINVAL;
1227         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1228
1229         /* Check to see whether len was rounded up from small -ve to zero */
1230         if (len_in && !len)
1231                 return -EINVAL;
1232
1233         end = start + len;
1234         if (end < start)
1235                 return -EINVAL;
1236
1237         if (end == start)
1238                 return 0;
1239
1240         return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1241                                  madvise_vma_anon_name);
1242 }
1243 #endif /* CONFIG_ANON_VMA_NAME */
1244 /*
1245  * The madvise(2) system call.
1246  *
1247  * Applications can use madvise() to advise the kernel how it should
1248  * handle paging I/O in this VM area.  The idea is to help the kernel
1249  * use appropriate read-ahead and caching techniques.  The information
1250  * provided is advisory only, and can be safely disregarded by the
1251  * kernel without affecting the correct operation of the application.
1252  *
1253  * behavior values:
1254  *  MADV_NORMAL - the default behavior is to read clusters.  This
1255  *              results in some read-ahead and read-behind.
1256  *  MADV_RANDOM - the system should read the minimum amount of data
1257  *              on any access, since it is unlikely that the appli-
1258  *              cation will need more than what it asks for.
1259  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1260  *              once, so they can be aggressively read ahead, and
1261  *              can be freed soon after they are accessed.
1262  *  MADV_WILLNEED - the application is notifying the system to read
1263  *              some pages ahead.
1264  *  MADV_DONTNEED - the application is finished with the given range,
1265  *              so the kernel can free resources associated with it.
1266  *  MADV_FREE - the application marks pages in the given range as lazy free,
1267  *              where actual purges are postponed until memory pressure happens.
1268  *  MADV_REMOVE - the application wants to free up the given range of
1269  *              pages and associated backing store.
1270  *  MADV_DONTFORK - omit this area from child's address space when forking:
1271  *              typically, to avoid COWing pages pinned by get_user_pages().
1272  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1273  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1274  *              range after a fork.
1275  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1276  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1277  *              were corrupted by unrecoverable hardware memory failure.
1278  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1279  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1280  *              this area with pages of identical content from other such areas.
1281  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1282  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1283  *              huge pages in the future. Existing pages might be coalesced and
1284  *              new pages might be allocated as THP.
1285  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1286  *              transparent huge pages so the existing pages will not be
1287  *              coalesced into THP and new pages will not be allocated as THP.
1288  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1289  *              from being included in its core dump.
1290  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1291  *  MADV_COLD - the application is not expected to use this memory soon,
1292  *              deactivate pages in this range so that they can be reclaimed
1293  *              easily if memory pressure happens.
1294  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1295  *              page out the pages in this range immediately.
1296  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1297  *              triggering read faults if required
1298  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1299  *              triggering write faults if required
1300  *
1301  * return values:
1302  *  zero    - success
1303  *  -EINVAL - start + len < 0, start is not page-aligned,
1304  *              "behavior" is not a valid value, or application
1305  *              is attempting to release locked or shared pages,
1306  *              or the specified address range includes file, Huge TLB,
1307  *              MAP_SHARED or VMPFNMAP range.
1308  *  -ENOMEM - addresses in the specified range are not currently
1309  *              mapped, or are outside the AS of the process.
1310  *  -EIO    - an I/O error occurred while paging in data.
1311  *  -EBADF  - map exists, but area maps something that isn't a file.
1312  *  -EAGAIN - a kernel resource was temporarily unavailable.
1313  */
1314 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1315 {
1316         unsigned long end;
1317         int error;
1318         int write;
1319         size_t len;
1320         struct blk_plug plug;
1321
1322         start = untagged_addr(start);
1323
1324         if (!madvise_behavior_valid(behavior))
1325                 return -EINVAL;
1326
1327         if (!PAGE_ALIGNED(start))
1328                 return -EINVAL;
1329         len = PAGE_ALIGN(len_in);
1330
1331         /* Check to see whether len was rounded up from small -ve to zero */
1332         if (len_in && !len)
1333                 return -EINVAL;
1334
1335         end = start + len;
1336         if (end < start)
1337                 return -EINVAL;
1338
1339         if (end == start)
1340                 return 0;
1341
1342 #ifdef CONFIG_MEMORY_FAILURE
1343         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1344                 return madvise_inject_error(behavior, start, start + len_in);
1345 #endif
1346
1347         write = madvise_need_mmap_write(behavior);
1348         if (write) {
1349                 if (mmap_write_lock_killable(mm))
1350                         return -EINTR;
1351         } else {
1352                 mmap_read_lock(mm);
1353         }
1354
1355         blk_start_plug(&plug);
1356         error = madvise_walk_vmas(mm, start, end, behavior,
1357                         madvise_vma_behavior);
1358         blk_finish_plug(&plug);
1359         if (write)
1360                 mmap_write_unlock(mm);
1361         else
1362                 mmap_read_unlock(mm);
1363
1364         return error;
1365 }
1366
1367 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1368 {
1369         return do_madvise(current->mm, start, len_in, behavior);
1370 }
1371
1372 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1373                 size_t, vlen, int, behavior, unsigned int, flags)
1374 {
1375         ssize_t ret;
1376         struct iovec iovstack[UIO_FASTIOV], iovec;
1377         struct iovec *iov = iovstack;
1378         struct iov_iter iter;
1379         struct task_struct *task;
1380         struct mm_struct *mm;
1381         size_t total_len;
1382         unsigned int f_flags;
1383
1384         if (flags != 0) {
1385                 ret = -EINVAL;
1386                 goto out;
1387         }
1388
1389         ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1390         if (ret < 0)
1391                 goto out;
1392
1393         task = pidfd_get_task(pidfd, &f_flags);
1394         if (IS_ERR(task)) {
1395                 ret = PTR_ERR(task);
1396                 goto free_iov;
1397         }
1398
1399         if (!process_madvise_behavior_valid(behavior)) {
1400                 ret = -EINVAL;
1401                 goto release_task;
1402         }
1403
1404         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1405         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1406         if (IS_ERR_OR_NULL(mm)) {
1407                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1408                 goto release_task;
1409         }
1410
1411         /*
1412          * Require CAP_SYS_NICE for influencing process performance. Note that
1413          * only non-destructive hints are currently supported.
1414          */
1415         if (!capable(CAP_SYS_NICE)) {
1416                 ret = -EPERM;
1417                 goto release_mm;
1418         }
1419
1420         total_len = iov_iter_count(&iter);
1421
1422         while (iov_iter_count(&iter)) {
1423                 iovec = iov_iter_iovec(&iter);
1424                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1425                                         iovec.iov_len, behavior);
1426                 if (ret < 0)
1427                         break;
1428                 iov_iter_advance(&iter, iovec.iov_len);
1429         }
1430
1431         if (ret == 0)
1432                 ret = total_len - iov_iter_count(&iter);
1433
1434 release_mm:
1435         mmput(mm);
1436 release_task:
1437         put_task_struct(task);
1438 free_iov:
1439         kfree(iov);
1440 out:
1441         return ret;
1442 }