• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/ksm.h>
21 #include <linux/fs.h>
22 #include <linux/file.h>
23 #include <linux/blkdev.h>
24 #include <linux/backing-dev.h>
25 #include <linux/pagewalk.h>
26 #include <linux/swap.h>
27 #include <linux/swapops.h>
28 #include <linux/shmem_fs.h>
29 #include <linux/mmu_notifier.h>
30 
31 #include <asm/tlb.h>
32 
33 #include "internal.h"
34 
35 struct madvise_walk_private {
36 	struct mmu_gather *tlb;
37 	bool pageout;
38 };
39 
40 /*
41  * Any behaviour which results in changes to the vma->vm_flags needs to
42  * take mmap_sem for writing. Others, which simply traverse vmas, need
43  * to only take it for reading.
44  */
madvise_need_mmap_write(int behavior)45 static int madvise_need_mmap_write(int behavior)
46 {
47 	switch (behavior) {
48 	case MADV_REMOVE:
49 	case MADV_WILLNEED:
50 	case MADV_DONTNEED:
51 	case MADV_COLD:
52 	case MADV_PAGEOUT:
53 	case MADV_FREE:
54 		return 0;
55 	default:
56 		/* be safe, default to 1. list exceptions explicitly */
57 		return 1;
58 	}
59 }
60 
61 /*
62  * We can potentially split a vm area into separate
63  * areas, each area with its own behavior.
64  */
madvise_behavior(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,int behavior)65 static long madvise_behavior(struct vm_area_struct *vma,
66 		     struct vm_area_struct **prev,
67 		     unsigned long start, unsigned long end, int behavior)
68 {
69 	struct mm_struct *mm = vma->vm_mm;
70 	int error = 0;
71 	pgoff_t pgoff;
72 	unsigned long new_flags = vma->vm_flags;
73 
74 	switch (behavior) {
75 	case MADV_NORMAL:
76 		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
77 		break;
78 	case MADV_SEQUENTIAL:
79 		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
80 		break;
81 	case MADV_RANDOM:
82 		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
83 		break;
84 	case MADV_DONTFORK:
85 		new_flags |= VM_DONTCOPY;
86 		break;
87 	case MADV_DOFORK:
88 		if (vma->vm_flags & VM_IO) {
89 			error = -EINVAL;
90 			goto out;
91 		}
92 		new_flags &= ~VM_DONTCOPY;
93 		break;
94 	case MADV_WIPEONFORK:
95 		/* MADV_WIPEONFORK is only supported on anonymous memory. */
96 		if (vma->vm_file || vma->vm_flags & VM_SHARED) {
97 			error = -EINVAL;
98 			goto out;
99 		}
100 		new_flags |= VM_WIPEONFORK;
101 		break;
102 	case MADV_KEEPONFORK:
103 		new_flags &= ~VM_WIPEONFORK;
104 		break;
105 	case MADV_DONTDUMP:
106 		new_flags |= VM_DONTDUMP;
107 		break;
108 	case MADV_DODUMP:
109 		if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL) {
110 			error = -EINVAL;
111 			goto out;
112 		}
113 		new_flags &= ~VM_DONTDUMP;
114 		break;
115 	case MADV_MERGEABLE:
116 	case MADV_UNMERGEABLE:
117 		error = ksm_madvise(vma, start, end, behavior, &new_flags);
118 		if (error)
119 			goto out_convert_errno;
120 		break;
121 	case MADV_HUGEPAGE:
122 	case MADV_NOHUGEPAGE:
123 		error = hugepage_madvise(vma, &new_flags, behavior);
124 		if (error)
125 			goto out_convert_errno;
126 		break;
127 	}
128 
129 	if (new_flags == vma->vm_flags) {
130 		*prev = vma;
131 		goto out;
132 	}
133 
134 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
135 	*prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
136 			  vma->vm_file, pgoff, vma_policy(vma),
137 			  vma->vm_userfaultfd_ctx, vma_get_anon_name(vma));
138 	if (*prev) {
139 		vma = *prev;
140 		goto success;
141 	}
142 
143 	*prev = vma;
144 
145 	if (start != vma->vm_start) {
146 		if (unlikely(mm->map_count >= sysctl_max_map_count)) {
147 			error = -ENOMEM;
148 			goto out;
149 		}
150 		error = __split_vma(mm, vma, start, 1);
151 		if (error)
152 			goto out_convert_errno;
153 	}
154 
155 	if (end != vma->vm_end) {
156 		if (unlikely(mm->map_count >= sysctl_max_map_count)) {
157 			error = -ENOMEM;
158 			goto out;
159 		}
160 		error = __split_vma(mm, vma, end, 0);
161 		if (error)
162 			goto out_convert_errno;
163 	}
164 
165 success:
166 	/*
167 	 * vm_flags is protected by the mmap_sem held in write mode.
168 	 */
169 	vma->vm_flags = new_flags;
170 
171 out_convert_errno:
172 	/*
173 	 * madvise() returns EAGAIN if kernel resources, such as
174 	 * slab, are temporarily unavailable.
175 	 */
176 	if (error == -ENOMEM)
177 		error = -EAGAIN;
178 out:
179 	return error;
180 }
181 
182 #ifdef CONFIG_SWAP
swapin_walk_pmd_entry(pmd_t * pmd,unsigned long start,unsigned long end,struct mm_walk * walk)183 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
184 	unsigned long end, struct mm_walk *walk)
185 {
186 	pte_t *orig_pte;
187 	struct vm_area_struct *vma = walk->private;
188 	unsigned long index;
189 
190 	if (pmd_none_or_trans_huge_or_clear_bad(pmd))
191 		return 0;
192 
193 	for (index = start; index != end; index += PAGE_SIZE) {
194 		pte_t pte;
195 		swp_entry_t entry;
196 		struct page *page;
197 		spinlock_t *ptl;
198 
199 		orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
200 		pte = *(orig_pte + ((index - start) / PAGE_SIZE));
201 		pte_unmap_unlock(orig_pte, ptl);
202 
203 		if (pte_present(pte) || pte_none(pte))
204 			continue;
205 		entry = pte_to_swp_entry(pte);
206 		if (unlikely(non_swap_entry(entry)))
207 			continue;
208 
209 		page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
210 							vma, index, false);
211 		if (page)
212 			put_page(page);
213 	}
214 
215 	return 0;
216 }
217 
218 static const struct mm_walk_ops swapin_walk_ops = {
219 	.pmd_entry		= swapin_walk_pmd_entry,
220 };
221 
force_shm_swapin_readahead(struct vm_area_struct * vma,unsigned long start,unsigned long end,struct address_space * mapping)222 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
223 		unsigned long start, unsigned long end,
224 		struct address_space *mapping)
225 {
226 	pgoff_t index;
227 	struct page *page;
228 	swp_entry_t swap;
229 
230 	for (; start < end; start += PAGE_SIZE) {
231 		index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
232 
233 		page = find_get_entry(mapping, index);
234 		if (!xa_is_value(page)) {
235 			if (page)
236 				put_page(page);
237 			continue;
238 		}
239 		swap = radix_to_swp_entry(page);
240 		page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
241 							NULL, 0, false);
242 		if (page)
243 			put_page(page);
244 	}
245 
246 	lru_add_drain();	/* Push any new pages onto the LRU now */
247 }
248 #endif		/* CONFIG_SWAP */
249 
250 /*
251  * Schedule all required I/O operations.  Do not wait for completion.
252  */
madvise_willneed(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end)253 static long madvise_willneed(struct vm_area_struct *vma,
254 			     struct vm_area_struct **prev,
255 			     unsigned long start, unsigned long end)
256 {
257 	struct file *file = vma->vm_file;
258 	loff_t offset;
259 
260 	*prev = vma;
261 #ifdef CONFIG_SWAP
262 	if (!file) {
263 		walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
264 		lru_add_drain(); /* Push any new pages onto the LRU now */
265 		return 0;
266 	}
267 
268 	if (shmem_mapping(file->f_mapping)) {
269 		force_shm_swapin_readahead(vma, start, end,
270 					file->f_mapping);
271 		return 0;
272 	}
273 #else
274 	if (!file)
275 		return -EBADF;
276 #endif
277 
278 	if (IS_DAX(file_inode(file))) {
279 		/* no bad return value, but ignore advice */
280 		return 0;
281 	}
282 
283 	/*
284 	 * Filesystem's fadvise may need to take various locks.  We need to
285 	 * explicitly grab a reference because the vma (and hence the
286 	 * vma's reference to the file) can go away as soon as we drop
287 	 * mmap_sem.
288 	 */
289 	*prev = NULL;	/* tell sys_madvise we drop mmap_sem */
290 	get_file(file);
291 	offset = (loff_t)(start - vma->vm_start)
292 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
293 	up_read(&current->mm->mmap_sem);
294 	vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
295 	fput(file);
296 	down_read(&current->mm->mmap_sem);
297 	return 0;
298 }
299 
madvise_cold_or_pageout_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)300 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
301 				unsigned long addr, unsigned long end,
302 				struct mm_walk *walk)
303 {
304 	struct madvise_walk_private *private = walk->private;
305 	struct mmu_gather *tlb = private->tlb;
306 	bool pageout = private->pageout;
307 	struct mm_struct *mm = tlb->mm;
308 	struct vm_area_struct *vma = walk->vma;
309 	pte_t *orig_pte, *pte, ptent;
310 	spinlock_t *ptl;
311 	struct page *page = NULL;
312 	LIST_HEAD(page_list);
313 
314 	if (fatal_signal_pending(current))
315 		return -EINTR;
316 
317 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
318 	if (pmd_trans_huge(*pmd)) {
319 		pmd_t orig_pmd;
320 		unsigned long next = pmd_addr_end(addr, end);
321 
322 		tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
323 		ptl = pmd_trans_huge_lock(pmd, vma);
324 		if (!ptl)
325 			return 0;
326 
327 		orig_pmd = *pmd;
328 		if (is_huge_zero_pmd(orig_pmd))
329 			goto huge_unlock;
330 
331 		if (unlikely(!pmd_present(orig_pmd))) {
332 			VM_BUG_ON(thp_migration_supported() &&
333 					!is_pmd_migration_entry(orig_pmd));
334 			goto huge_unlock;
335 		}
336 
337 		page = pmd_page(orig_pmd);
338 
339 		/* Do not interfere with other mappings of this page */
340 		if (page_mapcount(page) != 1)
341 			goto huge_unlock;
342 
343 		if (next - addr != HPAGE_PMD_SIZE) {
344 			int err;
345 
346 			get_page(page);
347 			spin_unlock(ptl);
348 			lock_page(page);
349 			err = split_huge_page(page);
350 			unlock_page(page);
351 			put_page(page);
352 			if (!err)
353 				goto regular_page;
354 			return 0;
355 		}
356 
357 		if (pmd_young(orig_pmd)) {
358 			pmdp_invalidate(vma, addr, pmd);
359 			orig_pmd = pmd_mkold(orig_pmd);
360 
361 			set_pmd_at(mm, addr, pmd, orig_pmd);
362 			tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
363 		}
364 
365 		ClearPageReferenced(page);
366 		test_and_clear_page_young(page);
367 		if (pageout) {
368 			if (!isolate_lru_page(page)) {
369 				if (PageUnevictable(page))
370 					putback_lru_page(page);
371 				else
372 					list_add(&page->lru, &page_list);
373 			}
374 		} else
375 			deactivate_page(page);
376 huge_unlock:
377 		spin_unlock(ptl);
378 		if (pageout)
379 			reclaim_pages(&page_list);
380 		return 0;
381 	}
382 
383 regular_page:
384 	if (pmd_trans_unstable(pmd))
385 		return 0;
386 #endif
387 	tlb_change_page_size(tlb, PAGE_SIZE);
388 	orig_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
389 	flush_tlb_batched_pending(mm);
390 	arch_enter_lazy_mmu_mode();
391 	for (; addr < end; pte++, addr += PAGE_SIZE) {
392 		ptent = *pte;
393 
394 		if (pte_none(ptent))
395 			continue;
396 
397 		if (!pte_present(ptent))
398 			continue;
399 
400 		page = vm_normal_page(vma, addr, ptent);
401 		if (!page)
402 			continue;
403 
404 		/*
405 		 * Creating a THP page is expensive so split it only if we
406 		 * are sure it's worth. Split it if we are only owner.
407 		 */
408 		if (PageTransCompound(page)) {
409 			if (page_mapcount(page) != 1)
410 				break;
411 			get_page(page);
412 			if (!trylock_page(page)) {
413 				put_page(page);
414 				break;
415 			}
416 			pte_unmap_unlock(orig_pte, ptl);
417 			if (split_huge_page(page)) {
418 				unlock_page(page);
419 				put_page(page);
420 				pte_offset_map_lock(mm, pmd, addr, &ptl);
421 				break;
422 			}
423 			unlock_page(page);
424 			put_page(page);
425 			pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
426 			pte--;
427 			addr -= PAGE_SIZE;
428 			continue;
429 		}
430 
431 		/*
432 		 * Do not interfere with other mappings of this page and
433 		 * non-LRU page.
434 		 */
435 		if (!PageLRU(page) || page_mapcount(page) != 1)
436 			continue;
437 
438 		VM_BUG_ON_PAGE(PageTransCompound(page), page);
439 
440 		if (pte_young(ptent)) {
441 			ptent = ptep_get_and_clear_full(mm, addr, pte,
442 							tlb->fullmm);
443 			ptent = pte_mkold(ptent);
444 			set_pte_at(mm, addr, pte, ptent);
445 			tlb_remove_tlb_entry(tlb, pte, addr);
446 		}
447 
448 		/*
449 		 * We are deactivating a page for accelerating reclaiming.
450 		 * VM couldn't reclaim the page unless we clear PG_young.
451 		 * As a side effect, it makes confuse idle-page tracking
452 		 * because they will miss recent referenced history.
453 		 */
454 		ClearPageReferenced(page);
455 		test_and_clear_page_young(page);
456 		if (pageout) {
457 			if (!isolate_lru_page(page)) {
458 				if (PageUnevictable(page))
459 					putback_lru_page(page);
460 				else
461 					list_add(&page->lru, &page_list);
462 			}
463 		} else
464 			deactivate_page(page);
465 	}
466 
467 	arch_leave_lazy_mmu_mode();
468 	pte_unmap_unlock(orig_pte, ptl);
469 	if (pageout)
470 		reclaim_pages(&page_list);
471 	cond_resched();
472 
473 	return 0;
474 }
475 
476 static const struct mm_walk_ops cold_walk_ops = {
477 	.pmd_entry = madvise_cold_or_pageout_pte_range,
478 };
479 
madvise_cold_page_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long addr,unsigned long end)480 static void madvise_cold_page_range(struct mmu_gather *tlb,
481 			     struct vm_area_struct *vma,
482 			     unsigned long addr, unsigned long end)
483 {
484 	struct madvise_walk_private walk_private = {
485 		.pageout = false,
486 		.tlb = tlb,
487 	};
488 
489 	tlb_start_vma(tlb, vma);
490 	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
491 	tlb_end_vma(tlb, vma);
492 }
493 
madvise_cold(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start_addr,unsigned long end_addr)494 static long madvise_cold(struct vm_area_struct *vma,
495 			struct vm_area_struct **prev,
496 			unsigned long start_addr, unsigned long end_addr)
497 {
498 	struct mm_struct *mm = vma->vm_mm;
499 	struct mmu_gather tlb;
500 
501 	*prev = vma;
502 	if (!can_madv_lru_vma(vma))
503 		return -EINVAL;
504 
505 	lru_add_drain();
506 	tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
507 	madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
508 	tlb_finish_mmu(&tlb, start_addr, end_addr);
509 
510 	return 0;
511 }
512 
madvise_pageout_page_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long addr,unsigned long end)513 static void madvise_pageout_page_range(struct mmu_gather *tlb,
514 			     struct vm_area_struct *vma,
515 			     unsigned long addr, unsigned long end)
516 {
517 	struct madvise_walk_private walk_private = {
518 		.pageout = true,
519 		.tlb = tlb,
520 	};
521 
522 	tlb_start_vma(tlb, vma);
523 	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
524 	tlb_end_vma(tlb, vma);
525 }
526 
can_do_pageout(struct vm_area_struct * vma)527 static inline bool can_do_pageout(struct vm_area_struct *vma)
528 {
529 	if (vma_is_anonymous(vma))
530 		return true;
531 	if (!vma->vm_file)
532 		return false;
533 	/*
534 	 * paging out pagecache only for non-anonymous mappings that correspond
535 	 * to the files the calling process could (if tried) open for writing;
536 	 * otherwise we'd be including shared non-exclusive mappings, which
537 	 * opens a side channel.
538 	 */
539 	return inode_owner_or_capable(file_inode(vma->vm_file)) ||
540 		inode_permission(file_inode(vma->vm_file), MAY_WRITE) == 0;
541 }
542 
madvise_pageout(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start_addr,unsigned long end_addr)543 static long madvise_pageout(struct vm_area_struct *vma,
544 			struct vm_area_struct **prev,
545 			unsigned long start_addr, unsigned long end_addr)
546 {
547 	struct mm_struct *mm = vma->vm_mm;
548 	struct mmu_gather tlb;
549 
550 	*prev = vma;
551 	if (!can_madv_lru_vma(vma))
552 		return -EINVAL;
553 
554 	if (!can_do_pageout(vma))
555 		return 0;
556 
557 	lru_add_drain();
558 	tlb_gather_mmu(&tlb, mm, start_addr, end_addr);
559 	madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
560 	tlb_finish_mmu(&tlb, start_addr, end_addr);
561 
562 	return 0;
563 }
564 
madvise_free_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)565 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
566 				unsigned long end, struct mm_walk *walk)
567 
568 {
569 	struct mmu_gather *tlb = walk->private;
570 	struct mm_struct *mm = tlb->mm;
571 	struct vm_area_struct *vma = walk->vma;
572 	spinlock_t *ptl;
573 	pte_t *orig_pte, *pte, ptent;
574 	struct page *page;
575 	int nr_swap = 0;
576 	unsigned long next;
577 
578 	next = pmd_addr_end(addr, end);
579 	if (pmd_trans_huge(*pmd))
580 		if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
581 			goto next;
582 
583 	if (pmd_trans_unstable(pmd))
584 		return 0;
585 
586 	tlb_change_page_size(tlb, PAGE_SIZE);
587 	orig_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
588 	flush_tlb_batched_pending(mm);
589 	arch_enter_lazy_mmu_mode();
590 	for (; addr != end; pte++, addr += PAGE_SIZE) {
591 		ptent = *pte;
592 
593 		if (pte_none(ptent))
594 			continue;
595 		/*
596 		 * If the pte has swp_entry, just clear page table to
597 		 * prevent swap-in which is more expensive rather than
598 		 * (page allocation + zeroing).
599 		 */
600 		if (!pte_present(ptent)) {
601 			swp_entry_t entry;
602 
603 			entry = pte_to_swp_entry(ptent);
604 			if (non_swap_entry(entry))
605 				continue;
606 			nr_swap--;
607 			free_swap_and_cache(entry);
608 			pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
609 			continue;
610 		}
611 
612 		page = vm_normal_page(vma, addr, ptent);
613 		if (!page)
614 			continue;
615 
616 		/*
617 		 * If pmd isn't transhuge but the page is THP and
618 		 * is owned by only this process, split it and
619 		 * deactivate all pages.
620 		 */
621 		if (PageTransCompound(page)) {
622 			if (page_mapcount(page) != 1)
623 				goto out;
624 			get_page(page);
625 			if (!trylock_page(page)) {
626 				put_page(page);
627 				goto out;
628 			}
629 			pte_unmap_unlock(orig_pte, ptl);
630 			if (split_huge_page(page)) {
631 				unlock_page(page);
632 				put_page(page);
633 				pte_offset_map_lock(mm, pmd, addr, &ptl);
634 				goto out;
635 			}
636 			unlock_page(page);
637 			put_page(page);
638 			pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
639 			pte--;
640 			addr -= PAGE_SIZE;
641 			continue;
642 		}
643 
644 		VM_BUG_ON_PAGE(PageTransCompound(page), page);
645 
646 		if (PageSwapCache(page) || PageDirty(page)) {
647 			if (!trylock_page(page))
648 				continue;
649 			/*
650 			 * If page is shared with others, we couldn't clear
651 			 * PG_dirty of the page.
652 			 */
653 			if (page_mapcount(page) != 1) {
654 				unlock_page(page);
655 				continue;
656 			}
657 
658 			if (PageSwapCache(page) && !try_to_free_swap(page)) {
659 				unlock_page(page);
660 				continue;
661 			}
662 
663 			ClearPageDirty(page);
664 			unlock_page(page);
665 		}
666 
667 		if (pte_young(ptent) || pte_dirty(ptent)) {
668 			/*
669 			 * Some of architecture(ex, PPC) don't update TLB
670 			 * with set_pte_at and tlb_remove_tlb_entry so for
671 			 * the portability, remap the pte with old|clean
672 			 * after pte clearing.
673 			 */
674 			ptent = ptep_get_and_clear_full(mm, addr, pte,
675 							tlb->fullmm);
676 
677 			ptent = pte_mkold(ptent);
678 			ptent = pte_mkclean(ptent);
679 			set_pte_at(mm, addr, pte, ptent);
680 			tlb_remove_tlb_entry(tlb, pte, addr);
681 		}
682 		mark_page_lazyfree(page);
683 	}
684 out:
685 	if (nr_swap) {
686 		if (current->mm == mm)
687 			sync_mm_rss(mm);
688 
689 		add_mm_counter(mm, MM_SWAPENTS, nr_swap);
690 	}
691 	arch_leave_lazy_mmu_mode();
692 	pte_unmap_unlock(orig_pte, ptl);
693 	cond_resched();
694 next:
695 	return 0;
696 }
697 
698 static const struct mm_walk_ops madvise_free_walk_ops = {
699 	.pmd_entry		= madvise_free_pte_range,
700 };
701 
madvise_free_single_vma(struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr)702 static int madvise_free_single_vma(struct vm_area_struct *vma,
703 			unsigned long start_addr, unsigned long end_addr)
704 {
705 	struct mm_struct *mm = vma->vm_mm;
706 	struct mmu_notifier_range range;
707 	struct mmu_gather tlb;
708 
709 	/* MADV_FREE works for only anon vma at the moment */
710 	if (!vma_is_anonymous(vma))
711 		return -EINVAL;
712 
713 	range.start = max(vma->vm_start, start_addr);
714 	if (range.start >= vma->vm_end)
715 		return -EINVAL;
716 	range.end = min(vma->vm_end, end_addr);
717 	if (range.end <= vma->vm_start)
718 		return -EINVAL;
719 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
720 				range.start, range.end);
721 
722 	lru_add_drain();
723 	tlb_gather_mmu(&tlb, mm, range.start, range.end);
724 	update_hiwater_rss(mm);
725 
726 	mmu_notifier_invalidate_range_start(&range);
727 	tlb_start_vma(&tlb, vma);
728 	walk_page_range(vma->vm_mm, range.start, range.end,
729 			&madvise_free_walk_ops, &tlb);
730 	tlb_end_vma(&tlb, vma);
731 	mmu_notifier_invalidate_range_end(&range);
732 	tlb_finish_mmu(&tlb, range.start, range.end);
733 
734 	return 0;
735 }
736 
737 /*
738  * Application no longer needs these pages.  If the pages are dirty,
739  * it's OK to just throw them away.  The app will be more careful about
740  * data it wants to keep.  Be sure to free swap resources too.  The
741  * zap_page_range call sets things up for shrink_active_list to actually free
742  * these pages later if no one else has touched them in the meantime,
743  * although we could add these pages to a global reuse list for
744  * shrink_active_list to pick up before reclaiming other pages.
745  *
746  * NB: This interface discards data rather than pushes it out to swap,
747  * as some implementations do.  This has performance implications for
748  * applications like large transactional databases which want to discard
749  * pages in anonymous maps after committing to backing store the data
750  * that was kept in them.  There is no reason to write this data out to
751  * the swap area if the application is discarding it.
752  *
753  * An interface that causes the system to free clean pages and flush
754  * dirty pages is already available as msync(MS_INVALIDATE).
755  */
madvise_dontneed_single_vma(struct vm_area_struct * vma,unsigned long start,unsigned long end)756 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
757 					unsigned long start, unsigned long end)
758 {
759 	zap_page_range(vma, start, end - start);
760 	return 0;
761 }
762 
madvise_dontneed_free(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,int behavior)763 static long madvise_dontneed_free(struct vm_area_struct *vma,
764 				  struct vm_area_struct **prev,
765 				  unsigned long start, unsigned long end,
766 				  int behavior)
767 {
768 	*prev = vma;
769 	if (!can_madv_lru_vma(vma))
770 		return -EINVAL;
771 
772 	if (!userfaultfd_remove(vma, start, end)) {
773 		*prev = NULL; /* mmap_sem has been dropped, prev is stale */
774 
775 		down_read(&current->mm->mmap_sem);
776 		vma = find_vma(current->mm, start);
777 		if (!vma)
778 			return -ENOMEM;
779 		if (start < vma->vm_start) {
780 			/*
781 			 * This "vma" under revalidation is the one
782 			 * with the lowest vma->vm_start where start
783 			 * is also < vma->vm_end. If start <
784 			 * vma->vm_start it means an hole materialized
785 			 * in the user address space within the
786 			 * virtual range passed to MADV_DONTNEED
787 			 * or MADV_FREE.
788 			 */
789 			return -ENOMEM;
790 		}
791 		if (!can_madv_lru_vma(vma))
792 			return -EINVAL;
793 		if (end > vma->vm_end) {
794 			/*
795 			 * Don't fail if end > vma->vm_end. If the old
796 			 * vma was splitted while the mmap_sem was
797 			 * released the effect of the concurrent
798 			 * operation may not cause madvise() to
799 			 * have an undefined result. There may be an
800 			 * adjacent next vma that we'll walk
801 			 * next. userfaultfd_remove() will generate an
802 			 * UFFD_EVENT_REMOVE repetition on the
803 			 * end-vma->vm_end range, but the manager can
804 			 * handle a repetition fine.
805 			 */
806 			end = vma->vm_end;
807 		}
808 		VM_WARN_ON(start >= end);
809 	}
810 
811 	if (behavior == MADV_DONTNEED)
812 		return madvise_dontneed_single_vma(vma, start, end);
813 	else if (behavior == MADV_FREE)
814 		return madvise_free_single_vma(vma, start, end);
815 	else
816 		return -EINVAL;
817 }
818 
819 /*
820  * Application wants to free up the pages and associated backing store.
821  * This is effectively punching a hole into the middle of a file.
822  */
madvise_remove(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end)823 static long madvise_remove(struct vm_area_struct *vma,
824 				struct vm_area_struct **prev,
825 				unsigned long start, unsigned long end)
826 {
827 	loff_t offset;
828 	int error;
829 	struct file *f;
830 
831 	*prev = NULL;	/* tell sys_madvise we drop mmap_sem */
832 
833 	if (vma->vm_flags & VM_LOCKED)
834 		return -EINVAL;
835 
836 	f = vma->vm_file;
837 
838 	if (!f || !f->f_mapping || !f->f_mapping->host) {
839 			return -EINVAL;
840 	}
841 
842 	if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
843 		return -EACCES;
844 
845 	offset = (loff_t)(start - vma->vm_start)
846 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
847 
848 	/*
849 	 * Filesystem's fallocate may need to take i_mutex.  We need to
850 	 * explicitly grab a reference because the vma (and hence the
851 	 * vma's reference to the file) can go away as soon as we drop
852 	 * mmap_sem.
853 	 */
854 	get_file(f);
855 	if (userfaultfd_remove(vma, start, end)) {
856 		/* mmap_sem was not released by userfaultfd_remove() */
857 		up_read(&current->mm->mmap_sem);
858 	}
859 	error = vfs_fallocate(f,
860 				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
861 				offset, end - start);
862 	fput(f);
863 	down_read(&current->mm->mmap_sem);
864 	return error;
865 }
866 
867 #ifdef CONFIG_MEMORY_FAILURE
868 /*
869  * Error injection support for memory error handling.
870  */
madvise_inject_error(int behavior,unsigned long start,unsigned long end)871 static int madvise_inject_error(int behavior,
872 		unsigned long start, unsigned long end)
873 {
874 	struct page *page;
875 	struct zone *zone;
876 	unsigned int order;
877 
878 	if (!capable(CAP_SYS_ADMIN))
879 		return -EPERM;
880 
881 
882 	for (; start < end; start += PAGE_SIZE << order) {
883 		unsigned long pfn;
884 		int ret;
885 
886 		ret = get_user_pages_fast(start, 1, 0, &page);
887 		if (ret != 1)
888 			return ret;
889 		pfn = page_to_pfn(page);
890 
891 		/*
892 		 * When soft offlining hugepages, after migrating the page
893 		 * we dissolve it, therefore in the second loop "page" will
894 		 * no longer be a compound page, and order will be 0.
895 		 */
896 		order = compound_order(compound_head(page));
897 
898 		if (PageHWPoison(page)) {
899 			put_page(page);
900 			continue;
901 		}
902 
903 		if (behavior == MADV_SOFT_OFFLINE) {
904 			pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
905 					pfn, start);
906 
907 			ret = soft_offline_page(page, MF_COUNT_INCREASED);
908 			if (ret)
909 				return ret;
910 			continue;
911 		}
912 
913 		pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
914 				pfn, start);
915 
916 		/*
917 		 * Drop the page reference taken by get_user_pages_fast(). In
918 		 * the absence of MF_COUNT_INCREASED the memory_failure()
919 		 * routine is responsible for pinning the page to prevent it
920 		 * from being released back to the page allocator.
921 		 */
922 		put_page(page);
923 		ret = memory_failure(pfn, 0);
924 		if (ret)
925 			return ret;
926 	}
927 
928 	/* Ensure that all poisoned pages are removed from per-cpu lists */
929 	for_each_populated_zone(zone)
930 		drain_all_pages(zone);
931 
932 	return 0;
933 }
934 #endif
935 
936 static long
madvise_vma(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,int behavior)937 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
938 		unsigned long start, unsigned long end, int behavior)
939 {
940 	switch (behavior) {
941 	case MADV_REMOVE:
942 		return madvise_remove(vma, prev, start, end);
943 	case MADV_WILLNEED:
944 		return madvise_willneed(vma, prev, start, end);
945 	case MADV_COLD:
946 		return madvise_cold(vma, prev, start, end);
947 	case MADV_PAGEOUT:
948 		return madvise_pageout(vma, prev, start, end);
949 	case MADV_FREE:
950 	case MADV_DONTNEED:
951 		return madvise_dontneed_free(vma, prev, start, end, behavior);
952 	default:
953 		return madvise_behavior(vma, prev, start, end, behavior);
954 	}
955 }
956 
957 static bool
madvise_behavior_valid(int behavior)958 madvise_behavior_valid(int behavior)
959 {
960 	switch (behavior) {
961 	case MADV_DOFORK:
962 	case MADV_DONTFORK:
963 	case MADV_NORMAL:
964 	case MADV_SEQUENTIAL:
965 	case MADV_RANDOM:
966 	case MADV_REMOVE:
967 	case MADV_WILLNEED:
968 	case MADV_DONTNEED:
969 	case MADV_FREE:
970 	case MADV_COLD:
971 	case MADV_PAGEOUT:
972 #ifdef CONFIG_KSM
973 	case MADV_MERGEABLE:
974 	case MADV_UNMERGEABLE:
975 #endif
976 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
977 	case MADV_HUGEPAGE:
978 	case MADV_NOHUGEPAGE:
979 #endif
980 	case MADV_DONTDUMP:
981 	case MADV_DODUMP:
982 	case MADV_WIPEONFORK:
983 	case MADV_KEEPONFORK:
984 #ifdef CONFIG_MEMORY_FAILURE
985 	case MADV_SOFT_OFFLINE:
986 	case MADV_HWPOISON:
987 #endif
988 		return true;
989 
990 	default:
991 		return false;
992 	}
993 }
994 
995 /*
996  * The madvise(2) system call.
997  *
998  * Applications can use madvise() to advise the kernel how it should
999  * handle paging I/O in this VM area.  The idea is to help the kernel
1000  * use appropriate read-ahead and caching techniques.  The information
1001  * provided is advisory only, and can be safely disregarded by the
1002  * kernel without affecting the correct operation of the application.
1003  *
1004  * behavior values:
1005  *  MADV_NORMAL - the default behavior is to read clusters.  This
1006  *		results in some read-ahead and read-behind.
1007  *  MADV_RANDOM - the system should read the minimum amount of data
1008  *		on any access, since it is unlikely that the appli-
1009  *		cation will need more than what it asks for.
1010  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1011  *		once, so they can be aggressively read ahead, and
1012  *		can be freed soon after they are accessed.
1013  *  MADV_WILLNEED - the application is notifying the system to read
1014  *		some pages ahead.
1015  *  MADV_DONTNEED - the application is finished with the given range,
1016  *		so the kernel can free resources associated with it.
1017  *  MADV_FREE - the application marks pages in the given range as lazy free,
1018  *		where actual purges are postponed until memory pressure happens.
1019  *  MADV_REMOVE - the application wants to free up the given range of
1020  *		pages and associated backing store.
1021  *  MADV_DONTFORK - omit this area from child's address space when forking:
1022  *		typically, to avoid COWing pages pinned by get_user_pages().
1023  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1024  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1025  *              range after a fork.
1026  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1027  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1028  *		were corrupted by unrecoverable hardware memory failure.
1029  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1030  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1031  *		this area with pages of identical content from other such areas.
1032  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1033  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1034  *		huge pages in the future. Existing pages might be coalesced and
1035  *		new pages might be allocated as THP.
1036  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1037  *		transparent huge pages so the existing pages will not be
1038  *		coalesced into THP and new pages will not be allocated as THP.
1039  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1040  *		from being included in its core dump.
1041  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1042  *
1043  * return values:
1044  *  zero    - success
1045  *  -EINVAL - start + len < 0, start is not page-aligned,
1046  *		"behavior" is not a valid value, or application
1047  *		is attempting to release locked or shared pages,
1048  *		or the specified address range includes file, Huge TLB,
1049  *		MAP_SHARED or VMPFNMAP range.
1050  *  -ENOMEM - addresses in the specified range are not currently
1051  *		mapped, or are outside the AS of the process.
1052  *  -EIO    - an I/O error occurred while paging in data.
1053  *  -EBADF  - map exists, but area maps something that isn't a file.
1054  *  -EAGAIN - a kernel resource was temporarily unavailable.
1055  */
SYSCALL_DEFINE3(madvise,unsigned long,start,size_t,len_in,int,behavior)1056 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1057 {
1058 	unsigned long end, tmp;
1059 	struct vm_area_struct *vma, *prev;
1060 	int unmapped_error = 0;
1061 	int error = -EINVAL;
1062 	int write;
1063 	size_t len;
1064 	struct blk_plug plug;
1065 
1066 	start = untagged_addr(start);
1067 
1068 	if (!madvise_behavior_valid(behavior))
1069 		return error;
1070 
1071 	if (start & ~PAGE_MASK)
1072 		return error;
1073 	len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1074 
1075 	/* Check to see whether len was rounded up from small -ve to zero */
1076 	if (len_in && !len)
1077 		return error;
1078 
1079 	end = start + len;
1080 	if (end < start)
1081 		return error;
1082 
1083 	error = 0;
1084 	if (end == start)
1085 		return error;
1086 
1087 #ifdef CONFIG_MEMORY_FAILURE
1088 	if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1089 		return madvise_inject_error(behavior, start, start + len_in);
1090 #endif
1091 
1092 	write = madvise_need_mmap_write(behavior);
1093 	if (write) {
1094 		if (down_write_killable(&current->mm->mmap_sem))
1095 			return -EINTR;
1096 	} else {
1097 		down_read(&current->mm->mmap_sem);
1098 	}
1099 
1100 	/*
1101 	 * If the interval [start,end) covers some unmapped address
1102 	 * ranges, just ignore them, but return -ENOMEM at the end.
1103 	 * - different from the way of handling in mlock etc.
1104 	 */
1105 	vma = find_vma_prev(current->mm, start, &prev);
1106 	if (vma && start > vma->vm_start)
1107 		prev = vma;
1108 
1109 	blk_start_plug(&plug);
1110 	for (;;) {
1111 		/* Still start < end. */
1112 		error = -ENOMEM;
1113 		if (!vma)
1114 			goto out;
1115 
1116 		/* Here start < (end|vma->vm_end). */
1117 		if (start < vma->vm_start) {
1118 			unmapped_error = -ENOMEM;
1119 			start = vma->vm_start;
1120 			if (start >= end)
1121 				goto out;
1122 		}
1123 
1124 		/* Here vma->vm_start <= start < (end|vma->vm_end) */
1125 		tmp = vma->vm_end;
1126 		if (end < tmp)
1127 			tmp = end;
1128 
1129 		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1130 		error = madvise_vma(vma, &prev, start, tmp, behavior);
1131 		if (error)
1132 			goto out;
1133 		start = tmp;
1134 		if (prev && start < prev->vm_end)
1135 			start = prev->vm_end;
1136 		error = unmapped_error;
1137 		if (start >= end)
1138 			goto out;
1139 		if (prev)
1140 			vma = prev->vm_next;
1141 		else	/* madvise_remove dropped mmap_sem */
1142 			vma = find_vma(current->mm, start);
1143 	}
1144 out:
1145 	blk_finish_plug(&plug);
1146 	if (write)
1147 		up_write(&current->mm->mmap_sem);
1148 	else
1149 		up_read(&current->mm->mmap_sem);
1150 
1151 	return error;
1152 }
1153