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
:
64 /* be safe, default to 1. list exceptions explicitly */
69 #ifdef CONFIG_ANON_VMA_NAME
70 struct anon_vma_name
*anon_vma_name_alloc(const char *name
)
72 struct anon_vma_name
*anon_name
;
75 /* Add 1 for NUL terminator at the end of the anon_name->name */
76 count
= strlen(name
) + 1;
77 anon_name
= kmalloc(struct_size(anon_name
, name
, count
), GFP_KERNEL
);
79 kref_init(&anon_name
->kref
);
80 memcpy(anon_name
->name
, name
, count
);
86 void anon_vma_name_free(struct kref
*kref
)
88 struct anon_vma_name
*anon_name
=
89 container_of(kref
, struct anon_vma_name
, kref
);
93 struct anon_vma_name
*anon_vma_name(struct vm_area_struct
*vma
)
95 mmap_assert_locked(vma
->vm_mm
);
100 return vma
->anon_name
;
103 /* mmap_lock should be write-locked */
104 static int replace_anon_vma_name(struct vm_area_struct
*vma
,
105 struct anon_vma_name
*anon_name
)
107 struct anon_vma_name
*orig_name
= anon_vma_name(vma
);
110 vma
->anon_name
= NULL
;
111 anon_vma_name_put(orig_name
);
115 if (anon_vma_name_eq(orig_name
, anon_name
))
118 vma
->anon_name
= anon_vma_name_reuse(anon_name
);
119 anon_vma_name_put(orig_name
);
123 #else /* CONFIG_ANON_VMA_NAME */
124 static int replace_anon_vma_name(struct vm_area_struct
*vma
,
125 struct anon_vma_name
*anon_name
)
132 #endif /* CONFIG_ANON_VMA_NAME */
134 * Update the vm_flags on region of a vma, splitting it or merging it as
135 * necessary. Must be called with mmap_sem held for writing;
136 * Caller should ensure anon_name stability by raising its refcount even when
137 * anon_name belongs to a valid vma because this function might free that vma.
139 static int madvise_update_vma(struct vm_area_struct
*vma
,
140 struct vm_area_struct
**prev
, unsigned long start
,
141 unsigned long end
, unsigned long new_flags
,
142 struct anon_vma_name
*anon_name
)
144 struct mm_struct
*mm
= vma
->vm_mm
;
148 if (new_flags
== vma
->vm_flags
&& anon_vma_name_eq(anon_vma_name(vma
), anon_name
)) {
153 pgoff
= vma
->vm_pgoff
+ ((start
- vma
->vm_start
) >> PAGE_SHIFT
);
154 *prev
= vma_merge(mm
, *prev
, start
, end
, new_flags
, vma
->anon_vma
,
155 vma
->vm_file
, pgoff
, vma_policy(vma
),
156 vma
->vm_userfaultfd_ctx
, anon_name
);
164 if (start
!= vma
->vm_start
) {
165 if (unlikely(mm
->map_count
>= sysctl_max_map_count
))
167 error
= __split_vma(mm
, vma
, start
, 1);
172 if (end
!= vma
->vm_end
) {
173 if (unlikely(mm
->map_count
>= sysctl_max_map_count
))
175 error
= __split_vma(mm
, vma
, end
, 0);
182 * vm_flags is protected by the mmap_lock held in write mode.
184 vma
->vm_flags
= new_flags
;
186 error
= replace_anon_vma_name(vma
, anon_name
);
195 static int swapin_walk_pmd_entry(pmd_t
*pmd
, unsigned long start
,
196 unsigned long end
, struct mm_walk
*walk
)
198 struct vm_area_struct
*vma
= walk
->private;
200 struct swap_iocb
*splug
= NULL
;
202 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
205 for (index
= start
; index
!= end
; index
+= PAGE_SIZE
) {
212 ptep
= pte_offset_map_lock(vma
->vm_mm
, pmd
, index
, &ptl
);
214 pte_unmap_unlock(ptep
, ptl
);
216 if (!is_swap_pte(pte
))
218 entry
= pte_to_swp_entry(pte
);
219 if (unlikely(non_swap_entry(entry
)))
222 page
= read_swap_cache_async(entry
, GFP_HIGHUSER_MOVABLE
,
223 vma
, index
, false, &splug
);
227 swap_read_unplug(splug
);
232 static const struct mm_walk_ops swapin_walk_ops
= {
233 .pmd_entry
= swapin_walk_pmd_entry
,
236 static void force_shm_swapin_readahead(struct vm_area_struct
*vma
,
237 unsigned long start
, unsigned long end
,
238 struct address_space
*mapping
)
240 XA_STATE(xas
, &mapping
->i_pages
, linear_page_index(vma
, start
));
241 pgoff_t end_index
= linear_page_index(vma
, end
+ PAGE_SIZE
- 1);
243 struct swap_iocb
*splug
= NULL
;
246 xas_for_each(&xas
, page
, end_index
) {
249 if (!xa_is_value(page
))
251 swap
= radix_to_swp_entry(page
);
252 /* There might be swapin error entries in shmem mapping. */
253 if (non_swap_entry(swap
))
258 page
= read_swap_cache_async(swap
, GFP_HIGHUSER_MOVABLE
,
259 NULL
, 0, false, &splug
);
266 swap_read_unplug(splug
);
268 lru_add_drain(); /* Push any new pages onto the LRU now */
270 #endif /* CONFIG_SWAP */
273 * Schedule all required I/O operations. Do not wait for completion.
275 static long madvise_willneed(struct vm_area_struct
*vma
,
276 struct vm_area_struct
**prev
,
277 unsigned long start
, unsigned long end
)
279 struct mm_struct
*mm
= vma
->vm_mm
;
280 struct file
*file
= vma
->vm_file
;
286 walk_page_range(vma
->vm_mm
, start
, end
, &swapin_walk_ops
, vma
);
287 lru_add_drain(); /* Push any new pages onto the LRU now */
291 if (shmem_mapping(file
->f_mapping
)) {
292 force_shm_swapin_readahead(vma
, start
, end
,
301 if (IS_DAX(file_inode(file
))) {
302 /* no bad return value, but ignore advice */
307 * Filesystem's fadvise may need to take various locks. We need to
308 * explicitly grab a reference because the vma (and hence the
309 * vma's reference to the file) can go away as soon as we drop
312 *prev
= NULL
; /* tell sys_madvise we drop mmap_lock */
314 offset
= (loff_t
)(start
- vma
->vm_start
)
315 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
316 mmap_read_unlock(mm
);
317 vfs_fadvise(file
, offset
, end
- start
, POSIX_FADV_WILLNEED
);
323 static int madvise_cold_or_pageout_pte_range(pmd_t
*pmd
,
324 unsigned long addr
, unsigned long end
,
325 struct mm_walk
*walk
)
327 struct madvise_walk_private
*private = walk
->private;
328 struct mmu_gather
*tlb
= private->tlb
;
329 bool pageout
= private->pageout
;
330 struct mm_struct
*mm
= tlb
->mm
;
331 struct vm_area_struct
*vma
= walk
->vma
;
332 pte_t
*orig_pte
, *pte
, ptent
;
334 struct page
*page
= NULL
;
335 LIST_HEAD(page_list
);
337 if (fatal_signal_pending(current
))
340 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
341 if (pmd_trans_huge(*pmd
)) {
343 unsigned long next
= pmd_addr_end(addr
, end
);
345 tlb_change_page_size(tlb
, HPAGE_PMD_SIZE
);
346 ptl
= pmd_trans_huge_lock(pmd
, vma
);
351 if (is_huge_zero_pmd(orig_pmd
))
354 if (unlikely(!pmd_present(orig_pmd
))) {
355 VM_BUG_ON(thp_migration_supported() &&
356 !is_pmd_migration_entry(orig_pmd
));
360 page
= pmd_page(orig_pmd
);
362 /* Do not interfere with other mappings of this page */
363 if (page_mapcount(page
) != 1)
366 if (next
- addr
!= HPAGE_PMD_SIZE
) {
372 err
= split_huge_page(page
);
380 if (pmd_young(orig_pmd
)) {
381 pmdp_invalidate(vma
, addr
, pmd
);
382 orig_pmd
= pmd_mkold(orig_pmd
);
384 set_pmd_at(mm
, addr
, pmd
, orig_pmd
);
385 tlb_remove_pmd_tlb_entry(tlb
, pmd
, addr
);
388 ClearPageReferenced(page
);
389 test_and_clear_page_young(page
);
391 if (!isolate_lru_page(page
)) {
392 if (PageUnevictable(page
))
393 putback_lru_page(page
);
395 list_add(&page
->lru
, &page_list
);
398 deactivate_page(page
);
402 reclaim_pages(&page_list
);
407 if (pmd_trans_unstable(pmd
))
410 tlb_change_page_size(tlb
, PAGE_SIZE
);
411 orig_pte
= pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, addr
, &ptl
);
412 flush_tlb_batched_pending(mm
);
413 arch_enter_lazy_mmu_mode();
414 for (; addr
< end
; pte
++, addr
+= PAGE_SIZE
) {
420 if (!pte_present(ptent
))
423 page
= vm_normal_page(vma
, addr
, ptent
);
424 if (!page
|| is_zone_device_page(page
))
428 * Creating a THP page is expensive so split it only if we
429 * are sure it's worth. Split it if we are only owner.
431 if (PageTransCompound(page
)) {
432 if (page_mapcount(page
) != 1)
435 if (!trylock_page(page
)) {
439 pte_unmap_unlock(orig_pte
, ptl
);
440 if (split_huge_page(page
)) {
443 orig_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
448 orig_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
454 /* Do not interfere with other mappings of this page */
455 if (page_mapcount(page
) != 1)
458 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
460 if (pte_young(ptent
)) {
461 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
463 ptent
= pte_mkold(ptent
);
464 set_pte_at(mm
, addr
, pte
, ptent
);
465 tlb_remove_tlb_entry(tlb
, pte
, addr
);
469 * We are deactivating a page for accelerating reclaiming.
470 * VM couldn't reclaim the page unless we clear PG_young.
471 * As a side effect, it makes confuse idle-page tracking
472 * because they will miss recent referenced history.
474 ClearPageReferenced(page
);
475 test_and_clear_page_young(page
);
477 if (!isolate_lru_page(page
)) {
478 if (PageUnevictable(page
))
479 putback_lru_page(page
);
481 list_add(&page
->lru
, &page_list
);
484 deactivate_page(page
);
487 arch_leave_lazy_mmu_mode();
488 pte_unmap_unlock(orig_pte
, ptl
);
490 reclaim_pages(&page_list
);
496 static const struct mm_walk_ops cold_walk_ops
= {
497 .pmd_entry
= madvise_cold_or_pageout_pte_range
,
500 static void madvise_cold_page_range(struct mmu_gather
*tlb
,
501 struct vm_area_struct
*vma
,
502 unsigned long addr
, unsigned long end
)
504 struct madvise_walk_private walk_private
= {
509 tlb_start_vma(tlb
, vma
);
510 walk_page_range(vma
->vm_mm
, addr
, end
, &cold_walk_ops
, &walk_private
);
511 tlb_end_vma(tlb
, vma
);
514 static inline bool can_madv_lru_vma(struct vm_area_struct
*vma
)
516 return !(vma
->vm_flags
& (VM_LOCKED
|VM_PFNMAP
|VM_HUGETLB
));
519 static long madvise_cold(struct vm_area_struct
*vma
,
520 struct vm_area_struct
**prev
,
521 unsigned long start_addr
, unsigned long end_addr
)
523 struct mm_struct
*mm
= vma
->vm_mm
;
524 struct mmu_gather tlb
;
527 if (!can_madv_lru_vma(vma
))
531 tlb_gather_mmu(&tlb
, mm
);
532 madvise_cold_page_range(&tlb
, vma
, start_addr
, end_addr
);
533 tlb_finish_mmu(&tlb
);
538 static void madvise_pageout_page_range(struct mmu_gather
*tlb
,
539 struct vm_area_struct
*vma
,
540 unsigned long addr
, unsigned long end
)
542 struct madvise_walk_private walk_private
= {
547 tlb_start_vma(tlb
, vma
);
548 walk_page_range(vma
->vm_mm
, addr
, end
, &cold_walk_ops
, &walk_private
);
549 tlb_end_vma(tlb
, vma
);
552 static inline bool can_do_pageout(struct vm_area_struct
*vma
)
554 if (vma_is_anonymous(vma
))
559 * paging out pagecache only for non-anonymous mappings that correspond
560 * to the files the calling process could (if tried) open for writing;
561 * otherwise we'd be including shared non-exclusive mappings, which
562 * opens a side channel.
564 return inode_owner_or_capable(&init_user_ns
,
565 file_inode(vma
->vm_file
)) ||
566 file_permission(vma
->vm_file
, MAY_WRITE
) == 0;
569 static long madvise_pageout(struct vm_area_struct
*vma
,
570 struct vm_area_struct
**prev
,
571 unsigned long start_addr
, unsigned long end_addr
)
573 struct mm_struct
*mm
= vma
->vm_mm
;
574 struct mmu_gather tlb
;
577 if (!can_madv_lru_vma(vma
))
580 if (!can_do_pageout(vma
))
584 tlb_gather_mmu(&tlb
, mm
);
585 madvise_pageout_page_range(&tlb
, vma
, start_addr
, end_addr
);
586 tlb_finish_mmu(&tlb
);
591 static int madvise_free_pte_range(pmd_t
*pmd
, unsigned long addr
,
592 unsigned long end
, struct mm_walk
*walk
)
595 struct mmu_gather
*tlb
= walk
->private;
596 struct mm_struct
*mm
= tlb
->mm
;
597 struct vm_area_struct
*vma
= walk
->vma
;
599 pte_t
*orig_pte
, *pte
, ptent
;
604 next
= pmd_addr_end(addr
, end
);
605 if (pmd_trans_huge(*pmd
))
606 if (madvise_free_huge_pmd(tlb
, vma
, pmd
, addr
, next
))
609 if (pmd_trans_unstable(pmd
))
612 tlb_change_page_size(tlb
, PAGE_SIZE
);
613 orig_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
614 flush_tlb_batched_pending(mm
);
615 arch_enter_lazy_mmu_mode();
616 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
622 * If the pte has swp_entry, just clear page table to
623 * prevent swap-in which is more expensive rather than
624 * (page allocation + zeroing).
626 if (!pte_present(ptent
)) {
629 entry
= pte_to_swp_entry(ptent
);
630 if (!non_swap_entry(entry
)) {
632 free_swap_and_cache(entry
);
633 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
634 } else if (is_hwpoison_entry(entry
) ||
635 is_swapin_error_entry(entry
)) {
636 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
641 page
= vm_normal_page(vma
, addr
, ptent
);
642 if (!page
|| is_zone_device_page(page
))
646 * If pmd isn't transhuge but the page is THP and
647 * is owned by only this process, split it and
648 * deactivate all pages.
650 if (PageTransCompound(page
)) {
651 if (page_mapcount(page
) != 1)
654 if (!trylock_page(page
)) {
658 pte_unmap_unlock(orig_pte
, ptl
);
659 if (split_huge_page(page
)) {
662 orig_pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
667 orig_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
673 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
675 if (PageSwapCache(page
) || PageDirty(page
)) {
676 if (!trylock_page(page
))
679 * If page is shared with others, we couldn't clear
680 * PG_dirty of the page.
682 if (page_mapcount(page
) != 1) {
687 if (PageSwapCache(page
) && !try_to_free_swap(page
)) {
692 ClearPageDirty(page
);
696 if (pte_young(ptent
) || pte_dirty(ptent
)) {
698 * Some of architecture(ex, PPC) don't update TLB
699 * with set_pte_at and tlb_remove_tlb_entry so for
700 * the portability, remap the pte with old|clean
701 * after pte clearing.
703 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
706 ptent
= pte_mkold(ptent
);
707 ptent
= pte_mkclean(ptent
);
708 set_pte_at(mm
, addr
, pte
, ptent
);
709 tlb_remove_tlb_entry(tlb
, pte
, addr
);
711 mark_page_lazyfree(page
);
715 if (current
->mm
== mm
)
718 add_mm_counter(mm
, MM_SWAPENTS
, nr_swap
);
720 arch_leave_lazy_mmu_mode();
721 pte_unmap_unlock(orig_pte
, ptl
);
727 static const struct mm_walk_ops madvise_free_walk_ops
= {
728 .pmd_entry
= madvise_free_pte_range
,
731 static int madvise_free_single_vma(struct vm_area_struct
*vma
,
732 unsigned long start_addr
, unsigned long end_addr
)
734 struct mm_struct
*mm
= vma
->vm_mm
;
735 struct mmu_notifier_range range
;
736 struct mmu_gather tlb
;
738 /* MADV_FREE works for only anon vma at the moment */
739 if (!vma_is_anonymous(vma
))
742 range
.start
= max(vma
->vm_start
, start_addr
);
743 if (range
.start
>= vma
->vm_end
)
745 range
.end
= min(vma
->vm_end
, end_addr
);
746 if (range
.end
<= vma
->vm_start
)
748 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
749 range
.start
, range
.end
);
752 tlb_gather_mmu(&tlb
, mm
);
753 update_hiwater_rss(mm
);
755 mmu_notifier_invalidate_range_start(&range
);
756 tlb_start_vma(&tlb
, vma
);
757 walk_page_range(vma
->vm_mm
, range
.start
, range
.end
,
758 &madvise_free_walk_ops
, &tlb
);
759 tlb_end_vma(&tlb
, vma
);
760 mmu_notifier_invalidate_range_end(&range
);
761 tlb_finish_mmu(&tlb
);
767 * Application no longer needs these pages. If the pages are dirty,
768 * it's OK to just throw them away. The app will be more careful about
769 * data it wants to keep. Be sure to free swap resources too. The
770 * zap_page_range call sets things up for shrink_active_list to actually free
771 * these pages later if no one else has touched them in the meantime,
772 * although we could add these pages to a global reuse list for
773 * shrink_active_list to pick up before reclaiming other pages.
775 * NB: This interface discards data rather than pushes it out to swap,
776 * as some implementations do. This has performance implications for
777 * applications like large transactional databases which want to discard
778 * pages in anonymous maps after committing to backing store the data
779 * that was kept in them. There is no reason to write this data out to
780 * the swap area if the application is discarding it.
782 * An interface that causes the system to free clean pages and flush
783 * dirty pages is already available as msync(MS_INVALIDATE).
785 static long madvise_dontneed_single_vma(struct vm_area_struct
*vma
,
786 unsigned long start
, unsigned long end
)
788 zap_page_range(vma
, start
, end
- start
);
792 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct
*vma
,
797 if (!is_vm_hugetlb_page(vma
)) {
798 unsigned int forbidden
= VM_PFNMAP
;
800 if (behavior
!= MADV_DONTNEED_LOCKED
)
801 forbidden
|= VM_LOCKED
;
803 return !(vma
->vm_flags
& forbidden
);
806 if (behavior
!= MADV_DONTNEED
&& behavior
!= MADV_DONTNEED_LOCKED
)
808 if (start
& ~huge_page_mask(hstate_vma(vma
)))
811 *end
= ALIGN(*end
, huge_page_size(hstate_vma(vma
)));
815 static long madvise_dontneed_free(struct vm_area_struct
*vma
,
816 struct vm_area_struct
**prev
,
817 unsigned long start
, unsigned long end
,
820 struct mm_struct
*mm
= vma
->vm_mm
;
823 if (!madvise_dontneed_free_valid_vma(vma
, start
, &end
, behavior
))
826 if (!userfaultfd_remove(vma
, start
, end
)) {
827 *prev
= NULL
; /* mmap_lock has been dropped, prev is stale */
830 vma
= find_vma(mm
, start
);
833 if (start
< vma
->vm_start
) {
835 * This "vma" under revalidation is the one
836 * with the lowest vma->vm_start where start
837 * is also < vma->vm_end. If start <
838 * vma->vm_start it means an hole materialized
839 * in the user address space within the
840 * virtual range passed to MADV_DONTNEED
846 * Potential end adjustment for hugetlb vma is OK as
847 * the check below keeps end within vma.
849 if (!madvise_dontneed_free_valid_vma(vma
, start
, &end
,
852 if (end
> vma
->vm_end
) {
854 * Don't fail if end > vma->vm_end. If the old
855 * vma was split while the mmap_lock was
856 * released the effect of the concurrent
857 * operation may not cause madvise() to
858 * have an undefined result. There may be an
859 * adjacent next vma that we'll walk
860 * next. userfaultfd_remove() will generate an
861 * UFFD_EVENT_REMOVE repetition on the
862 * end-vma->vm_end range, but the manager can
863 * handle a repetition fine.
867 VM_WARN_ON(start
>= end
);
870 if (behavior
== MADV_DONTNEED
|| behavior
== MADV_DONTNEED_LOCKED
)
871 return madvise_dontneed_single_vma(vma
, start
, end
);
872 else if (behavior
== MADV_FREE
)
873 return madvise_free_single_vma(vma
, start
, end
);
878 static long madvise_populate(struct vm_area_struct
*vma
,
879 struct vm_area_struct
**prev
,
880 unsigned long start
, unsigned long end
,
883 const bool write
= behavior
== MADV_POPULATE_WRITE
;
884 struct mm_struct
*mm
= vma
->vm_mm
;
885 unsigned long tmp_end
;
891 while (start
< end
) {
893 * We might have temporarily dropped the lock. For example,
894 * our VMA might have been split.
896 if (!vma
|| start
>= vma
->vm_end
) {
897 vma
= vma_lookup(mm
, start
);
902 tmp_end
= min_t(unsigned long, end
, vma
->vm_end
);
903 /* Populate (prefault) page tables readable/writable. */
904 pages
= faultin_vma_page_range(vma
, start
, tmp_end
, write
,
916 case -EINVAL
: /* Incompatible mappings / permissions. */
920 case -EFAULT
: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
923 pr_warn_once("%s: unhandled return value: %ld\n",
930 start
+= pages
* PAGE_SIZE
;
936 * Application wants to free up the pages and associated backing store.
937 * This is effectively punching a hole into the middle of a file.
939 static long madvise_remove(struct vm_area_struct
*vma
,
940 struct vm_area_struct
**prev
,
941 unsigned long start
, unsigned long end
)
946 struct mm_struct
*mm
= vma
->vm_mm
;
948 *prev
= NULL
; /* tell sys_madvise we drop mmap_lock */
950 if (vma
->vm_flags
& VM_LOCKED
)
955 if (!f
|| !f
->f_mapping
|| !f
->f_mapping
->host
) {
959 if ((vma
->vm_flags
& (VM_SHARED
|VM_WRITE
)) != (VM_SHARED
|VM_WRITE
))
962 offset
= (loff_t
)(start
- vma
->vm_start
)
963 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
966 * Filesystem's fallocate may need to take i_rwsem. We need to
967 * explicitly grab a reference because the vma (and hence the
968 * vma's reference to the file) can go away as soon as we drop
972 if (userfaultfd_remove(vma
, start
, end
)) {
973 /* mmap_lock was not released by userfaultfd_remove() */
974 mmap_read_unlock(mm
);
976 error
= vfs_fallocate(f
,
977 FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
,
978 offset
, end
- start
);
985 * Apply an madvise behavior to a region of a vma. madvise_update_vma
986 * will handle splitting a vm area into separate areas, each area with its own
989 static int madvise_vma_behavior(struct vm_area_struct
*vma
,
990 struct vm_area_struct
**prev
,
991 unsigned long start
, unsigned long end
,
992 unsigned long behavior
)
995 struct anon_vma_name
*anon_name
;
996 unsigned long new_flags
= vma
->vm_flags
;
1000 return madvise_remove(vma
, prev
, start
, end
);
1002 return madvise_willneed(vma
, prev
, start
, end
);
1004 return madvise_cold(vma
, prev
, start
, end
);
1006 return madvise_pageout(vma
, prev
, start
, end
);
1009 case MADV_DONTNEED_LOCKED
:
1010 return madvise_dontneed_free(vma
, prev
, start
, end
, behavior
);
1011 case MADV_POPULATE_READ
:
1012 case MADV_POPULATE_WRITE
:
1013 return madvise_populate(vma
, prev
, start
, end
, behavior
);
1015 new_flags
= new_flags
& ~VM_RAND_READ
& ~VM_SEQ_READ
;
1017 case MADV_SEQUENTIAL
:
1018 new_flags
= (new_flags
& ~VM_RAND_READ
) | VM_SEQ_READ
;
1021 new_flags
= (new_flags
& ~VM_SEQ_READ
) | VM_RAND_READ
;
1024 new_flags
|= VM_DONTCOPY
;
1027 if (vma
->vm_flags
& VM_IO
)
1029 new_flags
&= ~VM_DONTCOPY
;
1031 case MADV_WIPEONFORK
:
1032 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1033 if (vma
->vm_file
|| vma
->vm_flags
& VM_SHARED
)
1035 new_flags
|= VM_WIPEONFORK
;
1037 case MADV_KEEPONFORK
:
1038 new_flags
&= ~VM_WIPEONFORK
;
1041 new_flags
|= VM_DONTDUMP
;
1044 if (!is_vm_hugetlb_page(vma
) && new_flags
& VM_SPECIAL
)
1046 new_flags
&= ~VM_DONTDUMP
;
1048 case MADV_MERGEABLE
:
1049 case MADV_UNMERGEABLE
:
1050 error
= ksm_madvise(vma
, start
, end
, behavior
, &new_flags
);
1055 case MADV_NOHUGEPAGE
:
1056 error
= hugepage_madvise(vma
, &new_flags
, behavior
);
1062 anon_name
= anon_vma_name(vma
);
1063 anon_vma_name_get(anon_name
);
1064 error
= madvise_update_vma(vma
, prev
, start
, end
, new_flags
,
1066 anon_vma_name_put(anon_name
);
1070 * madvise() returns EAGAIN if kernel resources, such as
1071 * slab, are temporarily unavailable.
1073 if (error
== -ENOMEM
)
1078 #ifdef CONFIG_MEMORY_FAILURE
1080 * Error injection support for memory error handling.
1082 static int madvise_inject_error(int behavior
,
1083 unsigned long start
, unsigned long end
)
1087 if (!capable(CAP_SYS_ADMIN
))
1091 for (; start
< end
; start
+= size
) {
1096 ret
= get_user_pages_fast(start
, 1, 0, &page
);
1099 pfn
= page_to_pfn(page
);
1102 * When soft offlining hugepages, after migrating the page
1103 * we dissolve it, therefore in the second loop "page" will
1104 * no longer be a compound page.
1106 size
= page_size(compound_head(page
));
1108 if (behavior
== MADV_SOFT_OFFLINE
) {
1109 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1111 ret
= soft_offline_page(pfn
, MF_COUNT_INCREASED
);
1113 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1115 ret
= memory_failure(pfn
, MF_COUNT_INCREASED
| MF_SW_SIMULATED
);
1116 if (ret
== -EOPNOTSUPP
)
1129 madvise_behavior_valid(int behavior
)
1135 case MADV_SEQUENTIAL
:
1140 case MADV_DONTNEED_LOCKED
:
1144 case MADV_POPULATE_READ
:
1145 case MADV_POPULATE_WRITE
:
1147 case MADV_MERGEABLE
:
1148 case MADV_UNMERGEABLE
:
1150 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1152 case MADV_NOHUGEPAGE
:
1156 case MADV_WIPEONFORK
:
1157 case MADV_KEEPONFORK
:
1158 #ifdef CONFIG_MEMORY_FAILURE
1159 case MADV_SOFT_OFFLINE
:
1170 process_madvise_behavior_valid(int behavior
)
1183 * Walk the vmas in range [start,end), and call the visit function on each one.
1184 * The visit function will get start and end parameters that cover the overlap
1185 * between the current vma and the original range. Any unmapped regions in the
1186 * original range will result in this function returning -ENOMEM while still
1187 * calling the visit function on all of the existing vmas in the range.
1188 * Must be called with the mmap_lock held for reading or writing.
1191 int madvise_walk_vmas(struct mm_struct
*mm
, unsigned long start
,
1192 unsigned long end
, unsigned long arg
,
1193 int (*visit
)(struct vm_area_struct
*vma
,
1194 struct vm_area_struct
**prev
, unsigned long start
,
1195 unsigned long end
, unsigned long arg
))
1197 struct vm_area_struct
*vma
;
1198 struct vm_area_struct
*prev
;
1200 int unmapped_error
= 0;
1203 * If the interval [start,end) covers some unmapped address
1204 * ranges, just ignore them, but return -ENOMEM at the end.
1205 * - different from the way of handling in mlock etc.
1207 vma
= find_vma_prev(mm
, start
, &prev
);
1208 if (vma
&& start
> vma
->vm_start
)
1214 /* Still start < end. */
1218 /* Here start < (end|vma->vm_end). */
1219 if (start
< vma
->vm_start
) {
1220 unmapped_error
= -ENOMEM
;
1221 start
= vma
->vm_start
;
1226 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1231 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1232 error
= visit(vma
, &prev
, start
, tmp
, arg
);
1236 if (prev
&& start
< prev
->vm_end
)
1237 start
= prev
->vm_end
;
1241 vma
= prev
->vm_next
;
1242 else /* madvise_remove dropped mmap_lock */
1243 vma
= find_vma(mm
, start
);
1246 return unmapped_error
;
1249 #ifdef CONFIG_ANON_VMA_NAME
1250 static int madvise_vma_anon_name(struct vm_area_struct
*vma
,
1251 struct vm_area_struct
**prev
,
1252 unsigned long start
, unsigned long end
,
1253 unsigned long anon_name
)
1257 /* Only anonymous mappings can be named */
1261 error
= madvise_update_vma(vma
, prev
, start
, end
, vma
->vm_flags
,
1262 (struct anon_vma_name
*)anon_name
);
1265 * madvise() returns EAGAIN if kernel resources, such as
1266 * slab, are temporarily unavailable.
1268 if (error
== -ENOMEM
)
1273 int madvise_set_anon_name(struct mm_struct
*mm
, unsigned long start
,
1274 unsigned long len_in
, struct anon_vma_name
*anon_name
)
1279 if (start
& ~PAGE_MASK
)
1281 len
= (len_in
+ ~PAGE_MASK
) & PAGE_MASK
;
1283 /* Check to see whether len was rounded up from small -ve to zero */
1294 return madvise_walk_vmas(mm
, start
, end
, (unsigned long)anon_name
,
1295 madvise_vma_anon_name
);
1297 #endif /* CONFIG_ANON_VMA_NAME */
1299 * The madvise(2) system call.
1301 * Applications can use madvise() to advise the kernel how it should
1302 * handle paging I/O in this VM area. The idea is to help the kernel
1303 * use appropriate read-ahead and caching techniques. The information
1304 * provided is advisory only, and can be safely disregarded by the
1305 * kernel without affecting the correct operation of the application.
1308 * MADV_NORMAL - the default behavior is to read clusters. This
1309 * results in some read-ahead and read-behind.
1310 * MADV_RANDOM - the system should read the minimum amount of data
1311 * on any access, since it is unlikely that the appli-
1312 * cation will need more than what it asks for.
1313 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1314 * once, so they can be aggressively read ahead, and
1315 * can be freed soon after they are accessed.
1316 * MADV_WILLNEED - the application is notifying the system to read
1318 * MADV_DONTNEED - the application is finished with the given range,
1319 * so the kernel can free resources associated with it.
1320 * MADV_FREE - the application marks pages in the given range as lazy free,
1321 * where actual purges are postponed until memory pressure happens.
1322 * MADV_REMOVE - the application wants to free up the given range of
1323 * pages and associated backing store.
1324 * MADV_DONTFORK - omit this area from child's address space when forking:
1325 * typically, to avoid COWing pages pinned by get_user_pages().
1326 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1327 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1328 * range after a fork.
1329 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1330 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1331 * were corrupted by unrecoverable hardware memory failure.
1332 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1333 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1334 * this area with pages of identical content from other such areas.
1335 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1336 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1337 * huge pages in the future. Existing pages might be coalesced and
1338 * new pages might be allocated as THP.
1339 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1340 * transparent huge pages so the existing pages will not be
1341 * coalesced into THP and new pages will not be allocated as THP.
1342 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1343 * from being included in its core dump.
1344 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1345 * MADV_COLD - the application is not expected to use this memory soon,
1346 * deactivate pages in this range so that they can be reclaimed
1347 * easily if memory pressure happens.
1348 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1349 * page out the pages in this range immediately.
1350 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1351 * triggering read faults if required
1352 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1353 * triggering write faults if required
1357 * -EINVAL - start + len < 0, start is not page-aligned,
1358 * "behavior" is not a valid value, or application
1359 * is attempting to release locked or shared pages,
1360 * or the specified address range includes file, Huge TLB,
1361 * MAP_SHARED or VMPFNMAP range.
1362 * -ENOMEM - addresses in the specified range are not currently
1363 * mapped, or are outside the AS of the process.
1364 * -EIO - an I/O error occurred while paging in data.
1365 * -EBADF - map exists, but area maps something that isn't a file.
1366 * -EAGAIN - a kernel resource was temporarily unavailable.
1368 int do_madvise(struct mm_struct
*mm
, unsigned long start
, size_t len_in
, int behavior
)
1374 struct blk_plug plug
;
1376 start
= untagged_addr(start
);
1378 if (!madvise_behavior_valid(behavior
))
1381 if (!PAGE_ALIGNED(start
))
1383 len
= PAGE_ALIGN(len_in
);
1385 /* Check to see whether len was rounded up from small -ve to zero */
1396 #ifdef CONFIG_MEMORY_FAILURE
1397 if (behavior
== MADV_HWPOISON
|| behavior
== MADV_SOFT_OFFLINE
)
1398 return madvise_inject_error(behavior
, start
, start
+ len_in
);
1401 write
= madvise_need_mmap_write(behavior
);
1403 if (mmap_write_lock_killable(mm
))
1409 blk_start_plug(&plug
);
1410 error
= madvise_walk_vmas(mm
, start
, end
, behavior
,
1411 madvise_vma_behavior
);
1412 blk_finish_plug(&plug
);
1414 mmap_write_unlock(mm
);
1416 mmap_read_unlock(mm
);
1421 SYSCALL_DEFINE3(madvise
, unsigned long, start
, size_t, len_in
, int, behavior
)
1423 return do_madvise(current
->mm
, start
, len_in
, behavior
);
1426 SYSCALL_DEFINE5(process_madvise
, int, pidfd
, const struct iovec __user
*, vec
,
1427 size_t, vlen
, int, behavior
, unsigned int, flags
)
1430 struct iovec iovstack
[UIO_FASTIOV
], iovec
;
1431 struct iovec
*iov
= iovstack
;
1432 struct iov_iter iter
;
1433 struct task_struct
*task
;
1434 struct mm_struct
*mm
;
1436 unsigned int f_flags
;
1443 ret
= import_iovec(READ
, vec
, vlen
, ARRAY_SIZE(iovstack
), &iov
, &iter
);
1447 task
= pidfd_get_task(pidfd
, &f_flags
);
1449 ret
= PTR_ERR(task
);
1453 if (!process_madvise_behavior_valid(behavior
)) {
1458 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1459 mm
= mm_access(task
, PTRACE_MODE_READ_FSCREDS
);
1460 if (IS_ERR_OR_NULL(mm
)) {
1461 ret
= IS_ERR(mm
) ? PTR_ERR(mm
) : -ESRCH
;
1466 * Require CAP_SYS_NICE for influencing process performance. Note that
1467 * only non-destructive hints are currently supported.
1469 if (!capable(CAP_SYS_NICE
)) {
1474 total_len
= iov_iter_count(&iter
);
1476 while (iov_iter_count(&iter
)) {
1477 iovec
= iov_iter_iovec(&iter
);
1478 ret
= do_madvise(mm
, (unsigned long)iovec
.iov_base
,
1479 iovec
.iov_len
, behavior
);
1482 iov_iter_advance(&iter
, iovec
.iov_len
);
1485 ret
= (total_len
- iov_iter_count(&iter
)) ? : ret
;
1490 put_task_struct(task
);