Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/madvise.c
4 *
5 * Copyright (C) 1999 Linus Torvalds
6 * Copyright (C) 2002 Christoph Hellwig
7 */
8
9#include <linux/mman.h>
10#include <linux/pagemap.h>
11#include <linux/syscalls.h>
12#include <linux/mempolicy.h>
13#include <linux/page-isolation.h>
14#include <linux/page_idle.h>
15#include <linux/userfaultfd_k.h>
16#include <linux/hugetlb.h>
17#include <linux/falloc.h>
18#include <linux/fadvise.h>
19#include <linux/sched.h>
20#include <linux/sched/mm.h>
21#include <linux/mm_inline.h>
22#include <linux/string.h>
23#include <linux/uio.h>
24#include <linux/ksm.h>
25#include <linux/fs.h>
26#include <linux/file.h>
27#include <linux/blkdev.h>
28#include <linux/backing-dev.h>
29#include <linux/pagewalk.h>
30#include <linux/swap.h>
31#include <linux/swapops.h>
32#include <linux/shmem_fs.h>
33#include <linux/mmu_notifier.h>
34
35#include <asm/tlb.h>
36
37#include "internal.h"
38#include "swap.h"
39
40struct madvise_walk_private {
41 struct mmu_gather *tlb;
42 bool pageout;
43};
44
45/*
46 * Any behaviour which results in changes to the vma->vm_flags needs to
47 * take mmap_lock for writing. Others, which simply traverse vmas, need
48 * to only take it for reading.
49 */
50static int madvise_need_mmap_write(int behavior)
51{
52 switch (behavior) {
53 case MADV_REMOVE:
54 case MADV_WILLNEED:
55 case MADV_DONTNEED:
56 case MADV_DONTNEED_LOCKED:
57 case MADV_COLD:
58 case MADV_PAGEOUT:
59 case MADV_FREE:
60 case MADV_POPULATE_READ:
61 case MADV_POPULATE_WRITE:
62 case MADV_COLLAPSE:
63 return 0;
64 default:
65 /* be safe, default to 1. list exceptions explicitly */
66 return 1;
67 }
68}
69
70#ifdef CONFIG_ANON_VMA_NAME
71struct anon_vma_name *anon_vma_name_alloc(const char *name)
72{
73 struct anon_vma_name *anon_name;
74 size_t count;
75
76 /* Add 1 for NUL terminator at the end of the anon_name->name */
77 count = strlen(name) + 1;
78 anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
79 if (anon_name) {
80 kref_init(&anon_name->kref);
81 memcpy(anon_name->name, name, count);
82 }
83
84 return anon_name;
85}
86
87void anon_vma_name_free(struct kref *kref)
88{
89 struct anon_vma_name *anon_name =
90 container_of(kref, struct anon_vma_name, kref);
91 kfree(anon_name);
92}
93
94struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
95{
96 mmap_assert_locked(vma->vm_mm);
97
98 return vma->anon_name;
99}
100
101/* mmap_lock should be write-locked */
102static int replace_anon_vma_name(struct vm_area_struct *vma,
103 struct anon_vma_name *anon_name)
104{
105 struct anon_vma_name *orig_name = anon_vma_name(vma);
106
107 if (!anon_name) {
108 vma->anon_name = NULL;
109 anon_vma_name_put(orig_name);
110 return 0;
111 }
112
113 if (anon_vma_name_eq(orig_name, anon_name))
114 return 0;
115
116 vma->anon_name = anon_vma_name_reuse(anon_name);
117 anon_vma_name_put(orig_name);
118
119 return 0;
120}
121#else /* CONFIG_ANON_VMA_NAME */
122static int replace_anon_vma_name(struct vm_area_struct *vma,
123 struct anon_vma_name *anon_name)
124{
125 if (anon_name)
126 return -EINVAL;
127
128 return 0;
129}
130#endif /* CONFIG_ANON_VMA_NAME */
131/*
132 * Update the vm_flags on region of a vma, splitting it or merging it as
133 * necessary. Must be called with mmap_lock held for writing;
134 * Caller should ensure anon_name stability by raising its refcount even when
135 * anon_name belongs to a valid vma because this function might free that vma.
136 */
137static int madvise_update_vma(struct vm_area_struct *vma,
138 struct vm_area_struct **prev, unsigned long start,
139 unsigned long end, unsigned long new_flags,
140 struct anon_vma_name *anon_name)
141{
142 struct mm_struct *mm = vma->vm_mm;
143 int error;
144 VMA_ITERATOR(vmi, mm, start);
145
146 if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
147 *prev = vma;
148 return 0;
149 }
150
151 vma = vma_modify_flags_name(&vmi, *prev, vma, start, end, new_flags,
152 anon_name);
153 if (IS_ERR(vma))
154 return PTR_ERR(vma);
155
156 *prev = vma;
157
158 /* vm_flags is protected by the mmap_lock held in write mode. */
159 vma_start_write(vma);
160 vm_flags_reset(vma, new_flags);
161 if (!vma->vm_file || vma_is_anon_shmem(vma)) {
162 error = replace_anon_vma_name(vma, anon_name);
163 if (error)
164 return error;
165 }
166
167 return 0;
168}
169
170#ifdef CONFIG_SWAP
171static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
172 unsigned long end, struct mm_walk *walk)
173{
174 struct vm_area_struct *vma = walk->private;
175 struct swap_iocb *splug = NULL;
176 pte_t *ptep = NULL;
177 spinlock_t *ptl;
178 unsigned long addr;
179
180 for (addr = start; addr < end; addr += PAGE_SIZE) {
181 pte_t pte;
182 swp_entry_t entry;
183 struct folio *folio;
184
185 if (!ptep++) {
186 ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
187 if (!ptep)
188 break;
189 }
190
191 pte = ptep_get(ptep);
192 if (!is_swap_pte(pte))
193 continue;
194 entry = pte_to_swp_entry(pte);
195 if (unlikely(non_swap_entry(entry)))
196 continue;
197
198 pte_unmap_unlock(ptep, ptl);
199 ptep = NULL;
200
201 folio = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
202 vma, addr, &splug);
203 if (folio)
204 folio_put(folio);
205 }
206
207 if (ptep)
208 pte_unmap_unlock(ptep, ptl);
209 swap_read_unplug(splug);
210 cond_resched();
211
212 return 0;
213}
214
215static const struct mm_walk_ops swapin_walk_ops = {
216 .pmd_entry = swapin_walk_pmd_entry,
217 .walk_lock = PGWALK_RDLOCK,
218};
219
220static void shmem_swapin_range(struct vm_area_struct *vma,
221 unsigned long start, unsigned long end,
222 struct address_space *mapping)
223{
224 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
225 pgoff_t end_index = linear_page_index(vma, end) - 1;
226 struct folio *folio;
227 struct swap_iocb *splug = NULL;
228
229 rcu_read_lock();
230 xas_for_each(&xas, folio, end_index) {
231 unsigned long addr;
232 swp_entry_t entry;
233
234 if (!xa_is_value(folio))
235 continue;
236 entry = radix_to_swp_entry(folio);
237 /* There might be swapin error entries in shmem mapping. */
238 if (non_swap_entry(entry))
239 continue;
240
241 addr = vma->vm_start +
242 ((xas.xa_index - vma->vm_pgoff) << PAGE_SHIFT);
243 xas_pause(&xas);
244 rcu_read_unlock();
245
246 folio = read_swap_cache_async(entry, mapping_gfp_mask(mapping),
247 vma, addr, &splug);
248 if (folio)
249 folio_put(folio);
250
251 rcu_read_lock();
252 }
253 rcu_read_unlock();
254 swap_read_unplug(splug);
255}
256#endif /* CONFIG_SWAP */
257
258/*
259 * Schedule all required I/O operations. Do not wait for completion.
260 */
261static long madvise_willneed(struct vm_area_struct *vma,
262 struct vm_area_struct **prev,
263 unsigned long start, unsigned long end)
264{
265 struct mm_struct *mm = vma->vm_mm;
266 struct file *file = vma->vm_file;
267 loff_t offset;
268
269 *prev = vma;
270#ifdef CONFIG_SWAP
271 if (!file) {
272 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
273 lru_add_drain(); /* Push any new pages onto the LRU now */
274 return 0;
275 }
276
277 if (shmem_mapping(file->f_mapping)) {
278 shmem_swapin_range(vma, start, end, file->f_mapping);
279 lru_add_drain(); /* Push any new pages onto the LRU now */
280 return 0;
281 }
282#else
283 if (!file)
284 return -EBADF;
285#endif
286
287 if (IS_DAX(file_inode(file))) {
288 /* no bad return value, but ignore advice */
289 return 0;
290 }
291
292 /*
293 * Filesystem's fadvise may need to take various locks. We need to
294 * explicitly grab a reference because the vma (and hence the
295 * vma's reference to the file) can go away as soon as we drop
296 * mmap_lock.
297 */
298 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
299 get_file(file);
300 offset = (loff_t)(start - vma->vm_start)
301 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
302 mmap_read_unlock(mm);
303 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
304 fput(file);
305 mmap_read_lock(mm);
306 return 0;
307}
308
309static inline bool can_do_file_pageout(struct vm_area_struct *vma)
310{
311 if (!vma->vm_file)
312 return false;
313 /*
314 * paging out pagecache only for non-anonymous mappings that correspond
315 * to the files the calling process could (if tried) open for writing;
316 * otherwise we'd be including shared non-exclusive mappings, which
317 * opens a side channel.
318 */
319 return inode_owner_or_capable(&nop_mnt_idmap,
320 file_inode(vma->vm_file)) ||
321 file_permission(vma->vm_file, MAY_WRITE) == 0;
322}
323
324static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
325 unsigned long addr, unsigned long end,
326 struct mm_walk *walk)
327{
328 struct madvise_walk_private *private = walk->private;
329 struct mmu_gather *tlb = private->tlb;
330 bool pageout = private->pageout;
331 struct mm_struct *mm = tlb->mm;
332 struct vm_area_struct *vma = walk->vma;
333 pte_t *start_pte, *pte, ptent;
334 spinlock_t *ptl;
335 struct folio *folio = NULL;
336 LIST_HEAD(folio_list);
337 bool pageout_anon_only_filter;
338 unsigned int batch_count = 0;
339
340 if (fatal_signal_pending(current))
341 return -EINTR;
342
343 pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
344 !can_do_file_pageout(vma);
345
346#ifdef CONFIG_TRANSPARENT_HUGEPAGE
347 if (pmd_trans_huge(*pmd)) {
348 pmd_t orig_pmd;
349 unsigned long next = pmd_addr_end(addr, end);
350
351 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
352 ptl = pmd_trans_huge_lock(pmd, vma);
353 if (!ptl)
354 return 0;
355
356 orig_pmd = *pmd;
357 if (is_huge_zero_pmd(orig_pmd))
358 goto huge_unlock;
359
360 if (unlikely(!pmd_present(orig_pmd))) {
361 VM_BUG_ON(thp_migration_supported() &&
362 !is_pmd_migration_entry(orig_pmd));
363 goto huge_unlock;
364 }
365
366 folio = pfn_folio(pmd_pfn(orig_pmd));
367
368 /* Do not interfere with other mappings of this folio */
369 if (folio_estimated_sharers(folio) != 1)
370 goto huge_unlock;
371
372 if (pageout_anon_only_filter && !folio_test_anon(folio))
373 goto huge_unlock;
374
375 if (next - addr != HPAGE_PMD_SIZE) {
376 int err;
377
378 folio_get(folio);
379 spin_unlock(ptl);
380 folio_lock(folio);
381 err = split_folio(folio);
382 folio_unlock(folio);
383 folio_put(folio);
384 if (!err)
385 goto regular_folio;
386 return 0;
387 }
388
389 if (pmd_young(orig_pmd)) {
390 pmdp_invalidate(vma, addr, pmd);
391 orig_pmd = pmd_mkold(orig_pmd);
392
393 set_pmd_at(mm, addr, pmd, orig_pmd);
394 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
395 }
396
397 folio_clear_referenced(folio);
398 folio_test_clear_young(folio);
399 if (folio_test_active(folio))
400 folio_set_workingset(folio);
401 if (pageout) {
402 if (folio_isolate_lru(folio)) {
403 if (folio_test_unevictable(folio))
404 folio_putback_lru(folio);
405 else
406 list_add(&folio->lru, &folio_list);
407 }
408 } else
409 folio_deactivate(folio);
410huge_unlock:
411 spin_unlock(ptl);
412 if (pageout)
413 reclaim_pages(&folio_list);
414 return 0;
415 }
416
417regular_folio:
418#endif
419 tlb_change_page_size(tlb, PAGE_SIZE);
420restart:
421 start_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
422 if (!start_pte)
423 return 0;
424 flush_tlb_batched_pending(mm);
425 arch_enter_lazy_mmu_mode();
426 for (; addr < end; pte++, addr += PAGE_SIZE) {
427 ptent = ptep_get(pte);
428
429 if (++batch_count == SWAP_CLUSTER_MAX) {
430 batch_count = 0;
431 if (need_resched()) {
432 arch_leave_lazy_mmu_mode();
433 pte_unmap_unlock(start_pte, ptl);
434 cond_resched();
435 goto restart;
436 }
437 }
438
439 if (pte_none(ptent))
440 continue;
441
442 if (!pte_present(ptent))
443 continue;
444
445 folio = vm_normal_folio(vma, addr, ptent);
446 if (!folio || folio_is_zone_device(folio))
447 continue;
448
449 /*
450 * Creating a THP page is expensive so split it only if we
451 * are sure it's worth. Split it if we are only owner.
452 */
453 if (folio_test_large(folio)) {
454 int err;
455
456 if (folio_estimated_sharers(folio) != 1)
457 break;
458 if (pageout_anon_only_filter && !folio_test_anon(folio))
459 break;
460 if (!folio_trylock(folio))
461 break;
462 folio_get(folio);
463 arch_leave_lazy_mmu_mode();
464 pte_unmap_unlock(start_pte, ptl);
465 start_pte = NULL;
466 err = split_folio(folio);
467 folio_unlock(folio);
468 folio_put(folio);
469 if (err)
470 break;
471 start_pte = pte =
472 pte_offset_map_lock(mm, pmd, addr, &ptl);
473 if (!start_pte)
474 break;
475 arch_enter_lazy_mmu_mode();
476 pte--;
477 addr -= PAGE_SIZE;
478 continue;
479 }
480
481 /*
482 * Do not interfere with other mappings of this folio and
483 * non-LRU folio.
484 */
485 if (!folio_test_lru(folio) || folio_mapcount(folio) != 1)
486 continue;
487
488 if (pageout_anon_only_filter && !folio_test_anon(folio))
489 continue;
490
491 VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
492
493 if (pte_young(ptent)) {
494 ptent = ptep_get_and_clear_full(mm, addr, pte,
495 tlb->fullmm);
496 ptent = pte_mkold(ptent);
497 set_pte_at(mm, addr, pte, ptent);
498 tlb_remove_tlb_entry(tlb, pte, addr);
499 }
500
501 /*
502 * We are deactivating a folio for accelerating reclaiming.
503 * VM couldn't reclaim the folio unless we clear PG_young.
504 * As a side effect, it makes confuse idle-page tracking
505 * because they will miss recent referenced history.
506 */
507 folio_clear_referenced(folio);
508 folio_test_clear_young(folio);
509 if (folio_test_active(folio))
510 folio_set_workingset(folio);
511 if (pageout) {
512 if (folio_isolate_lru(folio)) {
513 if (folio_test_unevictable(folio))
514 folio_putback_lru(folio);
515 else
516 list_add(&folio->lru, &folio_list);
517 }
518 } else
519 folio_deactivate(folio);
520 }
521
522 if (start_pte) {
523 arch_leave_lazy_mmu_mode();
524 pte_unmap_unlock(start_pte, ptl);
525 }
526 if (pageout)
527 reclaim_pages(&folio_list);
528 cond_resched();
529
530 return 0;
531}
532
533static const struct mm_walk_ops cold_walk_ops = {
534 .pmd_entry = madvise_cold_or_pageout_pte_range,
535 .walk_lock = PGWALK_RDLOCK,
536};
537
538static void madvise_cold_page_range(struct mmu_gather *tlb,
539 struct vm_area_struct *vma,
540 unsigned long addr, unsigned long end)
541{
542 struct madvise_walk_private walk_private = {
543 .pageout = false,
544 .tlb = tlb,
545 };
546
547 tlb_start_vma(tlb, vma);
548 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
549 tlb_end_vma(tlb, vma);
550}
551
552static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
553{
554 return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
555}
556
557static long madvise_cold(struct vm_area_struct *vma,
558 struct vm_area_struct **prev,
559 unsigned long start_addr, unsigned long end_addr)
560{
561 struct mm_struct *mm = vma->vm_mm;
562 struct mmu_gather tlb;
563
564 *prev = vma;
565 if (!can_madv_lru_vma(vma))
566 return -EINVAL;
567
568 lru_add_drain();
569 tlb_gather_mmu(&tlb, mm);
570 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
571 tlb_finish_mmu(&tlb);
572
573 return 0;
574}
575
576static void madvise_pageout_page_range(struct mmu_gather *tlb,
577 struct vm_area_struct *vma,
578 unsigned long addr, unsigned long end)
579{
580 struct madvise_walk_private walk_private = {
581 .pageout = true,
582 .tlb = tlb,
583 };
584
585 tlb_start_vma(tlb, vma);
586 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
587 tlb_end_vma(tlb, vma);
588}
589
590static long madvise_pageout(struct vm_area_struct *vma,
591 struct vm_area_struct **prev,
592 unsigned long start_addr, unsigned long end_addr)
593{
594 struct mm_struct *mm = vma->vm_mm;
595 struct mmu_gather tlb;
596
597 *prev = vma;
598 if (!can_madv_lru_vma(vma))
599 return -EINVAL;
600
601 /*
602 * If the VMA belongs to a private file mapping, there can be private
603 * dirty pages which can be paged out if even this process is neither
604 * owner nor write capable of the file. We allow private file mappings
605 * further to pageout dirty anon pages.
606 */
607 if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
608 (vma->vm_flags & VM_MAYSHARE)))
609 return 0;
610
611 lru_add_drain();
612 tlb_gather_mmu(&tlb, mm);
613 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
614 tlb_finish_mmu(&tlb);
615
616 return 0;
617}
618
619static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
620 unsigned long end, struct mm_walk *walk)
621
622{
623 struct mmu_gather *tlb = walk->private;
624 struct mm_struct *mm = tlb->mm;
625 struct vm_area_struct *vma = walk->vma;
626 spinlock_t *ptl;
627 pte_t *start_pte, *pte, ptent;
628 struct folio *folio;
629 int nr_swap = 0;
630 unsigned long next;
631
632 next = pmd_addr_end(addr, end);
633 if (pmd_trans_huge(*pmd))
634 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
635 return 0;
636
637 tlb_change_page_size(tlb, PAGE_SIZE);
638 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
639 if (!start_pte)
640 return 0;
641 flush_tlb_batched_pending(mm);
642 arch_enter_lazy_mmu_mode();
643 for (; addr != end; pte++, addr += PAGE_SIZE) {
644 ptent = ptep_get(pte);
645
646 if (pte_none(ptent))
647 continue;
648 /*
649 * If the pte has swp_entry, just clear page table to
650 * prevent swap-in which is more expensive rather than
651 * (page allocation + zeroing).
652 */
653 if (!pte_present(ptent)) {
654 swp_entry_t entry;
655
656 entry = pte_to_swp_entry(ptent);
657 if (!non_swap_entry(entry)) {
658 nr_swap--;
659 free_swap_and_cache(entry);
660 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
661 } else if (is_hwpoison_entry(entry) ||
662 is_poisoned_swp_entry(entry)) {
663 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
664 }
665 continue;
666 }
667
668 folio = vm_normal_folio(vma, addr, ptent);
669 if (!folio || folio_is_zone_device(folio))
670 continue;
671
672 /*
673 * If pmd isn't transhuge but the folio is large and
674 * is owned by only this process, split it and
675 * deactivate all pages.
676 */
677 if (folio_test_large(folio)) {
678 int err;
679
680 if (folio_estimated_sharers(folio) != 1)
681 break;
682 if (!folio_trylock(folio))
683 break;
684 folio_get(folio);
685 arch_leave_lazy_mmu_mode();
686 pte_unmap_unlock(start_pte, ptl);
687 start_pte = NULL;
688 err = split_folio(folio);
689 folio_unlock(folio);
690 folio_put(folio);
691 if (err)
692 break;
693 start_pte = pte =
694 pte_offset_map_lock(mm, pmd, addr, &ptl);
695 if (!start_pte)
696 break;
697 arch_enter_lazy_mmu_mode();
698 pte--;
699 addr -= PAGE_SIZE;
700 continue;
701 }
702
703 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
704 if (!folio_trylock(folio))
705 continue;
706 /*
707 * If folio is shared with others, we mustn't clear
708 * the folio's dirty flag.
709 */
710 if (folio_mapcount(folio) != 1) {
711 folio_unlock(folio);
712 continue;
713 }
714
715 if (folio_test_swapcache(folio) &&
716 !folio_free_swap(folio)) {
717 folio_unlock(folio);
718 continue;
719 }
720
721 folio_clear_dirty(folio);
722 folio_unlock(folio);
723 }
724
725 if (pte_young(ptent) || pte_dirty(ptent)) {
726 /*
727 * Some of architecture(ex, PPC) don't update TLB
728 * with set_pte_at and tlb_remove_tlb_entry so for
729 * the portability, remap the pte with old|clean
730 * after pte clearing.
731 */
732 ptent = ptep_get_and_clear_full(mm, addr, pte,
733 tlb->fullmm);
734
735 ptent = pte_mkold(ptent);
736 ptent = pte_mkclean(ptent);
737 set_pte_at(mm, addr, pte, ptent);
738 tlb_remove_tlb_entry(tlb, pte, addr);
739 }
740 folio_mark_lazyfree(folio);
741 }
742
743 if (nr_swap)
744 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
745 if (start_pte) {
746 arch_leave_lazy_mmu_mode();
747 pte_unmap_unlock(start_pte, ptl);
748 }
749 cond_resched();
750
751 return 0;
752}
753
754static const struct mm_walk_ops madvise_free_walk_ops = {
755 .pmd_entry = madvise_free_pte_range,
756 .walk_lock = PGWALK_RDLOCK,
757};
758
759static int madvise_free_single_vma(struct vm_area_struct *vma,
760 unsigned long start_addr, unsigned long end_addr)
761{
762 struct mm_struct *mm = vma->vm_mm;
763 struct mmu_notifier_range range;
764 struct mmu_gather tlb;
765
766 /* MADV_FREE works for only anon vma at the moment */
767 if (!vma_is_anonymous(vma))
768 return -EINVAL;
769
770 range.start = max(vma->vm_start, start_addr);
771 if (range.start >= vma->vm_end)
772 return -EINVAL;
773 range.end = min(vma->vm_end, end_addr);
774 if (range.end <= vma->vm_start)
775 return -EINVAL;
776 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
777 range.start, range.end);
778
779 lru_add_drain();
780 tlb_gather_mmu(&tlb, mm);
781 update_hiwater_rss(mm);
782
783 mmu_notifier_invalidate_range_start(&range);
784 tlb_start_vma(&tlb, vma);
785 walk_page_range(vma->vm_mm, range.start, range.end,
786 &madvise_free_walk_ops, &tlb);
787 tlb_end_vma(&tlb, vma);
788 mmu_notifier_invalidate_range_end(&range);
789 tlb_finish_mmu(&tlb);
790
791 return 0;
792}
793
794/*
795 * Application no longer needs these pages. If the pages are dirty,
796 * it's OK to just throw them away. The app will be more careful about
797 * data it wants to keep. Be sure to free swap resources too. The
798 * zap_page_range_single call sets things up for shrink_active_list to actually
799 * free these pages later if no one else has touched them in the meantime,
800 * although we could add these pages to a global reuse list for
801 * shrink_active_list to pick up before reclaiming other pages.
802 *
803 * NB: This interface discards data rather than pushes it out to swap,
804 * as some implementations do. This has performance implications for
805 * applications like large transactional databases which want to discard
806 * pages in anonymous maps after committing to backing store the data
807 * that was kept in them. There is no reason to write this data out to
808 * the swap area if the application is discarding it.
809 *
810 * An interface that causes the system to free clean pages and flush
811 * dirty pages is already available as msync(MS_INVALIDATE).
812 */
813static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
814 unsigned long start, unsigned long end)
815{
816 zap_page_range_single(vma, start, end - start, NULL);
817 return 0;
818}
819
820static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
821 unsigned long start,
822 unsigned long *end,
823 int behavior)
824{
825 if (!is_vm_hugetlb_page(vma)) {
826 unsigned int forbidden = VM_PFNMAP;
827
828 if (behavior != MADV_DONTNEED_LOCKED)
829 forbidden |= VM_LOCKED;
830
831 return !(vma->vm_flags & forbidden);
832 }
833
834 if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
835 return false;
836 if (start & ~huge_page_mask(hstate_vma(vma)))
837 return false;
838
839 /*
840 * Madvise callers expect the length to be rounded up to PAGE_SIZE
841 * boundaries, and may be unaware that this VMA uses huge pages.
842 * Avoid unexpected data loss by rounding down the number of
843 * huge pages freed.
844 */
845 *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
846
847 return true;
848}
849
850static long madvise_dontneed_free(struct vm_area_struct *vma,
851 struct vm_area_struct **prev,
852 unsigned long start, unsigned long end,
853 int behavior)
854{
855 struct mm_struct *mm = vma->vm_mm;
856
857 *prev = vma;
858 if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
859 return -EINVAL;
860
861 if (start == end)
862 return 0;
863
864 if (!userfaultfd_remove(vma, start, end)) {
865 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
866
867 mmap_read_lock(mm);
868 vma = vma_lookup(mm, start);
869 if (!vma)
870 return -ENOMEM;
871 /*
872 * Potential end adjustment for hugetlb vma is OK as
873 * the check below keeps end within vma.
874 */
875 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
876 behavior))
877 return -EINVAL;
878 if (end > vma->vm_end) {
879 /*
880 * Don't fail if end > vma->vm_end. If the old
881 * vma was split while the mmap_lock was
882 * released the effect of the concurrent
883 * operation may not cause madvise() to
884 * have an undefined result. There may be an
885 * adjacent next vma that we'll walk
886 * next. userfaultfd_remove() will generate an
887 * UFFD_EVENT_REMOVE repetition on the
888 * end-vma->vm_end range, but the manager can
889 * handle a repetition fine.
890 */
891 end = vma->vm_end;
892 }
893 VM_WARN_ON(start >= end);
894 }
895
896 if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
897 return madvise_dontneed_single_vma(vma, start, end);
898 else if (behavior == MADV_FREE)
899 return madvise_free_single_vma(vma, start, end);
900 else
901 return -EINVAL;
902}
903
904static long madvise_populate(struct vm_area_struct *vma,
905 struct vm_area_struct **prev,
906 unsigned long start, unsigned long end,
907 int behavior)
908{
909 const bool write = behavior == MADV_POPULATE_WRITE;
910 struct mm_struct *mm = vma->vm_mm;
911 unsigned long tmp_end;
912 int locked = 1;
913 long pages;
914
915 *prev = vma;
916
917 while (start < end) {
918 /*
919 * We might have temporarily dropped the lock. For example,
920 * our VMA might have been split.
921 */
922 if (!vma || start >= vma->vm_end) {
923 vma = vma_lookup(mm, start);
924 if (!vma)
925 return -ENOMEM;
926 }
927
928 tmp_end = min_t(unsigned long, end, vma->vm_end);
929 /* Populate (prefault) page tables readable/writable. */
930 pages = faultin_vma_page_range(vma, start, tmp_end, write,
931 &locked);
932 if (!locked) {
933 mmap_read_lock(mm);
934 locked = 1;
935 *prev = NULL;
936 vma = NULL;
937 }
938 if (pages < 0) {
939 switch (pages) {
940 case -EINTR:
941 return -EINTR;
942 case -EINVAL: /* Incompatible mappings / permissions. */
943 return -EINVAL;
944 case -EHWPOISON:
945 return -EHWPOISON;
946 case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
947 return -EFAULT;
948 default:
949 pr_warn_once("%s: unhandled return value: %ld\n",
950 __func__, pages);
951 fallthrough;
952 case -ENOMEM:
953 return -ENOMEM;
954 }
955 }
956 start += pages * PAGE_SIZE;
957 }
958 return 0;
959}
960
961/*
962 * Application wants to free up the pages and associated backing store.
963 * This is effectively punching a hole into the middle of a file.
964 */
965static long madvise_remove(struct vm_area_struct *vma,
966 struct vm_area_struct **prev,
967 unsigned long start, unsigned long end)
968{
969 loff_t offset;
970 int error;
971 struct file *f;
972 struct mm_struct *mm = vma->vm_mm;
973
974 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
975
976 if (vma->vm_flags & VM_LOCKED)
977 return -EINVAL;
978
979 f = vma->vm_file;
980
981 if (!f || !f->f_mapping || !f->f_mapping->host) {
982 return -EINVAL;
983 }
984
985 if (!vma_is_shared_maywrite(vma))
986 return -EACCES;
987
988 offset = (loff_t)(start - vma->vm_start)
989 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
990
991 /*
992 * Filesystem's fallocate may need to take i_rwsem. We need to
993 * explicitly grab a reference because the vma (and hence the
994 * vma's reference to the file) can go away as soon as we drop
995 * mmap_lock.
996 */
997 get_file(f);
998 if (userfaultfd_remove(vma, start, end)) {
999 /* mmap_lock was not released by userfaultfd_remove() */
1000 mmap_read_unlock(mm);
1001 }
1002 error = vfs_fallocate(f,
1003 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1004 offset, end - start);
1005 fput(f);
1006 mmap_read_lock(mm);
1007 return error;
1008}
1009
1010/*
1011 * Apply an madvise behavior to a region of a vma. madvise_update_vma
1012 * will handle splitting a vm area into separate areas, each area with its own
1013 * behavior.
1014 */
1015static int madvise_vma_behavior(struct vm_area_struct *vma,
1016 struct vm_area_struct **prev,
1017 unsigned long start, unsigned long end,
1018 unsigned long behavior)
1019{
1020 int error;
1021 struct anon_vma_name *anon_name;
1022 unsigned long new_flags = vma->vm_flags;
1023
1024 switch (behavior) {
1025 case MADV_REMOVE:
1026 return madvise_remove(vma, prev, start, end);
1027 case MADV_WILLNEED:
1028 return madvise_willneed(vma, prev, start, end);
1029 case MADV_COLD:
1030 return madvise_cold(vma, prev, start, end);
1031 case MADV_PAGEOUT:
1032 return madvise_pageout(vma, prev, start, end);
1033 case MADV_FREE:
1034 case MADV_DONTNEED:
1035 case MADV_DONTNEED_LOCKED:
1036 return madvise_dontneed_free(vma, prev, start, end, behavior);
1037 case MADV_POPULATE_READ:
1038 case MADV_POPULATE_WRITE:
1039 return madvise_populate(vma, prev, start, end, behavior);
1040 case MADV_NORMAL:
1041 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1042 break;
1043 case MADV_SEQUENTIAL:
1044 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1045 break;
1046 case MADV_RANDOM:
1047 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1048 break;
1049 case MADV_DONTFORK:
1050 new_flags |= VM_DONTCOPY;
1051 break;
1052 case MADV_DOFORK:
1053 if (vma->vm_flags & VM_IO)
1054 return -EINVAL;
1055 new_flags &= ~VM_DONTCOPY;
1056 break;
1057 case MADV_WIPEONFORK:
1058 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1059 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1060 return -EINVAL;
1061 new_flags |= VM_WIPEONFORK;
1062 break;
1063 case MADV_KEEPONFORK:
1064 new_flags &= ~VM_WIPEONFORK;
1065 break;
1066 case MADV_DONTDUMP:
1067 new_flags |= VM_DONTDUMP;
1068 break;
1069 case MADV_DODUMP:
1070 if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1071 return -EINVAL;
1072 new_flags &= ~VM_DONTDUMP;
1073 break;
1074 case MADV_MERGEABLE:
1075 case MADV_UNMERGEABLE:
1076 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1077 if (error)
1078 goto out;
1079 break;
1080 case MADV_HUGEPAGE:
1081 case MADV_NOHUGEPAGE:
1082 error = hugepage_madvise(vma, &new_flags, behavior);
1083 if (error)
1084 goto out;
1085 break;
1086 case MADV_COLLAPSE:
1087 return madvise_collapse(vma, prev, start, end);
1088 }
1089
1090 anon_name = anon_vma_name(vma);
1091 anon_vma_name_get(anon_name);
1092 error = madvise_update_vma(vma, prev, start, end, new_flags,
1093 anon_name);
1094 anon_vma_name_put(anon_name);
1095
1096out:
1097 /*
1098 * madvise() returns EAGAIN if kernel resources, such as
1099 * slab, are temporarily unavailable.
1100 */
1101 if (error == -ENOMEM)
1102 error = -EAGAIN;
1103 return error;
1104}
1105
1106#ifdef CONFIG_MEMORY_FAILURE
1107/*
1108 * Error injection support for memory error handling.
1109 */
1110static int madvise_inject_error(int behavior,
1111 unsigned long start, unsigned long end)
1112{
1113 unsigned long size;
1114
1115 if (!capable(CAP_SYS_ADMIN))
1116 return -EPERM;
1117
1118
1119 for (; start < end; start += size) {
1120 unsigned long pfn;
1121 struct page *page;
1122 int ret;
1123
1124 ret = get_user_pages_fast(start, 1, 0, &page);
1125 if (ret != 1)
1126 return ret;
1127 pfn = page_to_pfn(page);
1128
1129 /*
1130 * When soft offlining hugepages, after migrating the page
1131 * we dissolve it, therefore in the second loop "page" will
1132 * no longer be a compound page.
1133 */
1134 size = page_size(compound_head(page));
1135
1136 if (behavior == MADV_SOFT_OFFLINE) {
1137 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1138 pfn, start);
1139 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1140 } else {
1141 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1142 pfn, start);
1143 ret = memory_failure(pfn, MF_COUNT_INCREASED | MF_SW_SIMULATED);
1144 if (ret == -EOPNOTSUPP)
1145 ret = 0;
1146 }
1147
1148 if (ret)
1149 return ret;
1150 }
1151
1152 return 0;
1153}
1154#endif
1155
1156static bool
1157madvise_behavior_valid(int behavior)
1158{
1159 switch (behavior) {
1160 case MADV_DOFORK:
1161 case MADV_DONTFORK:
1162 case MADV_NORMAL:
1163 case MADV_SEQUENTIAL:
1164 case MADV_RANDOM:
1165 case MADV_REMOVE:
1166 case MADV_WILLNEED:
1167 case MADV_DONTNEED:
1168 case MADV_DONTNEED_LOCKED:
1169 case MADV_FREE:
1170 case MADV_COLD:
1171 case MADV_PAGEOUT:
1172 case MADV_POPULATE_READ:
1173 case MADV_POPULATE_WRITE:
1174#ifdef CONFIG_KSM
1175 case MADV_MERGEABLE:
1176 case MADV_UNMERGEABLE:
1177#endif
1178#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1179 case MADV_HUGEPAGE:
1180 case MADV_NOHUGEPAGE:
1181 case MADV_COLLAPSE:
1182#endif
1183 case MADV_DONTDUMP:
1184 case MADV_DODUMP:
1185 case MADV_WIPEONFORK:
1186 case MADV_KEEPONFORK:
1187#ifdef CONFIG_MEMORY_FAILURE
1188 case MADV_SOFT_OFFLINE:
1189 case MADV_HWPOISON:
1190#endif
1191 return true;
1192
1193 default:
1194 return false;
1195 }
1196}
1197
1198static bool process_madvise_behavior_valid(int behavior)
1199{
1200 switch (behavior) {
1201 case MADV_COLD:
1202 case MADV_PAGEOUT:
1203 case MADV_WILLNEED:
1204 case MADV_COLLAPSE:
1205 return true;
1206 default:
1207 return false;
1208 }
1209}
1210
1211/*
1212 * Walk the vmas in range [start,end), and call the visit function on each one.
1213 * The visit function will get start and end parameters that cover the overlap
1214 * between the current vma and the original range. Any unmapped regions in the
1215 * original range will result in this function returning -ENOMEM while still
1216 * calling the visit function on all of the existing vmas in the range.
1217 * Must be called with the mmap_lock held for reading or writing.
1218 */
1219static
1220int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1221 unsigned long end, unsigned long arg,
1222 int (*visit)(struct vm_area_struct *vma,
1223 struct vm_area_struct **prev, unsigned long start,
1224 unsigned long end, unsigned long arg))
1225{
1226 struct vm_area_struct *vma;
1227 struct vm_area_struct *prev;
1228 unsigned long tmp;
1229 int unmapped_error = 0;
1230
1231 /*
1232 * If the interval [start,end) covers some unmapped address
1233 * ranges, just ignore them, but return -ENOMEM at the end.
1234 * - different from the way of handling in mlock etc.
1235 */
1236 vma = find_vma_prev(mm, start, &prev);
1237 if (vma && start > vma->vm_start)
1238 prev = vma;
1239
1240 for (;;) {
1241 int error;
1242
1243 /* Still start < end. */
1244 if (!vma)
1245 return -ENOMEM;
1246
1247 /* Here start < (end|vma->vm_end). */
1248 if (start < vma->vm_start) {
1249 unmapped_error = -ENOMEM;
1250 start = vma->vm_start;
1251 if (start >= end)
1252 break;
1253 }
1254
1255 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1256 tmp = vma->vm_end;
1257 if (end < tmp)
1258 tmp = end;
1259
1260 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1261 error = visit(vma, &prev, start, tmp, arg);
1262 if (error)
1263 return error;
1264 start = tmp;
1265 if (prev && start < prev->vm_end)
1266 start = prev->vm_end;
1267 if (start >= end)
1268 break;
1269 if (prev)
1270 vma = find_vma(mm, prev->vm_end);
1271 else /* madvise_remove dropped mmap_lock */
1272 vma = find_vma(mm, start);
1273 }
1274
1275 return unmapped_error;
1276}
1277
1278#ifdef CONFIG_ANON_VMA_NAME
1279static int madvise_vma_anon_name(struct vm_area_struct *vma,
1280 struct vm_area_struct **prev,
1281 unsigned long start, unsigned long end,
1282 unsigned long anon_name)
1283{
1284 int error;
1285
1286 /* Only anonymous mappings can be named */
1287 if (vma->vm_file && !vma_is_anon_shmem(vma))
1288 return -EBADF;
1289
1290 error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1291 (struct anon_vma_name *)anon_name);
1292
1293 /*
1294 * madvise() returns EAGAIN if kernel resources, such as
1295 * slab, are temporarily unavailable.
1296 */
1297 if (error == -ENOMEM)
1298 error = -EAGAIN;
1299 return error;
1300}
1301
1302int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1303 unsigned long len_in, struct anon_vma_name *anon_name)
1304{
1305 unsigned long end;
1306 unsigned long len;
1307
1308 if (start & ~PAGE_MASK)
1309 return -EINVAL;
1310 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1311
1312 /* Check to see whether len was rounded up from small -ve to zero */
1313 if (len_in && !len)
1314 return -EINVAL;
1315
1316 end = start + len;
1317 if (end < start)
1318 return -EINVAL;
1319
1320 if (end == start)
1321 return 0;
1322
1323 return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1324 madvise_vma_anon_name);
1325}
1326#endif /* CONFIG_ANON_VMA_NAME */
1327/*
1328 * The madvise(2) system call.
1329 *
1330 * Applications can use madvise() to advise the kernel how it should
1331 * handle paging I/O in this VM area. The idea is to help the kernel
1332 * use appropriate read-ahead and caching techniques. The information
1333 * provided is advisory only, and can be safely disregarded by the
1334 * kernel without affecting the correct operation of the application.
1335 *
1336 * behavior values:
1337 * MADV_NORMAL - the default behavior is to read clusters. This
1338 * results in some read-ahead and read-behind.
1339 * MADV_RANDOM - the system should read the minimum amount of data
1340 * on any access, since it is unlikely that the appli-
1341 * cation will need more than what it asks for.
1342 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1343 * once, so they can be aggressively read ahead, and
1344 * can be freed soon after they are accessed.
1345 * MADV_WILLNEED - the application is notifying the system to read
1346 * some pages ahead.
1347 * MADV_DONTNEED - the application is finished with the given range,
1348 * so the kernel can free resources associated with it.
1349 * MADV_FREE - the application marks pages in the given range as lazy free,
1350 * where actual purges are postponed until memory pressure happens.
1351 * MADV_REMOVE - the application wants to free up the given range of
1352 * pages and associated backing store.
1353 * MADV_DONTFORK - omit this area from child's address space when forking:
1354 * typically, to avoid COWing pages pinned by get_user_pages().
1355 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1356 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1357 * range after a fork.
1358 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1359 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1360 * were corrupted by unrecoverable hardware memory failure.
1361 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1362 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1363 * this area with pages of identical content from other such areas.
1364 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1365 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1366 * huge pages in the future. Existing pages might be coalesced and
1367 * new pages might be allocated as THP.
1368 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1369 * transparent huge pages so the existing pages will not be
1370 * coalesced into THP and new pages will not be allocated as THP.
1371 * MADV_COLLAPSE - synchronously coalesce pages into new THP.
1372 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1373 * from being included in its core dump.
1374 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1375 * MADV_COLD - the application is not expected to use this memory soon,
1376 * deactivate pages in this range so that they can be reclaimed
1377 * easily if memory pressure happens.
1378 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1379 * page out the pages in this range immediately.
1380 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1381 * triggering read faults if required
1382 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1383 * triggering write faults if required
1384 *
1385 * return values:
1386 * zero - success
1387 * -EINVAL - start + len < 0, start is not page-aligned,
1388 * "behavior" is not a valid value, or application
1389 * is attempting to release locked or shared pages,
1390 * or the specified address range includes file, Huge TLB,
1391 * MAP_SHARED or VMPFNMAP range.
1392 * -ENOMEM - addresses in the specified range are not currently
1393 * mapped, or are outside the AS of the process.
1394 * -EIO - an I/O error occurred while paging in data.
1395 * -EBADF - map exists, but area maps something that isn't a file.
1396 * -EAGAIN - a kernel resource was temporarily unavailable.
1397 */
1398int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1399{
1400 unsigned long end;
1401 int error;
1402 int write;
1403 size_t len;
1404 struct blk_plug plug;
1405
1406 if (!madvise_behavior_valid(behavior))
1407 return -EINVAL;
1408
1409 if (!PAGE_ALIGNED(start))
1410 return -EINVAL;
1411 len = PAGE_ALIGN(len_in);
1412
1413 /* Check to see whether len was rounded up from small -ve to zero */
1414 if (len_in && !len)
1415 return -EINVAL;
1416
1417 end = start + len;
1418 if (end < start)
1419 return -EINVAL;
1420
1421 if (end == start)
1422 return 0;
1423
1424#ifdef CONFIG_MEMORY_FAILURE
1425 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1426 return madvise_inject_error(behavior, start, start + len_in);
1427#endif
1428
1429 write = madvise_need_mmap_write(behavior);
1430 if (write) {
1431 if (mmap_write_lock_killable(mm))
1432 return -EINTR;
1433 } else {
1434 mmap_read_lock(mm);
1435 }
1436
1437 start = untagged_addr_remote(mm, start);
1438 end = start + len;
1439
1440 blk_start_plug(&plug);
1441 error = madvise_walk_vmas(mm, start, end, behavior,
1442 madvise_vma_behavior);
1443 blk_finish_plug(&plug);
1444 if (write)
1445 mmap_write_unlock(mm);
1446 else
1447 mmap_read_unlock(mm);
1448
1449 return error;
1450}
1451
1452SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1453{
1454 return do_madvise(current->mm, start, len_in, behavior);
1455}
1456
1457SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1458 size_t, vlen, int, behavior, unsigned int, flags)
1459{
1460 ssize_t ret;
1461 struct iovec iovstack[UIO_FASTIOV];
1462 struct iovec *iov = iovstack;
1463 struct iov_iter iter;
1464 struct task_struct *task;
1465 struct mm_struct *mm;
1466 size_t total_len;
1467 unsigned int f_flags;
1468
1469 if (flags != 0) {
1470 ret = -EINVAL;
1471 goto out;
1472 }
1473
1474 ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1475 if (ret < 0)
1476 goto out;
1477
1478 task = pidfd_get_task(pidfd, &f_flags);
1479 if (IS_ERR(task)) {
1480 ret = PTR_ERR(task);
1481 goto free_iov;
1482 }
1483
1484 if (!process_madvise_behavior_valid(behavior)) {
1485 ret = -EINVAL;
1486 goto release_task;
1487 }
1488
1489 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1490 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1491 if (IS_ERR_OR_NULL(mm)) {
1492 ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1493 goto release_task;
1494 }
1495
1496 /*
1497 * Require CAP_SYS_NICE for influencing process performance. Note that
1498 * only non-destructive hints are currently supported.
1499 */
1500 if (!capable(CAP_SYS_NICE)) {
1501 ret = -EPERM;
1502 goto release_mm;
1503 }
1504
1505 total_len = iov_iter_count(&iter);
1506
1507 while (iov_iter_count(&iter)) {
1508 ret = do_madvise(mm, (unsigned long)iter_iov_addr(&iter),
1509 iter_iov_len(&iter), behavior);
1510 if (ret < 0)
1511 break;
1512 iov_iter_advance(&iter, iter_iov_len(&iter));
1513 }
1514
1515 ret = (total_len - iov_iter_count(&iter)) ? : ret;
1516
1517release_mm:
1518 mmput(mm);
1519release_task:
1520 put_task_struct(task);
1521free_iov:
1522 kfree(iov);
1523out:
1524 return ret;
1525}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/madvise.c
4 *
5 * Copyright (C) 1999 Linus Torvalds
6 * Copyright (C) 2002 Christoph Hellwig
7 */
8
9#include <linux/mman.h>
10#include <linux/pagemap.h>
11#include <linux/syscalls.h>
12#include <linux/mempolicy.h>
13#include <linux/page-isolation.h>
14#include <linux/page_idle.h>
15#include <linux/userfaultfd_k.h>
16#include <linux/hugetlb.h>
17#include <linux/falloc.h>
18#include <linux/fadvise.h>
19#include <linux/sched.h>
20#include <linux/sched/mm.h>
21#include <linux/mm_inline.h>
22#include <linux/string.h>
23#include <linux/uio.h>
24#include <linux/ksm.h>
25#include <linux/fs.h>
26#include <linux/file.h>
27#include <linux/blkdev.h>
28#include <linux/backing-dev.h>
29#include <linux/pagewalk.h>
30#include <linux/swap.h>
31#include <linux/swapops.h>
32#include <linux/shmem_fs.h>
33#include <linux/mmu_notifier.h>
34
35#include <asm/tlb.h>
36
37#include "internal.h"
38#include "swap.h"
39
40/*
41 * Maximum number of attempts we make to install guard pages before we give up
42 * and return -ERESTARTNOINTR to have userspace try again.
43 */
44#define MAX_MADVISE_GUARD_RETRIES 3
45
46struct madvise_walk_private {
47 struct mmu_gather *tlb;
48 bool pageout;
49};
50
51/*
52 * Any behaviour which results in changes to the vma->vm_flags needs to
53 * take mmap_lock for writing. Others, which simply traverse vmas, need
54 * to only take it for reading.
55 */
56static int madvise_need_mmap_write(int behavior)
57{
58 switch (behavior) {
59 case MADV_REMOVE:
60 case MADV_WILLNEED:
61 case MADV_DONTNEED:
62 case MADV_DONTNEED_LOCKED:
63 case MADV_COLD:
64 case MADV_PAGEOUT:
65 case MADV_FREE:
66 case MADV_POPULATE_READ:
67 case MADV_POPULATE_WRITE:
68 case MADV_COLLAPSE:
69 case MADV_GUARD_INSTALL:
70 case MADV_GUARD_REMOVE:
71 return 0;
72 default:
73 /* be safe, default to 1. list exceptions explicitly */
74 return 1;
75 }
76}
77
78#ifdef CONFIG_ANON_VMA_NAME
79struct anon_vma_name *anon_vma_name_alloc(const char *name)
80{
81 struct anon_vma_name *anon_name;
82 size_t count;
83
84 /* Add 1 for NUL terminator at the end of the anon_name->name */
85 count = strlen(name) + 1;
86 anon_name = kmalloc(struct_size(anon_name, name, count), GFP_KERNEL);
87 if (anon_name) {
88 kref_init(&anon_name->kref);
89 memcpy(anon_name->name, name, count);
90 }
91
92 return anon_name;
93}
94
95void anon_vma_name_free(struct kref *kref)
96{
97 struct anon_vma_name *anon_name =
98 container_of(kref, struct anon_vma_name, kref);
99 kfree(anon_name);
100}
101
102struct anon_vma_name *anon_vma_name(struct vm_area_struct *vma)
103{
104 mmap_assert_locked(vma->vm_mm);
105
106 return vma->anon_name;
107}
108
109/* mmap_lock should be write-locked */
110static int replace_anon_vma_name(struct vm_area_struct *vma,
111 struct anon_vma_name *anon_name)
112{
113 struct anon_vma_name *orig_name = anon_vma_name(vma);
114
115 if (!anon_name) {
116 vma->anon_name = NULL;
117 anon_vma_name_put(orig_name);
118 return 0;
119 }
120
121 if (anon_vma_name_eq(orig_name, anon_name))
122 return 0;
123
124 vma->anon_name = anon_vma_name_reuse(anon_name);
125 anon_vma_name_put(orig_name);
126
127 return 0;
128}
129#else /* CONFIG_ANON_VMA_NAME */
130static int replace_anon_vma_name(struct vm_area_struct *vma,
131 struct anon_vma_name *anon_name)
132{
133 if (anon_name)
134 return -EINVAL;
135
136 return 0;
137}
138#endif /* CONFIG_ANON_VMA_NAME */
139/*
140 * Update the vm_flags on region of a vma, splitting it or merging it as
141 * necessary. Must be called with mmap_lock held for writing;
142 * Caller should ensure anon_name stability by raising its refcount even when
143 * anon_name belongs to a valid vma because this function might free that vma.
144 */
145static int madvise_update_vma(struct vm_area_struct *vma,
146 struct vm_area_struct **prev, unsigned long start,
147 unsigned long end, unsigned long new_flags,
148 struct anon_vma_name *anon_name)
149{
150 struct mm_struct *mm = vma->vm_mm;
151 int error;
152 VMA_ITERATOR(vmi, mm, start);
153
154 if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
155 *prev = vma;
156 return 0;
157 }
158
159 vma = vma_modify_flags_name(&vmi, *prev, vma, start, end, new_flags,
160 anon_name);
161 if (IS_ERR(vma))
162 return PTR_ERR(vma);
163
164 *prev = vma;
165
166 /* vm_flags is protected by the mmap_lock held in write mode. */
167 vma_start_write(vma);
168 vm_flags_reset(vma, new_flags);
169 if (!vma->vm_file || vma_is_anon_shmem(vma)) {
170 error = replace_anon_vma_name(vma, anon_name);
171 if (error)
172 return error;
173 }
174
175 return 0;
176}
177
178#ifdef CONFIG_SWAP
179static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
180 unsigned long end, struct mm_walk *walk)
181{
182 struct vm_area_struct *vma = walk->private;
183 struct swap_iocb *splug = NULL;
184 pte_t *ptep = NULL;
185 spinlock_t *ptl;
186 unsigned long addr;
187
188 for (addr = start; addr < end; addr += PAGE_SIZE) {
189 pte_t pte;
190 swp_entry_t entry;
191 struct folio *folio;
192
193 if (!ptep++) {
194 ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
195 if (!ptep)
196 break;
197 }
198
199 pte = ptep_get(ptep);
200 if (!is_swap_pte(pte))
201 continue;
202 entry = pte_to_swp_entry(pte);
203 if (unlikely(non_swap_entry(entry)))
204 continue;
205
206 pte_unmap_unlock(ptep, ptl);
207 ptep = NULL;
208
209 folio = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
210 vma, addr, &splug);
211 if (folio)
212 folio_put(folio);
213 }
214
215 if (ptep)
216 pte_unmap_unlock(ptep, ptl);
217 swap_read_unplug(splug);
218 cond_resched();
219
220 return 0;
221}
222
223static const struct mm_walk_ops swapin_walk_ops = {
224 .pmd_entry = swapin_walk_pmd_entry,
225 .walk_lock = PGWALK_RDLOCK,
226};
227
228static void shmem_swapin_range(struct vm_area_struct *vma,
229 unsigned long start, unsigned long end,
230 struct address_space *mapping)
231{
232 XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
233 pgoff_t end_index = linear_page_index(vma, end) - 1;
234 struct folio *folio;
235 struct swap_iocb *splug = NULL;
236
237 rcu_read_lock();
238 xas_for_each(&xas, folio, end_index) {
239 unsigned long addr;
240 swp_entry_t entry;
241
242 if (!xa_is_value(folio))
243 continue;
244 entry = radix_to_swp_entry(folio);
245 /* There might be swapin error entries in shmem mapping. */
246 if (non_swap_entry(entry))
247 continue;
248
249 addr = vma->vm_start +
250 ((xas.xa_index - vma->vm_pgoff) << PAGE_SHIFT);
251 xas_pause(&xas);
252 rcu_read_unlock();
253
254 folio = read_swap_cache_async(entry, mapping_gfp_mask(mapping),
255 vma, addr, &splug);
256 if (folio)
257 folio_put(folio);
258
259 rcu_read_lock();
260 }
261 rcu_read_unlock();
262 swap_read_unplug(splug);
263}
264#endif /* CONFIG_SWAP */
265
266/*
267 * Schedule all required I/O operations. Do not wait for completion.
268 */
269static long madvise_willneed(struct vm_area_struct *vma,
270 struct vm_area_struct **prev,
271 unsigned long start, unsigned long end)
272{
273 struct mm_struct *mm = vma->vm_mm;
274 struct file *file = vma->vm_file;
275 loff_t offset;
276
277 *prev = vma;
278#ifdef CONFIG_SWAP
279 if (!file) {
280 walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
281 lru_add_drain(); /* Push any new pages onto the LRU now */
282 return 0;
283 }
284
285 if (shmem_mapping(file->f_mapping)) {
286 shmem_swapin_range(vma, start, end, file->f_mapping);
287 lru_add_drain(); /* Push any new pages onto the LRU now */
288 return 0;
289 }
290#else
291 if (!file)
292 return -EBADF;
293#endif
294
295 if (IS_DAX(file_inode(file))) {
296 /* no bad return value, but ignore advice */
297 return 0;
298 }
299
300 /*
301 * Filesystem's fadvise may need to take various locks. We need to
302 * explicitly grab a reference because the vma (and hence the
303 * vma's reference to the file) can go away as soon as we drop
304 * mmap_lock.
305 */
306 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
307 get_file(file);
308 offset = (loff_t)(start - vma->vm_start)
309 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
310 mmap_read_unlock(mm);
311 vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
312 fput(file);
313 mmap_read_lock(mm);
314 return 0;
315}
316
317static inline bool can_do_file_pageout(struct vm_area_struct *vma)
318{
319 if (!vma->vm_file)
320 return false;
321 /*
322 * paging out pagecache only for non-anonymous mappings that correspond
323 * to the files the calling process could (if tried) open for writing;
324 * otherwise we'd be including shared non-exclusive mappings, which
325 * opens a side channel.
326 */
327 return inode_owner_or_capable(&nop_mnt_idmap,
328 file_inode(vma->vm_file)) ||
329 file_permission(vma->vm_file, MAY_WRITE) == 0;
330}
331
332static inline int madvise_folio_pte_batch(unsigned long addr, unsigned long end,
333 struct folio *folio, pte_t *ptep,
334 pte_t pte, bool *any_young,
335 bool *any_dirty)
336{
337 const fpb_t fpb_flags = FPB_IGNORE_DIRTY | FPB_IGNORE_SOFT_DIRTY;
338 int max_nr = (end - addr) / PAGE_SIZE;
339
340 return folio_pte_batch(folio, addr, ptep, pte, max_nr, fpb_flags, NULL,
341 any_young, any_dirty);
342}
343
344static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
345 unsigned long addr, unsigned long end,
346 struct mm_walk *walk)
347{
348 struct madvise_walk_private *private = walk->private;
349 struct mmu_gather *tlb = private->tlb;
350 bool pageout = private->pageout;
351 struct mm_struct *mm = tlb->mm;
352 struct vm_area_struct *vma = walk->vma;
353 pte_t *start_pte, *pte, ptent;
354 spinlock_t *ptl;
355 struct folio *folio = NULL;
356 LIST_HEAD(folio_list);
357 bool pageout_anon_only_filter;
358 unsigned int batch_count = 0;
359 int nr;
360
361 if (fatal_signal_pending(current))
362 return -EINTR;
363
364 pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
365 !can_do_file_pageout(vma);
366
367#ifdef CONFIG_TRANSPARENT_HUGEPAGE
368 if (pmd_trans_huge(*pmd)) {
369 pmd_t orig_pmd;
370 unsigned long next = pmd_addr_end(addr, end);
371
372 tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
373 ptl = pmd_trans_huge_lock(pmd, vma);
374 if (!ptl)
375 return 0;
376
377 orig_pmd = *pmd;
378 if (is_huge_zero_pmd(orig_pmd))
379 goto huge_unlock;
380
381 if (unlikely(!pmd_present(orig_pmd))) {
382 VM_BUG_ON(thp_migration_supported() &&
383 !is_pmd_migration_entry(orig_pmd));
384 goto huge_unlock;
385 }
386
387 folio = pmd_folio(orig_pmd);
388
389 /* Do not interfere with other mappings of this folio */
390 if (folio_likely_mapped_shared(folio))
391 goto huge_unlock;
392
393 if (pageout_anon_only_filter && !folio_test_anon(folio))
394 goto huge_unlock;
395
396 if (next - addr != HPAGE_PMD_SIZE) {
397 int err;
398
399 folio_get(folio);
400 spin_unlock(ptl);
401 folio_lock(folio);
402 err = split_folio(folio);
403 folio_unlock(folio);
404 folio_put(folio);
405 if (!err)
406 goto regular_folio;
407 return 0;
408 }
409
410 if (!pageout && pmd_young(orig_pmd)) {
411 pmdp_invalidate(vma, addr, pmd);
412 orig_pmd = pmd_mkold(orig_pmd);
413
414 set_pmd_at(mm, addr, pmd, orig_pmd);
415 tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
416 }
417
418 folio_clear_referenced(folio);
419 folio_test_clear_young(folio);
420 if (folio_test_active(folio))
421 folio_set_workingset(folio);
422 if (pageout) {
423 if (folio_isolate_lru(folio)) {
424 if (folio_test_unevictable(folio))
425 folio_putback_lru(folio);
426 else
427 list_add(&folio->lru, &folio_list);
428 }
429 } else
430 folio_deactivate(folio);
431huge_unlock:
432 spin_unlock(ptl);
433 if (pageout)
434 reclaim_pages(&folio_list);
435 return 0;
436 }
437
438regular_folio:
439#endif
440 tlb_change_page_size(tlb, PAGE_SIZE);
441restart:
442 start_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
443 if (!start_pte)
444 return 0;
445 flush_tlb_batched_pending(mm);
446 arch_enter_lazy_mmu_mode();
447 for (; addr < end; pte += nr, addr += nr * PAGE_SIZE) {
448 nr = 1;
449 ptent = ptep_get(pte);
450
451 if (++batch_count == SWAP_CLUSTER_MAX) {
452 batch_count = 0;
453 if (need_resched()) {
454 arch_leave_lazy_mmu_mode();
455 pte_unmap_unlock(start_pte, ptl);
456 cond_resched();
457 goto restart;
458 }
459 }
460
461 if (pte_none(ptent))
462 continue;
463
464 if (!pte_present(ptent))
465 continue;
466
467 folio = vm_normal_folio(vma, addr, ptent);
468 if (!folio || folio_is_zone_device(folio))
469 continue;
470
471 /*
472 * If we encounter a large folio, only split it if it is not
473 * fully mapped within the range we are operating on. Otherwise
474 * leave it as is so that it can be swapped out whole. If we
475 * fail to split a folio, leave it in place and advance to the
476 * next pte in the range.
477 */
478 if (folio_test_large(folio)) {
479 bool any_young;
480
481 nr = madvise_folio_pte_batch(addr, end, folio, pte,
482 ptent, &any_young, NULL);
483 if (any_young)
484 ptent = pte_mkyoung(ptent);
485
486 if (nr < folio_nr_pages(folio)) {
487 int err;
488
489 if (folio_likely_mapped_shared(folio))
490 continue;
491 if (pageout_anon_only_filter && !folio_test_anon(folio))
492 continue;
493 if (!folio_trylock(folio))
494 continue;
495 folio_get(folio);
496 arch_leave_lazy_mmu_mode();
497 pte_unmap_unlock(start_pte, ptl);
498 start_pte = NULL;
499 err = split_folio(folio);
500 folio_unlock(folio);
501 folio_put(folio);
502 start_pte = pte =
503 pte_offset_map_lock(mm, pmd, addr, &ptl);
504 if (!start_pte)
505 break;
506 arch_enter_lazy_mmu_mode();
507 if (!err)
508 nr = 0;
509 continue;
510 }
511 }
512
513 /*
514 * Do not interfere with other mappings of this folio and
515 * non-LRU folio. If we have a large folio at this point, we
516 * know it is fully mapped so if its mapcount is the same as its
517 * number of pages, it must be exclusive.
518 */
519 if (!folio_test_lru(folio) ||
520 folio_mapcount(folio) != folio_nr_pages(folio))
521 continue;
522
523 if (pageout_anon_only_filter && !folio_test_anon(folio))
524 continue;
525
526 if (!pageout && pte_young(ptent)) {
527 clear_young_dirty_ptes(vma, addr, pte, nr,
528 CYDP_CLEAR_YOUNG);
529 tlb_remove_tlb_entries(tlb, pte, nr, addr);
530 }
531
532 /*
533 * We are deactivating a folio for accelerating reclaiming.
534 * VM couldn't reclaim the folio unless we clear PG_young.
535 * As a side effect, it makes confuse idle-page tracking
536 * because they will miss recent referenced history.
537 */
538 folio_clear_referenced(folio);
539 folio_test_clear_young(folio);
540 if (folio_test_active(folio))
541 folio_set_workingset(folio);
542 if (pageout) {
543 if (folio_isolate_lru(folio)) {
544 if (folio_test_unevictable(folio))
545 folio_putback_lru(folio);
546 else
547 list_add(&folio->lru, &folio_list);
548 }
549 } else
550 folio_deactivate(folio);
551 }
552
553 if (start_pte) {
554 arch_leave_lazy_mmu_mode();
555 pte_unmap_unlock(start_pte, ptl);
556 }
557 if (pageout)
558 reclaim_pages(&folio_list);
559 cond_resched();
560
561 return 0;
562}
563
564static const struct mm_walk_ops cold_walk_ops = {
565 .pmd_entry = madvise_cold_or_pageout_pte_range,
566 .walk_lock = PGWALK_RDLOCK,
567};
568
569static void madvise_cold_page_range(struct mmu_gather *tlb,
570 struct vm_area_struct *vma,
571 unsigned long addr, unsigned long end)
572{
573 struct madvise_walk_private walk_private = {
574 .pageout = false,
575 .tlb = tlb,
576 };
577
578 tlb_start_vma(tlb, vma);
579 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
580 tlb_end_vma(tlb, vma);
581}
582
583static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
584{
585 return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
586}
587
588static long madvise_cold(struct vm_area_struct *vma,
589 struct vm_area_struct **prev,
590 unsigned long start_addr, unsigned long end_addr)
591{
592 struct mm_struct *mm = vma->vm_mm;
593 struct mmu_gather tlb;
594
595 *prev = vma;
596 if (!can_madv_lru_vma(vma))
597 return -EINVAL;
598
599 lru_add_drain();
600 tlb_gather_mmu(&tlb, mm);
601 madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
602 tlb_finish_mmu(&tlb);
603
604 return 0;
605}
606
607static void madvise_pageout_page_range(struct mmu_gather *tlb,
608 struct vm_area_struct *vma,
609 unsigned long addr, unsigned long end)
610{
611 struct madvise_walk_private walk_private = {
612 .pageout = true,
613 .tlb = tlb,
614 };
615
616 tlb_start_vma(tlb, vma);
617 walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
618 tlb_end_vma(tlb, vma);
619}
620
621static long madvise_pageout(struct vm_area_struct *vma,
622 struct vm_area_struct **prev,
623 unsigned long start_addr, unsigned long end_addr)
624{
625 struct mm_struct *mm = vma->vm_mm;
626 struct mmu_gather tlb;
627
628 *prev = vma;
629 if (!can_madv_lru_vma(vma))
630 return -EINVAL;
631
632 /*
633 * If the VMA belongs to a private file mapping, there can be private
634 * dirty pages which can be paged out if even this process is neither
635 * owner nor write capable of the file. We allow private file mappings
636 * further to pageout dirty anon pages.
637 */
638 if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
639 (vma->vm_flags & VM_MAYSHARE)))
640 return 0;
641
642 lru_add_drain();
643 tlb_gather_mmu(&tlb, mm);
644 madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
645 tlb_finish_mmu(&tlb);
646
647 return 0;
648}
649
650static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
651 unsigned long end, struct mm_walk *walk)
652
653{
654 const cydp_t cydp_flags = CYDP_CLEAR_YOUNG | CYDP_CLEAR_DIRTY;
655 struct mmu_gather *tlb = walk->private;
656 struct mm_struct *mm = tlb->mm;
657 struct vm_area_struct *vma = walk->vma;
658 spinlock_t *ptl;
659 pte_t *start_pte, *pte, ptent;
660 struct folio *folio;
661 int nr_swap = 0;
662 unsigned long next;
663 int nr, max_nr;
664
665 next = pmd_addr_end(addr, end);
666 if (pmd_trans_huge(*pmd))
667 if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
668 return 0;
669
670 tlb_change_page_size(tlb, PAGE_SIZE);
671 start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
672 if (!start_pte)
673 return 0;
674 flush_tlb_batched_pending(mm);
675 arch_enter_lazy_mmu_mode();
676 for (; addr != end; pte += nr, addr += PAGE_SIZE * nr) {
677 nr = 1;
678 ptent = ptep_get(pte);
679
680 if (pte_none(ptent))
681 continue;
682 /*
683 * If the pte has swp_entry, just clear page table to
684 * prevent swap-in which is more expensive rather than
685 * (page allocation + zeroing).
686 */
687 if (!pte_present(ptent)) {
688 swp_entry_t entry;
689
690 entry = pte_to_swp_entry(ptent);
691 if (!non_swap_entry(entry)) {
692 max_nr = (end - addr) / PAGE_SIZE;
693 nr = swap_pte_batch(pte, max_nr, ptent);
694 nr_swap -= nr;
695 free_swap_and_cache_nr(entry, nr);
696 clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
697 } else if (is_hwpoison_entry(entry) ||
698 is_poisoned_swp_entry(entry)) {
699 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
700 }
701 continue;
702 }
703
704 folio = vm_normal_folio(vma, addr, ptent);
705 if (!folio || folio_is_zone_device(folio))
706 continue;
707
708 /*
709 * If we encounter a large folio, only split it if it is not
710 * fully mapped within the range we are operating on. Otherwise
711 * leave it as is so that it can be marked as lazyfree. If we
712 * fail to split a folio, leave it in place and advance to the
713 * next pte in the range.
714 */
715 if (folio_test_large(folio)) {
716 bool any_young, any_dirty;
717
718 nr = madvise_folio_pte_batch(addr, end, folio, pte,
719 ptent, &any_young, &any_dirty);
720
721 if (nr < folio_nr_pages(folio)) {
722 int err;
723
724 if (folio_likely_mapped_shared(folio))
725 continue;
726 if (!folio_trylock(folio))
727 continue;
728 folio_get(folio);
729 arch_leave_lazy_mmu_mode();
730 pte_unmap_unlock(start_pte, ptl);
731 start_pte = NULL;
732 err = split_folio(folio);
733 folio_unlock(folio);
734 folio_put(folio);
735 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
736 start_pte = pte;
737 if (!start_pte)
738 break;
739 arch_enter_lazy_mmu_mode();
740 if (!err)
741 nr = 0;
742 continue;
743 }
744
745 if (any_young)
746 ptent = pte_mkyoung(ptent);
747 if (any_dirty)
748 ptent = pte_mkdirty(ptent);
749 }
750
751 if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
752 if (!folio_trylock(folio))
753 continue;
754 /*
755 * If we have a large folio at this point, we know it is
756 * fully mapped so if its mapcount is the same as its
757 * number of pages, it must be exclusive.
758 */
759 if (folio_mapcount(folio) != folio_nr_pages(folio)) {
760 folio_unlock(folio);
761 continue;
762 }
763
764 if (folio_test_swapcache(folio) &&
765 !folio_free_swap(folio)) {
766 folio_unlock(folio);
767 continue;
768 }
769
770 folio_clear_dirty(folio);
771 folio_unlock(folio);
772 }
773
774 if (pte_young(ptent) || pte_dirty(ptent)) {
775 clear_young_dirty_ptes(vma, addr, pte, nr, cydp_flags);
776 tlb_remove_tlb_entries(tlb, pte, nr, addr);
777 }
778 folio_mark_lazyfree(folio);
779 }
780
781 if (nr_swap)
782 add_mm_counter(mm, MM_SWAPENTS, nr_swap);
783 if (start_pte) {
784 arch_leave_lazy_mmu_mode();
785 pte_unmap_unlock(start_pte, ptl);
786 }
787 cond_resched();
788
789 return 0;
790}
791
792static const struct mm_walk_ops madvise_free_walk_ops = {
793 .pmd_entry = madvise_free_pte_range,
794 .walk_lock = PGWALK_RDLOCK,
795};
796
797static int madvise_free_single_vma(struct vm_area_struct *vma,
798 unsigned long start_addr, unsigned long end_addr)
799{
800 struct mm_struct *mm = vma->vm_mm;
801 struct mmu_notifier_range range;
802 struct mmu_gather tlb;
803
804 /* MADV_FREE works for only anon vma at the moment */
805 if (!vma_is_anonymous(vma))
806 return -EINVAL;
807
808 range.start = max(vma->vm_start, start_addr);
809 if (range.start >= vma->vm_end)
810 return -EINVAL;
811 range.end = min(vma->vm_end, end_addr);
812 if (range.end <= vma->vm_start)
813 return -EINVAL;
814 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
815 range.start, range.end);
816
817 lru_add_drain();
818 tlb_gather_mmu(&tlb, mm);
819 update_hiwater_rss(mm);
820
821 mmu_notifier_invalidate_range_start(&range);
822 tlb_start_vma(&tlb, vma);
823 walk_page_range(vma->vm_mm, range.start, range.end,
824 &madvise_free_walk_ops, &tlb);
825 tlb_end_vma(&tlb, vma);
826 mmu_notifier_invalidate_range_end(&range);
827 tlb_finish_mmu(&tlb);
828
829 return 0;
830}
831
832/*
833 * Application no longer needs these pages. If the pages are dirty,
834 * it's OK to just throw them away. The app will be more careful about
835 * data it wants to keep. Be sure to free swap resources too. The
836 * zap_page_range_single call sets things up for shrink_active_list to actually
837 * free these pages later if no one else has touched them in the meantime,
838 * although we could add these pages to a global reuse list for
839 * shrink_active_list to pick up before reclaiming other pages.
840 *
841 * NB: This interface discards data rather than pushes it out to swap,
842 * as some implementations do. This has performance implications for
843 * applications like large transactional databases which want to discard
844 * pages in anonymous maps after committing to backing store the data
845 * that was kept in them. There is no reason to write this data out to
846 * the swap area if the application is discarding it.
847 *
848 * An interface that causes the system to free clean pages and flush
849 * dirty pages is already available as msync(MS_INVALIDATE).
850 */
851static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
852 unsigned long start, unsigned long end)
853{
854 zap_page_range_single(vma, start, end - start, NULL);
855 return 0;
856}
857
858static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
859 unsigned long start,
860 unsigned long *end,
861 int behavior)
862{
863 if (!is_vm_hugetlb_page(vma)) {
864 unsigned int forbidden = VM_PFNMAP;
865
866 if (behavior != MADV_DONTNEED_LOCKED)
867 forbidden |= VM_LOCKED;
868
869 return !(vma->vm_flags & forbidden);
870 }
871
872 if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
873 return false;
874 if (start & ~huge_page_mask(hstate_vma(vma)))
875 return false;
876
877 /*
878 * Madvise callers expect the length to be rounded up to PAGE_SIZE
879 * boundaries, and may be unaware that this VMA uses huge pages.
880 * Avoid unexpected data loss by rounding down the number of
881 * huge pages freed.
882 */
883 *end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
884
885 return true;
886}
887
888static long madvise_dontneed_free(struct vm_area_struct *vma,
889 struct vm_area_struct **prev,
890 unsigned long start, unsigned long end,
891 int behavior)
892{
893 struct mm_struct *mm = vma->vm_mm;
894
895 *prev = vma;
896 if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
897 return -EINVAL;
898
899 if (start == end)
900 return 0;
901
902 if (!userfaultfd_remove(vma, start, end)) {
903 *prev = NULL; /* mmap_lock has been dropped, prev is stale */
904
905 mmap_read_lock(mm);
906 vma = vma_lookup(mm, start);
907 if (!vma)
908 return -ENOMEM;
909 /*
910 * Potential end adjustment for hugetlb vma is OK as
911 * the check below keeps end within vma.
912 */
913 if (!madvise_dontneed_free_valid_vma(vma, start, &end,
914 behavior))
915 return -EINVAL;
916 if (end > vma->vm_end) {
917 /*
918 * Don't fail if end > vma->vm_end. If the old
919 * vma was split while the mmap_lock was
920 * released the effect of the concurrent
921 * operation may not cause madvise() to
922 * have an undefined result. There may be an
923 * adjacent next vma that we'll walk
924 * next. userfaultfd_remove() will generate an
925 * UFFD_EVENT_REMOVE repetition on the
926 * end-vma->vm_end range, but the manager can
927 * handle a repetition fine.
928 */
929 end = vma->vm_end;
930 }
931 /*
932 * If the memory region between start and end was
933 * originally backed by 4kB pages and then remapped to
934 * be backed by hugepages while mmap_lock was dropped,
935 * the adjustment for hugetlb vma above may have rounded
936 * end down to the start address.
937 */
938 if (start == end)
939 return 0;
940 VM_WARN_ON(start > end);
941 }
942
943 if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
944 return madvise_dontneed_single_vma(vma, start, end);
945 else if (behavior == MADV_FREE)
946 return madvise_free_single_vma(vma, start, end);
947 else
948 return -EINVAL;
949}
950
951static long madvise_populate(struct mm_struct *mm, unsigned long start,
952 unsigned long end, int behavior)
953{
954 const bool write = behavior == MADV_POPULATE_WRITE;
955 int locked = 1;
956 long pages;
957
958 while (start < end) {
959 /* Populate (prefault) page tables readable/writable. */
960 pages = faultin_page_range(mm, start, end, write, &locked);
961 if (!locked) {
962 mmap_read_lock(mm);
963 locked = 1;
964 }
965 if (pages < 0) {
966 switch (pages) {
967 case -EINTR:
968 return -EINTR;
969 case -EINVAL: /* Incompatible mappings / permissions. */
970 return -EINVAL;
971 case -EHWPOISON:
972 return -EHWPOISON;
973 case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
974 return -EFAULT;
975 default:
976 pr_warn_once("%s: unhandled return value: %ld\n",
977 __func__, pages);
978 fallthrough;
979 case -ENOMEM: /* No VMA or out of memory. */
980 return -ENOMEM;
981 }
982 }
983 start += pages * PAGE_SIZE;
984 }
985 return 0;
986}
987
988/*
989 * Application wants to free up the pages and associated backing store.
990 * This is effectively punching a hole into the middle of a file.
991 */
992static long madvise_remove(struct vm_area_struct *vma,
993 struct vm_area_struct **prev,
994 unsigned long start, unsigned long end)
995{
996 loff_t offset;
997 int error;
998 struct file *f;
999 struct mm_struct *mm = vma->vm_mm;
1000
1001 *prev = NULL; /* tell sys_madvise we drop mmap_lock */
1002
1003 if (vma->vm_flags & VM_LOCKED)
1004 return -EINVAL;
1005
1006 f = vma->vm_file;
1007
1008 if (!f || !f->f_mapping || !f->f_mapping->host) {
1009 return -EINVAL;
1010 }
1011
1012 if (!vma_is_shared_maywrite(vma))
1013 return -EACCES;
1014
1015 offset = (loff_t)(start - vma->vm_start)
1016 + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
1017
1018 /*
1019 * Filesystem's fallocate may need to take i_rwsem. We need to
1020 * explicitly grab a reference because the vma (and hence the
1021 * vma's reference to the file) can go away as soon as we drop
1022 * mmap_lock.
1023 */
1024 get_file(f);
1025 if (userfaultfd_remove(vma, start, end)) {
1026 /* mmap_lock was not released by userfaultfd_remove() */
1027 mmap_read_unlock(mm);
1028 }
1029 error = vfs_fallocate(f,
1030 FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1031 offset, end - start);
1032 fput(f);
1033 mmap_read_lock(mm);
1034 return error;
1035}
1036
1037static bool is_valid_guard_vma(struct vm_area_struct *vma, bool allow_locked)
1038{
1039 vm_flags_t disallowed = VM_SPECIAL | VM_HUGETLB;
1040
1041 /*
1042 * A user could lock after setting a guard range but that's fine, as
1043 * they'd not be able to fault in. The issue arises when we try to zap
1044 * existing locked VMAs. We don't want to do that.
1045 */
1046 if (!allow_locked)
1047 disallowed |= VM_LOCKED;
1048
1049 if (!vma_is_anonymous(vma))
1050 return false;
1051
1052 if ((vma->vm_flags & (VM_MAYWRITE | disallowed)) != VM_MAYWRITE)
1053 return false;
1054
1055 return true;
1056}
1057
1058static bool is_guard_pte_marker(pte_t ptent)
1059{
1060 return is_pte_marker(ptent) &&
1061 is_guard_swp_entry(pte_to_swp_entry(ptent));
1062}
1063
1064static int guard_install_pud_entry(pud_t *pud, unsigned long addr,
1065 unsigned long next, struct mm_walk *walk)
1066{
1067 pud_t pudval = pudp_get(pud);
1068
1069 /* If huge return >0 so we abort the operation + zap. */
1070 return pud_trans_huge(pudval) || pud_devmap(pudval);
1071}
1072
1073static int guard_install_pmd_entry(pmd_t *pmd, unsigned long addr,
1074 unsigned long next, struct mm_walk *walk)
1075{
1076 pmd_t pmdval = pmdp_get(pmd);
1077
1078 /* If huge return >0 so we abort the operation + zap. */
1079 return pmd_trans_huge(pmdval) || pmd_devmap(pmdval);
1080}
1081
1082static int guard_install_pte_entry(pte_t *pte, unsigned long addr,
1083 unsigned long next, struct mm_walk *walk)
1084{
1085 pte_t pteval = ptep_get(pte);
1086 unsigned long *nr_pages = (unsigned long *)walk->private;
1087
1088 /* If there is already a guard page marker, we have nothing to do. */
1089 if (is_guard_pte_marker(pteval)) {
1090 (*nr_pages)++;
1091
1092 return 0;
1093 }
1094
1095 /* If populated return >0 so we abort the operation + zap. */
1096 return 1;
1097}
1098
1099static int guard_install_set_pte(unsigned long addr, unsigned long next,
1100 pte_t *ptep, struct mm_walk *walk)
1101{
1102 unsigned long *nr_pages = (unsigned long *)walk->private;
1103
1104 /* Simply install a PTE marker, this causes segfault on access. */
1105 *ptep = make_pte_marker(PTE_MARKER_GUARD);
1106 (*nr_pages)++;
1107
1108 return 0;
1109}
1110
1111static const struct mm_walk_ops guard_install_walk_ops = {
1112 .pud_entry = guard_install_pud_entry,
1113 .pmd_entry = guard_install_pmd_entry,
1114 .pte_entry = guard_install_pte_entry,
1115 .install_pte = guard_install_set_pte,
1116 .walk_lock = PGWALK_RDLOCK,
1117};
1118
1119static long madvise_guard_install(struct vm_area_struct *vma,
1120 struct vm_area_struct **prev,
1121 unsigned long start, unsigned long end)
1122{
1123 long err;
1124 int i;
1125
1126 *prev = vma;
1127 if (!is_valid_guard_vma(vma, /* allow_locked = */false))
1128 return -EINVAL;
1129
1130 /*
1131 * If we install guard markers, then the range is no longer
1132 * empty from a page table perspective and therefore it's
1133 * appropriate to have an anon_vma.
1134 *
1135 * This ensures that on fork, we copy page tables correctly.
1136 */
1137 err = anon_vma_prepare(vma);
1138 if (err)
1139 return err;
1140
1141 /*
1142 * Optimistically try to install the guard marker pages first. If any
1143 * non-guard pages are encountered, give up and zap the range before
1144 * trying again.
1145 *
1146 * We try a few times before giving up and releasing back to userland to
1147 * loop around, releasing locks in the process to avoid contention. This
1148 * would only happen if there was a great many racing page faults.
1149 *
1150 * In most cases we should simply install the guard markers immediately
1151 * with no zap or looping.
1152 */
1153 for (i = 0; i < MAX_MADVISE_GUARD_RETRIES; i++) {
1154 unsigned long nr_pages = 0;
1155
1156 /* Returns < 0 on error, == 0 if success, > 0 if zap needed. */
1157 err = walk_page_range_mm(vma->vm_mm, start, end,
1158 &guard_install_walk_ops, &nr_pages);
1159 if (err < 0)
1160 return err;
1161
1162 if (err == 0) {
1163 unsigned long nr_expected_pages = PHYS_PFN(end - start);
1164
1165 VM_WARN_ON(nr_pages != nr_expected_pages);
1166 return 0;
1167 }
1168
1169 /*
1170 * OK some of the range have non-guard pages mapped, zap
1171 * them. This leaves existing guard pages in place.
1172 */
1173 zap_page_range_single(vma, start, end - start, NULL);
1174 }
1175
1176 /*
1177 * We were unable to install the guard pages due to being raced by page
1178 * faults. This should not happen ordinarily. We return to userspace and
1179 * immediately retry, relieving lock contention.
1180 */
1181 return restart_syscall();
1182}
1183
1184static int guard_remove_pud_entry(pud_t *pud, unsigned long addr,
1185 unsigned long next, struct mm_walk *walk)
1186{
1187 pud_t pudval = pudp_get(pud);
1188
1189 /* If huge, cannot have guard pages present, so no-op - skip. */
1190 if (pud_trans_huge(pudval) || pud_devmap(pudval))
1191 walk->action = ACTION_CONTINUE;
1192
1193 return 0;
1194}
1195
1196static int guard_remove_pmd_entry(pmd_t *pmd, unsigned long addr,
1197 unsigned long next, struct mm_walk *walk)
1198{
1199 pmd_t pmdval = pmdp_get(pmd);
1200
1201 /* If huge, cannot have guard pages present, so no-op - skip. */
1202 if (pmd_trans_huge(pmdval) || pmd_devmap(pmdval))
1203 walk->action = ACTION_CONTINUE;
1204
1205 return 0;
1206}
1207
1208static int guard_remove_pte_entry(pte_t *pte, unsigned long addr,
1209 unsigned long next, struct mm_walk *walk)
1210{
1211 pte_t ptent = ptep_get(pte);
1212
1213 if (is_guard_pte_marker(ptent)) {
1214 /* Simply clear the PTE marker. */
1215 pte_clear_not_present_full(walk->mm, addr, pte, false);
1216 update_mmu_cache(walk->vma, addr, pte);
1217 }
1218
1219 return 0;
1220}
1221
1222static const struct mm_walk_ops guard_remove_walk_ops = {
1223 .pud_entry = guard_remove_pud_entry,
1224 .pmd_entry = guard_remove_pmd_entry,
1225 .pte_entry = guard_remove_pte_entry,
1226 .walk_lock = PGWALK_RDLOCK,
1227};
1228
1229static long madvise_guard_remove(struct vm_area_struct *vma,
1230 struct vm_area_struct **prev,
1231 unsigned long start, unsigned long end)
1232{
1233 *prev = vma;
1234 /*
1235 * We're ok with removing guards in mlock()'d ranges, as this is a
1236 * non-destructive action.
1237 */
1238 if (!is_valid_guard_vma(vma, /* allow_locked = */true))
1239 return -EINVAL;
1240
1241 return walk_page_range(vma->vm_mm, start, end,
1242 &guard_remove_walk_ops, NULL);
1243}
1244
1245/*
1246 * Apply an madvise behavior to a region of a vma. madvise_update_vma
1247 * will handle splitting a vm area into separate areas, each area with its own
1248 * behavior.
1249 */
1250static int madvise_vma_behavior(struct vm_area_struct *vma,
1251 struct vm_area_struct **prev,
1252 unsigned long start, unsigned long end,
1253 unsigned long behavior)
1254{
1255 int error;
1256 struct anon_vma_name *anon_name;
1257 unsigned long new_flags = vma->vm_flags;
1258
1259 if (unlikely(!can_modify_vma_madv(vma, behavior)))
1260 return -EPERM;
1261
1262 switch (behavior) {
1263 case MADV_REMOVE:
1264 return madvise_remove(vma, prev, start, end);
1265 case MADV_WILLNEED:
1266 return madvise_willneed(vma, prev, start, end);
1267 case MADV_COLD:
1268 return madvise_cold(vma, prev, start, end);
1269 case MADV_PAGEOUT:
1270 return madvise_pageout(vma, prev, start, end);
1271 case MADV_FREE:
1272 case MADV_DONTNEED:
1273 case MADV_DONTNEED_LOCKED:
1274 return madvise_dontneed_free(vma, prev, start, end, behavior);
1275 case MADV_NORMAL:
1276 new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1277 break;
1278 case MADV_SEQUENTIAL:
1279 new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1280 break;
1281 case MADV_RANDOM:
1282 new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1283 break;
1284 case MADV_DONTFORK:
1285 new_flags |= VM_DONTCOPY;
1286 break;
1287 case MADV_DOFORK:
1288 if (vma->vm_flags & VM_IO)
1289 return -EINVAL;
1290 new_flags &= ~VM_DONTCOPY;
1291 break;
1292 case MADV_WIPEONFORK:
1293 /* MADV_WIPEONFORK is only supported on anonymous memory. */
1294 if (vma->vm_file || vma->vm_flags & VM_SHARED)
1295 return -EINVAL;
1296 new_flags |= VM_WIPEONFORK;
1297 break;
1298 case MADV_KEEPONFORK:
1299 if (vma->vm_flags & VM_DROPPABLE)
1300 return -EINVAL;
1301 new_flags &= ~VM_WIPEONFORK;
1302 break;
1303 case MADV_DONTDUMP:
1304 new_flags |= VM_DONTDUMP;
1305 break;
1306 case MADV_DODUMP:
1307 if ((!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) ||
1308 (vma->vm_flags & VM_DROPPABLE))
1309 return -EINVAL;
1310 new_flags &= ~VM_DONTDUMP;
1311 break;
1312 case MADV_MERGEABLE:
1313 case MADV_UNMERGEABLE:
1314 error = ksm_madvise(vma, start, end, behavior, &new_flags);
1315 if (error)
1316 goto out;
1317 break;
1318 case MADV_HUGEPAGE:
1319 case MADV_NOHUGEPAGE:
1320 error = hugepage_madvise(vma, &new_flags, behavior);
1321 if (error)
1322 goto out;
1323 break;
1324 case MADV_COLLAPSE:
1325 return madvise_collapse(vma, prev, start, end);
1326 case MADV_GUARD_INSTALL:
1327 return madvise_guard_install(vma, prev, start, end);
1328 case MADV_GUARD_REMOVE:
1329 return madvise_guard_remove(vma, prev, start, end);
1330 }
1331
1332 anon_name = anon_vma_name(vma);
1333 anon_vma_name_get(anon_name);
1334 error = madvise_update_vma(vma, prev, start, end, new_flags,
1335 anon_name);
1336 anon_vma_name_put(anon_name);
1337
1338out:
1339 /*
1340 * madvise() returns EAGAIN if kernel resources, such as
1341 * slab, are temporarily unavailable.
1342 */
1343 if (error == -ENOMEM)
1344 error = -EAGAIN;
1345 return error;
1346}
1347
1348#ifdef CONFIG_MEMORY_FAILURE
1349/*
1350 * Error injection support for memory error handling.
1351 */
1352static int madvise_inject_error(int behavior,
1353 unsigned long start, unsigned long end)
1354{
1355 unsigned long size;
1356
1357 if (!capable(CAP_SYS_ADMIN))
1358 return -EPERM;
1359
1360
1361 for (; start < end; start += size) {
1362 unsigned long pfn;
1363 struct page *page;
1364 int ret;
1365
1366 ret = get_user_pages_fast(start, 1, 0, &page);
1367 if (ret != 1)
1368 return ret;
1369 pfn = page_to_pfn(page);
1370
1371 /*
1372 * When soft offlining hugepages, after migrating the page
1373 * we dissolve it, therefore in the second loop "page" will
1374 * no longer be a compound page.
1375 */
1376 size = page_size(compound_head(page));
1377
1378 if (behavior == MADV_SOFT_OFFLINE) {
1379 pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1380 pfn, start);
1381 ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1382 } else {
1383 pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1384 pfn, start);
1385 ret = memory_failure(pfn, MF_ACTION_REQUIRED | MF_COUNT_INCREASED | MF_SW_SIMULATED);
1386 if (ret == -EOPNOTSUPP)
1387 ret = 0;
1388 }
1389
1390 if (ret)
1391 return ret;
1392 }
1393
1394 return 0;
1395}
1396#endif
1397
1398static bool
1399madvise_behavior_valid(int behavior)
1400{
1401 switch (behavior) {
1402 case MADV_DOFORK:
1403 case MADV_DONTFORK:
1404 case MADV_NORMAL:
1405 case MADV_SEQUENTIAL:
1406 case MADV_RANDOM:
1407 case MADV_REMOVE:
1408 case MADV_WILLNEED:
1409 case MADV_DONTNEED:
1410 case MADV_DONTNEED_LOCKED:
1411 case MADV_FREE:
1412 case MADV_COLD:
1413 case MADV_PAGEOUT:
1414 case MADV_POPULATE_READ:
1415 case MADV_POPULATE_WRITE:
1416#ifdef CONFIG_KSM
1417 case MADV_MERGEABLE:
1418 case MADV_UNMERGEABLE:
1419#endif
1420#ifdef CONFIG_TRANSPARENT_HUGEPAGE
1421 case MADV_HUGEPAGE:
1422 case MADV_NOHUGEPAGE:
1423 case MADV_COLLAPSE:
1424#endif
1425 case MADV_DONTDUMP:
1426 case MADV_DODUMP:
1427 case MADV_WIPEONFORK:
1428 case MADV_KEEPONFORK:
1429 case MADV_GUARD_INSTALL:
1430 case MADV_GUARD_REMOVE:
1431#ifdef CONFIG_MEMORY_FAILURE
1432 case MADV_SOFT_OFFLINE:
1433 case MADV_HWPOISON:
1434#endif
1435 return true;
1436
1437 default:
1438 return false;
1439 }
1440}
1441
1442/* Can we invoke process_madvise() on a remote mm for the specified behavior? */
1443static bool process_madvise_remote_valid(int behavior)
1444{
1445 switch (behavior) {
1446 case MADV_COLD:
1447 case MADV_PAGEOUT:
1448 case MADV_WILLNEED:
1449 case MADV_COLLAPSE:
1450 return true;
1451 default:
1452 return false;
1453 }
1454}
1455
1456/*
1457 * Walk the vmas in range [start,end), and call the visit function on each one.
1458 * The visit function will get start and end parameters that cover the overlap
1459 * between the current vma and the original range. Any unmapped regions in the
1460 * original range will result in this function returning -ENOMEM while still
1461 * calling the visit function on all of the existing vmas in the range.
1462 * Must be called with the mmap_lock held for reading or writing.
1463 */
1464static
1465int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1466 unsigned long end, unsigned long arg,
1467 int (*visit)(struct vm_area_struct *vma,
1468 struct vm_area_struct **prev, unsigned long start,
1469 unsigned long end, unsigned long arg))
1470{
1471 struct vm_area_struct *vma;
1472 struct vm_area_struct *prev;
1473 unsigned long tmp;
1474 int unmapped_error = 0;
1475
1476 /*
1477 * If the interval [start,end) covers some unmapped address
1478 * ranges, just ignore them, but return -ENOMEM at the end.
1479 * - different from the way of handling in mlock etc.
1480 */
1481 vma = find_vma_prev(mm, start, &prev);
1482 if (vma && start > vma->vm_start)
1483 prev = vma;
1484
1485 for (;;) {
1486 int error;
1487
1488 /* Still start < end. */
1489 if (!vma)
1490 return -ENOMEM;
1491
1492 /* Here start < (end|vma->vm_end). */
1493 if (start < vma->vm_start) {
1494 unmapped_error = -ENOMEM;
1495 start = vma->vm_start;
1496 if (start >= end)
1497 break;
1498 }
1499
1500 /* Here vma->vm_start <= start < (end|vma->vm_end) */
1501 tmp = vma->vm_end;
1502 if (end < tmp)
1503 tmp = end;
1504
1505 /* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1506 error = visit(vma, &prev, start, tmp, arg);
1507 if (error)
1508 return error;
1509 start = tmp;
1510 if (prev && start < prev->vm_end)
1511 start = prev->vm_end;
1512 if (start >= end)
1513 break;
1514 if (prev)
1515 vma = find_vma(mm, prev->vm_end);
1516 else /* madvise_remove dropped mmap_lock */
1517 vma = find_vma(mm, start);
1518 }
1519
1520 return unmapped_error;
1521}
1522
1523#ifdef CONFIG_ANON_VMA_NAME
1524static int madvise_vma_anon_name(struct vm_area_struct *vma,
1525 struct vm_area_struct **prev,
1526 unsigned long start, unsigned long end,
1527 unsigned long anon_name)
1528{
1529 int error;
1530
1531 /* Only anonymous mappings can be named */
1532 if (vma->vm_file && !vma_is_anon_shmem(vma))
1533 return -EBADF;
1534
1535 error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1536 (struct anon_vma_name *)anon_name);
1537
1538 /*
1539 * madvise() returns EAGAIN if kernel resources, such as
1540 * slab, are temporarily unavailable.
1541 */
1542 if (error == -ENOMEM)
1543 error = -EAGAIN;
1544 return error;
1545}
1546
1547int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1548 unsigned long len_in, struct anon_vma_name *anon_name)
1549{
1550 unsigned long end;
1551 unsigned long len;
1552
1553 if (start & ~PAGE_MASK)
1554 return -EINVAL;
1555 len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1556
1557 /* Check to see whether len was rounded up from small -ve to zero */
1558 if (len_in && !len)
1559 return -EINVAL;
1560
1561 end = start + len;
1562 if (end < start)
1563 return -EINVAL;
1564
1565 if (end == start)
1566 return 0;
1567
1568 return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1569 madvise_vma_anon_name);
1570}
1571#endif /* CONFIG_ANON_VMA_NAME */
1572/*
1573 * The madvise(2) system call.
1574 *
1575 * Applications can use madvise() to advise the kernel how it should
1576 * handle paging I/O in this VM area. The idea is to help the kernel
1577 * use appropriate read-ahead and caching techniques. The information
1578 * provided is advisory only, and can be safely disregarded by the
1579 * kernel without affecting the correct operation of the application.
1580 *
1581 * behavior values:
1582 * MADV_NORMAL - the default behavior is to read clusters. This
1583 * results in some read-ahead and read-behind.
1584 * MADV_RANDOM - the system should read the minimum amount of data
1585 * on any access, since it is unlikely that the appli-
1586 * cation will need more than what it asks for.
1587 * MADV_SEQUENTIAL - pages in the given range will probably be accessed
1588 * once, so they can be aggressively read ahead, and
1589 * can be freed soon after they are accessed.
1590 * MADV_WILLNEED - the application is notifying the system to read
1591 * some pages ahead.
1592 * MADV_DONTNEED - the application is finished with the given range,
1593 * so the kernel can free resources associated with it.
1594 * MADV_FREE - the application marks pages in the given range as lazy free,
1595 * where actual purges are postponed until memory pressure happens.
1596 * MADV_REMOVE - the application wants to free up the given range of
1597 * pages and associated backing store.
1598 * MADV_DONTFORK - omit this area from child's address space when forking:
1599 * typically, to avoid COWing pages pinned by get_user_pages().
1600 * MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1601 * MADV_WIPEONFORK - present the child process with zero-filled memory in this
1602 * range after a fork.
1603 * MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1604 * MADV_HWPOISON - trigger memory error handler as if the given memory range
1605 * were corrupted by unrecoverable hardware memory failure.
1606 * MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1607 * MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1608 * this area with pages of identical content from other such areas.
1609 * MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1610 * MADV_HUGEPAGE - the application wants to back the given range by transparent
1611 * huge pages in the future. Existing pages might be coalesced and
1612 * new pages might be allocated as THP.
1613 * MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1614 * transparent huge pages so the existing pages will not be
1615 * coalesced into THP and new pages will not be allocated as THP.
1616 * MADV_COLLAPSE - synchronously coalesce pages into new THP.
1617 * MADV_DONTDUMP - the application wants to prevent pages in the given range
1618 * from being included in its core dump.
1619 * MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1620 * MADV_COLD - the application is not expected to use this memory soon,
1621 * deactivate pages in this range so that they can be reclaimed
1622 * easily if memory pressure happens.
1623 * MADV_PAGEOUT - the application is not expected to use this memory soon,
1624 * page out the pages in this range immediately.
1625 * MADV_POPULATE_READ - populate (prefault) page tables readable by
1626 * triggering read faults if required
1627 * MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1628 * triggering write faults if required
1629 *
1630 * return values:
1631 * zero - success
1632 * -EINVAL - start + len < 0, start is not page-aligned,
1633 * "behavior" is not a valid value, or application
1634 * is attempting to release locked or shared pages,
1635 * or the specified address range includes file, Huge TLB,
1636 * MAP_SHARED or VMPFNMAP range.
1637 * -ENOMEM - addresses in the specified range are not currently
1638 * mapped, or are outside the AS of the process.
1639 * -EIO - an I/O error occurred while paging in data.
1640 * -EBADF - map exists, but area maps something that isn't a file.
1641 * -EAGAIN - a kernel resource was temporarily unavailable.
1642 * -EPERM - memory is sealed.
1643 */
1644int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1645{
1646 unsigned long end;
1647 int error;
1648 int write;
1649 size_t len;
1650 struct blk_plug plug;
1651
1652 if (!madvise_behavior_valid(behavior))
1653 return -EINVAL;
1654
1655 if (!PAGE_ALIGNED(start))
1656 return -EINVAL;
1657 len = PAGE_ALIGN(len_in);
1658
1659 /* Check to see whether len was rounded up from small -ve to zero */
1660 if (len_in && !len)
1661 return -EINVAL;
1662
1663 end = start + len;
1664 if (end < start)
1665 return -EINVAL;
1666
1667 if (end == start)
1668 return 0;
1669
1670#ifdef CONFIG_MEMORY_FAILURE
1671 if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1672 return madvise_inject_error(behavior, start, start + len_in);
1673#endif
1674
1675 write = madvise_need_mmap_write(behavior);
1676 if (write) {
1677 if (mmap_write_lock_killable(mm))
1678 return -EINTR;
1679 } else {
1680 mmap_read_lock(mm);
1681 }
1682
1683 start = untagged_addr_remote(mm, start);
1684 end = start + len;
1685
1686 blk_start_plug(&plug);
1687 switch (behavior) {
1688 case MADV_POPULATE_READ:
1689 case MADV_POPULATE_WRITE:
1690 error = madvise_populate(mm, start, end, behavior);
1691 break;
1692 default:
1693 error = madvise_walk_vmas(mm, start, end, behavior,
1694 madvise_vma_behavior);
1695 break;
1696 }
1697 blk_finish_plug(&plug);
1698
1699 if (write)
1700 mmap_write_unlock(mm);
1701 else
1702 mmap_read_unlock(mm);
1703
1704 return error;
1705}
1706
1707SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1708{
1709 return do_madvise(current->mm, start, len_in, behavior);
1710}
1711
1712/* Perform an madvise operation over a vector of addresses and lengths. */
1713static ssize_t vector_madvise(struct mm_struct *mm, struct iov_iter *iter,
1714 int behavior)
1715{
1716 ssize_t ret = 0;
1717 size_t total_len;
1718
1719 total_len = iov_iter_count(iter);
1720
1721 while (iov_iter_count(iter)) {
1722 ret = do_madvise(mm, (unsigned long)iter_iov_addr(iter),
1723 iter_iov_len(iter), behavior);
1724 /*
1725 * An madvise operation is attempting to restart the syscall,
1726 * but we cannot proceed as it would not be correct to repeat
1727 * the operation in aggregate, and would be surprising to the
1728 * user.
1729 *
1730 * As we have already dropped locks, it is safe to just loop and
1731 * try again. We check for fatal signals in case we need exit
1732 * early anyway.
1733 */
1734 if (ret == -ERESTARTNOINTR) {
1735 if (fatal_signal_pending(current)) {
1736 ret = -EINTR;
1737 break;
1738 }
1739 continue;
1740 }
1741 if (ret < 0)
1742 break;
1743 iov_iter_advance(iter, iter_iov_len(iter));
1744 }
1745
1746 ret = (total_len - iov_iter_count(iter)) ? : ret;
1747
1748 return ret;
1749}
1750
1751SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1752 size_t, vlen, int, behavior, unsigned int, flags)
1753{
1754 ssize_t ret;
1755 struct iovec iovstack[UIO_FASTIOV];
1756 struct iovec *iov = iovstack;
1757 struct iov_iter iter;
1758 struct task_struct *task;
1759 struct mm_struct *mm;
1760 unsigned int f_flags;
1761
1762 if (flags != 0) {
1763 ret = -EINVAL;
1764 goto out;
1765 }
1766
1767 ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1768 if (ret < 0)
1769 goto out;
1770
1771 task = pidfd_get_task(pidfd, &f_flags);
1772 if (IS_ERR(task)) {
1773 ret = PTR_ERR(task);
1774 goto free_iov;
1775 }
1776
1777 /* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1778 mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1779 if (IS_ERR(mm)) {
1780 ret = PTR_ERR(mm);
1781 goto release_task;
1782 }
1783
1784 /*
1785 * We need only perform this check if we are attempting to manipulate a
1786 * remote process's address space.
1787 */
1788 if (mm != current->mm && !process_madvise_remote_valid(behavior)) {
1789 ret = -EINVAL;
1790 goto release_mm;
1791 }
1792
1793 /*
1794 * Require CAP_SYS_NICE for influencing process performance. Note that
1795 * only non-destructive hints are currently supported for remote
1796 * processes.
1797 */
1798 if (mm != current->mm && !capable(CAP_SYS_NICE)) {
1799 ret = -EPERM;
1800 goto release_mm;
1801 }
1802
1803 ret = vector_madvise(mm, &iter, behavior);
1804
1805release_mm:
1806 mmput(mm);
1807release_task:
1808 put_task_struct(task);
1809free_iov:
1810 kfree(iov);
1811out:
1812 return ret;
1813}