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/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 #include <linux/sched/mm.h>
  31
  32 #include <asm/tlb.h>
  33
  34 #include "internal.h"
  35
  36 struct madvise_walk_private {
  37         struct mmu_gather *tlb;
  38         bool pageout;
  39 };
  40
  41 /*
  42  * Any behaviour which results in changes to the vma->vm_flags needs to
  43  * take mmap_lock for writing. Others, which simply traverse vmas, need
  44  * to only take it for reading.
  45  */
  46 static int madvise_need_mmap_write(int behavior)
  47 {
  48         switch (behavior) {
  49         case MADV_REMOVE:
  50         case MADV_WILLNEED:
  51         case MADV_DONTNEED:
  52         case MADV_COLD:
  53         case MADV_PAGEOUT:
  54         case MADV_FREE:
  55                 return 0;
  56         default:
  57                 /* be safe, default to 1. list exceptions explicitly */
  58                 return 1;
  59         }
  60 }
  61
  62 /*
  63  * We can potentially split a vm area into separate
  64  * areas, each area with its own behavior.
  65  */
  66 static long madvise_behavior(struct vm_area_struct *vma,
  67                      struct vm_area_struct **prev,
  68                      unsigned long start, unsigned long end, int behavior)
  69 {
  70         struct mm_struct *mm = vma->vm_mm;
  71         int error = 0;
  72         pgoff_t pgoff;
  73         unsigned long new_flags = vma->vm_flags;
  74
  75         switch (behavior) {
  76         case MADV_NORMAL:
  77                 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
  78                 break;
  79         case MADV_SEQUENTIAL:
  80                 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
  81                 break;
  82         case MADV_RANDOM:
  83                 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
  84                 break;
  85         case MADV_DONTFORK:
  86                 new_flags |= VM_DONTCOPY;
  87                 break;
  88         case MADV_DOFORK:
  89                 if (vma->vm_flags & VM_IO) {
  90                         error = -EINVAL;
  91                         goto out;
  92                 }
  93                 new_flags &= ~VM_DONTCOPY;
  94                 break;
  95         case MADV_WIPEONFORK:
  96                 /* MADV_WIPEONFORK is only supported on anonymous memory. */
  97                 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
  98                         error = -EINVAL;
  99                         goto out;
 100                 }
 101                 new_flags |= VM_WIPEONFORK;
 102                 break;
 103         case MADV_KEEPONFORK:
 104                 new_flags &= ~VM_WIPEONFORK;
 105                 break;
 106         case MADV_DONTDUMP:
 107                 new_flags |= VM_DONTDUMP;
 108                 break;
 109         case MADV_DODUMP:
 110                 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
 111                         error = -EINVAL;
 112                         goto out;
 113                 }
 114                 new_flags &= ~VM_DONTDUMP;
 115                 break;
 116         case MADV_MERGEABLE:
 117         case MADV_UNMERGEABLE:
 118                 error = ksm_madvise(vma, start, end, behavior, &new_flags);
 119                 if (error)
 120                         goto out_convert_errno;
 121                 break;
 122         case MADV_HUGEPAGE:
 123         case MADV_NOHUGEPAGE:
 124                 error = hugepage_madvise(vma, &new_flags, behavior);
 125                 if (error)
 126                         goto out_convert_errno;
 127                 break;
 128         }
 129
 130         if (new_flags == vma->vm_flags) {
 131                 *prev = vma;
 132                 goto out;
 133         }
 134
 135         pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 136         *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
 137                           vma->vm_file, pgoff, vma_policy(vma),
 138                           vma->vm_userfaultfd_ctx);
 139         if (*prev) {
 140                 vma = *prev;
 141                 goto success;
 142         }
 143
 144         *prev = vma;
 145
 146         if (start != vma->vm_start) {
 147                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 148                         error = -ENOMEM;
 149                         goto out;
 150                 }
 151                 error = __split_vma(mm, vma, start, 1);
 152                 if (error)
 153                         goto out_convert_errno;
 154         }
 155
 156         if (end != vma->vm_end) {
 157                 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
 158                         error = -ENOMEM;
 159                         goto out;
 160                 }
 161                 error = __split_vma(mm, vma, end, 0);
 162                 if (error)
 163                         goto out_convert_errno;
 164         }
 165
 166 success:
 167         /*
 168          * vm_flags is protected by the mmap_lock held in write mode.
 169          */
 170         vma->vm_flags = new_flags;
 171
 172 out_convert_errno:
 173         /*
 174          * madvise() returns EAGAIN if kernel resources, such as
 175          * slab, are temporarily unavailable.
 176          */
 177         if (error == -ENOMEM)
 178                 error = -EAGAIN;
 179 out:
 180         return error;
 181 }
 182
 183 #ifdef CONFIG_SWAP
 184 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
 185         unsigned long end, struct mm_walk *walk)
 186 {
 187         pte_t *orig_pte;
 188         struct vm_area_struct *vma = walk->private;
 189         unsigned long index;
 190
 191         if (pmd_none_or_trans_huge_or_clear_bad(pmd))
 192                 return 0;
 193
 194         for (index = start; index != end; index += PAGE_SIZE) {
 195                 pte_t pte;
 196                 swp_entry_t entry;
 197                 struct page *page;
 198                 spinlock_t *ptl;
 199
 200                 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
 201                 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
 202                 pte_unmap_unlock(orig_pte, ptl);
 203
 204                 if (pte_present(pte) || pte_none(pte))
 205                         continue;
 206                 entry = pte_to_swp_entry(pte);
 207                 if (unlikely(non_swap_entry(entry)))
 208                         continue;
 209
 210                 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
 211                                                         vma, index, false);
 212                 if (page)
 213                         put_page(page);
 214         }
 215
 216         return 0;
 217 }
 218
 219 static const struct mm_walk_ops swapin_walk_ops = {
 220         .pmd_entry              = swapin_walk_pmd_entry,
 221 };
 222
 223 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
 224                 unsigned long start, unsigned long end,
 225                 struct address_space *mapping)
 226 {
 227         pgoff_t index;
 228         struct page *page;
 229         swp_entry_t swap;
 230
 231         for (; start < end; start += PAGE_SIZE) {
 232                 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 233
 234                 page = find_get_entry(mapping, index);
 235                 if (!xa_is_value(page)) {
 236                         if (page)
 237                                 put_page(page);
 238                         continue;
 239                 }
 240                 swap = radix_to_swp_entry(page);
 241                 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
 242                                                         NULL, 0, false);
 243                 if (page)
 244                         put_page(page);
 245         }
 246
 247         lru_add_drain();        /* Push any new pages onto the LRU now */
 248 }
 249 #endif          /* CONFIG_SWAP */
 250
 251 /*
 252  * Schedule all required I/O operations.  Do not wait for completion.
 253  */
 254 static long madvise_willneed(struct vm_area_struct *vma,
 255                              struct vm_area_struct **prev,
 256                              unsigned long start, unsigned long end)
 257 {
 258         struct file *file = vma->vm_file;
 259         loff_t offset;
 260
 261         *prev = vma;
 262 #ifdef CONFIG_SWAP
 263         if (!file) {
 264                 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
 265                 lru_add_drain(); /* Push any new pages onto the LRU now */
 266                 return 0;
 267         }
 268
 269         if (shmem_mapping(file->f_mapping)) {
 270                 force_shm_swapin_readahead(vma, start, end,
 271                                         file->f_mapping);
 272                 return 0;
 273         }
 274 #else
 275         if (!file)
 276                 return -EBADF;
 277 #endif
 278
 279         if (IS_DAX(file_inode(file))) {
 280                 /* no bad return value, but ignore advice */
 281                 return 0;
 282         }
 283
 284         /*
 285          * Filesystem's fadvise may need to take various locks.  We need to
 286          * explicitly grab a reference because the vma (and hence the
 287          * vma's reference to the file) can go away as soon as we drop
 288          * mmap_lock.
 289          */
 290         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 291         get_file(file);
 292         mmap_read_unlock(current->mm);
 293         offset = (loff_t)(start - vma->vm_start)
 294                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 295         vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
 296         fput(file);
 297         mmap_read_lock(current->mm);
 298         return 0;
 299 }
 300
 301 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
 302                                 unsigned long addr, unsigned long end,
 303                                 struct mm_walk *walk)
 304 {
 305         struct madvise_walk_private *private = walk->private;
 306         struct mmu_gather *tlb = private->tlb;
 307         bool pageout = private->pageout;
 308         struct mm_struct *mm = tlb->mm;
 309         struct vm_area_struct *vma = walk->vma;
 310         pte_t *orig_pte, *pte, ptent;
 311         spinlock_t *ptl;
 312         struct page *page = NULL;
 313         LIST_HEAD(page_list);
 314
 315         if (fatal_signal_pending(current))
 316                 return -EINTR;
 317
 318 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 319         if (pmd_trans_huge(*pmd)) {
 320                 pmd_t orig_pmd;
 321                 unsigned long next = pmd_addr_end(addr, end);
 322
 323                 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
 324                 ptl = pmd_trans_huge_lock(pmd, vma);
 325                 if (!ptl)
 326                         return 0;
 327
 328                 orig_pmd = *pmd;
 329                 if (is_huge_zero_pmd(orig_pmd))
 330                         goto huge_unlock;
 331
 332                 if (unlikely(!pmd_present(orig_pmd))) {
 333                         VM_BUG_ON(thp_migration_supported() &&
 334                                         !is_pmd_migration_entry(orig_pmd));
 335                         goto huge_unlock;
 336                 }
 337
 338                 page = pmd_page(orig_pmd);
 339
 340                 /* Do not interfere with other mappings of this page */
 341                 if (page_mapcount(page) != 1)
 342                         goto huge_unlock;
 343
 344                 if (next - addr != HPAGE_PMD_SIZE) {
 345                         int err;
 346
 347                         get_page(page);
 348                         spin_unlock(ptl);
 349                         lock_page(page);
 350                         err = split_huge_page(page);
 351                         unlock_page(page);
 352                         put_page(page);
 353                         if (!err)
 354                                 goto regular_page;
 355                         return 0;
 356                 }
 357
 358                 if (pmd_young(orig_pmd)) {
 359                         pmdp_invalidate(vma, addr, pmd);
 360                         orig_pmd = pmd_mkold(orig_pmd);
 361
 362                         set_pmd_at(mm, addr, pmd, orig_pmd);
 363                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 364                 }
 365
 366                 ClearPageReferenced(page);
 367                 test_and_clear_page_young(page);
 368                 if (pageout) {
 369                         if (!isolate_lru_page(page)) {
 370                                 if (PageUnevictable(page))
 371                                         putback_lru_page(page);
 372                                 else
 373                                         list_add(&page->lru, &page_list);
 374                         }
 375                 } else
 376                         deactivate_page(page);
 377 huge_unlock:
 378                 spin_unlock(ptl);
 379                 if (pageout)
 380                         reclaim_pages(&page_list);
 381                 return 0;
 382         }
 383
 384         if (pmd_trans_unstable(pmd))
 385                 return 0;
 386 regular_page:
 387 #endif
 388         tlb_change_page_size(tlb, PAGE_SIZE);
 389         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 390         flush_tlb_batched_pending(mm);
 391         arch_enter_lazy_mmu_mode();
 392         for (; addr < end; pte++, addr += PAGE_SIZE) {
 393                 ptent = *pte;
 394
 395                 if (pte_none(ptent))
 396                         continue;
 397
 398                 if (!pte_present(ptent))
 399                         continue;
 400
 401                 page = vm_normal_page(vma, addr, ptent);
 402                 if (!page)
 403                         continue;
 404
 405                 /*
 406                  * Creating a THP page is expensive so split it only if we
 407                  * are sure it's worth. Split it if we are only owner.
 408                  */
 409                 if (PageTransCompound(page)) {
 410                         if (page_mapcount(page) != 1)
 411                                 break;
 412                         get_page(page);
 413                         if (!trylock_page(page)) {
 414                                 put_page(page);
 415                                 break;
 416                         }
 417                         pte_unmap_unlock(orig_pte, ptl);
 418                         if (split_huge_page(page)) {
 419                                 unlock_page(page);
 420                                 put_page(page);
 421                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 422                                 break;
 423                         }
 424                         unlock_page(page);
 425                         put_page(page);
 426                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 427                         pte--;
 428                         addr -= PAGE_SIZE;
 429                         continue;
 430                 }
 431
 432                 /* Do not interfere with other mappings of this page */
 433                 if (page_mapcount(page) != 1)
 434                         continue;
 435
 436                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 437
 438                 if (pte_young(ptent)) {
 439                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 440                                                         tlb->fullmm);
 441                         ptent = pte_mkold(ptent);
 442                         set_pte_at(mm, addr, pte, ptent);
 443                         tlb_remove_tlb_entry(tlb, pte, addr);
 444                 }
 445
 446                 /*
 447                  * We are deactivating a page for accelerating reclaiming.
 448                  * VM couldn't reclaim the page unless we clear PG_young.
 449                  * As a side effect, it makes confuse idle-page tracking
 450                  * because they will miss recent referenced history.
 451                  */
 452                 ClearPageReferenced(page);
 453                 test_and_clear_page_young(page);
 454                 if (pageout) {
 455                         if (!isolate_lru_page(page)) {
 456                                 if (PageUnevictable(page))
 457                                         putback_lru_page(page);
 458                                 else
 459                                         list_add(&page->lru, &page_list);
 460                         }
 461                 } else
 462                         deactivate_page(page);
 463         }
 464
 465         arch_leave_lazy_mmu_mode();
 466         pte_unmap_unlock(orig_pte, ptl);
 467         if (pageout)
 468                 reclaim_pages(&page_list);
 469         cond_resched();
 470
 471         return 0;
 472 }
 473
 474 static const struct mm_walk_ops cold_walk_ops = {
 475         .pmd_entry = madvise_cold_or_pageout_pte_range,
 476 };
 477
 478 static void madvise_cold_page_range(struct mmu_gather *tlb,
 479                              struct vm_area_struct *vma,
 480                              unsigned long addr, unsigned long end)
 481 {
 482         struct madvise_walk_private walk_private = {
 483                 .pageout = false,
 484                 .tlb = tlb,
 485         };
 486
 487         tlb_start_vma(tlb, vma);
 488         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 489         tlb_end_vma(tlb, vma);
 490 }
 491
 492 static long madvise_cold(struct vm_area_struct *vma,
 493                         struct vm_area_struct **prev,
 494                         unsigned long start_addr, unsigned long end_addr)
 495 {
 496         struct mm_struct *mm = vma->vm_mm;
 497         struct mmu_gather tlb;
 498
 499         *prev = vma;
 500         if (!can_madv_lru_vma(vma))
 501                 return -EINVAL;
 502
 503         lru_add_drain();
 504         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 505         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 506         tlb_finish_mmu(&tlb, start_addr, end_addr);
 507
 508         return 0;
 509 }
 510
 511 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 512                              struct vm_area_struct *vma,
 513                              unsigned long addr, unsigned long end)
 514 {
 515         struct madvise_walk_private walk_private = {
 516                 .pageout = true,
 517                 .tlb = tlb,
 518         };
 519
 520         tlb_start_vma(tlb, vma);
 521         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 522         tlb_end_vma(tlb, vma);
 523 }
 524
 525 static inline bool can_do_pageout(struct vm_area_struct *vma)
 526 {
 527         if (vma_is_anonymous(vma))
 528                 return true;
 529         if (!vma->vm_file)
 530                 return false;
 531         /*
 532          * paging out pagecache only for non-anonymous mappings that correspond
 533          * to the files the calling process could (if tried) open for writing;
 534          * otherwise we'd be including shared non-exclusive mappings, which
 535          * opens a side channel.
 536          */
 537         return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 538                 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 539 }
 540
 541 static long madvise_pageout(struct vm_area_struct *vma,
 542                         struct vm_area_struct **prev,
 543                         unsigned long start_addr, unsigned long end_addr)
 544 {
 545         struct mm_struct *mm = vma->vm_mm;
 546         struct mmu_gather tlb;
 547
 548         *prev = vma;
 549         if (!can_madv_lru_vma(vma))
 550                 return -EINVAL;
 551
 552         if (!can_do_pageout(vma))
 553                 return 0;
 554
 555         lru_add_drain();
 556         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 557         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 558         tlb_finish_mmu(&tlb, start_addr, end_addr);
 559
 560         return 0;
 561 }
 562
 563 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 564                                 unsigned long end, struct mm_walk *walk)
 565
 566 {
 567         struct mmu_gather *tlb = walk->private;
 568         struct mm_struct *mm = tlb->mm;
 569         struct vm_area_struct *vma = walk->vma;
 570         spinlock_t *ptl;
 571         pte_t *orig_pte, *pte, ptent;
 572         struct page *page;
 573         int nr_swap = 0;
 574         unsigned long next;
 575
 576         next = pmd_addr_end(addr, end);
 577         if (pmd_trans_huge(*pmd))
 578                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 579                         goto next;
 580
 581         if (pmd_trans_unstable(pmd))
 582                 return 0;
 583
 584         tlb_change_page_size(tlb, PAGE_SIZE);
 585         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 586         flush_tlb_batched_pending(mm);
 587         arch_enter_lazy_mmu_mode();
 588         for (; addr != end; pte++, addr += PAGE_SIZE) {
 589                 ptent = *pte;
 590
 591                 if (pte_none(ptent))
 592                         continue;
 593                 /*
 594                  * If the pte has swp_entry, just clear page table to
 595                  * prevent swap-in which is more expensive rather than
 596                  * (page allocation + zeroing).
 597                  */
 598                 if (!pte_present(ptent)) {
 599                         swp_entry_t entry;
 600
 601                         entry = pte_to_swp_entry(ptent);
 602                         if (non_swap_entry(entry))
 603                                 continue;
 604                         nr_swap--;
 605                         free_swap_and_cache(entry);
 606                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 607                         continue;
 608                 }
 609
 610                 page = vm_normal_page(vma, addr, ptent);
 611                 if (!page)
 612                         continue;
 613
 614                 /*
 615                  * If pmd isn't transhuge but the page is THP and
 616                  * is owned by only this process, split it and
 617                  * deactivate all pages.
 618                  */
 619                 if (PageTransCompound(page)) {
 620                         if (page_mapcount(page) != 1)
 621                                 goto out;
 622                         get_page(page);
 623                         if (!trylock_page(page)) {
 624                                 put_page(page);
 625                                 goto out;
 626                         }
 627                         pte_unmap_unlock(orig_pte, ptl);
 628                         if (split_huge_page(page)) {
 629                                 unlock_page(page);
 630                                 put_page(page);
 631                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 632                                 goto out;
 633                         }
 634                         unlock_page(page);
 635                         put_page(page);
 636                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 637                         pte--;
 638                         addr -= PAGE_SIZE;
 639                         continue;
 640                 }
 641
 642                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 643
 644                 if (PageSwapCache(page) || PageDirty(page)) {
 645                         if (!trylock_page(page))
 646                                 continue;
 647                         /*
 648                          * If page is shared with others, we couldn't clear
 649                          * PG_dirty of the page.
 650                          */
 651                         if (page_mapcount(page) != 1) {
 652                                 unlock_page(page);
 653                                 continue;
 654                         }
 655
 656                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 657                                 unlock_page(page);
 658                                 continue;
 659                         }
 660
 661                         ClearPageDirty(page);
 662                         unlock_page(page);
 663                 }
 664
 665                 if (pte_young(ptent) || pte_dirty(ptent)) {
 666                         /*
 667                          * Some of architecture(ex, PPC) don't update TLB
 668                          * with set_pte_at and tlb_remove_tlb_entry so for
 669                          * the portability, remap the pte with old|clean
 670                          * after pte clearing.
 671                          */
 672                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 673                                                         tlb->fullmm);
 674
 675                         ptent = pte_mkold(ptent);
 676                         ptent = pte_mkclean(ptent);
 677                         set_pte_at(mm, addr, pte, ptent);
 678                         tlb_remove_tlb_entry(tlb, pte, addr);
 679                 }
 680                 mark_page_lazyfree(page);
 681         }
 682 out:
 683         if (nr_swap) {
 684                 if (current->mm == mm)
 685                         sync_mm_rss(mm);
 686
 687                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 688         }
 689         arch_leave_lazy_mmu_mode();
 690         pte_unmap_unlock(orig_pte, ptl);
 691         cond_resched();
 692 next:
 693         return 0;
 694 }
 695
 696 static const struct mm_walk_ops madvise_free_walk_ops = {
 697         .pmd_entry              = madvise_free_pte_range,
 698 };
 699
 700 static int madvise_free_single_vma(struct vm_area_struct *vma,
 701                         unsigned long start_addr, unsigned long end_addr)
 702 {
 703         struct mm_struct *mm = vma->vm_mm;
 704         struct mmu_notifier_range range;
 705         struct mmu_gather tlb;
 706
 707         /* MADV_FREE works for only anon vma at the moment */
 708         if (!vma_is_anonymous(vma))
 709                 return -EINVAL;
 710
 711         range.start = max(vma->vm_start, start_addr);
 712         if (range.start >= vma->vm_end)
 713                 return -EINVAL;
 714         range.end = min(vma->vm_end, end_addr);
 715         if (range.end <= vma->vm_start)
 716                 return -EINVAL;
 717         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 718                                 range.start, range.end);
 719
 720         lru_add_drain();
 721         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 722         update_hiwater_rss(mm);
 723
 724         mmu_notifier_invalidate_range_start(&range);
 725         tlb_start_vma(&tlb, vma);
 726         walk_page_range(vma->vm_mm, range.start, range.end,
 727                         &madvise_free_walk_ops, &tlb);
 728         tlb_end_vma(&tlb, vma);
 729         mmu_notifier_invalidate_range_end(&range);
 730         tlb_finish_mmu(&tlb, range.start, range.end);
 731
 732         return 0;
 733 }
 734
 735 /*
 736  * Application no longer needs these pages.  If the pages are dirty,
 737  * it's OK to just throw them away.  The app will be more careful about
 738  * data it wants to keep.  Be sure to free swap resources too.  The
 739  * zap_page_range call sets things up for shrink_active_list to actually free
 740  * these pages later if no one else has touched them in the meantime,
 741  * although we could add these pages to a global reuse list for
 742  * shrink_active_list to pick up before reclaiming other pages.
 743  *
 744  * NB: This interface discards data rather than pushes it out to swap,
 745  * as some implementations do.  This has performance implications for
 746  * applications like large transactional databases which want to discard
 747  * pages in anonymous maps after committing to backing store the data
 748  * that was kept in them.  There is no reason to write this data out to
 749  * the swap area if the application is discarding it.
 750  *
 751  * An interface that causes the system to free clean pages and flush
 752  * dirty pages is already available as msync(MS_INVALIDATE).
 753  */
 754 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 755                                         unsigned long start, unsigned long end)
 756 {
 757         zap_page_range(vma, start, end - start);
 758         return 0;
 759 }
 760
 761 static long madvise_dontneed_free(struct vm_area_struct *vma,
 762                                   struct vm_area_struct **prev,
 763                                   unsigned long start, unsigned long end,
 764                                   int behavior)
 765 {
 766         *prev = vma;
 767         if (!can_madv_lru_vma(vma))
 768                 return -EINVAL;
 769
 770         if (!userfaultfd_remove(vma, start, end)) {
 771                 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
 772
 773                 mmap_read_lock(current->mm);
 774                 vma = find_vma(current->mm, start);
 775                 if (!vma)
 776                         return -ENOMEM;
 777                 if (start < vma->vm_start) {
 778                         /*
 779                          * This "vma" under revalidation is the one
 780                          * with the lowest vma->vm_start where start
 781                          * is also < vma->vm_end. If start <
 782                          * vma->vm_start it means an hole materialized
 783                          * in the user address space within the
 784                          * virtual range passed to MADV_DONTNEED
 785                          * or MADV_FREE.
 786                          */
 787                         return -ENOMEM;
 788                 }
 789                 if (!can_madv_lru_vma(vma))
 790                         return -EINVAL;
 791                 if (end > vma->vm_end) {
 792                         /*
 793                          * Don't fail if end > vma->vm_end. If the old
 794                          * vma was splitted while the mmap_lock was
 795                          * released the effect of the concurrent
 796                          * operation may not cause madvise() to
 797                          * have an undefined result. There may be an
 798                          * adjacent next vma that we'll walk
 799                          * next. userfaultfd_remove() will generate an
 800                          * UFFD_EVENT_REMOVE repetition on the
 801                          * end-vma->vm_end range, but the manager can
 802                          * handle a repetition fine.
 803                          */
 804                         end = vma->vm_end;
 805                 }
 806                 VM_WARN_ON(start >= end);
 807         }
 808
 809         if (behavior == MADV_DONTNEED)
 810                 return madvise_dontneed_single_vma(vma, start, end);
 811         else if (behavior == MADV_FREE)
 812                 return madvise_free_single_vma(vma, start, end);
 813         else
 814                 return -EINVAL;
 815 }
 816
 817 /*
 818  * Application wants to free up the pages and associated backing store.
 819  * This is effectively punching a hole into the middle of a file.
 820  */
 821 static long madvise_remove(struct vm_area_struct *vma,
 822                                 struct vm_area_struct **prev,
 823                                 unsigned long start, unsigned long end)
 824 {
 825         loff_t offset;
 826         int error;
 827         struct file *f;
 828
 829         *prev = NULL;   /* tell sys_madvise we drop mmap_lock */
 830
 831         if (vma->vm_flags & VM_LOCKED)
 832                 return -EINVAL;
 833
 834         f = vma->vm_file;
 835
 836         if (!f || !f->f_mapping || !f->f_mapping->host) {
 837                         return -EINVAL;
 838         }
 839
 840         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 841                 return -EACCES;
 842
 843         offset = (loff_t)(start - vma->vm_start)
 844                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 845
 846         /*
 847          * Filesystem's fallocate may need to take i_mutex.  We need to
 848          * explicitly grab a reference because the vma (and hence the
 849          * vma's reference to the file) can go away as soon as we drop
 850          * mmap_lock.
 851          */
 852         get_file(f);
 853         if (userfaultfd_remove(vma, start, end)) {
 854                 /* mmap_lock was not released by userfaultfd_remove() */
 855                 mmap_read_unlock(current->mm);
 856         }
 857         error = vfs_fallocate(f,
 858                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 859                                 offset, end - start);
 860         fput(f);
 861         mmap_read_lock(current->mm);
 862         return error;
 863 }
 864
 865 #ifdef CONFIG_MEMORY_FAILURE
 866 /*
 867  * Error injection support for memory error handling.
 868  */
 869 static int madvise_inject_error(int behavior,
 870                 unsigned long start, unsigned long end)
 871 {
 872         struct page *page;
 873         struct zone *zone;
 874         unsigned long size;
 875
 876         if (!capable(CAP_SYS_ADMIN))
 877                 return -EPERM;
 878
 879
 880         for (; start < end; start += size) {
 881                 unsigned long pfn;
 882                 int ret;
 883
 884                 ret = get_user_pages_fast(start, 1, 0, &page);
 885                 if (ret != 1)
 886                         return ret;
 887                 pfn = page_to_pfn(page);
 888
 889                 /*
 890                  * When soft offlining hugepages, after migrating the page
 891                  * we dissolve it, therefore in the second loop "page" will
 892                  * no longer be a compound page.
 893                  */
 894                 size = page_size(compound_head(page));
 895
 896                 if (PageHWPoison(page)) {
 897                         put_page(page);
 898                         continue;
 899                 }
 900
 901                 if (behavior == MADV_SOFT_OFFLINE) {
 902                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 903                                         pfn, start);
 904
 905                         ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
 906                         if (ret)
 907                                 return ret;
 908                         continue;
 909                 }
 910
 911                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 912                                 pfn, start);
 913
 914                 /*
 915                  * Drop the page reference taken by get_user_pages_fast(). In
 916                  * the absence of MF_COUNT_INCREASED the memory_failure()
 917                  * routine is responsible for pinning the page to prevent it
 918                  * from being released back to the page allocator.
 919                  */
 920                 put_page(page);
 921                 ret = memory_failure(pfn, 0);
 922                 if (ret)
 923                         return ret;
 924         }
 925
 926         /* Ensure that all poisoned pages are removed from per-cpu lists */
 927         for_each_populated_zone(zone)
 928                 drain_all_pages(zone);
 929
 930         return 0;
 931 }
 932 #endif
 933
 934 static long
 935 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 936                 unsigned long start, unsigned long end, int behavior)
 937 {
 938         switch (behavior) {
 939         case MADV_REMOVE:
 940                 return madvise_remove(vma, prev, start, end);
 941         case MADV_WILLNEED:
 942                 return madvise_willneed(vma, prev, start, end);
 943         case MADV_COLD:
 944                 return madvise_cold(vma, prev, start, end);
 945         case MADV_PAGEOUT:
 946                 return madvise_pageout(vma, prev, start, end);
 947         case MADV_FREE:
 948         case MADV_DONTNEED:
 949                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 950         default:
 951                 return madvise_behavior(vma, prev, start, end, behavior);
 952         }
 953 }
 954
 955 static bool
 956 madvise_behavior_valid(int behavior)
 957 {
 958         switch (behavior) {
 959         case MADV_DOFORK:
 960         case MADV_DONTFORK:
 961         case MADV_NORMAL:
 962         case MADV_SEQUENTIAL:
 963         case MADV_RANDOM:
 964         case MADV_REMOVE:
 965         case MADV_WILLNEED:
 966         case MADV_DONTNEED:
 967         case MADV_FREE:
 968         case MADV_COLD:
 969         case MADV_PAGEOUT:
 970 #ifdef CONFIG_KSM
 971         case MADV_MERGEABLE:
 972         case MADV_UNMERGEABLE:
 973 #endif
 974 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 975         case MADV_HUGEPAGE:
 976         case MADV_NOHUGEPAGE:
 977 #endif
 978         case MADV_DONTDUMP:
 979         case MADV_DODUMP:
 980         case MADV_WIPEONFORK:
 981         case MADV_KEEPONFORK:
 982 #ifdef CONFIG_MEMORY_FAILURE
 983         case MADV_SOFT_OFFLINE:
 984         case MADV_HWPOISON:
 985 #endif
 986                 return true;
 987
 988         default:
 989                 return false;
 990         }
 991 }
 992
 993 /*
 994  * The madvise(2) system call.
 995  *
 996  * Applications can use madvise() to advise the kernel how it should
 997  * handle paging I/O in this VM area.  The idea is to help the kernel
 998  * use appropriate read-ahead and caching techniques.  The information
 999  * provided is advisory only, and can be safely disregarded by the
1000  * kernel without affecting the correct operation of the application.
1001  *
1002  * behavior values:
1003  *  MADV_NORMAL - the default behavior is to read clusters.  This
1004  *              results in some read-ahead and read-behind.
1005  *  MADV_RANDOM - the system should read the minimum amount of data
1006  *              on any access, since it is unlikely that the appli-
1007  *              cation will need more than what it asks for.
1008  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1009  *              once, so they can be aggressively read ahead, and
1010  *              can be freed soon after they are accessed.
1011  *  MADV_WILLNEED - the application is notifying the system to read
1012  *              some pages ahead.
1013  *  MADV_DONTNEED - the application is finished with the given range,
1014  *              so the kernel can free resources associated with it.
1015  *  MADV_FREE - the application marks pages in the given range as lazy free,
1016  *              where actual purges are postponed until memory pressure happens.
1017  *  MADV_REMOVE - the application wants to free up the given range of
1018  *              pages and associated backing store.
1019  *  MADV_DONTFORK - omit this area from child's address space when forking:
1020  *              typically, to avoid COWing pages pinned by get_user_pages().
1021  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1022  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1023  *              range after a fork.
1024  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1025  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1026  *              were corrupted by unrecoverable hardware memory failure.
1027  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1028  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1029  *              this area with pages of identical content from other such areas.
1030  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1031  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1032  *              huge pages in the future. Existing pages might be coalesced and
1033  *              new pages might be allocated as THP.
1034  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1035  *              transparent huge pages so the existing pages will not be
1036  *              coalesced into THP and new pages will not be allocated as THP.
1037  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1038  *              from being included in its core dump.
1039  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1040  *
1041  * return values:
1042  *  zero    - success
1043  *  -EINVAL - start + len < 0, start is not page-aligned,
1044  *              "behavior" is not a valid value, or application
1045  *              is attempting to release locked or shared pages,
1046  *              or the specified address range includes file, Huge TLB,
1047  *              MAP_SHARED or VMPFNMAP range.
1048  *  -ENOMEM - addresses in the specified range are not currently
1049  *              mapped, or are outside the AS of the process.
1050  *  -EIO    - an I/O error occurred while paging in data.
1051  *  -EBADF  - map exists, but area maps something that isn't a file.
1052  *  -EAGAIN - a kernel resource was temporarily unavailable.
1053  */
1054 int do_madvise(unsigned long start, size_t len_in, int behavior)
1055 {
1056         unsigned long end, tmp;
1057         struct vm_area_struct *vma, *prev;
1058         int unmapped_error = 0;
1059         int error = -EINVAL;
1060         int write;
1061         size_t len;
1062         struct blk_plug plug;
1063
1064         start = untagged_addr(start);
1065
1066         if (!madvise_behavior_valid(behavior))
1067                 return error;
1068
1069         if (!PAGE_ALIGNED(start))
1070                 return error;
1071         len = PAGE_ALIGN(len_in);
1072
1073         /* Check to see whether len was rounded up from small -ve to zero */
1074         if (len_in && !len)
1075                 return error;
1076
1077         end = start + len;
1078         if (end < start)
1079                 return error;
1080
1081         error = 0;
1082         if (end == start)
1083                 return error;
1084
1085 #ifdef CONFIG_MEMORY_FAILURE
1086         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1087                 return madvise_inject_error(behavior, start, start + len_in);
1088 #endif
1089
1090         write = madvise_need_mmap_write(behavior);
1091         if (write) {
1092                 if (mmap_write_lock_killable(current->mm))
1093                         return -EINTR;
1094
1095                 /*
1096                  * We may have stolen the mm from another process
1097                  * that is undergoing core dumping.
1098                  *
1099                  * Right now that's io_ring, in the future it may
1100                  * be remote process management and not "current"
1101                  * at all.
1102                  *
1103                  * We need to fix core dumping to not do this,
1104                  * but for now we have the mmget_still_valid()
1105                  * model.
1106                  */
1107                 if (!mmget_still_valid(current->mm)) {
1108                         mmap_write_unlock(current->mm);
1109                         return -EINTR;
1110                 }
1111         } else {
1112                 mmap_read_lock(current->mm);
1113         }
1114
1115         /*
1116          * If the interval [start,end) covers some unmapped address
1117          * ranges, just ignore them, but return -ENOMEM at the end.
1118          * - different from the way of handling in mlock etc.
1119          */
1120         vma = find_vma_prev(current->mm, start, &prev);
1121         if (vma && start > vma->vm_start)
1122                 prev = vma;
1123
1124         blk_start_plug(&plug);
1125         for (;;) {
1126                 /* Still start < end. */
1127                 error = -ENOMEM;
1128                 if (!vma)
1129                         goto out;
1130
1131                 /* Here start < (end|vma->vm_end). */
1132                 if (start < vma->vm_start) {
1133                         unmapped_error = -ENOMEM;
1134                         start = vma->vm_start;
1135                         if (start >= end)
1136                                 goto out;
1137                 }
1138
1139                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1140                 tmp = vma->vm_end;
1141                 if (end < tmp)
1142                         tmp = end;
1143
1144                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1145                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1146                 if (error)
1147                         goto out;
1148                 start = tmp;
1149                 if (prev && start < prev->vm_end)
1150                         start = prev->vm_end;
1151                 error = unmapped_error;
1152                 if (start >= end)
1153                         goto out;
1154                 if (prev)
1155                         vma = prev->vm_next;
1156                 else    /* madvise_remove dropped mmap_lock */
1157                         vma = find_vma(current->mm, start);
1158         }
1159 out:
1160         blk_finish_plug(&plug);
1161         if (write)
1162                 mmap_write_unlock(current->mm);
1163         else
1164                 mmap_read_unlock(current->mm);
1165
1166         return error;
1167 }
1168
1169 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1170 {
1171         return do_madvise(start, len_in, behavior);
1172 }