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         vma->anon_name = anon_vma_name_reuse(anon_name);
 117         anon_vma_name_put(orig_name);
 118
 119         return 0;
 120 }
 121 #else /* CONFIG_ANON_VMA_NAME */
 122 static int replace_anon_vma_name(struct vm_area_struct *vma,
 123                                  struct anon_vma_name *anon_name)
 124 {
 125         if (anon_name)
 126                 return -EINVAL;
 127
 128         return 0;
 129 }
 130 #endif /* CONFIG_ANON_VMA_NAME */
 131 /*
 132  * Update the vm_flags on region of a vma, splitting it or merging it as
 133  * necessary.  Must be called with mmap_sem held for writing;
 134  * Caller should ensure anon_name stability by raising its refcount even when
 135  * anon_name belongs to a valid vma because this function might free that vma.
 136  */
 137 static int madvise_update_vma(struct vm_area_struct *vma,
 138                               struct vm_area_struct **prev, unsigned long start,
 139                               unsigned long end, unsigned long new_flags,
 140                               struct anon_vma_name *anon_name)
 141 {
 142         struct mm_struct *mm = vma->vm_mm;
 143         int error;
 144         pgoff_t pgoff;
 145
 146         if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
 147                 *prev = vma;
 148                 return 0;
 149         }
 150
 151         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 152         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 153                           vma->vm_file, pgoff, vma_policy(vma),
 154                           vma->vm_userfaultfd_ctx, anon_name);
 155         if (*prev) {
 156                 vma = *prev;
 157                 goto success;
 158         }
 159
 160         *prev = vma;
 161
 162         if (start != vma->vm_start) {
 163                 if (unlikely(mm->map_count >= sysctl_max_map_count))
 164                         return -ENOMEM;
 165                 error = __split_vma(mm, vma, start, 1);
 166                 if (error)
 167                         return error;
 168         }
 169
 170         if (end != vma->vm_end) {
 171                 if (unlikely(mm->map_count >= sysctl_max_map_count))
 172                         return -ENOMEM;
 173                 error = __split_vma(mm, vma, end, 0);
 174                 if (error)
 175                         return error;
 176         }
 177
 178 success:
 179         /*
 180          * vm_flags is protected by the mmap_lock held in write mode.
 181          */
 182         vma->vm_flags = new_flags;
 183         if (!vma->vm_file) {
 184                 error = replace_anon_vma_name(vma, anon_name);
 185                 if (error)
 186                         return error;
 187         }
 188
 189         return 0;
 190 }
 191
 192 #ifdef CONFIG_SWAP
 193 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 194         unsigned long end, struct mm_walk *walk)
 195 {
 196         pte_t *orig_pte;
 197         struct vm_area_struct *vma = walk->private;
 198         unsigned long index;
 199
 200         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 201                 return 0;
 202
 203         for (index = start; index != end; index += PAGE_SIZE) {
 204                 pte_t pte;
 205                 swp_entry_t entry;
 206                 struct page *page;
 207                 spinlock_t *ptl;
 208
 209                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 210                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 211                 pte_unmap_unlock(orig_pte, ptl);
 212
 213                 if (pte_present(pte) || pte_none(pte))
 214                         continue;
 215                 entry = pte_to_swp_entry(pte);
 216                 if (unlikely(non_swap_entry(entry)))
 217                         continue;
 218
 219                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 220                                                         vma, index, false);
 221                 if (page)
 222                         put_page(page);
 223         }
 224
 225         return 0;
 226 }
 227
 228 static const struct mm_walk_ops swapin_walk_ops = {
 229         .pmd_entry              = swapin_walk_pmd_entry,
 230 };
 231
 232 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 233                 unsigned long start, unsigned long end,
 234                 struct address_space *mapping)
 235 {
 236         XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
 237         pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
 238         struct page *page;
 239
 240         rcu_read_lock();
 241         xas_for_each(&xas, page, end_index) {
 242                 swp_entry_t swap;
 243
 244                 if (!xa_is_value(page))
 245                         continue;
 246                 xas_pause(&xas);
 247                 rcu_read_unlock();
 248
 249                 swap = radix_to_swp_entry(page);
 250                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 251                                                         NULL, 0, false);
 252                 if (page)
 253                         put_page(page);
 254
 255                 rcu_read_lock();
 256         }
 257         rcu_read_unlock();
 258
 259         lru_add_drain();        /* Push any new pages onto the LRU now */
 260 }
 261 #endif          /* CONFIG_SWAP */
 262
 263 /*
 264  * Schedule all required I/O operations.  Do not wait for completion.
 265  */
 266 static long madvise_willneed(struct vm_area_struct *vma,
 267                              struct vm_area_struct **prev,
 268                              unsigned long start, unsigned long end)
 269 {
 270         struct mm_struct *mm = vma->vm_mm;
 271         struct file *file = vma->vm_file;
 272         loff_t offset;
 273
 274         *prev = vma;
 275 #ifdef CONFIG_SWAP
 276         if (!file) {
 277                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 278                 lru_add_drain(); /* Push any new pages onto the LRU now */
 279                 return 0;
 280         }
 281
 282         if (shmem_mapping(file->f_mapping)) {
 283                 force_shm_swapin_readahead(vma, start, end,
 284                                         file->f_mapping);
 285                 return 0;
 286         }
 287 #else
 288         if (!file)
 289                 return -EBADF;
 290 #endif
 291
 292         if (IS_DAX(file_inode(file))) {
 293                 /* no bad return value, but ignore advice */
 294                 return 0;
 295         }
 296
 297         /*
 298          * Filesystem's fadvise may need to take various locks.  We need to
 299          * explicitly grab a reference because the vma (and hence the
 300          * vma's reference to the file) can go away as soon as we drop
 301          * mmap_lock.
 302          */
 303         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 304         get_file(file);
 305         offset = (loff_t)(start - vma->vm_start)
 306                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 307         mmap_read_unlock(mm);
 308         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 309         fput(file);
 310         mmap_read_lock(mm);
 311         return 0;
 312 }
 313
 314 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 315                                 unsigned long addr, unsigned long end,
 316                                 struct mm_walk *walk)
 317 {
 318         struct madvise_walk_private *private = walk->private;
 319         struct mmu_gather *tlb = private->tlb;
 320         bool pageout = private->pageout;
 321         struct mm_struct *mm = tlb->mm;
 322         struct vm_area_struct *vma = walk->vma;
 323         pte_t *orig_pte, *pte, ptent;
 324         spinlock_t *ptl;
 325         struct page *page = NULL;
 326         LIST_HEAD(page_list);
 327
 328         if (fatal_signal_pending(current))
 329                 return -EINTR;
 330
 331 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 332         if (pmd_trans_huge(*pmd)) {
 333                 pmd_t orig_pmd;
 334                 unsigned long next = pmd_addr_end(addr, end);
 335
 336                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 337                 ptl = pmd_trans_huge_lock(pmd, vma);
 338                 if (!ptl)
 339                         return 0;
 340
 341                 orig_pmd = *pmd;
 342                 if (is_huge_zero_pmd(orig_pmd))
 343                         goto huge_unlock;
 344
 345                 if (unlikely(!pmd_present(orig_pmd))) {
 346                         VM_BUG_ON(thp_migration_supported() &&
 347                                         !is_pmd_migration_entry(orig_pmd));
 348                         goto huge_unlock;
 349                 }
 350
 351                 page = pmd_page(orig_pmd);
 352
 353                 /* Do not interfere with other mappings of this page */
 354                 if (page_mapcount(page) != 1)
 355                         goto huge_unlock;
 356
 357                 if (next - addr != HPAGE_PMD_SIZE) {
 358                         int err;
 359
 360                         get_page(page);
 361                         spin_unlock(ptl);
 362                         lock_page(page);
 363                         err = split_huge_page(page);
 364                         unlock_page(page);
 365                         put_page(page);
 366                         if (!err)
 367                                 goto regular_page;
 368                         return 0;
 369                 }
 370
 371                 if (pmd_young(orig_pmd)) {
 372                         pmdp_invalidate(vma, addr, pmd);
 373                         orig_pmd = pmd_mkold(orig_pmd);
 374
 375                         set_pmd_at(mm, addr, pmd, orig_pmd);
 376                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 377                 }
 378
 379                 ClearPageReferenced(page);
 380                 test_and_clear_page_young(page);
 381                 if (pageout) {
 382                         if (!isolate_lru_page(page)) {
 383                                 if (PageUnevictable(page))
 384                                         putback_lru_page(page);
 385                                 else
 386                                         list_add(&page->lru, &page_list);
 387                         }
 388                 } else
 389                         deactivate_page(page);
 390 huge_unlock:
 391                 spin_unlock(ptl);
 392                 if (pageout)
 393                         reclaim_pages(&page_list);
 394                 return 0;
 395         }
 396
 397 regular_page:
 398         if (pmd_trans_unstable(pmd))
 399                 return 0;
 400 #endif
 401         tlb_change_page_size(tlb, PAGE_SIZE);
 402         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 403         flush_tlb_batched_pending(mm);
 404         arch_enter_lazy_mmu_mode();
 405         for (; addr < end; pte++, addr += PAGE_SIZE) {
 406                 ptent = *pte;
 407
 408                 if (pte_none(ptent))
 409                         continue;
 410
 411                 if (!pte_present(ptent))
 412                         continue;
 413
 414                 page = vm_normal_page(vma, addr, ptent);
 415                 if (!page)
 416                         continue;
 417
 418                 /*
 419                  * Creating a THP page is expensive so split it only if we
 420                  * are sure it's worth. Split it if we are only owner.
 421                  */
 422                 if (PageTransCompound(page)) {
 423                         if (page_mapcount(page) != 1)
 424                                 break;
 425                         get_page(page);
 426                         if (!trylock_page(page)) {
 427                                 put_page(page);
 428                                 break;
 429                         }
 430                         pte_unmap_unlock(orig_pte, ptl);
 431                         if (split_huge_page(page)) {
 432                                 unlock_page(page);
 433                                 put_page(page);
 434                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 435                                 break;
 436                         }
 437                         unlock_page(page);
 438                         put_page(page);
 439                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 440                         pte--;
 441                         addr -= PAGE_SIZE;
 442                         continue;
 443                 }
 444
 445                 /* Do not interfere with other mappings of this page */
 446                 if (page_mapcount(page) != 1)
 447                         continue;
 448
 449                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 450
 451                 if (pte_young(ptent)) {
 452                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 453                                                         tlb->fullmm);
 454                         ptent = pte_mkold(ptent);
 455                         set_pte_at(mm, addr, pte, ptent);
 456                         tlb_remove_tlb_entry(tlb, pte, addr);
 457                 }
 458
 459                 /*
 460                  * We are deactivating a page for accelerating reclaiming.
 461                  * VM couldn't reclaim the page unless we clear PG_young.
 462                  * As a side effect, it makes confuse idle-page tracking
 463                  * because they will miss recent referenced history.
 464                  */
 465                 ClearPageReferenced(page);
 466                 test_and_clear_page_young(page);
 467                 if (pageout) {
 468                         if (!isolate_lru_page(page)) {
 469                                 if (PageUnevictable(page))
 470                                         putback_lru_page(page);
 471                                 else
 472                                         list_add(&page->lru, &page_list);
 473                         }
 474                 } else
 475                         deactivate_page(page);
 476         }
 477
 478         arch_leave_lazy_mmu_mode();
 479         pte_unmap_unlock(orig_pte, ptl);
 480         if (pageout)
 481                 reclaim_pages(&page_list);
 482         cond_resched();
 483
 484         return 0;
 485 }
 486
 487 static const struct mm_walk_ops cold_walk_ops = {
 488         .pmd_entry = madvise_cold_or_pageout_pte_range,
 489 };
 490
 491 static void madvise_cold_page_range(struct mmu_gather *tlb,
 492                              struct vm_area_struct *vma,
 493                              unsigned long addr, unsigned long end)
 494 {
 495         struct madvise_walk_private walk_private = {
 496                 .pageout = false,
 497                 .tlb = tlb,
 498         };
 499
 500         tlb_start_vma(tlb, vma);
 501         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 502         tlb_end_vma(tlb, vma);
 503 }
 504
 505 static long madvise_cold(struct vm_area_struct *vma,
 506                         struct vm_area_struct **prev,
 507                         unsigned long start_addr, unsigned long end_addr)
 508 {
 509         struct mm_struct *mm = vma->vm_mm;
 510         struct mmu_gather tlb;
 511
 512         *prev = vma;
 513         if (!can_madv_lru_vma(vma))
 514                 return -EINVAL;
 515
 516         lru_add_drain();
 517         tlb_gather_mmu(&tlb, mm);
 518         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 519         tlb_finish_mmu(&tlb);
 520
 521         return 0;
 522 }
 523
 524 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 525                              struct vm_area_struct *vma,
 526                              unsigned long addr, unsigned long end)
 527 {
 528         struct madvise_walk_private walk_private = {
 529                 .pageout = true,
 530                 .tlb = tlb,
 531         };
 532
 533         tlb_start_vma(tlb, vma);
 534         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 535         tlb_end_vma(tlb, vma);
 536 }
 537
 538 static inline bool can_do_pageout(struct vm_area_struct *vma)
 539 {
 540         if (vma_is_anonymous(vma))
 541                 return true;
 542         if (!vma->vm_file)
 543                 return false;
 544         /*
 545          * paging out pagecache only for non-anonymous mappings that correspond
 546          * to the files the calling process could (if tried) open for writing;
 547          * otherwise we'd be including shared non-exclusive mappings, which
 548          * opens a side channel.
 549          */
 550         return inode_owner_or_capable(&init_user_ns,
 551                                       file_inode(vma->vm_file)) ||
 552                file_permission(vma->vm_file, MAY_WRITE) == 0;
 553 }
 554
 555 static long madvise_pageout(struct vm_area_struct *vma,
 556                         struct vm_area_struct **prev,
 557                         unsigned long start_addr, unsigned long end_addr)
 558 {
 559         struct mm_struct *mm = vma->vm_mm;
 560         struct mmu_gather tlb;
 561
 562         *prev = vma;
 563         if (!can_madv_lru_vma(vma))
 564                 return -EINVAL;
 565
 566         if (!can_do_pageout(vma))
 567                 return 0;
 568
 569         lru_add_drain();
 570         tlb_gather_mmu(&tlb, mm);
 571         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 572         tlb_finish_mmu(&tlb);
 573
 574         return 0;
 575 }
 576
 577 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 578                                 unsigned long end, struct mm_walk *walk)
 579
 580 {
 581         struct mmu_gather *tlb = walk->private;
 582         struct mm_struct *mm = tlb->mm;
 583         struct vm_area_struct *vma = walk->vma;
 584         spinlock_t *ptl;
 585         pte_t *orig_pte, *pte, ptent;
 586         struct page *page;
 587         int nr_swap = 0;
 588         unsigned long next;
 589
 590         next = pmd_addr_end(addr, end);
 591         if (pmd_trans_huge(*pmd))
 592                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 593                         goto next;
 594
 595         if (pmd_trans_unstable(pmd))
 596                 return 0;
 597
 598         tlb_change_page_size(tlb, PAGE_SIZE);
 599         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 600         flush_tlb_batched_pending(mm);
 601         arch_enter_lazy_mmu_mode();
 602         for (; addr != end; pte++, addr += PAGE_SIZE) {
 603                 ptent = *pte;
 604
 605                 if (pte_none(ptent))
 606                         continue;
 607                 /*
 608                  * If the pte has swp_entry, just clear page table to
 609                  * prevent swap-in which is more expensive rather than
 610                  * (page allocation + zeroing).
 611                  */
 612                 if (!pte_present(ptent)) {
 613                         swp_entry_t entry;
 614
 615                         entry = pte_to_swp_entry(ptent);
 616                         if (non_swap_entry(entry))
 617                                 continue;
 618                         nr_swap--;
 619                         free_swap_and_cache(entry);
 620                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 621                         continue;
 622                 }
 623
 624                 page = vm_normal_page(vma, addr, ptent);
 625                 if (!page)
 626                         continue;
 627
 628                 /*
 629                  * If pmd isn't transhuge but the page is THP and
 630                  * is owned by only this process, split it and
 631                  * deactivate all pages.
 632                  */
 633                 if (PageTransCompound(page)) {
 634                         if (page_mapcount(page) != 1)
 635                                 goto out;
 636                         get_page(page);
 637                         if (!trylock_page(page)) {
 638                                 put_page(page);
 639                                 goto out;
 640                         }
 641                         pte_unmap_unlock(orig_pte, ptl);
 642                         if (split_huge_page(page)) {
 643                                 unlock_page(page);
 644                                 put_page(page);
 645                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 646                                 goto out;
 647                         }
 648                         unlock_page(page);
 649                         put_page(page);
 650                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 651                         pte--;
 652                         addr -= PAGE_SIZE;
 653                         continue;
 654                 }
 655
 656                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 657
 658                 if (PageSwapCache(page) || PageDirty(page)) {
 659                         if (!trylock_page(page))
 660                                 continue;
 661                         /*
 662                          * If page is shared with others, we couldn't clear
 663                          * PG_dirty of the page.
 664                          */
 665                         if (page_mapcount(page) != 1) {
 666                                 unlock_page(page);
 667                                 continue;
 668                         }
 669
 670                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 671                                 unlock_page(page);
 672                                 continue;
 673                         }
 674
 675                         ClearPageDirty(page);
 676                         unlock_page(page);
 677                 }
 678
 679                 if (pte_young(ptent) || pte_dirty(ptent)) {
 680                         /*
 681                          * Some of architecture(ex, PPC) don't update TLB
 682                          * with set_pte_at and tlb_remove_tlb_entry so for
 683                          * the portability, remap the pte with old|clean
 684                          * after pte clearing.
 685                          */
 686                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 687                                                         tlb->fullmm);
 688
 689                         ptent = pte_mkold(ptent);
 690                         ptent = pte_mkclean(ptent);
 691                         set_pte_at(mm, addr, pte, ptent);
 692                         tlb_remove_tlb_entry(tlb, pte, addr);
 693                 }
 694                 mark_page_lazyfree(page);
 695         }
 696 out:
 697         if (nr_swap) {
 698                 if (current->mm == mm)
 699                         sync_mm_rss(mm);
 700
 701                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 702         }
 703         arch_leave_lazy_mmu_mode();
 704         pte_unmap_unlock(orig_pte, ptl);
 705         cond_resched();
 706 next:
 707         return 0;
 708 }
 709
 710 static const struct mm_walk_ops madvise_free_walk_ops = {
 711         .pmd_entry              = madvise_free_pte_range,
 712 };
 713
 714 static int madvise_free_single_vma(struct vm_area_struct *vma,
 715                         unsigned long start_addr, unsigned long end_addr)
 716 {
 717         struct mm_struct *mm = vma->vm_mm;
 718         struct mmu_notifier_range range;
 719         struct mmu_gather tlb;
 720
 721         /* MADV_FREE works for only anon vma at the moment */
 722         if (!vma_is_anonymous(vma))
 723                 return -EINVAL;
 724
 725         range.start = max(vma->vm_start, start_addr);
 726         if (range.start >= vma->vm_end)
 727                 return -EINVAL;
 728         range.end = min(vma->vm_end, end_addr);
 729         if (range.end <= vma->vm_start)
 730                 return -EINVAL;
 731         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 732                                 range.start, range.end);
 733
 734         lru_add_drain();
 735         tlb_gather_mmu(&tlb, mm);
 736         update_hiwater_rss(mm);
 737
 738         mmu_notifier_invalidate_range_start(&range);
 739         tlb_start_vma(&tlb, vma);
 740         walk_page_range(vma->vm_mm, range.start, range.end,
 741                         &madvise_free_walk_ops, &tlb);
 742         tlb_end_vma(&tlb, vma);
 743         mmu_notifier_invalidate_range_end(&range);
 744         tlb_finish_mmu(&tlb);
 745
 746         return 0;
 747 }
 748
 749 /*
 750  * Application no longer needs these pages.  If the pages are dirty,
 751  * it's OK to just throw them away.  The app will be more careful about
 752  * data it wants to keep.  Be sure to free swap resources too.  The
 753  * zap_page_range call sets things up for shrink_active_list to actually free
 754  * these pages later if no one else has touched them in the meantime,
 755  * although we could add these pages to a global reuse list for
 756  * shrink_active_list to pick up before reclaiming other pages.
 757  *
 758  * NB: This interface discards data rather than pushes it out to swap,
 759  * as some implementations do.  This has performance implications for
 760  * applications like large transactional databases which want to discard
 761  * pages in anonymous maps after committing to backing store the data
 762  * that was kept in them.  There is no reason to write this data out to
 763  * the swap area if the application is discarding it.
 764  *
 765  * An interface that causes the system to free clean pages and flush
 766  * dirty pages is already available as msync(MS_INVALIDATE).
 767  */
 768 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 769                                         unsigned long start, unsigned long end)
 770 {
 771         zap_page_range(vma, start, end - start);
 772         return 0;
 773 }
 774
 775 static long madvise_dontneed_free(struct vm_area_struct *vma,
 776                                   struct vm_area_struct **prev,
 777                                   unsigned long start, unsigned long end,
 778                                   int behavior)
 779 {
 780         struct mm_struct *mm = vma->vm_mm;
 781
 782         *prev = vma;
 783         if (!can_madv_lru_vma(vma))
 784                 return -EINVAL;
 785
 786         if (!userfaultfd_remove(vma, start, end)) {
 787                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 788
 789                 mmap_read_lock(mm);
 790                 vma = find_vma(mm, start);
 791                 if (!vma)
 792                         return -ENOMEM;
 793                 if (start < vma->vm_start) {
 794                         /*
 795                          * This "vma" under revalidation is the one
 796                          * with the lowest vma->vm_start where start
 797                          * is also < vma->vm_end. If start <
 798                          * vma->vm_start it means an hole materialized
 799                          * in the user address space within the
 800                          * virtual range passed to MADV_DONTNEED
 801                          * or MADV_FREE.
 802                          */
 803                         return -ENOMEM;
 804                 }
 805                 if (!can_madv_lru_vma(vma))
 806                         return -EINVAL;
 807                 if (end > vma->vm_end) {
 808                         /*
 809                          * Don't fail if end > vma->vm_end. If the old
 810                          * vma was split while the mmap_lock was
 811                          * released the effect of the concurrent
 812                          * operation may not cause madvise() to
 813                          * have an undefined result. There may be an
 814                          * adjacent next vma that we'll walk
 815                          * next. userfaultfd_remove() will generate an
 816                          * UFFD_EVENT_REMOVE repetition on the
 817                          * end-vma->vm_end range, but the manager can
 818                          * handle a repetition fine.
 819                          */
 820                         end = vma->vm_end;
 821                 }
 822                 VM_WARN_ON(start >= end);
 823         }
 824
 825         if (behavior == MADV_DONTNEED)
 826                 return madvise_dontneed_single_vma(vma, start, end);
 827         else if (behavior == MADV_FREE)
 828                 return madvise_free_single_vma(vma, start, end);
 829         else
 830                 return -EINVAL;
 831 }
 832
 833 static long madvise_populate(struct vm_area_struct *vma,
 834                              struct vm_area_struct **prev,
 835                              unsigned long start, unsigned long end,
 836                              int behavior)
 837 {
 838         const bool write = behavior == MADV_POPULATE_WRITE;
 839         struct mm_struct *mm = vma->vm_mm;
 840         unsigned long tmp_end;
 841         int locked = 1;
 842         long pages;
 843
 844         *prev = vma;
 845
 846         while (start < end) {
 847                 /*
 848                  * We might have temporarily dropped the lock. For example,
 849                  * our VMA might have been split.
 850                  */
 851                 if (!vma || start >= vma->vm_end) {
 852                         vma = find_vma(mm, start);
 853                         if (!vma || start < vma->vm_start)
 854                                 return -ENOMEM;
 855                 }
 856
 857                 tmp_end = min_t(unsigned long, end, vma->vm_end);
 858                 /* Populate (prefault) page tables readable/writable. */
 859                 pages = faultin_vma_page_range(vma, start, tmp_end, write,
 860                                                &locked);
 861                 if (!locked) {
 862                         mmap_read_lock(mm);
 863                         locked = 1;
 864                         *prev = NULL;
 865                         vma = NULL;
 866                 }
 867                 if (pages < 0) {
 868                         switch (pages) {
 869                         case -EINTR:
 870                                 return -EINTR;
 871                         case -EINVAL: /* Incompatible mappings / permissions. */
 872                                 return -EINVAL;
 873                         case -EHWPOISON:
 874                                 return -EHWPOISON;
 875                         case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
 876                                 return -EFAULT;
 877                         default:
 878                                 pr_warn_once("%s: unhandled return value: %ld\n",
 879                                              __func__, pages);
 880                                 fallthrough;
 881                         case -ENOMEM:
 882                                 return -ENOMEM;
 883                         }
 884                 }
 885                 start += pages * PAGE_SIZE;
 886         }
 887         return 0;
 888 }
 889
 890 /*
 891  * Application wants to free up the pages and associated backing store.
 892  * This is effectively punching a hole into the middle of a file.
 893  */
 894 static long madvise_remove(struct vm_area_struct *vma,
 895                                 struct vm_area_struct **prev,
 896                                 unsigned long start, unsigned long end)
 897 {
 898         loff_t offset;
 899         int error;
 900         struct file *f;
 901         struct mm_struct *mm = vma->vm_mm;
 902
 903         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 904
 905         if (vma->vm_flags & VM_LOCKED)
 906                 return -EINVAL;
 907
 908         f = vma->vm_file;
 909
 910         if (!f || !f->f_mapping || !f->f_mapping->host) {
 911                         return -EINVAL;
 912         }
 913
 914         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 915                 return -EACCES;
 916
 917         offset = (loff_t)(start - vma->vm_start)
 918                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 919
 920         /*
 921          * Filesystem's fallocate may need to take i_rwsem.  We need to
 922          * explicitly grab a reference because the vma (and hence the
 923          * vma's reference to the file) can go away as soon as we drop
 924          * mmap_lock.
 925          */
 926         get_file(f);
 927         if (userfaultfd_remove(vma, start, end)) {
 928                 /* mmap_lock was not released by userfaultfd_remove() */
 929                 mmap_read_unlock(mm);
 930         }
 931         error = vfs_fallocate(f,
 932                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 933                                 offset, end - start);
 934         fput(f);
 935         mmap_read_lock(mm);
 936         return error;
 937 }
 938
 939 /*
 940  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
 941  * will handle splitting a vm area into separate areas, each area with its own
 942  * behavior.
 943  */
 944 static int madvise_vma_behavior(struct vm_area_struct *vma,
 945                                 struct vm_area_struct **prev,
 946                                 unsigned long start, unsigned long end,
 947                                 unsigned long behavior)
 948 {
 949         int error;
 950         struct anon_vma_name *anon_name;
 951         unsigned long new_flags = vma->vm_flags;
 952
 953         switch (behavior) {
 954         case MADV_REMOVE:
 955                 return madvise_remove(vma, prev, start, end);
 956         case MADV_WILLNEED:
 957                 return madvise_willneed(vma, prev, start, end);
 958         case MADV_COLD:
 959                 return madvise_cold(vma, prev, start, end);
 960         case MADV_PAGEOUT:
 961                 return madvise_pageout(vma, prev, start, end);
 962         case MADV_FREE:
 963         case MADV_DONTNEED:
 964                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 965         case MADV_POPULATE_READ:
 966         case MADV_POPULATE_WRITE:
 967                 return madvise_populate(vma, prev, start, end, behavior);
 968         case MADV_NORMAL:
 969                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
 970                 break;
 971         case MADV_SEQUENTIAL:
 972                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
 973                 break;
 974         case MADV_RANDOM:
 975                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
 976                 break;
 977         case MADV_DONTFORK:
 978                 new_flags |= VM_DONTCOPY;
 979                 break;
 980         case MADV_DOFORK:
 981                 if (vma->vm_flags & VM_IO)
 982                         return -EINVAL;
 983                 new_flags &= ~VM_DONTCOPY;
 984                 break;
 985         case MADV_WIPEONFORK:
 986                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
 987                 if (vma->vm_file || vma->vm_flags & VM_SHARED)
 988                         return -EINVAL;
 989                 new_flags |= VM_WIPEONFORK;
 990                 break;
 991         case MADV_KEEPONFORK:
 992                 new_flags &= ~VM_WIPEONFORK;
 993                 break;
 994         case MADV_DONTDUMP:
 995                 new_flags |= VM_DONTDUMP;
 996                 break;
 997         case MADV_DODUMP:
 998                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
 999                         return -EINVAL;
1000                 new_flags &= ~VM_DONTDUMP;
1001                 break;
1002         case MADV_MERGEABLE:
1003         case MADV_UNMERGEABLE:
1004                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1005                 if (error)
1006                         goto out;
1007                 break;
1008         case MADV_HUGEPAGE:
1009         case MADV_NOHUGEPAGE:
1010                 error = hugepage_madvise(vma, &new_flags, behavior);
1011                 if (error)
1012                         goto out;
1013                 break;
1014         }
1015
1016         anon_name = anon_vma_name(vma);
1017         anon_vma_name_get(anon_name);
1018         error = madvise_update_vma(vma, prev, start, end, new_flags,
1019                                    anon_name);
1020         anon_vma_name_put(anon_name);
1021
1022 out:
1023         /*
1024          * madvise() returns EAGAIN if kernel resources, such as
1025          * slab, are temporarily unavailable.
1026          */
1027         if (error == -ENOMEM)
1028                 error = -EAGAIN;
1029         return error;
1030 }
1031
1032 #ifdef CONFIG_MEMORY_FAILURE
1033 /*
1034  * Error injection support for memory error handling.
1035  */
1036 static int madvise_inject_error(int behavior,
1037                 unsigned long start, unsigned long end)
1038 {
1039         unsigned long size;
1040
1041         if (!capable(CAP_SYS_ADMIN))
1042                 return -EPERM;
1043
1044
1045         for (; start < end; start += size) {
1046                 unsigned long pfn;
1047                 struct page *page;
1048                 int ret;
1049
1050                 ret = get_user_pages_fast(start, 1, 0, &page);
1051                 if (ret != 1)
1052                         return ret;
1053                 pfn = page_to_pfn(page);
1054
1055                 /*
1056                  * When soft offlining hugepages, after migrating the page
1057                  * we dissolve it, therefore in the second loop "page" will
1058                  * no longer be a compound page.
1059                  */
1060                 size = page_size(compound_head(page));
1061
1062                 if (behavior == MADV_SOFT_OFFLINE) {
1063                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1064                                  pfn, start);
1065                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1066                 } else {
1067                         pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1068                                  pfn, start);
1069                         ret = memory_failure(pfn, MF_COUNT_INCREASED);
1070                         if (ret == -EOPNOTSUPP)
1071                                 ret = 0;
1072                 }
1073
1074                 if (ret)
1075                         return ret;
1076         }
1077
1078         return 0;
1079 }
1080 #endif
1081
1082 static bool
1083 madvise_behavior_valid(int behavior)
1084 {
1085         switch (behavior) {
1086         case MADV_DOFORK:
1087         case MADV_DONTFORK:
1088         case MADV_NORMAL:
1089         case MADV_SEQUENTIAL:
1090         case MADV_RANDOM:
1091         case MADV_REMOVE:
1092         case MADV_WILLNEED:
1093         case MADV_DONTNEED:
1094         case MADV_FREE:
1095         case MADV_COLD:
1096         case MADV_PAGEOUT:
1097         case MADV_POPULATE_READ:
1098         case MADV_POPULATE_WRITE:
1099 #ifdef CONFIG_KSM
1100         case MADV_MERGEABLE:
1101         case MADV_UNMERGEABLE:
1102 #endif
1103 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1104         case MADV_HUGEPAGE:
1105         case MADV_NOHUGEPAGE:
1106 #endif
1107         case MADV_DONTDUMP:
1108         case MADV_DODUMP:
1109         case MADV_WIPEONFORK:
1110         case MADV_KEEPONFORK:
1111 #ifdef CONFIG_MEMORY_FAILURE
1112         case MADV_SOFT_OFFLINE:
1113         case MADV_HWPOISON:
1114 #endif
1115                 return true;
1116
1117         default:
1118                 return false;
1119         }
1120 }
1121
1122 static bool
1123 process_madvise_behavior_valid(int behavior)
1124 {
1125         switch (behavior) {
1126         case MADV_COLD:
1127         case MADV_PAGEOUT:
1128         case MADV_WILLNEED:
1129                 return true;
1130         default:
1131                 return false;
1132         }
1133 }
1134
1135 /*
1136  * Walk the vmas in range [start,end), and call the visit function on each one.
1137  * The visit function will get start and end parameters that cover the overlap
1138  * between the current vma and the original range.  Any unmapped regions in the
1139  * original range will result in this function returning -ENOMEM while still
1140  * calling the visit function on all of the existing vmas in the range.
1141  * Must be called with the mmap_lock held for reading or writing.
1142  */
1143 static
1144 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1145                       unsigned long end, unsigned long arg,
1146                       int (*visit)(struct vm_area_struct *vma,
1147                                    struct vm_area_struct **prev, unsigned long start,
1148                                    unsigned long end, unsigned long arg))
1149 {
1150         struct vm_area_struct *vma;
1151         struct vm_area_struct *prev;
1152         unsigned long tmp;
1153         int unmapped_error = 0;
1154
1155         /*
1156          * If the interval [start,end) covers some unmapped address
1157          * ranges, just ignore them, but return -ENOMEM at the end.
1158          * - different from the way of handling in mlock etc.
1159          */
1160         vma = find_vma_prev(mm, start, &prev);
1161         if (vma && start > vma->vm_start)
1162                 prev = vma;
1163
1164         for (;;) {
1165                 int error;
1166
1167                 /* Still start < end. */
1168                 if (!vma)
1169                         return -ENOMEM;
1170
1171                 /* Here start < (end|vma->vm_end). */
1172                 if (start < vma->vm_start) {
1173                         unmapped_error = -ENOMEM;
1174                         start = vma->vm_start;
1175                         if (start >= end)
1176                                 break;
1177                 }
1178
1179                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1180                 tmp = vma->vm_end;
1181                 if (end < tmp)
1182                         tmp = end;
1183
1184                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1185                 error = visit(vma, &prev, start, tmp, arg);
1186                 if (error)
1187                         return error;
1188                 start = tmp;
1189                 if (prev && start < prev->vm_end)
1190                         start = prev->vm_end;
1191                 if (start >= end)
1192                         break;
1193                 if (prev)
1194                         vma = prev->vm_next;
1195                 else    /* madvise_remove dropped mmap_lock */
1196                         vma = find_vma(mm, start);
1197         }
1198
1199         return unmapped_error;
1200 }
1201
1202 #ifdef CONFIG_ANON_VMA_NAME
1203 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1204                                  struct vm_area_struct **prev,
1205                                  unsigned long start, unsigned long end,
1206                                  unsigned long anon_name)
1207 {
1208         int error;
1209
1210         /* Only anonymous mappings can be named */
1211         if (vma->vm_file)
1212                 return -EBADF;
1213
1214         error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1215                                    (struct anon_vma_name *)anon_name);
1216
1217         /*
1218          * madvise() returns EAGAIN if kernel resources, such as
1219          * slab, are temporarily unavailable.
1220          */
1221         if (error == -ENOMEM)
1222                 error = -EAGAIN;
1223         return error;
1224 }
1225
1226 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1227                           unsigned long len_in, struct anon_vma_name *anon_name)
1228 {
1229         unsigned long end;
1230         unsigned long len;
1231
1232         if (start & ~PAGE_MASK)
1233                 return -EINVAL;
1234         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1235
1236         /* Check to see whether len was rounded up from small -ve to zero */
1237         if (len_in && !len)
1238                 return -EINVAL;
1239
1240         end = start + len;
1241         if (end < start)
1242                 return -EINVAL;
1243
1244         if (end == start)
1245                 return 0;
1246
1247         return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1248                                  madvise_vma_anon_name);
1249 }
1250 #endif /* CONFIG_ANON_VMA_NAME */
1251 /*
1252  * The madvise(2) system call.
1253  *
1254  * Applications can use madvise() to advise the kernel how it should
1255  * handle paging I/O in this VM area.  The idea is to help the kernel
1256  * use appropriate read-ahead and caching techniques.  The information
1257  * provided is advisory only, and can be safely disregarded by the
1258  * kernel without affecting the correct operation of the application.
1259  *
1260  * behavior values:
1261  *  MADV_NORMAL - the default behavior is to read clusters.  This
1262  *              results in some read-ahead and read-behind.
1263  *  MADV_RANDOM - the system should read the minimum amount of data
1264  *              on any access, since it is unlikely that the appli-
1265  *              cation will need more than what it asks for.
1266  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1267  *              once, so they can be aggressively read ahead, and
1268  *              can be freed soon after they are accessed.
1269  *  MADV_WILLNEED - the application is notifying the system to read
1270  *              some pages ahead.
1271  *  MADV_DONTNEED - the application is finished with the given range,
1272  *              so the kernel can free resources associated with it.
1273  *  MADV_FREE - the application marks pages in the given range as lazy free,
1274  *              where actual purges are postponed until memory pressure happens.
1275  *  MADV_REMOVE - the application wants to free up the given range of
1276  *              pages and associated backing store.
1277  *  MADV_DONTFORK - omit this area from child's address space when forking:
1278  *              typically, to avoid COWing pages pinned by get_user_pages().
1279  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1280  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1281  *              range after a fork.
1282  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1283  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1284  *              were corrupted by unrecoverable hardware memory failure.
1285  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1286  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1287  *              this area with pages of identical content from other such areas.
1288  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1289  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1290  *              huge pages in the future. Existing pages might be coalesced and
1291  *              new pages might be allocated as THP.
1292  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1293  *              transparent huge pages so the existing pages will not be
1294  *              coalesced into THP and new pages will not be allocated as THP.
1295  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1296  *              from being included in its core dump.
1297  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1298  *  MADV_COLD - the application is not expected to use this memory soon,
1299  *              deactivate pages in this range so that they can be reclaimed
1300  *              easily if memory pressure happens.
1301  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1302  *              page out the pages in this range immediately.
1303  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1304  *              triggering read faults if required
1305  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1306  *              triggering write faults if required
1307  *
1308  * return values:
1309  *  zero    - success
1310  *  -EINVAL - start + len < 0, start is not page-aligned,
1311  *              "behavior" is not a valid value, or application
1312  *              is attempting to release locked or shared pages,
1313  *              or the specified address range includes file, Huge TLB,
1314  *              MAP_SHARED or VMPFNMAP range.
1315  *  -ENOMEM - addresses in the specified range are not currently
1316  *              mapped, or are outside the AS of the process.
1317  *  -EIO    - an I/O error occurred while paging in data.
1318  *  -EBADF  - map exists, but area maps something that isn't a file.
1319  *  -EAGAIN - a kernel resource was temporarily unavailable.
1320  */
1321 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1322 {
1323         unsigned long end;
1324         int error;
1325         int write;
1326         size_t len;
1327         struct blk_plug plug;
1328
1329         start = untagged_addr(start);
1330
1331         if (!madvise_behavior_valid(behavior))
1332                 return -EINVAL;
1333
1334         if (!PAGE_ALIGNED(start))
1335                 return -EINVAL;
1336         len = PAGE_ALIGN(len_in);
1337
1338         /* Check to see whether len was rounded up from small -ve to zero */
1339         if (len_in && !len)
1340                 return -EINVAL;
1341
1342         end = start + len;
1343         if (end < start)
1344                 return -EINVAL;
1345
1346         if (end == start)
1347                 return 0;
1348
1349 #ifdef CONFIG_MEMORY_FAILURE
1350         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1351                 return madvise_inject_error(behavior, start, start + len_in);
1352 #endif
1353
1354         write = madvise_need_mmap_write(behavior);
1355         if (write) {
1356                 if (mmap_write_lock_killable(mm))
1357                         return -EINTR;
1358         } else {
1359                 mmap_read_lock(mm);
1360         }
1361
1362         blk_start_plug(&plug);
1363         error = madvise_walk_vmas(mm, start, end, behavior,
1364                         madvise_vma_behavior);
1365         blk_finish_plug(&plug);
1366         if (write)
1367                 mmap_write_unlock(mm);
1368         else
1369                 mmap_read_unlock(mm);
1370
1371         return error;
1372 }
1373
1374 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1375 {
1376         return do_madvise(current->mm, start, len_in, behavior);
1377 }
1378
1379 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1380                 size_t, vlen, int, behavior, unsigned int, flags)
1381 {
1382         ssize_t ret;
1383         struct iovec iovstack[UIO_FASTIOV], iovec;
1384         struct iovec *iov = iovstack;
1385         struct iov_iter iter;
1386         struct task_struct *task;
1387         struct mm_struct *mm;
1388         size_t total_len;
1389         unsigned int f_flags;
1390
1391         if (flags != 0) {
1392                 ret = -EINVAL;
1393                 goto out;
1394         }
1395
1396         ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1397         if (ret < 0)
1398                 goto out;
1399
1400         task = pidfd_get_task(pidfd, &f_flags);
1401         if (IS_ERR(task)) {
1402                 ret = PTR_ERR(task);
1403                 goto free_iov;
1404         }
1405
1406         if (!process_madvise_behavior_valid(behavior)) {
1407                 ret = -EINVAL;
1408                 goto release_task;
1409         }
1410
1411         /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1412         mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1413         if (IS_ERR_OR_NULL(mm)) {
1414                 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1415                 goto release_task;
1416         }
1417
1418         /*
1419          * Require CAP_SYS_NICE for influencing process performance. Note that
1420          * only non-destructive hints are currently supported.
1421          */
1422         if (!capable(CAP_SYS_NICE)) {
1423                 ret = -EPERM;
1424                 goto release_mm;
1425         }
1426
1427         total_len = iov_iter_count(&iter);
1428
1429         while (iov_iter_count(&iter)) {
1430                 iovec = iov_iter_iovec(&iter);
1431                 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1432                                         iovec.iov_len, behavior);
1433                 if (ret < 0)
1434                         break;
1435                 iov_iter_advance(&iter, iovec.iov_len);
1436         }
1437
1438         if (ret == 0)
1439                 ret = total_len - iov_iter_count(&iter);
1440
1441 release_mm:
1442         mmput(mm);
1443 release_task:
1444         put_task_struct(task);
1445 free_iov:
1446         kfree(iov);
1447 out:
1448         return ret;
1449 }