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/*
2 * mm/truncate.c - code for taking down pages from address_spaces
3 *
4 * Copyright (C) 2002, Linus Torvalds
5 *
6 * 10Sep2002 Andrew Morton
7 * Initial version.
8 */
9
10#include <linux/kernel.h>
11#include <linux/backing-dev.h>
12#include <linux/dax.h>
13#include <linux/gfp.h>
14#include <linux/mm.h>
15#include <linux/swap.h>
16#include <linux/export.h>
17#include <linux/pagemap.h>
18#include <linux/highmem.h>
19#include <linux/pagevec.h>
20#include <linux/task_io_accounting_ops.h>
21#include <linux/buffer_head.h> /* grr. try_to_release_page,
22 do_invalidatepage */
23#include <linux/shmem_fs.h>
24#include <linux/cleancache.h>
25#include <linux/rmap.h>
26#include "internal.h"
27
28/*
29 * Regular page slots are stabilized by the page lock even without the tree
30 * itself locked. These unlocked entries need verification under the tree
31 * lock.
32 */
33static inline void __clear_shadow_entry(struct address_space *mapping,
34 pgoff_t index, void *entry)
35{
36 struct radix_tree_node *node;
37 void **slot;
38
39 if (!__radix_tree_lookup(&mapping->i_pages, index, &node, &slot))
40 return;
41 if (*slot != entry)
42 return;
43 __radix_tree_replace(&mapping->i_pages, node, slot, NULL,
44 workingset_update_node);
45 mapping->nrexceptional--;
46}
47
48static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
49 void *entry)
50{
51 xa_lock_irq(&mapping->i_pages);
52 __clear_shadow_entry(mapping, index, entry);
53 xa_unlock_irq(&mapping->i_pages);
54}
55
56/*
57 * Unconditionally remove exceptional entries. Usually called from truncate
58 * path. Note that the pagevec may be altered by this function by removing
59 * exceptional entries similar to what pagevec_remove_exceptionals does.
60 */
61static void truncate_exceptional_pvec_entries(struct address_space *mapping,
62 struct pagevec *pvec, pgoff_t *indices,
63 pgoff_t end)
64{
65 int i, j;
66 bool dax, lock;
67
68 /* Handled by shmem itself */
69 if (shmem_mapping(mapping))
70 return;
71
72 for (j = 0; j < pagevec_count(pvec); j++)
73 if (radix_tree_exceptional_entry(pvec->pages[j]))
74 break;
75
76 if (j == pagevec_count(pvec))
77 return;
78
79 dax = dax_mapping(mapping);
80 lock = !dax && indices[j] < end;
81 if (lock)
82 xa_lock_irq(&mapping->i_pages);
83
84 for (i = j; i < pagevec_count(pvec); i++) {
85 struct page *page = pvec->pages[i];
86 pgoff_t index = indices[i];
87
88 if (!radix_tree_exceptional_entry(page)) {
89 pvec->pages[j++] = page;
90 continue;
91 }
92
93 if (index >= end)
94 continue;
95
96 if (unlikely(dax)) {
97 dax_delete_mapping_entry(mapping, index);
98 continue;
99 }
100
101 __clear_shadow_entry(mapping, index, page);
102 }
103
104 if (lock)
105 xa_unlock_irq(&mapping->i_pages);
106 pvec->nr = j;
107}
108
109/*
110 * Invalidate exceptional entry if easily possible. This handles exceptional
111 * entries for invalidate_inode_pages().
112 */
113static int invalidate_exceptional_entry(struct address_space *mapping,
114 pgoff_t index, void *entry)
115{
116 /* Handled by shmem itself, or for DAX we do nothing. */
117 if (shmem_mapping(mapping) || dax_mapping(mapping))
118 return 1;
119 clear_shadow_entry(mapping, index, entry);
120 return 1;
121}
122
123/*
124 * Invalidate exceptional entry if clean. This handles exceptional entries for
125 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
126 */
127static int invalidate_exceptional_entry2(struct address_space *mapping,
128 pgoff_t index, void *entry)
129{
130 /* Handled by shmem itself */
131 if (shmem_mapping(mapping))
132 return 1;
133 if (dax_mapping(mapping))
134 return dax_invalidate_mapping_entry_sync(mapping, index);
135 clear_shadow_entry(mapping, index, entry);
136 return 1;
137}
138
139/**
140 * do_invalidatepage - invalidate part or all of a page
141 * @page: the page which is affected
142 * @offset: start of the range to invalidate
143 * @length: length of the range to invalidate
144 *
145 * do_invalidatepage() is called when all or part of the page has become
146 * invalidated by a truncate operation.
147 *
148 * do_invalidatepage() does not have to release all buffers, but it must
149 * ensure that no dirty buffer is left outside @offset and that no I/O
150 * is underway against any of the blocks which are outside the truncation
151 * point. Because the caller is about to free (and possibly reuse) those
152 * blocks on-disk.
153 */
154void do_invalidatepage(struct page *page, unsigned int offset,
155 unsigned int length)
156{
157 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
158
159 invalidatepage = page->mapping->a_ops->invalidatepage;
160#ifdef CONFIG_BLOCK
161 if (!invalidatepage)
162 invalidatepage = block_invalidatepage;
163#endif
164 if (invalidatepage)
165 (*invalidatepage)(page, offset, length);
166}
167
168/*
169 * If truncate cannot remove the fs-private metadata from the page, the page
170 * becomes orphaned. It will be left on the LRU and may even be mapped into
171 * user pagetables if we're racing with filemap_fault().
172 *
173 * We need to bale out if page->mapping is no longer equal to the original
174 * mapping. This happens a) when the VM reclaimed the page while we waited on
175 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
176 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
177 */
178static void
179truncate_cleanup_page(struct address_space *mapping, struct page *page)
180{
181 if (page_mapped(page)) {
182 pgoff_t nr = PageTransHuge(page) ? HPAGE_PMD_NR : 1;
183 unmap_mapping_pages(mapping, page->index, nr, false);
184 }
185
186 if (page_has_private(page))
187 do_invalidatepage(page, 0, PAGE_SIZE);
188
189 /*
190 * Some filesystems seem to re-dirty the page even after
191 * the VM has canceled the dirty bit (eg ext3 journaling).
192 * Hence dirty accounting check is placed after invalidation.
193 */
194 cancel_dirty_page(page);
195 ClearPageMappedToDisk(page);
196}
197
198/*
199 * This is for invalidate_mapping_pages(). That function can be called at
200 * any time, and is not supposed to throw away dirty pages. But pages can
201 * be marked dirty at any time too, so use remove_mapping which safely
202 * discards clean, unused pages.
203 *
204 * Returns non-zero if the page was successfully invalidated.
205 */
206static int
207invalidate_complete_page(struct address_space *mapping, struct page *page)
208{
209 int ret;
210
211 if (page->mapping != mapping)
212 return 0;
213
214 if (page_has_private(page) && !try_to_release_page(page, 0))
215 return 0;
216
217 ret = remove_mapping(mapping, page);
218
219 return ret;
220}
221
222int truncate_inode_page(struct address_space *mapping, struct page *page)
223{
224 VM_BUG_ON_PAGE(PageTail(page), page);
225
226 if (page->mapping != mapping)
227 return -EIO;
228
229 truncate_cleanup_page(mapping, page);
230 delete_from_page_cache(page);
231 return 0;
232}
233
234/*
235 * Used to get rid of pages on hardware memory corruption.
236 */
237int generic_error_remove_page(struct address_space *mapping, struct page *page)
238{
239 if (!mapping)
240 return -EINVAL;
241 /*
242 * Only punch for normal data pages for now.
243 * Handling other types like directories would need more auditing.
244 */
245 if (!S_ISREG(mapping->host->i_mode))
246 return -EIO;
247 return truncate_inode_page(mapping, page);
248}
249EXPORT_SYMBOL(generic_error_remove_page);
250
251/*
252 * Safely invalidate one page from its pagecache mapping.
253 * It only drops clean, unused pages. The page must be locked.
254 *
255 * Returns 1 if the page is successfully invalidated, otherwise 0.
256 */
257int invalidate_inode_page(struct page *page)
258{
259 struct address_space *mapping = page_mapping(page);
260 if (!mapping)
261 return 0;
262 if (PageDirty(page) || PageWriteback(page))
263 return 0;
264 if (page_mapped(page))
265 return 0;
266 return invalidate_complete_page(mapping, page);
267}
268
269/**
270 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
271 * @mapping: mapping to truncate
272 * @lstart: offset from which to truncate
273 * @lend: offset to which to truncate (inclusive)
274 *
275 * Truncate the page cache, removing the pages that are between
276 * specified offsets (and zeroing out partial pages
277 * if lstart or lend + 1 is not page aligned).
278 *
279 * Truncate takes two passes - the first pass is nonblocking. It will not
280 * block on page locks and it will not block on writeback. The second pass
281 * will wait. This is to prevent as much IO as possible in the affected region.
282 * The first pass will remove most pages, so the search cost of the second pass
283 * is low.
284 *
285 * We pass down the cache-hot hint to the page freeing code. Even if the
286 * mapping is large, it is probably the case that the final pages are the most
287 * recently touched, and freeing happens in ascending file offset order.
288 *
289 * Note that since ->invalidatepage() accepts range to invalidate
290 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
291 * page aligned properly.
292 */
293void truncate_inode_pages_range(struct address_space *mapping,
294 loff_t lstart, loff_t lend)
295{
296 pgoff_t start; /* inclusive */
297 pgoff_t end; /* exclusive */
298 unsigned int partial_start; /* inclusive */
299 unsigned int partial_end; /* exclusive */
300 struct pagevec pvec;
301 pgoff_t indices[PAGEVEC_SIZE];
302 pgoff_t index;
303 int i;
304
305 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
306 goto out;
307
308 /* Offsets within partial pages */
309 partial_start = lstart & (PAGE_SIZE - 1);
310 partial_end = (lend + 1) & (PAGE_SIZE - 1);
311
312 /*
313 * 'start' and 'end' always covers the range of pages to be fully
314 * truncated. Partial pages are covered with 'partial_start' at the
315 * start of the range and 'partial_end' at the end of the range.
316 * Note that 'end' is exclusive while 'lend' is inclusive.
317 */
318 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
319 if (lend == -1)
320 /*
321 * lend == -1 indicates end-of-file so we have to set 'end'
322 * to the highest possible pgoff_t and since the type is
323 * unsigned we're using -1.
324 */
325 end = -1;
326 else
327 end = (lend + 1) >> PAGE_SHIFT;
328
329 pagevec_init(&pvec);
330 index = start;
331 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
332 min(end - index, (pgoff_t)PAGEVEC_SIZE),
333 indices)) {
334 /*
335 * Pagevec array has exceptional entries and we may also fail
336 * to lock some pages. So we store pages that can be deleted
337 * in a new pagevec.
338 */
339 struct pagevec locked_pvec;
340
341 pagevec_init(&locked_pvec);
342 for (i = 0; i < pagevec_count(&pvec); i++) {
343 struct page *page = pvec.pages[i];
344
345 /* We rely upon deletion not changing page->index */
346 index = indices[i];
347 if (index >= end)
348 break;
349
350 if (radix_tree_exceptional_entry(page))
351 continue;
352
353 if (!trylock_page(page))
354 continue;
355 WARN_ON(page_to_index(page) != index);
356 if (PageWriteback(page)) {
357 unlock_page(page);
358 continue;
359 }
360 if (page->mapping != mapping) {
361 unlock_page(page);
362 continue;
363 }
364 pagevec_add(&locked_pvec, page);
365 }
366 for (i = 0; i < pagevec_count(&locked_pvec); i++)
367 truncate_cleanup_page(mapping, locked_pvec.pages[i]);
368 delete_from_page_cache_batch(mapping, &locked_pvec);
369 for (i = 0; i < pagevec_count(&locked_pvec); i++)
370 unlock_page(locked_pvec.pages[i]);
371 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
372 pagevec_release(&pvec);
373 cond_resched();
374 index++;
375 }
376 if (partial_start) {
377 struct page *page = find_lock_page(mapping, start - 1);
378 if (page) {
379 unsigned int top = PAGE_SIZE;
380 if (start > end) {
381 /* Truncation within a single page */
382 top = partial_end;
383 partial_end = 0;
384 }
385 wait_on_page_writeback(page);
386 zero_user_segment(page, partial_start, top);
387 cleancache_invalidate_page(mapping, page);
388 if (page_has_private(page))
389 do_invalidatepage(page, partial_start,
390 top - partial_start);
391 unlock_page(page);
392 put_page(page);
393 }
394 }
395 if (partial_end) {
396 struct page *page = find_lock_page(mapping, end);
397 if (page) {
398 wait_on_page_writeback(page);
399 zero_user_segment(page, 0, partial_end);
400 cleancache_invalidate_page(mapping, page);
401 if (page_has_private(page))
402 do_invalidatepage(page, 0,
403 partial_end);
404 unlock_page(page);
405 put_page(page);
406 }
407 }
408 /*
409 * If the truncation happened within a single page no pages
410 * will be released, just zeroed, so we can bail out now.
411 */
412 if (start >= end)
413 goto out;
414
415 index = start;
416 for ( ; ; ) {
417 cond_resched();
418 if (!pagevec_lookup_entries(&pvec, mapping, index,
419 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
420 /* If all gone from start onwards, we're done */
421 if (index == start)
422 break;
423 /* Otherwise restart to make sure all gone */
424 index = start;
425 continue;
426 }
427 if (index == start && indices[0] >= end) {
428 /* All gone out of hole to be punched, we're done */
429 pagevec_remove_exceptionals(&pvec);
430 pagevec_release(&pvec);
431 break;
432 }
433
434 for (i = 0; i < pagevec_count(&pvec); i++) {
435 struct page *page = pvec.pages[i];
436
437 /* We rely upon deletion not changing page->index */
438 index = indices[i];
439 if (index >= end) {
440 /* Restart punch to make sure all gone */
441 index = start - 1;
442 break;
443 }
444
445 if (radix_tree_exceptional_entry(page))
446 continue;
447
448 lock_page(page);
449 WARN_ON(page_to_index(page) != index);
450 wait_on_page_writeback(page);
451 truncate_inode_page(mapping, page);
452 unlock_page(page);
453 }
454 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
455 pagevec_release(&pvec);
456 index++;
457 }
458
459out:
460 cleancache_invalidate_inode(mapping);
461}
462EXPORT_SYMBOL(truncate_inode_pages_range);
463
464/**
465 * truncate_inode_pages - truncate *all* the pages from an offset
466 * @mapping: mapping to truncate
467 * @lstart: offset from which to truncate
468 *
469 * Called under (and serialised by) inode->i_mutex.
470 *
471 * Note: When this function returns, there can be a page in the process of
472 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
473 * mapping->nrpages can be non-zero when this function returns even after
474 * truncation of the whole mapping.
475 */
476void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
477{
478 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
479}
480EXPORT_SYMBOL(truncate_inode_pages);
481
482/**
483 * truncate_inode_pages_final - truncate *all* pages before inode dies
484 * @mapping: mapping to truncate
485 *
486 * Called under (and serialized by) inode->i_mutex.
487 *
488 * Filesystems have to use this in the .evict_inode path to inform the
489 * VM that this is the final truncate and the inode is going away.
490 */
491void truncate_inode_pages_final(struct address_space *mapping)
492{
493 unsigned long nrexceptional;
494 unsigned long nrpages;
495
496 /*
497 * Page reclaim can not participate in regular inode lifetime
498 * management (can't call iput()) and thus can race with the
499 * inode teardown. Tell it when the address space is exiting,
500 * so that it does not install eviction information after the
501 * final truncate has begun.
502 */
503 mapping_set_exiting(mapping);
504
505 /*
506 * When reclaim installs eviction entries, it increases
507 * nrexceptional first, then decreases nrpages. Make sure we see
508 * this in the right order or we might miss an entry.
509 */
510 nrpages = mapping->nrpages;
511 smp_rmb();
512 nrexceptional = mapping->nrexceptional;
513
514 if (nrpages || nrexceptional) {
515 /*
516 * As truncation uses a lockless tree lookup, cycle
517 * the tree lock to make sure any ongoing tree
518 * modification that does not see AS_EXITING is
519 * completed before starting the final truncate.
520 */
521 xa_lock_irq(&mapping->i_pages);
522 xa_unlock_irq(&mapping->i_pages);
523
524 truncate_inode_pages(mapping, 0);
525 }
526}
527EXPORT_SYMBOL(truncate_inode_pages_final);
528
529/**
530 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
531 * @mapping: the address_space which holds the pages to invalidate
532 * @start: the offset 'from' which to invalidate
533 * @end: the offset 'to' which to invalidate (inclusive)
534 *
535 * This function only removes the unlocked pages, if you want to
536 * remove all the pages of one inode, you must call truncate_inode_pages.
537 *
538 * invalidate_mapping_pages() will not block on IO activity. It will not
539 * invalidate pages which are dirty, locked, under writeback or mapped into
540 * pagetables.
541 */
542unsigned long invalidate_mapping_pages(struct address_space *mapping,
543 pgoff_t start, pgoff_t end)
544{
545 pgoff_t indices[PAGEVEC_SIZE];
546 struct pagevec pvec;
547 pgoff_t index = start;
548 unsigned long ret;
549 unsigned long count = 0;
550 int i;
551
552 pagevec_init(&pvec);
553 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
554 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
555 indices)) {
556 for (i = 0; i < pagevec_count(&pvec); i++) {
557 struct page *page = pvec.pages[i];
558
559 /* We rely upon deletion not changing page->index */
560 index = indices[i];
561 if (index > end)
562 break;
563
564 if (radix_tree_exceptional_entry(page)) {
565 invalidate_exceptional_entry(mapping, index,
566 page);
567 continue;
568 }
569
570 if (!trylock_page(page))
571 continue;
572
573 WARN_ON(page_to_index(page) != index);
574
575 /* Middle of THP: skip */
576 if (PageTransTail(page)) {
577 unlock_page(page);
578 continue;
579 } else if (PageTransHuge(page)) {
580 index += HPAGE_PMD_NR - 1;
581 i += HPAGE_PMD_NR - 1;
582 /*
583 * 'end' is in the middle of THP. Don't
584 * invalidate the page as the part outside of
585 * 'end' could be still useful.
586 */
587 if (index > end) {
588 unlock_page(page);
589 continue;
590 }
591 }
592
593 ret = invalidate_inode_page(page);
594 unlock_page(page);
595 /*
596 * Invalidation is a hint that the page is no longer
597 * of interest and try to speed up its reclaim.
598 */
599 if (!ret)
600 deactivate_file_page(page);
601 count += ret;
602 }
603 pagevec_remove_exceptionals(&pvec);
604 pagevec_release(&pvec);
605 cond_resched();
606 index++;
607 }
608 return count;
609}
610EXPORT_SYMBOL(invalidate_mapping_pages);
611
612/*
613 * This is like invalidate_complete_page(), except it ignores the page's
614 * refcount. We do this because invalidate_inode_pages2() needs stronger
615 * invalidation guarantees, and cannot afford to leave pages behind because
616 * shrink_page_list() has a temp ref on them, or because they're transiently
617 * sitting in the lru_cache_add() pagevecs.
618 */
619static int
620invalidate_complete_page2(struct address_space *mapping, struct page *page)
621{
622 unsigned long flags;
623
624 if (page->mapping != mapping)
625 return 0;
626
627 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
628 return 0;
629
630 xa_lock_irqsave(&mapping->i_pages, flags);
631 if (PageDirty(page))
632 goto failed;
633
634 BUG_ON(page_has_private(page));
635 __delete_from_page_cache(page, NULL);
636 xa_unlock_irqrestore(&mapping->i_pages, flags);
637
638 if (mapping->a_ops->freepage)
639 mapping->a_ops->freepage(page);
640
641 put_page(page); /* pagecache ref */
642 return 1;
643failed:
644 xa_unlock_irqrestore(&mapping->i_pages, flags);
645 return 0;
646}
647
648static int do_launder_page(struct address_space *mapping, struct page *page)
649{
650 if (!PageDirty(page))
651 return 0;
652 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
653 return 0;
654 return mapping->a_ops->launder_page(page);
655}
656
657/**
658 * invalidate_inode_pages2_range - remove range of pages from an address_space
659 * @mapping: the address_space
660 * @start: the page offset 'from' which to invalidate
661 * @end: the page offset 'to' which to invalidate (inclusive)
662 *
663 * Any pages which are found to be mapped into pagetables are unmapped prior to
664 * invalidation.
665 *
666 * Returns -EBUSY if any pages could not be invalidated.
667 */
668int invalidate_inode_pages2_range(struct address_space *mapping,
669 pgoff_t start, pgoff_t end)
670{
671 pgoff_t indices[PAGEVEC_SIZE];
672 struct pagevec pvec;
673 pgoff_t index;
674 int i;
675 int ret = 0;
676 int ret2 = 0;
677 int did_range_unmap = 0;
678
679 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
680 goto out;
681
682 pagevec_init(&pvec);
683 index = start;
684 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
685 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
686 indices)) {
687 for (i = 0; i < pagevec_count(&pvec); i++) {
688 struct page *page = pvec.pages[i];
689
690 /* We rely upon deletion not changing page->index */
691 index = indices[i];
692 if (index > end)
693 break;
694
695 if (radix_tree_exceptional_entry(page)) {
696 if (!invalidate_exceptional_entry2(mapping,
697 index, page))
698 ret = -EBUSY;
699 continue;
700 }
701
702 lock_page(page);
703 WARN_ON(page_to_index(page) != index);
704 if (page->mapping != mapping) {
705 unlock_page(page);
706 continue;
707 }
708 wait_on_page_writeback(page);
709 if (page_mapped(page)) {
710 if (!did_range_unmap) {
711 /*
712 * Zap the rest of the file in one hit.
713 */
714 unmap_mapping_pages(mapping, index,
715 (1 + end - index), false);
716 did_range_unmap = 1;
717 } else {
718 /*
719 * Just zap this page
720 */
721 unmap_mapping_pages(mapping, index,
722 1, false);
723 }
724 }
725 BUG_ON(page_mapped(page));
726 ret2 = do_launder_page(mapping, page);
727 if (ret2 == 0) {
728 if (!invalidate_complete_page2(mapping, page))
729 ret2 = -EBUSY;
730 }
731 if (ret2 < 0)
732 ret = ret2;
733 unlock_page(page);
734 }
735 pagevec_remove_exceptionals(&pvec);
736 pagevec_release(&pvec);
737 cond_resched();
738 index++;
739 }
740 /*
741 * For DAX we invalidate page tables after invalidating radix tree. We
742 * could invalidate page tables while invalidating each entry however
743 * that would be expensive. And doing range unmapping before doesn't
744 * work as we have no cheap way to find whether radix tree entry didn't
745 * get remapped later.
746 */
747 if (dax_mapping(mapping)) {
748 unmap_mapping_pages(mapping, start, end - start + 1, false);
749 }
750out:
751 cleancache_invalidate_inode(mapping);
752 return ret;
753}
754EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
755
756/**
757 * invalidate_inode_pages2 - remove all pages from an address_space
758 * @mapping: the address_space
759 *
760 * Any pages which are found to be mapped into pagetables are unmapped prior to
761 * invalidation.
762 *
763 * Returns -EBUSY if any pages could not be invalidated.
764 */
765int invalidate_inode_pages2(struct address_space *mapping)
766{
767 return invalidate_inode_pages2_range(mapping, 0, -1);
768}
769EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
770
771/**
772 * truncate_pagecache - unmap and remove pagecache that has been truncated
773 * @inode: inode
774 * @newsize: new file size
775 *
776 * inode's new i_size must already be written before truncate_pagecache
777 * is called.
778 *
779 * This function should typically be called before the filesystem
780 * releases resources associated with the freed range (eg. deallocates
781 * blocks). This way, pagecache will always stay logically coherent
782 * with on-disk format, and the filesystem would not have to deal with
783 * situations such as writepage being called for a page that has already
784 * had its underlying blocks deallocated.
785 */
786void truncate_pagecache(struct inode *inode, loff_t newsize)
787{
788 struct address_space *mapping = inode->i_mapping;
789 loff_t holebegin = round_up(newsize, PAGE_SIZE);
790
791 /*
792 * unmap_mapping_range is called twice, first simply for
793 * efficiency so that truncate_inode_pages does fewer
794 * single-page unmaps. However after this first call, and
795 * before truncate_inode_pages finishes, it is possible for
796 * private pages to be COWed, which remain after
797 * truncate_inode_pages finishes, hence the second
798 * unmap_mapping_range call must be made for correctness.
799 */
800 unmap_mapping_range(mapping, holebegin, 0, 1);
801 truncate_inode_pages(mapping, newsize);
802 unmap_mapping_range(mapping, holebegin, 0, 1);
803}
804EXPORT_SYMBOL(truncate_pagecache);
805
806/**
807 * truncate_setsize - update inode and pagecache for a new file size
808 * @inode: inode
809 * @newsize: new file size
810 *
811 * truncate_setsize updates i_size and performs pagecache truncation (if
812 * necessary) to @newsize. It will be typically be called from the filesystem's
813 * setattr function when ATTR_SIZE is passed in.
814 *
815 * Must be called with a lock serializing truncates and writes (generally
816 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
817 * specific block truncation has been performed.
818 */
819void truncate_setsize(struct inode *inode, loff_t newsize)
820{
821 loff_t oldsize = inode->i_size;
822
823 i_size_write(inode, newsize);
824 if (newsize > oldsize)
825 pagecache_isize_extended(inode, oldsize, newsize);
826 truncate_pagecache(inode, newsize);
827}
828EXPORT_SYMBOL(truncate_setsize);
829
830/**
831 * pagecache_isize_extended - update pagecache after extension of i_size
832 * @inode: inode for which i_size was extended
833 * @from: original inode size
834 * @to: new inode size
835 *
836 * Handle extension of inode size either caused by extending truncate or by
837 * write starting after current i_size. We mark the page straddling current
838 * i_size RO so that page_mkwrite() is called on the nearest write access to
839 * the page. This way filesystem can be sure that page_mkwrite() is called on
840 * the page before user writes to the page via mmap after the i_size has been
841 * changed.
842 *
843 * The function must be called after i_size is updated so that page fault
844 * coming after we unlock the page will already see the new i_size.
845 * The function must be called while we still hold i_mutex - this not only
846 * makes sure i_size is stable but also that userspace cannot observe new
847 * i_size value before we are prepared to store mmap writes at new inode size.
848 */
849void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
850{
851 int bsize = i_blocksize(inode);
852 loff_t rounded_from;
853 struct page *page;
854 pgoff_t index;
855
856 WARN_ON(to > inode->i_size);
857
858 if (from >= to || bsize == PAGE_SIZE)
859 return;
860 /* Page straddling @from will not have any hole block created? */
861 rounded_from = round_up(from, bsize);
862 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
863 return;
864
865 index = from >> PAGE_SHIFT;
866 page = find_lock_page(inode->i_mapping, index);
867 /* Page not cached? Nothing to do */
868 if (!page)
869 return;
870 /*
871 * See clear_page_dirty_for_io() for details why set_page_dirty()
872 * is needed.
873 */
874 if (page_mkclean(page))
875 set_page_dirty(page);
876 unlock_page(page);
877 put_page(page);
878}
879EXPORT_SYMBOL(pagecache_isize_extended);
880
881/**
882 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
883 * @inode: inode
884 * @lstart: offset of beginning of hole
885 * @lend: offset of last byte of hole
886 *
887 * This function should typically be called before the filesystem
888 * releases resources associated with the freed range (eg. deallocates
889 * blocks). This way, pagecache will always stay logically coherent
890 * with on-disk format, and the filesystem would not have to deal with
891 * situations such as writepage being called for a page that has already
892 * had its underlying blocks deallocated.
893 */
894void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
895{
896 struct address_space *mapping = inode->i_mapping;
897 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
898 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
899 /*
900 * This rounding is currently just for example: unmap_mapping_range
901 * expands its hole outwards, whereas we want it to contract the hole
902 * inwards. However, existing callers of truncate_pagecache_range are
903 * doing their own page rounding first. Note that unmap_mapping_range
904 * allows holelen 0 for all, and we allow lend -1 for end of file.
905 */
906
907 /*
908 * Unlike in truncate_pagecache, unmap_mapping_range is called only
909 * once (before truncating pagecache), and without "even_cows" flag:
910 * hole-punching should not remove private COWed pages from the hole.
911 */
912 if ((u64)unmap_end > (u64)unmap_start)
913 unmap_mapping_range(mapping, unmap_start,
914 1 + unmap_end - unmap_start, 0);
915 truncate_inode_pages_range(mapping, lstart, lend);
916}
917EXPORT_SYMBOL(truncate_pagecache_range);