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