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/pagewalk.h>
24 #include <linux/swap.h>
25 #include <linux/swapops.h>
26 #include <linux/shmem_fs.h>
27 #include <linux/mmu_notifier.h>
34 * Any behaviour which results in changes to the vma->vm_flags needs to
35 * take mmap_sem for writing. Others, which simply traverse vmas, need
36 * to only take it for reading.
38 static int madvise_need_mmap_write(int behavior)
47 /* be safe, default to 1. list exceptions explicitly */
53 * We can potentially split a vm area into separate
54 * areas, each area with its own behavior.
56 static long madvise_behavior(struct vm_area_struct *vma,
57 struct vm_area_struct **prev,
58 unsigned long start, unsigned long end, int behavior)
60 struct mm_struct *mm = vma->vm_mm;
63 unsigned long new_flags = vma->vm_flags;
67 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
70 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
73 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
76 new_flags |= VM_DONTCOPY;
79 if (vma->vm_flags & VM_IO) {
83 new_flags &= ~VM_DONTCOPY;
86 /* MADV_WIPEONFORK is only supported on anonymous memory. */
87 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
91 new_flags |= VM_WIPEONFORK;
94 new_flags &= ~VM_WIPEONFORK;
97 new_flags |= VM_DONTDUMP;
100 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
104 new_flags &= ~VM_DONTDUMP;
107 case MADV_UNMERGEABLE:
108 error = ksm_madvise(vma, start, end, behavior, &new_flags);
111 * madvise() returns EAGAIN if kernel resources, such as
112 * slab, are temporarily unavailable.
114 if (error == -ENOMEM)
120 case MADV_NOHUGEPAGE:
121 error = hugepage_madvise(vma, &new_flags, behavior);
124 * madvise() returns EAGAIN if kernel resources, such as
125 * slab, are temporarily unavailable.
127 if (error == -ENOMEM)
134 if (new_flags == vma->vm_flags) {
139 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
140 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
141 vma->vm_file, pgoff, vma_policy(vma),
142 vma->vm_userfaultfd_ctx);
150 if (start != vma->vm_start) {
151 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
155 error = __split_vma(mm, vma, start, 1);
158 * madvise() returns EAGAIN if kernel resources, such as
159 * slab, are temporarily unavailable.
161 if (error == -ENOMEM)
167 if (end != vma->vm_end) {
168 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
172 error = __split_vma(mm, vma, end, 0);
175 * madvise() returns EAGAIN if kernel resources, such as
176 * slab, are temporarily unavailable.
178 if (error == -ENOMEM)
186 * vm_flags is protected by the mmap_sem held in write mode.
188 vma->vm_flags = new_flags;
194 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
195 unsigned long end, struct mm_walk *walk)
198 struct vm_area_struct *vma = walk->private;
201 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
204 for (index = start; index != end; index += PAGE_SIZE) {
210 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
211 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
212 pte_unmap_unlock(orig_pte, ptl);
214 if (pte_present(pte) || pte_none(pte))
216 entry = pte_to_swp_entry(pte);
217 if (unlikely(non_swap_entry(entry)))
220 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
229 static void force_swapin_readahead(struct vm_area_struct *vma,
230 unsigned long start, unsigned long end)
232 struct mm_walk walk = {
234 .pmd_entry = swapin_walk_pmd_entry,
238 walk_page_range(start, end, &walk);
240 lru_add_drain(); /* Push any new pages onto the LRU now */
243 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
244 unsigned long start, unsigned long end,
245 struct address_space *mapping)
251 for (; start < end; start += PAGE_SIZE) {
252 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
254 page = find_get_entry(mapping, index);
255 if (!xa_is_value(page)) {
260 swap = radix_to_swp_entry(page);
261 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
267 lru_add_drain(); /* Push any new pages onto the LRU now */
269 #endif /* CONFIG_SWAP */
272 * Schedule all required I/O operations. Do not wait for completion.
274 static long madvise_willneed(struct vm_area_struct *vma,
275 struct vm_area_struct **prev,
276 unsigned long start, unsigned long end)
278 struct file *file = vma->vm_file;
283 force_swapin_readahead(vma, start, end);
287 if (shmem_mapping(file->f_mapping)) {
288 force_shm_swapin_readahead(vma, start, end,
297 if (IS_DAX(file_inode(file))) {
298 /* no bad return value, but ignore advice */
302 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
303 if (end > vma->vm_end)
305 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
307 force_page_cache_readahead(file->f_mapping, file, start, end - start);
311 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
312 unsigned long end, struct mm_walk *walk)
315 struct mmu_gather *tlb = walk->private;
316 struct mm_struct *mm = tlb->mm;
317 struct vm_area_struct *vma = walk->vma;
319 pte_t *orig_pte, *pte, ptent;
324 next = pmd_addr_end(addr, end);
325 if (pmd_trans_huge(*pmd))
326 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
329 if (pmd_trans_unstable(pmd))
332 tlb_change_page_size(tlb, PAGE_SIZE);
333 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
334 flush_tlb_batched_pending(mm);
335 arch_enter_lazy_mmu_mode();
336 for (; addr != end; pte++, addr += PAGE_SIZE) {
342 * If the pte has swp_entry, just clear page table to
343 * prevent swap-in which is more expensive rather than
344 * (page allocation + zeroing).
346 if (!pte_present(ptent)) {
349 entry = pte_to_swp_entry(ptent);
350 if (non_swap_entry(entry))
353 free_swap_and_cache(entry);
354 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
358 page = vm_normal_page(vma, addr, ptent);
363 * If pmd isn't transhuge but the page is THP and
364 * is owned by only this process, split it and
365 * deactivate all pages.
367 if (PageTransCompound(page)) {
368 if (page_mapcount(page) != 1)
371 if (!trylock_page(page)) {
375 pte_unmap_unlock(orig_pte, ptl);
376 if (split_huge_page(page)) {
379 pte_offset_map_lock(mm, pmd, addr, &ptl);
384 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
390 VM_BUG_ON_PAGE(PageTransCompound(page), page);
392 if (PageSwapCache(page) || PageDirty(page)) {
393 if (!trylock_page(page))
396 * If page is shared with others, we couldn't clear
397 * PG_dirty of the page.
399 if (page_mapcount(page) != 1) {
404 if (PageSwapCache(page) && !try_to_free_swap(page)) {
409 ClearPageDirty(page);
413 if (pte_young(ptent) || pte_dirty(ptent)) {
415 * Some of architecture(ex, PPC) don't update TLB
416 * with set_pte_at and tlb_remove_tlb_entry so for
417 * the portability, remap the pte with old|clean
418 * after pte clearing.
420 ptent = ptep_get_and_clear_full(mm, addr, pte,
423 ptent = pte_mkold(ptent);
424 ptent = pte_mkclean(ptent);
425 set_pte_at(mm, addr, pte, ptent);
426 tlb_remove_tlb_entry(tlb, pte, addr);
428 mark_page_lazyfree(page);
432 if (current->mm == mm)
435 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
437 arch_leave_lazy_mmu_mode();
438 pte_unmap_unlock(orig_pte, ptl);
444 static void madvise_free_page_range(struct mmu_gather *tlb,
445 struct vm_area_struct *vma,
446 unsigned long addr, unsigned long end)
448 struct mm_walk free_walk = {
449 .pmd_entry = madvise_free_pte_range,
454 tlb_start_vma(tlb, vma);
455 walk_page_range(addr, end, &free_walk);
456 tlb_end_vma(tlb, vma);
459 static int madvise_free_single_vma(struct vm_area_struct *vma,
460 unsigned long start_addr, unsigned long end_addr)
462 struct mm_struct *mm = vma->vm_mm;
463 struct mmu_notifier_range range;
464 struct mmu_gather tlb;
466 /* MADV_FREE works for only anon vma at the moment */
467 if (!vma_is_anonymous(vma))
470 range.start = max(vma->vm_start, start_addr);
471 if (range.start >= vma->vm_end)
473 range.end = min(vma->vm_end, end_addr);
474 if (range.end <= vma->vm_start)
476 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
477 range.start, range.end);
480 tlb_gather_mmu(&tlb, mm, range.start, range.end);
481 update_hiwater_rss(mm);
483 mmu_notifier_invalidate_range_start(&range);
484 madvise_free_page_range(&tlb, vma, range.start, range.end);
485 mmu_notifier_invalidate_range_end(&range);
486 tlb_finish_mmu(&tlb, range.start, range.end);
492 * Application no longer needs these pages. If the pages are dirty,
493 * it's OK to just throw them away. The app will be more careful about
494 * data it wants to keep. Be sure to free swap resources too. The
495 * zap_page_range call sets things up for shrink_active_list to actually free
496 * these pages later if no one else has touched them in the meantime,
497 * although we could add these pages to a global reuse list for
498 * shrink_active_list to pick up before reclaiming other pages.
500 * NB: This interface discards data rather than pushes it out to swap,
501 * as some implementations do. This has performance implications for
502 * applications like large transactional databases which want to discard
503 * pages in anonymous maps after committing to backing store the data
504 * that was kept in them. There is no reason to write this data out to
505 * the swap area if the application is discarding it.
507 * An interface that causes the system to free clean pages and flush
508 * dirty pages is already available as msync(MS_INVALIDATE).
510 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
511 unsigned long start, unsigned long end)
513 zap_page_range(vma, start, end - start);
517 static long madvise_dontneed_free(struct vm_area_struct *vma,
518 struct vm_area_struct **prev,
519 unsigned long start, unsigned long end,
523 if (!can_madv_dontneed_vma(vma))
526 if (!userfaultfd_remove(vma, start, end)) {
527 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
529 down_read(¤t->mm->mmap_sem);
530 vma = find_vma(current->mm, start);
533 if (start < vma->vm_start) {
535 * This "vma" under revalidation is the one
536 * with the lowest vma->vm_start where start
537 * is also < vma->vm_end. If start <
538 * vma->vm_start it means an hole materialized
539 * in the user address space within the
540 * virtual range passed to MADV_DONTNEED
545 if (!can_madv_dontneed_vma(vma))
547 if (end > vma->vm_end) {
549 * Don't fail if end > vma->vm_end. If the old
550 * vma was splitted while the mmap_sem was
551 * released the effect of the concurrent
552 * operation may not cause madvise() to
553 * have an undefined result. There may be an
554 * adjacent next vma that we'll walk
555 * next. userfaultfd_remove() will generate an
556 * UFFD_EVENT_REMOVE repetition on the
557 * end-vma->vm_end range, but the manager can
558 * handle a repetition fine.
562 VM_WARN_ON(start >= end);
565 if (behavior == MADV_DONTNEED)
566 return madvise_dontneed_single_vma(vma, start, end);
567 else if (behavior == MADV_FREE)
568 return madvise_free_single_vma(vma, start, end);
574 * Application wants to free up the pages and associated backing store.
575 * This is effectively punching a hole into the middle of a file.
577 static long madvise_remove(struct vm_area_struct *vma,
578 struct vm_area_struct **prev,
579 unsigned long start, unsigned long end)
585 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
587 if (vma->vm_flags & VM_LOCKED)
592 if (!f || !f->f_mapping || !f->f_mapping->host) {
596 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
599 offset = (loff_t)(start - vma->vm_start)
600 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
603 * Filesystem's fallocate may need to take i_mutex. We need to
604 * explicitly grab a reference because the vma (and hence the
605 * vma's reference to the file) can go away as soon as we drop
609 if (userfaultfd_remove(vma, start, end)) {
610 /* mmap_sem was not released by userfaultfd_remove() */
611 up_read(¤t->mm->mmap_sem);
613 error = vfs_fallocate(f,
614 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
615 offset, end - start);
617 down_read(¤t->mm->mmap_sem);
621 #ifdef CONFIG_MEMORY_FAILURE
623 * Error injection support for memory error handling.
625 static int madvise_inject_error(int behavior,
626 unsigned long start, unsigned long end)
632 if (!capable(CAP_SYS_ADMIN))
636 for (; start < end; start += PAGE_SIZE << order) {
640 ret = get_user_pages_fast(start, 1, 0, &page);
643 pfn = page_to_pfn(page);
646 * When soft offlining hugepages, after migrating the page
647 * we dissolve it, therefore in the second loop "page" will
648 * no longer be a compound page, and order will be 0.
650 order = compound_order(compound_head(page));
652 if (PageHWPoison(page)) {
657 if (behavior == MADV_SOFT_OFFLINE) {
658 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
661 ret = soft_offline_page(page, MF_COUNT_INCREASED);
667 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
671 * Drop the page reference taken by get_user_pages_fast(). In
672 * the absence of MF_COUNT_INCREASED the memory_failure()
673 * routine is responsible for pinning the page to prevent it
674 * from being released back to the page allocator.
677 ret = memory_failure(pfn, 0);
682 /* Ensure that all poisoned pages are removed from per-cpu lists */
683 for_each_populated_zone(zone)
684 drain_all_pages(zone);
691 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
692 unsigned long start, unsigned long end, int behavior)
696 return madvise_remove(vma, prev, start, end);
698 return madvise_willneed(vma, prev, start, end);
701 return madvise_dontneed_free(vma, prev, start, end, behavior);
703 return madvise_behavior(vma, prev, start, end, behavior);
708 madvise_behavior_valid(int behavior)
714 case MADV_SEQUENTIAL:
722 case MADV_UNMERGEABLE:
724 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
726 case MADV_NOHUGEPAGE:
730 case MADV_WIPEONFORK:
731 case MADV_KEEPONFORK:
732 #ifdef CONFIG_MEMORY_FAILURE
733 case MADV_SOFT_OFFLINE:
744 * The madvise(2) system call.
746 * Applications can use madvise() to advise the kernel how it should
747 * handle paging I/O in this VM area. The idea is to help the kernel
748 * use appropriate read-ahead and caching techniques. The information
749 * provided is advisory only, and can be safely disregarded by the
750 * kernel without affecting the correct operation of the application.
753 * MADV_NORMAL - the default behavior is to read clusters. This
754 * results in some read-ahead and read-behind.
755 * MADV_RANDOM - the system should read the minimum amount of data
756 * on any access, since it is unlikely that the appli-
757 * cation will need more than what it asks for.
758 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
759 * once, so they can be aggressively read ahead, and
760 * can be freed soon after they are accessed.
761 * MADV_WILLNEED - the application is notifying the system to read
763 * MADV_DONTNEED - the application is finished with the given range,
764 * so the kernel can free resources associated with it.
765 * MADV_FREE - the application marks pages in the given range as lazy free,
766 * where actual purges are postponed until memory pressure happens.
767 * MADV_REMOVE - the application wants to free up the given range of
768 * pages and associated backing store.
769 * MADV_DONTFORK - omit this area from child's address space when forking:
770 * typically, to avoid COWing pages pinned by get_user_pages().
771 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
772 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
773 * range after a fork.
774 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
775 * MADV_HWPOISON - trigger memory error handler as if the given memory range
776 * were corrupted by unrecoverable hardware memory failure.
777 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
778 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
779 * this area with pages of identical content from other such areas.
780 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
781 * MADV_HUGEPAGE - the application wants to back the given range by transparent
782 * huge pages in the future. Existing pages might be coalesced and
783 * new pages might be allocated as THP.
784 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
785 * transparent huge pages so the existing pages will not be
786 * coalesced into THP and new pages will not be allocated as THP.
787 * MADV_DONTDUMP - the application wants to prevent pages in the given range
788 * from being included in its core dump.
789 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
793 * -EINVAL - start + len < 0, start is not page-aligned,
794 * "behavior" is not a valid value, or application
795 * is attempting to release locked or shared pages,
796 * or the specified address range includes file, Huge TLB,
797 * MAP_SHARED or VMPFNMAP range.
798 * -ENOMEM - addresses in the specified range are not currently
799 * mapped, or are outside the AS of the process.
800 * -EIO - an I/O error occurred while paging in data.
801 * -EBADF - map exists, but area maps something that isn't a file.
802 * -EAGAIN - a kernel resource was temporarily unavailable.
804 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
806 unsigned long end, tmp;
807 struct vm_area_struct *vma, *prev;
808 int unmapped_error = 0;
812 struct blk_plug plug;
814 if (!madvise_behavior_valid(behavior))
817 if (start & ~PAGE_MASK)
819 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
821 /* Check to see whether len was rounded up from small -ve to zero */
833 #ifdef CONFIG_MEMORY_FAILURE
834 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
835 return madvise_inject_error(behavior, start, start + len_in);
838 write = madvise_need_mmap_write(behavior);
840 if (down_write_killable(¤t->mm->mmap_sem))
843 down_read(¤t->mm->mmap_sem);
847 * If the interval [start,end) covers some unmapped address
848 * ranges, just ignore them, but return -ENOMEM at the end.
849 * - different from the way of handling in mlock etc.
851 vma = find_vma_prev(current->mm, start, &prev);
852 if (vma && start > vma->vm_start)
855 blk_start_plug(&plug);
857 /* Still start < end. */
862 /* Here start < (end|vma->vm_end). */
863 if (start < vma->vm_start) {
864 unmapped_error = -ENOMEM;
865 start = vma->vm_start;
870 /* Here vma->vm_start <= start < (end|vma->vm_end) */
875 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
876 error = madvise_vma(vma, &prev, start, tmp, behavior);
880 if (prev && start < prev->vm_end)
881 start = prev->vm_end;
882 error = unmapped_error;
887 else /* madvise_remove dropped mmap_sem */
888 vma = find_vma(current->mm, start);
891 blk_finish_plug(&plug);
893 up_write(¤t->mm->mmap_sem);
895 up_read(¤t->mm->mmap_sem);