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/mm_inline.h>
22 #include <linux/string.h>
23 #include <linux/uio.h>
24 #include <linux/ksm.h>
26 #include <linux/file.h>
27 #include <linux/blkdev.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagewalk.h>
30 #include <linux/swap.h>
31 #include <linux/swapops.h>
32 #include <linux/shmem_fs.h>
33 #include <linux/mmu_notifier.h>
40 struct madvise_walk_private {
41 struct mmu_gather *tlb;
46 * Any behaviour which results in changes to the vma->vm_flags needs to
47 * take mmap_lock for writing. Others, which simply traverse vmas, need
48 * to only take it for reading.
50 static int madvise_need_mmap_write(int behavior)
56 case MADV_DONTNEED_LOCKED:
60 case MADV_POPULATE_READ:
61 case MADV_POPULATE_WRITE:
65 /* be safe, default to 1. list exceptions explicitly */
70 #ifdef CONFIG_ANON_VMA_NAME
71 struct anon_vma_name *anon_vma_name_alloc(const char *name)
73 struct anon_vma_name *anon_name;
76 /* Add 1 for NUL terminator at the end of the anon_name->name */
77 count = strlen(name) + 1;
78 anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
80 kref_init(&anon_name->kref);
81 memcpy(anon_name->name, name, count);
87 void anon_vma_name_free(struct kref *kref)
89 struct anon_vma_name *anon_name =
90 container_of(kref, struct anon_vma_name, kref);
94 struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
96 mmap_assert_locked(vma->vm_mm);
98 return vma->anon_name;
101 /* mmap_lock should be write-locked */
102 static int replace_anon_vma_name(struct vm_area_struct *vma,
103 struct anon_vma_name *anon_name)
105 struct anon_vma_name *orig_name = anon_vma_name(vma);
108 vma->anon_name = NULL;
109 anon_vma_name_put(orig_name);
113 if (anon_vma_name_eq(orig_name, anon_name))
116 vma->anon_name = anon_vma_name_reuse(anon_name);
117 anon_vma_name_put(orig_name);
121 #else /* CONFIG_ANON_VMA_NAME */
122 static int replace_anon_vma_name(struct vm_area_struct *vma,
123 struct anon_vma_name *anon_name)
130 #endif /* CONFIG_ANON_VMA_NAME */
132 * Update the vm_flags on region of a vma, splitting it or merging it as
133 * necessary. Must be called with mmap_lock held for writing;
134 * Caller should ensure anon_name stability by raising its refcount even when
135 * anon_name belongs to a valid vma because this function might free that vma.
137 static int madvise_update_vma(struct vm_area_struct *vma,
138 struct vm_area_struct **prev, unsigned long start,
139 unsigned long end, unsigned long new_flags,
140 struct anon_vma_name *anon_name)
142 struct mm_struct *mm = vma->vm_mm;
145 VMA_ITERATOR(vmi, mm, start);
147 if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
152 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
153 *prev = vma_merge(&vmi, mm, *prev, start, end, new_flags,
154 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
155 vma->vm_userfaultfd_ctx, anon_name);
163 if (start != vma->vm_start) {
164 error = split_vma(&vmi, vma, start, 1);
169 if (end != vma->vm_end) {
170 error = split_vma(&vmi, vma, end, 0);
177 * vm_flags is protected by the mmap_lock held in write mode.
179 vm_flags_reset(vma, new_flags);
180 if (!vma->vm_file || vma_is_anon_shmem(vma)) {
181 error = replace_anon_vma_name(vma, anon_name);
190 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
191 unsigned long end, struct mm_walk *walk)
193 struct vm_area_struct *vma = walk->private;
195 struct swap_iocb *splug = NULL;
197 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
200 for (index = start; index != end; index += PAGE_SIZE) {
207 ptep = pte_offset_map_lock(vma->vm_mm, pmd, index, &ptl);
209 pte_unmap_unlock(ptep, ptl);
211 if (!is_swap_pte(pte))
213 entry = pte_to_swp_entry(pte);
214 if (unlikely(non_swap_entry(entry)))
217 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
218 vma, index, false, &splug);
222 swap_read_unplug(splug);
228 static const struct mm_walk_ops swapin_walk_ops = {
229 .pmd_entry = swapin_walk_pmd_entry,
230 .walk_lock = PGWALK_RDLOCK,
233 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
234 unsigned long start, unsigned long end,
235 struct address_space *mapping)
237 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
238 pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
240 struct swap_iocb *splug = NULL;
243 xas_for_each(&xas, page, end_index) {
246 if (!xa_is_value(page))
248 swap = radix_to_swp_entry(page);
249 /* There might be swapin error entries in shmem mapping. */
250 if (non_swap_entry(swap))
255 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
256 NULL, 0, false, &splug);
263 swap_read_unplug(splug);
265 lru_add_drain(); /* Push any new pages onto the LRU now */
267 #endif /* CONFIG_SWAP */
270 * Schedule all required I/O operations. Do not wait for completion.
272 static long madvise_willneed(struct vm_area_struct *vma,
273 struct vm_area_struct **prev,
274 unsigned long start, unsigned long end)
276 struct mm_struct *mm = vma->vm_mm;
277 struct file *file = vma->vm_file;
283 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
284 lru_add_drain(); /* Push any new pages onto the LRU now */
288 if (shmem_mapping(file->f_mapping)) {
289 force_shm_swapin_readahead(vma, start, end,
298 if (IS_DAX(file_inode(file))) {
299 /* no bad return value, but ignore advice */
304 * Filesystem's fadvise may need to take various locks. We need to
305 * explicitly grab a reference because the vma (and hence the
306 * vma's reference to the file) can go away as soon as we drop
309 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
311 offset = (loff_t)(start - vma->vm_start)
312 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
313 mmap_read_unlock(mm);
314 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
320 static inline bool can_do_file_pageout(struct vm_area_struct *vma)
325 * paging out pagecache only for non-anonymous mappings that correspond
326 * to the files the calling process could (if tried) open for writing;
327 * otherwise we'd be including shared non-exclusive mappings, which
328 * opens a side channel.
330 return inode_owner_or_capable(&nop_mnt_idmap,
331 file_inode(vma->vm_file)) ||
332 file_permission(vma->vm_file, MAY_WRITE) == 0;
335 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
336 unsigned long addr, unsigned long end,
337 struct mm_walk *walk)
339 struct madvise_walk_private *private = walk->private;
340 struct mmu_gather *tlb = private->tlb;
341 bool pageout = private->pageout;
342 struct mm_struct *mm = tlb->mm;
343 struct vm_area_struct *vma = walk->vma;
344 pte_t *orig_pte, *pte, ptent;
346 struct folio *folio = NULL;
347 LIST_HEAD(folio_list);
348 bool pageout_anon_only_filter;
350 if (fatal_signal_pending(current))
353 pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
354 !can_do_file_pageout(vma);
356 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
357 if (pmd_trans_huge(*pmd)) {
359 unsigned long next = pmd_addr_end(addr, end);
361 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
362 ptl = pmd_trans_huge_lock(pmd, vma);
367 if (is_huge_zero_pmd(orig_pmd))
370 if (unlikely(!pmd_present(orig_pmd))) {
371 VM_BUG_ON(thp_migration_supported() &&
372 !is_pmd_migration_entry(orig_pmd));
376 folio = pfn_folio(pmd_pfn(orig_pmd));
378 /* Do not interfere with other mappings of this folio */
379 if (folio_estimated_sharers(folio) != 1)
382 if (pageout_anon_only_filter && !folio_test_anon(folio))
385 if (next - addr != HPAGE_PMD_SIZE) {
391 err = split_folio(folio);
399 if (pmd_young(orig_pmd)) {
400 pmdp_invalidate(vma, addr, pmd);
401 orig_pmd = pmd_mkold(orig_pmd);
403 set_pmd_at(mm, addr, pmd, orig_pmd);
404 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
407 folio_clear_referenced(folio);
408 folio_test_clear_young(folio);
410 if (folio_isolate_lru(folio)) {
411 if (folio_test_unevictable(folio))
412 folio_putback_lru(folio);
414 list_add(&folio->lru, &folio_list);
417 folio_deactivate(folio);
421 reclaim_pages(&folio_list);
426 if (pmd_trans_unstable(pmd))
429 tlb_change_page_size(tlb, PAGE_SIZE);
430 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
431 flush_tlb_batched_pending(mm);
432 arch_enter_lazy_mmu_mode();
433 for (; addr < end; pte++, addr += PAGE_SIZE) {
439 if (!pte_present(ptent))
442 folio = vm_normal_folio(vma, addr, ptent);
443 if (!folio || folio_is_zone_device(folio))
447 * Creating a THP page is expensive so split it only if we
448 * are sure it's worth. Split it if we are only owner.
450 if (folio_test_large(folio)) {
451 if (folio_estimated_sharers(folio) != 1)
453 if (pageout_anon_only_filter && !folio_test_anon(folio))
456 if (!folio_trylock(folio)) {
460 pte_unmap_unlock(orig_pte, ptl);
461 if (split_folio(folio)) {
464 orig_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
469 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
476 * Do not interfere with other mappings of this folio and
479 if (!folio_test_lru(folio) || folio_mapcount(folio) != 1)
482 if (pageout_anon_only_filter && !folio_test_anon(folio))
485 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
487 if (pte_young(ptent)) {
488 ptent = ptep_get_and_clear_full(mm, addr, pte,
490 ptent = pte_mkold(ptent);
491 set_pte_at(mm, addr, pte, ptent);
492 tlb_remove_tlb_entry(tlb, pte, addr);
496 * We are deactivating a folio for accelerating reclaiming.
497 * VM couldn't reclaim the folio unless we clear PG_young.
498 * As a side effect, it makes confuse idle-page tracking
499 * because they will miss recent referenced history.
501 folio_clear_referenced(folio);
502 folio_test_clear_young(folio);
504 if (folio_isolate_lru(folio)) {
505 if (folio_test_unevictable(folio))
506 folio_putback_lru(folio);
508 list_add(&folio->lru, &folio_list);
511 folio_deactivate(folio);
514 arch_leave_lazy_mmu_mode();
515 pte_unmap_unlock(orig_pte, ptl);
517 reclaim_pages(&folio_list);
523 static const struct mm_walk_ops cold_walk_ops = {
524 .pmd_entry = madvise_cold_or_pageout_pte_range,
525 .walk_lock = PGWALK_RDLOCK,
528 static void madvise_cold_page_range(struct mmu_gather *tlb,
529 struct vm_area_struct *vma,
530 unsigned long addr, unsigned long end)
532 struct madvise_walk_private walk_private = {
537 tlb_start_vma(tlb, vma);
538 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
539 tlb_end_vma(tlb, vma);
542 static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
544 return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
547 static long madvise_cold(struct vm_area_struct *vma,
548 struct vm_area_struct **prev,
549 unsigned long start_addr, unsigned long end_addr)
551 struct mm_struct *mm = vma->vm_mm;
552 struct mmu_gather tlb;
555 if (!can_madv_lru_vma(vma))
559 tlb_gather_mmu(&tlb, mm);
560 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
561 tlb_finish_mmu(&tlb);
566 static void madvise_pageout_page_range(struct mmu_gather *tlb,
567 struct vm_area_struct *vma,
568 unsigned long addr, unsigned long end)
570 struct madvise_walk_private walk_private = {
575 tlb_start_vma(tlb, vma);
576 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
577 tlb_end_vma(tlb, vma);
580 static long madvise_pageout(struct vm_area_struct *vma,
581 struct vm_area_struct **prev,
582 unsigned long start_addr, unsigned long end_addr)
584 struct mm_struct *mm = vma->vm_mm;
585 struct mmu_gather tlb;
588 if (!can_madv_lru_vma(vma))
592 * If the VMA belongs to a private file mapping, there can be private
593 * dirty pages which can be paged out if even this process is neither
594 * owner nor write capable of the file. We allow private file mappings
595 * further to pageout dirty anon pages.
597 if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
598 (vma->vm_flags & VM_MAYSHARE)))
602 tlb_gather_mmu(&tlb, mm);
603 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
604 tlb_finish_mmu(&tlb);
609 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
610 unsigned long end, struct mm_walk *walk)
613 struct mmu_gather *tlb = walk->private;
614 struct mm_struct *mm = tlb->mm;
615 struct vm_area_struct *vma = walk->vma;
617 pte_t *orig_pte, *pte, ptent;
622 next = pmd_addr_end(addr, end);
623 if (pmd_trans_huge(*pmd))
624 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
627 if (pmd_trans_unstable(pmd))
630 tlb_change_page_size(tlb, PAGE_SIZE);
631 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
632 flush_tlb_batched_pending(mm);
633 arch_enter_lazy_mmu_mode();
634 for (; addr != end; pte++, addr += PAGE_SIZE) {
640 * If the pte has swp_entry, just clear page table to
641 * prevent swap-in which is more expensive rather than
642 * (page allocation + zeroing).
644 if (!pte_present(ptent)) {
647 entry = pte_to_swp_entry(ptent);
648 if (!non_swap_entry(entry)) {
650 free_swap_and_cache(entry);
651 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
652 } else if (is_hwpoison_entry(entry) ||
653 is_swapin_error_entry(entry)) {
654 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
659 folio = vm_normal_folio(vma, addr, ptent);
660 if (!folio || folio_is_zone_device(folio))
664 * If pmd isn't transhuge but the folio is large and
665 * is owned by only this process, split it and
666 * deactivate all pages.
668 if (folio_test_large(folio)) {
669 if (folio_estimated_sharers(folio) != 1)
672 if (!folio_trylock(folio)) {
676 pte_unmap_unlock(orig_pte, ptl);
677 if (split_folio(folio)) {
680 orig_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
685 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
691 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
692 if (!folio_trylock(folio))
695 * If folio is shared with others, we mustn't clear
696 * the folio's dirty flag.
698 if (folio_mapcount(folio) != 1) {
703 if (folio_test_swapcache(folio) &&
704 !folio_free_swap(folio)) {
709 folio_clear_dirty(folio);
713 if (pte_young(ptent) || pte_dirty(ptent)) {
715 * Some of architecture(ex, PPC) don't update TLB
716 * with set_pte_at and tlb_remove_tlb_entry so for
717 * the portability, remap the pte with old|clean
718 * after pte clearing.
720 ptent = ptep_get_and_clear_full(mm, addr, pte,
723 ptent = pte_mkold(ptent);
724 ptent = pte_mkclean(ptent);
725 set_pte_at(mm, addr, pte, ptent);
726 tlb_remove_tlb_entry(tlb, pte, addr);
728 folio_mark_lazyfree(folio);
732 if (current->mm == mm)
735 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
737 arch_leave_lazy_mmu_mode();
738 pte_unmap_unlock(orig_pte, ptl);
744 static const struct mm_walk_ops madvise_free_walk_ops = {
745 .pmd_entry = madvise_free_pte_range,
746 .walk_lock = PGWALK_RDLOCK,
749 static int madvise_free_single_vma(struct vm_area_struct *vma,
750 unsigned long start_addr, unsigned long end_addr)
752 struct mm_struct *mm = vma->vm_mm;
753 struct mmu_notifier_range range;
754 struct mmu_gather tlb;
756 /* MADV_FREE works for only anon vma at the moment */
757 if (!vma_is_anonymous(vma))
760 range.start = max(vma->vm_start, start_addr);
761 if (range.start >= vma->vm_end)
763 range.end = min(vma->vm_end, end_addr);
764 if (range.end <= vma->vm_start)
766 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
767 range.start, range.end);
770 tlb_gather_mmu(&tlb, mm);
771 update_hiwater_rss(mm);
773 mmu_notifier_invalidate_range_start(&range);
774 tlb_start_vma(&tlb, vma);
775 walk_page_range(vma->vm_mm, range.start, range.end,
776 &madvise_free_walk_ops, &tlb);
777 tlb_end_vma(&tlb, vma);
778 mmu_notifier_invalidate_range_end(&range);
779 tlb_finish_mmu(&tlb);
785 * Application no longer needs these pages. If the pages are dirty,
786 * it's OK to just throw them away. The app will be more careful about
787 * data it wants to keep. Be sure to free swap resources too. The
788 * zap_page_range_single call sets things up for shrink_active_list to actually
789 * free these pages later if no one else has touched them in the meantime,
790 * although we could add these pages to a global reuse list for
791 * shrink_active_list to pick up before reclaiming other pages.
793 * NB: This interface discards data rather than pushes it out to swap,
794 * as some implementations do. This has performance implications for
795 * applications like large transactional databases which want to discard
796 * pages in anonymous maps after committing to backing store the data
797 * that was kept in them. There is no reason to write this data out to
798 * the swap area if the application is discarding it.
800 * An interface that causes the system to free clean pages and flush
801 * dirty pages is already available as msync(MS_INVALIDATE).
803 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
804 unsigned long start, unsigned long end)
806 zap_page_range_single(vma, start, end - start, NULL);
810 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
815 if (!is_vm_hugetlb_page(vma)) {
816 unsigned int forbidden = VM_PFNMAP;
818 if (behavior != MADV_DONTNEED_LOCKED)
819 forbidden |= VM_LOCKED;
821 return !(vma->vm_flags & forbidden);
824 if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
826 if (start & ~huge_page_mask(hstate_vma(vma)))
830 * Madvise callers expect the length to be rounded up to PAGE_SIZE
831 * boundaries, and may be unaware that this VMA uses huge pages.
832 * Avoid unexpected data loss by rounding down the number of
835 *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
840 static long madvise_dontneed_free(struct vm_area_struct *vma,
841 struct vm_area_struct **prev,
842 unsigned long start, unsigned long end,
845 struct mm_struct *mm = vma->vm_mm;
848 if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
854 if (!userfaultfd_remove(vma, start, end)) {
855 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
858 vma = vma_lookup(mm, start);
862 * Potential end adjustment for hugetlb vma is OK as
863 * the check below keeps end within vma.
865 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
868 if (end > vma->vm_end) {
870 * Don't fail if end > vma->vm_end. If the old
871 * vma was split while the mmap_lock was
872 * released the effect of the concurrent
873 * operation may not cause madvise() to
874 * have an undefined result. There may be an
875 * adjacent next vma that we'll walk
876 * next. userfaultfd_remove() will generate an
877 * UFFD_EVENT_REMOVE repetition on the
878 * end-vma->vm_end range, but the manager can
879 * handle a repetition fine.
883 VM_WARN_ON(start >= end);
886 if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
887 return madvise_dontneed_single_vma(vma, start, end);
888 else if (behavior == MADV_FREE)
889 return madvise_free_single_vma(vma, start, end);
894 static long madvise_populate(struct vm_area_struct *vma,
895 struct vm_area_struct **prev,
896 unsigned long start, unsigned long end,
899 const bool write = behavior == MADV_POPULATE_WRITE;
900 struct mm_struct *mm = vma->vm_mm;
901 unsigned long tmp_end;
907 while (start < end) {
909 * We might have temporarily dropped the lock. For example,
910 * our VMA might have been split.
912 if (!vma || start >= vma->vm_end) {
913 vma = vma_lookup(mm, start);
918 tmp_end = min_t(unsigned long, end, vma->vm_end);
919 /* Populate (prefault) page tables readable/writable. */
920 pages = faultin_vma_page_range(vma, start, tmp_end, write,
932 case -EINVAL: /* Incompatible mappings / permissions. */
936 case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
939 pr_warn_once("%s: unhandled return value: %ld\n",
946 start += pages * PAGE_SIZE;
952 * Application wants to free up the pages and associated backing store.
953 * This is effectively punching a hole into the middle of a file.
955 static long madvise_remove(struct vm_area_struct *vma,
956 struct vm_area_struct **prev,
957 unsigned long start, unsigned long end)
962 struct mm_struct *mm = vma->vm_mm;
964 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
966 if (vma->vm_flags & VM_LOCKED)
971 if (!f || !f->f_mapping || !f->f_mapping->host) {
975 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
978 offset = (loff_t)(start - vma->vm_start)
979 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
982 * Filesystem's fallocate may need to take i_rwsem. We need to
983 * explicitly grab a reference because the vma (and hence the
984 * vma's reference to the file) can go away as soon as we drop
988 if (userfaultfd_remove(vma, start, end)) {
989 /* mmap_lock was not released by userfaultfd_remove() */
990 mmap_read_unlock(mm);
992 error = vfs_fallocate(f,
993 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
994 offset, end - start);
1001 * Apply an madvise behavior to a region of a vma. madvise_update_vma
1002 * will handle splitting a vm area into separate areas, each area with its own
1005 static int madvise_vma_behavior(struct vm_area_struct *vma,
1006 struct vm_area_struct **prev,
1007 unsigned long start, unsigned long end,
1008 unsigned long behavior)
1011 struct anon_vma_name *anon_name;
1012 unsigned long new_flags = vma->vm_flags;
1016 return madvise_remove(vma, prev, start, end);
1018 return madvise_willneed(vma, prev, start, end);
1020 return madvise_cold(vma, prev, start, end);
1022 return madvise_pageout(vma, prev, start, end);
1025 case MADV_DONTNEED_LOCKED:
1026 return madvise_dontneed_free(vma, prev, start, end, behavior);
1027 case MADV_POPULATE_READ:
1028 case MADV_POPULATE_WRITE:
1029 return madvise_populate(vma, prev, start, end, behavior);
1031 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1033 case MADV_SEQUENTIAL:
1034 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1037 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1040 new_flags |= VM_DONTCOPY;
1043 if (vma->vm_flags & VM_IO)
1045 new_flags &= ~VM_DONTCOPY;
1047 case MADV_WIPEONFORK:
1048 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1049 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1051 new_flags |= VM_WIPEONFORK;
1053 case MADV_KEEPONFORK:
1054 new_flags &= ~VM_WIPEONFORK;
1057 new_flags |= VM_DONTDUMP;
1060 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1062 new_flags &= ~VM_DONTDUMP;
1064 case MADV_MERGEABLE:
1065 case MADV_UNMERGEABLE:
1066 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1071 case MADV_NOHUGEPAGE:
1072 error = hugepage_madvise(vma, &new_flags, behavior);
1077 return madvise_collapse(vma, prev, start, end);
1080 anon_name = anon_vma_name(vma);
1081 anon_vma_name_get(anon_name);
1082 error = madvise_update_vma(vma, prev, start, end, new_flags,
1084 anon_vma_name_put(anon_name);
1088 * madvise() returns EAGAIN if kernel resources, such as
1089 * slab, are temporarily unavailable.
1091 if (error == -ENOMEM)
1096 #ifdef CONFIG_MEMORY_FAILURE
1098 * Error injection support for memory error handling.
1100 static int madvise_inject_error(int behavior,
1101 unsigned long start, unsigned long end)
1105 if (!capable(CAP_SYS_ADMIN))
1109 for (; start < end; start += size) {
1114 ret = get_user_pages_fast(start, 1, 0, &page);
1117 pfn = page_to_pfn(page);
1120 * When soft offlining hugepages, after migrating the page
1121 * we dissolve it, therefore in the second loop "page" will
1122 * no longer be a compound page.
1124 size = page_size(compound_head(page));
1126 if (behavior == MADV_SOFT_OFFLINE) {
1127 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1129 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1131 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1133 ret = memory_failure(pfn, MF_COUNT_INCREASED | MF_SW_SIMULATED);
1134 if (ret == -EOPNOTSUPP)
1147 madvise_behavior_valid(int behavior)
1153 case MADV_SEQUENTIAL:
1158 case MADV_DONTNEED_LOCKED:
1162 case MADV_POPULATE_READ:
1163 case MADV_POPULATE_WRITE:
1165 case MADV_MERGEABLE:
1166 case MADV_UNMERGEABLE:
1168 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1170 case MADV_NOHUGEPAGE:
1175 case MADV_WIPEONFORK:
1176 case MADV_KEEPONFORK:
1177 #ifdef CONFIG_MEMORY_FAILURE
1178 case MADV_SOFT_OFFLINE:
1188 static bool process_madvise_behavior_valid(int behavior)
1202 * Walk the vmas in range [start,end), and call the visit function on each one.
1203 * The visit function will get start and end parameters that cover the overlap
1204 * between the current vma and the original range. Any unmapped regions in the
1205 * original range will result in this function returning -ENOMEM while still
1206 * calling the visit function on all of the existing vmas in the range.
1207 * Must be called with the mmap_lock held for reading or writing.
1210 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1211 unsigned long end, unsigned long arg,
1212 int (*visit)(struct vm_area_struct *vma,
1213 struct vm_area_struct **prev, unsigned long start,
1214 unsigned long end, unsigned long arg))
1216 struct vm_area_struct *vma;
1217 struct vm_area_struct *prev;
1219 int unmapped_error = 0;
1222 * If the interval [start,end) covers some unmapped address
1223 * ranges, just ignore them, but return -ENOMEM at the end.
1224 * - different from the way of handling in mlock etc.
1226 vma = find_vma_prev(mm, start, &prev);
1227 if (vma && start > vma->vm_start)
1233 /* Still start < end. */
1237 /* Here start < (end|vma->vm_end). */
1238 if (start < vma->vm_start) {
1239 unmapped_error = -ENOMEM;
1240 start = vma->vm_start;
1245 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1250 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1251 error = visit(vma, &prev, start, tmp, arg);
1255 if (prev && start < prev->vm_end)
1256 start = prev->vm_end;
1260 vma = find_vma(mm, prev->vm_end);
1261 else /* madvise_remove dropped mmap_lock */
1262 vma = find_vma(mm, start);
1265 return unmapped_error;
1268 #ifdef CONFIG_ANON_VMA_NAME
1269 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1270 struct vm_area_struct **prev,
1271 unsigned long start, unsigned long end,
1272 unsigned long anon_name)
1276 /* Only anonymous mappings can be named */
1277 if (vma->vm_file && !vma_is_anon_shmem(vma))
1280 error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1281 (struct anon_vma_name *)anon_name);
1284 * madvise() returns EAGAIN if kernel resources, such as
1285 * slab, are temporarily unavailable.
1287 if (error == -ENOMEM)
1292 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1293 unsigned long len_in, struct anon_vma_name *anon_name)
1298 if (start & ~PAGE_MASK)
1300 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1302 /* Check to see whether len was rounded up from small -ve to zero */
1313 return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1314 madvise_vma_anon_name);
1316 #endif /* CONFIG_ANON_VMA_NAME */
1318 * The madvise(2) system call.
1320 * Applications can use madvise() to advise the kernel how it should
1321 * handle paging I/O in this VM area. The idea is to help the kernel
1322 * use appropriate read-ahead and caching techniques. The information
1323 * provided is advisory only, and can be safely disregarded by the
1324 * kernel without affecting the correct operation of the application.
1327 * MADV_NORMAL - the default behavior is to read clusters. This
1328 * results in some read-ahead and read-behind.
1329 * MADV_RANDOM - the system should read the minimum amount of data
1330 * on any access, since it is unlikely that the appli-
1331 * cation will need more than what it asks for.
1332 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1333 * once, so they can be aggressively read ahead, and
1334 * can be freed soon after they are accessed.
1335 * MADV_WILLNEED - the application is notifying the system to read
1337 * MADV_DONTNEED - the application is finished with the given range,
1338 * so the kernel can free resources associated with it.
1339 * MADV_FREE - the application marks pages in the given range as lazy free,
1340 * where actual purges are postponed until memory pressure happens.
1341 * MADV_REMOVE - the application wants to free up the given range of
1342 * pages and associated backing store.
1343 * MADV_DONTFORK - omit this area from child's address space when forking:
1344 * typically, to avoid COWing pages pinned by get_user_pages().
1345 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1346 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1347 * range after a fork.
1348 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1349 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1350 * were corrupted by unrecoverable hardware memory failure.
1351 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1352 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1353 * this area with pages of identical content from other such areas.
1354 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1355 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1356 * huge pages in the future. Existing pages might be coalesced and
1357 * new pages might be allocated as THP.
1358 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1359 * transparent huge pages so the existing pages will not be
1360 * coalesced into THP and new pages will not be allocated as THP.
1361 * MADV_COLLAPSE - synchronously coalesce pages into new THP.
1362 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1363 * from being included in its core dump.
1364 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1365 * MADV_COLD - the application is not expected to use this memory soon,
1366 * deactivate pages in this range so that they can be reclaimed
1367 * easily if memory pressure happens.
1368 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1369 * page out the pages in this range immediately.
1370 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1371 * triggering read faults if required
1372 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1373 * triggering write faults if required
1377 * -EINVAL - start + len < 0, start is not page-aligned,
1378 * "behavior" is not a valid value, or application
1379 * is attempting to release locked or shared pages,
1380 * or the specified address range includes file, Huge TLB,
1381 * MAP_SHARED or VMPFNMAP range.
1382 * -ENOMEM - addresses in the specified range are not currently
1383 * mapped, or are outside the AS of the process.
1384 * -EIO - an I/O error occurred while paging in data.
1385 * -EBADF - map exists, but area maps something that isn't a file.
1386 * -EAGAIN - a kernel resource was temporarily unavailable.
1388 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1394 struct blk_plug plug;
1396 if (!madvise_behavior_valid(behavior))
1399 if (!PAGE_ALIGNED(start))
1401 len = PAGE_ALIGN(len_in);
1403 /* Check to see whether len was rounded up from small -ve to zero */
1414 #ifdef CONFIG_MEMORY_FAILURE
1415 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1416 return madvise_inject_error(behavior, start, start + len_in);
1419 write = madvise_need_mmap_write(behavior);
1421 if (mmap_write_lock_killable(mm))
1427 start = untagged_addr_remote(mm, start);
1430 blk_start_plug(&plug);
1431 error = madvise_walk_vmas(mm, start, end, behavior,
1432 madvise_vma_behavior);
1433 blk_finish_plug(&plug);
1435 mmap_write_unlock(mm);
1437 mmap_read_unlock(mm);
1442 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1444 return do_madvise(current->mm, start, len_in, behavior);
1447 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1448 size_t, vlen, int, behavior, unsigned int, flags)
1451 struct iovec iovstack[UIO_FASTIOV];
1452 struct iovec *iov = iovstack;
1453 struct iov_iter iter;
1454 struct task_struct *task;
1455 struct mm_struct *mm;
1457 unsigned int f_flags;
1464 ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1468 task = pidfd_get_task(pidfd, &f_flags);
1470 ret = PTR_ERR(task);
1474 if (!process_madvise_behavior_valid(behavior)) {
1479 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1480 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1481 if (IS_ERR_OR_NULL(mm)) {
1482 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1487 * Require CAP_SYS_NICE for influencing process performance. Note that
1488 * only non-destructive hints are currently supported.
1490 if (!capable(CAP_SYS_NICE)) {
1495 total_len = iov_iter_count(&iter);
1497 while (iov_iter_count(&iter)) {
1498 ret = do_madvise(mm, (unsigned long)iter_iov_addr(&iter),
1499 iter_iov_len(&iter), behavior);
1502 iov_iter_advance(&iter, iter_iov_len(&iter));
1505 ret = (total_len - iov_iter_count(&iter)) ? : ret;
1510 put_task_struct(task);