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>
27 struct madvise_free_private {
28 struct vm_area_struct *vma;
29 struct mmu_gather *tlb;
33 * Any behaviour which results in changes to the vma->vm_flags needs to
34 * take mmap_sem for writing. Others, which simply traverse vmas, need
35 * to only take it for reading.
37 static int madvise_need_mmap_write(int behavior)
46 /* be safe, default to 1. list exceptions explicitly */
52 * We can potentially split a vm area into separate
53 * areas, each area with its own behavior.
55 static long madvise_behavior(struct vm_area_struct *vma,
56 struct vm_area_struct **prev,
57 unsigned long start, unsigned long end, int behavior)
59 struct mm_struct *mm = vma->vm_mm;
62 unsigned long new_flags = vma->vm_flags;
66 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
69 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
72 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
75 new_flags |= VM_DONTCOPY;
78 if (vma->vm_flags & VM_IO) {
82 new_flags &= ~VM_DONTCOPY;
85 new_flags |= VM_DONTDUMP;
88 if (new_flags & VM_SPECIAL) {
92 new_flags &= ~VM_DONTDUMP;
95 case MADV_UNMERGEABLE:
96 error = ksm_madvise(vma, start, end, behavior, &new_flags);
101 case MADV_NOHUGEPAGE:
102 error = hugepage_madvise(vma, &new_flags, behavior);
108 if (new_flags == vma->vm_flags) {
113 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
114 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
115 vma->vm_file, pgoff, vma_policy(vma),
116 vma->vm_userfaultfd_ctx);
124 if (start != vma->vm_start) {
125 error = split_vma(mm, vma, start, 1);
130 if (end != vma->vm_end) {
131 error = split_vma(mm, vma, end, 0);
138 * vm_flags is protected by the mmap_sem held in write mode.
140 vma->vm_flags = new_flags;
143 if (error == -ENOMEM)
149 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
150 unsigned long end, struct mm_walk *walk)
153 struct vm_area_struct *vma = walk->private;
156 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
159 for (index = start; index != end; index += PAGE_SIZE) {
165 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
166 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
167 pte_unmap_unlock(orig_pte, ptl);
169 if (pte_present(pte) || pte_none(pte))
171 entry = pte_to_swp_entry(pte);
172 if (unlikely(non_swap_entry(entry)))
175 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
178 page_cache_release(page);
184 static void force_swapin_readahead(struct vm_area_struct *vma,
185 unsigned long start, unsigned long end)
187 struct mm_walk walk = {
189 .pmd_entry = swapin_walk_pmd_entry,
193 walk_page_range(start, end, &walk);
195 lru_add_drain(); /* Push any new pages onto the LRU now */
198 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
199 unsigned long start, unsigned long end,
200 struct address_space *mapping)
206 for (; start < end; start += PAGE_SIZE) {
207 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
209 page = find_get_entry(mapping, index);
210 if (!radix_tree_exceptional_entry(page)) {
212 page_cache_release(page);
215 swap = radix_to_swp_entry(page);
216 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
219 page_cache_release(page);
222 lru_add_drain(); /* Push any new pages onto the LRU now */
224 #endif /* CONFIG_SWAP */
227 * Schedule all required I/O operations. Do not wait for completion.
229 static long madvise_willneed(struct vm_area_struct *vma,
230 struct vm_area_struct **prev,
231 unsigned long start, unsigned long end)
233 struct file *file = vma->vm_file;
238 force_swapin_readahead(vma, start, end);
242 if (shmem_mapping(file->f_mapping)) {
244 force_shm_swapin_readahead(vma, start, end,
253 if (IS_DAX(file_inode(file))) {
254 /* no bad return value, but ignore advice */
259 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
260 if (end > vma->vm_end)
262 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
264 force_page_cache_readahead(file->f_mapping, file, start, end - start);
268 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
269 unsigned long end, struct mm_walk *walk)
272 struct madvise_free_private *fp = walk->private;
273 struct mmu_gather *tlb = fp->tlb;
274 struct mm_struct *mm = tlb->mm;
275 struct vm_area_struct *vma = fp->vma;
280 split_huge_page_pmd(vma, addr, pmd);
281 if (pmd_trans_unstable(pmd))
284 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
285 arch_enter_lazy_mmu_mode();
286 for (; addr != end; pte++, addr += PAGE_SIZE) {
289 if (!pte_present(ptent))
292 page = vm_normal_page(vma, addr, ptent);
296 if (PageSwapCache(page)) {
297 if (!trylock_page(page))
300 if (!try_to_free_swap(page)) {
305 ClearPageDirty(page);
310 * Some of architecture(ex, PPC) don't update TLB
311 * with set_pte_at and tlb_remove_tlb_entry so for
312 * the portability, remap the pte with old|clean
313 * after pte clearing.
315 ptent = ptep_get_and_clear_full(mm, addr, pte,
317 ptent = pte_mkold(ptent);
318 ptent = pte_mkclean(ptent);
319 set_pte_at(mm, addr, pte, ptent);
320 tlb_remove_tlb_entry(tlb, pte, addr);
322 arch_leave_lazy_mmu_mode();
323 pte_unmap_unlock(pte - 1, ptl);
328 static void madvise_free_page_range(struct mmu_gather *tlb,
329 struct vm_area_struct *vma,
330 unsigned long addr, unsigned long end)
332 struct madvise_free_private fp = {
337 struct mm_walk free_walk = {
338 .pmd_entry = madvise_free_pte_range,
344 tlb_start_vma(tlb, vma);
345 walk_page_range(addr, end, &free_walk);
346 tlb_end_vma(tlb, vma);
349 static int madvise_free_single_vma(struct vm_area_struct *vma,
350 unsigned long start_addr, unsigned long end_addr)
352 unsigned long start, end;
353 struct mm_struct *mm = vma->vm_mm;
354 struct mmu_gather tlb;
356 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
359 /* MADV_FREE works for only anon vma at the moment */
363 start = max(vma->vm_start, start_addr);
364 if (start >= vma->vm_end)
366 end = min(vma->vm_end, end_addr);
367 if (end <= vma->vm_start)
371 tlb_gather_mmu(&tlb, mm, start, end);
372 update_hiwater_rss(mm);
374 mmu_notifier_invalidate_range_start(mm, start, end);
375 madvise_free_page_range(&tlb, vma, start, end);
376 mmu_notifier_invalidate_range_end(mm, start, end);
377 tlb_finish_mmu(&tlb, start, end);
382 static long madvise_free(struct vm_area_struct *vma,
383 struct vm_area_struct **prev,
384 unsigned long start, unsigned long end)
387 return madvise_free_single_vma(vma, start, end);
391 * Application no longer needs these pages. If the pages are dirty,
392 * it's OK to just throw them away. The app will be more careful about
393 * data it wants to keep. Be sure to free swap resources too. The
394 * zap_page_range call sets things up for shrink_active_list to actually free
395 * these pages later if no one else has touched them in the meantime,
396 * although we could add these pages to a global reuse list for
397 * shrink_active_list to pick up before reclaiming other pages.
399 * NB: This interface discards data rather than pushes it out to swap,
400 * as some implementations do. This has performance implications for
401 * applications like large transactional databases which want to discard
402 * pages in anonymous maps after committing to backing store the data
403 * that was kept in them. There is no reason to write this data out to
404 * the swap area if the application is discarding it.
406 * An interface that causes the system to free clean pages and flush
407 * dirty pages is already available as msync(MS_INVALIDATE).
409 static long madvise_dontneed(struct vm_area_struct *vma,
410 struct vm_area_struct **prev,
411 unsigned long start, unsigned long end)
414 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
417 zap_page_range(vma, start, end - start, NULL);
422 * Application wants to free up the pages and associated backing store.
423 * This is effectively punching a hole into the middle of a file.
425 static long madvise_remove(struct vm_area_struct *vma,
426 struct vm_area_struct **prev,
427 unsigned long start, unsigned long end)
433 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
435 if (vma->vm_flags & VM_LOCKED)
440 if (!f || !f->f_mapping || !f->f_mapping->host) {
444 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
447 offset = (loff_t)(start - vma->vm_start)
448 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
451 * Filesystem's fallocate may need to take i_mutex. We need to
452 * explicitly grab a reference because the vma (and hence the
453 * vma's reference to the file) can go away as soon as we drop
457 up_read(¤t->mm->mmap_sem);
458 error = vfs_fallocate(f,
459 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
460 offset, end - start);
462 down_read(¤t->mm->mmap_sem);
466 #ifdef CONFIG_MEMORY_FAILURE
468 * Error injection support for memory error handling.
470 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
473 if (!capable(CAP_SYS_ADMIN))
475 for (; start < end; start += PAGE_SIZE <<
476 compound_order(compound_head(p))) {
479 ret = get_user_pages_fast(start, 1, 0, &p);
483 if (PageHWPoison(p)) {
487 if (bhv == MADV_SOFT_OFFLINE) {
488 pr_info("Soft offlining page %#lx at %#lx\n",
489 page_to_pfn(p), start);
490 ret = soft_offline_page(p, MF_COUNT_INCREASED);
495 pr_info("Injecting memory failure for page %#lx at %#lx\n",
496 page_to_pfn(p), start);
497 /* Ignore return value for now */
498 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
505 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
506 unsigned long start, unsigned long end, int behavior)
510 return madvise_remove(vma, prev, start, end);
512 return madvise_willneed(vma, prev, start, end);
515 * XXX: In this implementation, MADV_FREE works like
516 * MADV_DONTNEED on swapless system or full swap.
518 if (get_nr_swap_pages() > 0)
519 return madvise_free(vma, prev, start, end);
522 return madvise_dontneed(vma, prev, start, end);
524 return madvise_behavior(vma, prev, start, end, behavior);
529 madvise_behavior_valid(int behavior)
535 case MADV_SEQUENTIAL:
543 case MADV_UNMERGEABLE:
545 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
547 case MADV_NOHUGEPAGE:
559 * The madvise(2) system call.
561 * Applications can use madvise() to advise the kernel how it should
562 * handle paging I/O in this VM area. The idea is to help the kernel
563 * use appropriate read-ahead and caching techniques. The information
564 * provided is advisory only, and can be safely disregarded by the
565 * kernel without affecting the correct operation of the application.
568 * MADV_NORMAL - the default behavior is to read clusters. This
569 * results in some read-ahead and read-behind.
570 * MADV_RANDOM - the system should read the minimum amount of data
571 * on any access, since it is unlikely that the appli-
572 * cation will need more than what it asks for.
573 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
574 * once, so they can be aggressively read ahead, and
575 * can be freed soon after they are accessed.
576 * MADV_WILLNEED - the application is notifying the system to read
578 * MADV_DONTNEED - the application is finished with the given range,
579 * so the kernel can free resources associated with it.
580 * MADV_REMOVE - the application wants to free up the given range of
581 * pages and associated backing store.
582 * MADV_DONTFORK - omit this area from child's address space when forking:
583 * typically, to avoid COWing pages pinned by get_user_pages().
584 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
585 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
586 * this area with pages of identical content from other such areas.
587 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
591 * -EINVAL - start + len < 0, start is not page-aligned,
592 * "behavior" is not a valid value, or application
593 * is attempting to release locked or shared pages.
594 * -ENOMEM - addresses in the specified range are not currently
595 * mapped, or are outside the AS of the process.
596 * -EIO - an I/O error occurred while paging in data.
597 * -EBADF - map exists, but area maps something that isn't a file.
598 * -EAGAIN - a kernel resource was temporarily unavailable.
600 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
602 unsigned long end, tmp;
603 struct vm_area_struct *vma, *prev;
604 int unmapped_error = 0;
608 struct blk_plug plug;
610 #ifdef CONFIG_MEMORY_FAILURE
611 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
612 return madvise_hwpoison(behavior, start, start+len_in);
614 if (!madvise_behavior_valid(behavior))
617 if (start & ~PAGE_MASK)
619 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
621 /* Check to see whether len was rounded up from small -ve to zero */
633 write = madvise_need_mmap_write(behavior);
635 down_write(¤t->mm->mmap_sem);
637 down_read(¤t->mm->mmap_sem);
640 * If the interval [start,end) covers some unmapped address
641 * ranges, just ignore them, but return -ENOMEM at the end.
642 * - different from the way of handling in mlock etc.
644 vma = find_vma_prev(current->mm, start, &prev);
645 if (vma && start > vma->vm_start)
648 blk_start_plug(&plug);
650 /* Still start < end. */
655 /* Here start < (end|vma->vm_end). */
656 if (start < vma->vm_start) {
657 unmapped_error = -ENOMEM;
658 start = vma->vm_start;
663 /* Here vma->vm_start <= start < (end|vma->vm_end) */
668 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
669 error = madvise_vma(vma, &prev, start, tmp, behavior);
673 if (prev && start < prev->vm_end)
674 start = prev->vm_end;
675 error = unmapped_error;
680 else /* madvise_remove dropped mmap_sem */
681 vma = find_vma(current->mm, start);
684 blk_finish_plug(&plug);
686 up_write(¤t->mm->mmap_sem);
688 up_read(¤t->mm->mmap_sem);