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);
440 * Do not interfere with other mappings of this page and
443 if (!PageLRU(page) || page_mapcount(page) != 1)
446 VM_BUG_ON_PAGE(PageTransCompound(page), page);
448 if (pte_young(ptent)) {
449 ptent = ptep_get_and_clear_full(mm, addr, pte,
451 ptent = pte_mkold(ptent);
452 set_pte_at(mm, addr, pte, ptent);
453 tlb_remove_tlb_entry(tlb, pte, addr);
457 * We are deactivating a page for accelerating reclaiming.
458 * VM couldn't reclaim the page unless we clear PG_young.
459 * As a side effect, it makes confuse idle-page tracking
460 * because they will miss recent referenced history.
462 ClearPageReferenced(page);
463 test_and_clear_page_young(page);
465 if (!isolate_lru_page(page)) {
466 if (PageUnevictable(page))
467 putback_lru_page(page);
469 list_add(&page->lru, &page_list);
472 deactivate_page(page);
475 arch_leave_lazy_mmu_mode();
476 pte_unmap_unlock(orig_pte, ptl);
478 reclaim_pages(&page_list);
484 static const struct mm_walk_ops cold_walk_ops = {
485 .pmd_entry = madvise_cold_or_pageout_pte_range,
488 static void madvise_cold_page_range(struct mmu_gather *tlb,
489 struct vm_area_struct *vma,
490 unsigned long addr, unsigned long end)
492 struct madvise_walk_private walk_private = {
497 tlb_start_vma(tlb, vma);
498 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
499 tlb_end_vma(tlb, vma);
502 static long madvise_cold(struct vm_area_struct *vma,
503 struct vm_area_struct **prev,
504 unsigned long start_addr, unsigned long end_addr)
506 struct mm_struct *mm = vma->vm_mm;
507 struct mmu_gather tlb;
510 if (!can_madv_lru_vma(vma))
514 tlb_gather_mmu(&tlb, mm);
515 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
516 tlb_finish_mmu(&tlb);
521 static void madvise_pageout_page_range(struct mmu_gather *tlb,
522 struct vm_area_struct *vma,
523 unsigned long addr, unsigned long end)
525 struct madvise_walk_private walk_private = {
530 tlb_start_vma(tlb, vma);
531 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
532 tlb_end_vma(tlb, vma);
535 static inline bool can_do_pageout(struct vm_area_struct *vma)
537 if (vma_is_anonymous(vma))
542 * paging out pagecache only for non-anonymous mappings that correspond
543 * to the files the calling process could (if tried) open for writing;
544 * otherwise we'd be including shared non-exclusive mappings, which
545 * opens a side channel.
547 return inode_owner_or_capable(&init_user_ns,
548 file_inode(vma->vm_file)) ||
549 file_permission(vma->vm_file, MAY_WRITE) == 0;
552 static long madvise_pageout(struct vm_area_struct *vma,
553 struct vm_area_struct **prev,
554 unsigned long start_addr, unsigned long end_addr)
556 struct mm_struct *mm = vma->vm_mm;
557 struct mmu_gather tlb;
560 if (!can_madv_lru_vma(vma))
563 if (!can_do_pageout(vma))
567 tlb_gather_mmu(&tlb, mm);
568 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
569 tlb_finish_mmu(&tlb);
574 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
575 unsigned long end, struct mm_walk *walk)
578 struct mmu_gather *tlb = walk->private;
579 struct mm_struct *mm = tlb->mm;
580 struct vm_area_struct *vma = walk->vma;
582 pte_t *orig_pte, *pte, ptent;
587 next = pmd_addr_end(addr, end);
588 if (pmd_trans_huge(*pmd))
589 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
592 if (pmd_trans_unstable(pmd))
595 tlb_change_page_size(tlb, PAGE_SIZE);
596 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
597 flush_tlb_batched_pending(mm);
598 arch_enter_lazy_mmu_mode();
599 for (; addr != end; pte++, addr += PAGE_SIZE) {
605 * If the pte has swp_entry, just clear page table to
606 * prevent swap-in which is more expensive rather than
607 * (page allocation + zeroing).
609 if (!pte_present(ptent)) {
612 entry = pte_to_swp_entry(ptent);
613 if (non_swap_entry(entry))
616 free_swap_and_cache(entry);
617 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
621 page = vm_normal_page(vma, addr, ptent);
626 * If pmd isn't transhuge but the page is THP and
627 * is owned by only this process, split it and
628 * deactivate all pages.
630 if (PageTransCompound(page)) {
631 if (page_mapcount(page) != 1)
634 if (!trylock_page(page)) {
638 pte_unmap_unlock(orig_pte, ptl);
639 if (split_huge_page(page)) {
642 pte_offset_map_lock(mm, pmd, addr, &ptl);
647 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
653 VM_BUG_ON_PAGE(PageTransCompound(page), page);
655 if (PageSwapCache(page) || PageDirty(page)) {
656 if (!trylock_page(page))
659 * If page is shared with others, we couldn't clear
660 * PG_dirty of the page.
662 if (page_mapcount(page) != 1) {
667 if (PageSwapCache(page) && !try_to_free_swap(page)) {
672 ClearPageDirty(page);
676 if (pte_young(ptent) || pte_dirty(ptent)) {
678 * Some of architecture(ex, PPC) don't update TLB
679 * with set_pte_at and tlb_remove_tlb_entry so for
680 * the portability, remap the pte with old|clean
681 * after pte clearing.
683 ptent = ptep_get_and_clear_full(mm, addr, pte,
686 ptent = pte_mkold(ptent);
687 ptent = pte_mkclean(ptent);
688 set_pte_at(mm, addr, pte, ptent);
689 tlb_remove_tlb_entry(tlb, pte, addr);
691 mark_page_lazyfree(page);
695 if (current->mm == mm)
698 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
700 arch_leave_lazy_mmu_mode();
701 pte_unmap_unlock(orig_pte, ptl);
707 static const struct mm_walk_ops madvise_free_walk_ops = {
708 .pmd_entry = madvise_free_pte_range,
711 static int madvise_free_single_vma(struct vm_area_struct *vma,
712 unsigned long start_addr, unsigned long end_addr)
714 struct mm_struct *mm = vma->vm_mm;
715 struct mmu_notifier_range range;
716 struct mmu_gather tlb;
718 /* MADV_FREE works for only anon vma at the moment */
719 if (!vma_is_anonymous(vma))
722 range.start = max(vma->vm_start, start_addr);
723 if (range.start >= vma->vm_end)
725 range.end = min(vma->vm_end, end_addr);
726 if (range.end <= vma->vm_start)
728 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
729 range.start, range.end);
732 tlb_gather_mmu(&tlb, mm);
733 update_hiwater_rss(mm);
735 mmu_notifier_invalidate_range_start(&range);
736 tlb_start_vma(&tlb, vma);
737 walk_page_range(vma->vm_mm, range.start, range.end,
738 &madvise_free_walk_ops, &tlb);
739 tlb_end_vma(&tlb, vma);
740 mmu_notifier_invalidate_range_end(&range);
741 tlb_finish_mmu(&tlb);
747 * Application no longer needs these pages. If the pages are dirty,
748 * it's OK to just throw them away. The app will be more careful about
749 * data it wants to keep. Be sure to free swap resources too. The
750 * zap_page_range call sets things up for shrink_active_list to actually free
751 * these pages later if no one else has touched them in the meantime,
752 * although we could add these pages to a global reuse list for
753 * shrink_active_list to pick up before reclaiming other pages.
755 * NB: This interface discards data rather than pushes it out to swap,
756 * as some implementations do. This has performance implications for
757 * applications like large transactional databases which want to discard
758 * pages in anonymous maps after committing to backing store the data
759 * that was kept in them. There is no reason to write this data out to
760 * the swap area if the application is discarding it.
762 * An interface that causes the system to free clean pages and flush
763 * dirty pages is already available as msync(MS_INVALIDATE).
765 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
766 unsigned long start, unsigned long end)
768 zap_page_range(vma, start, end - start);
772 static long madvise_dontneed_free(struct vm_area_struct *vma,
773 struct vm_area_struct **prev,
774 unsigned long start, unsigned long end,
777 struct mm_struct *mm = vma->vm_mm;
780 if (!can_madv_lru_vma(vma))
783 if (!userfaultfd_remove(vma, start, end)) {
784 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
787 vma = find_vma(mm, start);
790 if (start < vma->vm_start) {
792 * This "vma" under revalidation is the one
793 * with the lowest vma->vm_start where start
794 * is also < vma->vm_end. If start <
795 * vma->vm_start it means an hole materialized
796 * in the user address space within the
797 * virtual range passed to MADV_DONTNEED
802 if (!can_madv_lru_vma(vma))
804 if (end > vma->vm_end) {
806 * Don't fail if end > vma->vm_end. If the old
807 * vma was split while the mmap_lock was
808 * released the effect of the concurrent
809 * operation may not cause madvise() to
810 * have an undefined result. There may be an
811 * adjacent next vma that we'll walk
812 * next. userfaultfd_remove() will generate an
813 * UFFD_EVENT_REMOVE repetition on the
814 * end-vma->vm_end range, but the manager can
815 * handle a repetition fine.
819 VM_WARN_ON(start >= end);
822 if (behavior == MADV_DONTNEED)
823 return madvise_dontneed_single_vma(vma, start, end);
824 else if (behavior == MADV_FREE)
825 return madvise_free_single_vma(vma, start, end);
830 static long madvise_populate(struct vm_area_struct *vma,
831 struct vm_area_struct **prev,
832 unsigned long start, unsigned long end,
835 const bool write = behavior == MADV_POPULATE_WRITE;
836 struct mm_struct *mm = vma->vm_mm;
837 unsigned long tmp_end;
843 while (start < end) {
845 * We might have temporarily dropped the lock. For example,
846 * our VMA might have been split.
848 if (!vma || start >= vma->vm_end) {
849 vma = find_vma(mm, start);
850 if (!vma || start < vma->vm_start)
854 tmp_end = min_t(unsigned long, end, vma->vm_end);
855 /* Populate (prefault) page tables readable/writable. */
856 pages = faultin_vma_page_range(vma, start, tmp_end, write,
868 case -EINVAL: /* Incompatible mappings / permissions. */
872 case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
875 pr_warn_once("%s: unhandled return value: %ld\n",
882 start += pages * PAGE_SIZE;
888 * Application wants to free up the pages and associated backing store.
889 * This is effectively punching a hole into the middle of a file.
891 static long madvise_remove(struct vm_area_struct *vma,
892 struct vm_area_struct **prev,
893 unsigned long start, unsigned long end)
898 struct mm_struct *mm = vma->vm_mm;
900 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
902 if (vma->vm_flags & VM_LOCKED)
907 if (!f || !f->f_mapping || !f->f_mapping->host) {
911 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
914 offset = (loff_t)(start - vma->vm_start)
915 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
918 * Filesystem's fallocate may need to take i_rwsem. We need to
919 * explicitly grab a reference because the vma (and hence the
920 * vma's reference to the file) can go away as soon as we drop
924 if (userfaultfd_remove(vma, start, end)) {
925 /* mmap_lock was not released by userfaultfd_remove() */
926 mmap_read_unlock(mm);
928 error = vfs_fallocate(f,
929 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
930 offset, end - start);
936 #ifdef CONFIG_MEMORY_FAILURE
938 * Error injection support for memory error handling.
940 static int madvise_inject_error(int behavior,
941 unsigned long start, unsigned long end)
945 if (!capable(CAP_SYS_ADMIN))
949 for (; start < end; start += size) {
954 ret = get_user_pages_fast(start, 1, 0, &page);
957 pfn = page_to_pfn(page);
960 * When soft offlining hugepages, after migrating the page
961 * we dissolve it, therefore in the second loop "page" will
962 * no longer be a compound page.
964 size = page_size(compound_head(page));
966 if (behavior == MADV_SOFT_OFFLINE) {
967 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
969 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
971 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
973 ret = memory_failure(pfn, MF_COUNT_INCREASED);
974 if (ret == -EOPNOTSUPP)
987 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
988 unsigned long start, unsigned long end, int behavior)
992 return madvise_remove(vma, prev, start, end);
994 return madvise_willneed(vma, prev, start, end);
996 return madvise_cold(vma, prev, start, end);
998 return madvise_pageout(vma, prev, start, end);
1001 return madvise_dontneed_free(vma, prev, start, end, behavior);
1002 case MADV_POPULATE_READ:
1003 case MADV_POPULATE_WRITE:
1004 return madvise_populate(vma, prev, start, end, behavior);
1006 return madvise_behavior(vma, prev, start, end, behavior);
1011 madvise_behavior_valid(int behavior)
1017 case MADV_SEQUENTIAL:
1025 case MADV_POPULATE_READ:
1026 case MADV_POPULATE_WRITE:
1028 case MADV_MERGEABLE:
1029 case MADV_UNMERGEABLE:
1031 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1033 case MADV_NOHUGEPAGE:
1037 case MADV_WIPEONFORK:
1038 case MADV_KEEPONFORK:
1039 #ifdef CONFIG_MEMORY_FAILURE
1040 case MADV_SOFT_OFFLINE:
1051 process_madvise_behavior_valid(int behavior)
1064 * The madvise(2) system call.
1066 * Applications can use madvise() to advise the kernel how it should
1067 * handle paging I/O in this VM area. The idea is to help the kernel
1068 * use appropriate read-ahead and caching techniques. The information
1069 * provided is advisory only, and can be safely disregarded by the
1070 * kernel without affecting the correct operation of the application.
1073 * MADV_NORMAL - the default behavior is to read clusters. This
1074 * results in some read-ahead and read-behind.
1075 * MADV_RANDOM - the system should read the minimum amount of data
1076 * on any access, since it is unlikely that the appli-
1077 * cation will need more than what it asks for.
1078 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1079 * once, so they can be aggressively read ahead, and
1080 * can be freed soon after they are accessed.
1081 * MADV_WILLNEED - the application is notifying the system to read
1083 * MADV_DONTNEED - the application is finished with the given range,
1084 * so the kernel can free resources associated with it.
1085 * MADV_FREE - the application marks pages in the given range as lazy free,
1086 * where actual purges are postponed until memory pressure happens.
1087 * MADV_REMOVE - the application wants to free up the given range of
1088 * pages and associated backing store.
1089 * MADV_DONTFORK - omit this area from child's address space when forking:
1090 * typically, to avoid COWing pages pinned by get_user_pages().
1091 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1092 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1093 * range after a fork.
1094 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1095 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1096 * were corrupted by unrecoverable hardware memory failure.
1097 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1098 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1099 * this area with pages of identical content from other such areas.
1100 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1101 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1102 * huge pages in the future. Existing pages might be coalesced and
1103 * new pages might be allocated as THP.
1104 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1105 * transparent huge pages so the existing pages will not be
1106 * coalesced into THP and new pages will not be allocated as THP.
1107 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1108 * from being included in its core dump.
1109 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1110 * MADV_COLD - the application is not expected to use this memory soon,
1111 * deactivate pages in this range so that they can be reclaimed
1112 * easily if memory pressure happens.
1113 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1114 * page out the pages in this range immediately.
1115 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1116 * triggering read faults if required
1117 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1118 * triggering write faults if required
1122 * -EINVAL - start + len < 0, start is not page-aligned,
1123 * "behavior" is not a valid value, or application
1124 * is attempting to release locked or shared pages,
1125 * or the specified address range includes file, Huge TLB,
1126 * MAP_SHARED or VMPFNMAP range.
1127 * -ENOMEM - addresses in the specified range are not currently
1128 * mapped, or are outside the AS of the process.
1129 * -EIO - an I/O error occurred while paging in data.
1130 * -EBADF - map exists, but area maps something that isn't a file.
1131 * -EAGAIN - a kernel resource was temporarily unavailable.
1133 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1135 unsigned long end, tmp;
1136 struct vm_area_struct *vma, *prev;
1137 int unmapped_error = 0;
1138 int error = -EINVAL;
1141 struct blk_plug plug;
1143 start = untagged_addr(start);
1145 if (!madvise_behavior_valid(behavior))
1148 if (!PAGE_ALIGNED(start))
1150 len = PAGE_ALIGN(len_in);
1152 /* Check to see whether len was rounded up from small -ve to zero */
1164 #ifdef CONFIG_MEMORY_FAILURE
1165 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1166 return madvise_inject_error(behavior, start, start + len_in);
1169 write = madvise_need_mmap_write(behavior);
1171 if (mmap_write_lock_killable(mm))
1178 * If the interval [start,end) covers some unmapped address
1179 * ranges, just ignore them, but return -ENOMEM at the end.
1180 * - different from the way of handling in mlock etc.
1182 vma = find_vma_prev(mm, start, &prev);
1183 if (vma && start > vma->vm_start)
1186 blk_start_plug(&plug);
1188 /* Still start < end. */
1193 /* Here start < (end|vma->vm_end). */
1194 if (start < vma->vm_start) {
1195 unmapped_error = -ENOMEM;
1196 start = vma->vm_start;
1201 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1206 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1207 error = madvise_vma(vma, &prev, start, tmp, behavior);
1211 if (prev && start < prev->vm_end)
1212 start = prev->vm_end;
1213 error = unmapped_error;
1217 vma = prev->vm_next;
1218 else /* madvise_remove dropped mmap_lock */
1219 vma = find_vma(mm, start);
1222 blk_finish_plug(&plug);
1224 mmap_write_unlock(mm);
1226 mmap_read_unlock(mm);
1231 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1233 return do_madvise(current->mm, start, len_in, behavior);
1236 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1237 size_t, vlen, int, behavior, unsigned int, flags)
1240 struct iovec iovstack[UIO_FASTIOV], iovec;
1241 struct iovec *iov = iovstack;
1242 struct iov_iter iter;
1244 struct task_struct *task;
1245 struct mm_struct *mm;
1247 unsigned int f_flags;
1254 ret = import_iovec(READ, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1258 pid = pidfd_get_pid(pidfd, &f_flags);
1264 task = get_pid_task(pid, PIDTYPE_PID);
1270 if (!process_madvise_behavior_valid(behavior)) {
1275 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1276 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1277 if (IS_ERR_OR_NULL(mm)) {
1278 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1283 * Require CAP_SYS_NICE for influencing process performance. Note that
1284 * only non-destructive hints are currently supported.
1286 if (!capable(CAP_SYS_NICE)) {
1291 total_len = iov_iter_count(&iter);
1293 while (iov_iter_count(&iter)) {
1294 iovec = iov_iter_iovec(&iter);
1295 ret = do_madvise(mm, (unsigned long)iovec.iov_base,
1296 iovec.iov_len, behavior);
1299 iov_iter_advance(&iter, iovec.iov_len);
1302 ret = (total_len - iov_iter_count(&iter)) ? : ret;
1307 put_task_struct(task);