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 struct 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  */
madvise_need_mmap_write(int behavior)50 static 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
anon_vma_name_alloc(const char * name)71 struct 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 
anon_vma_name_free(struct kref * kref)87 void 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 
anon_vma_name(struct vm_area_struct * vma)94 struct 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 */
replace_anon_vma_name(struct vm_area_struct * vma,struct anon_vma_name * anon_name)102 static 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 */
replace_anon_vma_name(struct vm_area_struct * vma,struct anon_vma_name * anon_name)122 static 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  */
madvise_update_vma(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,unsigned long new_flags,struct anon_vma_name * anon_name)137 static 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 	pgoff_t pgoff;
145 	VMA_ITERATOR(vmi, mm, start);
146 
147 	if (new_flags == vma->vm_flags && anon_vma_name_eq(anon_vma_name(vma), anon_name)) {
148 		*prev = vma;
149 		return 0;
150 	}
151 
152 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
153 	*prev = vma_merge(&vmi, mm, *prev, start, end, new_flags,
154 			  vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma),
155 			  vma->vm_userfaultfd_ctx, anon_name);
156 	if (*prev) {
157 		vma = *prev;
158 		goto success;
159 	}
160 
161 	*prev = vma;
162 
163 	if (start != vma->vm_start) {
164 		error = split_vma(&vmi, vma, start, 1);
165 		if (error)
166 			return error;
167 	}
168 
169 	if (end != vma->vm_end) {
170 		error = split_vma(&vmi, vma, end, 0);
171 		if (error)
172 			return error;
173 	}
174 
175 success:
176 	/* vm_flags is protected by the mmap_lock held in write mode. */
177 	vma_start_write(vma);
178 	vm_flags_reset(vma, new_flags);
179 	if (!vma->vm_file || vma_is_anon_shmem(vma)) {
180 		error = replace_anon_vma_name(vma, anon_name);
181 		if (error)
182 			return error;
183 	}
184 
185 	return 0;
186 }
187 
188 #ifdef CONFIG_SWAP
swapin_walk_pmd_entry(pmd_t * pmd,unsigned long start,unsigned long end,struct mm_walk * walk)189 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
190 		unsigned long end, struct mm_walk *walk)
191 {
192 	struct vm_area_struct *vma = walk->private;
193 	struct swap_iocb *splug = NULL;
194 	pte_t *ptep = NULL;
195 	spinlock_t *ptl;
196 	unsigned long addr;
197 
198 	for (addr = start; addr < end; addr += PAGE_SIZE) {
199 		pte_t pte;
200 		swp_entry_t entry;
201 		struct page *page;
202 
203 		if (!ptep++) {
204 			ptep = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
205 			if (!ptep)
206 				break;
207 		}
208 
209 		pte = ptep_get(ptep);
210 		if (!is_swap_pte(pte))
211 			continue;
212 		entry = pte_to_swp_entry(pte);
213 		if (unlikely(non_swap_entry(entry)))
214 			continue;
215 
216 		pte_unmap_unlock(ptep, ptl);
217 		ptep = NULL;
218 
219 		page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
220 					     vma, addr, &splug);
221 		if (page)
222 			put_page(page);
223 	}
224 
225 	if (ptep)
226 		pte_unmap_unlock(ptep, ptl);
227 	swap_read_unplug(splug);
228 	cond_resched();
229 
230 	return 0;
231 }
232 
233 static const struct mm_walk_ops swapin_walk_ops = {
234 	.pmd_entry		= swapin_walk_pmd_entry,
235 	.walk_lock		= PGWALK_RDLOCK,
236 };
237 
shmem_swapin_range(struct vm_area_struct * vma,unsigned long start,unsigned long end,struct address_space * mapping)238 static void shmem_swapin_range(struct vm_area_struct *vma,
239 		unsigned long start, unsigned long end,
240 		struct address_space *mapping)
241 {
242 	XA_STATE(xas, &mapping->i_pages, linear_page_index(vma, start));
243 	pgoff_t end_index = linear_page_index(vma, end) - 1;
244 	struct page *page;
245 	struct swap_iocb *splug = NULL;
246 
247 	rcu_read_lock();
248 	xas_for_each(&xas, page, end_index) {
249 		unsigned long addr;
250 		swp_entry_t entry;
251 
252 		if (!xa_is_value(page))
253 			continue;
254 		entry = radix_to_swp_entry(page);
255 		/* There might be swapin error entries in shmem mapping. */
256 		if (non_swap_entry(entry))
257 			continue;
258 
259 		addr = vma->vm_start +
260 			((xas.xa_index - vma->vm_pgoff) << PAGE_SHIFT);
261 		xas_pause(&xas);
262 		rcu_read_unlock();
263 
264 		page = read_swap_cache_async(entry, mapping_gfp_mask(mapping),
265 					     vma, addr, &splug);
266 		if (page)
267 			put_page(page);
268 
269 		rcu_read_lock();
270 	}
271 	rcu_read_unlock();
272 	swap_read_unplug(splug);
273 }
274 #endif		/* CONFIG_SWAP */
275 
276 /*
277  * Schedule all required I/O operations.  Do not wait for completion.
278  */
madvise_willneed(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end)279 static long madvise_willneed(struct vm_area_struct *vma,
280 			     struct vm_area_struct **prev,
281 			     unsigned long start, unsigned long end)
282 {
283 	struct mm_struct *mm = vma->vm_mm;
284 	struct file *file = vma->vm_file;
285 	loff_t offset;
286 
287 	*prev = vma;
288 #ifdef CONFIG_SWAP
289 	if (!file) {
290 		walk_page_range(vma->vm_mm, start, end, &swapin_walk_ops, vma);
291 		lru_add_drain(); /* Push any new pages onto the LRU now */
292 		return 0;
293 	}
294 
295 	if (shmem_mapping(file->f_mapping)) {
296 		shmem_swapin_range(vma, start, end, file->f_mapping);
297 		lru_add_drain(); /* Push any new pages onto the LRU now */
298 		return 0;
299 	}
300 #else
301 	if (!file)
302 		return -EBADF;
303 #endif
304 
305 	if (IS_DAX(file_inode(file))) {
306 		/* no bad return value, but ignore advice */
307 		return 0;
308 	}
309 
310 	/*
311 	 * Filesystem's fadvise may need to take various locks.  We need to
312 	 * explicitly grab a reference because the vma (and hence the
313 	 * vma's reference to the file) can go away as soon as we drop
314 	 * mmap_lock.
315 	 */
316 	*prev = NULL;	/* tell sys_madvise we drop mmap_lock */
317 	get_file(file);
318 	offset = (loff_t)(start - vma->vm_start)
319 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
320 	mmap_read_unlock(mm);
321 	vfs_fadvise(file, offset, end - start, POSIX_FADV_WILLNEED);
322 	fput(file);
323 	mmap_read_lock(mm);
324 	return 0;
325 }
326 
can_do_file_pageout(struct vm_area_struct * vma)327 static inline bool can_do_file_pageout(struct vm_area_struct *vma)
328 {
329 	if (!vma->vm_file)
330 		return false;
331 	/*
332 	 * paging out pagecache only for non-anonymous mappings that correspond
333 	 * to the files the calling process could (if tried) open for writing;
334 	 * otherwise we'd be including shared non-exclusive mappings, which
335 	 * opens a side channel.
336 	 */
337 	return inode_owner_or_capable(&nop_mnt_idmap,
338 				      file_inode(vma->vm_file)) ||
339 	       file_permission(vma->vm_file, MAY_WRITE) == 0;
340 }
341 
madvise_cold_or_pageout_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)342 static int madvise_cold_or_pageout_pte_range(pmd_t *pmd,
343 				unsigned long addr, unsigned long end,
344 				struct mm_walk *walk)
345 {
346 	struct madvise_walk_private *private = walk->private;
347 	struct mmu_gather *tlb = private->tlb;
348 	bool pageout = private->pageout;
349 	struct mm_struct *mm = tlb->mm;
350 	struct vm_area_struct *vma = walk->vma;
351 	pte_t *start_pte, *pte, ptent;
352 	spinlock_t *ptl;
353 	struct folio *folio = NULL;
354 	LIST_HEAD(folio_list);
355 	bool pageout_anon_only_filter;
356 
357 	if (fatal_signal_pending(current))
358 		return -EINTR;
359 
360 	pageout_anon_only_filter = pageout && !vma_is_anonymous(vma) &&
361 					!can_do_file_pageout(vma);
362 
363 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
364 	if (pmd_trans_huge(*pmd)) {
365 		pmd_t orig_pmd;
366 		unsigned long next = pmd_addr_end(addr, end);
367 
368 		tlb_change_page_size(tlb, HPAGE_PMD_SIZE);
369 		ptl = pmd_trans_huge_lock(pmd, vma);
370 		if (!ptl)
371 			return 0;
372 
373 		orig_pmd = *pmd;
374 		if (is_huge_zero_pmd(orig_pmd))
375 			goto huge_unlock;
376 
377 		if (unlikely(!pmd_present(orig_pmd))) {
378 			VM_BUG_ON(thp_migration_supported() &&
379 					!is_pmd_migration_entry(orig_pmd));
380 			goto huge_unlock;
381 		}
382 
383 		folio = pfn_folio(pmd_pfn(orig_pmd));
384 
385 		/* Do not interfere with other mappings of this folio */
386 		if (folio_estimated_sharers(folio) != 1)
387 			goto huge_unlock;
388 
389 		if (pageout_anon_only_filter && !folio_test_anon(folio))
390 			goto huge_unlock;
391 
392 		if (next - addr != HPAGE_PMD_SIZE) {
393 			int err;
394 
395 			folio_get(folio);
396 			spin_unlock(ptl);
397 			folio_lock(folio);
398 			err = split_folio(folio);
399 			folio_unlock(folio);
400 			folio_put(folio);
401 			if (!err)
402 				goto regular_folio;
403 			return 0;
404 		}
405 
406 		if (pmd_young(orig_pmd)) {
407 			pmdp_invalidate(vma, addr, pmd);
408 			orig_pmd = pmd_mkold(orig_pmd);
409 
410 			set_pmd_at(mm, addr, pmd, orig_pmd);
411 			tlb_remove_pmd_tlb_entry(tlb, pmd, addr);
412 		}
413 
414 		folio_clear_referenced(folio);
415 		folio_test_clear_young(folio);
416 		if (folio_test_active(folio))
417 			folio_set_workingset(folio);
418 		if (pageout) {
419 			if (folio_isolate_lru(folio)) {
420 				if (folio_test_unevictable(folio))
421 					folio_putback_lru(folio);
422 				else
423 					list_add(&folio->lru, &folio_list);
424 			}
425 		} else
426 			folio_deactivate(folio);
427 huge_unlock:
428 		spin_unlock(ptl);
429 		if (pageout)
430 			reclaim_pages(&folio_list);
431 		return 0;
432 	}
433 
434 regular_folio:
435 #endif
436 	tlb_change_page_size(tlb, PAGE_SIZE);
437 	start_pte = pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
438 	if (!start_pte)
439 		return 0;
440 	flush_tlb_batched_pending(mm);
441 	arch_enter_lazy_mmu_mode();
442 	for (; addr < end; pte++, addr += PAGE_SIZE) {
443 		ptent = ptep_get(pte);
444 
445 		if (pte_none(ptent))
446 			continue;
447 
448 		if (!pte_present(ptent))
449 			continue;
450 
451 		folio = vm_normal_folio(vma, addr, ptent);
452 		if (!folio || folio_is_zone_device(folio))
453 			continue;
454 
455 		/*
456 		 * Creating a THP page is expensive so split it only if we
457 		 * are sure it's worth. Split it if we are only owner.
458 		 */
459 		if (folio_test_large(folio)) {
460 			int err;
461 
462 			if (folio_estimated_sharers(folio) != 1)
463 				break;
464 			if (pageout_anon_only_filter && !folio_test_anon(folio))
465 				break;
466 			if (!folio_trylock(folio))
467 				break;
468 			folio_get(folio);
469 			arch_leave_lazy_mmu_mode();
470 			pte_unmap_unlock(start_pte, ptl);
471 			start_pte = NULL;
472 			err = split_folio(folio);
473 			folio_unlock(folio);
474 			folio_put(folio);
475 			if (err)
476 				break;
477 			start_pte = pte =
478 				pte_offset_map_lock(mm, pmd, addr, &ptl);
479 			if (!start_pte)
480 				break;
481 			arch_enter_lazy_mmu_mode();
482 			pte--;
483 			addr -= PAGE_SIZE;
484 			continue;
485 		}
486 
487 		/*
488 		 * Do not interfere with other mappings of this folio and
489 		 * non-LRU folio.
490 		 */
491 		if (!folio_test_lru(folio) || folio_mapcount(folio) != 1)
492 			continue;
493 
494 		if (pageout_anon_only_filter && !folio_test_anon(folio))
495 			continue;
496 
497 		VM_BUG_ON_FOLIO(folio_test_large(folio), folio);
498 
499 		if (pte_young(ptent)) {
500 			ptent = ptep_get_and_clear_full(mm, addr, pte,
501 							tlb->fullmm);
502 			ptent = pte_mkold(ptent);
503 			set_pte_at(mm, addr, pte, ptent);
504 			tlb_remove_tlb_entry(tlb, pte, addr);
505 		}
506 
507 		/*
508 		 * We are deactivating a folio for accelerating reclaiming.
509 		 * VM couldn't reclaim the folio unless we clear PG_young.
510 		 * As a side effect, it makes confuse idle-page tracking
511 		 * because they will miss recent referenced history.
512 		 */
513 		folio_clear_referenced(folio);
514 		folio_test_clear_young(folio);
515 		if (folio_test_active(folio))
516 			folio_set_workingset(folio);
517 		if (pageout) {
518 			if (folio_isolate_lru(folio)) {
519 				if (folio_test_unevictable(folio))
520 					folio_putback_lru(folio);
521 				else
522 					list_add(&folio->lru, &folio_list);
523 			}
524 		} else
525 			folio_deactivate(folio);
526 	}
527 
528 	if (start_pte) {
529 		arch_leave_lazy_mmu_mode();
530 		pte_unmap_unlock(start_pte, ptl);
531 	}
532 	if (pageout)
533 		reclaim_pages(&folio_list);
534 	cond_resched();
535 
536 	return 0;
537 }
538 
539 static const struct mm_walk_ops cold_walk_ops = {
540 	.pmd_entry = madvise_cold_or_pageout_pte_range,
541 	.walk_lock = PGWALK_RDLOCK,
542 };
543 
madvise_cold_page_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long addr,unsigned long end)544 static void madvise_cold_page_range(struct mmu_gather *tlb,
545 			     struct vm_area_struct *vma,
546 			     unsigned long addr, unsigned long end)
547 {
548 	struct madvise_walk_private walk_private = {
549 		.pageout = false,
550 		.tlb = tlb,
551 	};
552 
553 	tlb_start_vma(tlb, vma);
554 	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
555 	tlb_end_vma(tlb, vma);
556 }
557 
can_madv_lru_vma(struct vm_area_struct * vma)558 static inline bool can_madv_lru_vma(struct vm_area_struct *vma)
559 {
560 	return !(vma->vm_flags & (VM_LOCKED|VM_PFNMAP|VM_HUGETLB));
561 }
562 
madvise_cold(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start_addr,unsigned long end_addr)563 static long madvise_cold(struct vm_area_struct *vma,
564 			struct vm_area_struct **prev,
565 			unsigned long start_addr, unsigned long end_addr)
566 {
567 	struct mm_struct *mm = vma->vm_mm;
568 	struct mmu_gather tlb;
569 
570 	*prev = vma;
571 	if (!can_madv_lru_vma(vma))
572 		return -EINVAL;
573 
574 	lru_add_drain();
575 	tlb_gather_mmu(&tlb, mm);
576 	madvise_cold_page_range(&tlb, vma, start_addr, end_addr);
577 	tlb_finish_mmu(&tlb);
578 
579 	return 0;
580 }
581 
madvise_pageout_page_range(struct mmu_gather * tlb,struct vm_area_struct * vma,unsigned long addr,unsigned long end)582 static void madvise_pageout_page_range(struct mmu_gather *tlb,
583 			     struct vm_area_struct *vma,
584 			     unsigned long addr, unsigned long end)
585 {
586 	struct madvise_walk_private walk_private = {
587 		.pageout = true,
588 		.tlb = tlb,
589 	};
590 
591 	tlb_start_vma(tlb, vma);
592 	walk_page_range(vma->vm_mm, addr, end, &cold_walk_ops, &walk_private);
593 	tlb_end_vma(tlb, vma);
594 }
595 
madvise_pageout(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start_addr,unsigned long end_addr)596 static long madvise_pageout(struct vm_area_struct *vma,
597 			struct vm_area_struct **prev,
598 			unsigned long start_addr, unsigned long end_addr)
599 {
600 	struct mm_struct *mm = vma->vm_mm;
601 	struct mmu_gather tlb;
602 
603 	*prev = vma;
604 	if (!can_madv_lru_vma(vma))
605 		return -EINVAL;
606 
607 	/*
608 	 * If the VMA belongs to a private file mapping, there can be private
609 	 * dirty pages which can be paged out if even this process is neither
610 	 * owner nor write capable of the file. We allow private file mappings
611 	 * further to pageout dirty anon pages.
612 	 */
613 	if (!vma_is_anonymous(vma) && (!can_do_file_pageout(vma) &&
614 				(vma->vm_flags & VM_MAYSHARE)))
615 		return 0;
616 
617 	lru_add_drain();
618 	tlb_gather_mmu(&tlb, mm);
619 	madvise_pageout_page_range(&tlb, vma, start_addr, end_addr);
620 	tlb_finish_mmu(&tlb);
621 
622 	return 0;
623 }
624 
madvise_free_pte_range(pmd_t * pmd,unsigned long addr,unsigned long end,struct mm_walk * walk)625 static int madvise_free_pte_range(pmd_t *pmd, unsigned long addr,
626 				unsigned long end, struct mm_walk *walk)
627 
628 {
629 	struct mmu_gather *tlb = walk->private;
630 	struct mm_struct *mm = tlb->mm;
631 	struct vm_area_struct *vma = walk->vma;
632 	spinlock_t *ptl;
633 	pte_t *start_pte, *pte, ptent;
634 	struct folio *folio;
635 	int nr_swap = 0;
636 	unsigned long next;
637 
638 	next = pmd_addr_end(addr, end);
639 	if (pmd_trans_huge(*pmd))
640 		if (madvise_free_huge_pmd(tlb, vma, pmd, addr, next))
641 			return 0;
642 
643 	tlb_change_page_size(tlb, PAGE_SIZE);
644 	start_pte = pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
645 	if (!start_pte)
646 		return 0;
647 	flush_tlb_batched_pending(mm);
648 	arch_enter_lazy_mmu_mode();
649 	for (; addr != end; pte++, addr += PAGE_SIZE) {
650 		ptent = ptep_get(pte);
651 
652 		if (pte_none(ptent))
653 			continue;
654 		/*
655 		 * If the pte has swp_entry, just clear page table to
656 		 * prevent swap-in which is more expensive rather than
657 		 * (page allocation + zeroing).
658 		 */
659 		if (!pte_present(ptent)) {
660 			swp_entry_t entry;
661 
662 			entry = pte_to_swp_entry(ptent);
663 			if (!non_swap_entry(entry)) {
664 				nr_swap--;
665 				free_swap_and_cache(entry);
666 				pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
667 			} else if (is_hwpoison_entry(entry) ||
668 				   is_poisoned_swp_entry(entry)) {
669 				pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
670 			}
671 			continue;
672 		}
673 
674 		folio = vm_normal_folio(vma, addr, ptent);
675 		if (!folio || folio_is_zone_device(folio))
676 			continue;
677 
678 		/*
679 		 * If pmd isn't transhuge but the folio is large and
680 		 * is owned by only this process, split it and
681 		 * deactivate all pages.
682 		 */
683 		if (folio_test_large(folio)) {
684 			int err;
685 
686 			if (folio_estimated_sharers(folio) != 1)
687 				break;
688 			if (!folio_trylock(folio))
689 				break;
690 			folio_get(folio);
691 			arch_leave_lazy_mmu_mode();
692 			pte_unmap_unlock(start_pte, ptl);
693 			start_pte = NULL;
694 			err = split_folio(folio);
695 			folio_unlock(folio);
696 			folio_put(folio);
697 			if (err)
698 				break;
699 			start_pte = pte =
700 				pte_offset_map_lock(mm, pmd, addr, &ptl);
701 			if (!start_pte)
702 				break;
703 			arch_enter_lazy_mmu_mode();
704 			pte--;
705 			addr -= PAGE_SIZE;
706 			continue;
707 		}
708 
709 		if (folio_test_swapcache(folio) || folio_test_dirty(folio)) {
710 			if (!folio_trylock(folio))
711 				continue;
712 			/*
713 			 * If folio is shared with others, we mustn't clear
714 			 * the folio's dirty flag.
715 			 */
716 			if (folio_mapcount(folio) != 1) {
717 				folio_unlock(folio);
718 				continue;
719 			}
720 
721 			if (folio_test_swapcache(folio) &&
722 			    !folio_free_swap(folio)) {
723 				folio_unlock(folio);
724 				continue;
725 			}
726 
727 			folio_clear_dirty(folio);
728 			folio_unlock(folio);
729 		}
730 
731 		if (pte_young(ptent) || pte_dirty(ptent)) {
732 			/*
733 			 * Some of architecture(ex, PPC) don't update TLB
734 			 * with set_pte_at and tlb_remove_tlb_entry so for
735 			 * the portability, remap the pte with old|clean
736 			 * after pte clearing.
737 			 */
738 			ptent = ptep_get_and_clear_full(mm, addr, pte,
739 							tlb->fullmm);
740 
741 			ptent = pte_mkold(ptent);
742 			ptent = pte_mkclean(ptent);
743 			set_pte_at(mm, addr, pte, ptent);
744 			tlb_remove_tlb_entry(tlb, pte, addr);
745 		}
746 		folio_mark_lazyfree(folio);
747 	}
748 
749 	if (nr_swap) {
750 		if (current->mm == mm)
751 			sync_mm_rss(mm);
752 		add_mm_counter(mm, MM_SWAPENTS, nr_swap);
753 	}
754 	if (start_pte) {
755 		arch_leave_lazy_mmu_mode();
756 		pte_unmap_unlock(start_pte, ptl);
757 	}
758 	cond_resched();
759 
760 	return 0;
761 }
762 
763 static const struct mm_walk_ops madvise_free_walk_ops = {
764 	.pmd_entry		= madvise_free_pte_range,
765 	.walk_lock		= PGWALK_RDLOCK,
766 };
767 
madvise_free_single_vma(struct vm_area_struct * vma,unsigned long start_addr,unsigned long end_addr)768 static int madvise_free_single_vma(struct vm_area_struct *vma,
769 			unsigned long start_addr, unsigned long end_addr)
770 {
771 	struct mm_struct *mm = vma->vm_mm;
772 	struct mmu_notifier_range range;
773 	struct mmu_gather tlb;
774 
775 	/* MADV_FREE works for only anon vma at the moment */
776 	if (!vma_is_anonymous(vma))
777 		return -EINVAL;
778 
779 	range.start = max(vma->vm_start, start_addr);
780 	if (range.start >= vma->vm_end)
781 		return -EINVAL;
782 	range.end = min(vma->vm_end, end_addr);
783 	if (range.end <= vma->vm_start)
784 		return -EINVAL;
785 	mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, mm,
786 				range.start, range.end);
787 
788 	lru_add_drain();
789 	tlb_gather_mmu(&tlb, mm);
790 	update_hiwater_rss(mm);
791 
792 	mmu_notifier_invalidate_range_start(&range);
793 	tlb_start_vma(&tlb, vma);
794 	walk_page_range(vma->vm_mm, range.start, range.end,
795 			&madvise_free_walk_ops, &tlb);
796 	tlb_end_vma(&tlb, vma);
797 	mmu_notifier_invalidate_range_end(&range);
798 	tlb_finish_mmu(&tlb);
799 
800 	return 0;
801 }
802 
803 /*
804  * Application no longer needs these pages.  If the pages are dirty,
805  * it's OK to just throw them away.  The app will be more careful about
806  * data it wants to keep.  Be sure to free swap resources too.  The
807  * zap_page_range_single call sets things up for shrink_active_list to actually
808  * free these pages later if no one else has touched them in the meantime,
809  * although we could add these pages to a global reuse list for
810  * shrink_active_list to pick up before reclaiming other pages.
811  *
812  * NB: This interface discards data rather than pushes it out to swap,
813  * as some implementations do.  This has performance implications for
814  * applications like large transactional databases which want to discard
815  * pages in anonymous maps after committing to backing store the data
816  * that was kept in them.  There is no reason to write this data out to
817  * the swap area if the application is discarding it.
818  *
819  * An interface that causes the system to free clean pages and flush
820  * dirty pages is already available as msync(MS_INVALIDATE).
821  */
madvise_dontneed_single_vma(struct vm_area_struct * vma,unsigned long start,unsigned long end)822 static long madvise_dontneed_single_vma(struct vm_area_struct *vma,
823 					unsigned long start, unsigned long end)
824 {
825 	zap_page_range_single(vma, start, end - start, NULL);
826 	return 0;
827 }
828 
madvise_dontneed_free_valid_vma(struct vm_area_struct * vma,unsigned long start,unsigned long * end,int behavior)829 static bool madvise_dontneed_free_valid_vma(struct vm_area_struct *vma,
830 					    unsigned long start,
831 					    unsigned long *end,
832 					    int behavior)
833 {
834 	if (!is_vm_hugetlb_page(vma)) {
835 		unsigned int forbidden = VM_PFNMAP;
836 
837 		if (behavior != MADV_DONTNEED_LOCKED)
838 			forbidden |= VM_LOCKED;
839 
840 		return !(vma->vm_flags & forbidden);
841 	}
842 
843 	if (behavior != MADV_DONTNEED && behavior != MADV_DONTNEED_LOCKED)
844 		return false;
845 	if (start & ~huge_page_mask(hstate_vma(vma)))
846 		return false;
847 
848 	/*
849 	 * Madvise callers expect the length to be rounded up to PAGE_SIZE
850 	 * boundaries, and may be unaware that this VMA uses huge pages.
851 	 * Avoid unexpected data loss by rounding down the number of
852 	 * huge pages freed.
853 	 */
854 	*end = ALIGN_DOWN(*end, huge_page_size(hstate_vma(vma)));
855 
856 	return true;
857 }
858 
madvise_dontneed_free(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,int behavior)859 static long madvise_dontneed_free(struct vm_area_struct *vma,
860 				  struct vm_area_struct **prev,
861 				  unsigned long start, unsigned long end,
862 				  int behavior)
863 {
864 	struct mm_struct *mm = vma->vm_mm;
865 
866 	*prev = vma;
867 	if (!madvise_dontneed_free_valid_vma(vma, start, &end, behavior))
868 		return -EINVAL;
869 
870 	if (start == end)
871 		return 0;
872 
873 	if (!userfaultfd_remove(vma, start, end)) {
874 		*prev = NULL; /* mmap_lock has been dropped, prev is stale */
875 
876 		mmap_read_lock(mm);
877 		vma = vma_lookup(mm, start);
878 		if (!vma)
879 			return -ENOMEM;
880 		/*
881 		 * Potential end adjustment for hugetlb vma is OK as
882 		 * the check below keeps end within vma.
883 		 */
884 		if (!madvise_dontneed_free_valid_vma(vma, start, &end,
885 						     behavior))
886 			return -EINVAL;
887 		if (end > vma->vm_end) {
888 			/*
889 			 * Don't fail if end > vma->vm_end. If the old
890 			 * vma was split while the mmap_lock was
891 			 * released the effect of the concurrent
892 			 * operation may not cause madvise() to
893 			 * have an undefined result. There may be an
894 			 * adjacent next vma that we'll walk
895 			 * next. userfaultfd_remove() will generate an
896 			 * UFFD_EVENT_REMOVE repetition on the
897 			 * end-vma->vm_end range, but the manager can
898 			 * handle a repetition fine.
899 			 */
900 			end = vma->vm_end;
901 		}
902 		VM_WARN_ON(start >= end);
903 	}
904 
905 	if (behavior == MADV_DONTNEED || behavior == MADV_DONTNEED_LOCKED)
906 		return madvise_dontneed_single_vma(vma, start, end);
907 	else if (behavior == MADV_FREE)
908 		return madvise_free_single_vma(vma, start, end);
909 	else
910 		return -EINVAL;
911 }
912 
madvise_populate(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,int behavior)913 static long madvise_populate(struct vm_area_struct *vma,
914 			     struct vm_area_struct **prev,
915 			     unsigned long start, unsigned long end,
916 			     int behavior)
917 {
918 	const bool write = behavior == MADV_POPULATE_WRITE;
919 	struct mm_struct *mm = vma->vm_mm;
920 	unsigned long tmp_end;
921 	int locked = 1;
922 	long pages;
923 
924 	*prev = vma;
925 
926 	while (start < end) {
927 		/*
928 		 * We might have temporarily dropped the lock. For example,
929 		 * our VMA might have been split.
930 		 */
931 		if (!vma || start >= vma->vm_end) {
932 			vma = vma_lookup(mm, start);
933 			if (!vma)
934 				return -ENOMEM;
935 		}
936 
937 		tmp_end = min_t(unsigned long, end, vma->vm_end);
938 		/* Populate (prefault) page tables readable/writable. */
939 		pages = faultin_vma_page_range(vma, start, tmp_end, write,
940 					       &locked);
941 		if (!locked) {
942 			mmap_read_lock(mm);
943 			locked = 1;
944 			*prev = NULL;
945 			vma = NULL;
946 		}
947 		if (pages < 0) {
948 			switch (pages) {
949 			case -EINTR:
950 				return -EINTR;
951 			case -EINVAL: /* Incompatible mappings / permissions. */
952 				return -EINVAL;
953 			case -EHWPOISON:
954 				return -EHWPOISON;
955 			case -EFAULT: /* VM_FAULT_SIGBUS or VM_FAULT_SIGSEGV */
956 				return -EFAULT;
957 			default:
958 				pr_warn_once("%s: unhandled return value: %ld\n",
959 					     __func__, pages);
960 				fallthrough;
961 			case -ENOMEM:
962 				return -ENOMEM;
963 			}
964 		}
965 		start += pages * PAGE_SIZE;
966 	}
967 	return 0;
968 }
969 
970 /*
971  * Application wants to free up the pages and associated backing store.
972  * This is effectively punching a hole into the middle of a file.
973  */
madvise_remove(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end)974 static long madvise_remove(struct vm_area_struct *vma,
975 				struct vm_area_struct **prev,
976 				unsigned long start, unsigned long end)
977 {
978 	loff_t offset;
979 	int error;
980 	struct file *f;
981 	struct mm_struct *mm = vma->vm_mm;
982 
983 	*prev = NULL;	/* tell sys_madvise we drop mmap_lock */
984 
985 	if (vma->vm_flags & VM_LOCKED)
986 		return -EINVAL;
987 
988 	f = vma->vm_file;
989 
990 	if (!f || !f->f_mapping || !f->f_mapping->host) {
991 			return -EINVAL;
992 	}
993 
994 	if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
995 		return -EACCES;
996 
997 	offset = (loff_t)(start - vma->vm_start)
998 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
999 
1000 	/*
1001 	 * Filesystem's fallocate may need to take i_rwsem.  We need to
1002 	 * explicitly grab a reference because the vma (and hence the
1003 	 * vma's reference to the file) can go away as soon as we drop
1004 	 * mmap_lock.
1005 	 */
1006 	get_file(f);
1007 	if (userfaultfd_remove(vma, start, end)) {
1008 		/* mmap_lock was not released by userfaultfd_remove() */
1009 		mmap_read_unlock(mm);
1010 	}
1011 	error = vfs_fallocate(f,
1012 				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
1013 				offset, end - start);
1014 	fput(f);
1015 	mmap_read_lock(mm);
1016 	return error;
1017 }
1018 
1019 /*
1020  * Apply an madvise behavior to a region of a vma.  madvise_update_vma
1021  * will handle splitting a vm area into separate areas, each area with its own
1022  * behavior.
1023  */
madvise_vma_behavior(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,unsigned long behavior)1024 static int madvise_vma_behavior(struct vm_area_struct *vma,
1025 				struct vm_area_struct **prev,
1026 				unsigned long start, unsigned long end,
1027 				unsigned long behavior)
1028 {
1029 	int error;
1030 	struct anon_vma_name *anon_name;
1031 	unsigned long new_flags = vma->vm_flags;
1032 
1033 	switch (behavior) {
1034 	case MADV_REMOVE:
1035 		return madvise_remove(vma, prev, start, end);
1036 	case MADV_WILLNEED:
1037 		return madvise_willneed(vma, prev, start, end);
1038 	case MADV_COLD:
1039 		return madvise_cold(vma, prev, start, end);
1040 	case MADV_PAGEOUT:
1041 		return madvise_pageout(vma, prev, start, end);
1042 	case MADV_FREE:
1043 	case MADV_DONTNEED:
1044 	case MADV_DONTNEED_LOCKED:
1045 		return madvise_dontneed_free(vma, prev, start, end, behavior);
1046 	case MADV_POPULATE_READ:
1047 	case MADV_POPULATE_WRITE:
1048 		return madvise_populate(vma, prev, start, end, behavior);
1049 	case MADV_NORMAL:
1050 		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
1051 		break;
1052 	case MADV_SEQUENTIAL:
1053 		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
1054 		break;
1055 	case MADV_RANDOM:
1056 		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
1057 		break;
1058 	case MADV_DONTFORK:
1059 		new_flags |= VM_DONTCOPY;
1060 		break;
1061 	case MADV_DOFORK:
1062 		if (vma->vm_flags & VM_IO)
1063 			return -EINVAL;
1064 		new_flags &= ~VM_DONTCOPY;
1065 		break;
1066 	case MADV_WIPEONFORK:
1067 		/* MADV_WIPEONFORK is only supported on anonymous memory. */
1068 		if (vma->vm_file || vma->vm_flags & VM_SHARED)
1069 			return -EINVAL;
1070 		new_flags |= VM_WIPEONFORK;
1071 		break;
1072 	case MADV_KEEPONFORK:
1073 		new_flags &= ~VM_WIPEONFORK;
1074 		break;
1075 	case MADV_DONTDUMP:
1076 		new_flags |= VM_DONTDUMP;
1077 		break;
1078 	case MADV_DODUMP:
1079 		if (!is_vm_hugetlb_page(vma) && new_flags & VM_SPECIAL)
1080 			return -EINVAL;
1081 		new_flags &= ~VM_DONTDUMP;
1082 		break;
1083 	case MADV_MERGEABLE:
1084 	case MADV_UNMERGEABLE:
1085 		error = ksm_madvise(vma, start, end, behavior, &new_flags);
1086 		if (error)
1087 			goto out;
1088 		break;
1089 	case MADV_HUGEPAGE:
1090 	case MADV_NOHUGEPAGE:
1091 		error = hugepage_madvise(vma, &new_flags, behavior);
1092 		if (error)
1093 			goto out;
1094 		break;
1095 	case MADV_COLLAPSE:
1096 		return madvise_collapse(vma, prev, start, end);
1097 	}
1098 
1099 	anon_name = anon_vma_name(vma);
1100 	anon_vma_name_get(anon_name);
1101 	error = madvise_update_vma(vma, prev, start, end, new_flags,
1102 				   anon_name);
1103 	anon_vma_name_put(anon_name);
1104 
1105 out:
1106 	/*
1107 	 * madvise() returns EAGAIN if kernel resources, such as
1108 	 * slab, are temporarily unavailable.
1109 	 */
1110 	if (error == -ENOMEM)
1111 		error = -EAGAIN;
1112 	return error;
1113 }
1114 
1115 #ifdef CONFIG_MEMORY_FAILURE
1116 /*
1117  * Error injection support for memory error handling.
1118  */
madvise_inject_error(int behavior,unsigned long start,unsigned long end)1119 static int madvise_inject_error(int behavior,
1120 		unsigned long start, unsigned long end)
1121 {
1122 	unsigned long size;
1123 
1124 	if (!capable(CAP_SYS_ADMIN))
1125 		return -EPERM;
1126 
1127 
1128 	for (; start < end; start += size) {
1129 		unsigned long pfn;
1130 		struct page *page;
1131 		int ret;
1132 
1133 		ret = get_user_pages_fast(start, 1, 0, &page);
1134 		if (ret != 1)
1135 			return ret;
1136 		pfn = page_to_pfn(page);
1137 
1138 		/*
1139 		 * When soft offlining hugepages, after migrating the page
1140 		 * we dissolve it, therefore in the second loop "page" will
1141 		 * no longer be a compound page.
1142 		 */
1143 		size = page_size(compound_head(page));
1144 
1145 		if (behavior == MADV_SOFT_OFFLINE) {
1146 			pr_info("Soft offlining pfn %#lx at process virtual address %#lx\n",
1147 				 pfn, start);
1148 			ret = soft_offline_page(pfn, MF_COUNT_INCREASED);
1149 		} else {
1150 			pr_info("Injecting memory failure for pfn %#lx at process virtual address %#lx\n",
1151 				 pfn, start);
1152 			ret = memory_failure(pfn, MF_COUNT_INCREASED | MF_SW_SIMULATED);
1153 			if (ret == -EOPNOTSUPP)
1154 				ret = 0;
1155 		}
1156 
1157 		if (ret)
1158 			return ret;
1159 	}
1160 
1161 	return 0;
1162 }
1163 #endif
1164 
1165 static bool
madvise_behavior_valid(int behavior)1166 madvise_behavior_valid(int behavior)
1167 {
1168 	switch (behavior) {
1169 	case MADV_DOFORK:
1170 	case MADV_DONTFORK:
1171 	case MADV_NORMAL:
1172 	case MADV_SEQUENTIAL:
1173 	case MADV_RANDOM:
1174 	case MADV_REMOVE:
1175 	case MADV_WILLNEED:
1176 	case MADV_DONTNEED:
1177 	case MADV_DONTNEED_LOCKED:
1178 	case MADV_FREE:
1179 	case MADV_COLD:
1180 	case MADV_PAGEOUT:
1181 	case MADV_POPULATE_READ:
1182 	case MADV_POPULATE_WRITE:
1183 #ifdef CONFIG_KSM
1184 	case MADV_MERGEABLE:
1185 	case MADV_UNMERGEABLE:
1186 #endif
1187 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1188 	case MADV_HUGEPAGE:
1189 	case MADV_NOHUGEPAGE:
1190 	case MADV_COLLAPSE:
1191 #endif
1192 	case MADV_DONTDUMP:
1193 	case MADV_DODUMP:
1194 	case MADV_WIPEONFORK:
1195 	case MADV_KEEPONFORK:
1196 #ifdef CONFIG_MEMORY_FAILURE
1197 	case MADV_SOFT_OFFLINE:
1198 	case MADV_HWPOISON:
1199 #endif
1200 		return true;
1201 
1202 	default:
1203 		return false;
1204 	}
1205 }
1206 
process_madvise_behavior_valid(int behavior)1207 static bool process_madvise_behavior_valid(int behavior)
1208 {
1209 	switch (behavior) {
1210 	case MADV_COLD:
1211 	case MADV_PAGEOUT:
1212 	case MADV_WILLNEED:
1213 	case MADV_COLLAPSE:
1214 		return true;
1215 	default:
1216 		return false;
1217 	}
1218 }
1219 
1220 /*
1221  * Walk the vmas in range [start,end), and call the visit function on each one.
1222  * The visit function will get start and end parameters that cover the overlap
1223  * between the current vma and the original range.  Any unmapped regions in the
1224  * original range will result in this function returning -ENOMEM while still
1225  * calling the visit function on all of the existing vmas in the range.
1226  * Must be called with the mmap_lock held for reading or writing.
1227  */
1228 static
madvise_walk_vmas(struct mm_struct * mm,unsigned long start,unsigned long end,unsigned long arg,int (* visit)(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,unsigned long arg))1229 int madvise_walk_vmas(struct mm_struct *mm, unsigned long start,
1230 		      unsigned long end, unsigned long arg,
1231 		      int (*visit)(struct vm_area_struct *vma,
1232 				   struct vm_area_struct **prev, unsigned long start,
1233 				   unsigned long end, unsigned long arg))
1234 {
1235 	struct vm_area_struct *vma;
1236 	struct vm_area_struct *prev;
1237 	unsigned long tmp;
1238 	int unmapped_error = 0;
1239 
1240 	/*
1241 	 * If the interval [start,end) covers some unmapped address
1242 	 * ranges, just ignore them, but return -ENOMEM at the end.
1243 	 * - different from the way of handling in mlock etc.
1244 	 */
1245 	vma = find_vma_prev(mm, start, &prev);
1246 	if (vma && start > vma->vm_start)
1247 		prev = vma;
1248 
1249 	for (;;) {
1250 		int error;
1251 
1252 		/* Still start < end. */
1253 		if (!vma)
1254 			return -ENOMEM;
1255 
1256 		/* Here start < (end|vma->vm_end). */
1257 		if (start < vma->vm_start) {
1258 			unmapped_error = -ENOMEM;
1259 			start = vma->vm_start;
1260 			if (start >= end)
1261 				break;
1262 		}
1263 
1264 		/* Here vma->vm_start <= start < (end|vma->vm_end) */
1265 		tmp = vma->vm_end;
1266 		if (end < tmp)
1267 			tmp = end;
1268 
1269 		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
1270 		error = visit(vma, &prev, start, tmp, arg);
1271 		if (error)
1272 			return error;
1273 		start = tmp;
1274 		if (prev && start < prev->vm_end)
1275 			start = prev->vm_end;
1276 		if (start >= end)
1277 			break;
1278 		if (prev)
1279 			vma = find_vma(mm, prev->vm_end);
1280 		else	/* madvise_remove dropped mmap_lock */
1281 			vma = find_vma(mm, start);
1282 	}
1283 
1284 	return unmapped_error;
1285 }
1286 
1287 #ifdef CONFIG_ANON_VMA_NAME
madvise_vma_anon_name(struct vm_area_struct * vma,struct vm_area_struct ** prev,unsigned long start,unsigned long end,unsigned long anon_name)1288 static int madvise_vma_anon_name(struct vm_area_struct *vma,
1289 				 struct vm_area_struct **prev,
1290 				 unsigned long start, unsigned long end,
1291 				 unsigned long anon_name)
1292 {
1293 	int error;
1294 
1295 	/* Only anonymous mappings can be named */
1296 	if (vma->vm_file && !vma_is_anon_shmem(vma))
1297 		return -EBADF;
1298 
1299 	error = madvise_update_vma(vma, prev, start, end, vma->vm_flags,
1300 				   (struct anon_vma_name *)anon_name);
1301 
1302 	/*
1303 	 * madvise() returns EAGAIN if kernel resources, such as
1304 	 * slab, are temporarily unavailable.
1305 	 */
1306 	if (error == -ENOMEM)
1307 		error = -EAGAIN;
1308 	return error;
1309 }
1310 
madvise_set_anon_name(struct mm_struct * mm,unsigned long start,unsigned long len_in,struct anon_vma_name * anon_name)1311 int madvise_set_anon_name(struct mm_struct *mm, unsigned long start,
1312 			  unsigned long len_in, struct anon_vma_name *anon_name)
1313 {
1314 	unsigned long end;
1315 	unsigned long len;
1316 
1317 	if (start & ~PAGE_MASK)
1318 		return -EINVAL;
1319 	len = (len_in + ~PAGE_MASK) & PAGE_MASK;
1320 
1321 	/* Check to see whether len was rounded up from small -ve to zero */
1322 	if (len_in && !len)
1323 		return -EINVAL;
1324 
1325 	end = start + len;
1326 	if (end < start)
1327 		return -EINVAL;
1328 
1329 	if (end == start)
1330 		return 0;
1331 
1332 	return madvise_walk_vmas(mm, start, end, (unsigned long)anon_name,
1333 				 madvise_vma_anon_name);
1334 }
1335 #endif /* CONFIG_ANON_VMA_NAME */
1336 /*
1337  * The madvise(2) system call.
1338  *
1339  * Applications can use madvise() to advise the kernel how it should
1340  * handle paging I/O in this VM area.  The idea is to help the kernel
1341  * use appropriate read-ahead and caching techniques.  The information
1342  * provided is advisory only, and can be safely disregarded by the
1343  * kernel without affecting the correct operation of the application.
1344  *
1345  * behavior values:
1346  *  MADV_NORMAL - the default behavior is to read clusters.  This
1347  *		results in some read-ahead and read-behind.
1348  *  MADV_RANDOM - the system should read the minimum amount of data
1349  *		on any access, since it is unlikely that the appli-
1350  *		cation will need more than what it asks for.
1351  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
1352  *		once, so they can be aggressively read ahead, and
1353  *		can be freed soon after they are accessed.
1354  *  MADV_WILLNEED - the application is notifying the system to read
1355  *		some pages ahead.
1356  *  MADV_DONTNEED - the application is finished with the given range,
1357  *		so the kernel can free resources associated with it.
1358  *  MADV_FREE - the application marks pages in the given range as lazy free,
1359  *		where actual purges are postponed until memory pressure happens.
1360  *  MADV_REMOVE - the application wants to free up the given range of
1361  *		pages and associated backing store.
1362  *  MADV_DONTFORK - omit this area from child's address space when forking:
1363  *		typically, to avoid COWing pages pinned by get_user_pages().
1364  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
1365  *  MADV_WIPEONFORK - present the child process with zero-filled memory in this
1366  *              range after a fork.
1367  *  MADV_KEEPONFORK - undo the effect of MADV_WIPEONFORK
1368  *  MADV_HWPOISON - trigger memory error handler as if the given memory range
1369  *		were corrupted by unrecoverable hardware memory failure.
1370  *  MADV_SOFT_OFFLINE - try to soft-offline the given range of memory.
1371  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
1372  *		this area with pages of identical content from other such areas.
1373  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
1374  *  MADV_HUGEPAGE - the application wants to back the given range by transparent
1375  *		huge pages in the future. Existing pages might be coalesced and
1376  *		new pages might be allocated as THP.
1377  *  MADV_NOHUGEPAGE - mark the given range as not worth being backed by
1378  *		transparent huge pages so the existing pages will not be
1379  *		coalesced into THP and new pages will not be allocated as THP.
1380  *  MADV_COLLAPSE - synchronously coalesce pages into new THP.
1381  *  MADV_DONTDUMP - the application wants to prevent pages in the given range
1382  *		from being included in its core dump.
1383  *  MADV_DODUMP - cancel MADV_DONTDUMP: no longer exclude from core dump.
1384  *  MADV_COLD - the application is not expected to use this memory soon,
1385  *		deactivate pages in this range so that they can be reclaimed
1386  *		easily if memory pressure happens.
1387  *  MADV_PAGEOUT - the application is not expected to use this memory soon,
1388  *		page out the pages in this range immediately.
1389  *  MADV_POPULATE_READ - populate (prefault) page tables readable by
1390  *		triggering read faults if required
1391  *  MADV_POPULATE_WRITE - populate (prefault) page tables writable by
1392  *		triggering write faults if required
1393  *
1394  * return values:
1395  *  zero    - success
1396  *  -EINVAL - start + len < 0, start is not page-aligned,
1397  *		"behavior" is not a valid value, or application
1398  *		is attempting to release locked or shared pages,
1399  *		or the specified address range includes file, Huge TLB,
1400  *		MAP_SHARED or VMPFNMAP range.
1401  *  -ENOMEM - addresses in the specified range are not currently
1402  *		mapped, or are outside the AS of the process.
1403  *  -EIO    - an I/O error occurred while paging in data.
1404  *  -EBADF  - map exists, but area maps something that isn't a file.
1405  *  -EAGAIN - a kernel resource was temporarily unavailable.
1406  */
do_madvise(struct mm_struct * mm,unsigned long start,size_t len_in,int behavior)1407 int do_madvise(struct mm_struct *mm, unsigned long start, size_t len_in, int behavior)
1408 {
1409 	unsigned long end;
1410 	int error;
1411 	int write;
1412 	size_t len;
1413 	struct blk_plug plug;
1414 
1415 	if (!madvise_behavior_valid(behavior))
1416 		return -EINVAL;
1417 
1418 	if (!PAGE_ALIGNED(start))
1419 		return -EINVAL;
1420 	len = PAGE_ALIGN(len_in);
1421 
1422 	/* Check to see whether len was rounded up from small -ve to zero */
1423 	if (len_in && !len)
1424 		return -EINVAL;
1425 
1426 	end = start + len;
1427 	if (end < start)
1428 		return -EINVAL;
1429 
1430 	if (end == start)
1431 		return 0;
1432 
1433 #ifdef CONFIG_MEMORY_FAILURE
1434 	if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
1435 		return madvise_inject_error(behavior, start, start + len_in);
1436 #endif
1437 
1438 	write = madvise_need_mmap_write(behavior);
1439 	if (write) {
1440 		if (mmap_write_lock_killable(mm))
1441 			return -EINTR;
1442 	} else {
1443 		mmap_read_lock(mm);
1444 	}
1445 
1446 	start = untagged_addr_remote(mm, start);
1447 	end = start + len;
1448 
1449 	blk_start_plug(&plug);
1450 	error = madvise_walk_vmas(mm, start, end, behavior,
1451 			madvise_vma_behavior);
1452 	blk_finish_plug(&plug);
1453 	if (write)
1454 		mmap_write_unlock(mm);
1455 	else
1456 		mmap_read_unlock(mm);
1457 
1458 	return error;
1459 }
1460 
SYSCALL_DEFINE3(madvise,unsigned long,start,size_t,len_in,int,behavior)1461 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
1462 {
1463 	return do_madvise(current->mm, start, len_in, behavior);
1464 }
1465 
SYSCALL_DEFINE5(process_madvise,int,pidfd,const struct iovec __user *,vec,size_t,vlen,int,behavior,unsigned int,flags)1466 SYSCALL_DEFINE5(process_madvise, int, pidfd, const struct iovec __user *, vec,
1467 		size_t, vlen, int, behavior, unsigned int, flags)
1468 {
1469 	ssize_t ret;
1470 	struct iovec iovstack[UIO_FASTIOV];
1471 	struct iovec *iov = iovstack;
1472 	struct iov_iter iter;
1473 	struct task_struct *task;
1474 	struct mm_struct *mm;
1475 	size_t total_len;
1476 	unsigned int f_flags;
1477 
1478 	if (flags != 0) {
1479 		ret = -EINVAL;
1480 		goto out;
1481 	}
1482 
1483 	ret = import_iovec(ITER_DEST, vec, vlen, ARRAY_SIZE(iovstack), &iov, &iter);
1484 	if (ret < 0)
1485 		goto out;
1486 
1487 	task = pidfd_get_task(pidfd, &f_flags);
1488 	if (IS_ERR(task)) {
1489 		ret = PTR_ERR(task);
1490 		goto free_iov;
1491 	}
1492 
1493 	if (!process_madvise_behavior_valid(behavior)) {
1494 		ret = -EINVAL;
1495 		goto release_task;
1496 	}
1497 
1498 	/* Require PTRACE_MODE_READ to avoid leaking ASLR metadata. */
1499 	mm = mm_access(task, PTRACE_MODE_READ_FSCREDS);
1500 	if (IS_ERR_OR_NULL(mm)) {
1501 		ret = IS_ERR(mm) ? PTR_ERR(mm) : -ESRCH;
1502 		goto release_task;
1503 	}
1504 
1505 	/*
1506 	 * Require CAP_SYS_NICE for influencing process performance. Note that
1507 	 * only non-destructive hints are currently supported.
1508 	 */
1509 	if (!capable(CAP_SYS_NICE)) {
1510 		ret = -EPERM;
1511 		goto release_mm;
1512 	}
1513 
1514 	total_len = iov_iter_count(&iter);
1515 
1516 	while (iov_iter_count(&iter)) {
1517 		ret = do_madvise(mm, (unsigned long)iter_iov_addr(&iter),
1518 					iter_iov_len(&iter), behavior);
1519 		if (ret < 0)
1520 			break;
1521 		iov_iter_advance(&iter, iter_iov_len(&iter));
1522 	}
1523 
1524 	ret = (total_len - iov_iter_count(&iter)) ? : ret;
1525 
1526 release_mm:
1527 	mmput(mm);
1528 release_task:
1529 	put_task_struct(task);
1530 free_iov:
1531 	kfree(iov);
1532 out:
1533 	return ret;
1534 }
1535