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);
111 goto out_convert_errno;
114 case MADV_NOHUGEPAGE:
115 error = hugepage_madvise(vma, &new_flags, behavior);
117 goto out_convert_errno;
121 if (new_flags == vma->vm_flags) {
126 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
127 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
128 vma->vm_file, pgoff, vma_policy(vma),
129 vma->vm_userfaultfd_ctx);
137 if (start != vma->vm_start) {
138 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
142 error = __split_vma(mm, vma, start, 1);
144 goto out_convert_errno;
147 if (end != vma->vm_end) {
148 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
152 error = __split_vma(mm, vma, end, 0);
154 goto out_convert_errno;
159 * vm_flags is protected by the mmap_sem held in write mode.
161 vma->vm_flags = new_flags;
165 * madvise() returns EAGAIN if kernel resources, such as
166 * slab, are temporarily unavailable.
168 if (error == -ENOMEM)
175 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
176 unsigned long end, struct mm_walk *walk)
179 struct vm_area_struct *vma = walk->private;
182 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
185 for (index = start; index != end; index += PAGE_SIZE) {
191 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
192 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
193 pte_unmap_unlock(orig_pte, ptl);
195 if (pte_present(pte) || pte_none(pte))
197 entry = pte_to_swp_entry(pte);
198 if (unlikely(non_swap_entry(entry)))
201 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
210 static const struct mm_walk_ops swapin_walk_ops = {
211 .pmd_entry = swapin_walk_pmd_entry,
214 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
215 unsigned long start, unsigned long end,
216 struct address_space *mapping)
222 for (; start < end; start += PAGE_SIZE) {
223 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
225 page = find_get_entry(mapping, index);
226 if (!xa_is_value(page)) {
231 swap = radix_to_swp_entry(page);
232 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
238 lru_add_drain(); /* Push any new pages onto the LRU now */
240 #endif /* CONFIG_SWAP */
243 * Schedule all required I/O operations. Do not wait for completion.
245 static long madvise_willneed(struct vm_area_struct *vma,
246 struct vm_area_struct **prev,
247 unsigned long start, unsigned long end)
249 struct file *file = vma->vm_file;
255 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
256 lru_add_drain(); /* Push any new pages onto the LRU now */
260 if (shmem_mapping(file->f_mapping)) {
261 force_shm_swapin_readahead(vma, start, end,
270 if (IS_DAX(file_inode(file))) {
271 /* no bad return value, but ignore advice */
276 * Filesystem's fadvise may need to take various locks. We need to
277 * explicitly grab a reference because the vma (and hence the
278 * vma's reference to the file) can go away as soon as we drop
281 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
283 up_read(¤t->mm->mmap_sem);
284 offset = (loff_t)(start - vma->vm_start)
285 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
286 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
288 down_read(¤t->mm->mmap_sem);
292 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
293 unsigned long end, struct mm_walk *walk)
296 struct mmu_gather *tlb = walk->private;
297 struct mm_struct *mm = tlb->mm;
298 struct vm_area_struct *vma = walk->vma;
300 pte_t *orig_pte, *pte, ptent;
305 next = pmd_addr_end(addr, end);
306 if (pmd_trans_huge(*pmd))
307 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
310 if (pmd_trans_unstable(pmd))
313 tlb_change_page_size(tlb, PAGE_SIZE);
314 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
315 flush_tlb_batched_pending(mm);
316 arch_enter_lazy_mmu_mode();
317 for (; addr != end; pte++, addr += PAGE_SIZE) {
323 * If the pte has swp_entry, just clear page table to
324 * prevent swap-in which is more expensive rather than
325 * (page allocation + zeroing).
327 if (!pte_present(ptent)) {
330 entry = pte_to_swp_entry(ptent);
331 if (non_swap_entry(entry))
334 free_swap_and_cache(entry);
335 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
339 page = vm_normal_page(vma, addr, ptent);
344 * If pmd isn't transhuge but the page is THP and
345 * is owned by only this process, split it and
346 * deactivate all pages.
348 if (PageTransCompound(page)) {
349 if (page_mapcount(page) != 1)
352 if (!trylock_page(page)) {
356 pte_unmap_unlock(orig_pte, ptl);
357 if (split_huge_page(page)) {
360 pte_offset_map_lock(mm, pmd, addr, &ptl);
365 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
371 VM_BUG_ON_PAGE(PageTransCompound(page), page);
373 if (PageSwapCache(page) || PageDirty(page)) {
374 if (!trylock_page(page))
377 * If page is shared with others, we couldn't clear
378 * PG_dirty of the page.
380 if (page_mapcount(page) != 1) {
385 if (PageSwapCache(page) && !try_to_free_swap(page)) {
390 ClearPageDirty(page);
394 if (pte_young(ptent) || pte_dirty(ptent)) {
396 * Some of architecture(ex, PPC) don't update TLB
397 * with set_pte_at and tlb_remove_tlb_entry so for
398 * the portability, remap the pte with old|clean
399 * after pte clearing.
401 ptent = ptep_get_and_clear_full(mm, addr, pte,
404 ptent = pte_mkold(ptent);
405 ptent = pte_mkclean(ptent);
406 set_pte_at(mm, addr, pte, ptent);
407 tlb_remove_tlb_entry(tlb, pte, addr);
409 mark_page_lazyfree(page);
413 if (current->mm == mm)
416 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
418 arch_leave_lazy_mmu_mode();
419 pte_unmap_unlock(orig_pte, ptl);
425 static const struct mm_walk_ops madvise_free_walk_ops = {
426 .pmd_entry = madvise_free_pte_range,
429 static int madvise_free_single_vma(struct vm_area_struct *vma,
430 unsigned long start_addr, unsigned long end_addr)
432 struct mm_struct *mm = vma->vm_mm;
433 struct mmu_notifier_range range;
434 struct mmu_gather tlb;
436 /* MADV_FREE works for only anon vma at the moment */
437 if (!vma_is_anonymous(vma))
440 range.start = max(vma->vm_start, start_addr);
441 if (range.start >= vma->vm_end)
443 range.end = min(vma->vm_end, end_addr);
444 if (range.end <= vma->vm_start)
446 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
447 range.start, range.end);
450 tlb_gather_mmu(&tlb, mm, range.start, range.end);
451 update_hiwater_rss(mm);
453 mmu_notifier_invalidate_range_start(&range);
454 tlb_start_vma(&tlb, vma);
455 walk_page_range(vma->vm_mm, range.start, range.end,
456 &madvise_free_walk_ops, &tlb);
457 tlb_end_vma(&tlb, vma);
458 mmu_notifier_invalidate_range_end(&range);
459 tlb_finish_mmu(&tlb, range.start, range.end);
465 * Application no longer needs these pages. If the pages are dirty,
466 * it's OK to just throw them away. The app will be more careful about
467 * data it wants to keep. Be sure to free swap resources too. The
468 * zap_page_range call sets things up for shrink_active_list to actually free
469 * these pages later if no one else has touched them in the meantime,
470 * although we could add these pages to a global reuse list for
471 * shrink_active_list to pick up before reclaiming other pages.
473 * NB: This interface discards data rather than pushes it out to swap,
474 * as some implementations do. This has performance implications for
475 * applications like large transactional databases which want to discard
476 * pages in anonymous maps after committing to backing store the data
477 * that was kept in them. There is no reason to write this data out to
478 * the swap area if the application is discarding it.
480 * An interface that causes the system to free clean pages and flush
481 * dirty pages is already available as msync(MS_INVALIDATE).
483 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
484 unsigned long start, unsigned long end)
486 zap_page_range(vma, start, end - start);
490 static long madvise_dontneed_free(struct vm_area_struct *vma,
491 struct vm_area_struct **prev,
492 unsigned long start, unsigned long end,
496 if (!can_madv_dontneed_vma(vma))
499 if (!userfaultfd_remove(vma, start, end)) {
500 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
502 down_read(¤t->mm->mmap_sem);
503 vma = find_vma(current->mm, start);
506 if (start < vma->vm_start) {
508 * This "vma" under revalidation is the one
509 * with the lowest vma->vm_start where start
510 * is also < vma->vm_end. If start <
511 * vma->vm_start it means an hole materialized
512 * in the user address space within the
513 * virtual range passed to MADV_DONTNEED
518 if (!can_madv_dontneed_vma(vma))
520 if (end > vma->vm_end) {
522 * Don't fail if end > vma->vm_end. If the old
523 * vma was splitted while the mmap_sem was
524 * released the effect of the concurrent
525 * operation may not cause madvise() to
526 * have an undefined result. There may be an
527 * adjacent next vma that we'll walk
528 * next. userfaultfd_remove() will generate an
529 * UFFD_EVENT_REMOVE repetition on the
530 * end-vma->vm_end range, but the manager can
531 * handle a repetition fine.
535 VM_WARN_ON(start >= end);
538 if (behavior == MADV_DONTNEED)
539 return madvise_dontneed_single_vma(vma, start, end);
540 else if (behavior == MADV_FREE)
541 return madvise_free_single_vma(vma, start, end);
547 * Application wants to free up the pages and associated backing store.
548 * This is effectively punching a hole into the middle of a file.
550 static long madvise_remove(struct vm_area_struct *vma,
551 struct vm_area_struct **prev,
552 unsigned long start, unsigned long end)
558 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
560 if (vma->vm_flags & VM_LOCKED)
565 if (!f || !f->f_mapping || !f->f_mapping->host) {
569 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
572 offset = (loff_t)(start - vma->vm_start)
573 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
576 * Filesystem's fallocate may need to take i_mutex. We need to
577 * explicitly grab a reference because the vma (and hence the
578 * vma's reference to the file) can go away as soon as we drop
582 if (userfaultfd_remove(vma, start, end)) {
583 /* mmap_sem was not released by userfaultfd_remove() */
584 up_read(¤t->mm->mmap_sem);
586 error = vfs_fallocate(f,
587 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
588 offset, end - start);
590 down_read(¤t->mm->mmap_sem);
594 #ifdef CONFIG_MEMORY_FAILURE
596 * Error injection support for memory error handling.
598 static int madvise_inject_error(int behavior,
599 unsigned long start, unsigned long end)
605 if (!capable(CAP_SYS_ADMIN))
609 for (; start < end; start += PAGE_SIZE << order) {
613 ret = get_user_pages_fast(start, 1, 0, &page);
616 pfn = page_to_pfn(page);
619 * When soft offlining hugepages, after migrating the page
620 * we dissolve it, therefore in the second loop "page" will
621 * no longer be a compound page, and order will be 0.
623 order = compound_order(compound_head(page));
625 if (PageHWPoison(page)) {
630 if (behavior == MADV_SOFT_OFFLINE) {
631 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
634 ret = soft_offline_page(page, MF_COUNT_INCREASED);
640 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
644 * Drop the page reference taken by get_user_pages_fast(). In
645 * the absence of MF_COUNT_INCREASED the memory_failure()
646 * routine is responsible for pinning the page to prevent it
647 * from being released back to the page allocator.
650 ret = memory_failure(pfn, 0);
655 /* Ensure that all poisoned pages are removed from per-cpu lists */
656 for_each_populated_zone(zone)
657 drain_all_pages(zone);
664 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
665 unsigned long start, unsigned long end, int behavior)
669 return madvise_remove(vma, prev, start, end);
671 return madvise_willneed(vma, prev, start, end);
674 return madvise_dontneed_free(vma, prev, start, end, behavior);
676 return madvise_behavior(vma, prev, start, end, behavior);
681 madvise_behavior_valid(int behavior)
687 case MADV_SEQUENTIAL:
695 case MADV_UNMERGEABLE:
697 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
699 case MADV_NOHUGEPAGE:
703 case MADV_WIPEONFORK:
704 case MADV_KEEPONFORK:
705 #ifdef CONFIG_MEMORY_FAILURE
706 case MADV_SOFT_OFFLINE:
717 * The madvise(2) system call.
719 * Applications can use madvise() to advise the kernel how it should
720 * handle paging I/O in this VM area. The idea is to help the kernel
721 * use appropriate read-ahead and caching techniques. The information
722 * provided is advisory only, and can be safely disregarded by the
723 * kernel without affecting the correct operation of the application.
726 * MADV_NORMAL - the default behavior is to read clusters. This
727 * results in some read-ahead and read-behind.
728 * MADV_RANDOM - the system should read the minimum amount of data
729 * on any access, since it is unlikely that the appli-
730 * cation will need more than what it asks for.
731 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
732 * once, so they can be aggressively read ahead, and
733 * can be freed soon after they are accessed.
734 * MADV_WILLNEED - the application is notifying the system to read
736 * MADV_DONTNEED - the application is finished with the given range,
737 * so the kernel can free resources associated with it.
738 * MADV_FREE - the application marks pages in the given range as lazy free,
739 * where actual purges are postponed until memory pressure happens.
740 * MADV_REMOVE - the application wants to free up the given range of
741 * pages and associated backing store.
742 * MADV_DONTFORK - omit this area from child's address space when forking:
743 * typically, to avoid COWing pages pinned by get_user_pages().
744 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
745 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
746 * range after a fork.
747 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
748 * MADV_HWPOISON - trigger memory error handler as if the given memory range
749 * were corrupted by unrecoverable hardware memory failure.
750 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
751 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
752 * this area with pages of identical content from other such areas.
753 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
754 * MADV_HUGEPAGE - the application wants to back the given range by transparent
755 * huge pages in the future. Existing pages might be coalesced and
756 * new pages might be allocated as THP.
757 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
758 * transparent huge pages so the existing pages will not be
759 * coalesced into THP and new pages will not be allocated as THP.
760 * MADV_DONTDUMP - the application wants to prevent pages in the given range
761 * from being included in its core dump.
762 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
766 * -EINVAL - start + len < 0, start is not page-aligned,
767 * "behavior" is not a valid value, or application
768 * is attempting to release locked or shared pages,
769 * or the specified address range includes file, Huge TLB,
770 * MAP_SHARED or VMPFNMAP range.
771 * -ENOMEM - addresses in the specified range are not currently
772 * mapped, or are outside the AS of the process.
773 * -EIO - an I/O error occurred while paging in data.
774 * -EBADF - map exists, but area maps something that isn't a file.
775 * -EAGAIN - a kernel resource was temporarily unavailable.
777 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
779 unsigned long end, tmp;
780 struct vm_area_struct *vma, *prev;
781 int unmapped_error = 0;
785 struct blk_plug plug;
787 if (!madvise_behavior_valid(behavior))
790 if (start & ~PAGE_MASK)
792 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
794 /* Check to see whether len was rounded up from small -ve to zero */
806 #ifdef CONFIG_MEMORY_FAILURE
807 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
808 return madvise_inject_error(behavior, start, start + len_in);
811 write = madvise_need_mmap_write(behavior);
813 if (down_write_killable(¤t->mm->mmap_sem))
816 down_read(¤t->mm->mmap_sem);
820 * If the interval [start,end) covers some unmapped address
821 * ranges, just ignore them, but return -ENOMEM at the end.
822 * - different from the way of handling in mlock etc.
824 vma = find_vma_prev(current->mm, start, &prev);
825 if (vma && start > vma->vm_start)
828 blk_start_plug(&plug);
830 /* Still start < end. */
835 /* Here start < (end|vma->vm_end). */
836 if (start < vma->vm_start) {
837 unmapped_error = -ENOMEM;
838 start = vma->vm_start;
843 /* Here vma->vm_start <= start < (end|vma->vm_end) */
848 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
849 error = madvise_vma(vma, &prev, start, tmp, behavior);
853 if (prev && start < prev->vm_end)
854 start = prev->vm_end;
855 error = unmapped_error;
860 else /* madvise_remove dropped mmap_sem */
861 vma = find_vma(current->mm, start);
864 blk_finish_plug(&plug);
866 up_write(¤t->mm->mmap_sem);
868 up_read(¤t->mm->mmap_sem);