xref: /linux/mm/madvise.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /*
2  *	linux/mm/madvise.c
3  *
4  * Copyright (C) 1999  Linus Torvalds
5  * Copyright (C) 2002  Christoph Hellwig
6  */
7 
8 #include <linux/mman.h>
9 #include <linux/pagemap.h>
10 #include <linux/syscalls.h>
11 #include <linux/mempolicy.h>
12 #include <linux/page-isolation.h>
13 #include <linux/hugetlb.h>
14 #include <linux/falloc.h>
15 #include <linux/sched.h>
16 #include <linux/ksm.h>
17 #include <linux/fs.h>
18 #include <linux/file.h>
19 #include <linux/blkdev.h>
20 #include <linux/backing-dev.h>
21 #include <linux/swap.h>
22 #include <linux/swapops.h>
23 
24 /*
25  * Any behaviour which results in changes to the vma->vm_flags needs to
26  * take mmap_sem for writing. Others, which simply traverse vmas, need
27  * to only take it for reading.
28  */
29 static int madvise_need_mmap_write(int behavior)
30 {
31 	switch (behavior) {
32 	case MADV_REMOVE:
33 	case MADV_WILLNEED:
34 	case MADV_DONTNEED:
35 		return 0;
36 	default:
37 		/* be safe, default to 1. list exceptions explicitly */
38 		return 1;
39 	}
40 }
41 
42 /*
43  * We can potentially split a vm area into separate
44  * areas, each area with its own behavior.
45  */
46 static long madvise_behavior(struct vm_area_struct *vma,
47 		     struct vm_area_struct **prev,
48 		     unsigned long start, unsigned long end, int behavior)
49 {
50 	struct mm_struct *mm = vma->vm_mm;
51 	int error = 0;
52 	pgoff_t pgoff;
53 	unsigned long new_flags = vma->vm_flags;
54 
55 	switch (behavior) {
56 	case MADV_NORMAL:
57 		new_flags = new_flags & ~VM_RAND_READ & ~VM_SEQ_READ;
58 		break;
59 	case MADV_SEQUENTIAL:
60 		new_flags = (new_flags & ~VM_RAND_READ) | VM_SEQ_READ;
61 		break;
62 	case MADV_RANDOM:
63 		new_flags = (new_flags & ~VM_SEQ_READ) | VM_RAND_READ;
64 		break;
65 	case MADV_DONTFORK:
66 		new_flags |= VM_DONTCOPY;
67 		break;
68 	case MADV_DOFORK:
69 		if (vma->vm_flags & VM_IO) {
70 			error = -EINVAL;
71 			goto out;
72 		}
73 		new_flags &= ~VM_DONTCOPY;
74 		break;
75 	case MADV_DONTDUMP:
76 		new_flags |= VM_DONTDUMP;
77 		break;
78 	case MADV_DODUMP:
79 		if (new_flags & VM_SPECIAL) {
80 			error = -EINVAL;
81 			goto out;
82 		}
83 		new_flags &= ~VM_DONTDUMP;
84 		break;
85 	case MADV_MERGEABLE:
86 	case MADV_UNMERGEABLE:
87 		error = ksm_madvise(vma, start, end, behavior, &new_flags);
88 		if (error)
89 			goto out;
90 		break;
91 	case MADV_HUGEPAGE:
92 	case MADV_NOHUGEPAGE:
93 		error = hugepage_madvise(vma, &new_flags, behavior);
94 		if (error)
95 			goto out;
96 		break;
97 	}
98 
99 	if (new_flags == vma->vm_flags) {
100 		*prev = vma;
101 		goto out;
102 	}
103 
104 	pgoff = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
105 	*prev = vma_merge(mm, *prev, start, end, new_flags, vma->anon_vma,
106 			  vma->vm_file, pgoff, vma_policy(vma),
107 			  vma->vm_userfaultfd_ctx);
108 	if (*prev) {
109 		vma = *prev;
110 		goto success;
111 	}
112 
113 	*prev = vma;
114 
115 	if (start != vma->vm_start) {
116 		error = split_vma(mm, vma, start, 1);
117 		if (error)
118 			goto out;
119 	}
120 
121 	if (end != vma->vm_end) {
122 		error = split_vma(mm, vma, end, 0);
123 		if (error)
124 			goto out;
125 	}
126 
127 success:
128 	/*
129 	 * vm_flags is protected by the mmap_sem held in write mode.
130 	 */
131 	vma->vm_flags = new_flags;
132 
133 out:
134 	if (error == -ENOMEM)
135 		error = -EAGAIN;
136 	return error;
137 }
138 
139 #ifdef CONFIG_SWAP
140 static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
141 	unsigned long end, struct mm_walk *walk)
142 {
143 	pte_t *orig_pte;
144 	struct vm_area_struct *vma = walk->private;
145 	unsigned long index;
146 
147 	if (pmd_none_or_trans_huge_or_clear_bad(pmd))
148 		return 0;
149 
150 	for (index = start; index != end; index += PAGE_SIZE) {
151 		pte_t pte;
152 		swp_entry_t entry;
153 		struct page *page;
154 		spinlock_t *ptl;
155 
156 		orig_pte = pte_offset_map_lock(vma->vm_mm, pmd, start, &ptl);
157 		pte = *(orig_pte + ((index - start) / PAGE_SIZE));
158 		pte_unmap_unlock(orig_pte, ptl);
159 
160 		if (pte_present(pte) || pte_none(pte))
161 			continue;
162 		entry = pte_to_swp_entry(pte);
163 		if (unlikely(non_swap_entry(entry)))
164 			continue;
165 
166 		page = read_swap_cache_async(entry, GFP_HIGHUSER_MOVABLE,
167 								vma, index);
168 		if (page)
169 			page_cache_release(page);
170 	}
171 
172 	return 0;
173 }
174 
175 static void force_swapin_readahead(struct vm_area_struct *vma,
176 		unsigned long start, unsigned long end)
177 {
178 	struct mm_walk walk = {
179 		.mm = vma->vm_mm,
180 		.pmd_entry = swapin_walk_pmd_entry,
181 		.private = vma,
182 	};
183 
184 	walk_page_range(start, end, &walk);
185 
186 	lru_add_drain();	/* Push any new pages onto the LRU now */
187 }
188 
189 static void force_shm_swapin_readahead(struct vm_area_struct *vma,
190 		unsigned long start, unsigned long end,
191 		struct address_space *mapping)
192 {
193 	pgoff_t index;
194 	struct page *page;
195 	swp_entry_t swap;
196 
197 	for (; start < end; start += PAGE_SIZE) {
198 		index = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
199 
200 		page = find_get_entry(mapping, index);
201 		if (!radix_tree_exceptional_entry(page)) {
202 			if (page)
203 				page_cache_release(page);
204 			continue;
205 		}
206 		swap = radix_to_swp_entry(page);
207 		page = read_swap_cache_async(swap, GFP_HIGHUSER_MOVABLE,
208 								NULL, 0);
209 		if (page)
210 			page_cache_release(page);
211 	}
212 
213 	lru_add_drain();	/* Push any new pages onto the LRU now */
214 }
215 #endif		/* CONFIG_SWAP */
216 
217 /*
218  * Schedule all required I/O operations.  Do not wait for completion.
219  */
220 static long madvise_willneed(struct vm_area_struct *vma,
221 			     struct vm_area_struct **prev,
222 			     unsigned long start, unsigned long end)
223 {
224 	struct file *file = vma->vm_file;
225 
226 #ifdef CONFIG_SWAP
227 	if (!file) {
228 		*prev = vma;
229 		force_swapin_readahead(vma, start, end);
230 		return 0;
231 	}
232 
233 	if (shmem_mapping(file->f_mapping)) {
234 		*prev = vma;
235 		force_shm_swapin_readahead(vma, start, end,
236 					file->f_mapping);
237 		return 0;
238 	}
239 #else
240 	if (!file)
241 		return -EBADF;
242 #endif
243 
244 	if (IS_DAX(file_inode(file))) {
245 		/* no bad return value, but ignore advice */
246 		return 0;
247 	}
248 
249 	*prev = vma;
250 	start = ((start - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
251 	if (end > vma->vm_end)
252 		end = vma->vm_end;
253 	end = ((end - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
254 
255 	force_page_cache_readahead(file->f_mapping, file, start, end - start);
256 	return 0;
257 }
258 
259 /*
260  * Application no longer needs these pages.  If the pages are dirty,
261  * it's OK to just throw them away.  The app will be more careful about
262  * data it wants to keep.  Be sure to free swap resources too.  The
263  * zap_page_range call sets things up for shrink_active_list to actually free
264  * these pages later if no one else has touched them in the meantime,
265  * although we could add these pages to a global reuse list for
266  * shrink_active_list to pick up before reclaiming other pages.
267  *
268  * NB: This interface discards data rather than pushes it out to swap,
269  * as some implementations do.  This has performance implications for
270  * applications like large transactional databases which want to discard
271  * pages in anonymous maps after committing to backing store the data
272  * that was kept in them.  There is no reason to write this data out to
273  * the swap area if the application is discarding it.
274  *
275  * An interface that causes the system to free clean pages and flush
276  * dirty pages is already available as msync(MS_INVALIDATE).
277  */
278 static long madvise_dontneed(struct vm_area_struct *vma,
279 			     struct vm_area_struct **prev,
280 			     unsigned long start, unsigned long end)
281 {
282 	*prev = vma;
283 	if (vma->vm_flags & (VM_LOCKED|VM_HUGETLB|VM_PFNMAP))
284 		return -EINVAL;
285 
286 	zap_page_range(vma, start, end - start, NULL);
287 	return 0;
288 }
289 
290 /*
291  * Application wants to free up the pages and associated backing store.
292  * This is effectively punching a hole into the middle of a file.
293  */
294 static long madvise_remove(struct vm_area_struct *vma,
295 				struct vm_area_struct **prev,
296 				unsigned long start, unsigned long end)
297 {
298 	loff_t offset;
299 	int error;
300 	struct file *f;
301 
302 	*prev = NULL;	/* tell sys_madvise we drop mmap_sem */
303 
304 	if (vma->vm_flags & VM_LOCKED)
305 		return -EINVAL;
306 
307 	f = vma->vm_file;
308 
309 	if (!f || !f->f_mapping || !f->f_mapping->host) {
310 			return -EINVAL;
311 	}
312 
313 	if ((vma->vm_flags & (VM_SHARED|VM_WRITE)) != (VM_SHARED|VM_WRITE))
314 		return -EACCES;
315 
316 	offset = (loff_t)(start - vma->vm_start)
317 			+ ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
318 
319 	/*
320 	 * Filesystem's fallocate may need to take i_mutex.  We need to
321 	 * explicitly grab a reference because the vma (and hence the
322 	 * vma's reference to the file) can go away as soon as we drop
323 	 * mmap_sem.
324 	 */
325 	get_file(f);
326 	up_read(&current->mm->mmap_sem);
327 	error = vfs_fallocate(f,
328 				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
329 				offset, end - start);
330 	fput(f);
331 	down_read(&current->mm->mmap_sem);
332 	return error;
333 }
334 
335 #ifdef CONFIG_MEMORY_FAILURE
336 /*
337  * Error injection support for memory error handling.
338  */
339 static int madvise_hwpoison(int bhv, unsigned long start, unsigned long end)
340 {
341 	struct page *p;
342 	if (!capable(CAP_SYS_ADMIN))
343 		return -EPERM;
344 	for (; start < end; start += PAGE_SIZE <<
345 				compound_order(compound_head(p))) {
346 		int ret;
347 
348 		ret = get_user_pages_fast(start, 1, 0, &p);
349 		if (ret != 1)
350 			return ret;
351 
352 		if (PageHWPoison(p)) {
353 			put_page(p);
354 			continue;
355 		}
356 		if (bhv == MADV_SOFT_OFFLINE) {
357 			pr_info("Soft offlining page %#lx at %#lx\n",
358 				page_to_pfn(p), start);
359 			ret = soft_offline_page(p, MF_COUNT_INCREASED);
360 			if (ret)
361 				return ret;
362 			continue;
363 		}
364 		pr_info("Injecting memory failure for page %#lx at %#lx\n",
365 		       page_to_pfn(p), start);
366 		/* Ignore return value for now */
367 		memory_failure(page_to_pfn(p), 0, MF_COUNT_INCREASED);
368 	}
369 	return 0;
370 }
371 #endif
372 
373 static long
374 madvise_vma(struct vm_area_struct *vma, struct vm_area_struct **prev,
375 		unsigned long start, unsigned long end, int behavior)
376 {
377 	switch (behavior) {
378 	case MADV_REMOVE:
379 		return madvise_remove(vma, prev, start, end);
380 	case MADV_WILLNEED:
381 		return madvise_willneed(vma, prev, start, end);
382 	case MADV_DONTNEED:
383 		return madvise_dontneed(vma, prev, start, end);
384 	default:
385 		return madvise_behavior(vma, prev, start, end, behavior);
386 	}
387 }
388 
389 static bool
390 madvise_behavior_valid(int behavior)
391 {
392 	switch (behavior) {
393 	case MADV_DOFORK:
394 	case MADV_DONTFORK:
395 	case MADV_NORMAL:
396 	case MADV_SEQUENTIAL:
397 	case MADV_RANDOM:
398 	case MADV_REMOVE:
399 	case MADV_WILLNEED:
400 	case MADV_DONTNEED:
401 #ifdef CONFIG_KSM
402 	case MADV_MERGEABLE:
403 	case MADV_UNMERGEABLE:
404 #endif
405 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
406 	case MADV_HUGEPAGE:
407 	case MADV_NOHUGEPAGE:
408 #endif
409 	case MADV_DONTDUMP:
410 	case MADV_DODUMP:
411 		return true;
412 
413 	default:
414 		return false;
415 	}
416 }
417 
418 /*
419  * The madvise(2) system call.
420  *
421  * Applications can use madvise() to advise the kernel how it should
422  * handle paging I/O in this VM area.  The idea is to help the kernel
423  * use appropriate read-ahead and caching techniques.  The information
424  * provided is advisory only, and can be safely disregarded by the
425  * kernel without affecting the correct operation of the application.
426  *
427  * behavior values:
428  *  MADV_NORMAL - the default behavior is to read clusters.  This
429  *		results in some read-ahead and read-behind.
430  *  MADV_RANDOM - the system should read the minimum amount of data
431  *		on any access, since it is unlikely that the appli-
432  *		cation will need more than what it asks for.
433  *  MADV_SEQUENTIAL - pages in the given range will probably be accessed
434  *		once, so they can be aggressively read ahead, and
435  *		can be freed soon after they are accessed.
436  *  MADV_WILLNEED - the application is notifying the system to read
437  *		some pages ahead.
438  *  MADV_DONTNEED - the application is finished with the given range,
439  *		so the kernel can free resources associated with it.
440  *  MADV_REMOVE - the application wants to free up the given range of
441  *		pages and associated backing store.
442  *  MADV_DONTFORK - omit this area from child's address space when forking:
443  *		typically, to avoid COWing pages pinned by get_user_pages().
444  *  MADV_DOFORK - cancel MADV_DONTFORK: no longer omit this area when forking.
445  *  MADV_MERGEABLE - the application recommends that KSM try to merge pages in
446  *		this area with pages of identical content from other such areas.
447  *  MADV_UNMERGEABLE- cancel MADV_MERGEABLE: no longer merge pages with others.
448  *
449  * return values:
450  *  zero    - success
451  *  -EINVAL - start + len < 0, start is not page-aligned,
452  *		"behavior" is not a valid value, or application
453  *		is attempting to release locked or shared pages.
454  *  -ENOMEM - addresses in the specified range are not currently
455  *		mapped, or are outside the AS of the process.
456  *  -EIO    - an I/O error occurred while paging in data.
457  *  -EBADF  - map exists, but area maps something that isn't a file.
458  *  -EAGAIN - a kernel resource was temporarily unavailable.
459  */
460 SYSCALL_DEFINE3(madvise, unsigned long, start, size_t, len_in, int, behavior)
461 {
462 	unsigned long end, tmp;
463 	struct vm_area_struct *vma, *prev;
464 	int unmapped_error = 0;
465 	int error = -EINVAL;
466 	int write;
467 	size_t len;
468 	struct blk_plug plug;
469 
470 #ifdef CONFIG_MEMORY_FAILURE
471 	if (behavior == MADV_HWPOISON || behavior == MADV_SOFT_OFFLINE)
472 		return madvise_hwpoison(behavior, start, start+len_in);
473 #endif
474 	if (!madvise_behavior_valid(behavior))
475 		return error;
476 
477 	if (start & ~PAGE_MASK)
478 		return error;
479 	len = (len_in + ~PAGE_MASK) & PAGE_MASK;
480 
481 	/* Check to see whether len was rounded up from small -ve to zero */
482 	if (len_in && !len)
483 		return error;
484 
485 	end = start + len;
486 	if (end < start)
487 		return error;
488 
489 	error = 0;
490 	if (end == start)
491 		return error;
492 
493 	write = madvise_need_mmap_write(behavior);
494 	if (write)
495 		down_write(&current->mm->mmap_sem);
496 	else
497 		down_read(&current->mm->mmap_sem);
498 
499 	/*
500 	 * If the interval [start,end) covers some unmapped address
501 	 * ranges, just ignore them, but return -ENOMEM at the end.
502 	 * - different from the way of handling in mlock etc.
503 	 */
504 	vma = find_vma_prev(current->mm, start, &prev);
505 	if (vma && start > vma->vm_start)
506 		prev = vma;
507 
508 	blk_start_plug(&plug);
509 	for (;;) {
510 		/* Still start < end. */
511 		error = -ENOMEM;
512 		if (!vma)
513 			goto out;
514 
515 		/* Here start < (end|vma->vm_end). */
516 		if (start < vma->vm_start) {
517 			unmapped_error = -ENOMEM;
518 			start = vma->vm_start;
519 			if (start >= end)
520 				goto out;
521 		}
522 
523 		/* Here vma->vm_start <= start < (end|vma->vm_end) */
524 		tmp = vma->vm_end;
525 		if (end < tmp)
526 			tmp = end;
527 
528 		/* Here vma->vm_start <= start < tmp <= (end|vma->vm_end). */
529 		error = madvise_vma(vma, &prev, start, tmp, behavior);
530 		if (error)
531 			goto out;
532 		start = tmp;
533 		if (prev && start < prev->vm_end)
534 			start = prev->vm_end;
535 		error = unmapped_error;
536 		if (start >= end)
537 			goto out;
538 		if (prev)
539 			vma = prev->vm_next;
540 		else	/* madvise_remove dropped mmap_sem */
541 			vma = find_vma(current->mm, start);
542 	}
543 out:
544 	blk_finish_plug(&plug);
545 	if (write)
546 		up_write(&current->mm->mmap_sem);
547 	else
548 		up_read(&current->mm->mmap_sem);
549 
550 	return error;
551 }
552