root/mm/madvise.c

DEFINITIONS

1. madvise_need_mmap_write
2. madvise_behavior
3. swapin_walk_pmd_entry
4. force_shm_swapin_readahead
5. madvise_willneed
6. madvise_cold_or_pageout_pte_range
7. madvise_cold_page_range
8. madvise_cold
9. madvise_pageout_page_range
10. can_do_pageout
11. madvise_pageout
12. madvise_free_pte_range
13. madvise_free_single_vma
14. madvise_dontneed_single_vma
15. madvise_dontneed_free
16. madvise_remove
17. madvise_inject_error
18. madvise_vma
19. madvise_behavior_valid
20. SYSCALL_DEFINE3

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