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