Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/truncate.c - code for taking down pages from address_spaces
4 *
5 * Copyright (C) 2002, Linus Torvalds
6 *
7 * 10Sep2002 Andrew Morton
8 * Initial version.
9 */
10
11#include <linux/kernel.h>
12#include <linux/backing-dev.h>
13#include <linux/dax.h>
14#include <linux/gfp.h>
15#include <linux/mm.h>
16#include <linux/swap.h>
17#include <linux/export.h>
18#include <linux/pagemap.h>
19#include <linux/highmem.h>
20#include <linux/pagevec.h>
21#include <linux/task_io_accounting_ops.h>
22#include <linux/shmem_fs.h>
23#include <linux/rmap.h>
24#include "internal.h"
25
26/*
27 * Regular page slots are stabilized by the page lock even without the tree
28 * itself locked. These unlocked entries need verification under the tree
29 * lock.
30 */
31static inline void __clear_shadow_entry(struct address_space *mapping,
32 pgoff_t index, void *entry)
33{
34 XA_STATE(xas, &mapping->i_pages, index);
35
36 xas_set_update(&xas, workingset_update_node);
37 if (xas_load(&xas) != entry)
38 return;
39 xas_store(&xas, NULL);
40}
41
42static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
43 void *entry)
44{
45 spin_lock(&mapping->host->i_lock);
46 xa_lock_irq(&mapping->i_pages);
47 __clear_shadow_entry(mapping, index, entry);
48 xa_unlock_irq(&mapping->i_pages);
49 if (mapping_shrinkable(mapping))
50 inode_add_lru(mapping->host);
51 spin_unlock(&mapping->host->i_lock);
52}
53
54/*
55 * Unconditionally remove exceptional entries. Usually called from truncate
56 * path. Note that the folio_batch may be altered by this function by removing
57 * exceptional entries similar to what folio_batch_remove_exceptionals() does.
58 */
59static void truncate_folio_batch_exceptionals(struct address_space *mapping,
60 struct folio_batch *fbatch, pgoff_t *indices)
61{
62 int i, j;
63 bool dax;
64
65 /* Handled by shmem itself */
66 if (shmem_mapping(mapping))
67 return;
68
69 for (j = 0; j < folio_batch_count(fbatch); j++)
70 if (xa_is_value(fbatch->folios[j]))
71 break;
72
73 if (j == folio_batch_count(fbatch))
74 return;
75
76 dax = dax_mapping(mapping);
77 if (!dax) {
78 spin_lock(&mapping->host->i_lock);
79 xa_lock_irq(&mapping->i_pages);
80 }
81
82 for (i = j; i < folio_batch_count(fbatch); i++) {
83 struct folio *folio = fbatch->folios[i];
84 pgoff_t index = indices[i];
85
86 if (!xa_is_value(folio)) {
87 fbatch->folios[j++] = folio;
88 continue;
89 }
90
91 if (unlikely(dax)) {
92 dax_delete_mapping_entry(mapping, index);
93 continue;
94 }
95
96 __clear_shadow_entry(mapping, index, folio);
97 }
98
99 if (!dax) {
100 xa_unlock_irq(&mapping->i_pages);
101 if (mapping_shrinkable(mapping))
102 inode_add_lru(mapping->host);
103 spin_unlock(&mapping->host->i_lock);
104 }
105 fbatch->nr = j;
106}
107
108/*
109 * Invalidate exceptional entry if easily possible. This handles exceptional
110 * entries for invalidate_inode_pages().
111 */
112static int invalidate_exceptional_entry(struct address_space *mapping,
113 pgoff_t index, void *entry)
114{
115 /* Handled by shmem itself, or for DAX we do nothing. */
116 if (shmem_mapping(mapping) || dax_mapping(mapping))
117 return 1;
118 clear_shadow_entry(mapping, index, entry);
119 return 1;
120}
121
122/*
123 * Invalidate exceptional entry if clean. This handles exceptional entries for
124 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
125 */
126static int invalidate_exceptional_entry2(struct address_space *mapping,
127 pgoff_t index, void *entry)
128{
129 /* Handled by shmem itself */
130 if (shmem_mapping(mapping))
131 return 1;
132 if (dax_mapping(mapping))
133 return dax_invalidate_mapping_entry_sync(mapping, index);
134 clear_shadow_entry(mapping, index, entry);
135 return 1;
136}
137
138/**
139 * folio_invalidate - Invalidate part or all of a folio.
140 * @folio: The folio which is affected.
141 * @offset: start of the range to invalidate
142 * @length: length of the range to invalidate
143 *
144 * folio_invalidate() is called when all or part of the folio has become
145 * invalidated by a truncate operation.
146 *
147 * folio_invalidate() does not have to release all buffers, but it must
148 * ensure that no dirty buffer is left outside @offset and that no I/O
149 * is underway against any of the blocks which are outside the truncation
150 * point. Because the caller is about to free (and possibly reuse) those
151 * blocks on-disk.
152 */
153void folio_invalidate(struct folio *folio, size_t offset, size_t length)
154{
155 const struct address_space_operations *aops = folio->mapping->a_ops;
156
157 if (aops->invalidate_folio)
158 aops->invalidate_folio(folio, offset, length);
159}
160EXPORT_SYMBOL_GPL(folio_invalidate);
161
162/*
163 * If truncate cannot remove the fs-private metadata from the page, the page
164 * becomes orphaned. It will be left on the LRU and may even be mapped into
165 * user pagetables if we're racing with filemap_fault().
166 *
167 * We need to bail out if page->mapping is no longer equal to the original
168 * mapping. This happens a) when the VM reclaimed the page while we waited on
169 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
170 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
171 */
172static void truncate_cleanup_folio(struct folio *folio)
173{
174 if (folio_mapped(folio))
175 unmap_mapping_folio(folio);
176
177 if (folio_has_private(folio))
178 folio_invalidate(folio, 0, folio_size(folio));
179
180 /*
181 * Some filesystems seem to re-dirty the page even after
182 * the VM has canceled the dirty bit (eg ext3 journaling).
183 * Hence dirty accounting check is placed after invalidation.
184 */
185 folio_cancel_dirty(folio);
186 folio_clear_mappedtodisk(folio);
187}
188
189int truncate_inode_folio(struct address_space *mapping, struct folio *folio)
190{
191 if (folio->mapping != mapping)
192 return -EIO;
193
194 truncate_cleanup_folio(folio);
195 filemap_remove_folio(folio);
196 return 0;
197}
198
199/*
200 * Handle partial folios. The folio may be entirely within the
201 * range if a split has raced with us. If not, we zero the part of the
202 * folio that's within the [start, end] range, and then split the folio if
203 * it's large. split_page_range() will discard pages which now lie beyond
204 * i_size, and we rely on the caller to discard pages which lie within a
205 * newly created hole.
206 *
207 * Returns false if splitting failed so the caller can avoid
208 * discarding the entire folio which is stubbornly unsplit.
209 */
210bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
211{
212 loff_t pos = folio_pos(folio);
213 unsigned int offset, length;
214
215 if (pos < start)
216 offset = start - pos;
217 else
218 offset = 0;
219 length = folio_size(folio);
220 if (pos + length <= (u64)end)
221 length = length - offset;
222 else
223 length = end + 1 - pos - offset;
224
225 folio_wait_writeback(folio);
226 if (length == folio_size(folio)) {
227 truncate_inode_folio(folio->mapping, folio);
228 return true;
229 }
230
231 /*
232 * We may be zeroing pages we're about to discard, but it avoids
233 * doing a complex calculation here, and then doing the zeroing
234 * anyway if the page split fails.
235 */
236 folio_zero_range(folio, offset, length);
237
238 if (folio_has_private(folio))
239 folio_invalidate(folio, offset, length);
240 if (!folio_test_large(folio))
241 return true;
242 if (split_folio(folio) == 0)
243 return true;
244 if (folio_test_dirty(folio))
245 return false;
246 truncate_inode_folio(folio->mapping, folio);
247 return true;
248}
249
250/*
251 * Used to get rid of pages on hardware memory corruption.
252 */
253int generic_error_remove_folio(struct address_space *mapping,
254 struct folio *folio)
255{
256 if (!mapping)
257 return -EINVAL;
258 /*
259 * Only punch for normal data pages for now.
260 * Handling other types like directories would need more auditing.
261 */
262 if (!S_ISREG(mapping->host->i_mode))
263 return -EIO;
264 return truncate_inode_folio(mapping, folio);
265}
266EXPORT_SYMBOL(generic_error_remove_folio);
267
268/**
269 * mapping_evict_folio() - Remove an unused folio from the page-cache.
270 * @mapping: The mapping this folio belongs to.
271 * @folio: The folio to remove.
272 *
273 * Safely remove one folio from the page cache.
274 * It only drops clean, unused folios.
275 *
276 * Context: Folio must be locked.
277 * Return: The number of pages successfully removed.
278 */
279long mapping_evict_folio(struct address_space *mapping, struct folio *folio)
280{
281 /* The page may have been truncated before it was locked */
282 if (!mapping)
283 return 0;
284 if (folio_test_dirty(folio) || folio_test_writeback(folio))
285 return 0;
286 /* The refcount will be elevated if any page in the folio is mapped */
287 if (folio_ref_count(folio) >
288 folio_nr_pages(folio) + folio_has_private(folio) + 1)
289 return 0;
290 if (!filemap_release_folio(folio, 0))
291 return 0;
292
293 return remove_mapping(mapping, folio);
294}
295
296/**
297 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
298 * @mapping: mapping to truncate
299 * @lstart: offset from which to truncate
300 * @lend: offset to which to truncate (inclusive)
301 *
302 * Truncate the page cache, removing the pages that are between
303 * specified offsets (and zeroing out partial pages
304 * if lstart or lend + 1 is not page aligned).
305 *
306 * Truncate takes two passes - the first pass is nonblocking. It will not
307 * block on page locks and it will not block on writeback. The second pass
308 * will wait. This is to prevent as much IO as possible in the affected region.
309 * The first pass will remove most pages, so the search cost of the second pass
310 * is low.
311 *
312 * We pass down the cache-hot hint to the page freeing code. Even if the
313 * mapping is large, it is probably the case that the final pages are the most
314 * recently touched, and freeing happens in ascending file offset order.
315 *
316 * Note that since ->invalidate_folio() accepts range to invalidate
317 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
318 * page aligned properly.
319 */
320void truncate_inode_pages_range(struct address_space *mapping,
321 loff_t lstart, loff_t lend)
322{
323 pgoff_t start; /* inclusive */
324 pgoff_t end; /* exclusive */
325 struct folio_batch fbatch;
326 pgoff_t indices[PAGEVEC_SIZE];
327 pgoff_t index;
328 int i;
329 struct folio *folio;
330 bool same_folio;
331
332 if (mapping_empty(mapping))
333 return;
334
335 /*
336 * 'start' and 'end' always covers the range of pages to be fully
337 * truncated. Partial pages are covered with 'partial_start' at the
338 * start of the range and 'partial_end' at the end of the range.
339 * Note that 'end' is exclusive while 'lend' is inclusive.
340 */
341 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
342 if (lend == -1)
343 /*
344 * lend == -1 indicates end-of-file so we have to set 'end'
345 * to the highest possible pgoff_t and since the type is
346 * unsigned we're using -1.
347 */
348 end = -1;
349 else
350 end = (lend + 1) >> PAGE_SHIFT;
351
352 folio_batch_init(&fbatch);
353 index = start;
354 while (index < end && find_lock_entries(mapping, &index, end - 1,
355 &fbatch, indices)) {
356 truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
357 for (i = 0; i < folio_batch_count(&fbatch); i++)
358 truncate_cleanup_folio(fbatch.folios[i]);
359 delete_from_page_cache_batch(mapping, &fbatch);
360 for (i = 0; i < folio_batch_count(&fbatch); i++)
361 folio_unlock(fbatch.folios[i]);
362 folio_batch_release(&fbatch);
363 cond_resched();
364 }
365
366 same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
367 folio = __filemap_get_folio(mapping, lstart >> PAGE_SHIFT, FGP_LOCK, 0);
368 if (!IS_ERR(folio)) {
369 same_folio = lend < folio_pos(folio) + folio_size(folio);
370 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
371 start = folio_next_index(folio);
372 if (same_folio)
373 end = folio->index;
374 }
375 folio_unlock(folio);
376 folio_put(folio);
377 folio = NULL;
378 }
379
380 if (!same_folio) {
381 folio = __filemap_get_folio(mapping, lend >> PAGE_SHIFT,
382 FGP_LOCK, 0);
383 if (!IS_ERR(folio)) {
384 if (!truncate_inode_partial_folio(folio, lstart, lend))
385 end = folio->index;
386 folio_unlock(folio);
387 folio_put(folio);
388 }
389 }
390
391 index = start;
392 while (index < end) {
393 cond_resched();
394 if (!find_get_entries(mapping, &index, end - 1, &fbatch,
395 indices)) {
396 /* If all gone from start onwards, we're done */
397 if (index == start)
398 break;
399 /* Otherwise restart to make sure all gone */
400 index = start;
401 continue;
402 }
403
404 for (i = 0; i < folio_batch_count(&fbatch); i++) {
405 struct folio *folio = fbatch.folios[i];
406
407 /* We rely upon deletion not changing page->index */
408
409 if (xa_is_value(folio))
410 continue;
411
412 folio_lock(folio);
413 VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
414 folio_wait_writeback(folio);
415 truncate_inode_folio(mapping, folio);
416 folio_unlock(folio);
417 }
418 truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
419 folio_batch_release(&fbatch);
420 }
421}
422EXPORT_SYMBOL(truncate_inode_pages_range);
423
424/**
425 * truncate_inode_pages - truncate *all* the pages from an offset
426 * @mapping: mapping to truncate
427 * @lstart: offset from which to truncate
428 *
429 * Called under (and serialised by) inode->i_rwsem and
430 * mapping->invalidate_lock.
431 *
432 * Note: When this function returns, there can be a page in the process of
433 * deletion (inside __filemap_remove_folio()) in the specified range. Thus
434 * mapping->nrpages can be non-zero when this function returns even after
435 * truncation of the whole mapping.
436 */
437void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
438{
439 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
440}
441EXPORT_SYMBOL(truncate_inode_pages);
442
443/**
444 * truncate_inode_pages_final - truncate *all* pages before inode dies
445 * @mapping: mapping to truncate
446 *
447 * Called under (and serialized by) inode->i_rwsem.
448 *
449 * Filesystems have to use this in the .evict_inode path to inform the
450 * VM that this is the final truncate and the inode is going away.
451 */
452void truncate_inode_pages_final(struct address_space *mapping)
453{
454 /*
455 * Page reclaim can not participate in regular inode lifetime
456 * management (can't call iput()) and thus can race with the
457 * inode teardown. Tell it when the address space is exiting,
458 * so that it does not install eviction information after the
459 * final truncate has begun.
460 */
461 mapping_set_exiting(mapping);
462
463 if (!mapping_empty(mapping)) {
464 /*
465 * As truncation uses a lockless tree lookup, cycle
466 * the tree lock to make sure any ongoing tree
467 * modification that does not see AS_EXITING is
468 * completed before starting the final truncate.
469 */
470 xa_lock_irq(&mapping->i_pages);
471 xa_unlock_irq(&mapping->i_pages);
472 }
473
474 truncate_inode_pages(mapping, 0);
475}
476EXPORT_SYMBOL(truncate_inode_pages_final);
477
478/**
479 * mapping_try_invalidate - Invalidate all the evictable folios of one inode
480 * @mapping: the address_space which holds the folios to invalidate
481 * @start: the offset 'from' which to invalidate
482 * @end: the offset 'to' which to invalidate (inclusive)
483 * @nr_failed: How many folio invalidations failed
484 *
485 * This function is similar to invalidate_mapping_pages(), except that it
486 * returns the number of folios which could not be evicted in @nr_failed.
487 */
488unsigned long mapping_try_invalidate(struct address_space *mapping,
489 pgoff_t start, pgoff_t end, unsigned long *nr_failed)
490{
491 pgoff_t indices[PAGEVEC_SIZE];
492 struct folio_batch fbatch;
493 pgoff_t index = start;
494 unsigned long ret;
495 unsigned long count = 0;
496 int i;
497
498 folio_batch_init(&fbatch);
499 while (find_lock_entries(mapping, &index, end, &fbatch, indices)) {
500 for (i = 0; i < folio_batch_count(&fbatch); i++) {
501 struct folio *folio = fbatch.folios[i];
502
503 /* We rely upon deletion not changing folio->index */
504
505 if (xa_is_value(folio)) {
506 count += invalidate_exceptional_entry(mapping,
507 indices[i], folio);
508 continue;
509 }
510
511 ret = mapping_evict_folio(mapping, folio);
512 folio_unlock(folio);
513 /*
514 * Invalidation is a hint that the folio is no longer
515 * of interest and try to speed up its reclaim.
516 */
517 if (!ret) {
518 deactivate_file_folio(folio);
519 /* Likely in the lru cache of a remote CPU */
520 if (nr_failed)
521 (*nr_failed)++;
522 }
523 count += ret;
524 }
525 folio_batch_remove_exceptionals(&fbatch);
526 folio_batch_release(&fbatch);
527 cond_resched();
528 }
529 return count;
530}
531
532/**
533 * invalidate_mapping_pages - Invalidate all clean, unlocked cache of one inode
534 * @mapping: the address_space which holds the cache to invalidate
535 * @start: the offset 'from' which to invalidate
536 * @end: the offset 'to' which to invalidate (inclusive)
537 *
538 * This function removes pages that are clean, unmapped and unlocked,
539 * as well as shadow entries. It will not block on IO activity.
540 *
541 * If you want to remove all the pages of one inode, regardless of
542 * their use and writeback state, use truncate_inode_pages().
543 *
544 * Return: The number of indices that had their contents invalidated
545 */
546unsigned long invalidate_mapping_pages(struct address_space *mapping,
547 pgoff_t start, pgoff_t end)
548{
549 return mapping_try_invalidate(mapping, start, end, NULL);
550}
551EXPORT_SYMBOL(invalidate_mapping_pages);
552
553/*
554 * This is like mapping_evict_folio(), except it ignores the folio's
555 * refcount. We do this because invalidate_inode_pages2() needs stronger
556 * invalidation guarantees, and cannot afford to leave folios behind because
557 * shrink_page_list() has a temp ref on them, or because they're transiently
558 * sitting in the folio_add_lru() caches.
559 */
560static int invalidate_complete_folio2(struct address_space *mapping,
561 struct folio *folio)
562{
563 if (folio->mapping != mapping)
564 return 0;
565
566 if (!filemap_release_folio(folio, GFP_KERNEL))
567 return 0;
568
569 spin_lock(&mapping->host->i_lock);
570 xa_lock_irq(&mapping->i_pages);
571 if (folio_test_dirty(folio))
572 goto failed;
573
574 BUG_ON(folio_has_private(folio));
575 __filemap_remove_folio(folio, NULL);
576 xa_unlock_irq(&mapping->i_pages);
577 if (mapping_shrinkable(mapping))
578 inode_add_lru(mapping->host);
579 spin_unlock(&mapping->host->i_lock);
580
581 filemap_free_folio(mapping, folio);
582 return 1;
583failed:
584 xa_unlock_irq(&mapping->i_pages);
585 spin_unlock(&mapping->host->i_lock);
586 return 0;
587}
588
589static int folio_launder(struct address_space *mapping, struct folio *folio)
590{
591 if (!folio_test_dirty(folio))
592 return 0;
593 if (folio->mapping != mapping || mapping->a_ops->launder_folio == NULL)
594 return 0;
595 return mapping->a_ops->launder_folio(folio);
596}
597
598/**
599 * invalidate_inode_pages2_range - remove range of pages from an address_space
600 * @mapping: the address_space
601 * @start: the page offset 'from' which to invalidate
602 * @end: the page offset 'to' which to invalidate (inclusive)
603 *
604 * Any pages which are found to be mapped into pagetables are unmapped prior to
605 * invalidation.
606 *
607 * Return: -EBUSY if any pages could not be invalidated.
608 */
609int invalidate_inode_pages2_range(struct address_space *mapping,
610 pgoff_t start, pgoff_t end)
611{
612 pgoff_t indices[PAGEVEC_SIZE];
613 struct folio_batch fbatch;
614 pgoff_t index;
615 int i;
616 int ret = 0;
617 int ret2 = 0;
618 int did_range_unmap = 0;
619
620 if (mapping_empty(mapping))
621 return 0;
622
623 folio_batch_init(&fbatch);
624 index = start;
625 while (find_get_entries(mapping, &index, end, &fbatch, indices)) {
626 for (i = 0; i < folio_batch_count(&fbatch); i++) {
627 struct folio *folio = fbatch.folios[i];
628
629 /* We rely upon deletion not changing folio->index */
630
631 if (xa_is_value(folio)) {
632 if (!invalidate_exceptional_entry2(mapping,
633 indices[i], folio))
634 ret = -EBUSY;
635 continue;
636 }
637
638 if (!did_range_unmap && folio_mapped(folio)) {
639 /*
640 * If folio is mapped, before taking its lock,
641 * zap the rest of the file in one hit.
642 */
643 unmap_mapping_pages(mapping, indices[i],
644 (1 + end - indices[i]), false);
645 did_range_unmap = 1;
646 }
647
648 folio_lock(folio);
649 if (unlikely(folio->mapping != mapping)) {
650 folio_unlock(folio);
651 continue;
652 }
653 VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
654 folio_wait_writeback(folio);
655
656 if (folio_mapped(folio))
657 unmap_mapping_folio(folio);
658 BUG_ON(folio_mapped(folio));
659
660 ret2 = folio_launder(mapping, folio);
661 if (ret2 == 0) {
662 if (!invalidate_complete_folio2(mapping, folio))
663 ret2 = -EBUSY;
664 }
665 if (ret2 < 0)
666 ret = ret2;
667 folio_unlock(folio);
668 }
669 folio_batch_remove_exceptionals(&fbatch);
670 folio_batch_release(&fbatch);
671 cond_resched();
672 }
673 /*
674 * For DAX we invalidate page tables after invalidating page cache. We
675 * could invalidate page tables while invalidating each entry however
676 * that would be expensive. And doing range unmapping before doesn't
677 * work as we have no cheap way to find whether page cache entry didn't
678 * get remapped later.
679 */
680 if (dax_mapping(mapping)) {
681 unmap_mapping_pages(mapping, start, end - start + 1, false);
682 }
683 return ret;
684}
685EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
686
687/**
688 * invalidate_inode_pages2 - remove all pages from an address_space
689 * @mapping: the address_space
690 *
691 * Any pages which are found to be mapped into pagetables are unmapped prior to
692 * invalidation.
693 *
694 * Return: -EBUSY if any pages could not be invalidated.
695 */
696int invalidate_inode_pages2(struct address_space *mapping)
697{
698 return invalidate_inode_pages2_range(mapping, 0, -1);
699}
700EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
701
702/**
703 * truncate_pagecache - unmap and remove pagecache that has been truncated
704 * @inode: inode
705 * @newsize: new file size
706 *
707 * inode's new i_size must already be written before truncate_pagecache
708 * is called.
709 *
710 * This function should typically be called before the filesystem
711 * releases resources associated with the freed range (eg. deallocates
712 * blocks). This way, pagecache will always stay logically coherent
713 * with on-disk format, and the filesystem would not have to deal with
714 * situations such as writepage being called for a page that has already
715 * had its underlying blocks deallocated.
716 */
717void truncate_pagecache(struct inode *inode, loff_t newsize)
718{
719 struct address_space *mapping = inode->i_mapping;
720 loff_t holebegin = round_up(newsize, PAGE_SIZE);
721
722 /*
723 * unmap_mapping_range is called twice, first simply for
724 * efficiency so that truncate_inode_pages does fewer
725 * single-page unmaps. However after this first call, and
726 * before truncate_inode_pages finishes, it is possible for
727 * private pages to be COWed, which remain after
728 * truncate_inode_pages finishes, hence the second
729 * unmap_mapping_range call must be made for correctness.
730 */
731 unmap_mapping_range(mapping, holebegin, 0, 1);
732 truncate_inode_pages(mapping, newsize);
733 unmap_mapping_range(mapping, holebegin, 0, 1);
734}
735EXPORT_SYMBOL(truncate_pagecache);
736
737/**
738 * truncate_setsize - update inode and pagecache for a new file size
739 * @inode: inode
740 * @newsize: new file size
741 *
742 * truncate_setsize updates i_size and performs pagecache truncation (if
743 * necessary) to @newsize. It will be typically be called from the filesystem's
744 * setattr function when ATTR_SIZE is passed in.
745 *
746 * Must be called with a lock serializing truncates and writes (generally
747 * i_rwsem but e.g. xfs uses a different lock) and before all filesystem
748 * specific block truncation has been performed.
749 */
750void truncate_setsize(struct inode *inode, loff_t newsize)
751{
752 loff_t oldsize = inode->i_size;
753
754 i_size_write(inode, newsize);
755 if (newsize > oldsize)
756 pagecache_isize_extended(inode, oldsize, newsize);
757 truncate_pagecache(inode, newsize);
758}
759EXPORT_SYMBOL(truncate_setsize);
760
761/**
762 * pagecache_isize_extended - update pagecache after extension of i_size
763 * @inode: inode for which i_size was extended
764 * @from: original inode size
765 * @to: new inode size
766 *
767 * Handle extension of inode size either caused by extending truncate or by
768 * write starting after current i_size. We mark the page straddling current
769 * i_size RO so that page_mkwrite() is called on the nearest write access to
770 * the page. This way filesystem can be sure that page_mkwrite() is called on
771 * the page before user writes to the page via mmap after the i_size has been
772 * changed.
773 *
774 * The function must be called after i_size is updated so that page fault
775 * coming after we unlock the page will already see the new i_size.
776 * The function must be called while we still hold i_rwsem - this not only
777 * makes sure i_size is stable but also that userspace cannot observe new
778 * i_size value before we are prepared to store mmap writes at new inode size.
779 */
780void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
781{
782 int bsize = i_blocksize(inode);
783 loff_t rounded_from;
784 struct page *page;
785 pgoff_t index;
786
787 WARN_ON(to > inode->i_size);
788
789 if (from >= to || bsize == PAGE_SIZE)
790 return;
791 /* Page straddling @from will not have any hole block created? */
792 rounded_from = round_up(from, bsize);
793 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
794 return;
795
796 index = from >> PAGE_SHIFT;
797 page = find_lock_page(inode->i_mapping, index);
798 /* Page not cached? Nothing to do */
799 if (!page)
800 return;
801 /*
802 * See clear_page_dirty_for_io() for details why set_page_dirty()
803 * is needed.
804 */
805 if (page_mkclean(page))
806 set_page_dirty(page);
807 unlock_page(page);
808 put_page(page);
809}
810EXPORT_SYMBOL(pagecache_isize_extended);
811
812/**
813 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
814 * @inode: inode
815 * @lstart: offset of beginning of hole
816 * @lend: offset of last byte of hole
817 *
818 * This function should typically be called before the filesystem
819 * releases resources associated with the freed range (eg. deallocates
820 * blocks). This way, pagecache will always stay logically coherent
821 * with on-disk format, and the filesystem would not have to deal with
822 * situations such as writepage being called for a page that has already
823 * had its underlying blocks deallocated.
824 */
825void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
826{
827 struct address_space *mapping = inode->i_mapping;
828 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
829 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
830 /*
831 * This rounding is currently just for example: unmap_mapping_range
832 * expands its hole outwards, whereas we want it to contract the hole
833 * inwards. However, existing callers of truncate_pagecache_range are
834 * doing their own page rounding first. Note that unmap_mapping_range
835 * allows holelen 0 for all, and we allow lend -1 for end of file.
836 */
837
838 /*
839 * Unlike in truncate_pagecache, unmap_mapping_range is called only
840 * once (before truncating pagecache), and without "even_cows" flag:
841 * hole-punching should not remove private COWed pages from the hole.
842 */
843 if ((u64)unmap_end > (u64)unmap_start)
844 unmap_mapping_range(mapping, unmap_start,
845 1 + unmap_end - unmap_start, 0);
846 truncate_inode_pages_range(mapping, lstart, lend);
847}
848EXPORT_SYMBOL(truncate_pagecache_range);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/truncate.c - code for taking down pages from address_spaces
4 *
5 * Copyright (C) 2002, Linus Torvalds
6 *
7 * 10Sep2002 Andrew Morton
8 * Initial version.
9 */
10
11#include <linux/kernel.h>
12#include <linux/backing-dev.h>
13#include <linux/dax.h>
14#include <linux/gfp.h>
15#include <linux/mm.h>
16#include <linux/swap.h>
17#include <linux/export.h>
18#include <linux/pagemap.h>
19#include <linux/highmem.h>
20#include <linux/pagevec.h>
21#include <linux/task_io_accounting_ops.h>
22#include <linux/buffer_head.h> /* grr. try_to_release_page */
23#include <linux/shmem_fs.h>
24#include <linux/rmap.h>
25#include "internal.h"
26
27/*
28 * Regular page slots are stabilized by the page lock even without the tree
29 * itself locked. These unlocked entries need verification under the tree
30 * lock.
31 */
32static inline void __clear_shadow_entry(struct address_space *mapping,
33 pgoff_t index, void *entry)
34{
35 XA_STATE(xas, &mapping->i_pages, index);
36
37 xas_set_update(&xas, workingset_update_node);
38 if (xas_load(&xas) != entry)
39 return;
40 xas_store(&xas, NULL);
41}
42
43static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
44 void *entry)
45{
46 spin_lock(&mapping->host->i_lock);
47 xa_lock_irq(&mapping->i_pages);
48 __clear_shadow_entry(mapping, index, entry);
49 xa_unlock_irq(&mapping->i_pages);
50 if (mapping_shrinkable(mapping))
51 inode_add_lru(mapping->host);
52 spin_unlock(&mapping->host->i_lock);
53}
54
55/*
56 * Unconditionally remove exceptional entries. Usually called from truncate
57 * path. Note that the folio_batch may be altered by this function by removing
58 * exceptional entries similar to what folio_batch_remove_exceptionals() does.
59 */
60static void truncate_folio_batch_exceptionals(struct address_space *mapping,
61 struct folio_batch *fbatch, pgoff_t *indices)
62{
63 int i, j;
64 bool dax;
65
66 /* Handled by shmem itself */
67 if (shmem_mapping(mapping))
68 return;
69
70 for (j = 0; j < folio_batch_count(fbatch); j++)
71 if (xa_is_value(fbatch->folios[j]))
72 break;
73
74 if (j == folio_batch_count(fbatch))
75 return;
76
77 dax = dax_mapping(mapping);
78 if (!dax) {
79 spin_lock(&mapping->host->i_lock);
80 xa_lock_irq(&mapping->i_pages);
81 }
82
83 for (i = j; i < folio_batch_count(fbatch); i++) {
84 struct folio *folio = fbatch->folios[i];
85 pgoff_t index = indices[i];
86
87 if (!xa_is_value(folio)) {
88 fbatch->folios[j++] = folio;
89 continue;
90 }
91
92 if (unlikely(dax)) {
93 dax_delete_mapping_entry(mapping, index);
94 continue;
95 }
96
97 __clear_shadow_entry(mapping, index, folio);
98 }
99
100 if (!dax) {
101 xa_unlock_irq(&mapping->i_pages);
102 if (mapping_shrinkable(mapping))
103 inode_add_lru(mapping->host);
104 spin_unlock(&mapping->host->i_lock);
105 }
106 fbatch->nr = j;
107}
108
109/*
110 * Invalidate exceptional entry if easily possible. This handles exceptional
111 * entries for invalidate_inode_pages().
112 */
113static int invalidate_exceptional_entry(struct address_space *mapping,
114 pgoff_t index, void *entry)
115{
116 /* Handled by shmem itself, or for DAX we do nothing. */
117 if (shmem_mapping(mapping) || dax_mapping(mapping))
118 return 1;
119 clear_shadow_entry(mapping, index, entry);
120 return 1;
121}
122
123/*
124 * Invalidate exceptional entry if clean. This handles exceptional entries for
125 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
126 */
127static int invalidate_exceptional_entry2(struct address_space *mapping,
128 pgoff_t index, void *entry)
129{
130 /* Handled by shmem itself */
131 if (shmem_mapping(mapping))
132 return 1;
133 if (dax_mapping(mapping))
134 return dax_invalidate_mapping_entry_sync(mapping, index);
135 clear_shadow_entry(mapping, index, entry);
136 return 1;
137}
138
139/**
140 * folio_invalidate - Invalidate part or all of a folio.
141 * @folio: The folio which is affected.
142 * @offset: start of the range to invalidate
143 * @length: length of the range to invalidate
144 *
145 * folio_invalidate() is called when all or part of the folio has become
146 * invalidated by a truncate operation.
147 *
148 * folio_invalidate() does not have to release all buffers, but it must
149 * ensure that no dirty buffer is left outside @offset and that no I/O
150 * is underway against any of the blocks which are outside the truncation
151 * point. Because the caller is about to free (and possibly reuse) those
152 * blocks on-disk.
153 */
154void folio_invalidate(struct folio *folio, size_t offset, size_t length)
155{
156 const struct address_space_operations *aops = folio->mapping->a_ops;
157
158 if (aops->invalidate_folio)
159 aops->invalidate_folio(folio, offset, length);
160}
161EXPORT_SYMBOL_GPL(folio_invalidate);
162
163/*
164 * If truncate cannot remove the fs-private metadata from the page, the page
165 * becomes orphaned. It will be left on the LRU and may even be mapped into
166 * user pagetables if we're racing with filemap_fault().
167 *
168 * We need to bail out if page->mapping is no longer equal to the original
169 * mapping. This happens a) when the VM reclaimed the page while we waited on
170 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
171 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
172 */
173static void truncate_cleanup_folio(struct folio *folio)
174{
175 if (folio_mapped(folio))
176 unmap_mapping_folio(folio);
177
178 if (folio_has_private(folio))
179 folio_invalidate(folio, 0, folio_size(folio));
180
181 /*
182 * Some filesystems seem to re-dirty the page even after
183 * the VM has canceled the dirty bit (eg ext3 journaling).
184 * Hence dirty accounting check is placed after invalidation.
185 */
186 folio_cancel_dirty(folio);
187 folio_clear_mappedtodisk(folio);
188}
189
190int truncate_inode_folio(struct address_space *mapping, struct folio *folio)
191{
192 if (folio->mapping != mapping)
193 return -EIO;
194
195 truncate_cleanup_folio(folio);
196 filemap_remove_folio(folio);
197 return 0;
198}
199
200/*
201 * Handle partial folios. The folio may be entirely within the
202 * range if a split has raced with us. If not, we zero the part of the
203 * folio that's within the [start, end] range, and then split the folio if
204 * it's large. split_page_range() will discard pages which now lie beyond
205 * i_size, and we rely on the caller to discard pages which lie within a
206 * newly created hole.
207 *
208 * Returns false if splitting failed so the caller can avoid
209 * discarding the entire folio which is stubbornly unsplit.
210 */
211bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
212{
213 loff_t pos = folio_pos(folio);
214 unsigned int offset, length;
215
216 if (pos < start)
217 offset = start - pos;
218 else
219 offset = 0;
220 length = folio_size(folio);
221 if (pos + length <= (u64)end)
222 length = length - offset;
223 else
224 length = end + 1 - pos - offset;
225
226 folio_wait_writeback(folio);
227 if (length == folio_size(folio)) {
228 truncate_inode_folio(folio->mapping, folio);
229 return true;
230 }
231
232 /*
233 * We may be zeroing pages we're about to discard, but it avoids
234 * doing a complex calculation here, and then doing the zeroing
235 * anyway if the page split fails.
236 */
237 folio_zero_range(folio, offset, length);
238
239 if (folio_has_private(folio))
240 folio_invalidate(folio, offset, length);
241 if (!folio_test_large(folio))
242 return true;
243 if (split_folio(folio) == 0)
244 return true;
245 if (folio_test_dirty(folio))
246 return false;
247 truncate_inode_folio(folio->mapping, folio);
248 return true;
249}
250
251/*
252 * Used to get rid of pages on hardware memory corruption.
253 */
254int generic_error_remove_page(struct address_space *mapping, struct page *page)
255{
256 VM_BUG_ON_PAGE(PageTail(page), page);
257
258 if (!mapping)
259 return -EINVAL;
260 /*
261 * Only punch for normal data pages for now.
262 * Handling other types like directories would need more auditing.
263 */
264 if (!S_ISREG(mapping->host->i_mode))
265 return -EIO;
266 return truncate_inode_folio(mapping, page_folio(page));
267}
268EXPORT_SYMBOL(generic_error_remove_page);
269
270static long mapping_evict_folio(struct address_space *mapping,
271 struct folio *folio)
272{
273 if (folio_test_dirty(folio) || folio_test_writeback(folio))
274 return 0;
275 /* The refcount will be elevated if any page in the folio is mapped */
276 if (folio_ref_count(folio) >
277 folio_nr_pages(folio) + folio_has_private(folio) + 1)
278 return 0;
279 if (folio_has_private(folio) && !filemap_release_folio(folio, 0))
280 return 0;
281
282 return remove_mapping(mapping, folio);
283}
284
285/**
286 * invalidate_inode_page() - Remove an unused page from the pagecache.
287 * @page: The page to remove.
288 *
289 * Safely invalidate one page from its pagecache mapping.
290 * It only drops clean, unused pages.
291 *
292 * Context: Page must be locked.
293 * Return: The number of pages successfully removed.
294 */
295long invalidate_inode_page(struct page *page)
296{
297 struct folio *folio = page_folio(page);
298 struct address_space *mapping = folio_mapping(folio);
299
300 /* The page may have been truncated before it was locked */
301 if (!mapping)
302 return 0;
303 return mapping_evict_folio(mapping, folio);
304}
305
306/**
307 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
308 * @mapping: mapping to truncate
309 * @lstart: offset from which to truncate
310 * @lend: offset to which to truncate (inclusive)
311 *
312 * Truncate the page cache, removing the pages that are between
313 * specified offsets (and zeroing out partial pages
314 * if lstart or lend + 1 is not page aligned).
315 *
316 * Truncate takes two passes - the first pass is nonblocking. It will not
317 * block on page locks and it will not block on writeback. The second pass
318 * will wait. This is to prevent as much IO as possible in the affected region.
319 * The first pass will remove most pages, so the search cost of the second pass
320 * is low.
321 *
322 * We pass down the cache-hot hint to the page freeing code. Even if the
323 * mapping is large, it is probably the case that the final pages are the most
324 * recently touched, and freeing happens in ascending file offset order.
325 *
326 * Note that since ->invalidate_folio() accepts range to invalidate
327 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
328 * page aligned properly.
329 */
330void truncate_inode_pages_range(struct address_space *mapping,
331 loff_t lstart, loff_t lend)
332{
333 pgoff_t start; /* inclusive */
334 pgoff_t end; /* exclusive */
335 struct folio_batch fbatch;
336 pgoff_t indices[PAGEVEC_SIZE];
337 pgoff_t index;
338 int i;
339 struct folio *folio;
340 bool same_folio;
341
342 if (mapping_empty(mapping))
343 return;
344
345 /*
346 * 'start' and 'end' always covers the range of pages to be fully
347 * truncated. Partial pages are covered with 'partial_start' at the
348 * start of the range and 'partial_end' at the end of the range.
349 * Note that 'end' is exclusive while 'lend' is inclusive.
350 */
351 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
352 if (lend == -1)
353 /*
354 * lend == -1 indicates end-of-file so we have to set 'end'
355 * to the highest possible pgoff_t and since the type is
356 * unsigned we're using -1.
357 */
358 end = -1;
359 else
360 end = (lend + 1) >> PAGE_SHIFT;
361
362 folio_batch_init(&fbatch);
363 index = start;
364 while (index < end && find_lock_entries(mapping, &index, end - 1,
365 &fbatch, indices)) {
366 truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
367 for (i = 0; i < folio_batch_count(&fbatch); i++)
368 truncate_cleanup_folio(fbatch.folios[i]);
369 delete_from_page_cache_batch(mapping, &fbatch);
370 for (i = 0; i < folio_batch_count(&fbatch); i++)
371 folio_unlock(fbatch.folios[i]);
372 folio_batch_release(&fbatch);
373 cond_resched();
374 }
375
376 same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
377 folio = __filemap_get_folio(mapping, lstart >> PAGE_SHIFT, FGP_LOCK, 0);
378 if (folio) {
379 same_folio = lend < folio_pos(folio) + folio_size(folio);
380 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
381 start = folio->index + folio_nr_pages(folio);
382 if (same_folio)
383 end = folio->index;
384 }
385 folio_unlock(folio);
386 folio_put(folio);
387 folio = NULL;
388 }
389
390 if (!same_folio)
391 folio = __filemap_get_folio(mapping, lend >> PAGE_SHIFT,
392 FGP_LOCK, 0);
393 if (folio) {
394 if (!truncate_inode_partial_folio(folio, lstart, lend))
395 end = folio->index;
396 folio_unlock(folio);
397 folio_put(folio);
398 }
399
400 index = start;
401 while (index < end) {
402 cond_resched();
403 if (!find_get_entries(mapping, &index, end - 1, &fbatch,
404 indices)) {
405 /* If all gone from start onwards, we're done */
406 if (index == start)
407 break;
408 /* Otherwise restart to make sure all gone */
409 index = start;
410 continue;
411 }
412
413 for (i = 0; i < folio_batch_count(&fbatch); i++) {
414 struct folio *folio = fbatch.folios[i];
415
416 /* We rely upon deletion not changing page->index */
417
418 if (xa_is_value(folio))
419 continue;
420
421 folio_lock(folio);
422 VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
423 folio_wait_writeback(folio);
424 truncate_inode_folio(mapping, folio);
425 folio_unlock(folio);
426 }
427 truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
428 folio_batch_release(&fbatch);
429 }
430}
431EXPORT_SYMBOL(truncate_inode_pages_range);
432
433/**
434 * truncate_inode_pages - truncate *all* the pages from an offset
435 * @mapping: mapping to truncate
436 * @lstart: offset from which to truncate
437 *
438 * Called under (and serialised by) inode->i_rwsem and
439 * mapping->invalidate_lock.
440 *
441 * Note: When this function returns, there can be a page in the process of
442 * deletion (inside __filemap_remove_folio()) in the specified range. Thus
443 * mapping->nrpages can be non-zero when this function returns even after
444 * truncation of the whole mapping.
445 */
446void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
447{
448 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
449}
450EXPORT_SYMBOL(truncate_inode_pages);
451
452/**
453 * truncate_inode_pages_final - truncate *all* pages before inode dies
454 * @mapping: mapping to truncate
455 *
456 * Called under (and serialized by) inode->i_rwsem.
457 *
458 * Filesystems have to use this in the .evict_inode path to inform the
459 * VM that this is the final truncate and the inode is going away.
460 */
461void truncate_inode_pages_final(struct address_space *mapping)
462{
463 /*
464 * Page reclaim can not participate in regular inode lifetime
465 * management (can't call iput()) and thus can race with the
466 * inode teardown. Tell it when the address space is exiting,
467 * so that it does not install eviction information after the
468 * final truncate has begun.
469 */
470 mapping_set_exiting(mapping);
471
472 if (!mapping_empty(mapping)) {
473 /*
474 * As truncation uses a lockless tree lookup, cycle
475 * the tree lock to make sure any ongoing tree
476 * modification that does not see AS_EXITING is
477 * completed before starting the final truncate.
478 */
479 xa_lock_irq(&mapping->i_pages);
480 xa_unlock_irq(&mapping->i_pages);
481 }
482
483 truncate_inode_pages(mapping, 0);
484}
485EXPORT_SYMBOL(truncate_inode_pages_final);
486
487/**
488 * invalidate_mapping_pagevec - Invalidate all the unlocked pages of one inode
489 * @mapping: the address_space which holds the pages to invalidate
490 * @start: the offset 'from' which to invalidate
491 * @end: the offset 'to' which to invalidate (inclusive)
492 * @nr_pagevec: invalidate failed page number for caller
493 *
494 * This helper is similar to invalidate_mapping_pages(), except that it accounts
495 * for pages that are likely on a pagevec and counts them in @nr_pagevec, which
496 * will be used by the caller.
497 */
498unsigned long invalidate_mapping_pagevec(struct address_space *mapping,
499 pgoff_t start, pgoff_t end, unsigned long *nr_pagevec)
500{
501 pgoff_t indices[PAGEVEC_SIZE];
502 struct folio_batch fbatch;
503 pgoff_t index = start;
504 unsigned long ret;
505 unsigned long count = 0;
506 int i;
507
508 folio_batch_init(&fbatch);
509 while (find_lock_entries(mapping, &index, end, &fbatch, indices)) {
510 for (i = 0; i < folio_batch_count(&fbatch); i++) {
511 struct folio *folio = fbatch.folios[i];
512
513 /* We rely upon deletion not changing folio->index */
514
515 if (xa_is_value(folio)) {
516 count += invalidate_exceptional_entry(mapping,
517 indices[i], folio);
518 continue;
519 }
520
521 ret = mapping_evict_folio(mapping, folio);
522 folio_unlock(folio);
523 /*
524 * Invalidation is a hint that the folio is no longer
525 * of interest and try to speed up its reclaim.
526 */
527 if (!ret) {
528 deactivate_file_folio(folio);
529 /* It is likely on the pagevec of a remote CPU */
530 if (nr_pagevec)
531 (*nr_pagevec)++;
532 }
533 count += ret;
534 }
535 folio_batch_remove_exceptionals(&fbatch);
536 folio_batch_release(&fbatch);
537 cond_resched();
538 }
539 return count;
540}
541
542/**
543 * invalidate_mapping_pages - Invalidate all clean, unlocked cache of one inode
544 * @mapping: the address_space which holds the cache to invalidate
545 * @start: the offset 'from' which to invalidate
546 * @end: the offset 'to' which to invalidate (inclusive)
547 *
548 * This function removes pages that are clean, unmapped and unlocked,
549 * as well as shadow entries. It will not block on IO activity.
550 *
551 * If you want to remove all the pages of one inode, regardless of
552 * their use and writeback state, use truncate_inode_pages().
553 *
554 * Return: the number of the cache entries that were invalidated
555 */
556unsigned long invalidate_mapping_pages(struct address_space *mapping,
557 pgoff_t start, pgoff_t end)
558{
559 return invalidate_mapping_pagevec(mapping, start, end, NULL);
560}
561EXPORT_SYMBOL(invalidate_mapping_pages);
562
563/*
564 * This is like invalidate_inode_page(), except it ignores the page's
565 * refcount. We do this because invalidate_inode_pages2() needs stronger
566 * invalidation guarantees, and cannot afford to leave pages behind because
567 * shrink_page_list() has a temp ref on them, or because they're transiently
568 * sitting in the folio_add_lru() pagevecs.
569 */
570static int invalidate_complete_folio2(struct address_space *mapping,
571 struct folio *folio)
572{
573 if (folio->mapping != mapping)
574 return 0;
575
576 if (folio_has_private(folio) &&
577 !filemap_release_folio(folio, GFP_KERNEL))
578 return 0;
579
580 spin_lock(&mapping->host->i_lock);
581 xa_lock_irq(&mapping->i_pages);
582 if (folio_test_dirty(folio))
583 goto failed;
584
585 BUG_ON(folio_has_private(folio));
586 __filemap_remove_folio(folio, NULL);
587 xa_unlock_irq(&mapping->i_pages);
588 if (mapping_shrinkable(mapping))
589 inode_add_lru(mapping->host);
590 spin_unlock(&mapping->host->i_lock);
591
592 filemap_free_folio(mapping, folio);
593 return 1;
594failed:
595 xa_unlock_irq(&mapping->i_pages);
596 spin_unlock(&mapping->host->i_lock);
597 return 0;
598}
599
600static int folio_launder(struct address_space *mapping, struct folio *folio)
601{
602 if (!folio_test_dirty(folio))
603 return 0;
604 if (folio->mapping != mapping || mapping->a_ops->launder_folio == NULL)
605 return 0;
606 return mapping->a_ops->launder_folio(folio);
607}
608
609/**
610 * invalidate_inode_pages2_range - remove range of pages from an address_space
611 * @mapping: the address_space
612 * @start: the page offset 'from' which to invalidate
613 * @end: the page offset 'to' which to invalidate (inclusive)
614 *
615 * Any pages which are found to be mapped into pagetables are unmapped prior to
616 * invalidation.
617 *
618 * Return: -EBUSY if any pages could not be invalidated.
619 */
620int invalidate_inode_pages2_range(struct address_space *mapping,
621 pgoff_t start, pgoff_t end)
622{
623 pgoff_t indices[PAGEVEC_SIZE];
624 struct folio_batch fbatch;
625 pgoff_t index;
626 int i;
627 int ret = 0;
628 int ret2 = 0;
629 int did_range_unmap = 0;
630
631 if (mapping_empty(mapping))
632 return 0;
633
634 folio_batch_init(&fbatch);
635 index = start;
636 while (find_get_entries(mapping, &index, end, &fbatch, indices)) {
637 for (i = 0; i < folio_batch_count(&fbatch); i++) {
638 struct folio *folio = fbatch.folios[i];
639
640 /* We rely upon deletion not changing folio->index */
641
642 if (xa_is_value(folio)) {
643 if (!invalidate_exceptional_entry2(mapping,
644 indices[i], folio))
645 ret = -EBUSY;
646 continue;
647 }
648
649 if (!did_range_unmap && folio_mapped(folio)) {
650 /*
651 * If folio is mapped, before taking its lock,
652 * zap the rest of the file in one hit.
653 */
654 unmap_mapping_pages(mapping, indices[i],
655 (1 + end - indices[i]), false);
656 did_range_unmap = 1;
657 }
658
659 folio_lock(folio);
660 VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
661 if (folio->mapping != mapping) {
662 folio_unlock(folio);
663 continue;
664 }
665 folio_wait_writeback(folio);
666
667 if (folio_mapped(folio))
668 unmap_mapping_folio(folio);
669 BUG_ON(folio_mapped(folio));
670
671 ret2 = folio_launder(mapping, folio);
672 if (ret2 == 0) {
673 if (!invalidate_complete_folio2(mapping, folio))
674 ret2 = -EBUSY;
675 }
676 if (ret2 < 0)
677 ret = ret2;
678 folio_unlock(folio);
679 }
680 folio_batch_remove_exceptionals(&fbatch);
681 folio_batch_release(&fbatch);
682 cond_resched();
683 }
684 /*
685 * For DAX we invalidate page tables after invalidating page cache. We
686 * could invalidate page tables while invalidating each entry however
687 * that would be expensive. And doing range unmapping before doesn't
688 * work as we have no cheap way to find whether page cache entry didn't
689 * get remapped later.
690 */
691 if (dax_mapping(mapping)) {
692 unmap_mapping_pages(mapping, start, end - start + 1, false);
693 }
694 return ret;
695}
696EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
697
698/**
699 * invalidate_inode_pages2 - remove all pages from an address_space
700 * @mapping: the address_space
701 *
702 * Any pages which are found to be mapped into pagetables are unmapped prior to
703 * invalidation.
704 *
705 * Return: -EBUSY if any pages could not be invalidated.
706 */
707int invalidate_inode_pages2(struct address_space *mapping)
708{
709 return invalidate_inode_pages2_range(mapping, 0, -1);
710}
711EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
712
713/**
714 * truncate_pagecache - unmap and remove pagecache that has been truncated
715 * @inode: inode
716 * @newsize: new file size
717 *
718 * inode's new i_size must already be written before truncate_pagecache
719 * is called.
720 *
721 * This function should typically be called before the filesystem
722 * releases resources associated with the freed range (eg. deallocates
723 * blocks). This way, pagecache will always stay logically coherent
724 * with on-disk format, and the filesystem would not have to deal with
725 * situations such as writepage being called for a page that has already
726 * had its underlying blocks deallocated.
727 */
728void truncate_pagecache(struct inode *inode, loff_t newsize)
729{
730 struct address_space *mapping = inode->i_mapping;
731 loff_t holebegin = round_up(newsize, PAGE_SIZE);
732
733 /*
734 * unmap_mapping_range is called twice, first simply for
735 * efficiency so that truncate_inode_pages does fewer
736 * single-page unmaps. However after this first call, and
737 * before truncate_inode_pages finishes, it is possible for
738 * private pages to be COWed, which remain after
739 * truncate_inode_pages finishes, hence the second
740 * unmap_mapping_range call must be made for correctness.
741 */
742 unmap_mapping_range(mapping, holebegin, 0, 1);
743 truncate_inode_pages(mapping, newsize);
744 unmap_mapping_range(mapping, holebegin, 0, 1);
745}
746EXPORT_SYMBOL(truncate_pagecache);
747
748/**
749 * truncate_setsize - update inode and pagecache for a new file size
750 * @inode: inode
751 * @newsize: new file size
752 *
753 * truncate_setsize updates i_size and performs pagecache truncation (if
754 * necessary) to @newsize. It will be typically be called from the filesystem's
755 * setattr function when ATTR_SIZE is passed in.
756 *
757 * Must be called with a lock serializing truncates and writes (generally
758 * i_rwsem but e.g. xfs uses a different lock) and before all filesystem
759 * specific block truncation has been performed.
760 */
761void truncate_setsize(struct inode *inode, loff_t newsize)
762{
763 loff_t oldsize = inode->i_size;
764
765 i_size_write(inode, newsize);
766 if (newsize > oldsize)
767 pagecache_isize_extended(inode, oldsize, newsize);
768 truncate_pagecache(inode, newsize);
769}
770EXPORT_SYMBOL(truncate_setsize);
771
772/**
773 * pagecache_isize_extended - update pagecache after extension of i_size
774 * @inode: inode for which i_size was extended
775 * @from: original inode size
776 * @to: new inode size
777 *
778 * Handle extension of inode size either caused by extending truncate or by
779 * write starting after current i_size. We mark the page straddling current
780 * i_size RO so that page_mkwrite() is called on the nearest write access to
781 * the page. This way filesystem can be sure that page_mkwrite() is called on
782 * the page before user writes to the page via mmap after the i_size has been
783 * changed.
784 *
785 * The function must be called after i_size is updated so that page fault
786 * coming after we unlock the page will already see the new i_size.
787 * The function must be called while we still hold i_rwsem - this not only
788 * makes sure i_size is stable but also that userspace cannot observe new
789 * i_size value before we are prepared to store mmap writes at new inode size.
790 */
791void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
792{
793 int bsize = i_blocksize(inode);
794 loff_t rounded_from;
795 struct page *page;
796 pgoff_t index;
797
798 WARN_ON(to > inode->i_size);
799
800 if (from >= to || bsize == PAGE_SIZE)
801 return;
802 /* Page straddling @from will not have any hole block created? */
803 rounded_from = round_up(from, bsize);
804 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
805 return;
806
807 index = from >> PAGE_SHIFT;
808 page = find_lock_page(inode->i_mapping, index);
809 /* Page not cached? Nothing to do */
810 if (!page)
811 return;
812 /*
813 * See clear_page_dirty_for_io() for details why set_page_dirty()
814 * is needed.
815 */
816 if (page_mkclean(page))
817 set_page_dirty(page);
818 unlock_page(page);
819 put_page(page);
820}
821EXPORT_SYMBOL(pagecache_isize_extended);
822
823/**
824 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
825 * @inode: inode
826 * @lstart: offset of beginning of hole
827 * @lend: offset of last byte of hole
828 *
829 * This function should typically be called before the filesystem
830 * releases resources associated with the freed range (eg. deallocates
831 * blocks). This way, pagecache will always stay logically coherent
832 * with on-disk format, and the filesystem would not have to deal with
833 * situations such as writepage being called for a page that has already
834 * had its underlying blocks deallocated.
835 */
836void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
837{
838 struct address_space *mapping = inode->i_mapping;
839 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
840 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
841 /*
842 * This rounding is currently just for example: unmap_mapping_range
843 * expands its hole outwards, whereas we want it to contract the hole
844 * inwards. However, existing callers of truncate_pagecache_range are
845 * doing their own page rounding first. Note that unmap_mapping_range
846 * allows holelen 0 for all, and we allow lend -1 for end of file.
847 */
848
849 /*
850 * Unlike in truncate_pagecache, unmap_mapping_range is called only
851 * once (before truncating pagecache), and without "even_cows" flag:
852 * hole-punching should not remove private COWed pages from the hole.
853 */
854 if ((u64)unmap_end > (u64)unmap_start)
855 unmap_mapping_range(mapping, unmap_start,
856 1 + unmap_end - unmap_start, 0);
857 truncate_inode_pages_range(mapping, lstart, lend);
858}
859EXPORT_SYMBOL(truncate_pagecache_range);