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/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);
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/cleancache.h>
24#include <linux/rmap.h>
25#include "internal.h"
26
27static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
28 void *entry)
29{
30 struct radix_tree_node *node;
31 void **slot;
32
33 spin_lock_irq(&mapping->tree_lock);
34 /*
35 * Regular page slots are stabilized by the page lock even
36 * without the tree itself locked. These unlocked entries
37 * need verification under the tree lock.
38 */
39 if (!__radix_tree_lookup(&mapping->page_tree, index, &node, &slot))
40 goto unlock;
41 if (*slot != entry)
42 goto unlock;
43 __radix_tree_replace(&mapping->page_tree, node, slot, NULL,
44 workingset_update_node, mapping);
45 mapping->nrexceptional--;
46unlock:
47 spin_unlock_irq(&mapping->tree_lock);
48}
49
50/*
51 * Unconditionally remove exceptional entry. Usually called from truncate path.
52 */
53static void truncate_exceptional_entry(struct address_space *mapping,
54 pgoff_t index, void *entry)
55{
56 /* Handled by shmem itself */
57 if (shmem_mapping(mapping))
58 return;
59
60 if (dax_mapping(mapping)) {
61 dax_delete_mapping_entry(mapping, index);
62 return;
63 }
64 clear_shadow_entry(mapping, index, entry);
65}
66
67/*
68 * Invalidate exceptional entry if easily possible. This handles exceptional
69 * entries for invalidate_inode_pages() so for DAX it evicts only unlocked and
70 * clean entries.
71 */
72static int invalidate_exceptional_entry(struct address_space *mapping,
73 pgoff_t index, void *entry)
74{
75 /* Handled by shmem itself */
76 if (shmem_mapping(mapping))
77 return 1;
78 if (dax_mapping(mapping))
79 return dax_invalidate_mapping_entry(mapping, index);
80 clear_shadow_entry(mapping, index, entry);
81 return 1;
82}
83
84/*
85 * Invalidate exceptional entry if clean. This handles exceptional entries for
86 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
87 */
88static int invalidate_exceptional_entry2(struct address_space *mapping,
89 pgoff_t index, void *entry)
90{
91 /* Handled by shmem itself */
92 if (shmem_mapping(mapping))
93 return 1;
94 if (dax_mapping(mapping))
95 return dax_invalidate_mapping_entry_sync(mapping, index);
96 clear_shadow_entry(mapping, index, entry);
97 return 1;
98}
99
100/**
101 * do_invalidatepage - invalidate part or all of a page
102 * @page: the page which is affected
103 * @offset: start of the range to invalidate
104 * @length: length of the range to invalidate
105 *
106 * do_invalidatepage() is called when all or part of the page has become
107 * invalidated by a truncate operation.
108 *
109 * do_invalidatepage() does not have to release all buffers, but it must
110 * ensure that no dirty buffer is left outside @offset and that no I/O
111 * is underway against any of the blocks which are outside the truncation
112 * point. Because the caller is about to free (and possibly reuse) those
113 * blocks on-disk.
114 */
115void do_invalidatepage(struct page *page, unsigned int offset,
116 unsigned int length)
117{
118 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
119
120 invalidatepage = page->mapping->a_ops->invalidatepage;
121#ifdef CONFIG_BLOCK
122 if (!invalidatepage)
123 invalidatepage = block_invalidatepage;
124#endif
125 if (invalidatepage)
126 (*invalidatepage)(page, offset, length);
127}
128
129/*
130 * If truncate cannot remove the fs-private metadata from the page, the page
131 * becomes orphaned. It will be left on the LRU and may even be mapped into
132 * user pagetables if we're racing with filemap_fault().
133 *
134 * We need to bale out if page->mapping is no longer equal to the original
135 * mapping. This happens a) when the VM reclaimed the page while we waited on
136 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
137 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
138 */
139static int
140truncate_complete_page(struct address_space *mapping, struct page *page)
141{
142 if (page->mapping != mapping)
143 return -EIO;
144
145 if (page_has_private(page))
146 do_invalidatepage(page, 0, PAGE_SIZE);
147
148 /*
149 * Some filesystems seem to re-dirty the page even after
150 * the VM has canceled the dirty bit (eg ext3 journaling).
151 * Hence dirty accounting check is placed after invalidation.
152 */
153 cancel_dirty_page(page);
154 ClearPageMappedToDisk(page);
155 delete_from_page_cache(page);
156 return 0;
157}
158
159/*
160 * This is for invalidate_mapping_pages(). That function can be called at
161 * any time, and is not supposed to throw away dirty pages. But pages can
162 * be marked dirty at any time too, so use remove_mapping which safely
163 * discards clean, unused pages.
164 *
165 * Returns non-zero if the page was successfully invalidated.
166 */
167static int
168invalidate_complete_page(struct address_space *mapping, struct page *page)
169{
170 int ret;
171
172 if (page->mapping != mapping)
173 return 0;
174
175 if (page_has_private(page) && !try_to_release_page(page, 0))
176 return 0;
177
178 ret = remove_mapping(mapping, page);
179
180 return ret;
181}
182
183int truncate_inode_page(struct address_space *mapping, struct page *page)
184{
185 loff_t holelen;
186 VM_BUG_ON_PAGE(PageTail(page), page);
187
188 holelen = PageTransHuge(page) ? HPAGE_PMD_SIZE : PAGE_SIZE;
189 if (page_mapped(page)) {
190 unmap_mapping_range(mapping,
191 (loff_t)page->index << PAGE_SHIFT,
192 holelen, 0);
193 }
194 return truncate_complete_page(mapping, page);
195}
196
197/*
198 * Used to get rid of pages on hardware memory corruption.
199 */
200int generic_error_remove_page(struct address_space *mapping, struct page *page)
201{
202 if (!mapping)
203 return -EINVAL;
204 /*
205 * Only punch for normal data pages for now.
206 * Handling other types like directories would need more auditing.
207 */
208 if (!S_ISREG(mapping->host->i_mode))
209 return -EIO;
210 return truncate_inode_page(mapping, page);
211}
212EXPORT_SYMBOL(generic_error_remove_page);
213
214/*
215 * Safely invalidate one page from its pagecache mapping.
216 * It only drops clean, unused pages. The page must be locked.
217 *
218 * Returns 1 if the page is successfully invalidated, otherwise 0.
219 */
220int invalidate_inode_page(struct page *page)
221{
222 struct address_space *mapping = page_mapping(page);
223 if (!mapping)
224 return 0;
225 if (PageDirty(page) || PageWriteback(page))
226 return 0;
227 if (page_mapped(page))
228 return 0;
229 return invalidate_complete_page(mapping, page);
230}
231
232/**
233 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
234 * @mapping: mapping to truncate
235 * @lstart: offset from which to truncate
236 * @lend: offset to which to truncate (inclusive)
237 *
238 * Truncate the page cache, removing the pages that are between
239 * specified offsets (and zeroing out partial pages
240 * if lstart or lend + 1 is not page aligned).
241 *
242 * Truncate takes two passes - the first pass is nonblocking. It will not
243 * block on page locks and it will not block on writeback. The second pass
244 * will wait. This is to prevent as much IO as possible in the affected region.
245 * The first pass will remove most pages, so the search cost of the second pass
246 * is low.
247 *
248 * We pass down the cache-hot hint to the page freeing code. Even if the
249 * mapping is large, it is probably the case that the final pages are the most
250 * recently touched, and freeing happens in ascending file offset order.
251 *
252 * Note that since ->invalidatepage() accepts range to invalidate
253 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
254 * page aligned properly.
255 */
256void truncate_inode_pages_range(struct address_space *mapping,
257 loff_t lstart, loff_t lend)
258{
259 pgoff_t start; /* inclusive */
260 pgoff_t end; /* exclusive */
261 unsigned int partial_start; /* inclusive */
262 unsigned int partial_end; /* exclusive */
263 struct pagevec pvec;
264 pgoff_t indices[PAGEVEC_SIZE];
265 pgoff_t index;
266 int i;
267
268 cleancache_invalidate_inode(mapping);
269 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
270 return;
271
272 /* Offsets within partial pages */
273 partial_start = lstart & (PAGE_SIZE - 1);
274 partial_end = (lend + 1) & (PAGE_SIZE - 1);
275
276 /*
277 * 'start' and 'end' always covers the range of pages to be fully
278 * truncated. Partial pages are covered with 'partial_start' at the
279 * start of the range and 'partial_end' at the end of the range.
280 * Note that 'end' is exclusive while 'lend' is inclusive.
281 */
282 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
283 if (lend == -1)
284 /*
285 * lend == -1 indicates end-of-file so we have to set 'end'
286 * to the highest possible pgoff_t and since the type is
287 * unsigned we're using -1.
288 */
289 end = -1;
290 else
291 end = (lend + 1) >> PAGE_SHIFT;
292
293 pagevec_init(&pvec, 0);
294 index = start;
295 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
296 min(end - index, (pgoff_t)PAGEVEC_SIZE),
297 indices)) {
298 for (i = 0; i < pagevec_count(&pvec); i++) {
299 struct page *page = pvec.pages[i];
300
301 /* We rely upon deletion not changing page->index */
302 index = indices[i];
303 if (index >= end)
304 break;
305
306 if (radix_tree_exceptional_entry(page)) {
307 truncate_exceptional_entry(mapping, index,
308 page);
309 continue;
310 }
311
312 if (!trylock_page(page))
313 continue;
314 WARN_ON(page_to_index(page) != index);
315 if (PageWriteback(page)) {
316 unlock_page(page);
317 continue;
318 }
319 truncate_inode_page(mapping, page);
320 unlock_page(page);
321 }
322 pagevec_remove_exceptionals(&pvec);
323 pagevec_release(&pvec);
324 cond_resched();
325 index++;
326 }
327
328 if (partial_start) {
329 struct page *page = find_lock_page(mapping, start - 1);
330 if (page) {
331 unsigned int top = PAGE_SIZE;
332 if (start > end) {
333 /* Truncation within a single page */
334 top = partial_end;
335 partial_end = 0;
336 }
337 wait_on_page_writeback(page);
338 zero_user_segment(page, partial_start, top);
339 cleancache_invalidate_page(mapping, page);
340 if (page_has_private(page))
341 do_invalidatepage(page, partial_start,
342 top - partial_start);
343 unlock_page(page);
344 put_page(page);
345 }
346 }
347 if (partial_end) {
348 struct page *page = find_lock_page(mapping, end);
349 if (page) {
350 wait_on_page_writeback(page);
351 zero_user_segment(page, 0, partial_end);
352 cleancache_invalidate_page(mapping, page);
353 if (page_has_private(page))
354 do_invalidatepage(page, 0,
355 partial_end);
356 unlock_page(page);
357 put_page(page);
358 }
359 }
360 /*
361 * If the truncation happened within a single page no pages
362 * will be released, just zeroed, so we can bail out now.
363 */
364 if (start >= end)
365 return;
366
367 index = start;
368 for ( ; ; ) {
369 cond_resched();
370 if (!pagevec_lookup_entries(&pvec, mapping, index,
371 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
372 /* If all gone from start onwards, we're done */
373 if (index == start)
374 break;
375 /* Otherwise restart to make sure all gone */
376 index = start;
377 continue;
378 }
379 if (index == start && indices[0] >= end) {
380 /* All gone out of hole to be punched, we're done */
381 pagevec_remove_exceptionals(&pvec);
382 pagevec_release(&pvec);
383 break;
384 }
385 for (i = 0; i < pagevec_count(&pvec); i++) {
386 struct page *page = pvec.pages[i];
387
388 /* We rely upon deletion not changing page->index */
389 index = indices[i];
390 if (index >= end) {
391 /* Restart punch to make sure all gone */
392 index = start - 1;
393 break;
394 }
395
396 if (radix_tree_exceptional_entry(page)) {
397 truncate_exceptional_entry(mapping, index,
398 page);
399 continue;
400 }
401
402 lock_page(page);
403 WARN_ON(page_to_index(page) != index);
404 wait_on_page_writeback(page);
405 truncate_inode_page(mapping, page);
406 unlock_page(page);
407 }
408 pagevec_remove_exceptionals(&pvec);
409 pagevec_release(&pvec);
410 index++;
411 }
412 cleancache_invalidate_inode(mapping);
413}
414EXPORT_SYMBOL(truncate_inode_pages_range);
415
416/**
417 * truncate_inode_pages - truncate *all* the pages from an offset
418 * @mapping: mapping to truncate
419 * @lstart: offset from which to truncate
420 *
421 * Called under (and serialised by) inode->i_mutex.
422 *
423 * Note: When this function returns, there can be a page in the process of
424 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
425 * mapping->nrpages can be non-zero when this function returns even after
426 * truncation of the whole mapping.
427 */
428void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
429{
430 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
431}
432EXPORT_SYMBOL(truncate_inode_pages);
433
434/**
435 * truncate_inode_pages_final - truncate *all* pages before inode dies
436 * @mapping: mapping to truncate
437 *
438 * Called under (and serialized by) inode->i_mutex.
439 *
440 * Filesystems have to use this in the .evict_inode path to inform the
441 * VM that this is the final truncate and the inode is going away.
442 */
443void truncate_inode_pages_final(struct address_space *mapping)
444{
445 unsigned long nrexceptional;
446 unsigned long nrpages;
447
448 /*
449 * Page reclaim can not participate in regular inode lifetime
450 * management (can't call iput()) and thus can race with the
451 * inode teardown. Tell it when the address space is exiting,
452 * so that it does not install eviction information after the
453 * final truncate has begun.
454 */
455 mapping_set_exiting(mapping);
456
457 /*
458 * When reclaim installs eviction entries, it increases
459 * nrexceptional first, then decreases nrpages. Make sure we see
460 * this in the right order or we might miss an entry.
461 */
462 nrpages = mapping->nrpages;
463 smp_rmb();
464 nrexceptional = mapping->nrexceptional;
465
466 if (nrpages || nrexceptional) {
467 /*
468 * As truncation uses a lockless tree lookup, cycle
469 * the tree lock to make sure any ongoing tree
470 * modification that does not see AS_EXITING is
471 * completed before starting the final truncate.
472 */
473 spin_lock_irq(&mapping->tree_lock);
474 spin_unlock_irq(&mapping->tree_lock);
475
476 truncate_inode_pages(mapping, 0);
477 }
478}
479EXPORT_SYMBOL(truncate_inode_pages_final);
480
481/**
482 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
483 * @mapping: the address_space which holds the pages to invalidate
484 * @start: the offset 'from' which to invalidate
485 * @end: the offset 'to' which to invalidate (inclusive)
486 *
487 * This function only removes the unlocked pages, if you want to
488 * remove all the pages of one inode, you must call truncate_inode_pages.
489 *
490 * invalidate_mapping_pages() will not block on IO activity. It will not
491 * invalidate pages which are dirty, locked, under writeback or mapped into
492 * pagetables.
493 */
494unsigned long invalidate_mapping_pages(struct address_space *mapping,
495 pgoff_t start, pgoff_t end)
496{
497 pgoff_t indices[PAGEVEC_SIZE];
498 struct pagevec pvec;
499 pgoff_t index = start;
500 unsigned long ret;
501 unsigned long count = 0;
502 int i;
503
504 pagevec_init(&pvec, 0);
505 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
506 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
507 indices)) {
508 for (i = 0; i < pagevec_count(&pvec); i++) {
509 struct page *page = pvec.pages[i];
510
511 /* We rely upon deletion not changing page->index */
512 index = indices[i];
513 if (index > end)
514 break;
515
516 if (radix_tree_exceptional_entry(page)) {
517 invalidate_exceptional_entry(mapping, index,
518 page);
519 continue;
520 }
521
522 if (!trylock_page(page))
523 continue;
524
525 WARN_ON(page_to_index(page) != index);
526
527 /* Middle of THP: skip */
528 if (PageTransTail(page)) {
529 unlock_page(page);
530 continue;
531 } else if (PageTransHuge(page)) {
532 index += HPAGE_PMD_NR - 1;
533 i += HPAGE_PMD_NR - 1;
534 /* 'end' is in the middle of THP */
535 if (index == round_down(end, HPAGE_PMD_NR))
536 continue;
537 }
538
539 ret = invalidate_inode_page(page);
540 unlock_page(page);
541 /*
542 * Invalidation is a hint that the page is no longer
543 * of interest and try to speed up its reclaim.
544 */
545 if (!ret)
546 deactivate_file_page(page);
547 count += ret;
548 }
549 pagevec_remove_exceptionals(&pvec);
550 pagevec_release(&pvec);
551 cond_resched();
552 index++;
553 }
554 return count;
555}
556EXPORT_SYMBOL(invalidate_mapping_pages);
557
558/*
559 * This is like invalidate_complete_page(), except it ignores the page's
560 * refcount. We do this because invalidate_inode_pages2() needs stronger
561 * invalidation guarantees, and cannot afford to leave pages behind because
562 * shrink_page_list() has a temp ref on them, or because they're transiently
563 * sitting in the lru_cache_add() pagevecs.
564 */
565static int
566invalidate_complete_page2(struct address_space *mapping, struct page *page)
567{
568 unsigned long flags;
569
570 if (page->mapping != mapping)
571 return 0;
572
573 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
574 return 0;
575
576 spin_lock_irqsave(&mapping->tree_lock, flags);
577 if (PageDirty(page))
578 goto failed;
579
580 BUG_ON(page_has_private(page));
581 __delete_from_page_cache(page, NULL);
582 spin_unlock_irqrestore(&mapping->tree_lock, flags);
583
584 if (mapping->a_ops->freepage)
585 mapping->a_ops->freepage(page);
586
587 put_page(page); /* pagecache ref */
588 return 1;
589failed:
590 spin_unlock_irqrestore(&mapping->tree_lock, flags);
591 return 0;
592}
593
594static int do_launder_page(struct address_space *mapping, struct page *page)
595{
596 if (!PageDirty(page))
597 return 0;
598 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
599 return 0;
600 return mapping->a_ops->launder_page(page);
601}
602
603/**
604 * invalidate_inode_pages2_range - remove range of pages from an address_space
605 * @mapping: the address_space
606 * @start: the page offset 'from' which to invalidate
607 * @end: the page offset 'to' which to invalidate (inclusive)
608 *
609 * Any pages which are found to be mapped into pagetables are unmapped prior to
610 * invalidation.
611 *
612 * Returns -EBUSY if any pages could not be invalidated.
613 */
614int invalidate_inode_pages2_range(struct address_space *mapping,
615 pgoff_t start, pgoff_t end)
616{
617 pgoff_t indices[PAGEVEC_SIZE];
618 struct pagevec pvec;
619 pgoff_t index;
620 int i;
621 int ret = 0;
622 int ret2 = 0;
623 int did_range_unmap = 0;
624
625 cleancache_invalidate_inode(mapping);
626 pagevec_init(&pvec, 0);
627 index = start;
628 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
629 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
630 indices)) {
631 for (i = 0; i < pagevec_count(&pvec); i++) {
632 struct page *page = pvec.pages[i];
633
634 /* We rely upon deletion not changing page->index */
635 index = indices[i];
636 if (index > end)
637 break;
638
639 if (radix_tree_exceptional_entry(page)) {
640 if (!invalidate_exceptional_entry2(mapping,
641 index, page))
642 ret = -EBUSY;
643 continue;
644 }
645
646 lock_page(page);
647 WARN_ON(page_to_index(page) != index);
648 if (page->mapping != mapping) {
649 unlock_page(page);
650 continue;
651 }
652 wait_on_page_writeback(page);
653 if (page_mapped(page)) {
654 if (!did_range_unmap) {
655 /*
656 * Zap the rest of the file in one hit.
657 */
658 unmap_mapping_range(mapping,
659 (loff_t)index << PAGE_SHIFT,
660 (loff_t)(1 + end - index)
661 << PAGE_SHIFT,
662 0);
663 did_range_unmap = 1;
664 } else {
665 /*
666 * Just zap this page
667 */
668 unmap_mapping_range(mapping,
669 (loff_t)index << PAGE_SHIFT,
670 PAGE_SIZE, 0);
671 }
672 }
673 BUG_ON(page_mapped(page));
674 ret2 = do_launder_page(mapping, page);
675 if (ret2 == 0) {
676 if (!invalidate_complete_page2(mapping, page))
677 ret2 = -EBUSY;
678 }
679 if (ret2 < 0)
680 ret = ret2;
681 unlock_page(page);
682 }
683 pagevec_remove_exceptionals(&pvec);
684 pagevec_release(&pvec);
685 cond_resched();
686 index++;
687 }
688 cleancache_invalidate_inode(mapping);
689 return ret;
690}
691EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
692
693/**
694 * invalidate_inode_pages2 - remove all pages from an address_space
695 * @mapping: the address_space
696 *
697 * Any pages which are found to be mapped into pagetables are unmapped prior to
698 * invalidation.
699 *
700 * Returns -EBUSY if any pages could not be invalidated.
701 */
702int invalidate_inode_pages2(struct address_space *mapping)
703{
704 return invalidate_inode_pages2_range(mapping, 0, -1);
705}
706EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
707
708/**
709 * truncate_pagecache - unmap and remove pagecache that has been truncated
710 * @inode: inode
711 * @newsize: new file size
712 *
713 * inode's new i_size must already be written before truncate_pagecache
714 * is called.
715 *
716 * This function should typically be called before the filesystem
717 * releases resources associated with the freed range (eg. deallocates
718 * blocks). This way, pagecache will always stay logically coherent
719 * with on-disk format, and the filesystem would not have to deal with
720 * situations such as writepage being called for a page that has already
721 * had its underlying blocks deallocated.
722 */
723void truncate_pagecache(struct inode *inode, loff_t newsize)
724{
725 struct address_space *mapping = inode->i_mapping;
726 loff_t holebegin = round_up(newsize, PAGE_SIZE);
727
728 /*
729 * unmap_mapping_range is called twice, first simply for
730 * efficiency so that truncate_inode_pages does fewer
731 * single-page unmaps. However after this first call, and
732 * before truncate_inode_pages finishes, it is possible for
733 * private pages to be COWed, which remain after
734 * truncate_inode_pages finishes, hence the second
735 * unmap_mapping_range call must be made for correctness.
736 */
737 unmap_mapping_range(mapping, holebegin, 0, 1);
738 truncate_inode_pages(mapping, newsize);
739 unmap_mapping_range(mapping, holebegin, 0, 1);
740}
741EXPORT_SYMBOL(truncate_pagecache);
742
743/**
744 * truncate_setsize - update inode and pagecache for a new file size
745 * @inode: inode
746 * @newsize: new file size
747 *
748 * truncate_setsize updates i_size and performs pagecache truncation (if
749 * necessary) to @newsize. It will be typically be called from the filesystem's
750 * setattr function when ATTR_SIZE is passed in.
751 *
752 * Must be called with a lock serializing truncates and writes (generally
753 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
754 * specific block truncation has been performed.
755 */
756void truncate_setsize(struct inode *inode, loff_t newsize)
757{
758 loff_t oldsize = inode->i_size;
759
760 i_size_write(inode, newsize);
761 if (newsize > oldsize)
762 pagecache_isize_extended(inode, oldsize, newsize);
763 truncate_pagecache(inode, newsize);
764}
765EXPORT_SYMBOL(truncate_setsize);
766
767/**
768 * pagecache_isize_extended - update pagecache after extension of i_size
769 * @inode: inode for which i_size was extended
770 * @from: original inode size
771 * @to: new inode size
772 *
773 * Handle extension of inode size either caused by extending truncate or by
774 * write starting after current i_size. We mark the page straddling current
775 * i_size RO so that page_mkwrite() is called on the nearest write access to
776 * the page. This way filesystem can be sure that page_mkwrite() is called on
777 * the page before user writes to the page via mmap after the i_size has been
778 * changed.
779 *
780 * The function must be called after i_size is updated so that page fault
781 * coming after we unlock the page will already see the new i_size.
782 * The function must be called while we still hold i_mutex - this not only
783 * makes sure i_size is stable but also that userspace cannot observe new
784 * i_size value before we are prepared to store mmap writes at new inode size.
785 */
786void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
787{
788 int bsize = 1 << inode->i_blkbits;
789 loff_t rounded_from;
790 struct page *page;
791 pgoff_t index;
792
793 WARN_ON(to > inode->i_size);
794
795 if (from >= to || bsize == PAGE_SIZE)
796 return;
797 /* Page straddling @from will not have any hole block created? */
798 rounded_from = round_up(from, bsize);
799 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
800 return;
801
802 index = from >> PAGE_SHIFT;
803 page = find_lock_page(inode->i_mapping, index);
804 /* Page not cached? Nothing to do */
805 if (!page)
806 return;
807 /*
808 * See clear_page_dirty_for_io() for details why set_page_dirty()
809 * is needed.
810 */
811 if (page_mkclean(page))
812 set_page_dirty(page);
813 unlock_page(page);
814 put_page(page);
815}
816EXPORT_SYMBOL(pagecache_isize_extended);
817
818/**
819 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
820 * @inode: inode
821 * @lstart: offset of beginning of hole
822 * @lend: offset of last byte of hole
823 *
824 * This function should typically be called before the filesystem
825 * releases resources associated with the freed range (eg. deallocates
826 * blocks). This way, pagecache will always stay logically coherent
827 * with on-disk format, and the filesystem would not have to deal with
828 * situations such as writepage being called for a page that has already
829 * had its underlying blocks deallocated.
830 */
831void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
832{
833 struct address_space *mapping = inode->i_mapping;
834 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
835 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
836 /*
837 * This rounding is currently just for example: unmap_mapping_range
838 * expands its hole outwards, whereas we want it to contract the hole
839 * inwards. However, existing callers of truncate_pagecache_range are
840 * doing their own page rounding first. Note that unmap_mapping_range
841 * allows holelen 0 for all, and we allow lend -1 for end of file.
842 */
843
844 /*
845 * Unlike in truncate_pagecache, unmap_mapping_range is called only
846 * once (before truncating pagecache), and without "even_cows" flag:
847 * hole-punching should not remove private COWed pages from the hole.
848 */
849 if ((u64)unmap_end > (u64)unmap_start)
850 unmap_mapping_range(mapping, unmap_start,
851 1 + unmap_end - unmap_start, 0);
852 truncate_inode_pages_range(mapping, lstart, lend);
853}
854EXPORT_SYMBOL(truncate_pagecache_range);