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/module.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
25
26/**
27 * do_invalidatepage - invalidate part or all of a page
28 * @page: the page which is affected
29 * @offset: the index of the truncation point
30 *
31 * do_invalidatepage() is called when all or part of the page has become
32 * invalidated by a truncate operation.
33 *
34 * do_invalidatepage() does not have to release all buffers, but it must
35 * ensure that no dirty buffer is left outside @offset and that no I/O
36 * is underway against any of the blocks which are outside the truncation
37 * point. Because the caller is about to free (and possibly reuse) those
38 * blocks on-disk.
39 */
40void do_invalidatepage(struct page *page, unsigned long offset)
41{
42 void (*invalidatepage)(struct page *, unsigned long);
43 invalidatepage = page->mapping->a_ops->invalidatepage;
44#ifdef CONFIG_BLOCK
45 if (!invalidatepage)
46 invalidatepage = block_invalidatepage;
47#endif
48 if (invalidatepage)
49 (*invalidatepage)(page, offset);
50}
51
52static inline void truncate_partial_page(struct page *page, unsigned partial)
53{
54 zero_user_segment(page, partial, PAGE_CACHE_SIZE);
55 cleancache_flush_page(page->mapping, page);
56 if (page_has_private(page))
57 do_invalidatepage(page, partial);
58}
59
60/*
61 * This cancels just the dirty bit on the kernel page itself, it
62 * does NOT actually remove dirty bits on any mmap's that may be
63 * around. It also leaves the page tagged dirty, so any sync
64 * activity will still find it on the dirty lists, and in particular,
65 * clear_page_dirty_for_io() will still look at the dirty bits in
66 * the VM.
67 *
68 * Doing this should *normally* only ever be done when a page
69 * is truncated, and is not actually mapped anywhere at all. However,
70 * fs/buffer.c does this when it notices that somebody has cleaned
71 * out all the buffers on a page without actually doing it through
72 * the VM. Can you say "ext3 is horribly ugly"? Tought you could.
73 */
74void cancel_dirty_page(struct page *page, unsigned int account_size)
75{
76 if (TestClearPageDirty(page)) {
77 struct address_space *mapping = page->mapping;
78 if (mapping && mapping_cap_account_dirty(mapping)) {
79 dec_zone_page_state(page, NR_FILE_DIRTY);
80 dec_bdi_stat(mapping->backing_dev_info,
81 BDI_RECLAIMABLE);
82 if (account_size)
83 task_io_account_cancelled_write(account_size);
84 }
85 }
86}
87EXPORT_SYMBOL(cancel_dirty_page);
88
89/*
90 * If truncate cannot remove the fs-private metadata from the page, the page
91 * becomes orphaned. It will be left on the LRU and may even be mapped into
92 * user pagetables if we're racing with filemap_fault().
93 *
94 * We need to bale out if page->mapping is no longer equal to the original
95 * mapping. This happens a) when the VM reclaimed the page while we waited on
96 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
97 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
98 */
99static int
100truncate_complete_page(struct address_space *mapping, struct page *page)
101{
102 if (page->mapping != mapping)
103 return -EIO;
104
105 if (page_has_private(page))
106 do_invalidatepage(page, 0);
107
108 cancel_dirty_page(page, PAGE_CACHE_SIZE);
109
110 clear_page_mlock(page);
111 ClearPageMappedToDisk(page);
112 delete_from_page_cache(page);
113 return 0;
114}
115
116/*
117 * This is for invalidate_mapping_pages(). That function can be called at
118 * any time, and is not supposed to throw away dirty pages. But pages can
119 * be marked dirty at any time too, so use remove_mapping which safely
120 * discards clean, unused pages.
121 *
122 * Returns non-zero if the page was successfully invalidated.
123 */
124static int
125invalidate_complete_page(struct address_space *mapping, struct page *page)
126{
127 int ret;
128
129 if (page->mapping != mapping)
130 return 0;
131
132 if (page_has_private(page) && !try_to_release_page(page, 0))
133 return 0;
134
135 clear_page_mlock(page);
136 ret = remove_mapping(mapping, page);
137
138 return ret;
139}
140
141int truncate_inode_page(struct address_space *mapping, struct page *page)
142{
143 if (page_mapped(page)) {
144 unmap_mapping_range(mapping,
145 (loff_t)page->index << PAGE_CACHE_SHIFT,
146 PAGE_CACHE_SIZE, 0);
147 }
148 return truncate_complete_page(mapping, page);
149}
150
151/*
152 * Used to get rid of pages on hardware memory corruption.
153 */
154int generic_error_remove_page(struct address_space *mapping, struct page *page)
155{
156 if (!mapping)
157 return -EINVAL;
158 /*
159 * Only punch for normal data pages for now.
160 * Handling other types like directories would need more auditing.
161 */
162 if (!S_ISREG(mapping->host->i_mode))
163 return -EIO;
164 return truncate_inode_page(mapping, page);
165}
166EXPORT_SYMBOL(generic_error_remove_page);
167
168/*
169 * Safely invalidate one page from its pagecache mapping.
170 * It only drops clean, unused pages. The page must be locked.
171 *
172 * Returns 1 if the page is successfully invalidated, otherwise 0.
173 */
174int invalidate_inode_page(struct page *page)
175{
176 struct address_space *mapping = page_mapping(page);
177 if (!mapping)
178 return 0;
179 if (PageDirty(page) || PageWriteback(page))
180 return 0;
181 if (page_mapped(page))
182 return 0;
183 return invalidate_complete_page(mapping, page);
184}
185
186/**
187 * truncate_inode_pages - truncate range of pages specified by start & end byte offsets
188 * @mapping: mapping to truncate
189 * @lstart: offset from which to truncate
190 * @lend: offset to which to truncate
191 *
192 * Truncate the page cache, removing the pages that are between
193 * specified offsets (and zeroing out partial page
194 * (if lstart is not page aligned)).
195 *
196 * Truncate takes two passes - the first pass is nonblocking. It will not
197 * block on page locks and it will not block on writeback. The second pass
198 * will wait. This is to prevent as much IO as possible in the affected region.
199 * The first pass will remove most pages, so the search cost of the second pass
200 * is low.
201 *
202 * We pass down the cache-hot hint to the page freeing code. Even if the
203 * mapping is large, it is probably the case that the final pages are the most
204 * recently touched, and freeing happens in ascending file offset order.
205 */
206void truncate_inode_pages_range(struct address_space *mapping,
207 loff_t lstart, loff_t lend)
208{
209 const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
210 const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
211 struct pagevec pvec;
212 pgoff_t index;
213 pgoff_t end;
214 int i;
215
216 cleancache_flush_inode(mapping);
217 if (mapping->nrpages == 0)
218 return;
219
220 BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
221 end = (lend >> PAGE_CACHE_SHIFT);
222
223 pagevec_init(&pvec, 0);
224 index = start;
225 while (index <= end && pagevec_lookup(&pvec, mapping, index,
226 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
227 mem_cgroup_uncharge_start();
228 for (i = 0; i < pagevec_count(&pvec); i++) {
229 struct page *page = pvec.pages[i];
230
231 /* We rely upon deletion not changing page->index */
232 index = page->index;
233 if (index > end)
234 break;
235
236 if (!trylock_page(page))
237 continue;
238 WARN_ON(page->index != index);
239 if (PageWriteback(page)) {
240 unlock_page(page);
241 continue;
242 }
243 truncate_inode_page(mapping, page);
244 unlock_page(page);
245 }
246 pagevec_release(&pvec);
247 mem_cgroup_uncharge_end();
248 cond_resched();
249 index++;
250 }
251
252 if (partial) {
253 struct page *page = find_lock_page(mapping, start - 1);
254 if (page) {
255 wait_on_page_writeback(page);
256 truncate_partial_page(page, partial);
257 unlock_page(page);
258 page_cache_release(page);
259 }
260 }
261
262 index = start;
263 for ( ; ; ) {
264 cond_resched();
265 if (!pagevec_lookup(&pvec, mapping, index,
266 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
267 if (index == start)
268 break;
269 index = start;
270 continue;
271 }
272 if (index == start && pvec.pages[0]->index > end) {
273 pagevec_release(&pvec);
274 break;
275 }
276 mem_cgroup_uncharge_start();
277 for (i = 0; i < pagevec_count(&pvec); i++) {
278 struct page *page = pvec.pages[i];
279
280 /* We rely upon deletion not changing page->index */
281 index = page->index;
282 if (index > end)
283 break;
284
285 lock_page(page);
286 WARN_ON(page->index != index);
287 wait_on_page_writeback(page);
288 truncate_inode_page(mapping, page);
289 unlock_page(page);
290 }
291 pagevec_release(&pvec);
292 mem_cgroup_uncharge_end();
293 index++;
294 }
295 cleancache_flush_inode(mapping);
296}
297EXPORT_SYMBOL(truncate_inode_pages_range);
298
299/**
300 * truncate_inode_pages - truncate *all* the pages from an offset
301 * @mapping: mapping to truncate
302 * @lstart: offset from which to truncate
303 *
304 * Called under (and serialised by) inode->i_mutex.
305 *
306 * Note: When this function returns, there can be a page in the process of
307 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
308 * mapping->nrpages can be non-zero when this function returns even after
309 * truncation of the whole mapping.
310 */
311void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
312{
313 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
314}
315EXPORT_SYMBOL(truncate_inode_pages);
316
317/**
318 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
319 * @mapping: the address_space which holds the pages to invalidate
320 * @start: the offset 'from' which to invalidate
321 * @end: the offset 'to' which to invalidate (inclusive)
322 *
323 * This function only removes the unlocked pages, if you want to
324 * remove all the pages of one inode, you must call truncate_inode_pages.
325 *
326 * invalidate_mapping_pages() will not block on IO activity. It will not
327 * invalidate pages which are dirty, locked, under writeback or mapped into
328 * pagetables.
329 */
330unsigned long invalidate_mapping_pages(struct address_space *mapping,
331 pgoff_t start, pgoff_t end)
332{
333 struct pagevec pvec;
334 pgoff_t index = start;
335 unsigned long ret;
336 unsigned long count = 0;
337 int i;
338
339 /*
340 * Note: this function may get called on a shmem/tmpfs mapping:
341 * pagevec_lookup() might then return 0 prematurely (because it
342 * got a gangful of swap entries); but it's hardly worth worrying
343 * about - it can rarely have anything to free from such a mapping
344 * (most pages are dirty), and already skips over any difficulties.
345 */
346
347 pagevec_init(&pvec, 0);
348 while (index <= end && pagevec_lookup(&pvec, mapping, index,
349 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
350 mem_cgroup_uncharge_start();
351 for (i = 0; i < pagevec_count(&pvec); i++) {
352 struct page *page = pvec.pages[i];
353
354 /* We rely upon deletion not changing page->index */
355 index = page->index;
356 if (index > end)
357 break;
358
359 if (!trylock_page(page))
360 continue;
361 WARN_ON(page->index != index);
362 ret = invalidate_inode_page(page);
363 unlock_page(page);
364 /*
365 * Invalidation is a hint that the page is no longer
366 * of interest and try to speed up its reclaim.
367 */
368 if (!ret)
369 deactivate_page(page);
370 count += ret;
371 }
372 pagevec_release(&pvec);
373 mem_cgroup_uncharge_end();
374 cond_resched();
375 index++;
376 }
377 return count;
378}
379EXPORT_SYMBOL(invalidate_mapping_pages);
380
381/*
382 * This is like invalidate_complete_page(), except it ignores the page's
383 * refcount. We do this because invalidate_inode_pages2() needs stronger
384 * invalidation guarantees, and cannot afford to leave pages behind because
385 * shrink_page_list() has a temp ref on them, or because they're transiently
386 * sitting in the lru_cache_add() pagevecs.
387 */
388static int
389invalidate_complete_page2(struct address_space *mapping, struct page *page)
390{
391 if (page->mapping != mapping)
392 return 0;
393
394 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
395 return 0;
396
397 spin_lock_irq(&mapping->tree_lock);
398 if (PageDirty(page))
399 goto failed;
400
401 clear_page_mlock(page);
402 BUG_ON(page_has_private(page));
403 __delete_from_page_cache(page);
404 spin_unlock_irq(&mapping->tree_lock);
405 mem_cgroup_uncharge_cache_page(page);
406
407 if (mapping->a_ops->freepage)
408 mapping->a_ops->freepage(page);
409
410 page_cache_release(page); /* pagecache ref */
411 return 1;
412failed:
413 spin_unlock_irq(&mapping->tree_lock);
414 return 0;
415}
416
417static int do_launder_page(struct address_space *mapping, struct page *page)
418{
419 if (!PageDirty(page))
420 return 0;
421 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
422 return 0;
423 return mapping->a_ops->launder_page(page);
424}
425
426/**
427 * invalidate_inode_pages2_range - remove range of pages from an address_space
428 * @mapping: the address_space
429 * @start: the page offset 'from' which to invalidate
430 * @end: the page offset 'to' which to invalidate (inclusive)
431 *
432 * Any pages which are found to be mapped into pagetables are unmapped prior to
433 * invalidation.
434 *
435 * Returns -EBUSY if any pages could not be invalidated.
436 */
437int invalidate_inode_pages2_range(struct address_space *mapping,
438 pgoff_t start, pgoff_t end)
439{
440 struct pagevec pvec;
441 pgoff_t index;
442 int i;
443 int ret = 0;
444 int ret2 = 0;
445 int did_range_unmap = 0;
446
447 cleancache_flush_inode(mapping);
448 pagevec_init(&pvec, 0);
449 index = start;
450 while (index <= end && pagevec_lookup(&pvec, mapping, index,
451 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
452 mem_cgroup_uncharge_start();
453 for (i = 0; i < pagevec_count(&pvec); i++) {
454 struct page *page = pvec.pages[i];
455
456 /* We rely upon deletion not changing page->index */
457 index = page->index;
458 if (index > end)
459 break;
460
461 lock_page(page);
462 WARN_ON(page->index != index);
463 if (page->mapping != mapping) {
464 unlock_page(page);
465 continue;
466 }
467 wait_on_page_writeback(page);
468 if (page_mapped(page)) {
469 if (!did_range_unmap) {
470 /*
471 * Zap the rest of the file in one hit.
472 */
473 unmap_mapping_range(mapping,
474 (loff_t)index << PAGE_CACHE_SHIFT,
475 (loff_t)(1 + end - index)
476 << PAGE_CACHE_SHIFT,
477 0);
478 did_range_unmap = 1;
479 } else {
480 /*
481 * Just zap this page
482 */
483 unmap_mapping_range(mapping,
484 (loff_t)index << PAGE_CACHE_SHIFT,
485 PAGE_CACHE_SIZE, 0);
486 }
487 }
488 BUG_ON(page_mapped(page));
489 ret2 = do_launder_page(mapping, page);
490 if (ret2 == 0) {
491 if (!invalidate_complete_page2(mapping, page))
492 ret2 = -EBUSY;
493 }
494 if (ret2 < 0)
495 ret = ret2;
496 unlock_page(page);
497 }
498 pagevec_release(&pvec);
499 mem_cgroup_uncharge_end();
500 cond_resched();
501 index++;
502 }
503 cleancache_flush_inode(mapping);
504 return ret;
505}
506EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
507
508/**
509 * invalidate_inode_pages2 - remove all pages from an address_space
510 * @mapping: the address_space
511 *
512 * Any pages which are found to be mapped into pagetables are unmapped prior to
513 * invalidation.
514 *
515 * Returns -EBUSY if any pages could not be invalidated.
516 */
517int invalidate_inode_pages2(struct address_space *mapping)
518{
519 return invalidate_inode_pages2_range(mapping, 0, -1);
520}
521EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
522
523/**
524 * truncate_pagecache - unmap and remove pagecache that has been truncated
525 * @inode: inode
526 * @oldsize: old file size
527 * @newsize: new file size
528 *
529 * inode's new i_size must already be written before truncate_pagecache
530 * is called.
531 *
532 * This function should typically be called before the filesystem
533 * releases resources associated with the freed range (eg. deallocates
534 * blocks). This way, pagecache will always stay logically coherent
535 * with on-disk format, and the filesystem would not have to deal with
536 * situations such as writepage being called for a page that has already
537 * had its underlying blocks deallocated.
538 */
539void truncate_pagecache(struct inode *inode, loff_t oldsize, loff_t newsize)
540{
541 struct address_space *mapping = inode->i_mapping;
542 loff_t holebegin = round_up(newsize, PAGE_SIZE);
543
544 /*
545 * unmap_mapping_range is called twice, first simply for
546 * efficiency so that truncate_inode_pages does fewer
547 * single-page unmaps. However after this first call, and
548 * before truncate_inode_pages finishes, it is possible for
549 * private pages to be COWed, which remain after
550 * truncate_inode_pages finishes, hence the second
551 * unmap_mapping_range call must be made for correctness.
552 */
553 unmap_mapping_range(mapping, holebegin, 0, 1);
554 truncate_inode_pages(mapping, newsize);
555 unmap_mapping_range(mapping, holebegin, 0, 1);
556}
557EXPORT_SYMBOL(truncate_pagecache);
558
559/**
560 * truncate_setsize - update inode and pagecache for a new file size
561 * @inode: inode
562 * @newsize: new file size
563 *
564 * truncate_setsize updates i_size and performs pagecache truncation (if
565 * necessary) to @newsize. It will be typically be called from the filesystem's
566 * setattr function when ATTR_SIZE is passed in.
567 *
568 * Must be called with inode_mutex held and before all filesystem specific
569 * block truncation has been performed.
570 */
571void truncate_setsize(struct inode *inode, loff_t newsize)
572{
573 loff_t oldsize;
574
575 oldsize = inode->i_size;
576 i_size_write(inode, newsize);
577
578 truncate_pagecache(inode, oldsize, newsize);
579}
580EXPORT_SYMBOL(truncate_setsize);
581
582/**
583 * vmtruncate - unmap mappings "freed" by truncate() syscall
584 * @inode: inode of the file used
585 * @newsize: file offset to start truncating
586 *
587 * This function is deprecated and truncate_setsize or truncate_pagecache
588 * should be used instead, together with filesystem specific block truncation.
589 */
590int vmtruncate(struct inode *inode, loff_t newsize)
591{
592 int error;
593
594 error = inode_newsize_ok(inode, newsize);
595 if (error)
596 return error;
597
598 truncate_setsize(inode, newsize);
599 if (inode->i_op->truncate)
600 inode->i_op->truncate(inode);
601 return 0;
602}
603EXPORT_SYMBOL(vmtruncate);
604
605int vmtruncate_range(struct inode *inode, loff_t lstart, loff_t lend)
606{
607 struct address_space *mapping = inode->i_mapping;
608 loff_t holebegin = round_up(lstart, PAGE_SIZE);
609 loff_t holelen = 1 + lend - holebegin;
610
611 /*
612 * If the underlying filesystem is not going to provide
613 * a way to truncate a range of blocks (punch a hole) -
614 * we should return failure right now.
615 */
616 if (!inode->i_op->truncate_range)
617 return -ENOSYS;
618
619 mutex_lock(&inode->i_mutex);
620 inode_dio_wait(inode);
621 unmap_mapping_range(mapping, holebegin, holelen, 1);
622 inode->i_op->truncate_range(inode, lstart, lend);
623 /* unmap again to remove racily COWed private pages */
624 unmap_mapping_range(mapping, holebegin, holelen, 1);
625 mutex_unlock(&inode->i_mutex);
626
627 return 0;
628}