4 * Copyright (C) 1999 Linus Torvalds
5 * Copyright (C) 2002 Christoph Hellwig
8 #include <linux/mman.h>
9 #include <linux/pagemap.h>
10 #include <linux/syscalls.h>
11 #include <linux/mempolicy.h>
12 #include <linux/page-isolation.h>
13 #include <linux/userfaultfd_k.h>
14 #include <linux/hugetlb.h>
15 #include <linux/falloc.h>
16 #include <linux/sched.h>
17 #include <linux/ksm.h>
19 #include <linux/file.h>
20 #include <linux/blkdev.h>
21 #include <linux/backing-dev.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/shmem_fs.h>
25 #include <linux/mmu_notifier.h>
32 * Any behaviour which results in changes to the vma->vm_flags needs to
33 * take mmap_sem for writing. Others, which simply traverse vmas, need
34 * to only take it for reading.
36 static int madvise_need_mmap_write(int behavior)
45 /* be safe, default to 1. list exceptions explicitly */
51 * We can potentially split a vm area into separate
52 * areas, each area with its own behavior.
54 static long madvise_behavior(struct vm_area_struct *vma,
55 struct vm_area_struct **prev,
56 unsigned long start, unsigned long end, int behavior)
58 struct mm_struct *mm = vma->vm_mm;
61 unsigned long new_flags = vma->vm_flags;
65 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
68 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
71 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
74 new_flags |= VM_DONTCOPY;
77 if (vma->vm_flags & VM_IO) {
81 new_flags &= ~VM_DONTCOPY;
84 new_flags |= VM_DONTDUMP;
87 if (new_flags & VM_SPECIAL) {
91 new_flags &= ~VM_DONTDUMP;
94 case MADV_UNMERGEABLE:
95 error = ksm_madvise(vma, start, end, behavior, &new_flags);
98 * madvise() returns EAGAIN if kernel resources, such as
99 * slab, are temporarily unavailable.
101 if (error == -ENOMEM)
107 case MADV_NOHUGEPAGE:
108 error = hugepage_madvise(vma, &new_flags, behavior);
111 * madvise() returns EAGAIN if kernel resources, such as
112 * slab, are temporarily unavailable.
114 if (error == -ENOMEM)
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);
145 * madvise() returns EAGAIN if kernel resources, such as
146 * slab, are temporarily unavailable.
148 if (error == -ENOMEM)
154 if (end != vma->vm_end) {
155 if (unlikely(mm->map_count >= sysctl_max_map_count)) {
159 error = __split_vma(mm, vma, end, 0);
162 * madvise() returns EAGAIN if kernel resources, such as
163 * slab, are temporarily unavailable.
165 if (error == -ENOMEM)
173 * vm_flags is protected by the mmap_sem held in write mode.
175 vma->vm_flags = new_flags;
181 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
182 unsigned long end, struct mm_walk *walk)
185 struct vm_area_struct *vma = walk->private;
188 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
191 for (index = start; index != end; index += PAGE_SIZE) {
197 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
198 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
199 pte_unmap_unlock(orig_pte, ptl);
201 if (pte_present(pte) || pte_none(pte))
203 entry = pte_to_swp_entry(pte);
204 if (unlikely(non_swap_entry(entry)))
207 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
216 static void force_swapin_readahead(struct vm_area_struct *vma,
217 unsigned long start, unsigned long end)
219 struct mm_walk walk = {
221 .pmd_entry = swapin_walk_pmd_entry,
225 walk_page_range(start, end, &walk);
227 lru_add_drain(); /* Push any new pages onto the LRU now */
230 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
231 unsigned long start, unsigned long end,
232 struct address_space *mapping)
238 for (; start < end; start += PAGE_SIZE) {
239 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
241 page = find_get_entry(mapping, index);
242 if (!radix_tree_exceptional_entry(page)) {
247 swap = radix_to_swp_entry(page);
248 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
254 lru_add_drain(); /* Push any new pages onto the LRU now */
256 #endif /* CONFIG_SWAP */
259 * Schedule all required I/O operations. Do not wait for completion.
261 static long madvise_willneed(struct vm_area_struct *vma,
262 struct vm_area_struct **prev,
263 unsigned long start, unsigned long end)
265 struct file *file = vma->vm_file;
270 force_swapin_readahead(vma, start, end);
274 if (shmem_mapping(file->f_mapping)) {
276 force_shm_swapin_readahead(vma, start, end,
285 if (IS_DAX(file_inode(file))) {
286 /* no bad return value, but ignore advice */
291 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
292 if (end > vma->vm_end)
294 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
296 force_page_cache_readahead(file->f_mapping, file, start, end - start);
300 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
301 unsigned long end, struct mm_walk *walk)
304 struct mmu_gather *tlb = walk->private;
305 struct mm_struct *mm = tlb->mm;
306 struct vm_area_struct *vma = walk->vma;
308 pte_t *orig_pte, *pte, ptent;
313 next = pmd_addr_end(addr, end);
314 if (pmd_trans_huge(*pmd))
315 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
318 if (pmd_trans_unstable(pmd))
321 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
322 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
323 arch_enter_lazy_mmu_mode();
324 for (; addr != end; pte++, addr += PAGE_SIZE) {
330 * If the pte has swp_entry, just clear page table to
331 * prevent swap-in which is more expensive rather than
332 * (page allocation + zeroing).
334 if (!pte_present(ptent)) {
337 entry = pte_to_swp_entry(ptent);
338 if (non_swap_entry(entry))
341 free_swap_and_cache(entry);
342 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
346 page = vm_normal_page(vma, addr, ptent);
351 * If pmd isn't transhuge but the page is THP and
352 * is owned by only this process, split it and
353 * deactivate all pages.
355 if (PageTransCompound(page)) {
356 if (page_mapcount(page) != 1)
359 if (!trylock_page(page)) {
363 pte_unmap_unlock(orig_pte, ptl);
364 if (split_huge_page(page)) {
367 pte_offset_map_lock(mm, pmd, addr, &ptl);
372 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
378 VM_BUG_ON_PAGE(PageTransCompound(page), page);
380 if (PageSwapCache(page) || PageDirty(page)) {
381 if (!trylock_page(page))
384 * If page is shared with others, we couldn't clear
385 * PG_dirty of the page.
387 if (page_mapcount(page) != 1) {
392 if (PageSwapCache(page) && !try_to_free_swap(page)) {
397 ClearPageDirty(page);
401 if (pte_young(ptent) || pte_dirty(ptent)) {
403 * Some of architecture(ex, PPC) don't update TLB
404 * with set_pte_at and tlb_remove_tlb_entry so for
405 * the portability, remap the pte with old|clean
406 * after pte clearing.
408 ptent = ptep_get_and_clear_full(mm, addr, pte,
411 ptent = pte_mkold(ptent);
412 ptent = pte_mkclean(ptent);
413 set_pte_at(mm, addr, pte, ptent);
414 tlb_remove_tlb_entry(tlb, pte, addr);
416 mark_page_lazyfree(page);
420 if (current->mm == mm)
423 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
425 arch_leave_lazy_mmu_mode();
426 pte_unmap_unlock(orig_pte, ptl);
432 static void madvise_free_page_range(struct mmu_gather *tlb,
433 struct vm_area_struct *vma,
434 unsigned long addr, unsigned long end)
436 struct mm_walk free_walk = {
437 .pmd_entry = madvise_free_pte_range,
442 tlb_start_vma(tlb, vma);
443 walk_page_range(addr, end, &free_walk);
444 tlb_end_vma(tlb, vma);
447 static int madvise_free_single_vma(struct vm_area_struct *vma,
448 unsigned long start_addr, unsigned long end_addr)
450 unsigned long start, end;
451 struct mm_struct *mm = vma->vm_mm;
452 struct mmu_gather tlb;
454 /* MADV_FREE works for only anon vma at the moment */
455 if (!vma_is_anonymous(vma))
458 start = max(vma->vm_start, start_addr);
459 if (start >= vma->vm_end)
461 end = min(vma->vm_end, end_addr);
462 if (end <= vma->vm_start)
466 tlb_gather_mmu(&tlb, mm, start, end);
467 update_hiwater_rss(mm);
469 mmu_notifier_invalidate_range_start(mm, start, end);
470 madvise_free_page_range(&tlb, vma, start, end);
471 mmu_notifier_invalidate_range_end(mm, start, end);
472 tlb_finish_mmu(&tlb, start, end);
478 * Application no longer needs these pages. If the pages are dirty,
479 * it's OK to just throw them away. The app will be more careful about
480 * data it wants to keep. Be sure to free swap resources too. The
481 * zap_page_range call sets things up for shrink_active_list to actually free
482 * these pages later if no one else has touched them in the meantime,
483 * although we could add these pages to a global reuse list for
484 * shrink_active_list to pick up before reclaiming other pages.
486 * NB: This interface discards data rather than pushes it out to swap,
487 * as some implementations do. This has performance implications for
488 * applications like large transactional databases which want to discard
489 * pages in anonymous maps after committing to backing store the data
490 * that was kept in them. There is no reason to write this data out to
491 * the swap area if the application is discarding it.
493 * An interface that causes the system to free clean pages and flush
494 * dirty pages is already available as msync(MS_INVALIDATE).
496 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
497 unsigned long start, unsigned long end)
499 zap_page_range(vma, start, end - start);
503 static long madvise_dontneed_free(struct vm_area_struct *vma,
504 struct vm_area_struct **prev,
505 unsigned long start, unsigned long end,
509 if (!can_madv_dontneed_vma(vma))
512 if (!userfaultfd_remove(vma, start, end)) {
513 *prev = NULL; /* mmap_sem has been dropped, prev is stale */
515 down_read(¤t->mm->mmap_sem);
516 vma = find_vma(current->mm, start);
519 if (start < vma->vm_start) {
521 * This "vma" under revalidation is the one
522 * with the lowest vma->vm_start where start
523 * is also < vma->vm_end. If start <
524 * vma->vm_start it means an hole materialized
525 * in the user address space within the
526 * virtual range passed to MADV_DONTNEED
531 if (!can_madv_dontneed_vma(vma))
533 if (end > vma->vm_end) {
535 * Don't fail if end > vma->vm_end. If the old
536 * vma was splitted while the mmap_sem was
537 * released the effect of the concurrent
538 * operation may not cause madvise() to
539 * have an undefined result. There may be an
540 * adjacent next vma that we'll walk
541 * next. userfaultfd_remove() will generate an
542 * UFFD_EVENT_REMOVE repetition on the
543 * end-vma->vm_end range, but the manager can
544 * handle a repetition fine.
548 VM_WARN_ON(start >= end);
551 if (behavior == MADV_DONTNEED)
552 return madvise_dontneed_single_vma(vma, start, end);
553 else if (behavior == MADV_FREE)
554 return madvise_free_single_vma(vma, start, end);
560 * Application wants to free up the pages and associated backing store.
561 * This is effectively punching a hole into the middle of a file.
563 static long madvise_remove(struct vm_area_struct *vma,
564 struct vm_area_struct **prev,
565 unsigned long start, unsigned long end)
571 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
573 if (vma->vm_flags & VM_LOCKED)
578 if (!f || !f->f_mapping || !f->f_mapping->host) {
582 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
585 offset = (loff_t)(start - vma->vm_start)
586 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
589 * Filesystem's fallocate may need to take i_mutex. We need to
590 * explicitly grab a reference because the vma (and hence the
591 * vma's reference to the file) can go away as soon as we drop
595 if (userfaultfd_remove(vma, start, end)) {
596 /* mmap_sem was not released by userfaultfd_remove() */
597 up_read(¤t->mm->mmap_sem);
599 error = vfs_fallocate(f,
600 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
601 offset, end - start);
603 down_read(¤t->mm->mmap_sem);
607 #ifdef CONFIG_MEMORY_FAILURE
609 * Error injection support for memory error handling.
611 static int madvise_inject_error(int behavior,
612 unsigned long start, unsigned long end)
616 if (!capable(CAP_SYS_ADMIN))
619 for (; start < end; start += PAGE_SIZE <<
620 compound_order(compound_head(page))) {
623 ret = get_user_pages_fast(start, 1, 0, &page);
627 if (PageHWPoison(page)) {
632 if (behavior == MADV_SOFT_OFFLINE) {
633 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
634 page_to_pfn(page), start);
636 ret = soft_offline_page(page, MF_COUNT_INCREASED);
641 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
642 page_to_pfn(page), start);
644 ret = memory_failure(page_to_pfn(page), 0, MF_COUNT_INCREASED);
653 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
654 unsigned long start, unsigned long end, int behavior)
658 return madvise_remove(vma, prev, start, end);
660 return madvise_willneed(vma, prev, start, end);
663 return madvise_dontneed_free(vma, prev, start, end, behavior);
665 return madvise_behavior(vma, prev, start, end, behavior);
670 madvise_behavior_valid(int behavior)
676 case MADV_SEQUENTIAL:
684 case MADV_UNMERGEABLE:
686 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
688 case MADV_NOHUGEPAGE:
692 #ifdef CONFIG_MEMORY_FAILURE
693 case MADV_SOFT_OFFLINE:
704 * The madvise(2) system call.
706 * Applications can use madvise() to advise the kernel how it should
707 * handle paging I/O in this VM area. The idea is to help the kernel
708 * use appropriate read-ahead and caching techniques. The information
709 * provided is advisory only, and can be safely disregarded by the
710 * kernel without affecting the correct operation of the application.
713 * MADV_NORMAL - the default behavior is to read clusters. This
714 * results in some read-ahead and read-behind.
715 * MADV_RANDOM - the system should read the minimum amount of data
716 * on any access, since it is unlikely that the appli-
717 * cation will need more than what it asks for.
718 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
719 * once, so they can be aggressively read ahead, and
720 * can be freed soon after they are accessed.
721 * MADV_WILLNEED - the application is notifying the system to read
723 * MADV_DONTNEED - the application is finished with the given range,
724 * so the kernel can free resources associated with it.
725 * MADV_FREE - the application marks pages in the given range as lazy free,
726 * where actual purges are postponed until memory pressure happens.
727 * MADV_REMOVE - the application wants to free up the given range of
728 * pages and associated backing store.
729 * MADV_DONTFORK - omit this area from child's address space when forking:
730 * typically, to avoid COWing pages pinned by get_user_pages().
731 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
732 * MADV_HWPOISON - trigger memory error handler as if the given memory range
733 * were corrupted by unrecoverable hardware memory failure.
734 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
735 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
736 * this area with pages of identical content from other such areas.
737 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
738 * MADV_HUGEPAGE - the application wants to back the given range by transparent
739 * huge pages in the future. Existing pages might be coalesced and
740 * new pages might be allocated as THP.
741 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
742 * transparent huge pages so the existing pages will not be
743 * coalesced into THP and new pages will not be allocated as THP.
744 * MADV_DONTDUMP - the application wants to prevent pages in the given range
745 * from being included in its core dump.
746 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
750 * -EINVAL - start + len < 0, start is not page-aligned,
751 * "behavior" is not a valid value, or application
752 * is attempting to release locked or shared pages.
753 * -ENOMEM - addresses in the specified range are not currently
754 * mapped, or are outside the AS of the process.
755 * -EIO - an I/O error occurred while paging in data.
756 * -EBADF - map exists, but area maps something that isn't a file.
757 * -EAGAIN - a kernel resource was temporarily unavailable.
759 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
761 unsigned long end, tmp;
762 struct vm_area_struct *vma, *prev;
763 int unmapped_error = 0;
767 struct blk_plug plug;
769 if (!madvise_behavior_valid(behavior))
772 if (start & ~PAGE_MASK)
774 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
776 /* Check to see whether len was rounded up from small -ve to zero */
788 #ifdef CONFIG_MEMORY_FAILURE
789 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
790 return madvise_inject_error(behavior, start, start + len_in);
793 write = madvise_need_mmap_write(behavior);
795 if (down_write_killable(¤t->mm->mmap_sem))
798 down_read(¤t->mm->mmap_sem);
802 * If the interval [start,end) covers some unmapped address
803 * ranges, just ignore them, but return -ENOMEM at the end.
804 * - different from the way of handling in mlock etc.
806 vma = find_vma_prev(current->mm, start, &prev);
807 if (vma && start > vma->vm_start)
810 blk_start_plug(&plug);
812 /* Still start < end. */
817 /* Here start < (end|vma->vm_end). */
818 if (start < vma->vm_start) {
819 unmapped_error = -ENOMEM;
820 start = vma->vm_start;
825 /* Here vma->vm_start <= start < (end|vma->vm_end) */
830 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
831 error = madvise_vma(vma, &prev, start, tmp, behavior);
835 if (prev && start < prev->vm_end)
836 start = prev->vm_end;
837 error = unmapped_error;
842 else /* madvise_remove dropped mmap_sem */
843 vma = find_vma(current->mm, start);
846 blk_finish_plug(&plug);
848 up_write(¤t->mm->mmap_sem);
850 up_read(¤t->mm->mmap_sem);