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/*
2 * mm/truncate.c - code for taking down pages from address_spaces
3 *
4 * Copyright (C) 2002, Linus Torvalds
5 *
6 * 10Sep2002 Andrew Morton
7 * Initial version.
8 */
9
10#include <linux/kernel.h>
11#include <linux/backing-dev.h>
12#include <linux/dax.h>
13#include <linux/gfp.h>
14#include <linux/mm.h>
15#include <linux/swap.h>
16#include <linux/export.h>
17#include <linux/pagemap.h>
18#include <linux/highmem.h>
19#include <linux/pagevec.h>
20#include <linux/task_io_accounting_ops.h>
21#include <linux/buffer_head.h> /* grr. try_to_release_page,
22 do_invalidatepage */
23#include <linux/shmem_fs.h>
24#include <linux/cleancache.h>
25#include <linux/rmap.h>
26#include "internal.h"
27
28/*
29 * Regular page slots are stabilized by the page lock even without the tree
30 * itself locked. These unlocked entries need verification under the tree
31 * lock.
32 */
33static inline void __clear_shadow_entry(struct address_space *mapping,
34 pgoff_t index, void *entry)
35{
36 struct radix_tree_node *node;
37 void **slot;
38
39 if (!__radix_tree_lookup(&mapping->i_pages, index, &node, &slot))
40 return;
41 if (*slot != entry)
42 return;
43 __radix_tree_replace(&mapping->i_pages, node, slot, NULL,
44 workingset_update_node);
45 mapping->nrexceptional--;
46}
47
48static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
49 void *entry)
50{
51 xa_lock_irq(&mapping->i_pages);
52 __clear_shadow_entry(mapping, index, entry);
53 xa_unlock_irq(&mapping->i_pages);
54}
55
56/*
57 * Unconditionally remove exceptional entries. Usually called from truncate
58 * path. Note that the pagevec may be altered by this function by removing
59 * exceptional entries similar to what pagevec_remove_exceptionals does.
60 */
61static void truncate_exceptional_pvec_entries(struct address_space *mapping,
62 struct pagevec *pvec, pgoff_t *indices,
63 pgoff_t end)
64{
65 int i, j;
66 bool dax, lock;
67
68 /* Handled by shmem itself */
69 if (shmem_mapping(mapping))
70 return;
71
72 for (j = 0; j < pagevec_count(pvec); j++)
73 if (radix_tree_exceptional_entry(pvec->pages[j]))
74 break;
75
76 if (j == pagevec_count(pvec))
77 return;
78
79 dax = dax_mapping(mapping);
80 lock = !dax && indices[j] < end;
81 if (lock)
82 xa_lock_irq(&mapping->i_pages);
83
84 for (i = j; i < pagevec_count(pvec); i++) {
85 struct page *page = pvec->pages[i];
86 pgoff_t index = indices[i];
87
88 if (!radix_tree_exceptional_entry(page)) {
89 pvec->pages[j++] = page;
90 continue;
91 }
92
93 if (index >= end)
94 continue;
95
96 if (unlikely(dax)) {
97 dax_delete_mapping_entry(mapping, index);
98 continue;
99 }
100
101 __clear_shadow_entry(mapping, index, page);
102 }
103
104 if (lock)
105 xa_unlock_irq(&mapping->i_pages);
106 pvec->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 * do_invalidatepage - invalidate part or all of a page
141 * @page: the page which is affected
142 * @offset: start of the range to invalidate
143 * @length: length of the range to invalidate
144 *
145 * do_invalidatepage() is called when all or part of the page has become
146 * invalidated by a truncate operation.
147 *
148 * do_invalidatepage() 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 do_invalidatepage(struct page *page, unsigned int offset,
155 unsigned int length)
156{
157 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
158
159 invalidatepage = page->mapping->a_ops->invalidatepage;
160#ifdef CONFIG_BLOCK
161 if (!invalidatepage)
162 invalidatepage = block_invalidatepage;
163#endif
164 if (invalidatepage)
165 (*invalidatepage)(page, offset, length);
166}
167
168/*
169 * If truncate cannot remove the fs-private metadata from the page, the page
170 * becomes orphaned. It will be left on the LRU and may even be mapped into
171 * user pagetables if we're racing with filemap_fault().
172 *
173 * We need to bale out if page->mapping is no longer equal to the original
174 * mapping. This happens a) when the VM reclaimed the page while we waited on
175 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
176 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
177 */
178static void
179truncate_cleanup_page(struct address_space *mapping, struct page *page)
180{
181 if (page_mapped(page)) {
182 pgoff_t nr = PageTransHuge(page) ? HPAGE_PMD_NR : 1;
183 unmap_mapping_pages(mapping, page->index, nr, false);
184 }
185
186 if (page_has_private(page))
187 do_invalidatepage(page, 0, PAGE_SIZE);
188
189 /*
190 * Some filesystems seem to re-dirty the page even after
191 * the VM has canceled the dirty bit (eg ext3 journaling).
192 * Hence dirty accounting check is placed after invalidation.
193 */
194 cancel_dirty_page(page);
195 ClearPageMappedToDisk(page);
196}
197
198/*
199 * This is for invalidate_mapping_pages(). That function can be called at
200 * any time, and is not supposed to throw away dirty pages. But pages can
201 * be marked dirty at any time too, so use remove_mapping which safely
202 * discards clean, unused pages.
203 *
204 * Returns non-zero if the page was successfully invalidated.
205 */
206static int
207invalidate_complete_page(struct address_space *mapping, struct page *page)
208{
209 int ret;
210
211 if (page->mapping != mapping)
212 return 0;
213
214 if (page_has_private(page) && !try_to_release_page(page, 0))
215 return 0;
216
217 ret = remove_mapping(mapping, page);
218
219 return ret;
220}
221
222int truncate_inode_page(struct address_space *mapping, struct page *page)
223{
224 VM_BUG_ON_PAGE(PageTail(page), page);
225
226 if (page->mapping != mapping)
227 return -EIO;
228
229 truncate_cleanup_page(mapping, page);
230 delete_from_page_cache(page);
231 return 0;
232}
233
234/*
235 * Used to get rid of pages on hardware memory corruption.
236 */
237int generic_error_remove_page(struct address_space *mapping, struct page *page)
238{
239 if (!mapping)
240 return -EINVAL;
241 /*
242 * Only punch for normal data pages for now.
243 * Handling other types like directories would need more auditing.
244 */
245 if (!S_ISREG(mapping->host->i_mode))
246 return -EIO;
247 return truncate_inode_page(mapping, page);
248}
249EXPORT_SYMBOL(generic_error_remove_page);
250
251/*
252 * Safely invalidate one page from its pagecache mapping.
253 * It only drops clean, unused pages. The page must be locked.
254 *
255 * Returns 1 if the page is successfully invalidated, otherwise 0.
256 */
257int invalidate_inode_page(struct page *page)
258{
259 struct address_space *mapping = page_mapping(page);
260 if (!mapping)
261 return 0;
262 if (PageDirty(page) || PageWriteback(page))
263 return 0;
264 if (page_mapped(page))
265 return 0;
266 return invalidate_complete_page(mapping, page);
267}
268
269/**
270 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
271 * @mapping: mapping to truncate
272 * @lstart: offset from which to truncate
273 * @lend: offset to which to truncate (inclusive)
274 *
275 * Truncate the page cache, removing the pages that are between
276 * specified offsets (and zeroing out partial pages
277 * if lstart or lend + 1 is not page aligned).
278 *
279 * Truncate takes two passes - the first pass is nonblocking. It will not
280 * block on page locks and it will not block on writeback. The second pass
281 * will wait. This is to prevent as much IO as possible in the affected region.
282 * The first pass will remove most pages, so the search cost of the second pass
283 * is low.
284 *
285 * We pass down the cache-hot hint to the page freeing code. Even if the
286 * mapping is large, it is probably the case that the final pages are the most
287 * recently touched, and freeing happens in ascending file offset order.
288 *
289 * Note that since ->invalidatepage() accepts range to invalidate
290 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
291 * page aligned properly.
292 */
293void truncate_inode_pages_range(struct address_space *mapping,
294 loff_t lstart, loff_t lend)
295{
296 pgoff_t start; /* inclusive */
297 pgoff_t end; /* exclusive */
298 unsigned int partial_start; /* inclusive */
299 unsigned int partial_end; /* exclusive */
300 struct pagevec pvec;
301 pgoff_t indices[PAGEVEC_SIZE];
302 pgoff_t index;
303 int i;
304
305 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
306 goto out;
307
308 /* Offsets within partial pages */
309 partial_start = lstart & (PAGE_SIZE - 1);
310 partial_end = (lend + 1) & (PAGE_SIZE - 1);
311
312 /*
313 * 'start' and 'end' always covers the range of pages to be fully
314 * truncated. Partial pages are covered with 'partial_start' at the
315 * start of the range and 'partial_end' at the end of the range.
316 * Note that 'end' is exclusive while 'lend' is inclusive.
317 */
318 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
319 if (lend == -1)
320 /*
321 * lend == -1 indicates end-of-file so we have to set 'end'
322 * to the highest possible pgoff_t and since the type is
323 * unsigned we're using -1.
324 */
325 end = -1;
326 else
327 end = (lend + 1) >> PAGE_SHIFT;
328
329 pagevec_init(&pvec);
330 index = start;
331 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
332 min(end - index, (pgoff_t)PAGEVEC_SIZE),
333 indices)) {
334 /*
335 * Pagevec array has exceptional entries and we may also fail
336 * to lock some pages. So we store pages that can be deleted
337 * in a new pagevec.
338 */
339 struct pagevec locked_pvec;
340
341 pagevec_init(&locked_pvec);
342 for (i = 0; i < pagevec_count(&pvec); i++) {
343 struct page *page = pvec.pages[i];
344
345 /* We rely upon deletion not changing page->index */
346 index = indices[i];
347 if (index >= end)
348 break;
349
350 if (radix_tree_exceptional_entry(page))
351 continue;
352
353 if (!trylock_page(page))
354 continue;
355 WARN_ON(page_to_index(page) != index);
356 if (PageWriteback(page)) {
357 unlock_page(page);
358 continue;
359 }
360 if (page->mapping != mapping) {
361 unlock_page(page);
362 continue;
363 }
364 pagevec_add(&locked_pvec, page);
365 }
366 for (i = 0; i < pagevec_count(&locked_pvec); i++)
367 truncate_cleanup_page(mapping, locked_pvec.pages[i]);
368 delete_from_page_cache_batch(mapping, &locked_pvec);
369 for (i = 0; i < pagevec_count(&locked_pvec); i++)
370 unlock_page(locked_pvec.pages[i]);
371 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
372 pagevec_release(&pvec);
373 cond_resched();
374 index++;
375 }
376 if (partial_start) {
377 struct page *page = find_lock_page(mapping, start - 1);
378 if (page) {
379 unsigned int top = PAGE_SIZE;
380 if (start > end) {
381 /* Truncation within a single page */
382 top = partial_end;
383 partial_end = 0;
384 }
385 wait_on_page_writeback(page);
386 zero_user_segment(page, partial_start, top);
387 cleancache_invalidate_page(mapping, page);
388 if (page_has_private(page))
389 do_invalidatepage(page, partial_start,
390 top - partial_start);
391 unlock_page(page);
392 put_page(page);
393 }
394 }
395 if (partial_end) {
396 struct page *page = find_lock_page(mapping, end);
397 if (page) {
398 wait_on_page_writeback(page);
399 zero_user_segment(page, 0, partial_end);
400 cleancache_invalidate_page(mapping, page);
401 if (page_has_private(page))
402 do_invalidatepage(page, 0,
403 partial_end);
404 unlock_page(page);
405 put_page(page);
406 }
407 }
408 /*
409 * If the truncation happened within a single page no pages
410 * will be released, just zeroed, so we can bail out now.
411 */
412 if (start >= end)
413 goto out;
414
415 index = start;
416 for ( ; ; ) {
417 cond_resched();
418 if (!pagevec_lookup_entries(&pvec, mapping, index,
419 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
420 /* If all gone from start onwards, we're done */
421 if (index == start)
422 break;
423 /* Otherwise restart to make sure all gone */
424 index = start;
425 continue;
426 }
427 if (index == start && indices[0] >= end) {
428 /* All gone out of hole to be punched, we're done */
429 pagevec_remove_exceptionals(&pvec);
430 pagevec_release(&pvec);
431 break;
432 }
433
434 for (i = 0; i < pagevec_count(&pvec); i++) {
435 struct page *page = pvec.pages[i];
436
437 /* We rely upon deletion not changing page->index */
438 index = indices[i];
439 if (index >= end) {
440 /* Restart punch to make sure all gone */
441 index = start - 1;
442 break;
443 }
444
445 if (radix_tree_exceptional_entry(page))
446 continue;
447
448 lock_page(page);
449 WARN_ON(page_to_index(page) != index);
450 wait_on_page_writeback(page);
451 truncate_inode_page(mapping, page);
452 unlock_page(page);
453 }
454 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
455 pagevec_release(&pvec);
456 index++;
457 }
458
459out:
460 cleancache_invalidate_inode(mapping);
461}
462EXPORT_SYMBOL(truncate_inode_pages_range);
463
464/**
465 * truncate_inode_pages - truncate *all* the pages from an offset
466 * @mapping: mapping to truncate
467 * @lstart: offset from which to truncate
468 *
469 * Called under (and serialised by) inode->i_mutex.
470 *
471 * Note: When this function returns, there can be a page in the process of
472 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
473 * mapping->nrpages can be non-zero when this function returns even after
474 * truncation of the whole mapping.
475 */
476void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
477{
478 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
479}
480EXPORT_SYMBOL(truncate_inode_pages);
481
482/**
483 * truncate_inode_pages_final - truncate *all* pages before inode dies
484 * @mapping: mapping to truncate
485 *
486 * Called under (and serialized by) inode->i_mutex.
487 *
488 * Filesystems have to use this in the .evict_inode path to inform the
489 * VM that this is the final truncate and the inode is going away.
490 */
491void truncate_inode_pages_final(struct address_space *mapping)
492{
493 unsigned long nrexceptional;
494 unsigned long nrpages;
495
496 /*
497 * Page reclaim can not participate in regular inode lifetime
498 * management (can't call iput()) and thus can race with the
499 * inode teardown. Tell it when the address space is exiting,
500 * so that it does not install eviction information after the
501 * final truncate has begun.
502 */
503 mapping_set_exiting(mapping);
504
505 /*
506 * When reclaim installs eviction entries, it increases
507 * nrexceptional first, then decreases nrpages. Make sure we see
508 * this in the right order or we might miss an entry.
509 */
510 nrpages = mapping->nrpages;
511 smp_rmb();
512 nrexceptional = mapping->nrexceptional;
513
514 if (nrpages || nrexceptional) {
515 /*
516 * As truncation uses a lockless tree lookup, cycle
517 * the tree lock to make sure any ongoing tree
518 * modification that does not see AS_EXITING is
519 * completed before starting the final truncate.
520 */
521 xa_lock_irq(&mapping->i_pages);
522 xa_unlock_irq(&mapping->i_pages);
523
524 truncate_inode_pages(mapping, 0);
525 }
526}
527EXPORT_SYMBOL(truncate_inode_pages_final);
528
529/**
530 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
531 * @mapping: the address_space which holds the pages to invalidate
532 * @start: the offset 'from' which to invalidate
533 * @end: the offset 'to' which to invalidate (inclusive)
534 *
535 * This function only removes the unlocked pages, if you want to
536 * remove all the pages of one inode, you must call truncate_inode_pages.
537 *
538 * invalidate_mapping_pages() will not block on IO activity. It will not
539 * invalidate pages which are dirty, locked, under writeback or mapped into
540 * pagetables.
541 */
542unsigned long invalidate_mapping_pages(struct address_space *mapping,
543 pgoff_t start, pgoff_t end)
544{
545 pgoff_t indices[PAGEVEC_SIZE];
546 struct pagevec pvec;
547 pgoff_t index = start;
548 unsigned long ret;
549 unsigned long count = 0;
550 int i;
551
552 pagevec_init(&pvec);
553 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
554 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
555 indices)) {
556 for (i = 0; i < pagevec_count(&pvec); i++) {
557 struct page *page = pvec.pages[i];
558
559 /* We rely upon deletion not changing page->index */
560 index = indices[i];
561 if (index > end)
562 break;
563
564 if (radix_tree_exceptional_entry(page)) {
565 invalidate_exceptional_entry(mapping, index,
566 page);
567 continue;
568 }
569
570 if (!trylock_page(page))
571 continue;
572
573 WARN_ON(page_to_index(page) != index);
574
575 /* Middle of THP: skip */
576 if (PageTransTail(page)) {
577 unlock_page(page);
578 continue;
579 } else if (PageTransHuge(page)) {
580 index += HPAGE_PMD_NR - 1;
581 i += HPAGE_PMD_NR - 1;
582 /*
583 * 'end' is in the middle of THP. Don't
584 * invalidate the page as the part outside of
585 * 'end' could be still useful.
586 */
587 if (index > end) {
588 unlock_page(page);
589 continue;
590 }
591 }
592
593 ret = invalidate_inode_page(page);
594 unlock_page(page);
595 /*
596 * Invalidation is a hint that the page is no longer
597 * of interest and try to speed up its reclaim.
598 */
599 if (!ret)
600 deactivate_file_page(page);
601 count += ret;
602 }
603 pagevec_remove_exceptionals(&pvec);
604 pagevec_release(&pvec);
605 cond_resched();
606 index++;
607 }
608 return count;
609}
610EXPORT_SYMBOL(invalidate_mapping_pages);
611
612/*
613 * This is like invalidate_complete_page(), except it ignores the page's
614 * refcount. We do this because invalidate_inode_pages2() needs stronger
615 * invalidation guarantees, and cannot afford to leave pages behind because
616 * shrink_page_list() has a temp ref on them, or because they're transiently
617 * sitting in the lru_cache_add() pagevecs.
618 */
619static int
620invalidate_complete_page2(struct address_space *mapping, struct page *page)
621{
622 unsigned long flags;
623
624 if (page->mapping != mapping)
625 return 0;
626
627 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
628 return 0;
629
630 xa_lock_irqsave(&mapping->i_pages, flags);
631 if (PageDirty(page))
632 goto failed;
633
634 BUG_ON(page_has_private(page));
635 __delete_from_page_cache(page, NULL);
636 xa_unlock_irqrestore(&mapping->i_pages, flags);
637
638 if (mapping->a_ops->freepage)
639 mapping->a_ops->freepage(page);
640
641 put_page(page); /* pagecache ref */
642 return 1;
643failed:
644 xa_unlock_irqrestore(&mapping->i_pages, flags);
645 return 0;
646}
647
648static int do_launder_page(struct address_space *mapping, struct page *page)
649{
650 if (!PageDirty(page))
651 return 0;
652 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
653 return 0;
654 return mapping->a_ops->launder_page(page);
655}
656
657/**
658 * invalidate_inode_pages2_range - remove range of pages from an address_space
659 * @mapping: the address_space
660 * @start: the page offset 'from' which to invalidate
661 * @end: the page offset 'to' which to invalidate (inclusive)
662 *
663 * Any pages which are found to be mapped into pagetables are unmapped prior to
664 * invalidation.
665 *
666 * Returns -EBUSY if any pages could not be invalidated.
667 */
668int invalidate_inode_pages2_range(struct address_space *mapping,
669 pgoff_t start, pgoff_t end)
670{
671 pgoff_t indices[PAGEVEC_SIZE];
672 struct pagevec pvec;
673 pgoff_t index;
674 int i;
675 int ret = 0;
676 int ret2 = 0;
677 int did_range_unmap = 0;
678
679 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
680 goto out;
681
682 pagevec_init(&pvec);
683 index = start;
684 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
685 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
686 indices)) {
687 for (i = 0; i < pagevec_count(&pvec); i++) {
688 struct page *page = pvec.pages[i];
689
690 /* We rely upon deletion not changing page->index */
691 index = indices[i];
692 if (index > end)
693 break;
694
695 if (radix_tree_exceptional_entry(page)) {
696 if (!invalidate_exceptional_entry2(mapping,
697 index, page))
698 ret = -EBUSY;
699 continue;
700 }
701
702 lock_page(page);
703 WARN_ON(page_to_index(page) != index);
704 if (page->mapping != mapping) {
705 unlock_page(page);
706 continue;
707 }
708 wait_on_page_writeback(page);
709 if (page_mapped(page)) {
710 if (!did_range_unmap) {
711 /*
712 * Zap the rest of the file in one hit.
713 */
714 unmap_mapping_pages(mapping, index,
715 (1 + end - index), false);
716 did_range_unmap = 1;
717 } else {
718 /*
719 * Just zap this page
720 */
721 unmap_mapping_pages(mapping, index,
722 1, false);
723 }
724 }
725 BUG_ON(page_mapped(page));
726 ret2 = do_launder_page(mapping, page);
727 if (ret2 == 0) {
728 if (!invalidate_complete_page2(mapping, page))
729 ret2 = -EBUSY;
730 }
731 if (ret2 < 0)
732 ret = ret2;
733 unlock_page(page);
734 }
735 pagevec_remove_exceptionals(&pvec);
736 pagevec_release(&pvec);
737 cond_resched();
738 index++;
739 }
740 /*
741 * For DAX we invalidate page tables after invalidating radix tree. We
742 * could invalidate page tables while invalidating each entry however
743 * that would be expensive. And doing range unmapping before doesn't
744 * work as we have no cheap way to find whether radix tree entry didn't
745 * get remapped later.
746 */
747 if (dax_mapping(mapping)) {
748 unmap_mapping_pages(mapping, start, end - start + 1, false);
749 }
750out:
751 cleancache_invalidate_inode(mapping);
752 return ret;
753}
754EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
755
756/**
757 * invalidate_inode_pages2 - remove all pages from an address_space
758 * @mapping: the address_space
759 *
760 * Any pages which are found to be mapped into pagetables are unmapped prior to
761 * invalidation.
762 *
763 * Returns -EBUSY if any pages could not be invalidated.
764 */
765int invalidate_inode_pages2(struct address_space *mapping)
766{
767 return invalidate_inode_pages2_range(mapping, 0, -1);
768}
769EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
770
771/**
772 * truncate_pagecache - unmap and remove pagecache that has been truncated
773 * @inode: inode
774 * @newsize: new file size
775 *
776 * inode's new i_size must already be written before truncate_pagecache
777 * is called.
778 *
779 * This function should typically be called before the filesystem
780 * releases resources associated with the freed range (eg. deallocates
781 * blocks). This way, pagecache will always stay logically coherent
782 * with on-disk format, and the filesystem would not have to deal with
783 * situations such as writepage being called for a page that has already
784 * had its underlying blocks deallocated.
785 */
786void truncate_pagecache(struct inode *inode, loff_t newsize)
787{
788 struct address_space *mapping = inode->i_mapping;
789 loff_t holebegin = round_up(newsize, PAGE_SIZE);
790
791 /*
792 * unmap_mapping_range is called twice, first simply for
793 * efficiency so that truncate_inode_pages does fewer
794 * single-page unmaps. However after this first call, and
795 * before truncate_inode_pages finishes, it is possible for
796 * private pages to be COWed, which remain after
797 * truncate_inode_pages finishes, hence the second
798 * unmap_mapping_range call must be made for correctness.
799 */
800 unmap_mapping_range(mapping, holebegin, 0, 1);
801 truncate_inode_pages(mapping, newsize);
802 unmap_mapping_range(mapping, holebegin, 0, 1);
803}
804EXPORT_SYMBOL(truncate_pagecache);
805
806/**
807 * truncate_setsize - update inode and pagecache for a new file size
808 * @inode: inode
809 * @newsize: new file size
810 *
811 * truncate_setsize updates i_size and performs pagecache truncation (if
812 * necessary) to @newsize. It will be typically be called from the filesystem's
813 * setattr function when ATTR_SIZE is passed in.
814 *
815 * Must be called with a lock serializing truncates and writes (generally
816 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
817 * specific block truncation has been performed.
818 */
819void truncate_setsize(struct inode *inode, loff_t newsize)
820{
821 loff_t oldsize = inode->i_size;
822
823 i_size_write(inode, newsize);
824 if (newsize > oldsize)
825 pagecache_isize_extended(inode, oldsize, newsize);
826 truncate_pagecache(inode, newsize);
827}
828EXPORT_SYMBOL(truncate_setsize);
829
830/**
831 * pagecache_isize_extended - update pagecache after extension of i_size
832 * @inode: inode for which i_size was extended
833 * @from: original inode size
834 * @to: new inode size
835 *
836 * Handle extension of inode size either caused by extending truncate or by
837 * write starting after current i_size. We mark the page straddling current
838 * i_size RO so that page_mkwrite() is called on the nearest write access to
839 * the page. This way filesystem can be sure that page_mkwrite() is called on
840 * the page before user writes to the page via mmap after the i_size has been
841 * changed.
842 *
843 * The function must be called after i_size is updated so that page fault
844 * coming after we unlock the page will already see the new i_size.
845 * The function must be called while we still hold i_mutex - this not only
846 * makes sure i_size is stable but also that userspace cannot observe new
847 * i_size value before we are prepared to store mmap writes at new inode size.
848 */
849void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
850{
851 int bsize = i_blocksize(inode);
852 loff_t rounded_from;
853 struct page *page;
854 pgoff_t index;
855
856 WARN_ON(to > inode->i_size);
857
858 if (from >= to || bsize == PAGE_SIZE)
859 return;
860 /* Page straddling @from will not have any hole block created? */
861 rounded_from = round_up(from, bsize);
862 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
863 return;
864
865 index = from >> PAGE_SHIFT;
866 page = find_lock_page(inode->i_mapping, index);
867 /* Page not cached? Nothing to do */
868 if (!page)
869 return;
870 /*
871 * See clear_page_dirty_for_io() for details why set_page_dirty()
872 * is needed.
873 */
874 if (page_mkclean(page))
875 set_page_dirty(page);
876 unlock_page(page);
877 put_page(page);
878}
879EXPORT_SYMBOL(pagecache_isize_extended);
880
881/**
882 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
883 * @inode: inode
884 * @lstart: offset of beginning of hole
885 * @lend: offset of last byte of hole
886 *
887 * This function should typically be called before the filesystem
888 * releases resources associated with the freed range (eg. deallocates
889 * blocks). This way, pagecache will always stay logically coherent
890 * with on-disk format, and the filesystem would not have to deal with
891 * situations such as writepage being called for a page that has already
892 * had its underlying blocks deallocated.
893 */
894void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
895{
896 struct address_space *mapping = inode->i_mapping;
897 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
898 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
899 /*
900 * This rounding is currently just for example: unmap_mapping_range
901 * expands its hole outwards, whereas we want it to contract the hole
902 * inwards. However, existing callers of truncate_pagecache_range are
903 * doing their own page rounding first. Note that unmap_mapping_range
904 * allows holelen 0 for all, and we allow lend -1 for end of file.
905 */
906
907 /*
908 * Unlike in truncate_pagecache, unmap_mapping_range is called only
909 * once (before truncating pagecache), and without "even_cows" flag:
910 * hole-punching should not remove private COWed pages from the hole.
911 */
912 if ((u64)unmap_end > (u64)unmap_start)
913 unmap_mapping_range(mapping, unmap_start,
914 1 + unmap_end - unmap_start, 0);
915 truncate_inode_pages_range(mapping, lstart, lend);
916}
917EXPORT_SYMBOL(truncate_pagecache_range);