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
  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                 if (next - addr != HPAGE_PMD_SIZE) {
 339                         int err;
 340
 341                         if (page_mapcount(page) != 1)
 342                                 goto huge_unlock;
 343
 344                         get_page(page);
 345                         spin_unlock(ptl);
 346                         lock_page(page);
 347                         err = split_huge_page(page);
 348                         unlock_page(page);
 349                         put_page(page);
 350                         if (!err)
 351                                 goto regular_page;
 352                         return 0;
 353                 }
 354
 355                 if (pmd_young(orig_pmd)) {
 356                         pmdp_invalidate(vma, addr, pmd);
 357                         orig_pmd = pmd_mkold(orig_pmd);
 358
 359                         set_pmd_at(mm, addr, pmd, orig_pmd);
 360                         tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
 361                 }
 362
 363                 ClearPageReferenced(page);
 364                 test_and_clear_page_young(page);
 365                 if (pageout) {
 366                         if (!isolate_lru_page(page)) {
 367                                 if (PageUnevictable(page))
 368                                         putback_lru_page(page);
 369                                 else
 370                                         list_add(&page->lru, &page_list);
 371                         }
 372                 } else
 373                         deactivate_page(page);
 374 huge_unlock:
 375                 spin_unlock(ptl);
 376                 if (pageout)
 377                         reclaim_pages(&page_list);
 378                 return 0;
 379         }
 380
 381         if (pmd_trans_unstable(pmd))
 382                 return 0;
 383 regular_page:
 384 #endif
 385         tlb_change_page_size(tlb, PAGE_SIZE);
 386         orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
 387         flush_tlb_batched_pending(mm);
 388         arch_enter_lazy_mmu_mode();
 389         for (; addr < end; pte++, addr += PAGE_SIZE) {
 390                 ptent = *pte;
 391
 392                 if (pte_none(ptent))
 393                         continue;
 394
 395                 if (!pte_present(ptent))
 396                         continue;
 397
 398                 page = vm_normal_page(vma, addr, ptent);
 399                 if (!page)
 400                         continue;
 401
 402                 /*
 403                  * Creating a THP page is expensive so split it only if we
 404                  * are sure it's worth. Split it if we are only owner.
 405                  */
 406                 if (PageTransCompound(page)) {
 407                         if (page_mapcount(page) != 1)
 408                                 break;
 409                         get_page(page);
 410                         if (!trylock_page(page)) {
 411                                 put_page(page);
 412                                 break;
 413                         }
 414                         pte_unmap_unlock(orig_pte, ptl);
 415                         if (split_huge_page(page)) {
 416                                 unlock_page(page);
 417                                 put_page(page);
 418                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 419                                 break;
 420                         }
 421                         unlock_page(page);
 422                         put_page(page);
 423                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 424                         pte--;
 425                         addr -= PAGE_SIZE;
 426                         continue;
 427                 }
 428
 429                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 430
 431                 if (pte_young(ptent)) {
 432                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 433                                                         tlb->fullmm);
 434                         ptent = pte_mkold(ptent);
 435                         set_pte_at(mm, addr, pte, ptent);
 436                         tlb_remove_tlb_entry(tlb, pte, addr);
 437                 }
 438
 439                 /*
 440                  * We are deactivating a page for accelerating reclaiming.
 441                  * VM couldn't reclaim the page unless we clear PG_young.
 442                  * As a side effect, it makes confuse idle-page tracking
 443                  * because they will miss recent referenced history.
 444                  */
 445                 ClearPageReferenced(page);
 446                 test_and_clear_page_young(page);
 447                 if (pageout) {
 448                         if (!isolate_lru_page(page)) {
 449                                 if (PageUnevictable(page))
 450                                         putback_lru_page(page);
 451                                 else
 452                                         list_add(&page->lru, &page_list);
 453                         }
 454                 } else
 455                         deactivate_page(page);
 456         }
 457
 458         arch_leave_lazy_mmu_mode();
 459         pte_unmap_unlock(orig_pte, ptl);
 460         if (pageout)
 461                 reclaim_pages(&page_list);
 462         cond_resched();
 463
 464         return 0;
 465 }
 466
 467 static const struct mm_walk_ops cold_walk_ops = {
 468         .pmd_entry = madvise_cold_or_pageout_pte_range,
 469 };
 470
 471 static void madvise_cold_page_range(struct mmu_gather *tlb,
 472                              struct vm_area_struct *vma,
 473                              unsigned long addr, unsigned long end)
 474 {
 475         struct madvise_walk_private walk_private = {
 476                 .pageout = false,
 477                 .tlb = tlb,
 478         };
 479
 480         tlb_start_vma(tlb, vma);
 481         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 482         tlb_end_vma(tlb, vma);
 483 }
 484
 485 static long madvise_cold(struct vm_area_struct *vma,
 486                         struct vm_area_struct **prev,
 487                         unsigned long start_addr, unsigned long end_addr)
 488 {
 489         struct mm_struct *mm = vma->vm_mm;
 490         struct mmu_gather tlb;
 491
 492         *prev = vma;
 493         if (!can_madv_lru_vma(vma))
 494                 return -EINVAL;
 495
 496         lru_add_drain();
 497         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 498         madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
 499         tlb_finish_mmu(&tlb, start_addr, end_addr);
 500
 501         return 0;
 502 }
 503
 504 static void madvise_pageout_page_range(struct mmu_gather *tlb,
 505                              struct vm_area_struct *vma,
 506                              unsigned long addr, unsigned long end)
 507 {
 508         struct madvise_walk_private walk_private = {
 509                 .pageout = true,
 510                 .tlb = tlb,
 511         };
 512
 513         tlb_start_vma(tlb, vma);
 514         walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
 515         tlb_end_vma(tlb, vma);
 516 }
 517
 518 static inline bool can_do_pageout(struct vm_area_struct *vma)
 519 {
 520         if (vma_is_anonymous(vma))
 521                 return true;
 522         if (!vma->vm_file)
 523                 return false;
 524         /*
 525          * paging out pagecache only for non-anonymous mappings that correspond
 526          * to the files the calling process could (if tried) open for writing;
 527          * otherwise we'd be including shared non-exclusive mappings, which
 528          * opens a side channel.
 529          */
 530         return inode_owner_or_capable(file_inode(vma->vm_file)) ||
 531                 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
 532 }
 533
 534 static long madvise_pageout(struct vm_area_struct *vma,
 535                         struct vm_area_struct **prev,
 536                         unsigned long start_addr, unsigned long end_addr)
 537 {
 538         struct mm_struct *mm = vma->vm_mm;
 539         struct mmu_gather tlb;
 540
 541         *prev = vma;
 542         if (!can_madv_lru_vma(vma))
 543                 return -EINVAL;
 544
 545         if (!can_do_pageout(vma))
 546                 return 0;
 547
 548         lru_add_drain();
 549         tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
 550         madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
 551         tlb_finish_mmu(&tlb, start_addr, end_addr);
 552
 553         return 0;
 554 }
 555
 556 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
 557                                 unsigned long end, struct mm_walk *walk)
 558
 559 {
 560         struct mmu_gather *tlb = walk->private;
 561         struct mm_struct *mm = tlb->mm;
 562         struct vm_area_struct *vma = walk->vma;
 563         spinlock_t *ptl;
 564         pte_t *orig_pte, *pte, ptent;
 565         struct page *page;
 566         int nr_swap = 0;
 567         unsigned long next;
 568
 569         next = pmd_addr_end(addr, end);
 570         if (pmd_trans_huge(*pmd))
 571                 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
 572                         goto next;
 573
 574         if (pmd_trans_unstable(pmd))
 575                 return 0;
 576
 577         tlb_change_page_size(tlb, PAGE_SIZE);
 578         orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 579         flush_tlb_batched_pending(mm);
 580         arch_enter_lazy_mmu_mode();
 581         for (; addr != end; pte++, addr += PAGE_SIZE) {
 582                 ptent = *pte;
 583
 584                 if (pte_none(ptent))
 585                         continue;
 586                 /*
 587                  * If the pte has swp_entry, just clear page table to
 588                  * prevent swap-in which is more expensive rather than
 589                  * (page allocation + zeroing).
 590                  */
 591                 if (!pte_present(ptent)) {
 592                         swp_entry_t entry;
 593
 594                         entry = pte_to_swp_entry(ptent);
 595                         if (non_swap_entry(entry))
 596                                 continue;
 597                         nr_swap--;
 598                         free_swap_and_cache(entry);
 599                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
 600                         continue;
 601                 }
 602
 603                 page = vm_normal_page(vma, addr, ptent);
 604                 if (!page)
 605                         continue;
 606
 607                 /*
 608                  * If pmd isn't transhuge but the page is THP and
 609                  * is owned by only this process, split it and
 610                  * deactivate all pages.
 611                  */
 612                 if (PageTransCompound(page)) {
 613                         if (page_mapcount(page) != 1)
 614                                 goto out;
 615                         get_page(page);
 616                         if (!trylock_page(page)) {
 617                                 put_page(page);
 618                                 goto out;
 619                         }
 620                         pte_unmap_unlock(orig_pte, ptl);
 621                         if (split_huge_page(page)) {
 622                                 unlock_page(page);
 623                                 put_page(page);
 624                                 pte_offset_map_lock(mm, pmd, addr, &ptl);
 625                                 goto out;
 626                         }
 627                         unlock_page(page);
 628                         put_page(page);
 629                         pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
 630                         pte--;
 631                         addr -= PAGE_SIZE;
 632                         continue;
 633                 }
 634
 635                 VM_BUG_ON_PAGE(PageTransCompound(page), page);
 636
 637                 if (PageSwapCache(page) || PageDirty(page)) {
 638                         if (!trylock_page(page))
 639                                 continue;
 640                         /*
 641                          * If page is shared with others, we couldn't clear
 642                          * PG_dirty of the page.
 643                          */
 644                         if (page_mapcount(page) != 1) {
 645                                 unlock_page(page);
 646                                 continue;
 647                         }
 648
 649                         if (PageSwapCache(page) && !try_to_free_swap(page)) {
 650                                 unlock_page(page);
 651                                 continue;
 652                         }
 653
 654                         ClearPageDirty(page);
 655                         unlock_page(page);
 656                 }
 657
 658                 if (pte_young(ptent) || pte_dirty(ptent)) {
 659                         /*
 660                          * Some of architecture(ex, PPC) don't update TLB
 661                          * with set_pte_at and tlb_remove_tlb_entry so for
 662                          * the portability, remap the pte with old|clean
 663                          * after pte clearing.
 664                          */
 665                         ptent = ptep_get_and_clear_full(mm, addr, pte,
 666                                                         tlb->fullmm);
 667
 668                         ptent = pte_mkold(ptent);
 669                         ptent = pte_mkclean(ptent);
 670                         set_pte_at(mm, addr, pte, ptent);
 671                         tlb_remove_tlb_entry(tlb, pte, addr);
 672                 }
 673                 mark_page_lazyfree(page);
 674         }
 675 out:
 676         if (nr_swap) {
 677                 if (current->mm == mm)
 678                         sync_mm_rss(mm);
 679
 680                 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
 681         }
 682         arch_leave_lazy_mmu_mode();
 683         pte_unmap_unlock(orig_pte, ptl);
 684         cond_resched();
 685 next:
 686         return 0;
 687 }
 688
 689 static const struct mm_walk_ops madvise_free_walk_ops = {
 690         .pmd_entry              = madvise_free_pte_range,
 691 };
 692
 693 static int madvise_free_single_vma(struct vm_area_struct *vma,
 694                         unsigned long start_addr, unsigned long end_addr)
 695 {
 696         struct mm_struct *mm = vma->vm_mm;
 697         struct mmu_notifier_range range;
 698         struct mmu_gather tlb;
 699
 700         /* MADV_FREE works for only anon vma at the moment */
 701         if (!vma_is_anonymous(vma))
 702                 return -EINVAL;
 703
 704         range.start = max(vma->vm_start, start_addr);
 705         if (range.start >= vma->vm_end)
 706                 return -EINVAL;
 707         range.end = min(vma->vm_end, end_addr);
 708         if (range.end <= vma->vm_start)
 709                 return -EINVAL;
 710         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
 711                                 range.start, range.end);
 712
 713         lru_add_drain();
 714         tlb_gather_mmu(&tlb, mm, range.start, range.end);
 715         update_hiwater_rss(mm);
 716
 717         mmu_notifier_invalidate_range_start(&range);
 718         tlb_start_vma(&tlb, vma);
 719         walk_page_range(vma->vm_mm, range.start, range.end,
 720                         &madvise_free_walk_ops, &tlb);
 721         tlb_end_vma(&tlb, vma);
 722         mmu_notifier_invalidate_range_end(&range);
 723         tlb_finish_mmu(&tlb, range.start, range.end);
 724
 725         return 0;
 726 }
 727
 728 /*
 729  * Application no longer needs these pages.  If the pages are dirty,
 730  * it's OK to just throw them away.  The app will be more careful about
 731  * data it wants to keep.  Be sure to free swap resources too.  The
 732  * zap_page_range call sets things up for shrink_active_list to actually free
 733  * these pages later if no one else has touched them in the meantime,
 734  * although we could add these pages to a global reuse list for
 735  * shrink_active_list to pick up before reclaiming other pages.
 736  *
 737  * NB: This interface discards data rather than pushes it out to swap,
 738  * as some implementations do.  This has performance implications for
 739  * applications like large transactional databases which want to discard
 740  * pages in anonymous maps after committing to backing store the data
 741  * that was kept in them.  There is no reason to write this data out to
 742  * the swap area if the application is discarding it.
 743  *
 744  * An interface that causes the system to free clean pages and flush
 745  * dirty pages is already available as msync(MS_INVALIDATE).
 746  */
 747 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
 748                                         unsigned long start, unsigned long end)
 749 {
 750         zap_page_range(vma, start, end - start);
 751         return 0;
 752 }
 753
 754 static long madvise_dontneed_free(struct vm_area_struct *vma,
 755                                   struct vm_area_struct **prev,
 756                                   unsigned long start, unsigned long end,
 757                                   int behavior)
 758 {
 759         *prev = vma;
 760         if (!can_madv_lru_vma(vma))
 761                 return -EINVAL;
 762
 763         if (!userfaultfd_remove(vma, start, end)) {
 764                 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
 765
 766                 down_read(&current->mm->mmap_sem);
 767                 vma = find_vma(current->mm, start);
 768                 if (!vma)
 769                         return -ENOMEM;
 770                 if (start < vma->vm_start) {
 771                         /*
 772                          * This "vma" under revalidation is the one
 773                          * with the lowest vma->vm_start where start
 774                          * is also < vma->vm_end. If start <
 775                          * vma->vm_start it means an hole materialized
 776                          * in the user address space within the
 777                          * virtual range passed to MADV_DONTNEED
 778                          * or MADV_FREE.
 779                          */
 780                         return -ENOMEM;
 781                 }
 782                 if (!can_madv_lru_vma(vma))
 783                         return -EINVAL;
 784                 if (end > vma->vm_end) {
 785                         /*
 786                          * Don't fail if end > vma->vm_end. If the old
 787                          * vma was splitted while the mmap_sem was
 788                          * released the effect of the concurrent
 789                          * operation may not cause madvise() to
 790                          * have an undefined result. There may be an
 791                          * adjacent next vma that we'll walk
 792                          * next. userfaultfd_remove() will generate an
 793                          * UFFD_EVENT_REMOVE repetition on the
 794                          * end-vma->vm_end range, but the manager can
 795                          * handle a repetition fine.
 796                          */
 797                         end = vma->vm_end;
 798                 }
 799                 VM_WARN_ON(start >= end);
 800         }
 801
 802         if (behavior == MADV_DONTNEED)
 803                 return madvise_dontneed_single_vma(vma, start, end);
 804         else if (behavior == MADV_FREE)
 805                 return madvise_free_single_vma(vma, start, end);
 806         else
 807                 return -EINVAL;
 808 }
 809
 810 /*
 811  * Application wants to free up the pages and associated backing store.
 812  * This is effectively punching a hole into the middle of a file.
 813  */
 814 static long madvise_remove(struct vm_area_struct *vma,
 815                                 struct vm_area_struct **prev,
 816                                 unsigned long start, unsigned long end)
 817 {
 818         loff_t offset;
 819         int error;
 820         struct file *f;
 821
 822         *prev = NULL;   /* tell sys_madvise we drop mmap_sem */
 823
 824         if (vma->vm_flags & VM_LOCKED)
 825                 return -EINVAL;
 826
 827         f = vma->vm_file;
 828
 829         if (!f || !f->f_mapping || !f->f_mapping->host) {
 830                         return -EINVAL;
 831         }
 832
 833         if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
 834                 return -EACCES;
 835
 836         offset = (loff_t)(start - vma->vm_start)
 837                         + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 838
 839         /*
 840          * Filesystem's fallocate may need to take i_mutex.  We need to
 841          * explicitly grab a reference because the vma (and hence the
 842          * vma's reference to the file) can go away as soon as we drop
 843          * mmap_sem.
 844          */
 845         get_file(f);
 846         if (userfaultfd_remove(vma, start, end)) {
 847                 /* mmap_sem was not released by userfaultfd_remove() */
 848                 up_read(&current->mm->mmap_sem);
 849         }
 850         error = vfs_fallocate(f,
 851                                 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
 852                                 offset, end - start);
 853         fput(f);
 854         down_read(&current->mm->mmap_sem);
 855         return error;
 856 }
 857
 858 #ifdef CONFIG_MEMORY_FAILURE
 859 /*
 860  * Error injection support for memory error handling.
 861  */
 862 static int madvise_inject_error(int behavior,
 863                 unsigned long start, unsigned long end)
 864 {
 865         struct page *page;
 866         struct zone *zone;
 867         unsigned int order;
 868
 869         if (!capable(CAP_SYS_ADMIN))
 870                 return -EPERM;
 871
 872
 873         for (; start < end; start += PAGE_SIZE << order) {
 874                 unsigned long pfn;
 875                 int ret;
 876
 877                 ret = get_user_pages_fast(start, 1, 0, &page);
 878                 if (ret != 1)
 879                         return ret;
 880                 pfn = page_to_pfn(page);
 881
 882                 /*
 883                  * When soft offlining hugepages, after migrating the page
 884                  * we dissolve it, therefore in the second loop "page" will
 885                  * no longer be a compound page, and order will be 0.
 886                  */
 887                 order = compound_order(compound_head(page));
 888
 889                 if (PageHWPoison(page)) {
 890                         put_page(page);
 891                         continue;
 892                 }
 893
 894                 if (behavior == MADV_SOFT_OFFLINE) {
 895                         pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
 896                                         pfn, start);
 897
 898                         ret = soft_offline_page(page, MF_COUNT_INCREASED);
 899                         if (ret)
 900                                 return ret;
 901                         continue;
 902                 }
 903
 904                 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
 905                                 pfn, start);
 906
 907                 /*
 908                  * Drop the page reference taken by get_user_pages_fast(). In
 909                  * the absence of MF_COUNT_INCREASED the memory_failure()
 910                  * routine is responsible for pinning the page to prevent it
 911                  * from being released back to the page allocator.
 912                  */
 913                 put_page(page);
 914                 ret = memory_failure(pfn, 0);
 915                 if (ret)
 916                         return ret;
 917         }
 918
 919         /* Ensure that all poisoned pages are removed from per-cpu lists */
 920         for_each_populated_zone(zone)
 921                 drain_all_pages(zone);
 922
 923         return 0;
 924 }
 925 #endif
 926
 927 static long
 928 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
 929                 unsigned long start, unsigned long end, int behavior)
 930 {
 931         switch (behavior) {
 932         case MADV_REMOVE:
 933                 return madvise_remove(vma, prev, start, end);
 934         case MADV_WILLNEED:
 935                 return madvise_willneed(vma, prev, start, end);
 936         case MADV_COLD:
 937                 return madvise_cold(vma, prev, start, end);
 938         case MADV_PAGEOUT:
 939                 return madvise_pageout(vma, prev, start, end);
 940         case MADV_FREE:
 941         case MADV_DONTNEED:
 942                 return madvise_dontneed_free(vma, prev, start, end, behavior);
 943         default:
 944                 return madvise_behavior(vma, prev, start, end, behavior);
 945         }
 946 }
 947
 948 static bool
 949 madvise_behavior_valid(int behavior)
 950 {
 951         switch (behavior) {
 952         case MADV_DOFORK:
 953         case MADV_DONTFORK:
 954         case MADV_NORMAL:
 955         case MADV_SEQUENTIAL:
 956         case MADV_RANDOM:
 957         case MADV_REMOVE:
 958         case MADV_WILLNEED:
 959         case MADV_DONTNEED:
 960         case MADV_FREE:
 961         case MADV_COLD:
 962         case MADV_PAGEOUT:
 963 #ifdef CONFIG_KSM
 964         case MADV_MERGEABLE:
 965         case MADV_UNMERGEABLE:
 966 #endif
 967 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
 968         case MADV_HUGEPAGE:
 969         case MADV_NOHUGEPAGE:
 970 #endif
 971         case MADV_DONTDUMP:
 972         case MADV_DODUMP:
 973         case MADV_WIPEONFORK:
 974         case MADV_KEEPONFORK:
 975 #ifdef CONFIG_MEMORY_FAILURE
 976         case MADV_SOFT_OFFLINE:
 977         case MADV_HWPOISON:
 978 #endif
 979                 return true;
 980
 981         default:
 982                 return false;
 983         }
 984 }
 985
 986 /*
 987  * The madvise(2) system call.
 988  *
 989  * Applications can use madvise() to advise the kernel how it should
 990  * handle paging I/O in this VM area.  The idea is to help the kernel
 991  * use appropriate read-ahead and caching techniques.  The information
 992  * provided is advisory only, and can be safely disregarded by the
 993  * kernel without affecting the correct operation of the application.
 994  *
 995  * behavior values:
 996  *  MADV_NORMAL - the default behavior is to read clusters.  This
 997  *              results in some read-ahead and read-behind.
 998  *  MADV_RANDOM - the system should read the minimum amount of data
 999  *              on any access, since it is unlikely that the appli-
1000  *              cation will need more than what it asks for.
1001  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1002  *              once, so they can be aggressively read ahead, and
1003  *              can be freed soon after they are accessed.
1004  *  MADV_WILLNEED - the application is notifying the system to read
1005  *              some pages ahead.
1006  *  MADV_DONTNEED - the application is finished with the given range,
1007  *              so the kernel can free resources associated with it.
1008  *  MADV_FREE - the application marks pages in the given range as lazy free,
1009  *              where actual purges are postponed until memory pressure happens.
1010  *  MADV_REMOVE - the application wants to free up the given range of
1011  *              pages and associated backing store.
1012  *  MADV_DONTFORK - omit this area from child's address space when forking:
1013  *              typically, to avoid COWing pages pinned by get_user_pages().
1014  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1015  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1016  *              range after a fork.
1017  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1018  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1019  *              were corrupted by unrecoverable hardware memory failure.
1020  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1021  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1022  *              this area with pages of identical content from other such areas.
1023  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1024  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1025  *              huge pages in the future. Existing pages might be coalesced and
1026  *              new pages might be allocated as THP.
1027  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1028  *              transparent huge pages so the existing pages will not be
1029  *              coalesced into THP and new pages will not be allocated as THP.
1030  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1031  *              from being included in its core dump.
1032  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1033  *
1034  * return values:
1035  *  zero    - success
1036  *  -EINVAL - start + len < 0, start is not page-aligned,
1037  *              "behavior" is not a valid value, or application
1038  *              is attempting to release locked or shared pages,
1039  *              or the specified address range includes file, Huge TLB,
1040  *              MAP_SHARED or VMPFNMAP range.
1041  *  -ENOMEM - addresses in the specified range are not currently
1042  *              mapped, or are outside the AS of the process.
1043  *  -EIO    - an I/O error occurred while paging in data.
1044  *  -EBADF  - map exists, but area maps something that isn't a file.
1045  *  -EAGAIN - a kernel resource was temporarily unavailable.
1046  */
1047 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1048 {
1049         unsigned long end, tmp;
1050         struct vm_area_struct *vma, *prev;
1051         int unmapped_error = 0;
1052         int error = -EINVAL;
1053         int write;
1054         size_t len;
1055         struct blk_plug plug;
1056
1057         start = untagged_addr(start);
1058
1059         if (!madvise_behavior_valid(behavior))
1060                 return error;
1061
1062         if (start & ~PAGE_MASK)
1063                 return error;
1064         len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1065
1066         /* Check to see whether len was rounded up from small -ve to zero */
1067         if (len_in && !len)
1068                 return error;
1069
1070         end = start + len;
1071         if (end < start)
1072                 return error;
1073
1074         error = 0;
1075         if (end == start)
1076                 return error;
1077
1078 #ifdef CONFIG_MEMORY_FAILURE
1079         if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1080                 return madvise_inject_error(behavior, start, start + len_in);
1081 #endif
1082
1083         write = madvise_need_mmap_write(behavior);
1084         if (write) {
1085                 if (down_write_killable(&current->mm->mmap_sem))
1086                         return -EINTR;
1087         } else {
1088                 down_read(&current->mm->mmap_sem);
1089         }
1090
1091         /*
1092          * If the interval [start,end) covers some unmapped address
1093          * ranges, just ignore them, but return -ENOMEM at the end.
1094          * - different from the way of handling in mlock etc.
1095          */
1096         vma = find_vma_prev(current->mm, start, &prev);
1097         if (vma && start > vma->vm_start)
1098                 prev = vma;
1099
1100         blk_start_plug(&plug);
1101         for (;;) {
1102                 /* Still start < end. */
1103                 error = -ENOMEM;
1104                 if (!vma)
1105                         goto out;
1106
1107                 /* Here start < (end|vma->vm_end). */
1108                 if (start < vma->vm_start) {
1109                         unmapped_error = -ENOMEM;
1110                         start = vma->vm_start;
1111                         if (start >= end)
1112                                 goto out;
1113                 }
1114
1115                 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1116                 tmp = vma->vm_end;
1117                 if (end < tmp)
1118                         tmp = end;
1119
1120                 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1121                 error = madvise_vma(vma, &prev, start, tmp, behavior);
1122                 if (error)
1123                         goto out;
1124                 start = tmp;
1125                 if (prev && start < prev->vm_end)
1126                         start = prev->vm_end;
1127                 error = unmapped_error;
1128                 if (start >= end)
1129                         goto out;
1130                 if (prev)
1131                         vma = prev->vm_next;
1132                 else    /* madvise_remove dropped mmap_sem */
1133                         vma = find_vma(current->mm, start);
1134         }
1135 out:
1136         blk_finish_plug(&plug);
1137         if (write)
1138                 up_write(&current->mm->mmap_sem);
1139         else
1140                 up_read(&current->mm->mmap_sem);
1141
1142         return error;
1143 }