Loading...
1/*
2 * linux/mm/madvise.c
3 *
4 * Copyright (C) 1999 Linus Torvalds
5 * Copyright (C) 2002 Christoph Hellwig
6 */
7
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>
17#include <linux/fs.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>
24
25#include <asm/tlb.h>
26
27/*
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.
31 */
32static int madvise_need_mmap_write(int behavior)
33{
34 switch (behavior) {
35 case MADV_REMOVE:
36 case MADV_WILLNEED:
37 case MADV_DONTNEED:
38 case MADV_FREE:
39 return 0;
40 default:
41 /* be safe, default to 1. list exceptions explicitly */
42 return 1;
43 }
44}
45
46/*
47 * We can potentially split a vm area into separate
48 * areas, each area with its own behavior.
49 */
50static long madvise_behavior(struct vm_area_struct *vma,
51 struct vm_area_struct **prev,
52 unsigned long start, unsigned long end, int behavior)
53{
54 struct mm_struct *mm = vma->vm_mm;
55 int error = 0;
56 pgoff_t pgoff;
57 unsigned long new_flags = vma->vm_flags;
58
59 switch (behavior) {
60 case MADV_NORMAL:
61 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
62 break;
63 case MADV_SEQUENTIAL:
64 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
65 break;
66 case MADV_RANDOM:
67 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
68 break;
69 case MADV_DONTFORK:
70 new_flags |= VM_DONTCOPY;
71 break;
72 case MADV_DOFORK:
73 if (vma->vm_flags & VM_IO) {
74 error = -EINVAL;
75 goto out;
76 }
77 new_flags &= ~VM_DONTCOPY;
78 break;
79 case MADV_DONTDUMP:
80 new_flags |= VM_DONTDUMP;
81 break;
82 case MADV_DODUMP:
83 if (new_flags & VM_SPECIAL) {
84 error = -EINVAL;
85 goto out;
86 }
87 new_flags &= ~VM_DONTDUMP;
88 break;
89 case MADV_MERGEABLE:
90 case MADV_UNMERGEABLE:
91 error = ksm_madvise(vma, start, end, behavior, &new_flags);
92 if (error)
93 goto out;
94 break;
95 case MADV_HUGEPAGE:
96 case MADV_NOHUGEPAGE:
97 error = hugepage_madvise(vma, &new_flags, behavior);
98 if (error)
99 goto out;
100 break;
101 }
102
103 if (new_flags == vma->vm_flags) {
104 *prev = vma;
105 goto out;
106 }
107
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);
112 if (*prev) {
113 vma = *prev;
114 goto success;
115 }
116
117 *prev = vma;
118
119 if (start != vma->vm_start) {
120 error = split_vma(mm, vma, start, 1);
121 if (error)
122 goto out;
123 }
124
125 if (end != vma->vm_end) {
126 error = split_vma(mm, vma, end, 0);
127 if (error)
128 goto out;
129 }
130
131success:
132 /*
133 * vm_flags is protected by the mmap_sem held in write mode.
134 */
135 vma->vm_flags = new_flags;
136
137out:
138 if (error == -ENOMEM)
139 error = -EAGAIN;
140 return error;
141}
142
143#ifdef CONFIG_SWAP
144static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
145 unsigned long end, struct mm_walk *walk)
146{
147 pte_t *orig_pte;
148 struct vm_area_struct *vma = walk->private;
149 unsigned long index;
150
151 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
152 return 0;
153
154 for (index = start; index != end; index += PAGE_SIZE) {
155 pte_t pte;
156 swp_entry_t entry;
157 struct page *page;
158 spinlock_t *ptl;
159
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);
163
164 if (pte_present(pte) || pte_none(pte))
165 continue;
166 entry = pte_to_swp_entry(pte);
167 if (unlikely(non_swap_entry(entry)))
168 continue;
169
170 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
171 vma, index);
172 if (page)
173 put_page(page);
174 }
175
176 return 0;
177}
178
179static void force_swapin_readahead(struct vm_area_struct *vma,
180 unsigned long start, unsigned long end)
181{
182 struct mm_walk walk = {
183 .mm = vma->vm_mm,
184 .pmd_entry = swapin_walk_pmd_entry,
185 .private = vma,
186 };
187
188 walk_page_range(start, end, &walk);
189
190 lru_add_drain(); /* Push any new pages onto the LRU now */
191}
192
193static void force_shm_swapin_readahead(struct vm_area_struct *vma,
194 unsigned long start, unsigned long end,
195 struct address_space *mapping)
196{
197 pgoff_t index;
198 struct page *page;
199 swp_entry_t swap;
200
201 for (; start < end; start += PAGE_SIZE) {
202 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
203
204 page = find_get_entry(mapping, index);
205 if (!radix_tree_exceptional_entry(page)) {
206 if (page)
207 put_page(page);
208 continue;
209 }
210 swap = radix_to_swp_entry(page);
211 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
212 NULL, 0);
213 if (page)
214 put_page(page);
215 }
216
217 lru_add_drain(); /* Push any new pages onto the LRU now */
218}
219#endif /* CONFIG_SWAP */
220
221/*
222 * Schedule all required I/O operations. Do not wait for completion.
223 */
224static long madvise_willneed(struct vm_area_struct *vma,
225 struct vm_area_struct **prev,
226 unsigned long start, unsigned long end)
227{
228 struct file *file = vma->vm_file;
229
230#ifdef CONFIG_SWAP
231 if (!file) {
232 *prev = vma;
233 force_swapin_readahead(vma, start, end);
234 return 0;
235 }
236
237 if (shmem_mapping(file->f_mapping)) {
238 *prev = vma;
239 force_shm_swapin_readahead(vma, start, end,
240 file->f_mapping);
241 return 0;
242 }
243#else
244 if (!file)
245 return -EBADF;
246#endif
247
248 if (IS_DAX(file_inode(file))) {
249 /* no bad return value, but ignore advice */
250 return 0;
251 }
252
253 *prev = vma;
254 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
255 if (end > vma->vm_end)
256 end = vma->vm_end;
257 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
258
259 force_page_cache_readahead(file->f_mapping, file, start, end - start);
260 return 0;
261}
262
263static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
264 unsigned long end, struct mm_walk *walk)
265
266{
267 struct mmu_gather *tlb = walk->private;
268 struct mm_struct *mm = tlb->mm;
269 struct vm_area_struct *vma = walk->vma;
270 spinlock_t *ptl;
271 pte_t *orig_pte, *pte, ptent;
272 struct page *page;
273 int nr_swap = 0;
274 unsigned long next;
275
276 next = pmd_addr_end(addr, end);
277 if (pmd_trans_huge(*pmd))
278 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
279 goto next;
280
281 if (pmd_trans_unstable(pmd))
282 return 0;
283
284 orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
285 arch_enter_lazy_mmu_mode();
286 for (; addr != end; pte++, addr += PAGE_SIZE) {
287 ptent = *pte;
288
289 if (pte_none(ptent))
290 continue;
291 /*
292 * If the pte has swp_entry, just clear page table to
293 * prevent swap-in which is more expensive rather than
294 * (page allocation + zeroing).
295 */
296 if (!pte_present(ptent)) {
297 swp_entry_t entry;
298
299 entry = pte_to_swp_entry(ptent);
300 if (non_swap_entry(entry))
301 continue;
302 nr_swap--;
303 free_swap_and_cache(entry);
304 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
305 continue;
306 }
307
308 page = vm_normal_page(vma, addr, ptent);
309 if (!page)
310 continue;
311
312 /*
313 * If pmd isn't transhuge but the page is THP and
314 * is owned by only this process, split it and
315 * deactivate all pages.
316 */
317 if (PageTransCompound(page)) {
318 if (page_mapcount(page) != 1)
319 goto out;
320 get_page(page);
321 if (!trylock_page(page)) {
322 put_page(page);
323 goto out;
324 }
325 pte_unmap_unlock(orig_pte, ptl);
326 if (split_huge_page(page)) {
327 unlock_page(page);
328 put_page(page);
329 pte_offset_map_lock(mm, pmd, addr, &ptl);
330 goto out;
331 }
332 put_page(page);
333 unlock_page(page);
334 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
335 pte--;
336 addr -= PAGE_SIZE;
337 continue;
338 }
339
340 VM_BUG_ON_PAGE(PageTransCompound(page), page);
341
342 if (PageSwapCache(page) || PageDirty(page)) {
343 if (!trylock_page(page))
344 continue;
345 /*
346 * If page is shared with others, we couldn't clear
347 * PG_dirty of the page.
348 */
349 if (page_mapcount(page) != 1) {
350 unlock_page(page);
351 continue;
352 }
353
354 if (PageSwapCache(page) && !try_to_free_swap(page)) {
355 unlock_page(page);
356 continue;
357 }
358
359 ClearPageDirty(page);
360 unlock_page(page);
361 }
362
363 if (pte_young(ptent) || pte_dirty(ptent)) {
364 /*
365 * Some of architecture(ex, PPC) don't update TLB
366 * with set_pte_at and tlb_remove_tlb_entry so for
367 * the portability, remap the pte with old|clean
368 * after pte clearing.
369 */
370 ptent = ptep_get_and_clear_full(mm, addr, pte,
371 tlb->fullmm);
372
373 ptent = pte_mkold(ptent);
374 ptent = pte_mkclean(ptent);
375 set_pte_at(mm, addr, pte, ptent);
376 if (PageActive(page))
377 deactivate_page(page);
378 tlb_remove_tlb_entry(tlb, pte, addr);
379 }
380 }
381out:
382 if (nr_swap) {
383 if (current->mm == mm)
384 sync_mm_rss(mm);
385
386 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
387 }
388 arch_leave_lazy_mmu_mode();
389 pte_unmap_unlock(orig_pte, ptl);
390 cond_resched();
391next:
392 return 0;
393}
394
395static void madvise_free_page_range(struct mmu_gather *tlb,
396 struct vm_area_struct *vma,
397 unsigned long addr, unsigned long end)
398{
399 struct mm_walk free_walk = {
400 .pmd_entry = madvise_free_pte_range,
401 .mm = vma->vm_mm,
402 .private = tlb,
403 };
404
405 tlb_start_vma(tlb, vma);
406 walk_page_range(addr, end, &free_walk);
407 tlb_end_vma(tlb, vma);
408}
409
410static int madvise_free_single_vma(struct vm_area_struct *vma,
411 unsigned long start_addr, unsigned long end_addr)
412{
413 unsigned long start, end;
414 struct mm_struct *mm = vma->vm_mm;
415 struct mmu_gather tlb;
416
417 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
418 return -EINVAL;
419
420 /* MADV_FREE works for only anon vma at the moment */
421 if (!vma_is_anonymous(vma))
422 return -EINVAL;
423
424 start = max(vma->vm_start, start_addr);
425 if (start >= vma->vm_end)
426 return -EINVAL;
427 end = min(vma->vm_end, end_addr);
428 if (end <= vma->vm_start)
429 return -EINVAL;
430
431 lru_add_drain();
432 tlb_gather_mmu(&tlb, mm, start, end);
433 update_hiwater_rss(mm);
434
435 mmu_notifier_invalidate_range_start(mm, start, end);
436 madvise_free_page_range(&tlb, vma, start, end);
437 mmu_notifier_invalidate_range_end(mm, start, end);
438 tlb_finish_mmu(&tlb, start, end);
439
440 return 0;
441}
442
443static long madvise_free(struct vm_area_struct *vma,
444 struct vm_area_struct **prev,
445 unsigned long start, unsigned long end)
446{
447 *prev = vma;
448 return madvise_free_single_vma(vma, start, end);
449}
450
451/*
452 * Application no longer needs these pages. If the pages are dirty,
453 * it's OK to just throw them away. The app will be more careful about
454 * data it wants to keep. Be sure to free swap resources too. The
455 * zap_page_range call sets things up for shrink_active_list to actually free
456 * these pages later if no one else has touched them in the meantime,
457 * although we could add these pages to a global reuse list for
458 * shrink_active_list to pick up before reclaiming other pages.
459 *
460 * NB: This interface discards data rather than pushes it out to swap,
461 * as some implementations do. This has performance implications for
462 * applications like large transactional databases which want to discard
463 * pages in anonymous maps after committing to backing store the data
464 * that was kept in them. There is no reason to write this data out to
465 * the swap area if the application is discarding it.
466 *
467 * An interface that causes the system to free clean pages and flush
468 * dirty pages is already available as msync(MS_INVALIDATE).
469 */
470static long madvise_dontneed(struct vm_area_struct *vma,
471 struct vm_area_struct **prev,
472 unsigned long start, unsigned long end)
473{
474 *prev = vma;
475 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
476 return -EINVAL;
477
478 zap_page_range(vma, start, end - start, NULL);
479 return 0;
480}
481
482/*
483 * Application wants to free up the pages and associated backing store.
484 * This is effectively punching a hole into the middle of a file.
485 */
486static long madvise_remove(struct vm_area_struct *vma,
487 struct vm_area_struct **prev,
488 unsigned long start, unsigned long end)
489{
490 loff_t offset;
491 int error;
492 struct file *f;
493
494 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
495
496 if (vma->vm_flags & VM_LOCKED)
497 return -EINVAL;
498
499 f = vma->vm_file;
500
501 if (!f || !f->f_mapping || !f->f_mapping->host) {
502 return -EINVAL;
503 }
504
505 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
506 return -EACCES;
507
508 offset = (loff_t)(start - vma->vm_start)
509 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
510
511 /*
512 * Filesystem's fallocate may need to take i_mutex. We need to
513 * explicitly grab a reference because the vma (and hence the
514 * vma's reference to the file) can go away as soon as we drop
515 * mmap_sem.
516 */
517 get_file(f);
518 up_read(¤t->mm->mmap_sem);
519 error = vfs_fallocate(f,
520 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
521 offset, end - start);
522 fput(f);
523 down_read(¤t->mm->mmap_sem);
524 return error;
525}
526
527#ifdef CONFIG_MEMORY_FAILURE
528/*
529 * Error injection support for memory error handling.
530 */
531static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
532{
533 struct page *p;
534 if (!capable(CAP_SYS_ADMIN))
535 return -EPERM;
536 for (; start < end; start += PAGE_SIZE <<
537 compound_order(compound_head(p))) {
538 int ret;
539
540 ret = get_user_pages_fast(start, 1, 0, &p);
541 if (ret != 1)
542 return ret;
543
544 if (PageHWPoison(p)) {
545 put_page(p);
546 continue;
547 }
548 if (bhv == MADV_SOFT_OFFLINE) {
549 pr_info("Soft offlining page %#lx at %#lx\n",
550 page_to_pfn(p), start);
551 ret = soft_offline_page(p, MF_COUNT_INCREASED);
552 if (ret)
553 return ret;
554 continue;
555 }
556 pr_info("Injecting memory failure for page %#lx at %#lx\n",
557 page_to_pfn(p), start);
558 ret = memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
559 if (ret)
560 return ret;
561 }
562 return 0;
563}
564#endif
565
566static long
567madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
568 unsigned long start, unsigned long end, int behavior)
569{
570 switch (behavior) {
571 case MADV_REMOVE:
572 return madvise_remove(vma, prev, start, end);
573 case MADV_WILLNEED:
574 return madvise_willneed(vma, prev, start, end);
575 case MADV_FREE:
576 /*
577 * XXX: In this implementation, MADV_FREE works like
578 * MADV_DONTNEED on swapless system or full swap.
579 */
580 if (get_nr_swap_pages() > 0)
581 return madvise_free(vma, prev, start, end);
582 /* passthrough */
583 case MADV_DONTNEED:
584 return madvise_dontneed(vma, prev, start, end);
585 default:
586 return madvise_behavior(vma, prev, start, end, behavior);
587 }
588}
589
590static bool
591madvise_behavior_valid(int behavior)
592{
593 switch (behavior) {
594 case MADV_DOFORK:
595 case MADV_DONTFORK:
596 case MADV_NORMAL:
597 case MADV_SEQUENTIAL:
598 case MADV_RANDOM:
599 case MADV_REMOVE:
600 case MADV_WILLNEED:
601 case MADV_DONTNEED:
602 case MADV_FREE:
603#ifdef CONFIG_KSM
604 case MADV_MERGEABLE:
605 case MADV_UNMERGEABLE:
606#endif
607#ifdef CONFIG_TRANSPARENT_HUGEPAGE
608 case MADV_HUGEPAGE:
609 case MADV_NOHUGEPAGE:
610#endif
611 case MADV_DONTDUMP:
612 case MADV_DODUMP:
613 return true;
614
615 default:
616 return false;
617 }
618}
619
620/*
621 * The madvise(2) system call.
622 *
623 * Applications can use madvise() to advise the kernel how it should
624 * handle paging I/O in this VM area. The idea is to help the kernel
625 * use appropriate read-ahead and caching techniques. The information
626 * provided is advisory only, and can be safely disregarded by the
627 * kernel without affecting the correct operation of the application.
628 *
629 * behavior values:
630 * MADV_NORMAL - the default behavior is to read clusters. This
631 * results in some read-ahead and read-behind.
632 * MADV_RANDOM - the system should read the minimum amount of data
633 * on any access, since it is unlikely that the appli-
634 * cation will need more than what it asks for.
635 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
636 * once, so they can be aggressively read ahead, and
637 * can be freed soon after they are accessed.
638 * MADV_WILLNEED - the application is notifying the system to read
639 * some pages ahead.
640 * MADV_DONTNEED - the application is finished with the given range,
641 * so the kernel can free resources associated with it.
642 * MADV_FREE - the application marks pages in the given range as lazy free,
643 * where actual purges are postponed until memory pressure happens.
644 * MADV_REMOVE - the application wants to free up the given range of
645 * pages and associated backing store.
646 * MADV_DONTFORK - omit this area from child's address space when forking:
647 * typically, to avoid COWing pages pinned by get_user_pages().
648 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
649 * MADV_HWPOISON - trigger memory error handler as if the given memory range
650 * were corrupted by unrecoverable hardware memory failure.
651 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
652 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
653 * this area with pages of identical content from other such areas.
654 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
655 * MADV_HUGEPAGE - the application wants to back the given range by transparent
656 * huge pages in the future. Existing pages might be coalesced and
657 * new pages might be allocated as THP.
658 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
659 * transparent huge pages so the existing pages will not be
660 * coalesced into THP and new pages will not be allocated as THP.
661 * MADV_DONTDUMP - the application wants to prevent pages in the given range
662 * from being included in its core dump.
663 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
664 *
665 * return values:
666 * zero - success
667 * -EINVAL - start + len < 0, start is not page-aligned,
668 * "behavior" is not a valid value, or application
669 * is attempting to release locked or shared pages.
670 * -ENOMEM - addresses in the specified range are not currently
671 * mapped, or are outside the AS of the process.
672 * -EIO - an I/O error occurred while paging in data.
673 * -EBADF - map exists, but area maps something that isn't a file.
674 * -EAGAIN - a kernel resource was temporarily unavailable.
675 */
676SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
677{
678 unsigned long end, tmp;
679 struct vm_area_struct *vma, *prev;
680 int unmapped_error = 0;
681 int error = -EINVAL;
682 int write;
683 size_t len;
684 struct blk_plug plug;
685
686#ifdef CONFIG_MEMORY_FAILURE
687 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
688 return madvise_hwpoison(behavior, start, start+len_in);
689#endif
690 if (!madvise_behavior_valid(behavior))
691 return error;
692
693 if (start & ~PAGE_MASK)
694 return error;
695 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
696
697 /* Check to see whether len was rounded up from small -ve to zero */
698 if (len_in && !len)
699 return error;
700
701 end = start + len;
702 if (end < start)
703 return error;
704
705 error = 0;
706 if (end == start)
707 return error;
708
709 write = madvise_need_mmap_write(behavior);
710 if (write)
711 down_write(¤t->mm->mmap_sem);
712 else
713 down_read(¤t->mm->mmap_sem);
714
715 /*
716 * If the interval [start,end) covers some unmapped address
717 * ranges, just ignore them, but return -ENOMEM at the end.
718 * - different from the way of handling in mlock etc.
719 */
720 vma = find_vma_prev(current->mm, start, &prev);
721 if (vma && start > vma->vm_start)
722 prev = vma;
723
724 blk_start_plug(&plug);
725 for (;;) {
726 /* Still start < end. */
727 error = -ENOMEM;
728 if (!vma)
729 goto out;
730
731 /* Here start < (end|vma->vm_end). */
732 if (start < vma->vm_start) {
733 unmapped_error = -ENOMEM;
734 start = vma->vm_start;
735 if (start >= end)
736 goto out;
737 }
738
739 /* Here vma->vm_start <= start < (end|vma->vm_end) */
740 tmp = vma->vm_end;
741 if (end < tmp)
742 tmp = end;
743
744 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
745 error = madvise_vma(vma, &prev, start, tmp, behavior);
746 if (error)
747 goto out;
748 start = tmp;
749 if (prev && start < prev->vm_end)
750 start = prev->vm_end;
751 error = unmapped_error;
752 if (start >= end)
753 goto out;
754 if (prev)
755 vma = prev->vm_next;
756 else /* madvise_remove dropped mmap_sem */
757 vma = find_vma(current->mm, start);
758 }
759out:
760 blk_finish_plug(&plug);
761 if (write)
762 up_write(¤t->mm->mmap_sem);
763 else
764 up_read(¤t->mm->mmap_sem);
765
766 return error;
767}
1/*
2 * linux/mm/madvise.c
3 *
4 * Copyright (C) 1999 Linus Torvalds
5 * Copyright (C) 2002 Christoph Hellwig
6 */
7
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>
17#include <linux/fs.h>
18#include <linux/file.h>
19#include <linux/blkdev.h>
20#include <linux/swap.h>
21#include <linux/swapops.h>
22
23/*
24 * Any behaviour which results in changes to the vma->vm_flags needs to
25 * take mmap_sem for writing. Others, which simply traverse vmas, need
26 * to only take it for reading.
27 */
28static int madvise_need_mmap_write(int behavior)
29{
30 switch (behavior) {
31 case MADV_REMOVE:
32 case MADV_WILLNEED:
33 case MADV_DONTNEED:
34 return 0;
35 default:
36 /* be safe, default to 1. list exceptions explicitly */
37 return 1;
38 }
39}
40
41/*
42 * We can potentially split a vm area into separate
43 * areas, each area with its own behavior.
44 */
45static long madvise_behavior(struct vm_area_struct *vma,
46 struct vm_area_struct **prev,
47 unsigned long start, unsigned long end, int behavior)
48{
49 struct mm_struct *mm = vma->vm_mm;
50 int error = 0;
51 pgoff_t pgoff;
52 unsigned long new_flags = vma->vm_flags;
53
54 switch (behavior) {
55 case MADV_NORMAL:
56 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
57 break;
58 case MADV_SEQUENTIAL:
59 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
60 break;
61 case MADV_RANDOM:
62 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
63 break;
64 case MADV_DONTFORK:
65 new_flags |= VM_DONTCOPY;
66 break;
67 case MADV_DOFORK:
68 if (vma->vm_flags & VM_IO) {
69 error = -EINVAL;
70 goto out;
71 }
72 new_flags &= ~VM_DONTCOPY;
73 break;
74 case MADV_DONTDUMP:
75 new_flags |= VM_DONTDUMP;
76 break;
77 case MADV_DODUMP:
78 if (new_flags & VM_SPECIAL) {
79 error = -EINVAL;
80 goto out;
81 }
82 new_flags &= ~VM_DONTDUMP;
83 break;
84 case MADV_MERGEABLE:
85 case MADV_UNMERGEABLE:
86 error = ksm_madvise(vma, start, end, behavior, &new_flags);
87 if (error)
88 goto out;
89 break;
90 case MADV_HUGEPAGE:
91 case MADV_NOHUGEPAGE:
92 error = hugepage_madvise(vma, &new_flags, behavior);
93 if (error)
94 goto out;
95 break;
96 }
97
98 if (new_flags == vma->vm_flags) {
99 *prev = vma;
100 goto out;
101 }
102
103 pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
104 *prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
105 vma->vm_file, pgoff, vma_policy(vma));
106 if (*prev) {
107 vma = *prev;
108 goto success;
109 }
110
111 *prev = vma;
112
113 if (start != vma->vm_start) {
114 error = split_vma(mm, vma, start, 1);
115 if (error)
116 goto out;
117 }
118
119 if (end != vma->vm_end) {
120 error = split_vma(mm, vma, end, 0);
121 if (error)
122 goto out;
123 }
124
125success:
126 /*
127 * vm_flags is protected by the mmap_sem held in write mode.
128 */
129 vma->vm_flags = new_flags;
130
131out:
132 if (error == -ENOMEM)
133 error = -EAGAIN;
134 return error;
135}
136
137#ifdef CONFIG_SWAP
138static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
139 unsigned long end, struct mm_walk *walk)
140{
141 pte_t *orig_pte;
142 struct vm_area_struct *vma = walk->private;
143 unsigned long index;
144
145 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
146 return 0;
147
148 for (index = start; index != end; index += PAGE_SIZE) {
149 pte_t pte;
150 swp_entry_t entry;
151 struct page *page;
152 spinlock_t *ptl;
153
154 orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
155 pte = *(orig_pte + ((index - start) / PAGE_SIZE));
156 pte_unmap_unlock(orig_pte, ptl);
157
158 if (pte_present(pte) || pte_none(pte) || pte_file(pte))
159 continue;
160 entry = pte_to_swp_entry(pte);
161 if (unlikely(non_swap_entry(entry)))
162 continue;
163
164 page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
165 vma, index);
166 if (page)
167 page_cache_release(page);
168 }
169
170 return 0;
171}
172
173static void force_swapin_readahead(struct vm_area_struct *vma,
174 unsigned long start, unsigned long end)
175{
176 struct mm_walk walk = {
177 .mm = vma->vm_mm,
178 .pmd_entry = swapin_walk_pmd_entry,
179 .private = vma,
180 };
181
182 walk_page_range(start, end, &walk);
183
184 lru_add_drain(); /* Push any new pages onto the LRU now */
185}
186
187static void force_shm_swapin_readahead(struct vm_area_struct *vma,
188 unsigned long start, unsigned long end,
189 struct address_space *mapping)
190{
191 pgoff_t index;
192 struct page *page;
193 swp_entry_t swap;
194
195 for (; start < end; start += PAGE_SIZE) {
196 index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
197
198 page = find_get_entry(mapping, index);
199 if (!radix_tree_exceptional_entry(page)) {
200 if (page)
201 page_cache_release(page);
202 continue;
203 }
204 swap = radix_to_swp_entry(page);
205 page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
206 NULL, 0);
207 if (page)
208 page_cache_release(page);
209 }
210
211 lru_add_drain(); /* Push any new pages onto the LRU now */
212}
213#endif /* CONFIG_SWAP */
214
215/*
216 * Schedule all required I/O operations. Do not wait for completion.
217 */
218static long madvise_willneed(struct vm_area_struct *vma,
219 struct vm_area_struct **prev,
220 unsigned long start, unsigned long end)
221{
222 struct file *file = vma->vm_file;
223
224#ifdef CONFIG_SWAP
225 if (!file || mapping_cap_swap_backed(file->f_mapping)) {
226 *prev = vma;
227 if (!file)
228 force_swapin_readahead(vma, start, end);
229 else
230 force_shm_swapin_readahead(vma, start, end,
231 file->f_mapping);
232 return 0;
233 }
234#endif
235
236 if (!file)
237 return -EBADF;
238
239 if (file->f_mapping->a_ops->get_xip_mem) {
240 /* no bad return value, but ignore advice */
241 return 0;
242 }
243
244 *prev = vma;
245 start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
246 if (end > vma->vm_end)
247 end = vma->vm_end;
248 end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
249
250 force_page_cache_readahead(file->f_mapping, file, start, end - start);
251 return 0;
252}
253
254/*
255 * Application no longer needs these pages. If the pages are dirty,
256 * it's OK to just throw them away. The app will be more careful about
257 * data it wants to keep. Be sure to free swap resources too. The
258 * zap_page_range call sets things up for shrink_active_list to actually free
259 * these pages later if no one else has touched them in the meantime,
260 * although we could add these pages to a global reuse list for
261 * shrink_active_list to pick up before reclaiming other pages.
262 *
263 * NB: This interface discards data rather than pushes it out to swap,
264 * as some implementations do. This has performance implications for
265 * applications like large transactional databases which want to discard
266 * pages in anonymous maps after committing to backing store the data
267 * that was kept in them. There is no reason to write this data out to
268 * the swap area if the application is discarding it.
269 *
270 * An interface that causes the system to free clean pages and flush
271 * dirty pages is already available as msync(MS_INVALIDATE).
272 */
273static long madvise_dontneed(struct vm_area_struct *vma,
274 struct vm_area_struct **prev,
275 unsigned long start, unsigned long end)
276{
277 *prev = vma;
278 if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
279 return -EINVAL;
280
281 if (unlikely(vma->vm_flags & VM_NONLINEAR)) {
282 struct zap_details details = {
283 .nonlinear_vma = vma,
284 .last_index = ULONG_MAX,
285 };
286 zap_page_range(vma, start, end - start, &details);
287 } else
288 zap_page_range(vma, start, end - start, NULL);
289 return 0;
290}
291
292/*
293 * Application wants to free up the pages and associated backing store.
294 * This is effectively punching a hole into the middle of a file.
295 *
296 * NOTE: Currently, only shmfs/tmpfs is supported for this operation.
297 * Other filesystems return -ENOSYS.
298 */
299static long madvise_remove(struct vm_area_struct *vma,
300 struct vm_area_struct **prev,
301 unsigned long start, unsigned long end)
302{
303 loff_t offset;
304 int error;
305 struct file *f;
306
307 *prev = NULL; /* tell sys_madvise we drop mmap_sem */
308
309 if (vma->vm_flags & (VM_LOCKED|VM_NONLINEAR|VM_HUGETLB))
310 return -EINVAL;
311
312 f = vma->vm_file;
313
314 if (!f || !f->f_mapping || !f->f_mapping->host) {
315 return -EINVAL;
316 }
317
318 if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
319 return -EACCES;
320
321 offset = (loff_t)(start - vma->vm_start)
322 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
323
324 /*
325 * Filesystem's fallocate may need to take i_mutex. We need to
326 * explicitly grab a reference because the vma (and hence the
327 * vma's reference to the file) can go away as soon as we drop
328 * mmap_sem.
329 */
330 get_file(f);
331 up_read(¤t->mm->mmap_sem);
332 error = do_fallocate(f,
333 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
334 offset, end - start);
335 fput(f);
336 down_read(¤t->mm->mmap_sem);
337 return error;
338}
339
340#ifdef CONFIG_MEMORY_FAILURE
341/*
342 * Error injection support for memory error handling.
343 */
344static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
345{
346 struct page *p;
347 if (!capable(CAP_SYS_ADMIN))
348 return -EPERM;
349 for (; start < end; start += PAGE_SIZE <<
350 compound_order(compound_head(p))) {
351 int ret;
352
353 ret = get_user_pages_fast(start, 1, 0, &p);
354 if (ret != 1)
355 return ret;
356
357 if (PageHWPoison(p)) {
358 put_page(p);
359 continue;
360 }
361 if (bhv == MADV_SOFT_OFFLINE) {
362 pr_info("Soft offlining page %#lx at %#lx\n",
363 page_to_pfn(p), start);
364 ret = soft_offline_page(p, MF_COUNT_INCREASED);
365 if (ret)
366 return ret;
367 continue;
368 }
369 pr_info("Injecting memory failure for page %#lx at %#lx\n",
370 page_to_pfn(p), start);
371 /* Ignore return value for now */
372 memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
373 }
374 return 0;
375}
376#endif
377
378static long
379madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
380 unsigned long start, unsigned long end, int behavior)
381{
382 switch (behavior) {
383 case MADV_REMOVE:
384 return madvise_remove(vma, prev, start, end);
385 case MADV_WILLNEED:
386 return madvise_willneed(vma, prev, start, end);
387 case MADV_DONTNEED:
388 return madvise_dontneed(vma, prev, start, end);
389 default:
390 return madvise_behavior(vma, prev, start, end, behavior);
391 }
392}
393
394static int
395madvise_behavior_valid(int behavior)
396{
397 switch (behavior) {
398 case MADV_DOFORK:
399 case MADV_DONTFORK:
400 case MADV_NORMAL:
401 case MADV_SEQUENTIAL:
402 case MADV_RANDOM:
403 case MADV_REMOVE:
404 case MADV_WILLNEED:
405 case MADV_DONTNEED:
406#ifdef CONFIG_KSM
407 case MADV_MERGEABLE:
408 case MADV_UNMERGEABLE:
409#endif
410#ifdef CONFIG_TRANSPARENT_HUGEPAGE
411 case MADV_HUGEPAGE:
412 case MADV_NOHUGEPAGE:
413#endif
414 case MADV_DONTDUMP:
415 case MADV_DODUMP:
416 return 1;
417
418 default:
419 return 0;
420 }
421}
422
423/*
424 * The madvise(2) system call.
425 *
426 * Applications can use madvise() to advise the kernel how it should
427 * handle paging I/O in this VM area. The idea is to help the kernel
428 * use appropriate read-ahead and caching techniques. The information
429 * provided is advisory only, and can be safely disregarded by the
430 * kernel without affecting the correct operation of the application.
431 *
432 * behavior values:
433 * MADV_NORMAL - the default behavior is to read clusters. This
434 * results in some read-ahead and read-behind.
435 * MADV_RANDOM - the system should read the minimum amount of data
436 * on any access, since it is unlikely that the appli-
437 * cation will need more than what it asks for.
438 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
439 * once, so they can be aggressively read ahead, and
440 * can be freed soon after they are accessed.
441 * MADV_WILLNEED - the application is notifying the system to read
442 * some pages ahead.
443 * MADV_DONTNEED - the application is finished with the given range,
444 * so the kernel can free resources associated with it.
445 * MADV_REMOVE - the application wants to free up the given range of
446 * pages and associated backing store.
447 * MADV_DONTFORK - omit this area from child's address space when forking:
448 * typically, to avoid COWing pages pinned by get_user_pages().
449 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
450 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
451 * this area with pages of identical content from other such areas.
452 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
453 *
454 * return values:
455 * zero - success
456 * -EINVAL - start + len < 0, start is not page-aligned,
457 * "behavior" is not a valid value, or application
458 * is attempting to release locked or shared pages.
459 * -ENOMEM - addresses in the specified range are not currently
460 * mapped, or are outside the AS of the process.
461 * -EIO - an I/O error occurred while paging in data.
462 * -EBADF - map exists, but area maps something that isn't a file.
463 * -EAGAIN - a kernel resource was temporarily unavailable.
464 */
465SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
466{
467 unsigned long end, tmp;
468 struct vm_area_struct *vma, *prev;
469 int unmapped_error = 0;
470 int error = -EINVAL;
471 int write;
472 size_t len;
473 struct blk_plug plug;
474
475#ifdef CONFIG_MEMORY_FAILURE
476 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
477 return madvise_hwpoison(behavior, start, start+len_in);
478#endif
479 if (!madvise_behavior_valid(behavior))
480 return error;
481
482 if (start & ~PAGE_MASK)
483 return error;
484 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
485
486 /* Check to see whether len was rounded up from small -ve to zero */
487 if (len_in && !len)
488 return error;
489
490 end = start + len;
491 if (end < start)
492 return error;
493
494 error = 0;
495 if (end == start)
496 return error;
497
498 write = madvise_need_mmap_write(behavior);
499 if (write)
500 down_write(¤t->mm->mmap_sem);
501 else
502 down_read(¤t->mm->mmap_sem);
503
504 /*
505 * If the interval [start,end) covers some unmapped address
506 * ranges, just ignore them, but return -ENOMEM at the end.
507 * - different from the way of handling in mlock etc.
508 */
509 vma = find_vma_prev(current->mm, start, &prev);
510 if (vma && start > vma->vm_start)
511 prev = vma;
512
513 blk_start_plug(&plug);
514 for (;;) {
515 /* Still start < end. */
516 error = -ENOMEM;
517 if (!vma)
518 goto out;
519
520 /* Here start < (end|vma->vm_end). */
521 if (start < vma->vm_start) {
522 unmapped_error = -ENOMEM;
523 start = vma->vm_start;
524 if (start >= end)
525 goto out;
526 }
527
528 /* Here vma->vm_start <= start < (end|vma->vm_end) */
529 tmp = vma->vm_end;
530 if (end < tmp)
531 tmp = end;
532
533 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
534 error = madvise_vma(vma, &prev, start, tmp, behavior);
535 if (error)
536 goto out;
537 start = tmp;
538 if (prev && start < prev->vm_end)
539 start = prev->vm_end;
540 error = unmapped_error;
541 if (start >= end)
542 goto out;
543 if (prev)
544 vma = prev->vm_next;
545 else /* madvise_remove dropped mmap_sem */
546 vma = find_vma(current->mm, start);
547 }
548out:
549 blk_finish_plug(&plug);
550 if (write)
551 up_write(¤t->mm->mmap_sem);
552 else
553 up_read(¤t->mm->mmap_sem);
554
555 return error;
556}