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/sched/mm.h>
21 #include <linux/uio.h>
22 #include <linux/ksm.h>
24 #include <linux/file.h>
25 #include <linux/blkdev.h>
26 #include <linux/backing-dev.h>
27 #include <linux/pagewalk.h>
28 #include <linux/swap.h>
29 #include <linux/swapops.h>
30 #include <linux/shmem_fs.h>
31 #include <linux/mmu_notifier.h>
37 struct madvise_walk_private {
38 struct mmu_gather *tlb;
43 * Any behaviour which results in changes to the vma->vm_flags needs to
44 * take mmap_lock for writing. Others, which simply traverse vmas, need
45 * to only take it for reading.
47 static int madvise_need_mmap_write(int behavior)
56 case MADV_POPULATE_READ:
57 case MADV_POPULATE_WRITE:
60 /* be safe, default to 1. list exceptions explicitly */
66 * We can potentially split a vm area into separate
67 * areas, each area with its own behavior.
69 static long madvise_behavior(struct vm_area_struct *vma,
70 struct vm_area_struct **prev,
71 unsigned long start, unsigned long end, int behavior)
73 struct mm_struct *mm = vma->vm_mm;
76 unsigned long new_flags = vma->vm_flags;
80 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
83 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
86 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
89 new_flags |= VM_DONTCOPY;
92 if (vma->vm_flags & VM_IO) {
96 new_flags &= ~VM_DONTCOPY;
99 /* MADV_WIPEONFORK is only supported on anonymous memory. */
100 if (vma->vm_file || vma->vm_flags & VM_SHARED) {
104 new_flags |= VM_WIPEONFORK;
106 case MADV_KEEPONFORK:
107 new_flags &= ~VM_WIPEONFORK;
110 new_flags |= VM_DONTDUMP;
113 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
117 new_flags &= ~VM_DONTDUMP;
120 case MADV_UNMERGEABLE:
121 error = ksm_madvise(vma, start, end, behavior, &new_flags);
123 goto out_convert_errno;
126 case MADV_NOHUGEPAGE:
127 error = hugepage_madvise(vma, &new_flags, behavior);
129 goto out_convert_errno;
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);
156 goto out_convert_errno;
159 if (end != vma->vm_end) {
160 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
164 error = __split_vma(mm, vma, end, 0);
166 goto out_convert_errno;
171 * vm_flags is protected by the mmap_lock held in write mode.
173 vma->vm_flags = new_flags;
177 * madvise() returns EAGAIN if kernel resources, such as
178 * slab, are temporarily unavailable.
180 if (error == -ENOMEM)
187 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
188 unsigned long end, struct mm_walk *walk)
191 struct vm_area_struct *vma = walk->private;
194 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
197 for (index = start; index != end; index += PAGE_SIZE) {
203 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
204 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
205 pte_unmap_unlock(orig_pte, ptl);
207 if (pte_present(pte) || pte_none(pte))
209 entry = pte_to_swp_entry(pte);
210 if (unlikely(non_swap_entry(entry)))
213 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
222 static const struct mm_walk_ops swapin_walk_ops = {
223 .pmd_entry = swapin_walk_pmd_entry,
226 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
227 unsigned long start, unsigned long end,
228 struct address_space *mapping)
230 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
231 pgoff_t end_index = linear_page_index(vma, end + PAGE_SIZE - 1);
235 xas_for_each(&xas, page, end_index) {
238 if (!xa_is_value(page))
243 swap = radix_to_swp_entry(page);
244 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
253 lru_add_drain(); /* Push any new pages onto the LRU now */
255 #endif /* CONFIG_SWAP */
258 * Schedule all required I/O operations. Do not wait for completion.
260 static long madvise_willneed(struct vm_area_struct *vma,
261 struct vm_area_struct **prev,
262 unsigned long start, unsigned long end)
264 struct mm_struct *mm = vma->vm_mm;
265 struct file *file = vma->vm_file;
271 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
272 lru_add_drain(); /* Push any new pages onto the LRU now */
276 if (shmem_mapping(file->f_mapping)) {
277 force_shm_swapin_readahead(vma, start, end,
286 if (IS_DAX(file_inode(file))) {
287 /* no bad return value, but ignore advice */
292 * Filesystem's fadvise may need to take various locks. We need to
293 * explicitly grab a reference because the vma (and hence the
294 * vma's reference to the file) can go away as soon as we drop
297 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
299 offset = (loff_t)(start - vma->vm_start)
300 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
301 mmap_read_unlock(mm);
302 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
308 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
309 unsigned long addr, unsigned long end,
310 struct mm_walk *walk)
312 struct madvise_walk_private *private = walk->private;
313 struct mmu_gather *tlb = private->tlb;
314 bool pageout = private->pageout;
315 struct mm_struct *mm = tlb->mm;
316 struct vm_area_struct *vma = walk->vma;
317 pte_t *orig_pte, *pte, ptent;
319 struct page *page = NULL;
320 LIST_HEAD(page_list);
322 if (fatal_signal_pending(current))
325 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
326 if (pmd_trans_huge(*pmd)) {
328 unsigned long next = pmd_addr_end(addr, end);
330 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
331 ptl = pmd_trans_huge_lock(pmd, vma);
336 if (is_huge_zero_pmd(orig_pmd))
339 if (unlikely(!pmd_present(orig_pmd))) {
340 VM_BUG_ON(thp_migration_supported() &&
341 !is_pmd_migration_entry(orig_pmd));
345 page = pmd_page(orig_pmd);
347 /* Do not interfere with other mappings of this page */
348 if (page_mapcount(page) != 1)
351 if (next - addr != HPAGE_PMD_SIZE) {
357 err = split_huge_page(page);
365 if (pmd_young(orig_pmd)) {
366 pmdp_invalidate(vma, addr, pmd);
367 orig_pmd = pmd_mkold(orig_pmd);
369 set_pmd_at(mm, addr, pmd, orig_pmd);
370 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
373 ClearPageReferenced(page);
374 test_and_clear_page_young(page);
376 if (!isolate_lru_page(page)) {
377 if (PageUnevictable(page))
378 putback_lru_page(page);
380 list_add(&page->lru, &page_list);
383 deactivate_page(page);
387 reclaim_pages(&page_list);
392 if (pmd_trans_unstable(pmd))
395 tlb_change_page_size(tlb, PAGE_SIZE);
396 orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
397 flush_tlb_batched_pending(mm);
398 arch_enter_lazy_mmu_mode();
399 for (; addr < end; pte++, addr += PAGE_SIZE) {
405 if (!pte_present(ptent))
408 page = vm_normal_page(vma, addr, ptent);
413 * Creating a THP page is expensive so split it only if we
414 * are sure it's worth. Split it if we are only owner.
416 if (PageTransCompound(page)) {
417 if (page_mapcount(page) != 1)
420 if (!trylock_page(page)) {
424 pte_unmap_unlock(orig_pte, ptl);
425 if (split_huge_page(page)) {
428 pte_offset_map_lock(mm, pmd, addr, &ptl);
433 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
439 /* Do not interfere with other mappings of this page */
440 if (page_mapcount(page) != 1)
443 VM_BUG_ON_PAGE(PageTransCompound(page), page);
445 if (pte_young(ptent)) {
446 ptent = ptep_get_and_clear_full(mm, addr, pte,
448 ptent = pte_mkold(ptent);
449 set_pte_at(mm, addr, pte, ptent);
450 tlb_remove_tlb_entry(tlb, pte, addr);
454 * We are deactivating a page for accelerating reclaiming.
455 * VM couldn't reclaim the page unless we clear PG_young.
456 * As a side effect, it makes confuse idle-page tracking
457 * because they will miss recent referenced history.
459 ClearPageReferenced(page);
460 test_and_clear_page_young(page);
462 if (!isolate_lru_page(page)) {
463 if (PageUnevictable(page))
464 putback_lru_page(page);
466 list_add(&page->lru, &page_list);
469 deactivate_page(page);
472 arch_leave_lazy_mmu_mode();
473 pte_unmap_unlock(orig_pte, ptl);
475 reclaim_pages(&page_list);
481 static const struct mm_walk_ops cold_walk_ops = {
482 .pmd_entry = madvise_cold_or_pageout_pte_range,
485 static void madvise_cold_page_range(struct mmu_gather *tlb,
486 struct vm_area_struct *vma,
487 unsigned long addr, unsigned long end)
489 struct madvise_walk_private walk_private = {
494 tlb_start_vma(tlb, vma);
495 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
496 tlb_end_vma(tlb, vma);
499 static long madvise_cold(struct vm_area_struct *vma,
500 struct vm_area_struct **prev,
501 unsigned long start_addr, unsigned long end_addr)
503 struct mm_struct *mm = vma->vm_mm;
504 struct mmu_gather tlb;
507 if (!can_madv_lru_vma(vma))
511 tlb_gather_mmu(&tlb, mm);
512 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
513 tlb_finish_mmu(&tlb);
518 static void madvise_pageout_page_range(struct mmu_gather *tlb,
519 struct vm_area_struct *vma,
520 unsigned long addr, unsigned long end)
522 struct madvise_walk_private walk_private = {
527 tlb_start_vma(tlb, vma);
528 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
529 tlb_end_vma(tlb, vma);
532 static inline bool can_do_pageout(struct vm_area_struct *vma)
534 if (vma_is_anonymous(vma))
539 * paging out pagecache only for non-anonymous mappings that correspond
540 * to the files the calling process could (if tried) open for writing;
541 * otherwise we'd be including shared non-exclusive mappings, which
542 * opens a side channel.
544 return inode_owner_or_capable(&init_user_ns,
545 file_inode(vma->vm_file)) ||
546 file_permission(vma->vm_file, MAY_WRITE) == 0;
549 static long madvise_pageout(struct vm_area_struct *vma,
550 struct vm_area_struct **prev,
551 unsigned long start_addr, unsigned long end_addr)
553 struct mm_struct *mm = vma->vm_mm;
554 struct mmu_gather tlb;
557 if (!can_madv_lru_vma(vma))
560 if (!can_do_pageout(vma))
564 tlb_gather_mmu(&tlb, mm);
565 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
566 tlb_finish_mmu(&tlb);
571 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
572 unsigned long end, struct mm_walk *walk)
575 struct mmu_gather *tlb = walk->private;
576 struct mm_struct *mm = tlb->mm;
577 struct vm_area_struct *vma = walk->vma;
579 pte_t *orig_pte, *pte, ptent;
584 next = pmd_addr_end(addr, end);
585 if (pmd_trans_huge(*pmd))
586 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
589 if (pmd_trans_unstable(pmd))
592 tlb_change_page_size(tlb, PAGE_SIZE);
593 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
594 flush_tlb_batched_pending(mm);
595 arch_enter_lazy_mmu_mode();
596 for (; addr != end; pte++, addr += PAGE_SIZE) {
602 * If the pte has swp_entry, just clear page table to
603 * prevent swap-in which is more expensive rather than
604 * (page allocation + zeroing).
606 if (!pte_present(ptent)) {
609 entry = pte_to_swp_entry(ptent);
610 if (non_swap_entry(entry))
613 free_swap_and_cache(entry);
614 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
618 page = vm_normal_page(vma, addr, ptent);
623 * If pmd isn't transhuge but the page is THP and
624 * is owned by only this process, split it and
625 * deactivate all pages.
627 if (PageTransCompound(page)) {
628 if (page_mapcount(page) != 1)
631 if (!trylock_page(page)) {
635 pte_unmap_unlock(orig_pte, ptl);
636 if (split_huge_page(page)) {
639 pte_offset_map_lock(mm, pmd, addr, &ptl);
644 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
650 VM_BUG_ON_PAGE(PageTransCompound(page), page);
652 if (PageSwapCache(page) || PageDirty(page)) {
653 if (!trylock_page(page))
656 * If page is shared with others, we couldn't clear
657 * PG_dirty of the page.
659 if (page_mapcount(page) != 1) {
664 if (PageSwapCache(page) && !try_to_free_swap(page)) {
669 ClearPageDirty(page);
673 if (pte_young(ptent) || pte_dirty(ptent)) {
675 * Some of architecture(ex, PPC) don't update TLB
676 * with set_pte_at and tlb_remove_tlb_entry so for
677 * the portability, remap the pte with old|clean
678 * after pte clearing.
680 ptent = ptep_get_and_clear_full(mm, addr, pte,
683 ptent = pte_mkold(ptent);
684 ptent = pte_mkclean(ptent);
685 set_pte_at(mm, addr, pte, ptent);
686 tlb_remove_tlb_entry(tlb, pte, addr);
688 mark_page_lazyfree(page);
692 if (current->mm == mm)
695 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
697 arch_leave_lazy_mmu_mode();
698 pte_unmap_unlock(orig_pte, ptl);
704 static const struct mm_walk_ops madvise_free_walk_ops = {
705 .pmd_entry = madvise_free_pte_range,
708 static int madvise_free_single_vma(struct vm_area_struct *vma,
709 unsigned long start_addr, unsigned long end_addr)
711 struct mm_struct *mm = vma->vm_mm;
712 struct mmu_notifier_range range;
713 struct mmu_gather tlb;
715 /* MADV_FREE works for only anon vma at the moment */
716 if (!vma_is_anonymous(vma))
719 range.start = max(vma->vm_start, start_addr);
720 if (range.start >= vma->vm_end)
722 range.end = min(vma->vm_end, end_addr);
723 if (range.end <= vma->vm_start)
725 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
726 range.start, range.end);
729 tlb_gather_mmu(&tlb, mm);
730 update_hiwater_rss(mm);
732 mmu_notifier_invalidate_range_start(&range);
733 tlb_start_vma(&tlb, vma);
734 walk_page_range(vma->vm_mm, range.start, range.end,
735 &madvise_free_walk_ops, &tlb);
736 tlb_end_vma(&tlb, vma);
737 mmu_notifier_invalidate_range_end(&range);
738 tlb_finish_mmu(&tlb);
744 * Application no longer needs these pages. If the pages are dirty,
745 * it's OK to just throw them away. The app will be more careful about
746 * data it wants to keep. Be sure to free swap resources too. The
747 * zap_page_range call sets things up for shrink_active_list to actually free
748 * these pages later if no one else has touched them in the meantime,
749 * although we could add these pages to a global reuse list for
750 * shrink_active_list to pick up before reclaiming other pages.
752 * NB: This interface discards data rather than pushes it out to swap,
753 * as some implementations do. This has performance implications for
754 * applications like large transactional databases which want to discard
755 * pages in anonymous maps after committing to backing store the data
756 * that was kept in them. There is no reason to write this data out to
757 * the swap area if the application is discarding it.
759 * An interface that causes the system to free clean pages and flush
760 * dirty pages is already available as msync(MS_INVALIDATE).
762 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
763 unsigned long start, unsigned long end)
765 zap_page_range(vma, start, end - start);
769 static long madvise_dontneed_free(struct vm_area_struct *vma,
770 struct vm_area_struct **prev,
771 unsigned long start, unsigned long end,
774 struct mm_struct *mm = vma->vm_mm;
777 if (!can_madv_lru_vma(vma))
780 if (!userfaultfd_remove(vma, start, end)) {
781 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
784 vma = find_vma(mm, start);
787 if (start < vma->vm_start) {
789 * This "vma" under revalidation is the one
790 * with the lowest vma->vm_start where start
791 * is also < vma->vm_end. If start <
792 * vma->vm_start it means an hole materialized
793 * in the user address space within the
794 * virtual range passed to MADV_DONTNEED
799 if (!can_madv_lru_vma(vma))
801 if (end > vma->vm_end) {
803 * Don't fail if end > vma->vm_end. If the old
804 * vma was split while the mmap_lock was
805 * released the effect of the concurrent
806 * operation may not cause madvise() to
807 * have an undefined result. There may be an
808 * adjacent next vma that we'll walk
809 * next. userfaultfd_remove() will generate an
810 * UFFD_EVENT_REMOVE repetition on the
811 * end-vma->vm_end range, but the manager can
812 * handle a repetition fine.
816 VM_WARN_ON(start >= end);
819 if (behavior == MADV_DONTNEED)
820 return madvise_dontneed_single_vma(vma, start, end);
821 else if (behavior == MADV_FREE)
822 return madvise_free_single_vma(vma, start, end);
827 static long madvise_populate(struct vm_area_struct *vma,
828 struct vm_area_struct **prev,
829 unsigned long start, unsigned long end,
832 const bool write = behavior == MADV_POPULATE_WRITE;
833 struct mm_struct *mm = vma->vm_mm;
834 unsigned long tmp_end;
840 while (start < end) {
842 * We might have temporarily dropped the lock. For example,
843 * our VMA might have been split.
845 if (!vma || start >= vma->vm_end) {
846 vma = find_vma(mm, start);
847 if (!vma || start < vma->vm_start)
851 tmp_end = min_t(unsigned long, end, vma->vm_end);
852 /* Populate (prefault) page tables readable/writable. */
853 pages = faultin_vma_page_range(vma, start, tmp_end, write,
865 case -EINVAL: /* Incompatible mappings / permissions. */
869 case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
872 pr_warn_once("%s: unhandled return value: %ld\n",
879 start += pages * PAGE_SIZE;
885 * Application wants to free up the pages and associated backing store.
886 * This is effectively punching a hole into the middle of a file.
888 static long madvise_remove(struct vm_area_struct *vma,
889 struct vm_area_struct **prev,
890 unsigned long start, unsigned long end)
895 struct mm_struct *mm = vma->vm_mm;
897 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
899 if (vma->vm_flags & VM_LOCKED)
904 if (!f || !f->f_mapping || !f->f_mapping->host) {
908 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
911 offset = (loff_t)(start - vma->vm_start)
912 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
915 * Filesystem's fallocate may need to take i_mutex. We need to
916 * explicitly grab a reference because the vma (and hence the
917 * vma's reference to the file) can go away as soon as we drop
921 if (userfaultfd_remove(vma, start, end)) {
922 /* mmap_lock was not released by userfaultfd_remove() */
923 mmap_read_unlock(mm);
925 error = vfs_fallocate(f,
926 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
927 offset, end - start);
933 #ifdef CONFIG_MEMORY_FAILURE
935 * Error injection support for memory error handling.
937 static int madvise_inject_error(int behavior,
938 unsigned long start, unsigned long end)
942 if (!capable(CAP_SYS_ADMIN))
946 for (; start < end; start += size) {
951 ret = get_user_pages_fast(start, 1, 0, &page);
954 pfn = page_to_pfn(page);
957 * When soft offlining hugepages, after migrating the page
958 * we dissolve it, therefore in the second loop "page" will
959 * no longer be a compound page.
961 size = page_size(compound_head(page));
963 if (behavior == MADV_SOFT_OFFLINE) {
964 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
966 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
968 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
970 ret = memory_failure(pfn, MF_COUNT_INCREASED);
982 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
983 unsigned long start, unsigned long end, int behavior)
987 return madvise_remove(vma, prev, start, end);
989 return madvise_willneed(vma, prev, start, end);
991 return madvise_cold(vma, prev, start, end);
993 return madvise_pageout(vma, prev, start, end);
996 return madvise_dontneed_free(vma, prev, start, end, behavior);
997 case MADV_POPULATE_READ:
998 case MADV_POPULATE_WRITE:
999 return madvise_populate(vma, prev, start, end, behavior);
1001 return madvise_behavior(vma, prev, start, end, behavior);
1006 madvise_behavior_valid(int behavior)
1012 case MADV_SEQUENTIAL:
1020 case MADV_POPULATE_READ:
1021 case MADV_POPULATE_WRITE:
1023 case MADV_MERGEABLE:
1024 case MADV_UNMERGEABLE:
1026 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1028 case MADV_NOHUGEPAGE:
1032 case MADV_WIPEONFORK:
1033 case MADV_KEEPONFORK:
1034 #ifdef CONFIG_MEMORY_FAILURE
1035 case MADV_SOFT_OFFLINE:
1046 process_madvise_behavior_valid(int behavior)
1058 * The madvise(2) system call.
1060 * Applications can use madvise() to advise the kernel how it should
1061 * handle paging I/O in this VM area. The idea is to help the kernel
1062 * use appropriate read-ahead and caching techniques. The information
1063 * provided is advisory only, and can be safely disregarded by the
1064 * kernel without affecting the correct operation of the application.
1067 * MADV_NORMAL - the default behavior is to read clusters. This
1068 * results in some read-ahead and read-behind.
1069 * MADV_RANDOM - the system should read the minimum amount of data
1070 * on any access, since it is unlikely that the appli-
1071 * cation will need more than what it asks for.
1072 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1073 * once, so they can be aggressively read ahead, and
1074 * can be freed soon after they are accessed.
1075 * MADV_WILLNEED - the application is notifying the system to read
1077 * MADV_DONTNEED - the application is finished with the given range,
1078 * so the kernel can free resources associated with it.
1079 * MADV_FREE - the application marks pages in the given range as lazy free,
1080 * where actual purges are postponed until memory pressure happens.
1081 * MADV_REMOVE - the application wants to free up the given range of
1082 * pages and associated backing store.
1083 * MADV_DONTFORK - omit this area from child's address space when forking:
1084 * typically, to avoid COWing pages pinned by get_user_pages().
1085 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1086 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1087 * range after a fork.
1088 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1089 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1090 * were corrupted by unrecoverable hardware memory failure.
1091 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1092 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1093 * this area with pages of identical content from other such areas.
1094 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1095 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1096 * huge pages in the future. Existing pages might be coalesced and
1097 * new pages might be allocated as THP.
1098 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1099 * transparent huge pages so the existing pages will not be
1100 * coalesced into THP and new pages will not be allocated as THP.
1101 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1102 * from being included in its core dump.
1103 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1104 * MADV_COLD - the application is not expected to use this memory soon,
1105 * deactivate pages in this range so that they can be reclaimed
1106 * easily if memory pressure happens.
1107 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1108 * page out the pages in this range immediately.
1109 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1110 * triggering read faults if required
1111 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1112 * triggering write faults if required
1116 * -EINVAL - start + len < 0, start is not page-aligned,
1117 * "behavior" is not a valid value, or application
1118 * is attempting to release locked or shared pages,
1119 * or the specified address range includes file, Huge TLB,
1120 * MAP_SHARED or VMPFNMAP range.
1121 * -ENOMEM - addresses in the specified range are not currently
1122 * mapped, or are outside the AS of the process.
1123 * -EIO - an I/O error occurred while paging in data.
1124 * -EBADF - map exists, but area maps something that isn't a file.
1125 * -EAGAIN - a kernel resource was temporarily unavailable.
1127 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1129 unsigned long end, tmp;
1130 struct vm_area_struct *vma, *prev;
1131 int unmapped_error = 0;
1132 int error = -EINVAL;
1135 struct blk_plug plug;
1137 start = untagged_addr(start);
1139 if (!madvise_behavior_valid(behavior))
1142 if (!PAGE_ALIGNED(start))
1144 len = PAGE_ALIGN(len_in);
1146 /* Check to see whether len was rounded up from small -ve to zero */
1158 #ifdef CONFIG_MEMORY_FAILURE
1159 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1160 return madvise_inject_error(behavior, start, start + len_in);
1163 write = madvise_need_mmap_write(behavior);
1165 if (mmap_write_lock_killable(mm))
1172 * If the interval [start,end) covers some unmapped address
1173 * ranges, just ignore them, but return -ENOMEM at the end.
1174 * - different from the way of handling in mlock etc.
1176 vma = find_vma_prev(mm, start, &prev);
1177 if (vma && start > vma->vm_start)
1180 blk_start_plug(&plug);
1182 /* Still start < end. */
1187 /* Here start < (end|vma->vm_end). */
1188 if (start < vma->vm_start) {
1189 unmapped_error = -ENOMEM;
1190 start = vma->vm_start;
1195 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1200 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1201 error = madvise_vma(vma, &prev, start, tmp, behavior);
1205 if (prev && start < prev->vm_end)
1206 start = prev->vm_end;
1207 error = unmapped_error;
1211 vma = prev->vm_next;
1212 else /* madvise_remove dropped mmap_lock */
1213 vma = find_vma(mm, start);
1216 blk_finish_plug(&plug);
1218 mmap_write_unlock(mm);
1220 mmap_read_unlock(mm);
1225 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1227 return do_madvise(current->mm, start, len_in, behavior);
1230 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1231 size_t, vlen, int, behavior, unsigned int, flags)
1234 struct iovec iovstack[UIO_FASTIOV], iovec;
1235 struct iovec *iov = iovstack;
1236 struct iov_iter iter;
1238 struct task_struct *task;
1239 struct mm_struct *mm;
1241 unsigned int f_flags;
1248 ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1252 pid = pidfd_get_pid(pidfd, &f_flags);
1258 task = get_pid_task(pid, PIDTYPE_PID);
1264 if (!process_madvise_behavior_valid(behavior)) {
1269 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1270 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1271 if (IS_ERR_OR_NULL(mm)) {
1272 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1277 * Require CAP_SYS_NICE for influencing process performance. Note that
1278 * only non-destructive hints are currently supported.
1280 if (!capable(CAP_SYS_NICE)) {
1285 total_len = iov_iter_count(&iter);
1287 while (iov_iter_count(&iter)) {
1288 iovec = iov_iter_iovec(&iter);
1289 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1290 iovec.iov_len, behavior);
1293 iov_iter_advance(&iter, iovec.iov_len);
1297 ret = total_len - iov_iter_count(&iter);
1302 put_task_struct(task);