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/userfaultfd_k.h>
15 #include <linux/hugetlb.h>
16 #include <linux/falloc.h>
17 #include <linux/fadvise.h>
18 #include <linux/sched.h>
19 #include <linux/ksm.h>
21 #include <linux/file.h>
22 #include <linux/blkdev.h>
23 #include <linux/backing-dev.h>
24 #include <linux/pagewalk.h>
25 #include <linux/swap.h>
26 #include <linux/swapops.h>
27 #include <linux/shmem_fs.h>
28 #include <linux/mmu_notifier.h>
35 * Any behaviour which results in changes to the vma->vm_flags needs to
36 * take mmap_sem for writing. Others, which simply traverse vmas, need
37 * to only take it for reading.
39 static int madvise_need_mmap_write(int behavior
)
48 /* be safe, default to 1. list exceptions explicitly */
54 * We can potentially split a vm area into separate
55 * areas, each area with its own behavior.
57 static long madvise_behavior(struct vm_area_struct
*vma
,
58 struct vm_area_struct
**prev
,
59 unsigned long start
, unsigned long end
, int behavior
)
61 struct mm_struct
*mm
= vma
->vm_mm
;
64 unsigned long new_flags
= vma
->vm_flags
;
68 new_flags
= new_flags
& ~VM_RAND_READ
& ~VM_SEQ_READ
;
71 new_flags
= (new_flags
& ~VM_RAND_READ
) | VM_SEQ_READ
;
74 new_flags
= (new_flags
& ~VM_SEQ_READ
) | VM_RAND_READ
;
77 new_flags
|= VM_DONTCOPY
;
80 if (vma
->vm_flags
& VM_IO
) {
84 new_flags
&= ~VM_DONTCOPY
;
87 /* MADV_WIPEONFORK is only supported on anonymous memory. */
88 if (vma
->vm_file
|| vma
->vm_flags
& VM_SHARED
) {
92 new_flags
|= VM_WIPEONFORK
;
95 new_flags
&= ~VM_WIPEONFORK
;
98 new_flags
|= VM_DONTDUMP
;
101 if (!is_vm_hugetlb_page(vma
) && new_flags
& VM_SPECIAL
) {
105 new_flags
&= ~VM_DONTDUMP
;
108 case MADV_UNMERGEABLE
:
109 error
= ksm_madvise(vma
, start
, end
, behavior
, &new_flags
);
112 * madvise() returns EAGAIN if kernel resources, such as
113 * slab, are temporarily unavailable.
115 if (error
== -ENOMEM
)
121 case MADV_NOHUGEPAGE
:
122 error
= hugepage_madvise(vma
, &new_flags
, behavior
);
125 * madvise() returns EAGAIN if kernel resources, such as
126 * slab, are temporarily unavailable.
128 if (error
== -ENOMEM
)
135 if (new_flags
== vma
->vm_flags
) {
140 pgoff
= vma
->vm_pgoff
+ ((start
- vma
->vm_start
) >> PAGE_SHIFT
);
141 *prev
= vma_merge(mm
, *prev
, start
, end
, new_flags
, vma
->anon_vma
,
142 vma
->vm_file
, pgoff
, vma_policy(vma
),
143 vma
->vm_userfaultfd_ctx
);
151 if (start
!= vma
->vm_start
) {
152 if (unlikely(mm
->map_count
>= sysctl_max_map_count
)) {
156 error
= __split_vma(mm
, vma
, start
, 1);
159 * madvise() returns EAGAIN if kernel resources, such as
160 * slab, are temporarily unavailable.
162 if (error
== -ENOMEM
)
168 if (end
!= vma
->vm_end
) {
169 if (unlikely(mm
->map_count
>= sysctl_max_map_count
)) {
173 error
= __split_vma(mm
, vma
, end
, 0);
176 * madvise() returns EAGAIN if kernel resources, such as
177 * slab, are temporarily unavailable.
179 if (error
== -ENOMEM
)
187 * vm_flags is protected by the mmap_sem held in write mode.
189 vma
->vm_flags
= new_flags
;
195 static int swapin_walk_pmd_entry(pmd_t
*pmd
, unsigned long start
,
196 unsigned long end
, struct mm_walk
*walk
)
199 struct vm_area_struct
*vma
= walk
->private;
202 if (pmd_none_or_trans_huge_or_clear_bad(pmd
))
205 for (index
= start
; index
!= end
; index
+= PAGE_SIZE
) {
211 orig_pte
= pte_offset_map_lock(vma
->vm_mm
, pmd
, start
, &ptl
);
212 pte
= *(orig_pte
+ ((index
- start
) / PAGE_SIZE
));
213 pte_unmap_unlock(orig_pte
, ptl
);
215 if (pte_present(pte
) || pte_none(pte
))
217 entry
= pte_to_swp_entry(pte
);
218 if (unlikely(non_swap_entry(entry
)))
221 page
= read_swap_cache_async(entry
, GFP_HIGHUSER_MOVABLE
,
230 static const struct mm_walk_ops swapin_walk_ops
= {
231 .pmd_entry
= swapin_walk_pmd_entry
,
234 static void force_shm_swapin_readahead(struct vm_area_struct
*vma
,
235 unsigned long start
, unsigned long end
,
236 struct address_space
*mapping
)
242 for (; start
< end
; start
+= PAGE_SIZE
) {
243 index
= ((start
- vma
->vm_start
) >> PAGE_SHIFT
) + vma
->vm_pgoff
;
245 page
= find_get_entry(mapping
, index
);
246 if (!xa_is_value(page
)) {
251 swap
= radix_to_swp_entry(page
);
252 page
= read_swap_cache_async(swap
, GFP_HIGHUSER_MOVABLE
,
258 lru_add_drain(); /* Push any new pages onto the LRU now */
260 #endif /* CONFIG_SWAP */
263 * Schedule all required I/O operations. Do not wait for completion.
265 static long madvise_willneed(struct vm_area_struct
*vma
,
266 struct vm_area_struct
**prev
,
267 unsigned long start
, unsigned long end
)
269 struct file
*file
= vma
->vm_file
;
275 walk_page_range(vma
->vm_mm
, start
, end
, &swapin_walk_ops
, vma
);
276 lru_add_drain(); /* Push any new pages onto the LRU now */
280 if (shmem_mapping(file
->f_mapping
)) {
281 force_shm_swapin_readahead(vma
, start
, end
,
290 if (IS_DAX(file_inode(file
))) {
291 /* no bad return value, but ignore advice */
296 * Filesystem's fadvise may need to take various locks. We need to
297 * explicitly grab a reference because the vma (and hence the
298 * vma's reference to the file) can go away as soon as we drop
301 *prev
= NULL
; /* tell sys_madvise we drop mmap_sem */
303 up_read(¤t
->mm
->mmap_sem
);
304 offset
= (loff_t
)(start
- vma
->vm_start
)
305 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
306 vfs_fadvise(file
, offset
, end
- start
, POSIX_FADV_WILLNEED
);
308 down_read(¤t
->mm
->mmap_sem
);
312 static int madvise_free_pte_range(pmd_t
*pmd
, unsigned long addr
,
313 unsigned long end
, struct mm_walk
*walk
)
316 struct mmu_gather
*tlb
= walk
->private;
317 struct mm_struct
*mm
= tlb
->mm
;
318 struct vm_area_struct
*vma
= walk
->vma
;
320 pte_t
*orig_pte
, *pte
, ptent
;
325 next
= pmd_addr_end(addr
, end
);
326 if (pmd_trans_huge(*pmd
))
327 if (madvise_free_huge_pmd(tlb
, vma
, pmd
, addr
, next
))
330 if (pmd_trans_unstable(pmd
))
333 tlb_change_page_size(tlb
, PAGE_SIZE
);
334 orig_pte
= pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
335 flush_tlb_batched_pending(mm
);
336 arch_enter_lazy_mmu_mode();
337 for (; addr
!= end
; pte
++, addr
+= PAGE_SIZE
) {
343 * If the pte has swp_entry, just clear page table to
344 * prevent swap-in which is more expensive rather than
345 * (page allocation + zeroing).
347 if (!pte_present(ptent
)) {
350 entry
= pte_to_swp_entry(ptent
);
351 if (non_swap_entry(entry
))
354 free_swap_and_cache(entry
);
355 pte_clear_not_present_full(mm
, addr
, pte
, tlb
->fullmm
);
359 page
= vm_normal_page(vma
, addr
, ptent
);
364 * If pmd isn't transhuge but the page is THP and
365 * is owned by only this process, split it and
366 * deactivate all pages.
368 if (PageTransCompound(page
)) {
369 if (page_mapcount(page
) != 1)
372 if (!trylock_page(page
)) {
376 pte_unmap_unlock(orig_pte
, ptl
);
377 if (split_huge_page(page
)) {
380 pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
385 pte
= pte_offset_map_lock(mm
, pmd
, addr
, &ptl
);
391 VM_BUG_ON_PAGE(PageTransCompound(page
), page
);
393 if (PageSwapCache(page
) || PageDirty(page
)) {
394 if (!trylock_page(page
))
397 * If page is shared with others, we couldn't clear
398 * PG_dirty of the page.
400 if (page_mapcount(page
) != 1) {
405 if (PageSwapCache(page
) && !try_to_free_swap(page
)) {
410 ClearPageDirty(page
);
414 if (pte_young(ptent
) || pte_dirty(ptent
)) {
416 * Some of architecture(ex, PPC) don't update TLB
417 * with set_pte_at and tlb_remove_tlb_entry so for
418 * the portability, remap the pte with old|clean
419 * after pte clearing.
421 ptent
= ptep_get_and_clear_full(mm
, addr
, pte
,
424 ptent
= pte_mkold(ptent
);
425 ptent
= pte_mkclean(ptent
);
426 set_pte_at(mm
, addr
, pte
, ptent
);
427 tlb_remove_tlb_entry(tlb
, pte
, addr
);
429 mark_page_lazyfree(page
);
433 if (current
->mm
== mm
)
436 add_mm_counter(mm
, MM_SWAPENTS
, nr_swap
);
438 arch_leave_lazy_mmu_mode();
439 pte_unmap_unlock(orig_pte
, ptl
);
445 static const struct mm_walk_ops madvise_free_walk_ops
= {
446 .pmd_entry
= madvise_free_pte_range
,
449 static int madvise_free_single_vma(struct vm_area_struct
*vma
,
450 unsigned long start_addr
, unsigned long end_addr
)
452 struct mm_struct
*mm
= vma
->vm_mm
;
453 struct mmu_notifier_range range
;
454 struct mmu_gather tlb
;
456 /* MADV_FREE works for only anon vma at the moment */
457 if (!vma_is_anonymous(vma
))
460 range
.start
= max(vma
->vm_start
, start_addr
);
461 if (range
.start
>= vma
->vm_end
)
463 range
.end
= min(vma
->vm_end
, end_addr
);
464 if (range
.end
<= vma
->vm_start
)
466 mmu_notifier_range_init(&range
, MMU_NOTIFY_CLEAR
, 0, vma
, mm
,
467 range
.start
, range
.end
);
470 tlb_gather_mmu(&tlb
, mm
, range
.start
, range
.end
);
471 update_hiwater_rss(mm
);
473 mmu_notifier_invalidate_range_start(&range
);
474 tlb_start_vma(&tlb
, vma
);
475 walk_page_range(vma
->vm_mm
, range
.start
, range
.end
,
476 &madvise_free_walk_ops
, &tlb
);
477 tlb_end_vma(&tlb
, vma
);
478 mmu_notifier_invalidate_range_end(&range
);
479 tlb_finish_mmu(&tlb
, range
.start
, range
.end
);
485 * Application no longer needs these pages. If the pages are dirty,
486 * it's OK to just throw them away. The app will be more careful about
487 * data it wants to keep. Be sure to free swap resources too. The
488 * zap_page_range call sets things up for shrink_active_list to actually free
489 * these pages later if no one else has touched them in the meantime,
490 * although we could add these pages to a global reuse list for
491 * shrink_active_list to pick up before reclaiming other pages.
493 * NB: This interface discards data rather than pushes it out to swap,
494 * as some implementations do. This has performance implications for
495 * applications like large transactional databases which want to discard
496 * pages in anonymous maps after committing to backing store the data
497 * that was kept in them. There is no reason to write this data out to
498 * the swap area if the application is discarding it.
500 * An interface that causes the system to free clean pages and flush
501 * dirty pages is already available as msync(MS_INVALIDATE).
503 static long madvise_dontneed_single_vma(struct vm_area_struct
*vma
,
504 unsigned long start
, unsigned long end
)
506 zap_page_range(vma
, start
, end
- start
);
510 static long madvise_dontneed_free(struct vm_area_struct
*vma
,
511 struct vm_area_struct
**prev
,
512 unsigned long start
, unsigned long end
,
516 if (!can_madv_dontneed_vma(vma
))
519 if (!userfaultfd_remove(vma
, start
, end
)) {
520 *prev
= NULL
; /* mmap_sem has been dropped, prev is stale */
522 down_read(¤t
->mm
->mmap_sem
);
523 vma
= find_vma(current
->mm
, start
);
526 if (start
< vma
->vm_start
) {
528 * This "vma" under revalidation is the one
529 * with the lowest vma->vm_start where start
530 * is also < vma->vm_end. If start <
531 * vma->vm_start it means an hole materialized
532 * in the user address space within the
533 * virtual range passed to MADV_DONTNEED
538 if (!can_madv_dontneed_vma(vma
))
540 if (end
> vma
->vm_end
) {
542 * Don't fail if end > vma->vm_end. If the old
543 * vma was splitted while the mmap_sem was
544 * released the effect of the concurrent
545 * operation may not cause madvise() to
546 * have an undefined result. There may be an
547 * adjacent next vma that we'll walk
548 * next. userfaultfd_remove() will generate an
549 * UFFD_EVENT_REMOVE repetition on the
550 * end-vma->vm_end range, but the manager can
551 * handle a repetition fine.
555 VM_WARN_ON(start
>= end
);
558 if (behavior
== MADV_DONTNEED
)
559 return madvise_dontneed_single_vma(vma
, start
, end
);
560 else if (behavior
== MADV_FREE
)
561 return madvise_free_single_vma(vma
, start
, end
);
567 * Application wants to free up the pages and associated backing store.
568 * This is effectively punching a hole into the middle of a file.
570 static long madvise_remove(struct vm_area_struct
*vma
,
571 struct vm_area_struct
**prev
,
572 unsigned long start
, unsigned long end
)
578 *prev
= NULL
; /* tell sys_madvise we drop mmap_sem */
580 if (vma
->vm_flags
& VM_LOCKED
)
585 if (!f
|| !f
->f_mapping
|| !f
->f_mapping
->host
) {
589 if ((vma
->vm_flags
& (VM_SHARED
|VM_WRITE
)) != (VM_SHARED
|VM_WRITE
))
592 offset
= (loff_t
)(start
- vma
->vm_start
)
593 + ((loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
);
596 * Filesystem's fallocate may need to take i_mutex. We need to
597 * explicitly grab a reference because the vma (and hence the
598 * vma's reference to the file) can go away as soon as we drop
602 if (userfaultfd_remove(vma
, start
, end
)) {
603 /* mmap_sem was not released by userfaultfd_remove() */
604 up_read(¤t
->mm
->mmap_sem
);
606 error
= vfs_fallocate(f
,
607 FALLOC_FL_PUNCH_HOLE
| FALLOC_FL_KEEP_SIZE
,
608 offset
, end
- start
);
610 down_read(¤t
->mm
->mmap_sem
);
614 #ifdef CONFIG_MEMORY_FAILURE
616 * Error injection support for memory error handling.
618 static int madvise_inject_error(int behavior
,
619 unsigned long start
, unsigned long end
)
625 if (!capable(CAP_SYS_ADMIN
))
629 for (; start
< end
; start
+= PAGE_SIZE
<< order
) {
633 ret
= get_user_pages_fast(start
, 1, 0, &page
);
636 pfn
= page_to_pfn(page
);
639 * When soft offlining hugepages, after migrating the page
640 * we dissolve it, therefore in the second loop "page" will
641 * no longer be a compound page, and order will be 0.
643 order
= compound_order(compound_head(page
));
645 if (PageHWPoison(page
)) {
650 if (behavior
== MADV_SOFT_OFFLINE
) {
651 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
654 ret
= soft_offline_page(page
, MF_COUNT_INCREASED
);
660 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
664 * Drop the page reference taken by get_user_pages_fast(). In
665 * the absence of MF_COUNT_INCREASED the memory_failure()
666 * routine is responsible for pinning the page to prevent it
667 * from being released back to the page allocator.
670 ret
= memory_failure(pfn
, 0);
675 /* Ensure that all poisoned pages are removed from per-cpu lists */
676 for_each_populated_zone(zone
)
677 drain_all_pages(zone
);
684 madvise_vma(struct vm_area_struct
*vma
, struct vm_area_struct
**prev
,
685 unsigned long start
, unsigned long end
, int behavior
)
689 return madvise_remove(vma
, prev
, start
, end
);
691 return madvise_willneed(vma
, prev
, start
, end
);
694 return madvise_dontneed_free(vma
, prev
, start
, end
, behavior
);
696 return madvise_behavior(vma
, prev
, start
, end
, behavior
);
701 madvise_behavior_valid(int behavior
)
707 case MADV_SEQUENTIAL
:
715 case MADV_UNMERGEABLE
:
717 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
719 case MADV_NOHUGEPAGE
:
723 case MADV_WIPEONFORK
:
724 case MADV_KEEPONFORK
:
725 #ifdef CONFIG_MEMORY_FAILURE
726 case MADV_SOFT_OFFLINE
:
737 * The madvise(2) system call.
739 * Applications can use madvise() to advise the kernel how it should
740 * handle paging I/O in this VM area. The idea is to help the kernel
741 * use appropriate read-ahead and caching techniques. The information
742 * provided is advisory only, and can be safely disregarded by the
743 * kernel without affecting the correct operation of the application.
746 * MADV_NORMAL - the default behavior is to read clusters. This
747 * results in some read-ahead and read-behind.
748 * MADV_RANDOM - the system should read the minimum amount of data
749 * on any access, since it is unlikely that the appli-
750 * cation will need more than what it asks for.
751 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
752 * once, so they can be aggressively read ahead, and
753 * can be freed soon after they are accessed.
754 * MADV_WILLNEED - the application is notifying the system to read
756 * MADV_DONTNEED - the application is finished with the given range,
757 * so the kernel can free resources associated with it.
758 * MADV_FREE - the application marks pages in the given range as lazy free,
759 * where actual purges are postponed until memory pressure happens.
760 * MADV_REMOVE - the application wants to free up the given range of
761 * pages and associated backing store.
762 * MADV_DONTFORK - omit this area from child's address space when forking:
763 * typically, to avoid COWing pages pinned by get_user_pages().
764 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
765 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
766 * range after a fork.
767 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
768 * MADV_HWPOISON - trigger memory error handler as if the given memory range
769 * were corrupted by unrecoverable hardware memory failure.
770 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
771 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
772 * this area with pages of identical content from other such areas.
773 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
774 * MADV_HUGEPAGE - the application wants to back the given range by transparent
775 * huge pages in the future. Existing pages might be coalesced and
776 * new pages might be allocated as THP.
777 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
778 * transparent huge pages so the existing pages will not be
779 * coalesced into THP and new pages will not be allocated as THP.
780 * MADV_DONTDUMP - the application wants to prevent pages in the given range
781 * from being included in its core dump.
782 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
786 * -EINVAL - start + len < 0, start is not page-aligned,
787 * "behavior" is not a valid value, or application
788 * is attempting to release locked or shared pages,
789 * or the specified address range includes file, Huge TLB,
790 * MAP_SHARED or VMPFNMAP range.
791 * -ENOMEM - addresses in the specified range are not currently
792 * mapped, or are outside the AS of the process.
793 * -EIO - an I/O error occurred while paging in data.
794 * -EBADF - map exists, but area maps something that isn't a file.
795 * -EAGAIN - a kernel resource was temporarily unavailable.
797 SYSCALL_DEFINE3(madvise
, unsigned long, start
, size_t, len_in
, int, behavior
)
799 unsigned long end
, tmp
;
800 struct vm_area_struct
*vma
, *prev
;
801 int unmapped_error
= 0;
805 struct blk_plug plug
;
807 if (!madvise_behavior_valid(behavior
))
810 if (start
& ~PAGE_MASK
)
812 len
= (len_in
+ ~PAGE_MASK
) & PAGE_MASK
;
814 /* Check to see whether len was rounded up from small -ve to zero */
826 #ifdef CONFIG_MEMORY_FAILURE
827 if (behavior
== MADV_HWPOISON
|| behavior
== MADV_SOFT_OFFLINE
)
828 return madvise_inject_error(behavior
, start
, start
+ len_in
);
831 write
= madvise_need_mmap_write(behavior
);
833 if (down_write_killable(¤t
->mm
->mmap_sem
))
836 down_read(¤t
->mm
->mmap_sem
);
840 * If the interval [start,end) covers some unmapped address
841 * ranges, just ignore them, but return -ENOMEM at the end.
842 * - different from the way of handling in mlock etc.
844 vma
= find_vma_prev(current
->mm
, start
, &prev
);
845 if (vma
&& start
> vma
->vm_start
)
848 blk_start_plug(&plug
);
850 /* Still start < end. */
855 /* Here start < (end|vma->vm_end). */
856 if (start
< vma
->vm_start
) {
857 unmapped_error
= -ENOMEM
;
858 start
= vma
->vm_start
;
863 /* Here vma->vm_start <= start < (end|vma->vm_end) */
868 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
869 error
= madvise_vma(vma
, &prev
, start
, tmp
, behavior
);
873 if (prev
&& start
< prev
->vm_end
)
874 start
= prev
->vm_end
;
875 error
= unmapped_error
;
880 else /* madvise_remove dropped mmap_sem */
881 vma
= find_vma(current
->mm
, start
);
884 blk_finish_plug(&plug
);
886 up_write(¤t
->mm
->mmap_sem
);
888 up_read(¤t
->mm
->mmap_sem
);