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/page_idle.h>
15 #include <linux/userfaultfd_k.h>
16 #include <linux/hugetlb.h>
17 #include <linux/falloc.h>
18 #include <linux/fadvise.h>
19 #include <linux/sched.h>
20 #include <linux/ksm.h>
22 #include <linux/file.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/pagewalk.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/mmu_notifier.h>
30 #include <linux/sched/mm.h>
36 struct madvise_walk_private {
37 struct mmu_gather *tlb;
42 * Any behaviour which results in changes to the vma->vm_flags needs to
43 * take mmap_lock for writing. Others, which simply traverse vmas, need
44 * to only take it for reading.
46 static int madvise_need_mmap_write(int behavior)
57 /* be safe, default to 1. list exceptions explicitly */
63 * We can potentially split a vm area into separate
64 * areas, each area with its own behavior.
66 static long madvise_behavior(struct vm_area_struct *vma,
67 struct vm_area_struct **prev,
68 unsigned long start, unsigned long end, int behavior)
70 struct mm_struct *mm = vma->vm_mm;
73 unsigned long new_flags = vma->vm_flags;
77 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
80 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
83 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
86 new_flags |= VM_DONTCOPY;
89 if (vma->vm_flags & VM_IO) {
93 new_flags &= ~VM_DONTCOPY;
96 /* MADV_WIPEONFORK is only supported on anonymous memory. */
97 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
101 new_flags |= VM_WIPEONFORK;
103 case MADV_KEEPONFORK:
104 new_flags &= ~VM_WIPEONFORK;
107 new_flags |= VM_DONTDUMP;
110 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
114 new_flags &= ~VM_DONTDUMP;
117 case MADV_UNMERGEABLE:
118 error = ksm_madvise(vma, start, end, behavior, &new_flags);
120 goto out_convert_errno;
123 case MADV_NOHUGEPAGE:
124 error = hugepage_madvise(vma, &new_flags, behavior);
126 goto out_convert_errno;
130 if (new_flags == vma->vm_flags) {
135 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
136 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
137 vma->vm_file, pgoff, vma_policy(vma),
138 vma->vm_userfaultfd_ctx);
146 if (start != vma->vm_start) {
147 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
151 error = __split_vma(mm, vma, start, 1);
153 goto out_convert_errno;
156 if (end != vma->vm_end) {
157 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
161 error = __split_vma(mm, vma, end, 0);
163 goto out_convert_errno;
168 * vm_flags is protected by the mmap_lock held in write mode.
170 vma->vm_flags = new_flags;
174 * madvise() returns EAGAIN if kernel resources, such as
175 * slab, are temporarily unavailable.
177 if (error == -ENOMEM)
184 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
185 unsigned long end, struct mm_walk *walk)
188 struct vm_area_struct *vma = walk->private;
191 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
194 for (index = start; index != end; index += PAGE_SIZE) {
200 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
201 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
202 pte_unmap_unlock(orig_pte, ptl);
204 if (pte_present(pte) || pte_none(pte))
206 entry = pte_to_swp_entry(pte);
207 if (unlikely(non_swap_entry(entry)))
210 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
219 static const struct mm_walk_ops swapin_walk_ops = {
220 .pmd_entry = swapin_walk_pmd_entry,
223 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
224 unsigned long start, unsigned long end,
225 struct address_space *mapping)
227 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
228 pgoff_t end_index = end / PAGE_SIZE;
232 xas_for_each(&xas, page, end_index) {
235 if (!xa_is_value(page))
240 swap = radix_to_swp_entry(page);
241 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
250 lru_add_drain(); /* Push any new pages onto the LRU now */
252 #endif /* CONFIG_SWAP */
255 * Schedule all required I/O operations. Do not wait for completion.
257 static long madvise_willneed(struct vm_area_struct *vma,
258 struct vm_area_struct **prev,
259 unsigned long start, unsigned long end)
261 struct file *file = vma->vm_file;
267 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
268 lru_add_drain(); /* Push any new pages onto the LRU now */
272 if (shmem_mapping(file->f_mapping)) {
273 force_shm_swapin_readahead(vma, start, end,
282 if (IS_DAX(file_inode(file))) {
283 /* no bad return value, but ignore advice */
288 * Filesystem's fadvise may need to take various locks. We need to
289 * explicitly grab a reference because the vma (and hence the
290 * vma's reference to the file) can go away as soon as we drop
293 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
295 offset = (loff_t)(start - vma->vm_start)
296 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
297 mmap_read_unlock(current->mm);
298 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
300 mmap_read_lock(current->mm);
304 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
305 unsigned long addr, unsigned long end,
306 struct mm_walk *walk)
308 struct madvise_walk_private *private = walk->private;
309 struct mmu_gather *tlb = private->tlb;
310 bool pageout = private->pageout;
311 struct mm_struct *mm = tlb->mm;
312 struct vm_area_struct *vma = walk->vma;
313 pte_t *orig_pte, *pte, ptent;
315 struct page *page = NULL;
316 LIST_HEAD(page_list);
318 if (fatal_signal_pending(current))
321 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
322 if (pmd_trans_huge(*pmd)) {
324 unsigned long next = pmd_addr_end(addr, end);
326 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
327 ptl = pmd_trans_huge_lock(pmd, vma);
332 if (is_huge_zero_pmd(orig_pmd))
335 if (unlikely(!pmd_present(orig_pmd))) {
336 VM_BUG_ON(thp_migration_supported() &&
337 !is_pmd_migration_entry(orig_pmd));
341 page = pmd_page(orig_pmd);
343 /* Do not interfere with other mappings of this page */
344 if (page_mapcount(page) != 1)
347 if (next - addr != HPAGE_PMD_SIZE) {
353 err = split_huge_page(page);
361 if (pmd_young(orig_pmd)) {
362 pmdp_invalidate(vma, addr, pmd);
363 orig_pmd = pmd_mkold(orig_pmd);
365 set_pmd_at(mm, addr, pmd, orig_pmd);
366 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
369 ClearPageReferenced(page);
370 test_and_clear_page_young(page);
372 if (!isolate_lru_page(page)) {
373 if (PageUnevictable(page))
374 putback_lru_page(page);
376 list_add(&page->lru, &page_list);
379 deactivate_page(page);
383 reclaim_pages(&page_list);
388 if (pmd_trans_unstable(pmd))
391 tlb_change_page_size(tlb, PAGE_SIZE);
392 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
393 flush_tlb_batched_pending(mm);
394 arch_enter_lazy_mmu_mode();
395 for (; addr < end; pte++, addr += PAGE_SIZE) {
401 if (!pte_present(ptent))
404 page = vm_normal_page(vma, addr, ptent);
409 * Creating a THP page is expensive so split it only if we
410 * are sure it's worth. Split it if we are only owner.
412 if (PageTransCompound(page)) {
413 if (page_mapcount(page) != 1)
416 if (!trylock_page(page)) {
420 pte_unmap_unlock(orig_pte, ptl);
421 if (split_huge_page(page)) {
424 pte_offset_map_lock(mm, pmd, addr, &ptl);
429 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
435 /* Do not interfere with other mappings of this page */
436 if (page_mapcount(page) != 1)
439 VM_BUG_ON_PAGE(PageTransCompound(page), page);
441 if (pte_young(ptent)) {
442 ptent = ptep_get_and_clear_full(mm, addr, pte,
444 ptent = pte_mkold(ptent);
445 set_pte_at(mm, addr, pte, ptent);
446 tlb_remove_tlb_entry(tlb, pte, addr);
450 * We are deactivating a page for accelerating reclaiming.
451 * VM couldn't reclaim the page unless we clear PG_young.
452 * As a side effect, it makes confuse idle-page tracking
453 * because they will miss recent referenced history.
455 ClearPageReferenced(page);
456 test_and_clear_page_young(page);
458 if (!isolate_lru_page(page)) {
459 if (PageUnevictable(page))
460 putback_lru_page(page);
462 list_add(&page->lru, &page_list);
465 deactivate_page(page);
468 arch_leave_lazy_mmu_mode();
469 pte_unmap_unlock(orig_pte, ptl);
471 reclaim_pages(&page_list);
477 static const struct mm_walk_ops cold_walk_ops = {
478 .pmd_entry = madvise_cold_or_pageout_pte_range,
481 static void madvise_cold_page_range(struct mmu_gather *tlb,
482 struct vm_area_struct *vma,
483 unsigned long addr, unsigned long end)
485 struct madvise_walk_private walk_private = {
490 tlb_start_vma(tlb, vma);
491 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
492 tlb_end_vma(tlb, vma);
495 static long madvise_cold(struct vm_area_struct *vma,
496 struct vm_area_struct **prev,
497 unsigned long start_addr, unsigned long end_addr)
499 struct mm_struct *mm = vma->vm_mm;
500 struct mmu_gather tlb;
503 if (!can_madv_lru_vma(vma))
507 tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
508 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
509 tlb_finish_mmu(&tlb, start_addr, end_addr);
514 static void madvise_pageout_page_range(struct mmu_gather *tlb,
515 struct vm_area_struct *vma,
516 unsigned long addr, unsigned long end)
518 struct madvise_walk_private walk_private = {
523 tlb_start_vma(tlb, vma);
524 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
525 tlb_end_vma(tlb, vma);
528 static inline bool can_do_pageout(struct vm_area_struct *vma)
530 if (vma_is_anonymous(vma))
535 * paging out pagecache only for non-anonymous mappings that correspond
536 * to the files the calling process could (if tried) open for writing;
537 * otherwise we'd be including shared non-exclusive mappings, which
538 * opens a side channel.
540 return inode_owner_or_capable(file_inode(vma->vm_file)) ||
541 inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
544 static long madvise_pageout(struct vm_area_struct *vma,
545 struct vm_area_struct **prev,
546 unsigned long start_addr, unsigned long end_addr)
548 struct mm_struct *mm = vma->vm_mm;
549 struct mmu_gather tlb;
552 if (!can_madv_lru_vma(vma))
555 if (!can_do_pageout(vma))
559 tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
560 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
561 tlb_finish_mmu(&tlb, start_addr, end_addr);
566 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
567 unsigned long end, struct mm_walk *walk)
570 struct mmu_gather *tlb = walk->private;
571 struct mm_struct *mm = tlb->mm;
572 struct vm_area_struct *vma = walk->vma;
574 pte_t *orig_pte, *pte, ptent;
579 next = pmd_addr_end(addr, end);
580 if (pmd_trans_huge(*pmd))
581 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
584 if (pmd_trans_unstable(pmd))
587 tlb_change_page_size(tlb, PAGE_SIZE);
588 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
589 flush_tlb_batched_pending(mm);
590 arch_enter_lazy_mmu_mode();
591 for (; addr != end; pte++, addr += PAGE_SIZE) {
597 * If the pte has swp_entry, just clear page table to
598 * prevent swap-in which is more expensive rather than
599 * (page allocation + zeroing).
601 if (!pte_present(ptent)) {
604 entry = pte_to_swp_entry(ptent);
605 if (non_swap_entry(entry))
608 free_swap_and_cache(entry);
609 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
613 page = vm_normal_page(vma, addr, ptent);
618 * If pmd isn't transhuge but the page is THP and
619 * is owned by only this process, split it and
620 * deactivate all pages.
622 if (PageTransCompound(page)) {
623 if (page_mapcount(page) != 1)
626 if (!trylock_page(page)) {
630 pte_unmap_unlock(orig_pte, ptl);
631 if (split_huge_page(page)) {
634 pte_offset_map_lock(mm, pmd, addr, &ptl);
639 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
645 VM_BUG_ON_PAGE(PageTransCompound(page), page);
647 if (PageSwapCache(page) || PageDirty(page)) {
648 if (!trylock_page(page))
651 * If page is shared with others, we couldn't clear
652 * PG_dirty of the page.
654 if (page_mapcount(page) != 1) {
659 if (PageSwapCache(page) && !try_to_free_swap(page)) {
664 ClearPageDirty(page);
668 if (pte_young(ptent) || pte_dirty(ptent)) {
670 * Some of architecture(ex, PPC) don't update TLB
671 * with set_pte_at and tlb_remove_tlb_entry so for
672 * the portability, remap the pte with old|clean
673 * after pte clearing.
675 ptent = ptep_get_and_clear_full(mm, addr, pte,
678 ptent = pte_mkold(ptent);
679 ptent = pte_mkclean(ptent);
680 set_pte_at(mm, addr, pte, ptent);
681 tlb_remove_tlb_entry(tlb, pte, addr);
683 mark_page_lazyfree(page);
687 if (current->mm == mm)
690 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
692 arch_leave_lazy_mmu_mode();
693 pte_unmap_unlock(orig_pte, ptl);
699 static const struct mm_walk_ops madvise_free_walk_ops = {
700 .pmd_entry = madvise_free_pte_range,
703 static int madvise_free_single_vma(struct vm_area_struct *vma,
704 unsigned long start_addr, unsigned long end_addr)
706 struct mm_struct *mm = vma->vm_mm;
707 struct mmu_notifier_range range;
708 struct mmu_gather tlb;
710 /* MADV_FREE works for only anon vma at the moment */
711 if (!vma_is_anonymous(vma))
714 range.start = max(vma->vm_start, start_addr);
715 if (range.start >= vma->vm_end)
717 range.end = min(vma->vm_end, end_addr);
718 if (range.end <= vma->vm_start)
720 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
721 range.start, range.end);
724 tlb_gather_mmu(&tlb, mm, range.start, range.end);
725 update_hiwater_rss(mm);
727 mmu_notifier_invalidate_range_start(&range);
728 tlb_start_vma(&tlb, vma);
729 walk_page_range(vma->vm_mm, range.start, range.end,
730 &madvise_free_walk_ops, &tlb);
731 tlb_end_vma(&tlb, vma);
732 mmu_notifier_invalidate_range_end(&range);
733 tlb_finish_mmu(&tlb, range.start, range.end);
739 * Application no longer needs these pages. If the pages are dirty,
740 * it's OK to just throw them away. The app will be more careful about
741 * data it wants to keep. Be sure to free swap resources too. The
742 * zap_page_range call sets things up for shrink_active_list to actually free
743 * these pages later if no one else has touched them in the meantime,
744 * although we could add these pages to a global reuse list for
745 * shrink_active_list to pick up before reclaiming other pages.
747 * NB: This interface discards data rather than pushes it out to swap,
748 * as some implementations do. This has performance implications for
749 * applications like large transactional databases which want to discard
750 * pages in anonymous maps after committing to backing store the data
751 * that was kept in them. There is no reason to write this data out to
752 * the swap area if the application is discarding it.
754 * An interface that causes the system to free clean pages and flush
755 * dirty pages is already available as msync(MS_INVALIDATE).
757 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
758 unsigned long start, unsigned long end)
760 zap_page_range(vma, start, end - start);
764 static long madvise_dontneed_free(struct vm_area_struct *vma,
765 struct vm_area_struct **prev,
766 unsigned long start, unsigned long end,
770 if (!can_madv_lru_vma(vma))
773 if (!userfaultfd_remove(vma, start, end)) {
774 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
776 mmap_read_lock(current->mm);
777 vma = find_vma(current->mm, start);
780 if (start < vma->vm_start) {
782 * This "vma" under revalidation is the one
783 * with the lowest vma->vm_start where start
784 * is also < vma->vm_end. If start <
785 * vma->vm_start it means an hole materialized
786 * in the user address space within the
787 * virtual range passed to MADV_DONTNEED
792 if (!can_madv_lru_vma(vma))
794 if (end > vma->vm_end) {
796 * Don't fail if end > vma->vm_end. If the old
797 * vma was splitted while the mmap_lock was
798 * released the effect of the concurrent
799 * operation may not cause madvise() to
800 * have an undefined result. There may be an
801 * adjacent next vma that we'll walk
802 * next. userfaultfd_remove() will generate an
803 * UFFD_EVENT_REMOVE repetition on the
804 * end-vma->vm_end range, but the manager can
805 * handle a repetition fine.
809 VM_WARN_ON(start >= end);
812 if (behavior == MADV_DONTNEED)
813 return madvise_dontneed_single_vma(vma, start, end);
814 else if (behavior == MADV_FREE)
815 return madvise_free_single_vma(vma, start, end);
821 * Application wants to free up the pages and associated backing store.
822 * This is effectively punching a hole into the middle of a file.
824 static long madvise_remove(struct vm_area_struct *vma,
825 struct vm_area_struct **prev,
826 unsigned long start, unsigned long end)
832 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
834 if (vma->vm_flags & VM_LOCKED)
839 if (!f || !f->f_mapping || !f->f_mapping->host) {
843 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
846 offset = (loff_t)(start - vma->vm_start)
847 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
850 * Filesystem's fallocate may need to take i_mutex. We need to
851 * explicitly grab a reference because the vma (and hence the
852 * vma's reference to the file) can go away as soon as we drop
856 if (userfaultfd_remove(vma, start, end)) {
857 /* mmap_lock was not released by userfaultfd_remove() */
858 mmap_read_unlock(current->mm);
860 error = vfs_fallocate(f,
861 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
862 offset, end - start);
864 mmap_read_lock(current->mm);
868 #ifdef CONFIG_MEMORY_FAILURE
870 * Error injection support for memory error handling.
872 static int madvise_inject_error(int behavior,
873 unsigned long start, unsigned long end)
878 if (!capable(CAP_SYS_ADMIN))
882 for (; start < end; start += size) {
887 ret = get_user_pages_fast(start, 1, 0, &page);
890 pfn = page_to_pfn(page);
893 * When soft offlining hugepages, after migrating the page
894 * we dissolve it, therefore in the second loop "page" will
895 * no longer be a compound page.
897 size = page_size(compound_head(page));
899 if (behavior == MADV_SOFT_OFFLINE) {
900 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
902 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
904 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
907 * Drop the page reference taken by get_user_pages_fast(). In
908 * the absence of MF_COUNT_INCREASED the memory_failure()
909 * routine is responsible for pinning the page to prevent it
910 * from being released back to the page allocator.
913 ret = memory_failure(pfn, 0);
920 /* Ensure that all poisoned pages are removed from per-cpu lists */
921 for_each_populated_zone(zone)
922 drain_all_pages(zone);
929 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
930 unsigned long start, unsigned long end, int behavior)
934 return madvise_remove(vma, prev, start, end);
936 return madvise_willneed(vma, prev, start, end);
938 return madvise_cold(vma, prev, start, end);
940 return madvise_pageout(vma, prev, start, end);
943 return madvise_dontneed_free(vma, prev, start, end, behavior);
945 return madvise_behavior(vma, prev, start, end, behavior);
950 madvise_behavior_valid(int behavior)
956 case MADV_SEQUENTIAL:
966 case MADV_UNMERGEABLE:
968 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
970 case MADV_NOHUGEPAGE:
974 case MADV_WIPEONFORK:
975 case MADV_KEEPONFORK:
976 #ifdef CONFIG_MEMORY_FAILURE
977 case MADV_SOFT_OFFLINE:
988 * The madvise(2) system call.
990 * Applications can use madvise() to advise the kernel how it should
991 * handle paging I/O in this VM area. The idea is to help the kernel
992 * use appropriate read-ahead and caching techniques. The information
993 * provided is advisory only, and can be safely disregarded by the
994 * kernel without affecting the correct operation of the application.
997 * MADV_NORMAL - the default behavior is to read clusters. This
998 * results in some read-ahead and read-behind.
999 * MADV_RANDOM - the system should read the minimum amount of data
1000 * on any access, since it is unlikely that the appli-
1001 * cation will need more than what it asks for.
1002 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1003 * once, so they can be aggressively read ahead, and
1004 * can be freed soon after they are accessed.
1005 * MADV_WILLNEED - the application is notifying the system to read
1007 * MADV_DONTNEED - the application is finished with the given range,
1008 * so the kernel can free resources associated with it.
1009 * MADV_FREE - the application marks pages in the given range as lazy free,
1010 * where actual purges are postponed until memory pressure happens.
1011 * MADV_REMOVE - the application wants to free up the given range of
1012 * pages and associated backing store.
1013 * MADV_DONTFORK - omit this area from child's address space when forking:
1014 * typically, to avoid COWing pages pinned by get_user_pages().
1015 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1016 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1017 * range after a fork.
1018 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1019 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1020 * were corrupted by unrecoverable hardware memory failure.
1021 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1022 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1023 * this area with pages of identical content from other such areas.
1024 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1025 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1026 * huge pages in the future. Existing pages might be coalesced and
1027 * new pages might be allocated as THP.
1028 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1029 * transparent huge pages so the existing pages will not be
1030 * coalesced into THP and new pages will not be allocated as THP.
1031 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1032 * from being included in its core dump.
1033 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1037 * -EINVAL - start + len < 0, start is not page-aligned,
1038 * "behavior" is not a valid value, or application
1039 * is attempting to release locked or shared pages,
1040 * or the specified address range includes file, Huge TLB,
1041 * MAP_SHARED or VMPFNMAP range.
1042 * -ENOMEM - addresses in the specified range are not currently
1043 * mapped, or are outside the AS of the process.
1044 * -EIO - an I/O error occurred while paging in data.
1045 * -EBADF - map exists, but area maps something that isn't a file.
1046 * -EAGAIN - a kernel resource was temporarily unavailable.
1048 int do_madvise(unsigned long start, size_t len_in, int behavior)
1050 unsigned long end, tmp;
1051 struct vm_area_struct *vma, *prev;
1052 int unmapped_error = 0;
1053 int error = -EINVAL;
1056 struct blk_plug plug;
1058 start = untagged_addr(start);
1060 if (!madvise_behavior_valid(behavior))
1063 if (!PAGE_ALIGNED(start))
1065 len = PAGE_ALIGN(len_in);
1067 /* Check to see whether len was rounded up from small -ve to zero */
1079 #ifdef CONFIG_MEMORY_FAILURE
1080 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1081 return madvise_inject_error(behavior, start, start + len_in);
1084 write = madvise_need_mmap_write(behavior);
1086 if (mmap_write_lock_killable(current->mm))
1090 * We may have stolen the mm from another process
1091 * that is undergoing core dumping.
1093 * Right now that's io_ring, in the future it may
1094 * be remote process management and not "current"
1097 * We need to fix core dumping to not do this,
1098 * but for now we have the mmget_still_valid()
1101 if (!mmget_still_valid(current->mm)) {
1102 mmap_write_unlock(current->mm);
1106 mmap_read_lock(current->mm);
1110 * If the interval [start,end) covers some unmapped address
1111 * ranges, just ignore them, but return -ENOMEM at the end.
1112 * - different from the way of handling in mlock etc.
1114 vma = find_vma_prev(current->mm, start, &prev);
1115 if (vma && start > vma->vm_start)
1118 blk_start_plug(&plug);
1120 /* Still start < end. */
1125 /* Here start < (end|vma->vm_end). */
1126 if (start < vma->vm_start) {
1127 unmapped_error = -ENOMEM;
1128 start = vma->vm_start;
1133 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1138 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1139 error = madvise_vma(vma, &prev, start, tmp, behavior);
1143 if (prev && start < prev->vm_end)
1144 start = prev->vm_end;
1145 error = unmapped_error;
1149 vma = prev->vm_next;
1150 else /* madvise_remove dropped mmap_lock */
1151 vma = find_vma(current->mm, start);
1154 blk_finish_plug(&plug);
1156 mmap_write_unlock(current->mm);
1158 mmap_read_unlock(current->mm);
1163 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1165 return do_madvise(start, len_in, behavior);