1 // SPDX-License-Identifier: GPL-2.0
5 * Copyright (C) 1999 Linus Torvalds
6 * Copyright (C) 2002 Christoph Hellwig
9 #include <linux/mman.h>
10 #include <linux/pagemap.h>
11 #include <linux/syscalls.h>
12 #include <linux/mempolicy.h>
13 #include <linux/page-isolation.h>
14 #include <linux/page_idle.h>
15 #include <linux/userfaultfd_k.h>
16 #include <linux/hugetlb.h>
17 #include <linux/falloc.h>
18 #include <linux/fadvise.h>
19 #include <linux/sched.h>
20 #include <linux/sched/mm.h>
21 #include <linux/uio.h>
22 #include <linux/ksm.h>
24 #include <linux/file.h>
25 #include <linux/blkdev.h>
26 #include <linux/backing-dev.h>
27 #include <linux/pagewalk.h>
28 #include <linux/swap.h>
29 #include <linux/swapops.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/mmu_notifier.h>
37 struct madvise_walk_private {
38 struct mmu_gather *tlb;
43 * Any behaviour which results in changes to the vma->vm_flags needs to
44 * take mmap_lock for writing. Others, which simply traverse vmas, need
45 * to only take it for reading.
47 static int madvise_need_mmap_write(int behavior)
58 /* be safe, default to 1. list exceptions explicitly */
64 * We can potentially split a vm area into separate
65 * areas, each area with its own behavior.
67 static long madvise_behavior(struct vm_area_struct *vma,
68 struct vm_area_struct **prev,
69 unsigned long start, unsigned long end, int behavior)
71 struct mm_struct *mm = vma->vm_mm;
74 unsigned long new_flags = vma->vm_flags;
78 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
81 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
84 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
87 new_flags |= VM_DONTCOPY;
90 if (vma->vm_flags & VM_IO) {
94 new_flags &= ~VM_DONTCOPY;
97 /* MADV_WIPEONFORK is only supported on anonymous memory. */
98 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
102 new_flags |= VM_WIPEONFORK;
104 case MADV_KEEPONFORK:
105 new_flags &= ~VM_WIPEONFORK;
108 new_flags |= VM_DONTDUMP;
111 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
115 new_flags &= ~VM_DONTDUMP;
118 case MADV_UNMERGEABLE:
119 error = ksm_madvise(vma, start, end, behavior, &new_flags);
121 goto out_convert_errno;
124 case MADV_NOHUGEPAGE:
125 error = hugepage_madvise(vma, &new_flags, behavior);
127 goto out_convert_errno;
131 if (new_flags == vma->vm_flags) {
136 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
137 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
138 vma->vm_file, pgoff, vma_policy(vma),
139 vma->vm_userfaultfd_ctx);
147 if (start != vma->vm_start) {
148 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
152 error = __split_vma(mm, vma, start, 1);
154 goto out_convert_errno;
157 if (end != vma->vm_end) {
158 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
162 error = __split_vma(mm, vma, end, 0);
164 goto out_convert_errno;
169 * vm_flags is protected by the mmap_lock held in write mode.
171 vma->vm_flags = new_flags;
175 * madvise() returns EAGAIN if kernel resources, such as
176 * slab, are temporarily unavailable.
178 if (error == -ENOMEM)
185 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
186 unsigned long end, struct mm_walk *walk)
189 struct vm_area_struct *vma = walk->private;
192 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
195 for (index = start; index != end; index += PAGE_SIZE) {
201 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
202 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
203 pte_unmap_unlock(orig_pte, ptl);
205 if (pte_present(pte) || pte_none(pte))
207 entry = pte_to_swp_entry(pte);
208 if (unlikely(non_swap_entry(entry)))
211 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
220 static const struct mm_walk_ops swapin_walk_ops = {
221 .pmd_entry = swapin_walk_pmd_entry,
224 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
225 unsigned long start, unsigned long end,
226 struct address_space *mapping)
228 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
229 pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
233 xas_for_each(&xas, page, end_index) {
236 if (!xa_is_value(page))
241 swap = radix_to_swp_entry(page);
242 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
251 lru_add_drain(); /* Push any new pages onto the LRU now */
253 #endif /* CONFIG_SWAP */
256 * Schedule all required I/O operations. Do not wait for completion.
258 static long madvise_willneed(struct vm_area_struct *vma,
259 struct vm_area_struct **prev,
260 unsigned long start, unsigned long end)
262 struct mm_struct *mm = vma->vm_mm;
263 struct file *file = vma->vm_file;
269 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
270 lru_add_drain(); /* Push any new pages onto the LRU now */
274 if (shmem_mapping(file->f_mapping)) {
275 force_shm_swapin_readahead(vma, start, end,
284 if (IS_DAX(file_inode(file))) {
285 /* no bad return value, but ignore advice */
290 * Filesystem's fadvise may need to take various locks. We need to
291 * explicitly grab a reference because the vma (and hence the
292 * vma's reference to the file) can go away as soon as we drop
295 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
297 offset = (loff_t)(start - vma->vm_start)
298 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
299 mmap_read_unlock(mm);
300 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
306 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
307 unsigned long addr, unsigned long end,
308 struct mm_walk *walk)
310 struct madvise_walk_private *private = walk->private;
311 struct mmu_gather *tlb = private->tlb;
312 bool pageout = private->pageout;
313 struct mm_struct *mm = tlb->mm;
314 struct vm_area_struct *vma = walk->vma;
315 pte_t *orig_pte, *pte, ptent;
317 struct page *page = NULL;
318 LIST_HEAD(page_list);
320 if (fatal_signal_pending(current))
323 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
324 if (pmd_trans_huge(*pmd)) {
326 unsigned long next = pmd_addr_end(addr, end);
328 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
329 ptl = pmd_trans_huge_lock(pmd, vma);
334 if (is_huge_zero_pmd(orig_pmd))
337 if (unlikely(!pmd_present(orig_pmd))) {
338 VM_BUG_ON(thp_migration_supported() &&
339 !is_pmd_migration_entry(orig_pmd));
343 page = pmd_page(orig_pmd);
345 /* Do not interfere with other mappings of this page */
346 if (page_mapcount(page) != 1)
349 if (next - addr != HPAGE_PMD_SIZE) {
355 err = split_huge_page(page);
363 if (pmd_young(orig_pmd)) {
364 pmdp_invalidate(vma, addr, pmd);
365 orig_pmd = pmd_mkold(orig_pmd);
367 set_pmd_at(mm, addr, pmd, orig_pmd);
368 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
371 ClearPageReferenced(page);
372 test_and_clear_page_young(page);
374 if (!isolate_lru_page(page)) {
375 if (PageUnevictable(page))
376 putback_lru_page(page);
378 list_add(&page->lru, &page_list);
381 deactivate_page(page);
385 reclaim_pages(&page_list);
390 if (pmd_trans_unstable(pmd))
393 tlb_change_page_size(tlb, PAGE_SIZE);
394 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
395 flush_tlb_batched_pending(mm);
396 arch_enter_lazy_mmu_mode();
397 for (; addr < end; pte++, addr += PAGE_SIZE) {
403 if (!pte_present(ptent))
406 page = vm_normal_page(vma, addr, ptent);
411 * Creating a THP page is expensive so split it only if we
412 * are sure it's worth. Split it if we are only owner.
414 if (PageTransCompound(page)) {
415 if (page_mapcount(page) != 1)
418 if (!trylock_page(page)) {
422 pte_unmap_unlock(orig_pte, ptl);
423 if (split_huge_page(page)) {
426 pte_offset_map_lock(mm, pmd, addr, &ptl);
431 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
438 * Do not interfere with other mappings of this page and
441 if (!PageLRU(page) || page_mapcount(page) != 1)
444 VM_BUG_ON_PAGE(PageTransCompound(page), page);
446 if (pte_young(ptent)) {
447 ptent = ptep_get_and_clear_full(mm, addr, pte,
449 ptent = pte_mkold(ptent);
450 set_pte_at(mm, addr, pte, ptent);
451 tlb_remove_tlb_entry(tlb, pte, addr);
455 * We are deactivating a page for accelerating reclaiming.
456 * VM couldn't reclaim the page unless we clear PG_young.
457 * As a side effect, it makes confuse idle-page tracking
458 * because they will miss recent referenced history.
460 ClearPageReferenced(page);
461 test_and_clear_page_young(page);
463 if (!isolate_lru_page(page)) {
464 if (PageUnevictable(page))
465 putback_lru_page(page);
467 list_add(&page->lru, &page_list);
470 deactivate_page(page);
473 arch_leave_lazy_mmu_mode();
474 pte_unmap_unlock(orig_pte, ptl);
476 reclaim_pages(&page_list);
482 static const struct mm_walk_ops cold_walk_ops = {
483 .pmd_entry = madvise_cold_or_pageout_pte_range,
486 static void madvise_cold_page_range(struct mmu_gather *tlb,
487 struct vm_area_struct *vma,
488 unsigned long addr, unsigned long end)
490 struct madvise_walk_private walk_private = {
495 tlb_start_vma(tlb, vma);
496 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
497 tlb_end_vma(tlb, vma);
500 static long madvise_cold(struct vm_area_struct *vma,
501 struct vm_area_struct **prev,
502 unsigned long start_addr, unsigned long end_addr)
504 struct mm_struct *mm = vma->vm_mm;
505 struct mmu_gather tlb;
508 if (!can_madv_lru_vma(vma))
512 tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
513 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
514 tlb_finish_mmu(&tlb, start_addr, end_addr);
519 static void madvise_pageout_page_range(struct mmu_gather *tlb,
520 struct vm_area_struct *vma,
521 unsigned long addr, unsigned long end)
523 struct madvise_walk_private walk_private = {
528 tlb_start_vma(tlb, vma);
529 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
530 tlb_end_vma(tlb, vma);
533 static inline bool can_do_pageout(struct vm_area_struct *vma)
535 if (vma_is_anonymous(vma))
540 * paging out pagecache only for non-anonymous mappings that correspond
541 * to the files the calling process could (if tried) open for writing;
542 * otherwise we'd be including shared non-exclusive mappings, which
543 * opens a side channel.
545 return inode_owner_or_capable(file_inode(vma->vm_file)) ||
546 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
549 static long madvise_pageout(struct vm_area_struct *vma,
550 struct vm_area_struct **prev,
551 unsigned long start_addr, unsigned long end_addr)
553 struct mm_struct *mm = vma->vm_mm;
554 struct mmu_gather tlb;
557 if (!can_madv_lru_vma(vma))
560 if (!can_do_pageout(vma))
564 tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
565 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
566 tlb_finish_mmu(&tlb, start_addr, end_addr);
571 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
572 unsigned long end, struct mm_walk *walk)
575 struct mmu_gather *tlb = walk->private;
576 struct mm_struct *mm = tlb->mm;
577 struct vm_area_struct *vma = walk->vma;
579 pte_t *orig_pte, *pte, ptent;
584 next = pmd_addr_end(addr, end);
585 if (pmd_trans_huge(*pmd))
586 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
589 if (pmd_trans_unstable(pmd))
592 tlb_change_page_size(tlb, PAGE_SIZE);
593 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
594 flush_tlb_batched_pending(mm);
595 arch_enter_lazy_mmu_mode();
596 for (; addr != end; pte++, addr += PAGE_SIZE) {
602 * If the pte has swp_entry, just clear page table to
603 * prevent swap-in which is more expensive rather than
604 * (page allocation + zeroing).
606 if (!pte_present(ptent)) {
609 entry = pte_to_swp_entry(ptent);
610 if (non_swap_entry(entry))
613 free_swap_and_cache(entry);
614 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
618 page = vm_normal_page(vma, addr, ptent);
623 * If pmd isn't transhuge but the page is THP and
624 * is owned by only this process, split it and
625 * deactivate all pages.
627 if (PageTransCompound(page)) {
628 if (page_mapcount(page) != 1)
631 if (!trylock_page(page)) {
635 pte_unmap_unlock(orig_pte, ptl);
636 if (split_huge_page(page)) {
639 pte_offset_map_lock(mm, pmd, addr, &ptl);
644 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
650 VM_BUG_ON_PAGE(PageTransCompound(page), page);
652 if (PageSwapCache(page) || PageDirty(page)) {
653 if (!trylock_page(page))
656 * If page is shared with others, we couldn't clear
657 * PG_dirty of the page.
659 if (page_mapcount(page) != 1) {
664 if (PageSwapCache(page) && !try_to_free_swap(page)) {
669 ClearPageDirty(page);
673 if (pte_young(ptent) || pte_dirty(ptent)) {
675 * Some of architecture(ex, PPC) don't update TLB
676 * with set_pte_at and tlb_remove_tlb_entry so for
677 * the portability, remap the pte with old|clean
678 * after pte clearing.
680 ptent = ptep_get_and_clear_full(mm, addr, pte,
683 ptent = pte_mkold(ptent);
684 ptent = pte_mkclean(ptent);
685 set_pte_at(mm, addr, pte, ptent);
686 tlb_remove_tlb_entry(tlb, pte, addr);
688 mark_page_lazyfree(page);
692 if (current->mm == mm)
695 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
697 arch_leave_lazy_mmu_mode();
698 pte_unmap_unlock(orig_pte, ptl);
704 static const struct mm_walk_ops madvise_free_walk_ops = {
705 .pmd_entry = madvise_free_pte_range,
708 static int madvise_free_single_vma(struct vm_area_struct *vma,
709 unsigned long start_addr, unsigned long end_addr)
711 struct mm_struct *mm = vma->vm_mm;
712 struct mmu_notifier_range range;
713 struct mmu_gather tlb;
715 /* MADV_FREE works for only anon vma at the moment */
716 if (!vma_is_anonymous(vma))
719 range.start = max(vma->vm_start, start_addr);
720 if (range.start >= vma->vm_end)
722 range.end = min(vma->vm_end, end_addr);
723 if (range.end <= vma->vm_start)
725 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
726 range.start, range.end);
729 tlb_gather_mmu(&tlb, mm, range.start, range.end);
730 update_hiwater_rss(mm);
732 mmu_notifier_invalidate_range_start(&range);
733 tlb_start_vma(&tlb, vma);
734 walk_page_range(vma->vm_mm, range.start, range.end,
735 &madvise_free_walk_ops, &tlb);
736 tlb_end_vma(&tlb, vma);
737 mmu_notifier_invalidate_range_end(&range);
738 tlb_finish_mmu(&tlb, range.start, range.end);
744 * Application no longer needs these pages. If the pages are dirty,
745 * it's OK to just throw them away. The app will be more careful about
746 * data it wants to keep. Be sure to free swap resources too. The
747 * zap_page_range call sets things up for shrink_active_list to actually free
748 * these pages later if no one else has touched them in the meantime,
749 * although we could add these pages to a global reuse list for
750 * shrink_active_list to pick up before reclaiming other pages.
752 * NB: This interface discards data rather than pushes it out to swap,
753 * as some implementations do. This has performance implications for
754 * applications like large transactional databases which want to discard
755 * pages in anonymous maps after committing to backing store the data
756 * that was kept in them. There is no reason to write this data out to
757 * the swap area if the application is discarding it.
759 * An interface that causes the system to free clean pages and flush
760 * dirty pages is already available as msync(MS_INVALIDATE).
762 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
763 unsigned long start, unsigned long end)
765 zap_page_range(vma, start, end - start);
769 static long madvise_dontneed_free(struct vm_area_struct *vma,
770 struct vm_area_struct **prev,
771 unsigned long start, unsigned long end,
774 struct mm_struct *mm = vma->vm_mm;
777 if (!can_madv_lru_vma(vma))
780 if (!userfaultfd_remove(vma, start, end)) {
781 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
784 vma = find_vma(mm, start);
787 if (start < vma->vm_start) {
789 * This "vma" under revalidation is the one
790 * with the lowest vma->vm_start where start
791 * is also < vma->vm_end. If start <
792 * vma->vm_start it means an hole materialized
793 * in the user address space within the
794 * virtual range passed to MADV_DONTNEED
799 if (!can_madv_lru_vma(vma))
801 if (end > vma->vm_end) {
803 * Don't fail if end > vma->vm_end. If the old
804 * vma was splitted while the mmap_lock was
805 * released the effect of the concurrent
806 * operation may not cause madvise() to
807 * have an undefined result. There may be an
808 * adjacent next vma that we'll walk
809 * next. userfaultfd_remove() will generate an
810 * UFFD_EVENT_REMOVE repetition on the
811 * end-vma->vm_end range, but the manager can
812 * handle a repetition fine.
816 VM_WARN_ON(start >= end);
819 if (behavior == MADV_DONTNEED)
820 return madvise_dontneed_single_vma(vma, start, end);
821 else if (behavior == MADV_FREE)
822 return madvise_free_single_vma(vma, start, end);
828 * Application wants to free up the pages and associated backing store.
829 * This is effectively punching a hole into the middle of a file.
831 static long madvise_remove(struct vm_area_struct *vma,
832 struct vm_area_struct **prev,
833 unsigned long start, unsigned long end)
838 struct mm_struct *mm = vma->vm_mm;
840 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
842 if (vma->vm_flags & VM_LOCKED)
847 if (!f || !f->f_mapping || !f->f_mapping->host) {
851 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
854 offset = (loff_t)(start - vma->vm_start)
855 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
858 * Filesystem's fallocate may need to take i_mutex. We need to
859 * explicitly grab a reference because the vma (and hence the
860 * vma's reference to the file) can go away as soon as we drop
864 if (userfaultfd_remove(vma, start, end)) {
865 /* mmap_lock was not released by userfaultfd_remove() */
866 mmap_read_unlock(mm);
868 error = vfs_fallocate(f,
869 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
870 offset, end - start);
876 #ifdef CONFIG_MEMORY_FAILURE
878 * Error injection support for memory error handling.
880 static int madvise_inject_error(int behavior,
881 unsigned long start, unsigned long end)
886 if (!capable(CAP_SYS_ADMIN))
890 for (; start < end; start += size) {
895 ret = get_user_pages_fast(start, 1, 0, &page);
898 pfn = page_to_pfn(page);
901 * When soft offlining hugepages, after migrating the page
902 * we dissolve it, therefore in the second loop "page" will
903 * no longer be a compound page.
905 size = page_size(compound_head(page));
907 if (behavior == MADV_SOFT_OFFLINE) {
908 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
910 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
912 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
914 ret = memory_failure(pfn, MF_COUNT_INCREASED);
921 /* Ensure that all poisoned pages are removed from per-cpu lists */
922 for_each_populated_zone(zone)
923 drain_all_pages(zone);
930 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
931 unsigned long start, unsigned long end, int behavior)
935 return madvise_remove(vma, prev, start, end);
937 return madvise_willneed(vma, prev, start, end);
939 return madvise_cold(vma, prev, start, end);
941 return madvise_pageout(vma, prev, start, end);
944 return madvise_dontneed_free(vma, prev, start, end, behavior);
946 return madvise_behavior(vma, prev, start, end, behavior);
951 madvise_behavior_valid(int behavior)
957 case MADV_SEQUENTIAL:
967 case MADV_UNMERGEABLE:
969 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
971 case MADV_NOHUGEPAGE:
975 case MADV_WIPEONFORK:
976 case MADV_KEEPONFORK:
977 #ifdef CONFIG_MEMORY_FAILURE
978 case MADV_SOFT_OFFLINE:
989 process_madvise_behavior_valid(int behavior)
1001 * The madvise(2) system call.
1003 * Applications can use madvise() to advise the kernel how it should
1004 * handle paging I/O in this VM area. The idea is to help the kernel
1005 * use appropriate read-ahead and caching techniques. The information
1006 * provided is advisory only, and can be safely disregarded by the
1007 * kernel without affecting the correct operation of the application.
1010 * MADV_NORMAL - the default behavior is to read clusters. This
1011 * results in some read-ahead and read-behind.
1012 * MADV_RANDOM - the system should read the minimum amount of data
1013 * on any access, since it is unlikely that the appli-
1014 * cation will need more than what it asks for.
1015 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1016 * once, so they can be aggressively read ahead, and
1017 * can be freed soon after they are accessed.
1018 * MADV_WILLNEED - the application is notifying the system to read
1020 * MADV_DONTNEED - the application is finished with the given range,
1021 * so the kernel can free resources associated with it.
1022 * MADV_FREE - the application marks pages in the given range as lazy free,
1023 * where actual purges are postponed until memory pressure happens.
1024 * MADV_REMOVE - the application wants to free up the given range of
1025 * pages and associated backing store.
1026 * MADV_DONTFORK - omit this area from child's address space when forking:
1027 * typically, to avoid COWing pages pinned by get_user_pages().
1028 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1029 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1030 * range after a fork.
1031 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1032 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1033 * were corrupted by unrecoverable hardware memory failure.
1034 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1035 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1036 * this area with pages of identical content from other such areas.
1037 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1038 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1039 * huge pages in the future. Existing pages might be coalesced and
1040 * new pages might be allocated as THP.
1041 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1042 * transparent huge pages so the existing pages will not be
1043 * coalesced into THP and new pages will not be allocated as THP.
1044 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1045 * from being included in its core dump.
1046 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1047 * MADV_COLD - the application is not expected to use this memory soon,
1048 * deactivate pages in this range so that they can be reclaimed
1049 * easily if memory pressure hanppens.
1050 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1051 * page out the pages in this range immediately.
1055 * -EINVAL - start + len < 0, start is not page-aligned,
1056 * "behavior" is not a valid value, or application
1057 * is attempting to release locked or shared pages,
1058 * or the specified address range includes file, Huge TLB,
1059 * MAP_SHARED or VMPFNMAP range.
1060 * -ENOMEM - addresses in the specified range are not currently
1061 * mapped, or are outside the AS of the process.
1062 * -EIO - an I/O error occurred while paging in data.
1063 * -EBADF - map exists, but area maps something that isn't a file.
1064 * -EAGAIN - a kernel resource was temporarily unavailable.
1066 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1068 unsigned long end, tmp;
1069 struct vm_area_struct *vma, *prev;
1070 int unmapped_error = 0;
1071 int error = -EINVAL;
1074 struct blk_plug plug;
1076 start = untagged_addr(start);
1078 if (!madvise_behavior_valid(behavior))
1081 if (!PAGE_ALIGNED(start))
1083 len = PAGE_ALIGN(len_in);
1085 /* Check to see whether len was rounded up from small -ve to zero */
1097 #ifdef CONFIG_MEMORY_FAILURE
1098 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1099 return madvise_inject_error(behavior, start, start + len_in);
1102 write = madvise_need_mmap_write(behavior);
1104 if (mmap_write_lock_killable(mm))
1111 * If the interval [start,end) covers some unmapped address
1112 * ranges, just ignore them, but return -ENOMEM at the end.
1113 * - different from the way of handling in mlock etc.
1115 vma = find_vma_prev(mm, start, &prev);
1116 if (vma && start > vma->vm_start)
1119 blk_start_plug(&plug);
1121 /* Still start < end. */
1126 /* Here start < (end|vma->vm_end). */
1127 if (start < vma->vm_start) {
1128 unmapped_error = -ENOMEM;
1129 start = vma->vm_start;
1134 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1139 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1140 error = madvise_vma(vma, &prev, start, tmp, behavior);
1144 if (prev && start < prev->vm_end)
1145 start = prev->vm_end;
1146 error = unmapped_error;
1150 vma = prev->vm_next;
1151 else /* madvise_remove dropped mmap_lock */
1152 vma = find_vma(mm, start);
1155 blk_finish_plug(&plug);
1157 mmap_write_unlock(mm);
1159 mmap_read_unlock(mm);
1164 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1166 return do_madvise(current->mm, start, len_in, behavior);
1169 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1170 size_t, vlen, int, behavior, unsigned int, flags)
1173 struct iovec iovstack[UIO_FASTIOV], iovec;
1174 struct iovec *iov = iovstack;
1175 struct iov_iter iter;
1177 struct task_struct *task;
1178 struct mm_struct *mm;
1180 unsigned int f_flags;
1187 ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1191 pid = pidfd_get_pid(pidfd, &f_flags);
1197 task = get_pid_task(pid, PIDTYPE_PID);
1203 if (!process_madvise_behavior_valid(behavior)) {
1208 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1209 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1210 if (IS_ERR_OR_NULL(mm)) {
1211 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1216 * Require CAP_SYS_NICE for influencing process performance. Note that
1217 * only non-destructive hints are currently supported.
1219 if (!capable(CAP_SYS_NICE)) {
1224 total_len = iov_iter_count(&iter);
1226 while (iov_iter_count(&iter)) {
1227 iovec = iov_iter_iovec(&iter);
1228 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1229 iovec.iov_len, behavior);
1232 iov_iter_advance(&iter, iovec.iov_len);
1235 ret = (total_len - iov_iter_count(&iter)) ? : ret;
1240 put_task_struct(task);