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/hugetlb.h>
14 #include <linux/falloc.h>
15 #include <linux/sched.h>
16 #include <linux/ksm.h>
18 #include <linux/file.h>
19 #include <linux/blkdev.h>
20 #include <linux/backing-dev.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 #include <linux/mmu_notifier.h>
28 * Any behaviour which results in changes to the vma->vm_flags needs to
29 * take mmap_sem for writing. Others, which simply traverse vmas, need
30 * to only take it for reading.
32 static int madvise_need_mmap_write(int behavior)
41 /* be safe, default to 1. list exceptions explicitly */
47 * We can potentially split a vm area into separate
48 * areas, each area with its own behavior.
50 static long madvise_behavior(struct vm_area_struct *vma,
51 struct vm_area_struct **prev,
52 unsigned long start, unsigned long end, int behavior)
54 struct mm_struct *mm = vma->vm_mm;
57 unsigned long new_flags = vma->vm_flags;
61 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
64 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
67 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
70 new_flags |= VM_DONTCOPY;
73 if (vma->vm_flags & VM_IO) {
77 new_flags &= ~VM_DONTCOPY;
80 new_flags |= VM_DONTDUMP;
83 if (new_flags & VM_SPECIAL) {
87 new_flags &= ~VM_DONTDUMP;
90 case MADV_UNMERGEABLE:
91 error = ksm_madvise(vma, start, end, behavior, &new_flags);
97 error = hugepage_madvise(vma, &new_flags, behavior);
103 if (new_flags == vma->vm_flags) {
108 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
109 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
110 vma->vm_file, pgoff, vma_policy(vma),
111 vma->vm_userfaultfd_ctx);
119 if (start != vma->vm_start) {
120 error = split_vma(mm, vma, start, 1);
125 if (end != vma->vm_end) {
126 error = split_vma(mm, vma, end, 0);
133 * vm_flags is protected by the mmap_sem held in write mode.
135 vma->vm_flags = new_flags;
138 if (error == -ENOMEM)
144 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
145 unsigned long end, struct mm_walk *walk)
148 struct vm_area_struct *vma = walk->private;
151 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
154 for (index = start; index != end; index += PAGE_SIZE) {
160 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
161 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
162 pte_unmap_unlock(orig_pte, ptl);
164 if (pte_present(pte) || pte_none(pte))
166 entry = pte_to_swp_entry(pte);
167 if (unlikely(non_swap_entry(entry)))
170 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
173 page_cache_release(page);
179 static void force_swapin_readahead(struct vm_area_struct *vma,
180 unsigned long start, unsigned long end)
182 struct mm_walk walk = {
184 .pmd_entry = swapin_walk_pmd_entry,
188 walk_page_range(start, end, &walk);
190 lru_add_drain(); /* Push any new pages onto the LRU now */
193 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
194 unsigned long start, unsigned long end,
195 struct address_space *mapping)
201 for (; start < end; start += PAGE_SIZE) {
202 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
204 page = find_get_entry(mapping, index);
205 if (!radix_tree_exceptional_entry(page)) {
207 page_cache_release(page);
210 swap = radix_to_swp_entry(page);
211 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
214 page_cache_release(page);
217 lru_add_drain(); /* Push any new pages onto the LRU now */
219 #endif /* CONFIG_SWAP */
222 * Schedule all required I/O operations. Do not wait for completion.
224 static long madvise_willneed(struct vm_area_struct *vma,
225 struct vm_area_struct **prev,
226 unsigned long start, unsigned long end)
228 struct file *file = vma->vm_file;
233 force_swapin_readahead(vma, start, end);
237 if (shmem_mapping(file->f_mapping)) {
239 force_shm_swapin_readahead(vma, start, end,
248 if (IS_DAX(file_inode(file))) {
249 /* no bad return value, but ignore advice */
254 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
255 if (end > vma->vm_end)
257 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
259 force_page_cache_readahead(file->f_mapping, file, start, end - start);
263 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
264 unsigned long end, struct mm_walk *walk)
267 struct mmu_gather *tlb = walk->private;
268 struct mm_struct *mm = tlb->mm;
269 struct vm_area_struct *vma = walk->vma;
277 next = pmd_addr_end(addr, end);
278 if (pmd_trans_huge(*pmd)) {
279 if (next - addr != HPAGE_PMD_SIZE)
280 split_huge_pmd(vma, pmd, addr);
281 else if (!madvise_free_huge_pmd(tlb, vma, pmd, addr))
286 if (pmd_trans_unstable(pmd))
289 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
290 arch_enter_lazy_mmu_mode();
291 for (; addr != end; pte++, addr += PAGE_SIZE) {
297 * If the pte has swp_entry, just clear page table to
298 * prevent swap-in which is more expensive rather than
299 * (page allocation + zeroing).
301 if (!pte_present(ptent)) {
302 entry = pte_to_swp_entry(ptent);
303 if (non_swap_entry(entry))
306 free_swap_and_cache(entry);
307 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
311 page = vm_normal_page(vma, addr, ptent);
315 if (PageSwapCache(page) || PageDirty(page)) {
316 if (!trylock_page(page))
319 * If page is shared with others, we couldn't clear
320 * PG_dirty of the page.
322 if (page_count(page) != 1 + !!PageSwapCache(page)) {
327 if (PageSwapCache(page) && !try_to_free_swap(page)) {
332 ClearPageDirty(page);
337 * Some of architecture(ex, PPC) don't update TLB
338 * with set_pte_at and tlb_remove_tlb_entry so for
339 * the portability, remap the pte with old|clean
340 * after pte clearing.
342 ptent = ptep_get_and_clear_full(mm, addr, pte,
344 ptent = pte_mkold(ptent);
345 ptent = pte_mkclean(ptent);
346 set_pte_at(mm, addr, pte, ptent);
347 if (PageActive(page))
348 deactivate_page(page);
349 tlb_remove_tlb_entry(tlb, pte, addr);
353 if (current->mm == mm)
356 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
359 arch_leave_lazy_mmu_mode();
360 pte_unmap_unlock(pte - 1, ptl);
366 static void madvise_free_page_range(struct mmu_gather *tlb,
367 struct vm_area_struct *vma,
368 unsigned long addr, unsigned long end)
370 struct mm_walk free_walk = {
371 .pmd_entry = madvise_free_pte_range,
377 tlb_start_vma(tlb, vma);
378 walk_page_range(addr, end, &free_walk);
379 tlb_end_vma(tlb, vma);
382 static int madvise_free_single_vma(struct vm_area_struct *vma,
383 unsigned long start_addr, unsigned long end_addr)
385 unsigned long start, end;
386 struct mm_struct *mm = vma->vm_mm;
387 struct mmu_gather tlb;
389 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
392 /* MADV_FREE works for only anon vma at the moment */
393 if (!vma_is_anonymous(vma))
396 start = max(vma->vm_start, start_addr);
397 if (start >= vma->vm_end)
399 end = min(vma->vm_end, end_addr);
400 if (end <= vma->vm_start)
404 tlb_gather_mmu(&tlb, mm, start, end);
405 update_hiwater_rss(mm);
407 mmu_notifier_invalidate_range_start(mm, start, end);
408 madvise_free_page_range(&tlb, vma, start, end);
409 mmu_notifier_invalidate_range_end(mm, start, end);
410 tlb_finish_mmu(&tlb, start, end);
415 static long madvise_free(struct vm_area_struct *vma,
416 struct vm_area_struct **prev,
417 unsigned long start, unsigned long end)
420 return madvise_free_single_vma(vma, start, end);
424 * Application no longer needs these pages. If the pages are dirty,
425 * it's OK to just throw them away. The app will be more careful about
426 * data it wants to keep. Be sure to free swap resources too. The
427 * zap_page_range call sets things up for shrink_active_list to actually free
428 * these pages later if no one else has touched them in the meantime,
429 * although we could add these pages to a global reuse list for
430 * shrink_active_list to pick up before reclaiming other pages.
432 * NB: This interface discards data rather than pushes it out to swap,
433 * as some implementations do. This has performance implications for
434 * applications like large transactional databases which want to discard
435 * pages in anonymous maps after committing to backing store the data
436 * that was kept in them. There is no reason to write this data out to
437 * the swap area if the application is discarding it.
439 * An interface that causes the system to free clean pages and flush
440 * dirty pages is already available as msync(MS_INVALIDATE).
442 static long madvise_dontneed(struct vm_area_struct *vma,
443 struct vm_area_struct **prev,
444 unsigned long start, unsigned long end)
447 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
450 zap_page_range(vma, start, end - start, NULL);
455 * Application wants to free up the pages and associated backing store.
456 * This is effectively punching a hole into the middle of a file.
458 static long madvise_remove(struct vm_area_struct *vma,
459 struct vm_area_struct **prev,
460 unsigned long start, unsigned long end)
466 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
468 if (vma->vm_flags & VM_LOCKED)
473 if (!f || !f->f_mapping || !f->f_mapping->host) {
477 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
480 offset = (loff_t)(start - vma->vm_start)
481 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
484 * Filesystem's fallocate may need to take i_mutex. We need to
485 * explicitly grab a reference because the vma (and hence the
486 * vma's reference to the file) can go away as soon as we drop
490 up_read(¤t->mm->mmap_sem);
491 error = vfs_fallocate(f,
492 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
493 offset, end - start);
495 down_read(¤t->mm->mmap_sem);
499 #ifdef CONFIG_MEMORY_FAILURE
501 * Error injection support for memory error handling.
503 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
506 if (!capable(CAP_SYS_ADMIN))
508 for (; start < end; start += PAGE_SIZE <<
509 compound_order(compound_head(p))) {
512 ret = get_user_pages_fast(start, 1, 0, &p);
516 if (PageHWPoison(p)) {
520 if (bhv == MADV_SOFT_OFFLINE) {
521 pr_info("Soft offlining page %#lx at %#lx\n",
522 page_to_pfn(p), start);
523 ret = soft_offline_page(p, MF_COUNT_INCREASED);
528 pr_info("Injecting memory failure for page %#lx at %#lx\n",
529 page_to_pfn(p), start);
530 /* Ignore return value for now */
531 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
538 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
539 unsigned long start, unsigned long end, int behavior)
543 return madvise_remove(vma, prev, start, end);
545 return madvise_willneed(vma, prev, start, end);
548 * XXX: In this implementation, MADV_FREE works like
549 * MADV_DONTNEED on swapless system or full swap.
551 if (get_nr_swap_pages() > 0)
552 return madvise_free(vma, prev, start, end);
555 return madvise_dontneed(vma, prev, start, end);
557 return madvise_behavior(vma, prev, start, end, behavior);
562 madvise_behavior_valid(int behavior)
568 case MADV_SEQUENTIAL:
576 case MADV_UNMERGEABLE:
578 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
580 case MADV_NOHUGEPAGE:
592 * The madvise(2) system call.
594 * Applications can use madvise() to advise the kernel how it should
595 * handle paging I/O in this VM area. The idea is to help the kernel
596 * use appropriate read-ahead and caching techniques. The information
597 * provided is advisory only, and can be safely disregarded by the
598 * kernel without affecting the correct operation of the application.
601 * MADV_NORMAL - the default behavior is to read clusters. This
602 * results in some read-ahead and read-behind.
603 * MADV_RANDOM - the system should read the minimum amount of data
604 * on any access, since it is unlikely that the appli-
605 * cation will need more than what it asks for.
606 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
607 * once, so they can be aggressively read ahead, and
608 * can be freed soon after they are accessed.
609 * MADV_WILLNEED - the application is notifying the system to read
611 * MADV_DONTNEED - the application is finished with the given range,
612 * so the kernel can free resources associated with it.
613 * MADV_REMOVE - the application wants to free up the given range of
614 * pages and associated backing store.
615 * MADV_DONTFORK - omit this area from child's address space when forking:
616 * typically, to avoid COWing pages pinned by get_user_pages().
617 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
618 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
619 * this area with pages of identical content from other such areas.
620 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
624 * -EINVAL - start + len < 0, start is not page-aligned,
625 * "behavior" is not a valid value, or application
626 * is attempting to release locked or shared pages.
627 * -ENOMEM - addresses in the specified range are not currently
628 * mapped, or are outside the AS of the process.
629 * -EIO - an I/O error occurred while paging in data.
630 * -EBADF - map exists, but area maps something that isn't a file.
631 * -EAGAIN - a kernel resource was temporarily unavailable.
633 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
635 unsigned long end, tmp;
636 struct vm_area_struct *vma, *prev;
637 int unmapped_error = 0;
641 struct blk_plug plug;
643 #ifdef CONFIG_MEMORY_FAILURE
644 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
645 return madvise_hwpoison(behavior, start, start+len_in);
647 if (!madvise_behavior_valid(behavior))
650 if (start & ~PAGE_MASK)
652 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
654 /* Check to see whether len was rounded up from small -ve to zero */
666 write = madvise_need_mmap_write(behavior);
668 down_write(¤t->mm->mmap_sem);
670 down_read(¤t->mm->mmap_sem);
673 * If the interval [start,end) covers some unmapped address
674 * ranges, just ignore them, but return -ENOMEM at the end.
675 * - different from the way of handling in mlock etc.
677 vma = find_vma_prev(current->mm, start, &prev);
678 if (vma && start > vma->vm_start)
681 blk_start_plug(&plug);
683 /* Still start < end. */
688 /* Here start < (end|vma->vm_end). */
689 if (start < vma->vm_start) {
690 unmapped_error = -ENOMEM;
691 start = vma->vm_start;
696 /* Here vma->vm_start <= start < (end|vma->vm_end) */
701 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
702 error = madvise_vma(vma, &prev, start, tmp, behavior);
706 if (prev && start < prev->vm_end)
707 start = prev->vm_end;
708 error = unmapped_error;
713 else /* madvise_remove dropped mmap_sem */
714 vma = find_vma(current->mm, start);
717 blk_finish_plug(&plug);
719 up_write(¤t->mm->mmap_sem);
721 up_read(¤t->mm->mmap_sem);