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>
41 * Maximum number of attempts we make to install guard pages before we give up
42 * and return -ERESTARTNOINTR to have userspace try again.
44 #define MAX_MADVISE_GUARD_RETRIES 3
46 struct madvise_walk_private {
47 struct mmu_gather *tlb;
52 * Any behaviour which results in changes to the vma->vm_flags needs to
53 * take mmap_lock for writing. Others, which simply traverse vmas, need
54 * to only take it for reading.
56 static int madvise_need_mmap_write(int behavior)
62 case MADV_DONTNEED_LOCKED:
66 case MADV_POPULATE_READ:
67 case MADV_POPULATE_WRITE:
69 case MADV_GUARD_INSTALL:
70 case MADV_GUARD_REMOVE:
73 /* be safe, default to 1. list exceptions explicitly */
78 #ifdef CONFIG_ANON_VMA_NAME
79 struct anon_vma_name *anon_vma_name_alloc(const char *name)
81 struct anon_vma_name *anon_name;
84 /* Add 1 for NUL terminator at the end of the anon_name->name */
85 count = strlen(name) + 1;
86 anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
88 kref_init(&anon_name->kref);
89 memcpy(anon_name->name, name, count);
95 void anon_vma_name_free(struct kref *kref)
97 struct anon_vma_name *anon_name =
98 container_of(kref, struct anon_vma_name, kref);
102 struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
104 mmap_assert_locked(vma->vm_mm);
106 return vma->anon_name;
109 /* mmap_lock should be write-locked */
110 static int replace_anon_vma_name(struct vm_area_struct *vma,
111 struct anon_vma_name *anon_name)
113 struct anon_vma_name *orig_name = anon_vma_name(vma);
116 vma->anon_name = NULL;
117 anon_vma_name_put(orig_name);
121 if (anon_vma_name_eq(orig_name, anon_name))
124 vma->anon_name = anon_vma_name_reuse(anon_name);
125 anon_vma_name_put(orig_name);
129 #else /* CONFIG_ANON_VMA_NAME */
130 static int replace_anon_vma_name(struct vm_area_struct *vma,
131 struct anon_vma_name *anon_name)
138 #endif /* CONFIG_ANON_VMA_NAME */
140 * Update the vm_flags on region of a vma, splitting it or merging it as
141 * necessary. Must be called with mmap_lock held for writing;
142 * Caller should ensure anon_name stability by raising its refcount even when
143 * anon_name belongs to a valid vma because this function might free that vma.
145 static int madvise_update_vma(struct vm_area_struct *vma,
146 struct vm_area_struct **prev, unsigned long start,
147 unsigned long end, unsigned long new_flags,
148 struct anon_vma_name *anon_name)
150 struct mm_struct *mm = vma->vm_mm;
152 VMA_ITERATOR(vmi, mm, start);
154 if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
159 vma = vma_modify_flags_name(&vmi, *prev, vma, start, end, new_flags,
166 /* vm_flags is protected by the mmap_lock held in write mode. */
167 vma_start_write(vma);
168 vm_flags_reset(vma, new_flags);
169 if (!vma->vm_file || vma_is_anon_shmem(vma)) {
170 error = replace_anon_vma_name(vma, anon_name);
179 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
180 unsigned long end, struct mm_walk *walk)
182 struct vm_area_struct *vma = walk->private;
183 struct swap_iocb *splug = NULL;
188 for (addr = start; addr < end; addr += PAGE_SIZE) {
194 ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
199 pte = ptep_get(ptep);
200 if (!is_swap_pte(pte))
202 entry = pte_to_swp_entry(pte);
203 if (unlikely(non_swap_entry(entry)))
206 pte_unmap_unlock(ptep, ptl);
209 folio = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
216 pte_unmap_unlock(ptep, ptl);
217 swap_read_unplug(splug);
223 static const struct mm_walk_ops swapin_walk_ops = {
224 .pmd_entry = swapin_walk_pmd_entry,
225 .walk_lock = PGWALK_RDLOCK,
228 static void shmem_swapin_range(struct vm_area_struct *vma,
229 unsigned long start, unsigned long end,
230 struct address_space *mapping)
232 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
233 pgoff_t end_index = linear_page_index(vma, end) - 1;
235 struct swap_iocb *splug = NULL;
238 xas_for_each(&xas, folio, end_index) {
242 if (!xa_is_value(folio))
244 entry = radix_to_swp_entry(folio);
245 /* There might be swapin error entries in shmem mapping. */
246 if (non_swap_entry(entry))
249 addr = vma->vm_start +
250 ((xas.xa_index - vma->vm_pgoff) << PAGE_SHIFT);
254 folio = read_swap_cache_async(entry, mapping_gfp_mask(mapping),
262 swap_read_unplug(splug);
264 #endif /* CONFIG_SWAP */
267 * Schedule all required I/O operations. Do not wait for completion.
269 static long madvise_willneed(struct vm_area_struct *vma,
270 struct vm_area_struct **prev,
271 unsigned long start, unsigned long end)
273 struct mm_struct *mm = vma->vm_mm;
274 struct file *file = vma->vm_file;
280 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
281 lru_add_drain(); /* Push any new pages onto the LRU now */
285 if (shmem_mapping(file->f_mapping)) {
286 shmem_swapin_range(vma, start, end, file->f_mapping);
287 lru_add_drain(); /* Push any new pages onto the LRU now */
295 if (IS_DAX(file_inode(file))) {
296 /* no bad return value, but ignore advice */
301 * Filesystem's fadvise may need to take various locks. We need to
302 * explicitly grab a reference because the vma (and hence the
303 * vma's reference to the file) can go away as soon as we drop
306 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
308 offset = (loff_t)(start - vma->vm_start)
309 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
310 mmap_read_unlock(mm);
311 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
317 static inline bool can_do_file_pageout(struct vm_area_struct *vma)
322 * paging out pagecache only for non-anonymous mappings that correspond
323 * to the files the calling process could (if tried) open for writing;
324 * otherwise we'd be including shared non-exclusive mappings, which
325 * opens a side channel.
327 return inode_owner_or_capable(&nop_mnt_idmap,
328 file_inode(vma->vm_file)) ||
329 file_permission(vma->vm_file, MAY_WRITE) == 0;
332 static inline int madvise_folio_pte_batch(unsigned long addr, unsigned long end,
333 struct folio *folio, pte_t *ptep,
334 pte_t pte, bool *any_young,
337 const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
338 int max_nr = (end - addr) / PAGE_SIZE;
340 return folio_pte_batch(folio, addr, ptep, pte, max_nr, fpb_flags, NULL,
341 any_young, any_dirty);
344 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
345 unsigned long addr, unsigned long end,
346 struct mm_walk *walk)
348 struct madvise_walk_private *private = walk->private;
349 struct mmu_gather *tlb = private->tlb;
350 bool pageout = private->pageout;
351 struct mm_struct *mm = tlb->mm;
352 struct vm_area_struct *vma = walk->vma;
353 pte_t *start_pte, *pte, ptent;
355 struct folio *folio = NULL;
356 LIST_HEAD(folio_list);
357 bool pageout_anon_only_filter;
358 unsigned int batch_count = 0;
361 if (fatal_signal_pending(current))
364 pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
365 !can_do_file_pageout(vma);
367 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
368 if (pmd_trans_huge(*pmd)) {
370 unsigned long next = pmd_addr_end(addr, end);
372 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
373 ptl = pmd_trans_huge_lock(pmd, vma);
378 if (is_huge_zero_pmd(orig_pmd))
381 if (unlikely(!pmd_present(orig_pmd))) {
382 VM_BUG_ON(thp_migration_supported() &&
383 !is_pmd_migration_entry(orig_pmd));
387 folio = pmd_folio(orig_pmd);
389 /* Do not interfere with other mappings of this folio */
390 if (folio_likely_mapped_shared(folio))
393 if (pageout_anon_only_filter && !folio_test_anon(folio))
396 if (next - addr != HPAGE_PMD_SIZE) {
402 err = split_folio(folio);
410 if (!pageout && pmd_young(orig_pmd)) {
411 pmdp_invalidate(vma, addr, pmd);
412 orig_pmd = pmd_mkold(orig_pmd);
414 set_pmd_at(mm, addr, pmd, orig_pmd);
415 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
418 folio_clear_referenced(folio);
419 folio_test_clear_young(folio);
420 if (folio_test_active(folio))
421 folio_set_workingset(folio);
423 if (folio_isolate_lru(folio)) {
424 if (folio_test_unevictable(folio))
425 folio_putback_lru(folio);
427 list_add(&folio->lru, &folio_list);
430 folio_deactivate(folio);
434 reclaim_pages(&folio_list);
440 tlb_change_page_size(tlb, PAGE_SIZE);
442 start_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
445 flush_tlb_batched_pending(mm);
446 arch_enter_lazy_mmu_mode();
447 for (; addr < end; pte += nr, addr += nr * PAGE_SIZE) {
449 ptent = ptep_get(pte);
451 if (++batch_count == SWAP_CLUSTER_MAX) {
453 if (need_resched()) {
454 arch_leave_lazy_mmu_mode();
455 pte_unmap_unlock(start_pte, ptl);
464 if (!pte_present(ptent))
467 folio = vm_normal_folio(vma, addr, ptent);
468 if (!folio || folio_is_zone_device(folio))
472 * If we encounter a large folio, only split it if it is not
473 * fully mapped within the range we are operating on. Otherwise
474 * leave it as is so that it can be swapped out whole. If we
475 * fail to split a folio, leave it in place and advance to the
476 * next pte in the range.
478 if (folio_test_large(folio)) {
481 nr = madvise_folio_pte_batch(addr, end, folio, pte,
482 ptent, &any_young, NULL);
484 ptent = pte_mkyoung(ptent);
486 if (nr < folio_nr_pages(folio)) {
489 if (folio_likely_mapped_shared(folio))
491 if (pageout_anon_only_filter && !folio_test_anon(folio))
493 if (!folio_trylock(folio))
496 arch_leave_lazy_mmu_mode();
497 pte_unmap_unlock(start_pte, ptl);
499 err = split_folio(folio);
503 pte_offset_map_lock(mm, pmd, addr, &ptl);
506 arch_enter_lazy_mmu_mode();
514 * Do not interfere with other mappings of this folio and
515 * non-LRU folio. If we have a large folio at this point, we
516 * know it is fully mapped so if its mapcount is the same as its
517 * number of pages, it must be exclusive.
519 if (!folio_test_lru(folio) ||
520 folio_mapcount(folio) != folio_nr_pages(folio))
523 if (pageout_anon_only_filter && !folio_test_anon(folio))
526 if (!pageout && pte_young(ptent)) {
527 clear_young_dirty_ptes(vma, addr, pte, nr,
529 tlb_remove_tlb_entries(tlb, pte, nr, addr);
533 * We are deactivating a folio for accelerating reclaiming.
534 * VM couldn't reclaim the folio unless we clear PG_young.
535 * As a side effect, it makes confuse idle-page tracking
536 * because they will miss recent referenced history.
538 folio_clear_referenced(folio);
539 folio_test_clear_young(folio);
540 if (folio_test_active(folio))
541 folio_set_workingset(folio);
543 if (folio_isolate_lru(folio)) {
544 if (folio_test_unevictable(folio))
545 folio_putback_lru(folio);
547 list_add(&folio->lru, &folio_list);
550 folio_deactivate(folio);
554 arch_leave_lazy_mmu_mode();
555 pte_unmap_unlock(start_pte, ptl);
558 reclaim_pages(&folio_list);
564 static const struct mm_walk_ops cold_walk_ops = {
565 .pmd_entry = madvise_cold_or_pageout_pte_range,
566 .walk_lock = PGWALK_RDLOCK,
569 static void madvise_cold_page_range(struct mmu_gather *tlb,
570 struct vm_area_struct *vma,
571 unsigned long addr, unsigned long end)
573 struct madvise_walk_private walk_private = {
578 tlb_start_vma(tlb, vma);
579 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
580 tlb_end_vma(tlb, vma);
583 static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
585 return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
588 static long madvise_cold(struct vm_area_struct *vma,
589 struct vm_area_struct **prev,
590 unsigned long start_addr, unsigned long end_addr)
592 struct mm_struct *mm = vma->vm_mm;
593 struct mmu_gather tlb;
596 if (!can_madv_lru_vma(vma))
600 tlb_gather_mmu(&tlb, mm);
601 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
602 tlb_finish_mmu(&tlb);
607 static void madvise_pageout_page_range(struct mmu_gather *tlb,
608 struct vm_area_struct *vma,
609 unsigned long addr, unsigned long end)
611 struct madvise_walk_private walk_private = {
616 tlb_start_vma(tlb, vma);
617 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
618 tlb_end_vma(tlb, vma);
621 static long madvise_pageout(struct vm_area_struct *vma,
622 struct vm_area_struct **prev,
623 unsigned long start_addr, unsigned long end_addr)
625 struct mm_struct *mm = vma->vm_mm;
626 struct mmu_gather tlb;
629 if (!can_madv_lru_vma(vma))
633 * If the VMA belongs to a private file mapping, there can be private
634 * dirty pages which can be paged out if even this process is neither
635 * owner nor write capable of the file. We allow private file mappings
636 * further to pageout dirty anon pages.
638 if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
639 (vma->vm_flags & VM_MAYSHARE)))
643 tlb_gather_mmu(&tlb, mm);
644 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
645 tlb_finish_mmu(&tlb);
650 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
651 unsigned long end, struct mm_walk *walk)
654 const cydp_t cydp_flags = CYDP_CLEAR_YOUNG | CYDP_CLEAR_DIRTY;
655 struct mmu_gather *tlb = walk->private;
656 struct mm_struct *mm = tlb->mm;
657 struct vm_area_struct *vma = walk->vma;
659 pte_t *start_pte, *pte, ptent;
665 next = pmd_addr_end(addr, end);
666 if (pmd_trans_huge(*pmd))
667 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
670 tlb_change_page_size(tlb, PAGE_SIZE);
671 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
674 flush_tlb_batched_pending(mm);
675 arch_enter_lazy_mmu_mode();
676 for (; addr != end; pte += nr, addr += PAGE_SIZE * nr) {
678 ptent = ptep_get(pte);
683 * If the pte has swp_entry, just clear page table to
684 * prevent swap-in which is more expensive rather than
685 * (page allocation + zeroing).
687 if (!pte_present(ptent)) {
690 entry = pte_to_swp_entry(ptent);
691 if (!non_swap_entry(entry)) {
692 max_nr = (end - addr) / PAGE_SIZE;
693 nr = swap_pte_batch(pte, max_nr, ptent);
695 free_swap_and_cache_nr(entry, nr);
696 clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
697 } else if (is_hwpoison_entry(entry) ||
698 is_poisoned_swp_entry(entry)) {
699 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
704 folio = vm_normal_folio(vma, addr, ptent);
705 if (!folio || folio_is_zone_device(folio))
709 * If we encounter a large folio, only split it if it is not
710 * fully mapped within the range we are operating on. Otherwise
711 * leave it as is so that it can be marked as lazyfree. If we
712 * fail to split a folio, leave it in place and advance to the
713 * next pte in the range.
715 if (folio_test_large(folio)) {
716 bool any_young, any_dirty;
718 nr = madvise_folio_pte_batch(addr, end, folio, pte,
719 ptent, &any_young, &any_dirty);
721 if (nr < folio_nr_pages(folio)) {
724 if (folio_likely_mapped_shared(folio))
726 if (!folio_trylock(folio))
729 arch_leave_lazy_mmu_mode();
730 pte_unmap_unlock(start_pte, ptl);
732 err = split_folio(folio);
735 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
739 arch_enter_lazy_mmu_mode();
746 ptent = pte_mkyoung(ptent);
748 ptent = pte_mkdirty(ptent);
751 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
752 if (!folio_trylock(folio))
755 * If we have a large folio at this point, we know it is
756 * fully mapped so if its mapcount is the same as its
757 * number of pages, it must be exclusive.
759 if (folio_mapcount(folio) != folio_nr_pages(folio)) {
764 if (folio_test_swapcache(folio) &&
765 !folio_free_swap(folio)) {
770 folio_clear_dirty(folio);
774 if (pte_young(ptent) || pte_dirty(ptent)) {
775 clear_young_dirty_ptes(vma, addr, pte, nr, cydp_flags);
776 tlb_remove_tlb_entries(tlb, pte, nr, addr);
778 folio_mark_lazyfree(folio);
782 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
784 arch_leave_lazy_mmu_mode();
785 pte_unmap_unlock(start_pte, ptl);
792 static const struct mm_walk_ops madvise_free_walk_ops = {
793 .pmd_entry = madvise_free_pte_range,
794 .walk_lock = PGWALK_RDLOCK,
797 static int madvise_free_single_vma(struct vm_area_struct *vma,
798 unsigned long start_addr, unsigned long end_addr)
800 struct mm_struct *mm = vma->vm_mm;
801 struct mmu_notifier_range range;
802 struct mmu_gather tlb;
804 /* MADV_FREE works for only anon vma at the moment */
805 if (!vma_is_anonymous(vma))
808 range.start = max(vma->vm_start, start_addr);
809 if (range.start >= vma->vm_end)
811 range.end = min(vma->vm_end, end_addr);
812 if (range.end <= vma->vm_start)
814 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
815 range.start, range.end);
818 tlb_gather_mmu(&tlb, mm);
819 update_hiwater_rss(mm);
821 mmu_notifier_invalidate_range_start(&range);
822 tlb_start_vma(&tlb, vma);
823 walk_page_range(vma->vm_mm, range.start, range.end,
824 &madvise_free_walk_ops, &tlb);
825 tlb_end_vma(&tlb, vma);
826 mmu_notifier_invalidate_range_end(&range);
827 tlb_finish_mmu(&tlb);
833 * Application no longer needs these pages. If the pages are dirty,
834 * it's OK to just throw them away. The app will be more careful about
835 * data it wants to keep. Be sure to free swap resources too. The
836 * zap_page_range_single call sets things up for shrink_active_list to actually
837 * free these pages later if no one else has touched them in the meantime,
838 * although we could add these pages to a global reuse list for
839 * shrink_active_list to pick up before reclaiming other pages.
841 * NB: This interface discards data rather than pushes it out to swap,
842 * as some implementations do. This has performance implications for
843 * applications like large transactional databases which want to discard
844 * pages in anonymous maps after committing to backing store the data
845 * that was kept in them. There is no reason to write this data out to
846 * the swap area if the application is discarding it.
848 * An interface that causes the system to free clean pages and flush
849 * dirty pages is already available as msync(MS_INVALIDATE).
851 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
852 unsigned long start, unsigned long end)
854 struct zap_details details = {
859 zap_page_range_single(vma, start, end - start, &details);
863 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
868 if (!is_vm_hugetlb_page(vma)) {
869 unsigned int forbidden = VM_PFNMAP;
871 if (behavior != MADV_DONTNEED_LOCKED)
872 forbidden |= VM_LOCKED;
874 return !(vma->vm_flags & forbidden);
877 if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
879 if (start & ~huge_page_mask(hstate_vma(vma)))
883 * Madvise callers expect the length to be rounded up to PAGE_SIZE
884 * boundaries, and may be unaware that this VMA uses huge pages.
885 * Avoid unexpected data loss by rounding down the number of
888 *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
893 static long madvise_dontneed_free(struct vm_area_struct *vma,
894 struct vm_area_struct **prev,
895 unsigned long start, unsigned long end,
898 struct mm_struct *mm = vma->vm_mm;
901 if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
907 if (!userfaultfd_remove(vma, start, end)) {
908 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
911 vma = vma_lookup(mm, start);
915 * Potential end adjustment for hugetlb vma is OK as
916 * the check below keeps end within vma.
918 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
921 if (end > vma->vm_end) {
923 * Don't fail if end > vma->vm_end. If the old
924 * vma was split while the mmap_lock was
925 * released the effect of the concurrent
926 * operation may not cause madvise() to
927 * have an undefined result. There may be an
928 * adjacent next vma that we'll walk
929 * next. userfaultfd_remove() will generate an
930 * UFFD_EVENT_REMOVE repetition on the
931 * end-vma->vm_end range, but the manager can
932 * handle a repetition fine.
936 VM_WARN_ON(start >= end);
939 if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
940 return madvise_dontneed_single_vma(vma, start, end);
941 else if (behavior == MADV_FREE)
942 return madvise_free_single_vma(vma, start, end);
947 static long madvise_populate(struct mm_struct *mm, unsigned long start,
948 unsigned long end, int behavior)
950 const bool write = behavior == MADV_POPULATE_WRITE;
954 while (start < end) {
955 /* Populate (prefault) page tables readable/writable. */
956 pages = faultin_page_range(mm, start, end, write, &locked);
965 case -EINVAL: /* Incompatible mappings / permissions. */
969 case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
972 pr_warn_once("%s: unhandled return value: %ld\n",
975 case -ENOMEM: /* No VMA or out of memory. */
979 start += pages * PAGE_SIZE;
985 * Application wants to free up the pages and associated backing store.
986 * This is effectively punching a hole into the middle of a file.
988 static long madvise_remove(struct vm_area_struct *vma,
989 struct vm_area_struct **prev,
990 unsigned long start, unsigned long end)
995 struct mm_struct *mm = vma->vm_mm;
997 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
999 if (vma->vm_flags & VM_LOCKED)
1004 if (!f || !f->f_mapping || !f->f_mapping->host) {
1008 if (!vma_is_shared_maywrite(vma))
1011 offset = (loff_t)(start - vma->vm_start)
1012 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
1015 * Filesystem's fallocate may need to take i_rwsem. We need to
1016 * explicitly grab a reference because the vma (and hence the
1017 * vma's reference to the file) can go away as soon as we drop
1021 if (userfaultfd_remove(vma, start, end)) {
1022 /* mmap_lock was not released by userfaultfd_remove() */
1023 mmap_read_unlock(mm);
1025 error = vfs_fallocate(f,
1026 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1027 offset, end - start);
1033 static bool is_valid_guard_vma(struct vm_area_struct *vma, bool allow_locked)
1035 vm_flags_t disallowed = VM_SPECIAL | VM_HUGETLB;
1038 * A user could lock after setting a guard range but that's fine, as
1039 * they'd not be able to fault in. The issue arises when we try to zap
1040 * existing locked VMAs. We don't want to do that.
1043 disallowed |= VM_LOCKED;
1045 if (!vma_is_anonymous(vma))
1048 if ((vma->vm_flags & (VM_MAYWRITE | disallowed)) != VM_MAYWRITE)
1054 static bool is_guard_pte_marker(pte_t ptent)
1056 return is_pte_marker(ptent) &&
1057 is_guard_swp_entry(pte_to_swp_entry(ptent));
1060 static int guard_install_pud_entry(pud_t *pud, unsigned long addr,
1061 unsigned long next, struct mm_walk *walk)
1063 pud_t pudval = pudp_get(pud);
1065 /* If huge return >0 so we abort the operation + zap. */
1066 return pud_trans_huge(pudval) || pud_devmap(pudval);
1069 static int guard_install_pmd_entry(pmd_t *pmd, unsigned long addr,
1070 unsigned long next, struct mm_walk *walk)
1072 pmd_t pmdval = pmdp_get(pmd);
1074 /* If huge return >0 so we abort the operation + zap. */
1075 return pmd_trans_huge(pmdval) || pmd_devmap(pmdval);
1078 static int guard_install_pte_entry(pte_t *pte, unsigned long addr,
1079 unsigned long next, struct mm_walk *walk)
1081 pte_t pteval = ptep_get(pte);
1082 unsigned long *nr_pages = (unsigned long *)walk->private;
1084 /* If there is already a guard page marker, we have nothing to do. */
1085 if (is_guard_pte_marker(pteval)) {
1091 /* If populated return >0 so we abort the operation + zap. */
1095 static int guard_install_set_pte(unsigned long addr, unsigned long next,
1096 pte_t *ptep, struct mm_walk *walk)
1098 unsigned long *nr_pages = (unsigned long *)walk->private;
1100 /* Simply install a PTE marker, this causes segfault on access. */
1101 *ptep = make_pte_marker(PTE_MARKER_GUARD);
1107 static const struct mm_walk_ops guard_install_walk_ops = {
1108 .pud_entry = guard_install_pud_entry,
1109 .pmd_entry = guard_install_pmd_entry,
1110 .pte_entry = guard_install_pte_entry,
1111 .install_pte = guard_install_set_pte,
1112 .walk_lock = PGWALK_RDLOCK,
1115 static long madvise_guard_install(struct vm_area_struct *vma,
1116 struct vm_area_struct **prev,
1117 unsigned long start, unsigned long end)
1123 if (!is_valid_guard_vma(vma, /* allow_locked = */false))
1127 * If we install guard markers, then the range is no longer
1128 * empty from a page table perspective and therefore it's
1129 * appropriate to have an anon_vma.
1131 * This ensures that on fork, we copy page tables correctly.
1133 err = anon_vma_prepare(vma);
1138 * Optimistically try to install the guard marker pages first. If any
1139 * non-guard pages are encountered, give up and zap the range before
1142 * We try a few times before giving up and releasing back to userland to
1143 * loop around, releasing locks in the process to avoid contention. This
1144 * would only happen if there was a great many racing page faults.
1146 * In most cases we should simply install the guard markers immediately
1147 * with no zap or looping.
1149 for (i = 0; i < MAX_MADVISE_GUARD_RETRIES; i++) {
1150 unsigned long nr_pages = 0;
1152 /* Returns < 0 on error, == 0 if success, > 0 if zap needed. */
1153 err = walk_page_range_mm(vma->vm_mm, start, end,
1154 &guard_install_walk_ops, &nr_pages);
1159 unsigned long nr_expected_pages = PHYS_PFN(end - start);
1161 VM_WARN_ON(nr_pages != nr_expected_pages);
1166 * OK some of the range have non-guard pages mapped, zap
1167 * them. This leaves existing guard pages in place.
1169 zap_page_range_single(vma, start, end - start, NULL);
1173 * We were unable to install the guard pages due to being raced by page
1174 * faults. This should not happen ordinarily. We return to userspace and
1175 * immediately retry, relieving lock contention.
1177 return restart_syscall();
1180 static int guard_remove_pud_entry(pud_t *pud, unsigned long addr,
1181 unsigned long next, struct mm_walk *walk)
1183 pud_t pudval = pudp_get(pud);
1185 /* If huge, cannot have guard pages present, so no-op - skip. */
1186 if (pud_trans_huge(pudval) || pud_devmap(pudval))
1187 walk->action = ACTION_CONTINUE;
1192 static int guard_remove_pmd_entry(pmd_t *pmd, unsigned long addr,
1193 unsigned long next, struct mm_walk *walk)
1195 pmd_t pmdval = pmdp_get(pmd);
1197 /* If huge, cannot have guard pages present, so no-op - skip. */
1198 if (pmd_trans_huge(pmdval) || pmd_devmap(pmdval))
1199 walk->action = ACTION_CONTINUE;
1204 static int guard_remove_pte_entry(pte_t *pte, unsigned long addr,
1205 unsigned long next, struct mm_walk *walk)
1207 pte_t ptent = ptep_get(pte);
1209 if (is_guard_pte_marker(ptent)) {
1210 /* Simply clear the PTE marker. */
1211 pte_clear_not_present_full(walk->mm, addr, pte, false);
1212 update_mmu_cache(walk->vma, addr, pte);
1218 static const struct mm_walk_ops guard_remove_walk_ops = {
1219 .pud_entry = guard_remove_pud_entry,
1220 .pmd_entry = guard_remove_pmd_entry,
1221 .pte_entry = guard_remove_pte_entry,
1222 .walk_lock = PGWALK_RDLOCK,
1225 static long madvise_guard_remove(struct vm_area_struct *vma,
1226 struct vm_area_struct **prev,
1227 unsigned long start, unsigned long end)
1231 * We're ok with removing guards in mlock()'d ranges, as this is a
1232 * non-destructive action.
1234 if (!is_valid_guard_vma(vma, /* allow_locked = */true))
1237 return walk_page_range(vma->vm_mm, start, end,
1238 &guard_remove_walk_ops, NULL);
1242 * Apply an madvise behavior to a region of a vma. madvise_update_vma
1243 * will handle splitting a vm area into separate areas, each area with its own
1246 static int madvise_vma_behavior(struct vm_area_struct *vma,
1247 struct vm_area_struct **prev,
1248 unsigned long start, unsigned long end,
1249 unsigned long behavior)
1252 struct anon_vma_name *anon_name;
1253 unsigned long new_flags = vma->vm_flags;
1255 if (unlikely(!can_modify_vma_madv(vma, behavior)))
1260 return madvise_remove(vma, prev, start, end);
1262 return madvise_willneed(vma, prev, start, end);
1264 return madvise_cold(vma, prev, start, end);
1266 return madvise_pageout(vma, prev, start, end);
1269 case MADV_DONTNEED_LOCKED:
1270 return madvise_dontneed_free(vma, prev, start, end, behavior);
1272 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1274 case MADV_SEQUENTIAL:
1275 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1278 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1281 new_flags |= VM_DONTCOPY;
1284 if (vma->vm_flags & VM_IO)
1286 new_flags &= ~VM_DONTCOPY;
1288 case MADV_WIPEONFORK:
1289 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1290 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1292 new_flags |= VM_WIPEONFORK;
1294 case MADV_KEEPONFORK:
1295 if (vma->vm_flags & VM_DROPPABLE)
1297 new_flags &= ~VM_WIPEONFORK;
1300 new_flags |= VM_DONTDUMP;
1303 if ((!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) ||
1304 (vma->vm_flags & VM_DROPPABLE))
1306 new_flags &= ~VM_DONTDUMP;
1308 case MADV_MERGEABLE:
1309 case MADV_UNMERGEABLE:
1310 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1315 case MADV_NOHUGEPAGE:
1316 error = hugepage_madvise(vma, &new_flags, behavior);
1321 return madvise_collapse(vma, prev, start, end);
1322 case MADV_GUARD_INSTALL:
1323 return madvise_guard_install(vma, prev, start, end);
1324 case MADV_GUARD_REMOVE:
1325 return madvise_guard_remove(vma, prev, start, end);
1328 anon_name = anon_vma_name(vma);
1329 anon_vma_name_get(anon_name);
1330 error = madvise_update_vma(vma, prev, start, end, new_flags,
1332 anon_vma_name_put(anon_name);
1336 * madvise() returns EAGAIN if kernel resources, such as
1337 * slab, are temporarily unavailable.
1339 if (error == -ENOMEM)
1344 #ifdef CONFIG_MEMORY_FAILURE
1346 * Error injection support for memory error handling.
1348 static int madvise_inject_error(int behavior,
1349 unsigned long start, unsigned long end)
1353 if (!capable(CAP_SYS_ADMIN))
1357 for (; start < end; start += size) {
1362 ret = get_user_pages_fast(start, 1, 0, &page);
1365 pfn = page_to_pfn(page);
1368 * When soft offlining hugepages, after migrating the page
1369 * we dissolve it, therefore in the second loop "page" will
1370 * no longer be a compound page.
1372 size = page_size(compound_head(page));
1374 if (behavior == MADV_SOFT_OFFLINE) {
1375 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1377 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1379 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1381 ret = memory_failure(pfn, MF_ACTION_REQUIRED | MF_COUNT_INCREASED | MF_SW_SIMULATED);
1382 if (ret == -EOPNOTSUPP)
1395 madvise_behavior_valid(int behavior)
1401 case MADV_SEQUENTIAL:
1406 case MADV_DONTNEED_LOCKED:
1410 case MADV_POPULATE_READ:
1411 case MADV_POPULATE_WRITE:
1413 case MADV_MERGEABLE:
1414 case MADV_UNMERGEABLE:
1416 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1418 case MADV_NOHUGEPAGE:
1423 case MADV_WIPEONFORK:
1424 case MADV_KEEPONFORK:
1425 case MADV_GUARD_INSTALL:
1426 case MADV_GUARD_REMOVE:
1427 #ifdef CONFIG_MEMORY_FAILURE
1428 case MADV_SOFT_OFFLINE:
1438 /* Can we invoke process_madvise() on a remote mm for the specified behavior? */
1439 static bool process_madvise_remote_valid(int behavior)
1453 * Walk the vmas in range [start,end), and call the visit function on each one.
1454 * The visit function will get start and end parameters that cover the overlap
1455 * between the current vma and the original range. Any unmapped regions in the
1456 * original range will result in this function returning -ENOMEM while still
1457 * calling the visit function on all of the existing vmas in the range.
1458 * Must be called with the mmap_lock held for reading or writing.
1461 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1462 unsigned long end, unsigned long arg,
1463 int (*visit)(struct vm_area_struct *vma,
1464 struct vm_area_struct **prev, unsigned long start,
1465 unsigned long end, unsigned long arg))
1467 struct vm_area_struct *vma;
1468 struct vm_area_struct *prev;
1470 int unmapped_error = 0;
1473 * If the interval [start,end) covers some unmapped address
1474 * ranges, just ignore them, but return -ENOMEM at the end.
1475 * - different from the way of handling in mlock etc.
1477 vma = find_vma_prev(mm, start, &prev);
1478 if (vma && start > vma->vm_start)
1484 /* Still start < end. */
1488 /* Here start < (end|vma->vm_end). */
1489 if (start < vma->vm_start) {
1490 unmapped_error = -ENOMEM;
1491 start = vma->vm_start;
1496 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1501 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1502 error = visit(vma, &prev, start, tmp, arg);
1506 if (prev && start < prev->vm_end)
1507 start = prev->vm_end;
1511 vma = find_vma(mm, prev->vm_end);
1512 else /* madvise_remove dropped mmap_lock */
1513 vma = find_vma(mm, start);
1516 return unmapped_error;
1519 #ifdef CONFIG_ANON_VMA_NAME
1520 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1521 struct vm_area_struct **prev,
1522 unsigned long start, unsigned long end,
1523 unsigned long anon_name)
1527 /* Only anonymous mappings can be named */
1528 if (vma->vm_file && !vma_is_anon_shmem(vma))
1531 error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1532 (struct anon_vma_name *)anon_name);
1535 * madvise() returns EAGAIN if kernel resources, such as
1536 * slab, are temporarily unavailable.
1538 if (error == -ENOMEM)
1543 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1544 unsigned long len_in, struct anon_vma_name *anon_name)
1549 if (start & ~PAGE_MASK)
1551 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1553 /* Check to see whether len was rounded up from small -ve to zero */
1564 return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1565 madvise_vma_anon_name);
1567 #endif /* CONFIG_ANON_VMA_NAME */
1569 * The madvise(2) system call.
1571 * Applications can use madvise() to advise the kernel how it should
1572 * handle paging I/O in this VM area. The idea is to help the kernel
1573 * use appropriate read-ahead and caching techniques. The information
1574 * provided is advisory only, and can be safely disregarded by the
1575 * kernel without affecting the correct operation of the application.
1578 * MADV_NORMAL - the default behavior is to read clusters. This
1579 * results in some read-ahead and read-behind.
1580 * MADV_RANDOM - the system should read the minimum amount of data
1581 * on any access, since it is unlikely that the appli-
1582 * cation will need more than what it asks for.
1583 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1584 * once, so they can be aggressively read ahead, and
1585 * can be freed soon after they are accessed.
1586 * MADV_WILLNEED - the application is notifying the system to read
1588 * MADV_DONTNEED - the application is finished with the given range,
1589 * so the kernel can free resources associated with it.
1590 * MADV_FREE - the application marks pages in the given range as lazy free,
1591 * where actual purges are postponed until memory pressure happens.
1592 * MADV_REMOVE - the application wants to free up the given range of
1593 * pages and associated backing store.
1594 * MADV_DONTFORK - omit this area from child's address space when forking:
1595 * typically, to avoid COWing pages pinned by get_user_pages().
1596 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1597 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1598 * range after a fork.
1599 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1600 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1601 * were corrupted by unrecoverable hardware memory failure.
1602 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1603 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1604 * this area with pages of identical content from other such areas.
1605 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1606 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1607 * huge pages in the future. Existing pages might be coalesced and
1608 * new pages might be allocated as THP.
1609 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1610 * transparent huge pages so the existing pages will not be
1611 * coalesced into THP and new pages will not be allocated as THP.
1612 * MADV_COLLAPSE - synchronously coalesce pages into new THP.
1613 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1614 * from being included in its core dump.
1615 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1616 * MADV_COLD - the application is not expected to use this memory soon,
1617 * deactivate pages in this range so that they can be reclaimed
1618 * easily if memory pressure happens.
1619 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1620 * page out the pages in this range immediately.
1621 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1622 * triggering read faults if required
1623 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1624 * triggering write faults if required
1628 * -EINVAL - start + len < 0, start is not page-aligned,
1629 * "behavior" is not a valid value, or application
1630 * is attempting to release locked or shared pages,
1631 * or the specified address range includes file, Huge TLB,
1632 * MAP_SHARED or VMPFNMAP range.
1633 * -ENOMEM - addresses in the specified range are not currently
1634 * mapped, or are outside the AS of the process.
1635 * -EIO - an I/O error occurred while paging in data.
1636 * -EBADF - map exists, but area maps something that isn't a file.
1637 * -EAGAIN - a kernel resource was temporarily unavailable.
1638 * -EPERM - memory is sealed.
1640 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1646 struct blk_plug plug;
1648 if (!madvise_behavior_valid(behavior))
1651 if (!PAGE_ALIGNED(start))
1653 len = PAGE_ALIGN(len_in);
1655 /* Check to see whether len was rounded up from small -ve to zero */
1666 #ifdef CONFIG_MEMORY_FAILURE
1667 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1668 return madvise_inject_error(behavior, start, start + len_in);
1671 write = madvise_need_mmap_write(behavior);
1673 if (mmap_write_lock_killable(mm))
1679 start = untagged_addr_remote(mm, start);
1682 blk_start_plug(&plug);
1684 case MADV_POPULATE_READ:
1685 case MADV_POPULATE_WRITE:
1686 error = madvise_populate(mm, start, end, behavior);
1689 error = madvise_walk_vmas(mm, start, end, behavior,
1690 madvise_vma_behavior);
1693 blk_finish_plug(&plug);
1696 mmap_write_unlock(mm);
1698 mmap_read_unlock(mm);
1703 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1705 return do_madvise(current->mm, start, len_in, behavior);
1708 /* Perform an madvise operation over a vector of addresses and lengths. */
1709 static ssize_t vector_madvise(struct mm_struct *mm, struct iov_iter *iter,
1715 total_len = iov_iter_count(iter);
1717 while (iov_iter_count(iter)) {
1718 ret = do_madvise(mm, (unsigned long)iter_iov_addr(iter),
1719 iter_iov_len(iter), behavior);
1721 * An madvise operation is attempting to restart the syscall,
1722 * but we cannot proceed as it would not be correct to repeat
1723 * the operation in aggregate, and would be surprising to the
1726 * As we have already dropped locks, it is safe to just loop and
1727 * try again. We check for fatal signals in case we need exit
1730 if (ret == -ERESTARTNOINTR) {
1731 if (fatal_signal_pending(current)) {
1739 iov_iter_advance(iter, iter_iov_len(iter));
1742 ret = (total_len - iov_iter_count(iter)) ? : ret;
1747 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1748 size_t, vlen, int, behavior, unsigned int, flags)
1751 struct iovec iovstack[UIO_FASTIOV];
1752 struct iovec *iov = iovstack;
1753 struct iov_iter iter;
1754 struct task_struct *task;
1755 struct mm_struct *mm;
1756 unsigned int f_flags;
1763 ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1767 task = pidfd_get_task(pidfd, &f_flags);
1769 ret = PTR_ERR(task);
1773 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1774 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1781 * We need only perform this check if we are attempting to manipulate a
1782 * remote process's address space.
1784 if (mm != current->mm && !process_madvise_remote_valid(behavior)) {
1790 * Require CAP_SYS_NICE for influencing process performance. Note that
1791 * only non-destructive hints are currently supported for remote
1794 if (mm != current->mm && !capable(CAP_SYS_NICE)) {
1799 ret = vector_madvise(mm, &iter, behavior);
1804 put_task_struct(task);