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/mmu_notifier.h>
29 * Any behaviour which results in changes to the vma->vm_flags needs to
30 * take mmap_sem for writing. Others, which simply traverse vmas, need
31 * to only take it for reading.
33 static int madvise_need_mmap_write(int behavior)
42 /* be safe, default to 1. list exceptions explicitly */
48 * We can potentially split a vm area into separate
49 * areas, each area with its own behavior.
51 static long madvise_behavior(struct vm_area_struct *vma,
52 struct vm_area_struct **prev,
53 unsigned long start, unsigned long end, int behavior)
55 struct mm_struct *mm = vma->vm_mm;
58 unsigned long new_flags = vma->vm_flags;
62 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
65 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
68 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
71 new_flags |= VM_DONTCOPY;
74 if (vma->vm_flags & VM_IO) {
78 new_flags &= ~VM_DONTCOPY;
81 new_flags |= VM_DONTDUMP;
84 if (new_flags & VM_SPECIAL) {
88 new_flags &= ~VM_DONTDUMP;
91 case MADV_UNMERGEABLE:
92 error = ksm_madvise(vma, start, end, behavior, &new_flags);
98 error = hugepage_madvise(vma, &new_flags, behavior);
104 if (new_flags == vma->vm_flags) {
109 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
110 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
111 vma->vm_file, pgoff, vma_policy(vma),
112 vma->vm_userfaultfd_ctx);
120 if (start != vma->vm_start) {
121 error = split_vma(mm, vma, start, 1);
126 if (end != vma->vm_end) {
127 error = split_vma(mm, vma, end, 0);
134 * vm_flags is protected by the mmap_sem held in write mode.
136 vma->vm_flags = new_flags;
139 if (error == -ENOMEM)
145 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
146 unsigned long end, struct mm_walk *walk)
149 struct vm_area_struct *vma = walk->private;
152 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
155 for (index = start; index != end; index += PAGE_SIZE) {
161 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
162 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
163 pte_unmap_unlock(orig_pte, ptl);
165 if (pte_present(pte) || pte_none(pte))
167 entry = pte_to_swp_entry(pte);
168 if (unlikely(non_swap_entry(entry)))
171 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
180 static void force_swapin_readahead(struct vm_area_struct *vma,
181 unsigned long start, unsigned long end)
183 struct mm_walk walk = {
185 .pmd_entry = swapin_walk_pmd_entry,
189 walk_page_range(start, end, &walk);
191 lru_add_drain(); /* Push any new pages onto the LRU now */
194 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
195 unsigned long start, unsigned long end,
196 struct address_space *mapping)
202 for (; start < end; start += PAGE_SIZE) {
203 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
205 page = find_get_entry(mapping, index);
206 if (!radix_tree_exceptional_entry(page)) {
211 swap = radix_to_swp_entry(page);
212 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
218 lru_add_drain(); /* Push any new pages onto the LRU now */
220 #endif /* CONFIG_SWAP */
223 * Schedule all required I/O operations. Do not wait for completion.
225 static long madvise_willneed(struct vm_area_struct *vma,
226 struct vm_area_struct **prev,
227 unsigned long start, unsigned long end)
229 struct file *file = vma->vm_file;
234 force_swapin_readahead(vma, start, end);
238 if (shmem_mapping(file->f_mapping)) {
240 force_shm_swapin_readahead(vma, start, end,
249 if (IS_DAX(file_inode(file))) {
250 /* no bad return value, but ignore advice */
255 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
256 if (end > vma->vm_end)
258 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
260 force_page_cache_readahead(file->f_mapping, file, start, end - start);
264 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
265 unsigned long end, struct mm_walk *walk)
268 struct mmu_gather *tlb = walk->private;
269 struct mm_struct *mm = tlb->mm;
270 struct vm_area_struct *vma = walk->vma;
272 pte_t *orig_pte, *pte, ptent;
277 next = pmd_addr_end(addr, end);
278 if (pmd_trans_huge(*pmd))
279 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
282 if (pmd_trans_unstable(pmd))
285 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
286 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
287 arch_enter_lazy_mmu_mode();
288 for (; addr != end; pte++, addr += PAGE_SIZE) {
294 * If the pte has swp_entry, just clear page table to
295 * prevent swap-in which is more expensive rather than
296 * (page allocation + zeroing).
298 if (!pte_present(ptent)) {
301 entry = pte_to_swp_entry(ptent);
302 if (non_swap_entry(entry))
305 free_swap_and_cache(entry);
306 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
310 page = vm_normal_page(vma, addr, ptent);
315 * If pmd isn't transhuge but the page is THP and
316 * is owned by only this process, split it and
317 * deactivate all pages.
319 if (PageTransCompound(page)) {
320 if (page_mapcount(page) != 1)
323 if (!trylock_page(page)) {
327 pte_unmap_unlock(orig_pte, ptl);
328 if (split_huge_page(page)) {
331 pte_offset_map_lock(mm, pmd, addr, &ptl);
336 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
342 VM_BUG_ON_PAGE(PageTransCompound(page), page);
344 if (PageSwapCache(page) || PageDirty(page)) {
345 if (!trylock_page(page))
348 * If page is shared with others, we couldn't clear
349 * PG_dirty of the page.
351 if (page_mapcount(page) != 1) {
356 if (PageSwapCache(page) && !try_to_free_swap(page)) {
361 ClearPageDirty(page);
365 if (pte_young(ptent) || pte_dirty(ptent)) {
367 * Some of architecture(ex, PPC) don't update TLB
368 * with set_pte_at and tlb_remove_tlb_entry so for
369 * the portability, remap the pte with old|clean
370 * after pte clearing.
372 ptent = ptep_get_and_clear_full(mm, addr, pte,
375 ptent = pte_mkold(ptent);
376 ptent = pte_mkclean(ptent);
377 set_pte_at(mm, addr, pte, ptent);
378 if (PageActive(page))
379 deactivate_page(page);
380 tlb_remove_tlb_entry(tlb, pte, addr);
385 if (current->mm == mm)
388 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
390 arch_leave_lazy_mmu_mode();
391 pte_unmap_unlock(orig_pte, ptl);
397 static void madvise_free_page_range(struct mmu_gather *tlb,
398 struct vm_area_struct *vma,
399 unsigned long addr, unsigned long end)
401 struct mm_walk free_walk = {
402 .pmd_entry = madvise_free_pte_range,
407 tlb_start_vma(tlb, vma);
408 walk_page_range(addr, end, &free_walk);
409 tlb_end_vma(tlb, vma);
412 static int madvise_free_single_vma(struct vm_area_struct *vma,
413 unsigned long start_addr, unsigned long end_addr)
415 unsigned long start, end;
416 struct mm_struct *mm = vma->vm_mm;
417 struct mmu_gather tlb;
419 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
422 /* MADV_FREE works for only anon vma at the moment */
423 if (!vma_is_anonymous(vma))
426 start = max(vma->vm_start, start_addr);
427 if (start >= vma->vm_end)
429 end = min(vma->vm_end, end_addr);
430 if (end <= vma->vm_start)
434 tlb_gather_mmu(&tlb, mm, start, end);
435 update_hiwater_rss(mm);
437 mmu_notifier_invalidate_range_start(mm, start, end);
438 madvise_free_page_range(&tlb, vma, start, end);
439 mmu_notifier_invalidate_range_end(mm, start, end);
440 tlb_finish_mmu(&tlb, start, end);
445 static long madvise_free(struct vm_area_struct *vma,
446 struct vm_area_struct **prev,
447 unsigned long start, unsigned long end)
450 return madvise_free_single_vma(vma, start, end);
454 * Application no longer needs these pages. If the pages are dirty,
455 * it's OK to just throw them away. The app will be more careful about
456 * data it wants to keep. Be sure to free swap resources too. The
457 * zap_page_range call sets things up for shrink_active_list to actually free
458 * these pages later if no one else has touched them in the meantime,
459 * although we could add these pages to a global reuse list for
460 * shrink_active_list to pick up before reclaiming other pages.
462 * NB: This interface discards data rather than pushes it out to swap,
463 * as some implementations do. This has performance implications for
464 * applications like large transactional databases which want to discard
465 * pages in anonymous maps after committing to backing store the data
466 * that was kept in them. There is no reason to write this data out to
467 * the swap area if the application is discarding it.
469 * An interface that causes the system to free clean pages and flush
470 * dirty pages is already available as msync(MS_INVALIDATE).
472 static long madvise_dontneed(struct vm_area_struct *vma,
473 struct vm_area_struct **prev,
474 unsigned long start, unsigned long end)
477 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
480 madvise_userfault_dontneed(vma, prev, start, end);
481 zap_page_range(vma, start, end - start, NULL);
486 * Application wants to free up the pages and associated backing store.
487 * This is effectively punching a hole into the middle of a file.
489 static long madvise_remove(struct vm_area_struct *vma,
490 struct vm_area_struct **prev,
491 unsigned long start, unsigned long end)
497 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
499 if (vma->vm_flags & VM_LOCKED)
504 if (!f || !f->f_mapping || !f->f_mapping->host) {
508 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
511 offset = (loff_t)(start - vma->vm_start)
512 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
515 * Filesystem's fallocate may need to take i_mutex. We need to
516 * explicitly grab a reference because the vma (and hence the
517 * vma's reference to the file) can go away as soon as we drop
521 up_read(¤t->mm->mmap_sem);
522 error = vfs_fallocate(f,
523 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
524 offset, end - start);
526 down_read(¤t->mm->mmap_sem);
530 #ifdef CONFIG_MEMORY_FAILURE
532 * Error injection support for memory error handling.
534 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
537 if (!capable(CAP_SYS_ADMIN))
539 for (; start < end; start += PAGE_SIZE <<
540 compound_order(compound_head(p))) {
543 ret = get_user_pages_fast(start, 1, 0, &p);
547 if (PageHWPoison(p)) {
551 if (bhv == MADV_SOFT_OFFLINE) {
552 pr_info("Soft offlining page %#lx at %#lx\n",
553 page_to_pfn(p), start);
554 ret = soft_offline_page(p, MF_COUNT_INCREASED);
559 pr_info("Injecting memory failure for page %#lx at %#lx\n",
560 page_to_pfn(p), start);
561 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
570 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
571 unsigned long start, unsigned long end, int behavior)
575 return madvise_remove(vma, prev, start, end);
577 return madvise_willneed(vma, prev, start, end);
580 * XXX: In this implementation, MADV_FREE works like
581 * MADV_DONTNEED on swapless system or full swap.
583 if (get_nr_swap_pages() > 0)
584 return madvise_free(vma, prev, start, end);
587 return madvise_dontneed(vma, prev, start, end);
589 return madvise_behavior(vma, prev, start, end, behavior);
594 madvise_behavior_valid(int behavior)
600 case MADV_SEQUENTIAL:
608 case MADV_UNMERGEABLE:
610 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
612 case MADV_NOHUGEPAGE:
624 * The madvise(2) system call.
626 * Applications can use madvise() to advise the kernel how it should
627 * handle paging I/O in this VM area. The idea is to help the kernel
628 * use appropriate read-ahead and caching techniques. The information
629 * provided is advisory only, and can be safely disregarded by the
630 * kernel without affecting the correct operation of the application.
633 * MADV_NORMAL - the default behavior is to read clusters. This
634 * results in some read-ahead and read-behind.
635 * MADV_RANDOM - the system should read the minimum amount of data
636 * on any access, since it is unlikely that the appli-
637 * cation will need more than what it asks for.
638 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
639 * once, so they can be aggressively read ahead, and
640 * can be freed soon after they are accessed.
641 * MADV_WILLNEED - the application is notifying the system to read
643 * MADV_DONTNEED - the application is finished with the given range,
644 * so the kernel can free resources associated with it.
645 * MADV_FREE - the application marks pages in the given range as lazy free,
646 * where actual purges are postponed until memory pressure happens.
647 * MADV_REMOVE - the application wants to free up the given range of
648 * pages and associated backing store.
649 * MADV_DONTFORK - omit this area from child's address space when forking:
650 * typically, to avoid COWing pages pinned by get_user_pages().
651 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
652 * MADV_HWPOISON - trigger memory error handler as if the given memory range
653 * were corrupted by unrecoverable hardware memory failure.
654 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
655 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
656 * this area with pages of identical content from other such areas.
657 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
658 * MADV_HUGEPAGE - the application wants to back the given range by transparent
659 * huge pages in the future. Existing pages might be coalesced and
660 * new pages might be allocated as THP.
661 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
662 * transparent huge pages so the existing pages will not be
663 * coalesced into THP and new pages will not be allocated as THP.
664 * MADV_DONTDUMP - the application wants to prevent pages in the given range
665 * from being included in its core dump.
666 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
670 * -EINVAL - start + len < 0, start is not page-aligned,
671 * "behavior" is not a valid value, or application
672 * is attempting to release locked or shared pages.
673 * -ENOMEM - addresses in the specified range are not currently
674 * mapped, or are outside the AS of the process.
675 * -EIO - an I/O error occurred while paging in data.
676 * -EBADF - map exists, but area maps something that isn't a file.
677 * -EAGAIN - a kernel resource was temporarily unavailable.
679 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
681 unsigned long end, tmp;
682 struct vm_area_struct *vma, *prev;
683 int unmapped_error = 0;
687 struct blk_plug plug;
689 #ifdef CONFIG_MEMORY_FAILURE
690 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
691 return madvise_hwpoison(behavior, start, start+len_in);
693 if (!madvise_behavior_valid(behavior))
696 if (start & ~PAGE_MASK)
698 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
700 /* Check to see whether len was rounded up from small -ve to zero */
712 write = madvise_need_mmap_write(behavior);
714 if (down_write_killable(¤t->mm->mmap_sem))
717 down_read(¤t->mm->mmap_sem);
721 * If the interval [start,end) covers some unmapped address
722 * ranges, just ignore them, but return -ENOMEM at the end.
723 * - different from the way of handling in mlock etc.
725 vma = find_vma_prev(current->mm, start, &prev);
726 if (vma && start > vma->vm_start)
729 blk_start_plug(&plug);
731 /* Still start < end. */
736 /* Here start < (end|vma->vm_end). */
737 if (start < vma->vm_start) {
738 unmapped_error = -ENOMEM;
739 start = vma->vm_start;
744 /* Here vma->vm_start <= start < (end|vma->vm_end) */
749 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
750 error = madvise_vma(vma, &prev, start, tmp, behavior);
754 if (prev && start < prev->vm_end)
755 start = prev->vm_end;
756 error = unmapped_error;
761 else /* madvise_remove dropped mmap_sem */
762 vma = find_vma(current->mm, start);
765 blk_finish_plug(&plug);
767 up_write(¤t->mm->mmap_sem);
769 up_read(¤t->mm->mmap_sem);