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);
176 /* vm_flags is protected by the mmap_lock held in write mode. */
177 vma_start_write(vma
);
178 vm_flags_reset(vma
, new_flags
);
179 if (!vma
->vm_file
|| vma_is_anon_shmem(vma
)) {
180 error
= replace_anon_vma_name(vma
, anon_name
);
189 static int swapin_walk_pmd_entry(pmd_t
*pmd
, unsigned long start
,
190 unsigned long end
, struct mm_walk
*walk
)
192 struct vm_area_struct
*vma
= walk
->private;
193 struct swap_iocb
*splug
= NULL
;
198 for (addr
= start
; addr
< end
; addr
+= PAGE_SIZE
) {
204 ptep
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
209 pte
= ptep_get(ptep
);
210 if (!is_swap_pte(pte
))
212 entry
= pte_to_swp_entry(pte
);
213 if (unlikely(non_swap_entry(entry
)))
216 pte_unmap_unlock(ptep
, ptl
);
219 page
= read_swap_cache_async(entry
, GFP_HIGHUSER_MOVABLE
,
226 pte_unmap_unlock(ptep
, ptl
);
227 swap_read_unplug(splug
);
233 static const struct mm_walk_ops swapin_walk_ops
= {
234 .pmd_entry
= swapin_walk_pmd_entry
,
235 .walk_lock
= PGWALK_RDLOCK
,
238 static void shmem_swapin_range(struct vm_area_struct
*vma
,
239 unsigned long start
, unsigned long end
,
240 struct address_space
*mapping
)
242 XA_STATE(xas
, &mapping
->i_pages
, linear_page_index(vma
, start
));
243 pgoff_t end_index
= linear_page_index(vma
, end
) - 1;
245 struct swap_iocb
*splug
= NULL
;
248 xas_for_each(&xas
, page
, end_index
) {
252 if (!xa_is_value(page
))
254 entry
= radix_to_swp_entry(page
);
255 /* There might be swapin error entries in shmem mapping. */
256 if (non_swap_entry(entry
))
259 addr
= vma
->vm_start
+
260 ((xas
.xa_index
- vma
->vm_pgoff
) << PAGE_SHIFT
);
264 page
= read_swap_cache_async(entry
, mapping_gfp_mask(mapping
),
272 swap_read_unplug(splug
);
274 #endif /* CONFIG_SWAP */
277 * Schedule all required I/O operations. Do not wait for completion.
279 static long madvise_willneed(struct vm_area_struct
*vma
,
280 struct vm_area_struct
**prev
,
281 unsigned long start
, unsigned long end
)
283 struct mm_struct
*mm
= vma
->vm_mm
;
284 struct file
*file
= vma
->vm_file
;
290 walk_page_range(vma
->vm_mm
, start
, end
, &swapin_walk_ops
, vma
);
291 lru_add_drain(); /* Push any new pages onto the LRU now */
295 if (shmem_mapping(file
->f_mapping
)) {
296 shmem_swapin_range(vma
, start
, end
, file
->f_mapping
);
297 lru_add_drain(); /* Push any new pages onto the LRU now */
305 if (IS_DAX(file_inode(file
))) {
306 /* no bad return value, but ignore advice */
311 * Filesystem's fadvise may need to take various locks. We need to
312 * explicitly grab a reference because the vma (and hence the
313 * vma's reference to the file) can go away as soon as we drop
316 *prev
= NULL
; /* tell sys_madvise we drop mmap_lock */
318 offset
= (loff_t
)(start
- vma
->vm_start
)
319 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
320 mmap_read_unlock(mm
);
321 vfs_fadvise(file
, offset
, end
- start
, POSIX_FADV_WILLNEED
);
327 static inline bool can_do_file_pageout(struct vm_area_struct
*vma
)
332 * paging out pagecache only for non-anonymous mappings that correspond
333 * to the files the calling process could (if tried) open for writing;
334 * otherwise we'd be including shared non-exclusive mappings, which
335 * opens a side channel.
337 return inode_owner_or_capable(&nop_mnt_idmap
,
338 file_inode(vma
->vm_file
)) ||
339 file_permission(vma
->vm_file
, MAY_WRITE
) == 0;
342 static int madvise_cold_or_pageout_pte_range(pmd_t
*pmd
,
343 unsigned long addr
, unsigned long end
,
344 struct mm_walk
*walk
)
346 struct madvise_walk_private
*private = walk
->private;
347 struct mmu_gather
*tlb
= private->tlb
;
348 bool pageout
= private->pageout
;
349 struct mm_struct
*mm
= tlb
->mm
;
350 struct vm_area_struct
*vma
= walk
->vma
;
351 pte_t
*start_pte
, *pte
, ptent
;
353 struct folio
*folio
= NULL
;
354 LIST_HEAD(folio_list
);
355 bool pageout_anon_only_filter
;
357 if (fatal_signal_pending(current
))
360 pageout_anon_only_filter
= pageout
&& !vma_is_anonymous(vma
) &&
361 !can_do_file_pageout(vma
);
363 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
364 if (pmd_trans_huge(*pmd
)) {
366 unsigned long next
= pmd_addr_end(addr
, end
);
368 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
369 ptl
= pmd_trans_huge_lock(pmd
, vma
);
374 if (is_huge_zero_pmd(orig_pmd
))
377 if (unlikely(!pmd_present(orig_pmd
))) {
378 VM_BUG_ON(thp_migration_supported() &&
379 !is_pmd_migration_entry(orig_pmd
));
383 folio
= pfn_folio(pmd_pfn(orig_pmd
));
385 /* Do not interfere with other mappings of this folio */
386 if (folio_estimated_sharers(folio
) != 1)
389 if (pageout_anon_only_filter
&& !folio_test_anon(folio
))
392 if (next
- addr
!= HPAGE_PMD_SIZE
) {
398 err
= split_folio(folio
);
406 if (pmd_young(orig_pmd
)) {
407 pmdp_invalidate(vma
, addr
, pmd
);
408 orig_pmd
= pmd_mkold(orig_pmd
);
410 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
411 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
414 folio_clear_referenced(folio
);
415 folio_test_clear_young(folio
);
416 if (folio_test_active(folio
))
417 folio_set_workingset(folio
);
419 if (folio_isolate_lru(folio
)) {
420 if (folio_test_unevictable(folio
))
421 folio_putback_lru(folio
);
423 list_add(&folio
->lru
, &folio_list
);
426 folio_deactivate(folio
);
430 reclaim_pages(&folio_list
);
436 tlb_change_page_size(tlb
, PAGE_SIZE
);
437 start_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
440 flush_tlb_batched_pending(mm
);
441 arch_enter_lazy_mmu_mode();
442 for (; addr
< end
; pte
++, addr
+= PAGE_SIZE
) {
443 ptent
= ptep_get(pte
);
448 if (!pte_present(ptent
))
451 folio
= vm_normal_folio(vma
, addr
, ptent
);
452 if (!folio
|| folio_is_zone_device(folio
))
456 * Creating a THP page is expensive so split it only if we
457 * are sure it's worth. Split it if we are only owner.
459 if (folio_test_large(folio
)) {
462 if (folio_estimated_sharers(folio
) != 1)
464 if (pageout_anon_only_filter
&& !folio_test_anon(folio
))
466 if (!folio_trylock(folio
))
469 arch_leave_lazy_mmu_mode();
470 pte_unmap_unlock(start_pte
, ptl
);
472 err
= split_folio(folio
);
478 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
481 arch_enter_lazy_mmu_mode();
488 * Do not interfere with other mappings of this folio and
491 if (!folio_test_lru(folio
) || folio_mapcount(folio
) != 1)
494 if (pageout_anon_only_filter
&& !folio_test_anon(folio
))
497 VM_BUG_ON_FOLIO(folio_test_large(folio
), folio
);
499 if (pte_young(ptent
)) {
500 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
502 ptent
= pte_mkold(ptent
);
503 set_pte_at(mm
, addr
, pte
, ptent
);
504 tlb_remove_tlb_entry(tlb
, pte
, addr
);
508 * We are deactivating a folio for accelerating reclaiming.
509 * VM couldn't reclaim the folio unless we clear PG_young.
510 * As a side effect, it makes confuse idle-page tracking
511 * because they will miss recent referenced history.
513 folio_clear_referenced(folio
);
514 folio_test_clear_young(folio
);
515 if (folio_test_active(folio
))
516 folio_set_workingset(folio
);
518 if (folio_isolate_lru(folio
)) {
519 if (folio_test_unevictable(folio
))
520 folio_putback_lru(folio
);
522 list_add(&folio
->lru
, &folio_list
);
525 folio_deactivate(folio
);
529 arch_leave_lazy_mmu_mode();
530 pte_unmap_unlock(start_pte
, ptl
);
533 reclaim_pages(&folio_list
);
539 static const struct mm_walk_ops cold_walk_ops
= {
540 .pmd_entry
= madvise_cold_or_pageout_pte_range
,
541 .walk_lock
= PGWALK_RDLOCK
,
544 static void madvise_cold_page_range(struct mmu_gather
*tlb
,
545 struct vm_area_struct
*vma
,
546 unsigned long addr
, unsigned long end
)
548 struct madvise_walk_private walk_private
= {
553 tlb_start_vma(tlb
, vma
);
554 walk_page_range(vma
->vm_mm
, addr
, end
, &cold_walk_ops
, &walk_private
);
555 tlb_end_vma(tlb
, vma
);
558 static inline bool can_madv_lru_vma(struct vm_area_struct
*vma
)
560 return !(vma
->vm_flags
& (VM_LOCKED
|VM_PFNMAP
|VM_HUGETLB
));
563 static long madvise_cold(struct vm_area_struct
*vma
,
564 struct vm_area_struct
**prev
,
565 unsigned long start_addr
, unsigned long end_addr
)
567 struct mm_struct
*mm
= vma
->vm_mm
;
568 struct mmu_gather tlb
;
571 if (!can_madv_lru_vma(vma
))
575 tlb_gather_mmu(&tlb
, mm
);
576 madvise_cold_page_range(&tlb
, vma
, start_addr
, end_addr
);
577 tlb_finish_mmu(&tlb
);
582 static void madvise_pageout_page_range(struct mmu_gather
*tlb
,
583 struct vm_area_struct
*vma
,
584 unsigned long addr
, unsigned long end
)
586 struct madvise_walk_private walk_private
= {
591 tlb_start_vma(tlb
, vma
);
592 walk_page_range(vma
->vm_mm
, addr
, end
, &cold_walk_ops
, &walk_private
);
593 tlb_end_vma(tlb
, vma
);
596 static long madvise_pageout(struct vm_area_struct
*vma
,
597 struct vm_area_struct
**prev
,
598 unsigned long start_addr
, unsigned long end_addr
)
600 struct mm_struct
*mm
= vma
->vm_mm
;
601 struct mmu_gather tlb
;
604 if (!can_madv_lru_vma(vma
))
608 * If the VMA belongs to a private file mapping, there can be private
609 * dirty pages which can be paged out if even this process is neither
610 * owner nor write capable of the file. We allow private file mappings
611 * further to pageout dirty anon pages.
613 if (!vma_is_anonymous(vma
) && (!can_do_file_pageout(vma
) &&
614 (vma
->vm_flags
& VM_MAYSHARE
)))
618 tlb_gather_mmu(&tlb
, mm
);
619 madvise_pageout_page_range(&tlb
, vma
, start_addr
, end_addr
);
620 tlb_finish_mmu(&tlb
);
625 static int madvise_free_pte_range(pmd_t
*pmd
, unsigned long addr
,
626 unsigned long end
, struct mm_walk
*walk
)
629 struct mmu_gather
*tlb
= walk
->private;
630 struct mm_struct
*mm
= tlb
->mm
;
631 struct vm_area_struct
*vma
= walk
->vma
;
633 pte_t
*start_pte
, *pte
, ptent
;
638 next
= pmd_addr_end(addr
, end
);
639 if (pmd_trans_huge(*pmd
))
640 if (madvise_free_huge_pmd(tlb
, vma
, pmd
, addr
, next
))
643 tlb_change_page_size(tlb
, PAGE_SIZE
);
644 start_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
647 flush_tlb_batched_pending(mm
);
648 arch_enter_lazy_mmu_mode();
649 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
650 ptent
= ptep_get(pte
);
655 * If the pte has swp_entry, just clear page table to
656 * prevent swap-in which is more expensive rather than
657 * (page allocation + zeroing).
659 if (!pte_present(ptent
)) {
662 entry
= pte_to_swp_entry(ptent
);
663 if (!non_swap_entry(entry
)) {
665 free_swap_and_cache(entry
);
666 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
667 } else if (is_hwpoison_entry(entry
) ||
668 is_poisoned_swp_entry(entry
)) {
669 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
674 folio
= vm_normal_folio(vma
, addr
, ptent
);
675 if (!folio
|| folio_is_zone_device(folio
))
679 * If pmd isn't transhuge but the folio is large and
680 * is owned by only this process, split it and
681 * deactivate all pages.
683 if (folio_test_large(folio
)) {
686 if (folio_estimated_sharers(folio
) != 1)
688 if (!folio_trylock(folio
))
691 arch_leave_lazy_mmu_mode();
692 pte_unmap_unlock(start_pte
, ptl
);
694 err
= split_folio(folio
);
700 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
703 arch_enter_lazy_mmu_mode();
709 if (folio_test_swapcache(folio
) || folio_test_dirty(folio
)) {
710 if (!folio_trylock(folio
))
713 * If folio is shared with others, we mustn't clear
714 * the folio's dirty flag.
716 if (folio_mapcount(folio
) != 1) {
721 if (folio_test_swapcache(folio
) &&
722 !folio_free_swap(folio
)) {
727 folio_clear_dirty(folio
);
731 if (pte_young(ptent
) || pte_dirty(ptent
)) {
733 * Some of architecture(ex, PPC) don't update TLB
734 * with set_pte_at and tlb_remove_tlb_entry so for
735 * the portability, remap the pte with old|clean
736 * after pte clearing.
738 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
741 ptent
= pte_mkold(ptent
);
742 ptent
= pte_mkclean(ptent
);
743 set_pte_at(mm
, addr
, pte
, ptent
);
744 tlb_remove_tlb_entry(tlb
, pte
, addr
);
746 folio_mark_lazyfree(folio
);
750 if (current
->mm
== mm
)
752 add_mm_counter(mm
, MM_SWAPENTS
, nr_swap
);
755 arch_leave_lazy_mmu_mode();
756 pte_unmap_unlock(start_pte
, ptl
);
763 static const struct mm_walk_ops madvise_free_walk_ops
= {
764 .pmd_entry
= madvise_free_pte_range
,
765 .walk_lock
= PGWALK_RDLOCK
,
768 static int madvise_free_single_vma(struct vm_area_struct
*vma
,
769 unsigned long start_addr
, unsigned long end_addr
)
771 struct mm_struct
*mm
= vma
->vm_mm
;
772 struct mmu_notifier_range range
;
773 struct mmu_gather tlb
;
775 /* MADV_FREE works for only anon vma at the moment */
776 if (!vma_is_anonymous(vma
))
779 range
.start
= max(vma
->vm_start
, start_addr
);
780 if (range
.start
>= vma
->vm_end
)
782 range
.end
= min(vma
->vm_end
, end_addr
);
783 if (range
.end
<= vma
->vm_start
)
785 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, mm
,
786 range
.start
, range
.end
);
789 tlb_gather_mmu(&tlb
, mm
);
790 update_hiwater_rss(mm
);
792 mmu_notifier_invalidate_range_start(&range
);
793 tlb_start_vma(&tlb
, vma
);
794 walk_page_range(vma
->vm_mm
, range
.start
, range
.end
,
795 &madvise_free_walk_ops
, &tlb
);
796 tlb_end_vma(&tlb
, vma
);
797 mmu_notifier_invalidate_range_end(&range
);
798 tlb_finish_mmu(&tlb
);
804 * Application no longer needs these pages. If the pages are dirty,
805 * it's OK to just throw them away. The app will be more careful about
806 * data it wants to keep. Be sure to free swap resources too. The
807 * zap_page_range_single call sets things up for shrink_active_list to actually
808 * free these pages later if no one else has touched them in the meantime,
809 * although we could add these pages to a global reuse list for
810 * shrink_active_list to pick up before reclaiming other pages.
812 * NB: This interface discards data rather than pushes it out to swap,
813 * as some implementations do. This has performance implications for
814 * applications like large transactional databases which want to discard
815 * pages in anonymous maps after committing to backing store the data
816 * that was kept in them. There is no reason to write this data out to
817 * the swap area if the application is discarding it.
819 * An interface that causes the system to free clean pages and flush
820 * dirty pages is already available as msync(MS_INVALIDATE).
822 static long madvise_dontneed_single_vma(struct vm_area_struct
*vma
,
823 unsigned long start
, unsigned long end
)
825 zap_page_range_single(vma
, start
, end
- start
, NULL
);
829 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct
*vma
,
834 if (!is_vm_hugetlb_page(vma
)) {
835 unsigned int forbidden
= VM_PFNMAP
;
837 if (behavior
!= MADV_DONTNEED_LOCKED
)
838 forbidden
|= VM_LOCKED
;
840 return !(vma
->vm_flags
& forbidden
);
843 if (behavior
!= MADV_DONTNEED
&& behavior
!= MADV_DONTNEED_LOCKED
)
845 if (start
& ~huge_page_mask(hstate_vma(vma
)))
849 * Madvise callers expect the length to be rounded up to PAGE_SIZE
850 * boundaries, and may be unaware that this VMA uses huge pages.
851 * Avoid unexpected data loss by rounding down the number of
854 *end
= ALIGN_DOWN(*end
, huge_page_size(hstate_vma(vma
)));
859 static long madvise_dontneed_free(struct vm_area_struct
*vma
,
860 struct vm_area_struct
**prev
,
861 unsigned long start
, unsigned long end
,
864 struct mm_struct
*mm
= vma
->vm_mm
;
867 if (!madvise_dontneed_free_valid_vma(vma
, start
, &end
, behavior
))
873 if (!userfaultfd_remove(vma
, start
, end
)) {
874 *prev
= NULL
; /* mmap_lock has been dropped, prev is stale */
877 vma
= vma_lookup(mm
, start
);
881 * Potential end adjustment for hugetlb vma is OK as
882 * the check below keeps end within vma.
884 if (!madvise_dontneed_free_valid_vma(vma
, start
, &end
,
887 if (end
> vma
->vm_end
) {
889 * Don't fail if end > vma->vm_end. If the old
890 * vma was split while the mmap_lock was
891 * released the effect of the concurrent
892 * operation may not cause madvise() to
893 * have an undefined result. There may be an
894 * adjacent next vma that we'll walk
895 * next. userfaultfd_remove() will generate an
896 * UFFD_EVENT_REMOVE repetition on the
897 * end-vma->vm_end range, but the manager can
898 * handle a repetition fine.
902 VM_WARN_ON(start
>= end
);
905 if (behavior
== MADV_DONTNEED
|| behavior
== MADV_DONTNEED_LOCKED
)
906 return madvise_dontneed_single_vma(vma
, start
, end
);
907 else if (behavior
== MADV_FREE
)
908 return madvise_free_single_vma(vma
, start
, end
);
913 static long madvise_populate(struct vm_area_struct
*vma
,
914 struct vm_area_struct
**prev
,
915 unsigned long start
, unsigned long end
,
918 const bool write
= behavior
== MADV_POPULATE_WRITE
;
919 struct mm_struct
*mm
= vma
->vm_mm
;
920 unsigned long tmp_end
;
926 while (start
< end
) {
928 * We might have temporarily dropped the lock. For example,
929 * our VMA might have been split.
931 if (!vma
|| start
>= vma
->vm_end
) {
932 vma
= vma_lookup(mm
, start
);
937 tmp_end
= min_t(unsigned long, end
, vma
->vm_end
);
938 /* Populate (prefault) page tables readable/writable. */
939 pages
= faultin_vma_page_range(vma
, start
, tmp_end
, write
,
951 case -EINVAL
: /* Incompatible mappings / permissions. */
955 case -EFAULT
: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
958 pr_warn_once("%s: unhandled return value: %ld\n",
965 start
+= pages
* PAGE_SIZE
;
971 * Application wants to free up the pages and associated backing store.
972 * This is effectively punching a hole into the middle of a file.
974 static long madvise_remove(struct vm_area_struct
*vma
,
975 struct vm_area_struct
**prev
,
976 unsigned long start
, unsigned long end
)
981 struct mm_struct
*mm
= vma
->vm_mm
;
983 *prev
= NULL
; /* tell sys_madvise we drop mmap_lock */
985 if (vma
->vm_flags
& VM_LOCKED
)
990 if (!f
|| !f
->f_mapping
|| !f
->f_mapping
->host
) {
994 if ((vma
->vm_flags
& (VM_SHARED
|VM_WRITE
)) != (VM_SHARED
|VM_WRITE
))
997 offset
= (loff_t
)(start
- vma
->vm_start
)
998 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
1001 * Filesystem's fallocate may need to take i_rwsem. We need to
1002 * explicitly grab a reference because the vma (and hence the
1003 * vma's reference to the file) can go away as soon as we drop
1007 if (userfaultfd_remove(vma
, start
, end
)) {
1008 /* mmap_lock was not released by userfaultfd_remove() */
1009 mmap_read_unlock(mm
);
1011 error
= vfs_fallocate(f
,
1012 FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
,
1013 offset
, end
- start
);
1020 * Apply an madvise behavior to a region of a vma. madvise_update_vma
1021 * will handle splitting a vm area into separate areas, each area with its own
1024 static int madvise_vma_behavior(struct vm_area_struct
*vma
,
1025 struct vm_area_struct
**prev
,
1026 unsigned long start
, unsigned long end
,
1027 unsigned long behavior
)
1030 struct anon_vma_name
*anon_name
;
1031 unsigned long new_flags
= vma
->vm_flags
;
1035 return madvise_remove(vma
, prev
, start
, end
);
1037 return madvise_willneed(vma
, prev
, start
, end
);
1039 return madvise_cold(vma
, prev
, start
, end
);
1041 return madvise_pageout(vma
, prev
, start
, end
);
1044 case MADV_DONTNEED_LOCKED
:
1045 return madvise_dontneed_free(vma
, prev
, start
, end
, behavior
);
1046 case MADV_POPULATE_READ
:
1047 case MADV_POPULATE_WRITE
:
1048 return madvise_populate(vma
, prev
, start
, end
, behavior
);
1050 new_flags
= new_flags
& ~VM_RAND_READ
& ~VM_SEQ_READ
;
1052 case MADV_SEQUENTIAL
:
1053 new_flags
= (new_flags
& ~VM_RAND_READ
) | VM_SEQ_READ
;
1056 new_flags
= (new_flags
& ~VM_SEQ_READ
) | VM_RAND_READ
;
1059 new_flags
|= VM_DONTCOPY
;
1062 if (vma
->vm_flags
& VM_IO
)
1064 new_flags
&= ~VM_DONTCOPY
;
1066 case MADV_WIPEONFORK
:
1067 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1068 if (vma
->vm_file
|| vma
->vm_flags
& VM_SHARED
)
1070 new_flags
|= VM_WIPEONFORK
;
1072 case MADV_KEEPONFORK
:
1073 new_flags
&= ~VM_WIPEONFORK
;
1076 new_flags
|= VM_DONTDUMP
;
1079 if (!is_vm_hugetlb_page(vma
) && new_flags
& VM_SPECIAL
)
1081 new_flags
&= ~VM_DONTDUMP
;
1083 case MADV_MERGEABLE
:
1084 case MADV_UNMERGEABLE
:
1085 error
= ksm_madvise(vma
, start
, end
, behavior
, &new_flags
);
1090 case MADV_NOHUGEPAGE
:
1091 error
= hugepage_madvise(vma
, &new_flags
, behavior
);
1096 return madvise_collapse(vma
, prev
, start
, end
);
1099 anon_name
= anon_vma_name(vma
);
1100 anon_vma_name_get(anon_name
);
1101 error
= madvise_update_vma(vma
, prev
, start
, end
, new_flags
,
1103 anon_vma_name_put(anon_name
);
1107 * madvise() returns EAGAIN if kernel resources, such as
1108 * slab, are temporarily unavailable.
1110 if (error
== -ENOMEM
)
1115 #ifdef CONFIG_MEMORY_FAILURE
1117 * Error injection support for memory error handling.
1119 static int madvise_inject_error(int behavior
,
1120 unsigned long start
, unsigned long end
)
1124 if (!capable(CAP_SYS_ADMIN
))
1128 for (; start
< end
; start
+= size
) {
1133 ret
= get_user_pages_fast(start
, 1, 0, &page
);
1136 pfn
= page_to_pfn(page
);
1139 * When soft offlining hugepages, after migrating the page
1140 * we dissolve it, therefore in the second loop "page" will
1141 * no longer be a compound page.
1143 size
= page_size(compound_head(page
));
1145 if (behavior
== MADV_SOFT_OFFLINE
) {
1146 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1148 ret
= soft_offline_page(pfn
, MF_COUNT_INCREASED
);
1150 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1152 ret
= memory_failure(pfn
, MF_COUNT_INCREASED
| MF_SW_SIMULATED
);
1153 if (ret
== -EOPNOTSUPP
)
1166 madvise_behavior_valid(int behavior
)
1172 case MADV_SEQUENTIAL
:
1177 case MADV_DONTNEED_LOCKED
:
1181 case MADV_POPULATE_READ
:
1182 case MADV_POPULATE_WRITE
:
1184 case MADV_MERGEABLE
:
1185 case MADV_UNMERGEABLE
:
1187 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1189 case MADV_NOHUGEPAGE
:
1194 case MADV_WIPEONFORK
:
1195 case MADV_KEEPONFORK
:
1196 #ifdef CONFIG_MEMORY_FAILURE
1197 case MADV_SOFT_OFFLINE
:
1207 static bool process_madvise_behavior_valid(int behavior
)
1221 * Walk the vmas in range [start,end), and call the visit function on each one.
1222 * The visit function will get start and end parameters that cover the overlap
1223 * between the current vma and the original range. Any unmapped regions in the
1224 * original range will result in this function returning -ENOMEM while still
1225 * calling the visit function on all of the existing vmas in the range.
1226 * Must be called with the mmap_lock held for reading or writing.
1229 int madvise_walk_vmas(struct mm_struct
*mm
, unsigned long start
,
1230 unsigned long end
, unsigned long arg
,
1231 int (*visit
)(struct vm_area_struct
*vma
,
1232 struct vm_area_struct
**prev
, unsigned long start
,
1233 unsigned long end
, unsigned long arg
))
1235 struct vm_area_struct
*vma
;
1236 struct vm_area_struct
*prev
;
1238 int unmapped_error
= 0;
1241 * If the interval [start,end) covers some unmapped address
1242 * ranges, just ignore them, but return -ENOMEM at the end.
1243 * - different from the way of handling in mlock etc.
1245 vma
= find_vma_prev(mm
, start
, &prev
);
1246 if (vma
&& start
> vma
->vm_start
)
1252 /* Still start < end. */
1256 /* Here start < (end|vma->vm_end). */
1257 if (start
< vma
->vm_start
) {
1258 unmapped_error
= -ENOMEM
;
1259 start
= vma
->vm_start
;
1264 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1269 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1270 error
= visit(vma
, &prev
, start
, tmp
, arg
);
1274 if (prev
&& start
< prev
->vm_end
)
1275 start
= prev
->vm_end
;
1279 vma
= find_vma(mm
, prev
->vm_end
);
1280 else /* madvise_remove dropped mmap_lock */
1281 vma
= find_vma(mm
, start
);
1284 return unmapped_error
;
1287 #ifdef CONFIG_ANON_VMA_NAME
1288 static int madvise_vma_anon_name(struct vm_area_struct
*vma
,
1289 struct vm_area_struct
**prev
,
1290 unsigned long start
, unsigned long end
,
1291 unsigned long anon_name
)
1295 /* Only anonymous mappings can be named */
1296 if (vma
->vm_file
&& !vma_is_anon_shmem(vma
))
1299 error
= madvise_update_vma(vma
, prev
, start
, end
, vma
->vm_flags
,
1300 (struct anon_vma_name
*)anon_name
);
1303 * madvise() returns EAGAIN if kernel resources, such as
1304 * slab, are temporarily unavailable.
1306 if (error
== -ENOMEM
)
1311 int madvise_set_anon_name(struct mm_struct
*mm
, unsigned long start
,
1312 unsigned long len_in
, struct anon_vma_name
*anon_name
)
1317 if (start
& ~PAGE_MASK
)
1319 len
= (len_in
+ ~PAGE_MASK
) & PAGE_MASK
;
1321 /* Check to see whether len was rounded up from small -ve to zero */
1332 return madvise_walk_vmas(mm
, start
, end
, (unsigned long)anon_name
,
1333 madvise_vma_anon_name
);
1335 #endif /* CONFIG_ANON_VMA_NAME */
1337 * The madvise(2) system call.
1339 * Applications can use madvise() to advise the kernel how it should
1340 * handle paging I/O in this VM area. The idea is to help the kernel
1341 * use appropriate read-ahead and caching techniques. The information
1342 * provided is advisory only, and can be safely disregarded by the
1343 * kernel without affecting the correct operation of the application.
1346 * MADV_NORMAL - the default behavior is to read clusters. This
1347 * results in some read-ahead and read-behind.
1348 * MADV_RANDOM - the system should read the minimum amount of data
1349 * on any access, since it is unlikely that the appli-
1350 * cation will need more than what it asks for.
1351 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1352 * once, so they can be aggressively read ahead, and
1353 * can be freed soon after they are accessed.
1354 * MADV_WILLNEED - the application is notifying the system to read
1356 * MADV_DONTNEED - the application is finished with the given range,
1357 * so the kernel can free resources associated with it.
1358 * MADV_FREE - the application marks pages in the given range as lazy free,
1359 * where actual purges are postponed until memory pressure happens.
1360 * MADV_REMOVE - the application wants to free up the given range of
1361 * pages and associated backing store.
1362 * MADV_DONTFORK - omit this area from child's address space when forking:
1363 * typically, to avoid COWing pages pinned by get_user_pages().
1364 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1365 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1366 * range after a fork.
1367 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1368 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1369 * were corrupted by unrecoverable hardware memory failure.
1370 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1371 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1372 * this area with pages of identical content from other such areas.
1373 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1374 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1375 * huge pages in the future. Existing pages might be coalesced and
1376 * new pages might be allocated as THP.
1377 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1378 * transparent huge pages so the existing pages will not be
1379 * coalesced into THP and new pages will not be allocated as THP.
1380 * MADV_COLLAPSE - synchronously coalesce pages into new THP.
1381 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1382 * from being included in its core dump.
1383 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1384 * MADV_COLD - the application is not expected to use this memory soon,
1385 * deactivate pages in this range so that they can be reclaimed
1386 * easily if memory pressure happens.
1387 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1388 * page out the pages in this range immediately.
1389 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1390 * triggering read faults if required
1391 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1392 * triggering write faults if required
1396 * -EINVAL - start + len < 0, start is not page-aligned,
1397 * "behavior" is not a valid value, or application
1398 * is attempting to release locked or shared pages,
1399 * or the specified address range includes file, Huge TLB,
1400 * MAP_SHARED or VMPFNMAP range.
1401 * -ENOMEM - addresses in the specified range are not currently
1402 * mapped, or are outside the AS of the process.
1403 * -EIO - an I/O error occurred while paging in data.
1404 * -EBADF - map exists, but area maps something that isn't a file.
1405 * -EAGAIN - a kernel resource was temporarily unavailable.
1407 int do_madvise(struct mm_struct
*mm
, unsigned long start
, size_t len_in
, int behavior
)
1413 struct blk_plug plug
;
1415 if (!madvise_behavior_valid(behavior
))
1418 if (!PAGE_ALIGNED(start
))
1420 len
= PAGE_ALIGN(len_in
);
1422 /* Check to see whether len was rounded up from small -ve to zero */
1433 #ifdef CONFIG_MEMORY_FAILURE
1434 if (behavior
== MADV_HWPOISON
|| behavior
== MADV_SOFT_OFFLINE
)
1435 return madvise_inject_error(behavior
, start
, start
+ len_in
);
1438 write
= madvise_need_mmap_write(behavior
);
1440 if (mmap_write_lock_killable(mm
))
1446 start
= untagged_addr_remote(mm
, start
);
1449 blk_start_plug(&plug
);
1450 error
= madvise_walk_vmas(mm
, start
, end
, behavior
,
1451 madvise_vma_behavior
);
1452 blk_finish_plug(&plug
);
1454 mmap_write_unlock(mm
);
1456 mmap_read_unlock(mm
);
1461 SYSCALL_DEFINE3(madvise
, unsigned long, start
, size_t, len_in
, int, behavior
)
1463 return do_madvise(current
->mm
, start
, len_in
, behavior
);
1466 SYSCALL_DEFINE5(process_madvise
, int, pidfd
, const struct iovec __user
*, vec
,
1467 size_t, vlen
, int, behavior
, unsigned int, flags
)
1470 struct iovec iovstack
[UIO_FASTIOV
];
1471 struct iovec
*iov
= iovstack
;
1472 struct iov_iter iter
;
1473 struct task_struct
*task
;
1474 struct mm_struct
*mm
;
1476 unsigned int f_flags
;
1483 ret
= import_iovec(ITER_DEST
, vec
, vlen
, ARRAY_SIZE(iovstack
), &iov
, &iter
);
1487 task
= pidfd_get_task(pidfd
, &f_flags
);
1489 ret
= PTR_ERR(task
);
1493 if (!process_madvise_behavior_valid(behavior
)) {
1498 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1499 mm
= mm_access(task
, PTRACE_MODE_READ_FSCREDS
);
1500 if (IS_ERR_OR_NULL(mm
)) {
1501 ret
= IS_ERR(mm
) ? PTR_ERR(mm
) : -ESRCH
;
1506 * Require CAP_SYS_NICE for influencing process performance. Note that
1507 * only non-destructive hints are currently supported.
1509 if (!capable(CAP_SYS_NICE
)) {
1514 total_len
= iov_iter_count(&iter
);
1516 while (iov_iter_count(&iter
)) {
1517 ret
= do_madvise(mm
, (unsigned long)iter_iov_addr(&iter
),
1518 iter_iov_len(&iter
), behavior
);
1521 iov_iter_advance(&iter
, iter_iov_len(&iter
));
1524 ret
= (total_len
- iov_iter_count(&iter
)) ? : ret
;
1529 put_task_struct(task
);