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/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;
284 force_swapin_readahead(vma, start, end);
288 if (shmem_mapping(file->f_mapping)) {
289 force_shm_swapin_readahead(vma, start, end,
298 if (IS_DAX(file_inode(file))) {
299 /* no bad return value, but ignore advice */
304 * Filesystem's fadvise may need to take various locks. We need to
305 * explicitly grab a reference because the vma (and hence the
306 * vma's reference to the file) can go away as soon as we drop
309 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
311 up_read(¤t->mm->mmap_sem);
312 offset = (loff_t)(start - vma->vm_start)
313 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
314 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
316 down_read(¤t->mm->mmap_sem);
320 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
321 unsigned long end, struct mm_walk *walk)
324 struct mmu_gather *tlb = walk->private;
325 struct mm_struct *mm = tlb->mm;
326 struct vm_area_struct *vma = walk->vma;
328 pte_t *orig_pte, *pte, ptent;
333 next = pmd_addr_end(addr, end);
334 if (pmd_trans_huge(*pmd))
335 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
338 if (pmd_trans_unstable(pmd))
341 tlb_change_page_size(tlb, PAGE_SIZE);
342 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
343 flush_tlb_batched_pending(mm);
344 arch_enter_lazy_mmu_mode();
345 for (; addr != end; pte++, addr += PAGE_SIZE) {
351 * If the pte has swp_entry, just clear page table to
352 * prevent swap-in which is more expensive rather than
353 * (page allocation + zeroing).
355 if (!pte_present(ptent)) {
358 entry = pte_to_swp_entry(ptent);
359 if (non_swap_entry(entry))
362 free_swap_and_cache(entry);
363 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
367 page = vm_normal_page(vma, addr, ptent);
372 * If pmd isn't transhuge but the page is THP and
373 * is owned by only this process, split it and
374 * deactivate all pages.
376 if (PageTransCompound(page)) {
377 if (page_mapcount(page) != 1)
380 if (!trylock_page(page)) {
384 pte_unmap_unlock(orig_pte, ptl);
385 if (split_huge_page(page)) {
388 pte_offset_map_lock(mm, pmd, addr, &ptl);
393 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
399 VM_BUG_ON_PAGE(PageTransCompound(page), page);
401 if (PageSwapCache(page) || PageDirty(page)) {
402 if (!trylock_page(page))
405 * If page is shared with others, we couldn't clear
406 * PG_dirty of the page.
408 if (page_mapcount(page) != 1) {
413 if (PageSwapCache(page) && !try_to_free_swap(page)) {
418 ClearPageDirty(page);
422 if (pte_young(ptent) || pte_dirty(ptent)) {
424 * Some of architecture(ex, PPC) don't update TLB
425 * with set_pte_at and tlb_remove_tlb_entry so for
426 * the portability, remap the pte with old|clean
427 * after pte clearing.
429 ptent = ptep_get_and_clear_full(mm, addr, pte,
432 ptent = pte_mkold(ptent);
433 ptent = pte_mkclean(ptent);
434 set_pte_at(mm, addr, pte, ptent);
435 tlb_remove_tlb_entry(tlb, pte, addr);
437 mark_page_lazyfree(page);
441 if (current->mm == mm)
444 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
446 arch_leave_lazy_mmu_mode();
447 pte_unmap_unlock(orig_pte, ptl);
453 static void madvise_free_page_range(struct mmu_gather *tlb,
454 struct vm_area_struct *vma,
455 unsigned long addr, unsigned long end)
457 struct mm_walk free_walk = {
458 .pmd_entry = madvise_free_pte_range,
463 tlb_start_vma(tlb, vma);
464 walk_page_range(addr, end, &free_walk);
465 tlb_end_vma(tlb, vma);
468 static int madvise_free_single_vma(struct vm_area_struct *vma,
469 unsigned long start_addr, unsigned long end_addr)
471 struct mm_struct *mm = vma->vm_mm;
472 struct mmu_notifier_range range;
473 struct mmu_gather tlb;
475 /* MADV_FREE works for only anon vma at the moment */
476 if (!vma_is_anonymous(vma))
479 range.start = max(vma->vm_start, start_addr);
480 if (range.start >= vma->vm_end)
482 range.end = min(vma->vm_end, end_addr);
483 if (range.end <= vma->vm_start)
485 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
486 range.start, range.end);
489 tlb_gather_mmu(&tlb, mm, range.start, range.end);
490 update_hiwater_rss(mm);
492 mmu_notifier_invalidate_range_start(&range);
493 madvise_free_page_range(&tlb, vma, range.start, range.end);
494 mmu_notifier_invalidate_range_end(&range);
495 tlb_finish_mmu(&tlb, range.start, range.end);
501 * Application no longer needs these pages. If the pages are dirty,
502 * it's OK to just throw them away. The app will be more careful about
503 * data it wants to keep. Be sure to free swap resources too. The
504 * zap_page_range call sets things up for shrink_active_list to actually free
505 * these pages later if no one else has touched them in the meantime,
506 * although we could add these pages to a global reuse list for
507 * shrink_active_list to pick up before reclaiming other pages.
509 * NB: This interface discards data rather than pushes it out to swap,
510 * as some implementations do. This has performance implications for
511 * applications like large transactional databases which want to discard
512 * pages in anonymous maps after committing to backing store the data
513 * that was kept in them. There is no reason to write this data out to
514 * the swap area if the application is discarding it.
516 * An interface that causes the system to free clean pages and flush
517 * dirty pages is already available as msync(MS_INVALIDATE).
519 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
520 unsigned long start, unsigned long end)
522 zap_page_range(vma, start, end - start);
526 static long madvise_dontneed_free(struct vm_area_struct *vma,
527 struct vm_area_struct **prev,
528 unsigned long start, unsigned long end,
532 if (!can_madv_dontneed_vma(vma))
535 if (!userfaultfd_remove(vma, start, end)) {
536 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
538 down_read(¤t->mm->mmap_sem);
539 vma = find_vma(current->mm, start);
542 if (start < vma->vm_start) {
544 * This "vma" under revalidation is the one
545 * with the lowest vma->vm_start where start
546 * is also < vma->vm_end. If start <
547 * vma->vm_start it means an hole materialized
548 * in the user address space within the
549 * virtual range passed to MADV_DONTNEED
554 if (!can_madv_dontneed_vma(vma))
556 if (end > vma->vm_end) {
558 * Don't fail if end > vma->vm_end. If the old
559 * vma was splitted while the mmap_sem was
560 * released the effect of the concurrent
561 * operation may not cause madvise() to
562 * have an undefined result. There may be an
563 * adjacent next vma that we'll walk
564 * next. userfaultfd_remove() will generate an
565 * UFFD_EVENT_REMOVE repetition on the
566 * end-vma->vm_end range, but the manager can
567 * handle a repetition fine.
571 VM_WARN_ON(start >= end);
574 if (behavior == MADV_DONTNEED)
575 return madvise_dontneed_single_vma(vma, start, end);
576 else if (behavior == MADV_FREE)
577 return madvise_free_single_vma(vma, start, end);
583 * Application wants to free up the pages and associated backing store.
584 * This is effectively punching a hole into the middle of a file.
586 static long madvise_remove(struct vm_area_struct *vma,
587 struct vm_area_struct **prev,
588 unsigned long start, unsigned long end)
594 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
596 if (vma->vm_flags & VM_LOCKED)
601 if (!f || !f->f_mapping || !f->f_mapping->host) {
605 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
608 offset = (loff_t)(start - vma->vm_start)
609 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
612 * Filesystem's fallocate may need to take i_mutex. We need to
613 * explicitly grab a reference because the vma (and hence the
614 * vma's reference to the file) can go away as soon as we drop
618 if (userfaultfd_remove(vma, start, end)) {
619 /* mmap_sem was not released by userfaultfd_remove() */
620 up_read(¤t->mm->mmap_sem);
622 error = vfs_fallocate(f,
623 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
624 offset, end - start);
626 down_read(¤t->mm->mmap_sem);
630 #ifdef CONFIG_MEMORY_FAILURE
632 * Error injection support for memory error handling.
634 static int madvise_inject_error(int behavior,
635 unsigned long start, unsigned long end)
641 if (!capable(CAP_SYS_ADMIN))
645 for (; start < end; start += PAGE_SIZE << order) {
649 ret = get_user_pages_fast(start, 1, 0, &page);
652 pfn = page_to_pfn(page);
655 * When soft offlining hugepages, after migrating the page
656 * we dissolve it, therefore in the second loop "page" will
657 * no longer be a compound page, and order will be 0.
659 order = compound_order(compound_head(page));
661 if (PageHWPoison(page)) {
666 if (behavior == MADV_SOFT_OFFLINE) {
667 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
670 ret = soft_offline_page(page, MF_COUNT_INCREASED);
676 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
680 * Drop the page reference taken by get_user_pages_fast(). In
681 * the absence of MF_COUNT_INCREASED the memory_failure()
682 * routine is responsible for pinning the page to prevent it
683 * from being released back to the page allocator.
686 ret = memory_failure(pfn, 0);
691 /* Ensure that all poisoned pages are removed from per-cpu lists */
692 for_each_populated_zone(zone)
693 drain_all_pages(zone);
700 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
701 unsigned long start, unsigned long end, int behavior)
705 return madvise_remove(vma, prev, start, end);
707 return madvise_willneed(vma, prev, start, end);
710 return madvise_dontneed_free(vma, prev, start, end, behavior);
712 return madvise_behavior(vma, prev, start, end, behavior);
717 madvise_behavior_valid(int behavior)
723 case MADV_SEQUENTIAL:
731 case MADV_UNMERGEABLE:
733 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
735 case MADV_NOHUGEPAGE:
739 case MADV_WIPEONFORK:
740 case MADV_KEEPONFORK:
741 #ifdef CONFIG_MEMORY_FAILURE
742 case MADV_SOFT_OFFLINE:
753 * The madvise(2) system call.
755 * Applications can use madvise() to advise the kernel how it should
756 * handle paging I/O in this VM area. The idea is to help the kernel
757 * use appropriate read-ahead and caching techniques. The information
758 * provided is advisory only, and can be safely disregarded by the
759 * kernel without affecting the correct operation of the application.
762 * MADV_NORMAL - the default behavior is to read clusters. This
763 * results in some read-ahead and read-behind.
764 * MADV_RANDOM - the system should read the minimum amount of data
765 * on any access, since it is unlikely that the appli-
766 * cation will need more than what it asks for.
767 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
768 * once, so they can be aggressively read ahead, and
769 * can be freed soon after they are accessed.
770 * MADV_WILLNEED - the application is notifying the system to read
772 * MADV_DONTNEED - the application is finished with the given range,
773 * so the kernel can free resources associated with it.
774 * MADV_FREE - the application marks pages in the given range as lazy free,
775 * where actual purges are postponed until memory pressure happens.
776 * MADV_REMOVE - the application wants to free up the given range of
777 * pages and associated backing store.
778 * MADV_DONTFORK - omit this area from child's address space when forking:
779 * typically, to avoid COWing pages pinned by get_user_pages().
780 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
781 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
782 * range after a fork.
783 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
784 * MADV_HWPOISON - trigger memory error handler as if the given memory range
785 * were corrupted by unrecoverable hardware memory failure.
786 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
787 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
788 * this area with pages of identical content from other such areas.
789 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
790 * MADV_HUGEPAGE - the application wants to back the given range by transparent
791 * huge pages in the future. Existing pages might be coalesced and
792 * new pages might be allocated as THP.
793 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
794 * transparent huge pages so the existing pages will not be
795 * coalesced into THP and new pages will not be allocated as THP.
796 * MADV_DONTDUMP - the application wants to prevent pages in the given range
797 * from being included in its core dump.
798 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
802 * -EINVAL - start + len < 0, start is not page-aligned,
803 * "behavior" is not a valid value, or application
804 * is attempting to release locked or shared pages,
805 * or the specified address range includes file, Huge TLB,
806 * MAP_SHARED or VMPFNMAP range.
807 * -ENOMEM - addresses in the specified range are not currently
808 * mapped, or are outside the AS of the process.
809 * -EIO - an I/O error occurred while paging in data.
810 * -EBADF - map exists, but area maps something that isn't a file.
811 * -EAGAIN - a kernel resource was temporarily unavailable.
813 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
815 unsigned long end, tmp;
816 struct vm_area_struct *vma, *prev;
817 int unmapped_error = 0;
821 struct blk_plug plug;
823 if (!madvise_behavior_valid(behavior))
826 if (start & ~PAGE_MASK)
828 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
830 /* Check to see whether len was rounded up from small -ve to zero */
842 #ifdef CONFIG_MEMORY_FAILURE
843 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
844 return madvise_inject_error(behavior, start, start + len_in);
847 write = madvise_need_mmap_write(behavior);
849 if (down_write_killable(¤t->mm->mmap_sem))
852 down_read(¤t->mm->mmap_sem);
856 * If the interval [start,end) covers some unmapped address
857 * ranges, just ignore them, but return -ENOMEM at the end.
858 * - different from the way of handling in mlock etc.
860 vma = find_vma_prev(current->mm, start, &prev);
861 if (vma && start > vma->vm_start)
864 blk_start_plug(&plug);
866 /* Still start < end. */
871 /* Here start < (end|vma->vm_end). */
872 if (start < vma->vm_start) {
873 unmapped_error = -ENOMEM;
874 start = vma->vm_start;
879 /* Here vma->vm_start <= start < (end|vma->vm_end) */
884 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
885 error = madvise_vma(vma, &prev, start, tmp, behavior);
889 if (prev && start < prev->vm_end)
890 start = prev->vm_end;
891 error = unmapped_error;
896 else /* madvise_remove dropped mmap_sem */
897 vma = find_vma(current->mm, start);
900 blk_finish_plug(&plug);
902 up_write(¤t->mm->mmap_sem);
904 up_read(¤t->mm->mmap_sem);