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 do_invalidatepage */
24#include <linux/shmem_fs.h>
25#include <linux/cleancache.h>
26#include <linux/rmap.h>
27#include "internal.h"
28
29/*
30 * Regular page slots are stabilized by the page lock even without the tree
31 * itself locked. These unlocked entries need verification under the tree
32 * lock.
33 */
34static inline void __clear_shadow_entry(struct address_space *mapping,
35 pgoff_t index, void *entry)
36{
37 XA_STATE(xas, &mapping->i_pages, index);
38
39 xas_set_update(&xas, workingset_update_node);
40 if (xas_load(&xas) != entry)
41 return;
42 xas_store(&xas, NULL);
43 mapping->nrexceptional--;
44}
45
46static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
47 void *entry)
48{
49 xa_lock_irq(&mapping->i_pages);
50 __clear_shadow_entry(mapping, index, entry);
51 xa_unlock_irq(&mapping->i_pages);
52}
53
54/*
55 * Unconditionally remove exceptional entries. Usually called from truncate
56 * path. Note that the pagevec may be altered by this function by removing
57 * exceptional entries similar to what pagevec_remove_exceptionals does.
58 */
59static void truncate_exceptional_pvec_entries(struct address_space *mapping,
60 struct pagevec *pvec, pgoff_t *indices,
61 pgoff_t end)
62{
63 int i, j;
64 bool dax, lock;
65
66 /* Handled by shmem itself */
67 if (shmem_mapping(mapping))
68 return;
69
70 for (j = 0; j < pagevec_count(pvec); j++)
71 if (xa_is_value(pvec->pages[j]))
72 break;
73
74 if (j == pagevec_count(pvec))
75 return;
76
77 dax = dax_mapping(mapping);
78 lock = !dax && indices[j] < end;
79 if (lock)
80 xa_lock_irq(&mapping->i_pages);
81
82 for (i = j; i < pagevec_count(pvec); i++) {
83 struct page *page = pvec->pages[i];
84 pgoff_t index = indices[i];
85
86 if (!xa_is_value(page)) {
87 pvec->pages[j++] = page;
88 continue;
89 }
90
91 if (index >= end)
92 continue;
93
94 if (unlikely(dax)) {
95 dax_delete_mapping_entry(mapping, index);
96 continue;
97 }
98
99 __clear_shadow_entry(mapping, index, page);
100 }
101
102 if (lock)
103 xa_unlock_irq(&mapping->i_pages);
104 pvec->nr = j;
105}
106
107/*
108 * Invalidate exceptional entry if easily possible. This handles exceptional
109 * entries for invalidate_inode_pages().
110 */
111static int invalidate_exceptional_entry(struct address_space *mapping,
112 pgoff_t index, void *entry)
113{
114 /* Handled by shmem itself, or for DAX we do nothing. */
115 if (shmem_mapping(mapping) || dax_mapping(mapping))
116 return 1;
117 clear_shadow_entry(mapping, index, entry);
118 return 1;
119}
120
121/*
122 * Invalidate exceptional entry if clean. This handles exceptional entries for
123 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
124 */
125static int invalidate_exceptional_entry2(struct address_space *mapping,
126 pgoff_t index, void *entry)
127{
128 /* Handled by shmem itself */
129 if (shmem_mapping(mapping))
130 return 1;
131 if (dax_mapping(mapping))
132 return dax_invalidate_mapping_entry_sync(mapping, index);
133 clear_shadow_entry(mapping, index, entry);
134 return 1;
135}
136
137/**
138 * do_invalidatepage - invalidate part or all of a page
139 * @page: the page which is affected
140 * @offset: start of the range to invalidate
141 * @length: length of the range to invalidate
142 *
143 * do_invalidatepage() is called when all or part of the page has become
144 * invalidated by a truncate operation.
145 *
146 * do_invalidatepage() does not have to release all buffers, but it must
147 * ensure that no dirty buffer is left outside @offset and that no I/O
148 * is underway against any of the blocks which are outside the truncation
149 * point. Because the caller is about to free (and possibly reuse) those
150 * blocks on-disk.
151 */
152void do_invalidatepage(struct page *page, unsigned int offset,
153 unsigned int length)
154{
155 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
156
157 invalidatepage = page->mapping->a_ops->invalidatepage;
158#ifdef CONFIG_BLOCK
159 if (!invalidatepage)
160 invalidatepage = block_invalidatepage;
161#endif
162 if (invalidatepage)
163 (*invalidatepage)(page, offset, length);
164}
165
166/*
167 * If truncate cannot remove the fs-private metadata from the page, the page
168 * becomes orphaned. It will be left on the LRU and may even be mapped into
169 * user pagetables if we're racing with filemap_fault().
170 *
171 * We need to bale out if page->mapping is no longer equal to the original
172 * mapping. This happens a) when the VM reclaimed the page while we waited on
173 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
174 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
175 */
176static void
177truncate_cleanup_page(struct address_space *mapping, struct page *page)
178{
179 if (page_mapped(page)) {
180 pgoff_t nr = PageTransHuge(page) ? HPAGE_PMD_NR : 1;
181 unmap_mapping_pages(mapping, page->index, nr, false);
182 }
183
184 if (page_has_private(page))
185 do_invalidatepage(page, 0, PAGE_SIZE);
186
187 /*
188 * Some filesystems seem to re-dirty the page even after
189 * the VM has canceled the dirty bit (eg ext3 journaling).
190 * Hence dirty accounting check is placed after invalidation.
191 */
192 cancel_dirty_page(page);
193 ClearPageMappedToDisk(page);
194}
195
196/*
197 * This is for invalidate_mapping_pages(). That function can be called at
198 * any time, and is not supposed to throw away dirty pages. But pages can
199 * be marked dirty at any time too, so use remove_mapping which safely
200 * discards clean, unused pages.
201 *
202 * Returns non-zero if the page was successfully invalidated.
203 */
204static int
205invalidate_complete_page(struct address_space *mapping, struct page *page)
206{
207 int ret;
208
209 if (page->mapping != mapping)
210 return 0;
211
212 if (page_has_private(page) && !try_to_release_page(page, 0))
213 return 0;
214
215 ret = remove_mapping(mapping, page);
216
217 return ret;
218}
219
220int truncate_inode_page(struct address_space *mapping, struct page *page)
221{
222 VM_BUG_ON_PAGE(PageTail(page), page);
223
224 if (page->mapping != mapping)
225 return -EIO;
226
227 truncate_cleanup_page(mapping, page);
228 delete_from_page_cache(page);
229 return 0;
230}
231
232/*
233 * Used to get rid of pages on hardware memory corruption.
234 */
235int generic_error_remove_page(struct address_space *mapping, struct page *page)
236{
237 if (!mapping)
238 return -EINVAL;
239 /*
240 * Only punch for normal data pages for now.
241 * Handling other types like directories would need more auditing.
242 */
243 if (!S_ISREG(mapping->host->i_mode))
244 return -EIO;
245 return truncate_inode_page(mapping, page);
246}
247EXPORT_SYMBOL(generic_error_remove_page);
248
249/*
250 * Safely invalidate one page from its pagecache mapping.
251 * It only drops clean, unused pages. The page must be locked.
252 *
253 * Returns 1 if the page is successfully invalidated, otherwise 0.
254 */
255int invalidate_inode_page(struct page *page)
256{
257 struct address_space *mapping = page_mapping(page);
258 if (!mapping)
259 return 0;
260 if (PageDirty(page) || PageWriteback(page))
261 return 0;
262 if (page_mapped(page))
263 return 0;
264 return invalidate_complete_page(mapping, page);
265}
266
267/**
268 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
269 * @mapping: mapping to truncate
270 * @lstart: offset from which to truncate
271 * @lend: offset to which to truncate (inclusive)
272 *
273 * Truncate the page cache, removing the pages that are between
274 * specified offsets (and zeroing out partial pages
275 * if lstart or lend + 1 is not page aligned).
276 *
277 * Truncate takes two passes - the first pass is nonblocking. It will not
278 * block on page locks and it will not block on writeback. The second pass
279 * will wait. This is to prevent as much IO as possible in the affected region.
280 * The first pass will remove most pages, so the search cost of the second pass
281 * is low.
282 *
283 * We pass down the cache-hot hint to the page freeing code. Even if the
284 * mapping is large, it is probably the case that the final pages are the most
285 * recently touched, and freeing happens in ascending file offset order.
286 *
287 * Note that since ->invalidatepage() accepts range to invalidate
288 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
289 * page aligned properly.
290 */
291void truncate_inode_pages_range(struct address_space *mapping,
292 loff_t lstart, loff_t lend)
293{
294 pgoff_t start; /* inclusive */
295 pgoff_t end; /* exclusive */
296 unsigned int partial_start; /* inclusive */
297 unsigned int partial_end; /* exclusive */
298 struct pagevec pvec;
299 pgoff_t indices[PAGEVEC_SIZE];
300 pgoff_t index;
301 int i;
302
303 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
304 goto out;
305
306 /* Offsets within partial pages */
307 partial_start = lstart & (PAGE_SIZE - 1);
308 partial_end = (lend + 1) & (PAGE_SIZE - 1);
309
310 /*
311 * 'start' and 'end' always covers the range of pages to be fully
312 * truncated. Partial pages are covered with 'partial_start' at the
313 * start of the range and 'partial_end' at the end of the range.
314 * Note that 'end' is exclusive while 'lend' is inclusive.
315 */
316 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
317 if (lend == -1)
318 /*
319 * lend == -1 indicates end-of-file so we have to set 'end'
320 * to the highest possible pgoff_t and since the type is
321 * unsigned we're using -1.
322 */
323 end = -1;
324 else
325 end = (lend + 1) >> PAGE_SHIFT;
326
327 pagevec_init(&pvec);
328 index = start;
329 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
330 min(end - index, (pgoff_t)PAGEVEC_SIZE),
331 indices)) {
332 /*
333 * Pagevec array has exceptional entries and we may also fail
334 * to lock some pages. So we store pages that can be deleted
335 * in a new pagevec.
336 */
337 struct pagevec locked_pvec;
338
339 pagevec_init(&locked_pvec);
340 for (i = 0; i < pagevec_count(&pvec); i++) {
341 struct page *page = pvec.pages[i];
342
343 /* We rely upon deletion not changing page->index */
344 index = indices[i];
345 if (index >= end)
346 break;
347
348 if (xa_is_value(page))
349 continue;
350
351 if (!trylock_page(page))
352 continue;
353 WARN_ON(page_to_index(page) != index);
354 if (PageWriteback(page)) {
355 unlock_page(page);
356 continue;
357 }
358 if (page->mapping != mapping) {
359 unlock_page(page);
360 continue;
361 }
362 pagevec_add(&locked_pvec, page);
363 }
364 for (i = 0; i < pagevec_count(&locked_pvec); i++)
365 truncate_cleanup_page(mapping, locked_pvec.pages[i]);
366 delete_from_page_cache_batch(mapping, &locked_pvec);
367 for (i = 0; i < pagevec_count(&locked_pvec); i++)
368 unlock_page(locked_pvec.pages[i]);
369 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
370 pagevec_release(&pvec);
371 cond_resched();
372 index++;
373 }
374 if (partial_start) {
375 struct page *page = find_lock_page(mapping, start - 1);
376 if (page) {
377 unsigned int top = PAGE_SIZE;
378 if (start > end) {
379 /* Truncation within a single page */
380 top = partial_end;
381 partial_end = 0;
382 }
383 wait_on_page_writeback(page);
384 zero_user_segment(page, partial_start, top);
385 cleancache_invalidate_page(mapping, page);
386 if (page_has_private(page))
387 do_invalidatepage(page, partial_start,
388 top - partial_start);
389 unlock_page(page);
390 put_page(page);
391 }
392 }
393 if (partial_end) {
394 struct page *page = find_lock_page(mapping, end);
395 if (page) {
396 wait_on_page_writeback(page);
397 zero_user_segment(page, 0, partial_end);
398 cleancache_invalidate_page(mapping, page);
399 if (page_has_private(page))
400 do_invalidatepage(page, 0,
401 partial_end);
402 unlock_page(page);
403 put_page(page);
404 }
405 }
406 /*
407 * If the truncation happened within a single page no pages
408 * will be released, just zeroed, so we can bail out now.
409 */
410 if (start >= end)
411 goto out;
412
413 index = start;
414 for ( ; ; ) {
415 cond_resched();
416 if (!pagevec_lookup_entries(&pvec, mapping, index,
417 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
418 /* If all gone from start onwards, we're done */
419 if (index == start)
420 break;
421 /* Otherwise restart to make sure all gone */
422 index = start;
423 continue;
424 }
425 if (index == start && indices[0] >= end) {
426 /* All gone out of hole to be punched, we're done */
427 pagevec_remove_exceptionals(&pvec);
428 pagevec_release(&pvec);
429 break;
430 }
431
432 for (i = 0; i < pagevec_count(&pvec); i++) {
433 struct page *page = pvec.pages[i];
434
435 /* We rely upon deletion not changing page->index */
436 index = indices[i];
437 if (index >= end) {
438 /* Restart punch to make sure all gone */
439 index = start - 1;
440 break;
441 }
442
443 if (xa_is_value(page))
444 continue;
445
446 lock_page(page);
447 WARN_ON(page_to_index(page) != index);
448 wait_on_page_writeback(page);
449 truncate_inode_page(mapping, page);
450 unlock_page(page);
451 }
452 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
453 pagevec_release(&pvec);
454 index++;
455 }
456
457out:
458 cleancache_invalidate_inode(mapping);
459}
460EXPORT_SYMBOL(truncate_inode_pages_range);
461
462/**
463 * truncate_inode_pages - truncate *all* the pages from an offset
464 * @mapping: mapping to truncate
465 * @lstart: offset from which to truncate
466 *
467 * Called under (and serialised by) inode->i_mutex.
468 *
469 * Note: When this function returns, there can be a page in the process of
470 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
471 * mapping->nrpages can be non-zero when this function returns even after
472 * truncation of the whole mapping.
473 */
474void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
475{
476 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
477}
478EXPORT_SYMBOL(truncate_inode_pages);
479
480/**
481 * truncate_inode_pages_final - truncate *all* pages before inode dies
482 * @mapping: mapping to truncate
483 *
484 * Called under (and serialized by) inode->i_mutex.
485 *
486 * Filesystems have to use this in the .evict_inode path to inform the
487 * VM that this is the final truncate and the inode is going away.
488 */
489void truncate_inode_pages_final(struct address_space *mapping)
490{
491 unsigned long nrexceptional;
492 unsigned long nrpages;
493
494 /*
495 * Page reclaim can not participate in regular inode lifetime
496 * management (can't call iput()) and thus can race with the
497 * inode teardown. Tell it when the address space is exiting,
498 * so that it does not install eviction information after the
499 * final truncate has begun.
500 */
501 mapping_set_exiting(mapping);
502
503 /*
504 * When reclaim installs eviction entries, it increases
505 * nrexceptional first, then decreases nrpages. Make sure we see
506 * this in the right order or we might miss an entry.
507 */
508 nrpages = mapping->nrpages;
509 smp_rmb();
510 nrexceptional = mapping->nrexceptional;
511
512 if (nrpages || nrexceptional) {
513 /*
514 * As truncation uses a lockless tree lookup, cycle
515 * the tree lock to make sure any ongoing tree
516 * modification that does not see AS_EXITING is
517 * completed before starting the final truncate.
518 */
519 xa_lock_irq(&mapping->i_pages);
520 xa_unlock_irq(&mapping->i_pages);
521 }
522
523 /*
524 * Cleancache needs notification even if there are no pages or shadow
525 * entries.
526 */
527 truncate_inode_pages(mapping, 0);
528}
529EXPORT_SYMBOL(truncate_inode_pages_final);
530
531/**
532 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
533 * @mapping: the address_space which holds the pages to invalidate
534 * @start: the offset 'from' which to invalidate
535 * @end: the offset 'to' which to invalidate (inclusive)
536 *
537 * This function only removes the unlocked pages, if you want to
538 * remove all the pages of one inode, you must call truncate_inode_pages.
539 *
540 * invalidate_mapping_pages() will not block on IO activity. It will not
541 * invalidate pages which are dirty, locked, under writeback or mapped into
542 * pagetables.
543 *
544 * Return: the number of the pages that were invalidated
545 */
546unsigned long invalidate_mapping_pages(struct address_space *mapping,
547 pgoff_t start, pgoff_t end)
548{
549 pgoff_t indices[PAGEVEC_SIZE];
550 struct pagevec pvec;
551 pgoff_t index = start;
552 unsigned long ret;
553 unsigned long count = 0;
554 int i;
555
556 pagevec_init(&pvec);
557 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
558 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
559 indices)) {
560 for (i = 0; i < pagevec_count(&pvec); i++) {
561 struct page *page = pvec.pages[i];
562
563 /* We rely upon deletion not changing page->index */
564 index = indices[i];
565 if (index > end)
566 break;
567
568 if (xa_is_value(page)) {
569 invalidate_exceptional_entry(mapping, index,
570 page);
571 continue;
572 }
573
574 if (!trylock_page(page))
575 continue;
576
577 WARN_ON(page_to_index(page) != index);
578
579 /* Middle of THP: skip */
580 if (PageTransTail(page)) {
581 unlock_page(page);
582 continue;
583 } else if (PageTransHuge(page)) {
584 index += HPAGE_PMD_NR - 1;
585 i += HPAGE_PMD_NR - 1;
586 /*
587 * 'end' is in the middle of THP. Don't
588 * invalidate the page as the part outside of
589 * 'end' could be still useful.
590 */
591 if (index > end) {
592 unlock_page(page);
593 continue;
594 }
595
596 /* Take a pin outside pagevec */
597 get_page(page);
598
599 /*
600 * Drop extra pins before trying to invalidate
601 * the huge page.
602 */
603 pagevec_remove_exceptionals(&pvec);
604 pagevec_release(&pvec);
605 }
606
607 ret = invalidate_inode_page(page);
608 unlock_page(page);
609 /*
610 * Invalidation is a hint that the page is no longer
611 * of interest and try to speed up its reclaim.
612 */
613 if (!ret)
614 deactivate_file_page(page);
615 if (PageTransHuge(page))
616 put_page(page);
617 count += ret;
618 }
619 pagevec_remove_exceptionals(&pvec);
620 pagevec_release(&pvec);
621 cond_resched();
622 index++;
623 }
624 return count;
625}
626EXPORT_SYMBOL(invalidate_mapping_pages);
627
628/*
629 * This is like invalidate_complete_page(), except it ignores the page's
630 * refcount. We do this because invalidate_inode_pages2() needs stronger
631 * invalidation guarantees, and cannot afford to leave pages behind because
632 * shrink_page_list() has a temp ref on them, or because they're transiently
633 * sitting in the lru_cache_add() pagevecs.
634 */
635static int
636invalidate_complete_page2(struct address_space *mapping, struct page *page)
637{
638 unsigned long flags;
639
640 if (page->mapping != mapping)
641 return 0;
642
643 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
644 return 0;
645
646 xa_lock_irqsave(&mapping->i_pages, flags);
647 if (PageDirty(page))
648 goto failed;
649
650 BUG_ON(page_has_private(page));
651 __delete_from_page_cache(page, NULL);
652 xa_unlock_irqrestore(&mapping->i_pages, flags);
653
654 if (mapping->a_ops->freepage)
655 mapping->a_ops->freepage(page);
656
657 put_page(page); /* pagecache ref */
658 return 1;
659failed:
660 xa_unlock_irqrestore(&mapping->i_pages, flags);
661 return 0;
662}
663
664static int do_launder_page(struct address_space *mapping, struct page *page)
665{
666 if (!PageDirty(page))
667 return 0;
668 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
669 return 0;
670 return mapping->a_ops->launder_page(page);
671}
672
673/**
674 * invalidate_inode_pages2_range - remove range of pages from an address_space
675 * @mapping: the address_space
676 * @start: the page offset 'from' which to invalidate
677 * @end: the page offset 'to' which to invalidate (inclusive)
678 *
679 * Any pages which are found to be mapped into pagetables are unmapped prior to
680 * invalidation.
681 *
682 * Return: -EBUSY if any pages could not be invalidated.
683 */
684int invalidate_inode_pages2_range(struct address_space *mapping,
685 pgoff_t start, pgoff_t end)
686{
687 pgoff_t indices[PAGEVEC_SIZE];
688 struct pagevec pvec;
689 pgoff_t index;
690 int i;
691 int ret = 0;
692 int ret2 = 0;
693 int did_range_unmap = 0;
694
695 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
696 goto out;
697
698 pagevec_init(&pvec);
699 index = start;
700 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
701 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
702 indices)) {
703 for (i = 0; i < pagevec_count(&pvec); i++) {
704 struct page *page = pvec.pages[i];
705
706 /* We rely upon deletion not changing page->index */
707 index = indices[i];
708 if (index > end)
709 break;
710
711 if (xa_is_value(page)) {
712 if (!invalidate_exceptional_entry2(mapping,
713 index, page))
714 ret = -EBUSY;
715 continue;
716 }
717
718 lock_page(page);
719 WARN_ON(page_to_index(page) != index);
720 if (page->mapping != mapping) {
721 unlock_page(page);
722 continue;
723 }
724 wait_on_page_writeback(page);
725 if (page_mapped(page)) {
726 if (!did_range_unmap) {
727 /*
728 * Zap the rest of the file in one hit.
729 */
730 unmap_mapping_pages(mapping, index,
731 (1 + end - index), false);
732 did_range_unmap = 1;
733 } else {
734 /*
735 * Just zap this page
736 */
737 unmap_mapping_pages(mapping, index,
738 1, false);
739 }
740 }
741 BUG_ON(page_mapped(page));
742 ret2 = do_launder_page(mapping, page);
743 if (ret2 == 0) {
744 if (!invalidate_complete_page2(mapping, page))
745 ret2 = -EBUSY;
746 }
747 if (ret2 < 0)
748 ret = ret2;
749 unlock_page(page);
750 }
751 pagevec_remove_exceptionals(&pvec);
752 pagevec_release(&pvec);
753 cond_resched();
754 index++;
755 }
756 /*
757 * For DAX we invalidate page tables after invalidating page cache. We
758 * could invalidate page tables while invalidating each entry however
759 * that would be expensive. And doing range unmapping before doesn't
760 * work as we have no cheap way to find whether page cache entry didn't
761 * get remapped later.
762 */
763 if (dax_mapping(mapping)) {
764 unmap_mapping_pages(mapping, start, end - start + 1, false);
765 }
766out:
767 cleancache_invalidate_inode(mapping);
768 return ret;
769}
770EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
771
772/**
773 * invalidate_inode_pages2 - remove all pages from an address_space
774 * @mapping: the address_space
775 *
776 * Any pages which are found to be mapped into pagetables are unmapped prior to
777 * invalidation.
778 *
779 * Return: -EBUSY if any pages could not be invalidated.
780 */
781int invalidate_inode_pages2(struct address_space *mapping)
782{
783 return invalidate_inode_pages2_range(mapping, 0, -1);
784}
785EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
786
787/**
788 * truncate_pagecache - unmap and remove pagecache that has been truncated
789 * @inode: inode
790 * @newsize: new file size
791 *
792 * inode's new i_size must already be written before truncate_pagecache
793 * is called.
794 *
795 * This function should typically be called before the filesystem
796 * releases resources associated with the freed range (eg. deallocates
797 * blocks). This way, pagecache will always stay logically coherent
798 * with on-disk format, and the filesystem would not have to deal with
799 * situations such as writepage being called for a page that has already
800 * had its underlying blocks deallocated.
801 */
802void truncate_pagecache(struct inode *inode, loff_t newsize)
803{
804 struct address_space *mapping = inode->i_mapping;
805 loff_t holebegin = round_up(newsize, PAGE_SIZE);
806
807 /*
808 * unmap_mapping_range is called twice, first simply for
809 * efficiency so that truncate_inode_pages does fewer
810 * single-page unmaps. However after this first call, and
811 * before truncate_inode_pages finishes, it is possible for
812 * private pages to be COWed, which remain after
813 * truncate_inode_pages finishes, hence the second
814 * unmap_mapping_range call must be made for correctness.
815 */
816 unmap_mapping_range(mapping, holebegin, 0, 1);
817 truncate_inode_pages(mapping, newsize);
818 unmap_mapping_range(mapping, holebegin, 0, 1);
819}
820EXPORT_SYMBOL(truncate_pagecache);
821
822/**
823 * truncate_setsize - update inode and pagecache for a new file size
824 * @inode: inode
825 * @newsize: new file size
826 *
827 * truncate_setsize updates i_size and performs pagecache truncation (if
828 * necessary) to @newsize. It will be typically be called from the filesystem's
829 * setattr function when ATTR_SIZE is passed in.
830 *
831 * Must be called with a lock serializing truncates and writes (generally
832 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
833 * specific block truncation has been performed.
834 */
835void truncate_setsize(struct inode *inode, loff_t newsize)
836{
837 loff_t oldsize = inode->i_size;
838
839 i_size_write(inode, newsize);
840 if (newsize > oldsize)
841 pagecache_isize_extended(inode, oldsize, newsize);
842 truncate_pagecache(inode, newsize);
843}
844EXPORT_SYMBOL(truncate_setsize);
845
846/**
847 * pagecache_isize_extended - update pagecache after extension of i_size
848 * @inode: inode for which i_size was extended
849 * @from: original inode size
850 * @to: new inode size
851 *
852 * Handle extension of inode size either caused by extending truncate or by
853 * write starting after current i_size. We mark the page straddling current
854 * i_size RO so that page_mkwrite() is called on the nearest write access to
855 * the page. This way filesystem can be sure that page_mkwrite() is called on
856 * the page before user writes to the page via mmap after the i_size has been
857 * changed.
858 *
859 * The function must be called after i_size is updated so that page fault
860 * coming after we unlock the page will already see the new i_size.
861 * The function must be called while we still hold i_mutex - this not only
862 * makes sure i_size is stable but also that userspace cannot observe new
863 * i_size value before we are prepared to store mmap writes at new inode size.
864 */
865void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
866{
867 int bsize = i_blocksize(inode);
868 loff_t rounded_from;
869 struct page *page;
870 pgoff_t index;
871
872 WARN_ON(to > inode->i_size);
873
874 if (from >= to || bsize == PAGE_SIZE)
875 return;
876 /* Page straddling @from will not have any hole block created? */
877 rounded_from = round_up(from, bsize);
878 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
879 return;
880
881 index = from >> PAGE_SHIFT;
882 page = find_lock_page(inode->i_mapping, index);
883 /* Page not cached? Nothing to do */
884 if (!page)
885 return;
886 /*
887 * See clear_page_dirty_for_io() for details why set_page_dirty()
888 * is needed.
889 */
890 if (page_mkclean(page))
891 set_page_dirty(page);
892 unlock_page(page);
893 put_page(page);
894}
895EXPORT_SYMBOL(pagecache_isize_extended);
896
897/**
898 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
899 * @inode: inode
900 * @lstart: offset of beginning of hole
901 * @lend: offset of last byte of hole
902 *
903 * This function should typically be called before the filesystem
904 * releases resources associated with the freed range (eg. deallocates
905 * blocks). This way, pagecache will always stay logically coherent
906 * with on-disk format, and the filesystem would not have to deal with
907 * situations such as writepage being called for a page that has already
908 * had its underlying blocks deallocated.
909 */
910void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
911{
912 struct address_space *mapping = inode->i_mapping;
913 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
914 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
915 /*
916 * This rounding is currently just for example: unmap_mapping_range
917 * expands its hole outwards, whereas we want it to contract the hole
918 * inwards. However, existing callers of truncate_pagecache_range are
919 * doing their own page rounding first. Note that unmap_mapping_range
920 * allows holelen 0 for all, and we allow lend -1 for end of file.
921 */
922
923 /*
924 * Unlike in truncate_pagecache, unmap_mapping_range is called only
925 * once (before truncating pagecache), and without "even_cows" flag:
926 * hole-punching should not remove private COWed pages from the hole.
927 */
928 if ((u64)unmap_end > (u64)unmap_start)
929 unmap_mapping_range(mapping, unmap_start,
930 1 + unmap_end - unmap_start, 0);
931 truncate_inode_pages_range(mapping, lstart, lend);
932}
933EXPORT_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);