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/sched.h>
18 #include <linux/ksm.h>
20 #include <linux/file.h>
21 #include <linux/blkdev.h>
22 #include <linux/backing-dev.h>
23 #include <linux/swap.h>
24 #include <linux/swapops.h>
25 #include <linux/shmem_fs.h>
26 #include <linux/mmu_notifier.h>
33 * Any behaviour which results in changes to the vma->vm_flags needs to
34 * take mmap_sem for writing. Others, which simply traverse vmas, need
35 * to only take it for reading.
37 static int madvise_need_mmap_write(int behavior)
46 /* be safe, default to 1. list exceptions explicitly */
52 * We can potentially split a vm area into separate
53 * areas, each area with its own behavior.
55 static long madvise_behavior(struct vm_area_struct *vma,
56 struct vm_area_struct **prev,
57 unsigned long start, unsigned long end, int behavior)
59 struct mm_struct *mm = vma->vm_mm;
62 unsigned long new_flags = vma->vm_flags;
66 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
69 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
72 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
75 new_flags |= VM_DONTCOPY;
78 if (vma->vm_flags & VM_IO) {
82 new_flags &= ~VM_DONTCOPY;
85 /* MADV_WIPEONFORK is only supported on anonymous memory. */
86 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
90 new_flags |= VM_WIPEONFORK;
93 new_flags &= ~VM_WIPEONFORK;
96 new_flags |= VM_DONTDUMP;
99 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
103 new_flags &= ~VM_DONTDUMP;
106 case MADV_UNMERGEABLE:
107 error = ksm_madvise(vma, start, end, behavior, &new_flags);
110 * madvise() returns EAGAIN if kernel resources, such as
111 * slab, are temporarily unavailable.
113 if (error == -ENOMEM)
119 case MADV_NOHUGEPAGE:
120 error = hugepage_madvise(vma, &new_flags, behavior);
123 * madvise() returns EAGAIN if kernel resources, such as
124 * slab, are temporarily unavailable.
126 if (error == -ENOMEM)
133 if (new_flags == vma->vm_flags) {
138 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
139 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
140 vma->vm_file, pgoff, vma_policy(vma),
141 vma->vm_userfaultfd_ctx);
149 if (start != vma->vm_start) {
150 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
154 error = __split_vma(mm, vma, start, 1);
157 * madvise() returns EAGAIN if kernel resources, such as
158 * slab, are temporarily unavailable.
160 if (error == -ENOMEM)
166 if (end != vma->vm_end) {
167 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
171 error = __split_vma(mm, vma, end, 0);
174 * madvise() returns EAGAIN if kernel resources, such as
175 * slab, are temporarily unavailable.
177 if (error == -ENOMEM)
185 * vm_flags is protected by the mmap_sem held in write mode.
187 vma->vm_flags = new_flags;
193 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
194 unsigned long end, struct mm_walk *walk)
197 struct vm_area_struct *vma = walk->private;
200 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
203 for (index = start; index != end; index += PAGE_SIZE) {
209 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
210 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
211 pte_unmap_unlock(orig_pte, ptl);
213 if (pte_present(pte) || pte_none(pte))
215 entry = pte_to_swp_entry(pte);
216 if (unlikely(non_swap_entry(entry)))
219 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
228 static void force_swapin_readahead(struct vm_area_struct *vma,
229 unsigned long start, unsigned long end)
231 struct mm_walk walk = {
233 .pmd_entry = swapin_walk_pmd_entry,
237 walk_page_range(start, end, &walk);
239 lru_add_drain(); /* Push any new pages onto the LRU now */
242 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
243 unsigned long start, unsigned long end,
244 struct address_space *mapping)
250 for (; start < end; start += PAGE_SIZE) {
251 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
253 page = find_get_entry(mapping, index);
254 if (!xa_is_value(page)) {
259 swap = radix_to_swp_entry(page);
260 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
266 lru_add_drain(); /* Push any new pages onto the LRU now */
268 #endif /* CONFIG_SWAP */
271 * Schedule all required I/O operations. Do not wait for completion.
273 static long madvise_willneed(struct vm_area_struct *vma,
274 struct vm_area_struct **prev,
275 unsigned long start, unsigned long end)
277 struct file *file = vma->vm_file;
282 force_swapin_readahead(vma, start, end);
286 if (shmem_mapping(file->f_mapping)) {
287 force_shm_swapin_readahead(vma, start, end,
296 if (IS_DAX(file_inode(file))) {
297 /* no bad return value, but ignore advice */
301 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
302 if (end > vma->vm_end)
304 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
306 force_page_cache_readahead(file->f_mapping, file, start, end - start);
310 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
311 unsigned long end, struct mm_walk *walk)
314 struct mmu_gather *tlb = walk->private;
315 struct mm_struct *mm = tlb->mm;
316 struct vm_area_struct *vma = walk->vma;
318 pte_t *orig_pte, *pte, ptent;
323 next = pmd_addr_end(addr, end);
324 if (pmd_trans_huge(*pmd))
325 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
328 if (pmd_trans_unstable(pmd))
331 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
332 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
333 flush_tlb_batched_pending(mm);
334 arch_enter_lazy_mmu_mode();
335 for (; addr != end; pte++, addr += PAGE_SIZE) {
341 * If the pte has swp_entry, just clear page table to
342 * prevent swap-in which is more expensive rather than
343 * (page allocation + zeroing).
345 if (!pte_present(ptent)) {
348 entry = pte_to_swp_entry(ptent);
349 if (non_swap_entry(entry))
352 free_swap_and_cache(entry);
353 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
357 page = _vm_normal_page(vma, addr, ptent, true);
362 * If pmd isn't transhuge but the page is THP and
363 * is owned by only this process, split it and
364 * deactivate all pages.
366 if (PageTransCompound(page)) {
367 if (page_mapcount(page) != 1)
370 if (!trylock_page(page)) {
374 pte_unmap_unlock(orig_pte, ptl);
375 if (split_huge_page(page)) {
378 pte_offset_map_lock(mm, pmd, addr, &ptl);
383 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
389 VM_BUG_ON_PAGE(PageTransCompound(page), page);
391 if (PageSwapCache(page) || PageDirty(page)) {
392 if (!trylock_page(page))
395 * If page is shared with others, we couldn't clear
396 * PG_dirty of the page.
398 if (page_mapcount(page) != 1) {
403 if (PageSwapCache(page) && !try_to_free_swap(page)) {
408 ClearPageDirty(page);
412 if (pte_young(ptent) || pte_dirty(ptent)) {
414 * Some of architecture(ex, PPC) don't update TLB
415 * with set_pte_at and tlb_remove_tlb_entry so for
416 * the portability, remap the pte with old|clean
417 * after pte clearing.
419 ptent = ptep_get_and_clear_full(mm, addr, pte,
422 ptent = pte_mkold(ptent);
423 ptent = pte_mkclean(ptent);
424 set_pte_at(mm, addr, pte, ptent);
425 tlb_remove_tlb_entry(tlb, pte, addr);
427 mark_page_lazyfree(page);
431 if (current->mm == mm)
434 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
436 arch_leave_lazy_mmu_mode();
437 pte_unmap_unlock(orig_pte, ptl);
443 static void madvise_free_page_range(struct mmu_gather *tlb,
444 struct vm_area_struct *vma,
445 unsigned long addr, unsigned long end)
447 struct mm_walk free_walk = {
448 .pmd_entry = madvise_free_pte_range,
453 tlb_start_vma(tlb, vma);
454 walk_page_range(addr, end, &free_walk);
455 tlb_end_vma(tlb, vma);
458 static int madvise_free_single_vma(struct vm_area_struct *vma,
459 unsigned long start_addr, unsigned long end_addr)
461 struct mm_struct *mm = vma->vm_mm;
462 struct mmu_notifier_range range;
463 struct mmu_gather tlb;
465 /* MADV_FREE works for only anon vma at the moment */
466 if (!vma_is_anonymous(vma))
469 range.start = max(vma->vm_start, start_addr);
470 if (range.start >= vma->vm_end)
472 range.end = min(vma->vm_end, end_addr);
473 if (range.end <= vma->vm_start)
475 mmu_notifier_range_init(&range, mm, range.start, range.end);
478 tlb_gather_mmu(&tlb, mm, range.start, range.end);
479 update_hiwater_rss(mm);
481 mmu_notifier_invalidate_range_start(&range);
482 madvise_free_page_range(&tlb, vma, range.start, range.end);
483 mmu_notifier_invalidate_range_end(&range);
484 tlb_finish_mmu(&tlb, range.start, range.end);
490 * Application no longer needs these pages. If the pages are dirty,
491 * it's OK to just throw them away. The app will be more careful about
492 * data it wants to keep. Be sure to free swap resources too. The
493 * zap_page_range call sets things up for shrink_active_list to actually free
494 * these pages later if no one else has touched them in the meantime,
495 * although we could add these pages to a global reuse list for
496 * shrink_active_list to pick up before reclaiming other pages.
498 * NB: This interface discards data rather than pushes it out to swap,
499 * as some implementations do. This has performance implications for
500 * applications like large transactional databases which want to discard
501 * pages in anonymous maps after committing to backing store the data
502 * that was kept in them. There is no reason to write this data out to
503 * the swap area if the application is discarding it.
505 * An interface that causes the system to free clean pages and flush
506 * dirty pages is already available as msync(MS_INVALIDATE).
508 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
509 unsigned long start, unsigned long end)
511 zap_page_range(vma, start, end - start);
515 static long madvise_dontneed_free(struct vm_area_struct *vma,
516 struct vm_area_struct **prev,
517 unsigned long start, unsigned long end,
521 if (!can_madv_dontneed_vma(vma))
524 if (!userfaultfd_remove(vma, start, end)) {
525 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
527 down_read(¤t->mm->mmap_sem);
528 vma = find_vma(current->mm, start);
531 if (start < vma->vm_start) {
533 * This "vma" under revalidation is the one
534 * with the lowest vma->vm_start where start
535 * is also < vma->vm_end. If start <
536 * vma->vm_start it means an hole materialized
537 * in the user address space within the
538 * virtual range passed to MADV_DONTNEED
543 if (!can_madv_dontneed_vma(vma))
545 if (end > vma->vm_end) {
547 * Don't fail if end > vma->vm_end. If the old
548 * vma was splitted while the mmap_sem was
549 * released the effect of the concurrent
550 * operation may not cause madvise() to
551 * have an undefined result. There may be an
552 * adjacent next vma that we'll walk
553 * next. userfaultfd_remove() will generate an
554 * UFFD_EVENT_REMOVE repetition on the
555 * end-vma->vm_end range, but the manager can
556 * handle a repetition fine.
560 VM_WARN_ON(start >= end);
563 if (behavior == MADV_DONTNEED)
564 return madvise_dontneed_single_vma(vma, start, end);
565 else if (behavior == MADV_FREE)
566 return madvise_free_single_vma(vma, start, end);
572 * Application wants to free up the pages and associated backing store.
573 * This is effectively punching a hole into the middle of a file.
575 static long madvise_remove(struct vm_area_struct *vma,
576 struct vm_area_struct **prev,
577 unsigned long start, unsigned long end)
583 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
585 if (vma->vm_flags & VM_LOCKED)
590 if (!f || !f->f_mapping || !f->f_mapping->host) {
594 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
597 offset = (loff_t)(start - vma->vm_start)
598 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
601 * Filesystem's fallocate may need to take i_mutex. We need to
602 * explicitly grab a reference because the vma (and hence the
603 * vma's reference to the file) can go away as soon as we drop
607 if (userfaultfd_remove(vma, start, end)) {
608 /* mmap_sem was not released by userfaultfd_remove() */
609 up_read(¤t->mm->mmap_sem);
611 error = vfs_fallocate(f,
612 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
613 offset, end - start);
615 down_read(¤t->mm->mmap_sem);
619 #ifdef CONFIG_MEMORY_FAILURE
621 * Error injection support for memory error handling.
623 static int madvise_inject_error(int behavior,
624 unsigned long start, unsigned long end)
630 if (!capable(CAP_SYS_ADMIN))
634 for (; start < end; start += PAGE_SIZE << order) {
638 ret = get_user_pages_fast(start, 1, 0, &page);
641 pfn = page_to_pfn(page);
644 * When soft offlining hugepages, after migrating the page
645 * we dissolve it, therefore in the second loop "page" will
646 * no longer be a compound page, and order will be 0.
648 order = compound_order(compound_head(page));
650 if (PageHWPoison(page)) {
655 if (behavior == MADV_SOFT_OFFLINE) {
656 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
659 ret = soft_offline_page(page, MF_COUNT_INCREASED);
665 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
669 * Drop the page reference taken by get_user_pages_fast(). In
670 * the absence of MF_COUNT_INCREASED the memory_failure()
671 * routine is responsible for pinning the page to prevent it
672 * from being released back to the page allocator.
675 ret = memory_failure(pfn, 0);
680 /* Ensure that all poisoned pages are removed from per-cpu lists */
681 for_each_populated_zone(zone)
682 drain_all_pages(zone);
689 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
690 unsigned long start, unsigned long end, int behavior)
694 return madvise_remove(vma, prev, start, end);
696 return madvise_willneed(vma, prev, start, end);
699 return madvise_dontneed_free(vma, prev, start, end, behavior);
701 return madvise_behavior(vma, prev, start, end, behavior);
706 madvise_behavior_valid(int behavior)
712 case MADV_SEQUENTIAL:
720 case MADV_UNMERGEABLE:
722 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
724 case MADV_NOHUGEPAGE:
728 case MADV_WIPEONFORK:
729 case MADV_KEEPONFORK:
730 #ifdef CONFIG_MEMORY_FAILURE
731 case MADV_SOFT_OFFLINE:
742 * The madvise(2) system call.
744 * Applications can use madvise() to advise the kernel how it should
745 * handle paging I/O in this VM area. The idea is to help the kernel
746 * use appropriate read-ahead and caching techniques. The information
747 * provided is advisory only, and can be safely disregarded by the
748 * kernel without affecting the correct operation of the application.
751 * MADV_NORMAL - the default behavior is to read clusters. This
752 * results in some read-ahead and read-behind.
753 * MADV_RANDOM - the system should read the minimum amount of data
754 * on any access, since it is unlikely that the appli-
755 * cation will need more than what it asks for.
756 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
757 * once, so they can be aggressively read ahead, and
758 * can be freed soon after they are accessed.
759 * MADV_WILLNEED - the application is notifying the system to read
761 * MADV_DONTNEED - the application is finished with the given range,
762 * so the kernel can free resources associated with it.
763 * MADV_FREE - the application marks pages in the given range as lazy free,
764 * where actual purges are postponed until memory pressure happens.
765 * MADV_REMOVE - the application wants to free up the given range of
766 * pages and associated backing store.
767 * MADV_DONTFORK - omit this area from child's address space when forking:
768 * typically, to avoid COWing pages pinned by get_user_pages().
769 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
770 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
771 * range after a fork.
772 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
773 * MADV_HWPOISON - trigger memory error handler as if the given memory range
774 * were corrupted by unrecoverable hardware memory failure.
775 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
776 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
777 * this area with pages of identical content from other such areas.
778 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
779 * MADV_HUGEPAGE - the application wants to back the given range by transparent
780 * huge pages in the future. Existing pages might be coalesced and
781 * new pages might be allocated as THP.
782 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
783 * transparent huge pages so the existing pages will not be
784 * coalesced into THP and new pages will not be allocated as THP.
785 * MADV_DONTDUMP - the application wants to prevent pages in the given range
786 * from being included in its core dump.
787 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
791 * -EINVAL - start + len < 0, start is not page-aligned,
792 * "behavior" is not a valid value, or application
793 * is attempting to release locked or shared pages,
794 * or the specified address range includes file, Huge TLB,
795 * MAP_SHARED or VMPFNMAP range.
796 * -ENOMEM - addresses in the specified range are not currently
797 * mapped, or are outside the AS of the process.
798 * -EIO - an I/O error occurred while paging in data.
799 * -EBADF - map exists, but area maps something that isn't a file.
800 * -EAGAIN - a kernel resource was temporarily unavailable.
802 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
804 unsigned long end, tmp;
805 struct vm_area_struct *vma, *prev;
806 int unmapped_error = 0;
810 struct blk_plug plug;
812 if (!madvise_behavior_valid(behavior))
815 if (start & ~PAGE_MASK)
817 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
819 /* Check to see whether len was rounded up from small -ve to zero */
831 #ifdef CONFIG_MEMORY_FAILURE
832 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
833 return madvise_inject_error(behavior, start, start + len_in);
836 write = madvise_need_mmap_write(behavior);
838 if (down_write_killable(¤t->mm->mmap_sem))
841 down_read(¤t->mm->mmap_sem);
845 * If the interval [start,end) covers some unmapped address
846 * ranges, just ignore them, but return -ENOMEM at the end.
847 * - different from the way of handling in mlock etc.
849 vma = find_vma_prev(current->mm, start, &prev);
850 if (vma && start > vma->vm_start)
853 blk_start_plug(&plug);
855 /* Still start < end. */
860 /* Here start < (end|vma->vm_end). */
861 if (start < vma->vm_start) {
862 unmapped_error = -ENOMEM;
863 start = vma->vm_start;
868 /* Here vma->vm_start <= start < (end|vma->vm_end) */
873 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
874 error = madvise_vma(vma, &prev, start, tmp, behavior);
878 if (prev && start < prev->vm_end)
879 start = prev->vm_end;
880 error = unmapped_error;
885 else /* madvise_remove dropped mmap_sem */
886 vma = find_vma(current->mm, start);
889 blk_finish_plug(&plug);
891 up_write(¤t->mm->mmap_sem);
893 up_read(¤t->mm->mmap_sem);