Loading...
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_exceptional_entry(struct address_space *mapping,
28 pgoff_t index, void *entry)
29{
30 struct radix_tree_node *node;
31 void **slot;
32
33 /* Handled by shmem itself */
34 if (shmem_mapping(mapping))
35 return;
36
37 spin_lock_irq(&mapping->tree_lock);
38
39 if (dax_mapping(mapping)) {
40 if (radix_tree_delete_item(&mapping->page_tree, index, entry))
41 mapping->nrexceptional--;
42 } else {
43 /*
44 * Regular page slots are stabilized by the page lock even
45 * without the tree itself locked. These unlocked entries
46 * need verification under the tree lock.
47 */
48 if (!__radix_tree_lookup(&mapping->page_tree, index, &node,
49 &slot))
50 goto unlock;
51 if (*slot != entry)
52 goto unlock;
53 radix_tree_replace_slot(slot, NULL);
54 mapping->nrexceptional--;
55 if (!node)
56 goto unlock;
57 workingset_node_shadows_dec(node);
58 /*
59 * Don't track node without shadow entries.
60 *
61 * Avoid acquiring the list_lru lock if already untracked.
62 * The list_empty() test is safe as node->private_list is
63 * protected by mapping->tree_lock.
64 */
65 if (!workingset_node_shadows(node) &&
66 !list_empty(&node->private_list))
67 list_lru_del(&workingset_shadow_nodes,
68 &node->private_list);
69 __radix_tree_delete_node(&mapping->page_tree, node);
70 }
71unlock:
72 spin_unlock_irq(&mapping->tree_lock);
73}
74
75/**
76 * do_invalidatepage - invalidate part or all of a page
77 * @page: the page which is affected
78 * @offset: start of the range to invalidate
79 * @length: length of the range to invalidate
80 *
81 * do_invalidatepage() is called when all or part of the page has become
82 * invalidated by a truncate operation.
83 *
84 * do_invalidatepage() does not have to release all buffers, but it must
85 * ensure that no dirty buffer is left outside @offset and that no I/O
86 * is underway against any of the blocks which are outside the truncation
87 * point. Because the caller is about to free (and possibly reuse) those
88 * blocks on-disk.
89 */
90void do_invalidatepage(struct page *page, unsigned int offset,
91 unsigned int length)
92{
93 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
94
95 invalidatepage = page->mapping->a_ops->invalidatepage;
96#ifdef CONFIG_BLOCK
97 if (!invalidatepage)
98 invalidatepage = block_invalidatepage;
99#endif
100 if (invalidatepage)
101 (*invalidatepage)(page, offset, length);
102}
103
104/*
105 * If truncate cannot remove the fs-private metadata from the page, the page
106 * becomes orphaned. It will be left on the LRU and may even be mapped into
107 * user pagetables if we're racing with filemap_fault().
108 *
109 * We need to bale out if page->mapping is no longer equal to the original
110 * mapping. This happens a) when the VM reclaimed the page while we waited on
111 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
112 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
113 */
114static int
115truncate_complete_page(struct address_space *mapping, struct page *page)
116{
117 if (page->mapping != mapping)
118 return -EIO;
119
120 if (page_has_private(page))
121 do_invalidatepage(page, 0, PAGE_SIZE);
122
123 /*
124 * Some filesystems seem to re-dirty the page even after
125 * the VM has canceled the dirty bit (eg ext3 journaling).
126 * Hence dirty accounting check is placed after invalidation.
127 */
128 cancel_dirty_page(page);
129 ClearPageMappedToDisk(page);
130 delete_from_page_cache(page);
131 return 0;
132}
133
134/*
135 * This is for invalidate_mapping_pages(). That function can be called at
136 * any time, and is not supposed to throw away dirty pages. But pages can
137 * be marked dirty at any time too, so use remove_mapping which safely
138 * discards clean, unused pages.
139 *
140 * Returns non-zero if the page was successfully invalidated.
141 */
142static int
143invalidate_complete_page(struct address_space *mapping, struct page *page)
144{
145 int ret;
146
147 if (page->mapping != mapping)
148 return 0;
149
150 if (page_has_private(page) && !try_to_release_page(page, 0))
151 return 0;
152
153 ret = remove_mapping(mapping, page);
154
155 return ret;
156}
157
158int truncate_inode_page(struct address_space *mapping, struct page *page)
159{
160 if (page_mapped(page)) {
161 unmap_mapping_range(mapping,
162 (loff_t)page->index << PAGE_SHIFT,
163 PAGE_SIZE, 0);
164 }
165 return truncate_complete_page(mapping, page);
166}
167
168/*
169 * Used to get rid of pages on hardware memory corruption.
170 */
171int generic_error_remove_page(struct address_space *mapping, struct page *page)
172{
173 if (!mapping)
174 return -EINVAL;
175 /*
176 * Only punch for normal data pages for now.
177 * Handling other types like directories would need more auditing.
178 */
179 if (!S_ISREG(mapping->host->i_mode))
180 return -EIO;
181 return truncate_inode_page(mapping, page);
182}
183EXPORT_SYMBOL(generic_error_remove_page);
184
185/*
186 * Safely invalidate one page from its pagecache mapping.
187 * It only drops clean, unused pages. The page must be locked.
188 *
189 * Returns 1 if the page is successfully invalidated, otherwise 0.
190 */
191int invalidate_inode_page(struct page *page)
192{
193 struct address_space *mapping = page_mapping(page);
194 if (!mapping)
195 return 0;
196 if (PageDirty(page) || PageWriteback(page))
197 return 0;
198 if (page_mapped(page))
199 return 0;
200 return invalidate_complete_page(mapping, page);
201}
202
203/**
204 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
205 * @mapping: mapping to truncate
206 * @lstart: offset from which to truncate
207 * @lend: offset to which to truncate (inclusive)
208 *
209 * Truncate the page cache, removing the pages that are between
210 * specified offsets (and zeroing out partial pages
211 * if lstart or lend + 1 is not page aligned).
212 *
213 * Truncate takes two passes - the first pass is nonblocking. It will not
214 * block on page locks and it will not block on writeback. The second pass
215 * will wait. This is to prevent as much IO as possible in the affected region.
216 * The first pass will remove most pages, so the search cost of the second pass
217 * is low.
218 *
219 * We pass down the cache-hot hint to the page freeing code. Even if the
220 * mapping is large, it is probably the case that the final pages are the most
221 * recently touched, and freeing happens in ascending file offset order.
222 *
223 * Note that since ->invalidatepage() accepts range to invalidate
224 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
225 * page aligned properly.
226 */
227void truncate_inode_pages_range(struct address_space *mapping,
228 loff_t lstart, loff_t lend)
229{
230 pgoff_t start; /* inclusive */
231 pgoff_t end; /* exclusive */
232 unsigned int partial_start; /* inclusive */
233 unsigned int partial_end; /* exclusive */
234 struct pagevec pvec;
235 pgoff_t indices[PAGEVEC_SIZE];
236 pgoff_t index;
237 int i;
238
239 cleancache_invalidate_inode(mapping);
240 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
241 return;
242
243 /* Offsets within partial pages */
244 partial_start = lstart & (PAGE_SIZE - 1);
245 partial_end = (lend + 1) & (PAGE_SIZE - 1);
246
247 /*
248 * 'start' and 'end' always covers the range of pages to be fully
249 * truncated. Partial pages are covered with 'partial_start' at the
250 * start of the range and 'partial_end' at the end of the range.
251 * Note that 'end' is exclusive while 'lend' is inclusive.
252 */
253 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
254 if (lend == -1)
255 /*
256 * lend == -1 indicates end-of-file so we have to set 'end'
257 * to the highest possible pgoff_t and since the type is
258 * unsigned we're using -1.
259 */
260 end = -1;
261 else
262 end = (lend + 1) >> PAGE_SHIFT;
263
264 pagevec_init(&pvec, 0);
265 index = start;
266 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
267 min(end - index, (pgoff_t)PAGEVEC_SIZE),
268 indices)) {
269 for (i = 0; i < pagevec_count(&pvec); i++) {
270 struct page *page = pvec.pages[i];
271
272 /* We rely upon deletion not changing page->index */
273 index = indices[i];
274 if (index >= end)
275 break;
276
277 if (radix_tree_exceptional_entry(page)) {
278 clear_exceptional_entry(mapping, index, page);
279 continue;
280 }
281
282 if (!trylock_page(page))
283 continue;
284 WARN_ON(page->index != index);
285 if (PageWriteback(page)) {
286 unlock_page(page);
287 continue;
288 }
289 truncate_inode_page(mapping, page);
290 unlock_page(page);
291 }
292 pagevec_remove_exceptionals(&pvec);
293 pagevec_release(&pvec);
294 cond_resched();
295 index++;
296 }
297
298 if (partial_start) {
299 struct page *page = find_lock_page(mapping, start - 1);
300 if (page) {
301 unsigned int top = PAGE_SIZE;
302 if (start > end) {
303 /* Truncation within a single page */
304 top = partial_end;
305 partial_end = 0;
306 }
307 wait_on_page_writeback(page);
308 zero_user_segment(page, partial_start, top);
309 cleancache_invalidate_page(mapping, page);
310 if (page_has_private(page))
311 do_invalidatepage(page, partial_start,
312 top - partial_start);
313 unlock_page(page);
314 put_page(page);
315 }
316 }
317 if (partial_end) {
318 struct page *page = find_lock_page(mapping, end);
319 if (page) {
320 wait_on_page_writeback(page);
321 zero_user_segment(page, 0, partial_end);
322 cleancache_invalidate_page(mapping, page);
323 if (page_has_private(page))
324 do_invalidatepage(page, 0,
325 partial_end);
326 unlock_page(page);
327 put_page(page);
328 }
329 }
330 /*
331 * If the truncation happened within a single page no pages
332 * will be released, just zeroed, so we can bail out now.
333 */
334 if (start >= end)
335 return;
336
337 index = start;
338 for ( ; ; ) {
339 cond_resched();
340 if (!pagevec_lookup_entries(&pvec, mapping, index,
341 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
342 /* If all gone from start onwards, we're done */
343 if (index == start)
344 break;
345 /* Otherwise restart to make sure all gone */
346 index = start;
347 continue;
348 }
349 if (index == start && indices[0] >= end) {
350 /* All gone out of hole to be punched, we're done */
351 pagevec_remove_exceptionals(&pvec);
352 pagevec_release(&pvec);
353 break;
354 }
355 for (i = 0; i < pagevec_count(&pvec); i++) {
356 struct page *page = pvec.pages[i];
357
358 /* We rely upon deletion not changing page->index */
359 index = indices[i];
360 if (index >= end) {
361 /* Restart punch to make sure all gone */
362 index = start - 1;
363 break;
364 }
365
366 if (radix_tree_exceptional_entry(page)) {
367 clear_exceptional_entry(mapping, index, page);
368 continue;
369 }
370
371 lock_page(page);
372 WARN_ON(page->index != index);
373 wait_on_page_writeback(page);
374 truncate_inode_page(mapping, page);
375 unlock_page(page);
376 }
377 pagevec_remove_exceptionals(&pvec);
378 pagevec_release(&pvec);
379 index++;
380 }
381 cleancache_invalidate_inode(mapping);
382}
383EXPORT_SYMBOL(truncate_inode_pages_range);
384
385/**
386 * truncate_inode_pages - truncate *all* the pages from an offset
387 * @mapping: mapping to truncate
388 * @lstart: offset from which to truncate
389 *
390 * Called under (and serialised by) inode->i_mutex.
391 *
392 * Note: When this function returns, there can be a page in the process of
393 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
394 * mapping->nrpages can be non-zero when this function returns even after
395 * truncation of the whole mapping.
396 */
397void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
398{
399 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
400}
401EXPORT_SYMBOL(truncate_inode_pages);
402
403/**
404 * truncate_inode_pages_final - truncate *all* pages before inode dies
405 * @mapping: mapping to truncate
406 *
407 * Called under (and serialized by) inode->i_mutex.
408 *
409 * Filesystems have to use this in the .evict_inode path to inform the
410 * VM that this is the final truncate and the inode is going away.
411 */
412void truncate_inode_pages_final(struct address_space *mapping)
413{
414 unsigned long nrexceptional;
415 unsigned long nrpages;
416
417 /*
418 * Page reclaim can not participate in regular inode lifetime
419 * management (can't call iput()) and thus can race with the
420 * inode teardown. Tell it when the address space is exiting,
421 * so that it does not install eviction information after the
422 * final truncate has begun.
423 */
424 mapping_set_exiting(mapping);
425
426 /*
427 * When reclaim installs eviction entries, it increases
428 * nrexceptional first, then decreases nrpages. Make sure we see
429 * this in the right order or we might miss an entry.
430 */
431 nrpages = mapping->nrpages;
432 smp_rmb();
433 nrexceptional = mapping->nrexceptional;
434
435 if (nrpages || nrexceptional) {
436 /*
437 * As truncation uses a lockless tree lookup, cycle
438 * the tree lock to make sure any ongoing tree
439 * modification that does not see AS_EXITING is
440 * completed before starting the final truncate.
441 */
442 spin_lock_irq(&mapping->tree_lock);
443 spin_unlock_irq(&mapping->tree_lock);
444
445 truncate_inode_pages(mapping, 0);
446 }
447}
448EXPORT_SYMBOL(truncate_inode_pages_final);
449
450/**
451 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
452 * @mapping: the address_space which holds the pages to invalidate
453 * @start: the offset 'from' which to invalidate
454 * @end: the offset 'to' which to invalidate (inclusive)
455 *
456 * This function only removes the unlocked pages, if you want to
457 * remove all the pages of one inode, you must call truncate_inode_pages.
458 *
459 * invalidate_mapping_pages() will not block on IO activity. It will not
460 * invalidate pages which are dirty, locked, under writeback or mapped into
461 * pagetables.
462 */
463unsigned long invalidate_mapping_pages(struct address_space *mapping,
464 pgoff_t start, pgoff_t end)
465{
466 pgoff_t indices[PAGEVEC_SIZE];
467 struct pagevec pvec;
468 pgoff_t index = start;
469 unsigned long ret;
470 unsigned long count = 0;
471 int i;
472
473 pagevec_init(&pvec, 0);
474 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
475 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
476 indices)) {
477 for (i = 0; i < pagevec_count(&pvec); i++) {
478 struct page *page = pvec.pages[i];
479
480 /* We rely upon deletion not changing page->index */
481 index = indices[i];
482 if (index > end)
483 break;
484
485 if (radix_tree_exceptional_entry(page)) {
486 clear_exceptional_entry(mapping, index, page);
487 continue;
488 }
489
490 if (!trylock_page(page))
491 continue;
492 WARN_ON(page->index != index);
493 ret = invalidate_inode_page(page);
494 unlock_page(page);
495 /*
496 * Invalidation is a hint that the page is no longer
497 * of interest and try to speed up its reclaim.
498 */
499 if (!ret)
500 deactivate_file_page(page);
501 count += ret;
502 }
503 pagevec_remove_exceptionals(&pvec);
504 pagevec_release(&pvec);
505 cond_resched();
506 index++;
507 }
508 return count;
509}
510EXPORT_SYMBOL(invalidate_mapping_pages);
511
512/*
513 * This is like invalidate_complete_page(), except it ignores the page's
514 * refcount. We do this because invalidate_inode_pages2() needs stronger
515 * invalidation guarantees, and cannot afford to leave pages behind because
516 * shrink_page_list() has a temp ref on them, or because they're transiently
517 * sitting in the lru_cache_add() pagevecs.
518 */
519static int
520invalidate_complete_page2(struct address_space *mapping, struct page *page)
521{
522 unsigned long flags;
523
524 if (page->mapping != mapping)
525 return 0;
526
527 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
528 return 0;
529
530 spin_lock_irqsave(&mapping->tree_lock, flags);
531 if (PageDirty(page))
532 goto failed;
533
534 BUG_ON(page_has_private(page));
535 __delete_from_page_cache(page, NULL);
536 spin_unlock_irqrestore(&mapping->tree_lock, flags);
537
538 if (mapping->a_ops->freepage)
539 mapping->a_ops->freepage(page);
540
541 put_page(page); /* pagecache ref */
542 return 1;
543failed:
544 spin_unlock_irqrestore(&mapping->tree_lock, flags);
545 return 0;
546}
547
548static int do_launder_page(struct address_space *mapping, struct page *page)
549{
550 if (!PageDirty(page))
551 return 0;
552 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
553 return 0;
554 return mapping->a_ops->launder_page(page);
555}
556
557/**
558 * invalidate_inode_pages2_range - remove range of pages from an address_space
559 * @mapping: the address_space
560 * @start: the page offset 'from' which to invalidate
561 * @end: the page offset 'to' which to invalidate (inclusive)
562 *
563 * Any pages which are found to be mapped into pagetables are unmapped prior to
564 * invalidation.
565 *
566 * Returns -EBUSY if any pages could not be invalidated.
567 */
568int invalidate_inode_pages2_range(struct address_space *mapping,
569 pgoff_t start, pgoff_t end)
570{
571 pgoff_t indices[PAGEVEC_SIZE];
572 struct pagevec pvec;
573 pgoff_t index;
574 int i;
575 int ret = 0;
576 int ret2 = 0;
577 int did_range_unmap = 0;
578
579 cleancache_invalidate_inode(mapping);
580 pagevec_init(&pvec, 0);
581 index = start;
582 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
583 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
584 indices)) {
585 for (i = 0; i < pagevec_count(&pvec); i++) {
586 struct page *page = pvec.pages[i];
587
588 /* We rely upon deletion not changing page->index */
589 index = indices[i];
590 if (index > end)
591 break;
592
593 if (radix_tree_exceptional_entry(page)) {
594 clear_exceptional_entry(mapping, index, page);
595 continue;
596 }
597
598 lock_page(page);
599 WARN_ON(page->index != index);
600 if (page->mapping != mapping) {
601 unlock_page(page);
602 continue;
603 }
604 wait_on_page_writeback(page);
605 if (page_mapped(page)) {
606 if (!did_range_unmap) {
607 /*
608 * Zap the rest of the file in one hit.
609 */
610 unmap_mapping_range(mapping,
611 (loff_t)index << PAGE_SHIFT,
612 (loff_t)(1 + end - index)
613 << PAGE_SHIFT,
614 0);
615 did_range_unmap = 1;
616 } else {
617 /*
618 * Just zap this page
619 */
620 unmap_mapping_range(mapping,
621 (loff_t)index << PAGE_SHIFT,
622 PAGE_SIZE, 0);
623 }
624 }
625 BUG_ON(page_mapped(page));
626 ret2 = do_launder_page(mapping, page);
627 if (ret2 == 0) {
628 if (!invalidate_complete_page2(mapping, page))
629 ret2 = -EBUSY;
630 }
631 if (ret2 < 0)
632 ret = ret2;
633 unlock_page(page);
634 }
635 pagevec_remove_exceptionals(&pvec);
636 pagevec_release(&pvec);
637 cond_resched();
638 index++;
639 }
640 cleancache_invalidate_inode(mapping);
641 return ret;
642}
643EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
644
645/**
646 * invalidate_inode_pages2 - remove all pages from an address_space
647 * @mapping: the address_space
648 *
649 * Any pages which are found to be mapped into pagetables are unmapped prior to
650 * invalidation.
651 *
652 * Returns -EBUSY if any pages could not be invalidated.
653 */
654int invalidate_inode_pages2(struct address_space *mapping)
655{
656 return invalidate_inode_pages2_range(mapping, 0, -1);
657}
658EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
659
660/**
661 * truncate_pagecache - unmap and remove pagecache that has been truncated
662 * @inode: inode
663 * @newsize: new file size
664 *
665 * inode's new i_size must already be written before truncate_pagecache
666 * is called.
667 *
668 * This function should typically be called before the filesystem
669 * releases resources associated with the freed range (eg. deallocates
670 * blocks). This way, pagecache will always stay logically coherent
671 * with on-disk format, and the filesystem would not have to deal with
672 * situations such as writepage being called for a page that has already
673 * had its underlying blocks deallocated.
674 */
675void truncate_pagecache(struct inode *inode, loff_t newsize)
676{
677 struct address_space *mapping = inode->i_mapping;
678 loff_t holebegin = round_up(newsize, PAGE_SIZE);
679
680 /*
681 * unmap_mapping_range is called twice, first simply for
682 * efficiency so that truncate_inode_pages does fewer
683 * single-page unmaps. However after this first call, and
684 * before truncate_inode_pages finishes, it is possible for
685 * private pages to be COWed, which remain after
686 * truncate_inode_pages finishes, hence the second
687 * unmap_mapping_range call must be made for correctness.
688 */
689 unmap_mapping_range(mapping, holebegin, 0, 1);
690 truncate_inode_pages(mapping, newsize);
691 unmap_mapping_range(mapping, holebegin, 0, 1);
692}
693EXPORT_SYMBOL(truncate_pagecache);
694
695/**
696 * truncate_setsize - update inode and pagecache for a new file size
697 * @inode: inode
698 * @newsize: new file size
699 *
700 * truncate_setsize updates i_size and performs pagecache truncation (if
701 * necessary) to @newsize. It will be typically be called from the filesystem's
702 * setattr function when ATTR_SIZE is passed in.
703 *
704 * Must be called with a lock serializing truncates and writes (generally
705 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
706 * specific block truncation has been performed.
707 */
708void truncate_setsize(struct inode *inode, loff_t newsize)
709{
710 loff_t oldsize = inode->i_size;
711
712 i_size_write(inode, newsize);
713 if (newsize > oldsize)
714 pagecache_isize_extended(inode, oldsize, newsize);
715 truncate_pagecache(inode, newsize);
716}
717EXPORT_SYMBOL(truncate_setsize);
718
719/**
720 * pagecache_isize_extended - update pagecache after extension of i_size
721 * @inode: inode for which i_size was extended
722 * @from: original inode size
723 * @to: new inode size
724 *
725 * Handle extension of inode size either caused by extending truncate or by
726 * write starting after current i_size. We mark the page straddling current
727 * i_size RO so that page_mkwrite() is called on the nearest write access to
728 * the page. This way filesystem can be sure that page_mkwrite() is called on
729 * the page before user writes to the page via mmap after the i_size has been
730 * changed.
731 *
732 * The function must be called after i_size is updated so that page fault
733 * coming after we unlock the page will already see the new i_size.
734 * The function must be called while we still hold i_mutex - this not only
735 * makes sure i_size is stable but also that userspace cannot observe new
736 * i_size value before we are prepared to store mmap writes at new inode size.
737 */
738void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
739{
740 int bsize = 1 << inode->i_blkbits;
741 loff_t rounded_from;
742 struct page *page;
743 pgoff_t index;
744
745 WARN_ON(to > inode->i_size);
746
747 if (from >= to || bsize == PAGE_SIZE)
748 return;
749 /* Page straddling @from will not have any hole block created? */
750 rounded_from = round_up(from, bsize);
751 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
752 return;
753
754 index = from >> PAGE_SHIFT;
755 page = find_lock_page(inode->i_mapping, index);
756 /* Page not cached? Nothing to do */
757 if (!page)
758 return;
759 /*
760 * See clear_page_dirty_for_io() for details why set_page_dirty()
761 * is needed.
762 */
763 if (page_mkclean(page))
764 set_page_dirty(page);
765 unlock_page(page);
766 put_page(page);
767}
768EXPORT_SYMBOL(pagecache_isize_extended);
769
770/**
771 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
772 * @inode: inode
773 * @lstart: offset of beginning of hole
774 * @lend: offset of last byte of hole
775 *
776 * This function should typically be called before the filesystem
777 * releases resources associated with the freed range (eg. deallocates
778 * blocks). This way, pagecache will always stay logically coherent
779 * with on-disk format, and the filesystem would not have to deal with
780 * situations such as writepage being called for a page that has already
781 * had its underlying blocks deallocated.
782 */
783void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
784{
785 struct address_space *mapping = inode->i_mapping;
786 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
787 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
788 /*
789 * This rounding is currently just for example: unmap_mapping_range
790 * expands its hole outwards, whereas we want it to contract the hole
791 * inwards. However, existing callers of truncate_pagecache_range are
792 * doing their own page rounding first. Note that unmap_mapping_range
793 * allows holelen 0 for all, and we allow lend -1 for end of file.
794 */
795
796 /*
797 * Unlike in truncate_pagecache, unmap_mapping_range is called only
798 * once (before truncating pagecache), and without "even_cows" flag:
799 * hole-punching should not remove private COWed pages from the hole.
800 */
801 if ((u64)unmap_end > (u64)unmap_start)
802 unmap_mapping_range(mapping, unmap_start,
803 1 + unmap_end - unmap_start, 0);
804 truncate_inode_pages_range(mapping, lstart, lend);
805}
806EXPORT_SYMBOL(truncate_pagecache_range);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * mm/truncate.c - code for taking down pages from address_spaces
4 *
5 * Copyright (C) 2002, Linus Torvalds
6 *
7 * 10Sep2002 Andrew Morton
8 * Initial version.
9 */
10
11#include <linux/kernel.h>
12#include <linux/backing-dev.h>
13#include <linux/dax.h>
14#include <linux/gfp.h>
15#include <linux/mm.h>
16#include <linux/swap.h>
17#include <linux/export.h>
18#include <linux/pagemap.h>
19#include <linux/highmem.h>
20#include <linux/pagevec.h>
21#include <linux/task_io_accounting_ops.h>
22#include <linux/shmem_fs.h>
23#include <linux/rmap.h>
24#include "internal.h"
25
26static void clear_shadow_entries(struct address_space *mapping,
27 unsigned long start, unsigned long max)
28{
29 XA_STATE(xas, &mapping->i_pages, start);
30 struct folio *folio;
31
32 /* Handled by shmem itself, or for DAX we do nothing. */
33 if (shmem_mapping(mapping) || dax_mapping(mapping))
34 return;
35
36 xas_set_update(&xas, workingset_update_node);
37
38 spin_lock(&mapping->host->i_lock);
39 xas_lock_irq(&xas);
40
41 /* Clear all shadow entries from start to max */
42 xas_for_each(&xas, folio, max) {
43 if (xa_is_value(folio))
44 xas_store(&xas, NULL);
45 }
46
47 xas_unlock_irq(&xas);
48 if (mapping_shrinkable(mapping))
49 inode_add_lru(mapping->host);
50 spin_unlock(&mapping->host->i_lock);
51}
52
53/*
54 * Unconditionally remove exceptional entries. Usually called from truncate
55 * path. Note that the folio_batch may be altered by this function by removing
56 * exceptional entries similar to what folio_batch_remove_exceptionals() does.
57 * Please note that indices[] has entries in ascending order as guaranteed by
58 * either find_get_entries() or find_lock_entries().
59 */
60static void truncate_folio_batch_exceptionals(struct address_space *mapping,
61 struct folio_batch *fbatch, pgoff_t *indices)
62{
63 XA_STATE(xas, &mapping->i_pages, indices[0]);
64 int nr = folio_batch_count(fbatch);
65 struct folio *folio;
66 int i, j;
67
68 /* Handled by shmem itself */
69 if (shmem_mapping(mapping))
70 return;
71
72 for (j = 0; j < nr; j++)
73 if (xa_is_value(fbatch->folios[j]))
74 break;
75
76 if (j == nr)
77 return;
78
79 if (dax_mapping(mapping)) {
80 for (i = j; i < nr; i++) {
81 if (xa_is_value(fbatch->folios[i]))
82 dax_delete_mapping_entry(mapping, indices[i]);
83 }
84 goto out;
85 }
86
87 xas_set(&xas, indices[j]);
88 xas_set_update(&xas, workingset_update_node);
89
90 spin_lock(&mapping->host->i_lock);
91 xas_lock_irq(&xas);
92
93 xas_for_each(&xas, folio, indices[nr-1]) {
94 if (xa_is_value(folio))
95 xas_store(&xas, NULL);
96 }
97
98 xas_unlock_irq(&xas);
99 if (mapping_shrinkable(mapping))
100 inode_add_lru(mapping->host);
101 spin_unlock(&mapping->host->i_lock);
102out:
103 folio_batch_remove_exceptionals(fbatch);
104}
105
106/**
107 * folio_invalidate - Invalidate part or all of a folio.
108 * @folio: The folio which is affected.
109 * @offset: start of the range to invalidate
110 * @length: length of the range to invalidate
111 *
112 * folio_invalidate() is called when all or part of the folio has become
113 * invalidated by a truncate operation.
114 *
115 * folio_invalidate() does not have to release all buffers, but it must
116 * ensure that no dirty buffer is left outside @offset and that no I/O
117 * is underway against any of the blocks which are outside the truncation
118 * point. Because the caller is about to free (and possibly reuse) those
119 * blocks on-disk.
120 */
121void folio_invalidate(struct folio *folio, size_t offset, size_t length)
122{
123 const struct address_space_operations *aops = folio->mapping->a_ops;
124
125 if (aops->invalidate_folio)
126 aops->invalidate_folio(folio, offset, length);
127}
128EXPORT_SYMBOL_GPL(folio_invalidate);
129
130/*
131 * If truncate cannot remove the fs-private metadata from the page, the page
132 * becomes orphaned. It will be left on the LRU and may even be mapped into
133 * user pagetables if we're racing with filemap_fault().
134 *
135 * We need to bail out if page->mapping is no longer equal to the original
136 * mapping. This happens a) when the VM reclaimed the page while we waited on
137 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
138 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
139 */
140static void truncate_cleanup_folio(struct folio *folio)
141{
142 if (folio_mapped(folio))
143 unmap_mapping_folio(folio);
144
145 if (folio_needs_release(folio))
146 folio_invalidate(folio, 0, folio_size(folio));
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 folio_cancel_dirty(folio);
154}
155
156int truncate_inode_folio(struct address_space *mapping, struct folio *folio)
157{
158 if (folio->mapping != mapping)
159 return -EIO;
160
161 truncate_cleanup_folio(folio);
162 filemap_remove_folio(folio);
163 return 0;
164}
165
166/*
167 * Handle partial folios. The folio may be entirely within the
168 * range if a split has raced with us. If not, we zero the part of the
169 * folio that's within the [start, end] range, and then split the folio if
170 * it's large. split_page_range() will discard pages which now lie beyond
171 * i_size, and we rely on the caller to discard pages which lie within a
172 * newly created hole.
173 *
174 * Returns false if splitting failed so the caller can avoid
175 * discarding the entire folio which is stubbornly unsplit.
176 */
177bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
178{
179 loff_t pos = folio_pos(folio);
180 unsigned int offset, length;
181
182 if (pos < start)
183 offset = start - pos;
184 else
185 offset = 0;
186 length = folio_size(folio);
187 if (pos + length <= (u64)end)
188 length = length - offset;
189 else
190 length = end + 1 - pos - offset;
191
192 folio_wait_writeback(folio);
193 if (length == folio_size(folio)) {
194 truncate_inode_folio(folio->mapping, folio);
195 return true;
196 }
197
198 /*
199 * We may be zeroing pages we're about to discard, but it avoids
200 * doing a complex calculation here, and then doing the zeroing
201 * anyway if the page split fails.
202 */
203 if (!mapping_inaccessible(folio->mapping))
204 folio_zero_range(folio, offset, length);
205
206 if (folio_needs_release(folio))
207 folio_invalidate(folio, offset, length);
208 if (!folio_test_large(folio))
209 return true;
210 if (split_folio(folio) == 0)
211 return true;
212 if (folio_test_dirty(folio))
213 return false;
214 truncate_inode_folio(folio->mapping, folio);
215 return true;
216}
217
218/*
219 * Used to get rid of pages on hardware memory corruption.
220 */
221int generic_error_remove_folio(struct address_space *mapping,
222 struct folio *folio)
223{
224 if (!mapping)
225 return -EINVAL;
226 /*
227 * Only punch for normal data pages for now.
228 * Handling other types like directories would need more auditing.
229 */
230 if (!S_ISREG(mapping->host->i_mode))
231 return -EIO;
232 return truncate_inode_folio(mapping, folio);
233}
234EXPORT_SYMBOL(generic_error_remove_folio);
235
236/**
237 * mapping_evict_folio() - Remove an unused folio from the page-cache.
238 * @mapping: The mapping this folio belongs to.
239 * @folio: The folio to remove.
240 *
241 * Safely remove one folio from the page cache.
242 * It only drops clean, unused folios.
243 *
244 * Context: Folio must be locked.
245 * Return: The number of pages successfully removed.
246 */
247long mapping_evict_folio(struct address_space *mapping, struct folio *folio)
248{
249 /* The page may have been truncated before it was locked */
250 if (!mapping)
251 return 0;
252 if (folio_test_dirty(folio) || folio_test_writeback(folio))
253 return 0;
254 /* The refcount will be elevated if any page in the folio is mapped */
255 if (folio_ref_count(folio) >
256 folio_nr_pages(folio) + folio_has_private(folio) + 1)
257 return 0;
258 if (!filemap_release_folio(folio, 0))
259 return 0;
260
261 return remove_mapping(mapping, folio);
262}
263
264/**
265 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
266 * @mapping: mapping to truncate
267 * @lstart: offset from which to truncate
268 * @lend: offset to which to truncate (inclusive)
269 *
270 * Truncate the page cache, removing the pages that are between
271 * specified offsets (and zeroing out partial pages
272 * if lstart or lend + 1 is not page aligned).
273 *
274 * Truncate takes two passes - the first pass is nonblocking. It will not
275 * block on page locks and it will not block on writeback. The second pass
276 * will wait. This is to prevent as much IO as possible in the affected region.
277 * The first pass will remove most pages, so the search cost of the second pass
278 * is low.
279 *
280 * We pass down the cache-hot hint to the page freeing code. Even if the
281 * mapping is large, it is probably the case that the final pages are the most
282 * recently touched, and freeing happens in ascending file offset order.
283 *
284 * Note that since ->invalidate_folio() accepts range to invalidate
285 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
286 * page aligned properly.
287 */
288void truncate_inode_pages_range(struct address_space *mapping,
289 loff_t lstart, loff_t lend)
290{
291 pgoff_t start; /* inclusive */
292 pgoff_t end; /* exclusive */
293 struct folio_batch fbatch;
294 pgoff_t indices[PAGEVEC_SIZE];
295 pgoff_t index;
296 int i;
297 struct folio *folio;
298 bool same_folio;
299
300 if (mapping_empty(mapping))
301 return;
302
303 /*
304 * 'start' and 'end' always covers the range of pages to be fully
305 * truncated. Partial pages are covered with 'partial_start' at the
306 * start of the range and 'partial_end' at the end of the range.
307 * Note that 'end' is exclusive while 'lend' is inclusive.
308 */
309 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
310 if (lend == -1)
311 /*
312 * lend == -1 indicates end-of-file so we have to set 'end'
313 * to the highest possible pgoff_t and since the type is
314 * unsigned we're using -1.
315 */
316 end = -1;
317 else
318 end = (lend + 1) >> PAGE_SHIFT;
319
320 folio_batch_init(&fbatch);
321 index = start;
322 while (index < end && find_lock_entries(mapping, &index, end - 1,
323 &fbatch, indices)) {
324 truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
325 for (i = 0; i < folio_batch_count(&fbatch); i++)
326 truncate_cleanup_folio(fbatch.folios[i]);
327 delete_from_page_cache_batch(mapping, &fbatch);
328 for (i = 0; i < folio_batch_count(&fbatch); i++)
329 folio_unlock(fbatch.folios[i]);
330 folio_batch_release(&fbatch);
331 cond_resched();
332 }
333
334 same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
335 folio = __filemap_get_folio(mapping, lstart >> PAGE_SHIFT, FGP_LOCK, 0);
336 if (!IS_ERR(folio)) {
337 same_folio = lend < folio_pos(folio) + folio_size(folio);
338 if (!truncate_inode_partial_folio(folio, lstart, lend)) {
339 start = folio_next_index(folio);
340 if (same_folio)
341 end = folio->index;
342 }
343 folio_unlock(folio);
344 folio_put(folio);
345 folio = NULL;
346 }
347
348 if (!same_folio) {
349 folio = __filemap_get_folio(mapping, lend >> PAGE_SHIFT,
350 FGP_LOCK, 0);
351 if (!IS_ERR(folio)) {
352 if (!truncate_inode_partial_folio(folio, lstart, lend))
353 end = folio->index;
354 folio_unlock(folio);
355 folio_put(folio);
356 }
357 }
358
359 index = start;
360 while (index < end) {
361 cond_resched();
362 if (!find_get_entries(mapping, &index, end - 1, &fbatch,
363 indices)) {
364 /* If all gone from start onwards, we're done */
365 if (index == start)
366 break;
367 /* Otherwise restart to make sure all gone */
368 index = start;
369 continue;
370 }
371
372 for (i = 0; i < folio_batch_count(&fbatch); i++) {
373 struct folio *folio = fbatch.folios[i];
374
375 /* We rely upon deletion not changing page->index */
376
377 if (xa_is_value(folio))
378 continue;
379
380 folio_lock(folio);
381 VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
382 folio_wait_writeback(folio);
383 truncate_inode_folio(mapping, folio);
384 folio_unlock(folio);
385 }
386 truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
387 folio_batch_release(&fbatch);
388 }
389}
390EXPORT_SYMBOL(truncate_inode_pages_range);
391
392/**
393 * truncate_inode_pages - truncate *all* the pages from an offset
394 * @mapping: mapping to truncate
395 * @lstart: offset from which to truncate
396 *
397 * Called under (and serialised by) inode->i_rwsem and
398 * mapping->invalidate_lock.
399 *
400 * Note: When this function returns, there can be a page in the process of
401 * deletion (inside __filemap_remove_folio()) in the specified range. Thus
402 * mapping->nrpages can be non-zero when this function returns even after
403 * truncation of the whole mapping.
404 */
405void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
406{
407 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
408}
409EXPORT_SYMBOL(truncate_inode_pages);
410
411/**
412 * truncate_inode_pages_final - truncate *all* pages before inode dies
413 * @mapping: mapping to truncate
414 *
415 * Called under (and serialized by) inode->i_rwsem.
416 *
417 * Filesystems have to use this in the .evict_inode path to inform the
418 * VM that this is the final truncate and the inode is going away.
419 */
420void truncate_inode_pages_final(struct address_space *mapping)
421{
422 /*
423 * Page reclaim can not participate in regular inode lifetime
424 * management (can't call iput()) and thus can race with the
425 * inode teardown. Tell it when the address space is exiting,
426 * so that it does not install eviction information after the
427 * final truncate has begun.
428 */
429 mapping_set_exiting(mapping);
430
431 if (!mapping_empty(mapping)) {
432 /*
433 * As truncation uses a lockless tree lookup, cycle
434 * the tree lock to make sure any ongoing tree
435 * modification that does not see AS_EXITING is
436 * completed before starting the final truncate.
437 */
438 xa_lock_irq(&mapping->i_pages);
439 xa_unlock_irq(&mapping->i_pages);
440 }
441
442 truncate_inode_pages(mapping, 0);
443}
444EXPORT_SYMBOL(truncate_inode_pages_final);
445
446/**
447 * mapping_try_invalidate - Invalidate all the evictable folios of one inode
448 * @mapping: the address_space which holds the folios to invalidate
449 * @start: the offset 'from' which to invalidate
450 * @end: the offset 'to' which to invalidate (inclusive)
451 * @nr_failed: How many folio invalidations failed
452 *
453 * This function is similar to invalidate_mapping_pages(), except that it
454 * returns the number of folios which could not be evicted in @nr_failed.
455 */
456unsigned long mapping_try_invalidate(struct address_space *mapping,
457 pgoff_t start, pgoff_t end, unsigned long *nr_failed)
458{
459 pgoff_t indices[PAGEVEC_SIZE];
460 struct folio_batch fbatch;
461 pgoff_t index = start;
462 unsigned long ret;
463 unsigned long count = 0;
464 int i;
465
466 folio_batch_init(&fbatch);
467 while (find_lock_entries(mapping, &index, end, &fbatch, indices)) {
468 bool xa_has_values = false;
469 int nr = folio_batch_count(&fbatch);
470
471 for (i = 0; i < nr; i++) {
472 struct folio *folio = fbatch.folios[i];
473
474 /* We rely upon deletion not changing folio->index */
475
476 if (xa_is_value(folio)) {
477 xa_has_values = true;
478 count++;
479 continue;
480 }
481
482 ret = mapping_evict_folio(mapping, folio);
483 folio_unlock(folio);
484 /*
485 * Invalidation is a hint that the folio is no longer
486 * of interest and try to speed up its reclaim.
487 */
488 if (!ret) {
489 deactivate_file_folio(folio);
490 /* Likely in the lru cache of a remote CPU */
491 if (nr_failed)
492 (*nr_failed)++;
493 }
494 count += ret;
495 }
496
497 if (xa_has_values)
498 clear_shadow_entries(mapping, indices[0], indices[nr-1]);
499
500 folio_batch_remove_exceptionals(&fbatch);
501 folio_batch_release(&fbatch);
502 cond_resched();
503 }
504 return count;
505}
506
507/**
508 * invalidate_mapping_pages - Invalidate all clean, unlocked cache of one inode
509 * @mapping: the address_space which holds the cache to invalidate
510 * @start: the offset 'from' which to invalidate
511 * @end: the offset 'to' which to invalidate (inclusive)
512 *
513 * This function removes pages that are clean, unmapped and unlocked,
514 * as well as shadow entries. It will not block on IO activity.
515 *
516 * If you want to remove all the pages of one inode, regardless of
517 * their use and writeback state, use truncate_inode_pages().
518 *
519 * Return: The number of indices that had their contents invalidated
520 */
521unsigned long invalidate_mapping_pages(struct address_space *mapping,
522 pgoff_t start, pgoff_t end)
523{
524 return mapping_try_invalidate(mapping, start, end, NULL);
525}
526EXPORT_SYMBOL(invalidate_mapping_pages);
527
528/*
529 * This is like mapping_evict_folio(), except it ignores the folio's
530 * refcount. We do this because invalidate_inode_pages2() needs stronger
531 * invalidation guarantees, and cannot afford to leave folios behind because
532 * shrink_folio_list() has a temp ref on them, or because they're transiently
533 * sitting in the folio_add_lru() caches.
534 */
535static int invalidate_complete_folio2(struct address_space *mapping,
536 struct folio *folio)
537{
538 if (folio->mapping != mapping)
539 return 0;
540
541 if (!filemap_release_folio(folio, GFP_KERNEL))
542 return 0;
543
544 spin_lock(&mapping->host->i_lock);
545 xa_lock_irq(&mapping->i_pages);
546 if (folio_test_dirty(folio))
547 goto failed;
548
549 BUG_ON(folio_has_private(folio));
550 __filemap_remove_folio(folio, NULL);
551 xa_unlock_irq(&mapping->i_pages);
552 if (mapping_shrinkable(mapping))
553 inode_add_lru(mapping->host);
554 spin_unlock(&mapping->host->i_lock);
555
556 filemap_free_folio(mapping, folio);
557 return 1;
558failed:
559 xa_unlock_irq(&mapping->i_pages);
560 spin_unlock(&mapping->host->i_lock);
561 return 0;
562}
563
564static int folio_launder(struct address_space *mapping, struct folio *folio)
565{
566 if (!folio_test_dirty(folio))
567 return 0;
568 if (folio->mapping != mapping || mapping->a_ops->launder_folio == NULL)
569 return 0;
570 return mapping->a_ops->launder_folio(folio);
571}
572
573/**
574 * invalidate_inode_pages2_range - remove range of pages from an address_space
575 * @mapping: the address_space
576 * @start: the page offset 'from' which to invalidate
577 * @end: the page offset 'to' which to invalidate (inclusive)
578 *
579 * Any pages which are found to be mapped into pagetables are unmapped prior to
580 * invalidation.
581 *
582 * Return: -EBUSY if any pages could not be invalidated.
583 */
584int invalidate_inode_pages2_range(struct address_space *mapping,
585 pgoff_t start, pgoff_t end)
586{
587 pgoff_t indices[PAGEVEC_SIZE];
588 struct folio_batch fbatch;
589 pgoff_t index;
590 int i;
591 int ret = 0;
592 int ret2 = 0;
593 int did_range_unmap = 0;
594
595 if (mapping_empty(mapping))
596 return 0;
597
598 folio_batch_init(&fbatch);
599 index = start;
600 while (find_get_entries(mapping, &index, end, &fbatch, indices)) {
601 bool xa_has_values = false;
602 int nr = folio_batch_count(&fbatch);
603
604 for (i = 0; i < nr; i++) {
605 struct folio *folio = fbatch.folios[i];
606
607 /* We rely upon deletion not changing folio->index */
608
609 if (xa_is_value(folio)) {
610 xa_has_values = true;
611 if (dax_mapping(mapping) &&
612 !dax_invalidate_mapping_entry_sync(mapping, indices[i]))
613 ret = -EBUSY;
614 continue;
615 }
616
617 if (!did_range_unmap && folio_mapped(folio)) {
618 /*
619 * If folio is mapped, before taking its lock,
620 * zap the rest of the file in one hit.
621 */
622 unmap_mapping_pages(mapping, indices[i],
623 (1 + end - indices[i]), false);
624 did_range_unmap = 1;
625 }
626
627 folio_lock(folio);
628 if (unlikely(folio->mapping != mapping)) {
629 folio_unlock(folio);
630 continue;
631 }
632 VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
633 folio_wait_writeback(folio);
634
635 if (folio_mapped(folio))
636 unmap_mapping_folio(folio);
637 BUG_ON(folio_mapped(folio));
638
639 ret2 = folio_launder(mapping, folio);
640 if (ret2 == 0) {
641 if (!invalidate_complete_folio2(mapping, folio))
642 ret2 = -EBUSY;
643 }
644 if (ret2 < 0)
645 ret = ret2;
646 folio_unlock(folio);
647 }
648
649 if (xa_has_values)
650 clear_shadow_entries(mapping, indices[0], indices[nr-1]);
651
652 folio_batch_remove_exceptionals(&fbatch);
653 folio_batch_release(&fbatch);
654 cond_resched();
655 }
656 /*
657 * For DAX we invalidate page tables after invalidating page cache. We
658 * could invalidate page tables while invalidating each entry however
659 * that would be expensive. And doing range unmapping before doesn't
660 * work as we have no cheap way to find whether page cache entry didn't
661 * get remapped later.
662 */
663 if (dax_mapping(mapping)) {
664 unmap_mapping_pages(mapping, start, end - start + 1, false);
665 }
666 return ret;
667}
668EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
669
670/**
671 * invalidate_inode_pages2 - remove all pages from an address_space
672 * @mapping: the address_space
673 *
674 * Any pages which are found to be mapped into pagetables are unmapped prior to
675 * invalidation.
676 *
677 * Return: -EBUSY if any pages could not be invalidated.
678 */
679int invalidate_inode_pages2(struct address_space *mapping)
680{
681 return invalidate_inode_pages2_range(mapping, 0, -1);
682}
683EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
684
685/**
686 * truncate_pagecache - unmap and remove pagecache that has been truncated
687 * @inode: inode
688 * @newsize: new file size
689 *
690 * inode's new i_size must already be written before truncate_pagecache
691 * is called.
692 *
693 * This function should typically be called before the filesystem
694 * releases resources associated with the freed range (eg. deallocates
695 * blocks). This way, pagecache will always stay logically coherent
696 * with on-disk format, and the filesystem would not have to deal with
697 * situations such as writepage being called for a page that has already
698 * had its underlying blocks deallocated.
699 */
700void truncate_pagecache(struct inode *inode, loff_t newsize)
701{
702 struct address_space *mapping = inode->i_mapping;
703 loff_t holebegin = round_up(newsize, PAGE_SIZE);
704
705 /*
706 * unmap_mapping_range is called twice, first simply for
707 * efficiency so that truncate_inode_pages does fewer
708 * single-page unmaps. However after this first call, and
709 * before truncate_inode_pages finishes, it is possible for
710 * private pages to be COWed, which remain after
711 * truncate_inode_pages finishes, hence the second
712 * unmap_mapping_range call must be made for correctness.
713 */
714 unmap_mapping_range(mapping, holebegin, 0, 1);
715 truncate_inode_pages(mapping, newsize);
716 unmap_mapping_range(mapping, holebegin, 0, 1);
717}
718EXPORT_SYMBOL(truncate_pagecache);
719
720/**
721 * truncate_setsize - update inode and pagecache for a new file size
722 * @inode: inode
723 * @newsize: new file size
724 *
725 * truncate_setsize updates i_size and performs pagecache truncation (if
726 * necessary) to @newsize. It will be typically be called from the filesystem's
727 * setattr function when ATTR_SIZE is passed in.
728 *
729 * Must be called with a lock serializing truncates and writes (generally
730 * i_rwsem but e.g. xfs uses a different lock) and before all filesystem
731 * specific block truncation has been performed.
732 */
733void truncate_setsize(struct inode *inode, loff_t newsize)
734{
735 loff_t oldsize = inode->i_size;
736
737 i_size_write(inode, newsize);
738 if (newsize > oldsize)
739 pagecache_isize_extended(inode, oldsize, newsize);
740 truncate_pagecache(inode, newsize);
741}
742EXPORT_SYMBOL(truncate_setsize);
743
744/**
745 * pagecache_isize_extended - update pagecache after extension of i_size
746 * @inode: inode for which i_size was extended
747 * @from: original inode size
748 * @to: new inode size
749 *
750 * Handle extension of inode size either caused by extending truncate or
751 * by write starting after current i_size. We mark the page straddling
752 * current i_size RO so that page_mkwrite() is called on the first
753 * write access to the page. The filesystem will update its per-block
754 * information before user writes to the page via mmap after the i_size
755 * has been changed.
756 *
757 * The function must be called after i_size is updated so that page fault
758 * coming after we unlock the folio will already see the new i_size.
759 * The function must be called while we still hold i_rwsem - this not only
760 * makes sure i_size is stable but also that userspace cannot observe new
761 * i_size value before we are prepared to store mmap writes at new inode size.
762 */
763void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
764{
765 int bsize = i_blocksize(inode);
766 loff_t rounded_from;
767 struct folio *folio;
768
769 WARN_ON(to > inode->i_size);
770
771 if (from >= to || bsize >= PAGE_SIZE)
772 return;
773 /* Page straddling @from will not have any hole block created? */
774 rounded_from = round_up(from, bsize);
775 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
776 return;
777
778 folio = filemap_lock_folio(inode->i_mapping, from / PAGE_SIZE);
779 /* Folio not cached? Nothing to do */
780 if (IS_ERR(folio))
781 return;
782 /*
783 * See folio_clear_dirty_for_io() for details why folio_mark_dirty()
784 * is needed.
785 */
786 if (folio_mkclean(folio))
787 folio_mark_dirty(folio);
788
789 /*
790 * The post-eof range of the folio must be zeroed before it is exposed
791 * to the file. Writeback normally does this, but since i_size has been
792 * increased we handle it here.
793 */
794 if (folio_test_dirty(folio)) {
795 unsigned int offset, end;
796
797 offset = from - folio_pos(folio);
798 end = min_t(unsigned int, to - folio_pos(folio),
799 folio_size(folio));
800 folio_zero_segment(folio, offset, end);
801 }
802
803 folio_unlock(folio);
804 folio_put(folio);
805}
806EXPORT_SYMBOL(pagecache_isize_extended);
807
808/**
809 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
810 * @inode: inode
811 * @lstart: offset of beginning of hole
812 * @lend: offset of last byte of hole
813 *
814 * This function should typically be called before the filesystem
815 * releases resources associated with the freed range (eg. deallocates
816 * blocks). This way, pagecache will always stay logically coherent
817 * with on-disk format, and the filesystem would not have to deal with
818 * situations such as writepage being called for a page that has already
819 * had its underlying blocks deallocated.
820 */
821void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
822{
823 struct address_space *mapping = inode->i_mapping;
824 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
825 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
826 /*
827 * This rounding is currently just for example: unmap_mapping_range
828 * expands its hole outwards, whereas we want it to contract the hole
829 * inwards. However, existing callers of truncate_pagecache_range are
830 * doing their own page rounding first. Note that unmap_mapping_range
831 * allows holelen 0 for all, and we allow lend -1 for end of file.
832 */
833
834 /*
835 * Unlike in truncate_pagecache, unmap_mapping_range is called only
836 * once (before truncating pagecache), and without "even_cows" flag:
837 * hole-punching should not remove private COWed pages from the hole.
838 */
839 if ((u64)unmap_end > (u64)unmap_start)
840 unmap_mapping_range(mapping, unmap_start,
841 1 + unmap_end - unmap_start, 0);
842 truncate_inode_pages_range(mapping, lstart, lend);
843}
844EXPORT_SYMBOL(truncate_pagecache_range);