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