2 * mm/truncate.c - code for taking down pages from address_spaces
4 * Copyright (C) 2002, Linus Torvalds
6 * 10Sep2002 Andrew Morton
10 #include <linux/kernel.h>
11 #include <linux/backing-dev.h>
12 #include <linux/gfp.h>
14 #include <linux/swap.h>
15 #include <linux/export.h>
16 #include <linux/pagemap.h>
17 #include <linux/highmem.h>
18 #include <linux/pagevec.h>
19 #include <linux/task_io_accounting_ops.h>
20 #include <linux/buffer_head.h> /* grr. try_to_release_page,
22 #include <linux/cleancache.h>
23 #include <linux/rmap.h>
26 static void clear_exceptional_entry(struct address_space *mapping,
27 pgoff_t index, void *entry)
29 struct radix_tree_node *node;
32 /* Handled by shmem itself */
33 if (shmem_mapping(mapping))
36 spin_lock_irq(&mapping->tree_lock);
38 * Regular page slots are stabilized by the page lock even
39 * without the tree itself locked. These unlocked entries
40 * need verification under the tree lock.
42 if (!__radix_tree_lookup(&mapping->page_tree, index, &node, &slot))
46 radix_tree_replace_slot(slot, NULL);
50 workingset_node_shadows_dec(node);
52 * Don't track node without shadow entries.
54 * Avoid acquiring the list_lru lock if already untracked.
55 * The list_empty() test is safe as node->private_list is
56 * protected by mapping->tree_lock.
58 if (!workingset_node_shadows(node) &&
59 !list_empty(&node->private_list)) {
60 local_lock(workingset_shadow_lock);
61 list_lru_del(&__workingset_shadow_nodes, &node->private_list);
62 local_unlock(workingset_shadow_lock);
64 __radix_tree_delete_node(&mapping->page_tree, node);
66 spin_unlock_irq(&mapping->tree_lock);
70 * do_invalidatepage - invalidate part or all of a page
71 * @page: the page which is affected
72 * @offset: start of the range to invalidate
73 * @length: length of the range to invalidate
75 * do_invalidatepage() is called when all or part of the page has become
76 * invalidated by a truncate operation.
78 * do_invalidatepage() does not have to release all buffers, but it must
79 * ensure that no dirty buffer is left outside @offset and that no I/O
80 * is underway against any of the blocks which are outside the truncation
81 * point. Because the caller is about to free (and possibly reuse) those
84 void do_invalidatepage(struct page *page, unsigned int offset,
87 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
89 invalidatepage = page->mapping->a_ops->invalidatepage;
92 invalidatepage = block_invalidatepage;
95 (*invalidatepage)(page, offset, length);
99 * If truncate cannot remove the fs-private metadata from the page, the page
100 * becomes orphaned. It will be left on the LRU and may even be mapped into
101 * user pagetables if we're racing with filemap_fault().
103 * We need to bale out if page->mapping is no longer equal to the original
104 * mapping. This happens a) when the VM reclaimed the page while we waited on
105 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
106 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
109 truncate_complete_page(struct address_space *mapping, struct page *page)
111 if (page->mapping != mapping)
114 if (page_has_private(page))
115 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
118 * Some filesystems seem to re-dirty the page even after
119 * the VM has canceled the dirty bit (eg ext3 journaling).
120 * Hence dirty accounting check is placed after invalidation.
122 cancel_dirty_page(page);
123 ClearPageMappedToDisk(page);
124 delete_from_page_cache(page);
129 * This is for invalidate_mapping_pages(). That function can be called at
130 * any time, and is not supposed to throw away dirty pages. But pages can
131 * be marked dirty at any time too, so use remove_mapping which safely
132 * discards clean, unused pages.
134 * Returns non-zero if the page was successfully invalidated.
137 invalidate_complete_page(struct address_space *mapping, struct page *page)
141 if (page->mapping != mapping)
144 if (page_has_private(page) && !try_to_release_page(page, 0))
147 ret = remove_mapping(mapping, page);
152 int truncate_inode_page(struct address_space *mapping, struct page *page)
154 if (page_mapped(page)) {
155 unmap_mapping_range(mapping,
156 (loff_t)page->index << PAGE_CACHE_SHIFT,
159 return truncate_complete_page(mapping, page);
163 * Used to get rid of pages on hardware memory corruption.
165 int generic_error_remove_page(struct address_space *mapping, struct page *page)
170 * Only punch for normal data pages for now.
171 * Handling other types like directories would need more auditing.
173 if (!S_ISREG(mapping->host->i_mode))
175 return truncate_inode_page(mapping, page);
177 EXPORT_SYMBOL(generic_error_remove_page);
180 * Safely invalidate one page from its pagecache mapping.
181 * It only drops clean, unused pages. The page must be locked.
183 * Returns 1 if the page is successfully invalidated, otherwise 0.
185 int invalidate_inode_page(struct page *page)
187 struct address_space *mapping = page_mapping(page);
190 if (PageDirty(page) || PageWriteback(page))
192 if (page_mapped(page))
194 return invalidate_complete_page(mapping, page);
198 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
199 * @mapping: mapping to truncate
200 * @lstart: offset from which to truncate
201 * @lend: offset to which to truncate (inclusive)
203 * Truncate the page cache, removing the pages that are between
204 * specified offsets (and zeroing out partial pages
205 * if lstart or lend + 1 is not page aligned).
207 * Truncate takes two passes - the first pass is nonblocking. It will not
208 * block on page locks and it will not block on writeback. The second pass
209 * will wait. This is to prevent as much IO as possible in the affected region.
210 * The first pass will remove most pages, so the search cost of the second pass
213 * We pass down the cache-hot hint to the page freeing code. Even if the
214 * mapping is large, it is probably the case that the final pages are the most
215 * recently touched, and freeing happens in ascending file offset order.
217 * Note that since ->invalidatepage() accepts range to invalidate
218 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
219 * page aligned properly.
221 void truncate_inode_pages_range(struct address_space *mapping,
222 loff_t lstart, loff_t lend)
224 pgoff_t start; /* inclusive */
225 pgoff_t end; /* exclusive */
226 unsigned int partial_start; /* inclusive */
227 unsigned int partial_end; /* exclusive */
229 pgoff_t indices[PAGEVEC_SIZE];
233 cleancache_invalidate_inode(mapping);
234 if (mapping->nrpages == 0 && mapping->nrshadows == 0)
237 /* Offsets within partial pages */
238 partial_start = lstart & (PAGE_CACHE_SIZE - 1);
239 partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
242 * 'start' and 'end' always covers the range of pages to be fully
243 * truncated. Partial pages are covered with 'partial_start' at the
244 * start of the range and 'partial_end' at the end of the range.
245 * Note that 'end' is exclusive while 'lend' is inclusive.
247 start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
250 * lend == -1 indicates end-of-file so we have to set 'end'
251 * to the highest possible pgoff_t and since the type is
252 * unsigned we're using -1.
256 end = (lend + 1) >> PAGE_CACHE_SHIFT;
258 pagevec_init(&pvec, 0);
260 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
261 min(end - index, (pgoff_t)PAGEVEC_SIZE),
263 for (i = 0; i < pagevec_count(&pvec); i++) {
264 struct page *page = pvec.pages[i];
266 /* We rely upon deletion not changing page->index */
271 if (radix_tree_exceptional_entry(page)) {
272 clear_exceptional_entry(mapping, index, page);
276 if (!trylock_page(page))
278 WARN_ON(page->index != index);
279 if (PageWriteback(page)) {
283 truncate_inode_page(mapping, page);
286 pagevec_remove_exceptionals(&pvec);
287 pagevec_release(&pvec);
293 struct page *page = find_lock_page(mapping, start - 1);
295 unsigned int top = PAGE_CACHE_SIZE;
297 /* Truncation within a single page */
301 wait_on_page_writeback(page);
302 zero_user_segment(page, partial_start, top);
303 cleancache_invalidate_page(mapping, page);
304 if (page_has_private(page))
305 do_invalidatepage(page, partial_start,
306 top - partial_start);
308 page_cache_release(page);
312 struct page *page = find_lock_page(mapping, end);
314 wait_on_page_writeback(page);
315 zero_user_segment(page, 0, partial_end);
316 cleancache_invalidate_page(mapping, page);
317 if (page_has_private(page))
318 do_invalidatepage(page, 0,
321 page_cache_release(page);
325 * If the truncation happened within a single page no pages
326 * will be released, just zeroed, so we can bail out now.
334 if (!pagevec_lookup_entries(&pvec, mapping, index,
335 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
336 /* If all gone from start onwards, we're done */
339 /* Otherwise restart to make sure all gone */
343 if (index == start && indices[0] >= end) {
344 /* All gone out of hole to be punched, we're done */
345 pagevec_remove_exceptionals(&pvec);
346 pagevec_release(&pvec);
349 for (i = 0; i < pagevec_count(&pvec); i++) {
350 struct page *page = pvec.pages[i];
352 /* We rely upon deletion not changing page->index */
355 /* Restart punch to make sure all gone */
360 if (radix_tree_exceptional_entry(page)) {
361 clear_exceptional_entry(mapping, index, page);
366 WARN_ON(page->index != index);
367 wait_on_page_writeback(page);
368 truncate_inode_page(mapping, page);
371 pagevec_remove_exceptionals(&pvec);
372 pagevec_release(&pvec);
375 cleancache_invalidate_inode(mapping);
377 EXPORT_SYMBOL(truncate_inode_pages_range);
380 * truncate_inode_pages - truncate *all* the pages from an offset
381 * @mapping: mapping to truncate
382 * @lstart: offset from which to truncate
384 * Called under (and serialised by) inode->i_mutex.
386 * Note: When this function returns, there can be a page in the process of
387 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
388 * mapping->nrpages can be non-zero when this function returns even after
389 * truncation of the whole mapping.
391 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
393 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
395 EXPORT_SYMBOL(truncate_inode_pages);
398 * truncate_inode_pages_final - truncate *all* pages before inode dies
399 * @mapping: mapping to truncate
401 * Called under (and serialized by) inode->i_mutex.
403 * Filesystems have to use this in the .evict_inode path to inform the
404 * VM that this is the final truncate and the inode is going away.
406 void truncate_inode_pages_final(struct address_space *mapping)
408 unsigned long nrshadows;
409 unsigned long nrpages;
412 * Page reclaim can not participate in regular inode lifetime
413 * management (can't call iput()) and thus can race with the
414 * inode teardown. Tell it when the address space is exiting,
415 * so that it does not install eviction information after the
416 * final truncate has begun.
418 mapping_set_exiting(mapping);
421 * When reclaim installs eviction entries, it increases
422 * nrshadows first, then decreases nrpages. Make sure we see
423 * this in the right order or we might miss an entry.
425 nrpages = mapping->nrpages;
427 nrshadows = mapping->nrshadows;
429 if (nrpages || nrshadows) {
431 * As truncation uses a lockless tree lookup, cycle
432 * the tree lock to make sure any ongoing tree
433 * modification that does not see AS_EXITING is
434 * completed before starting the final truncate.
436 spin_lock_irq(&mapping->tree_lock);
437 spin_unlock_irq(&mapping->tree_lock);
439 truncate_inode_pages(mapping, 0);
442 EXPORT_SYMBOL(truncate_inode_pages_final);
445 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
446 * @mapping: the address_space which holds the pages to invalidate
447 * @start: the offset 'from' which to invalidate
448 * @end: the offset 'to' which to invalidate (inclusive)
450 * This function only removes the unlocked pages, if you want to
451 * remove all the pages of one inode, you must call truncate_inode_pages.
453 * invalidate_mapping_pages() will not block on IO activity. It will not
454 * invalidate pages which are dirty, locked, under writeback or mapped into
457 unsigned long invalidate_mapping_pages(struct address_space *mapping,
458 pgoff_t start, pgoff_t end)
460 pgoff_t indices[PAGEVEC_SIZE];
462 pgoff_t index = start;
464 unsigned long count = 0;
467 pagevec_init(&pvec, 0);
468 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
469 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
471 for (i = 0; i < pagevec_count(&pvec); i++) {
472 struct page *page = pvec.pages[i];
474 /* We rely upon deletion not changing page->index */
479 if (radix_tree_exceptional_entry(page)) {
480 clear_exceptional_entry(mapping, index, page);
484 if (!trylock_page(page))
486 WARN_ON(page->index != index);
487 ret = invalidate_inode_page(page);
490 * Invalidation is a hint that the page is no longer
491 * of interest and try to speed up its reclaim.
494 deactivate_file_page(page);
497 pagevec_remove_exceptionals(&pvec);
498 pagevec_release(&pvec);
504 EXPORT_SYMBOL(invalidate_mapping_pages);
507 * This is like invalidate_complete_page(), except it ignores the page's
508 * refcount. We do this because invalidate_inode_pages2() needs stronger
509 * invalidation guarantees, and cannot afford to leave pages behind because
510 * shrink_page_list() has a temp ref on them, or because they're transiently
511 * sitting in the lru_cache_add() pagevecs.
514 invalidate_complete_page2(struct address_space *mapping, struct page *page)
516 struct mem_cgroup *memcg;
519 if (page->mapping != mapping)
522 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
525 memcg = mem_cgroup_begin_page_stat(page);
526 spin_lock_irqsave(&mapping->tree_lock, flags);
530 BUG_ON(page_has_private(page));
531 __delete_from_page_cache(page, NULL, memcg);
532 spin_unlock_irqrestore(&mapping->tree_lock, flags);
533 mem_cgroup_end_page_stat(memcg);
535 if (mapping->a_ops->freepage)
536 mapping->a_ops->freepage(page);
538 page_cache_release(page); /* pagecache ref */
541 spin_unlock_irqrestore(&mapping->tree_lock, flags);
542 mem_cgroup_end_page_stat(memcg);
546 static int do_launder_page(struct address_space *mapping, struct page *page)
548 if (!PageDirty(page))
550 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
552 return mapping->a_ops->launder_page(page);
556 * invalidate_inode_pages2_range - remove range of pages from an address_space
557 * @mapping: the address_space
558 * @start: the page offset 'from' which to invalidate
559 * @end: the page offset 'to' which to invalidate (inclusive)
561 * Any pages which are found to be mapped into pagetables are unmapped prior to
564 * Returns -EBUSY if any pages could not be invalidated.
566 int invalidate_inode_pages2_range(struct address_space *mapping,
567 pgoff_t start, pgoff_t end)
569 pgoff_t indices[PAGEVEC_SIZE];
575 int did_range_unmap = 0;
577 cleancache_invalidate_inode(mapping);
578 pagevec_init(&pvec, 0);
580 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
581 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
583 for (i = 0; i < pagevec_count(&pvec); i++) {
584 struct page *page = pvec.pages[i];
586 /* We rely upon deletion not changing page->index */
591 if (radix_tree_exceptional_entry(page)) {
592 clear_exceptional_entry(mapping, index, page);
597 WARN_ON(page->index != index);
598 if (page->mapping != mapping) {
602 wait_on_page_writeback(page);
603 if (page_mapped(page)) {
604 if (!did_range_unmap) {
606 * Zap the rest of the file in one hit.
608 unmap_mapping_range(mapping,
609 (loff_t)index << PAGE_CACHE_SHIFT,
610 (loff_t)(1 + end - index)
618 unmap_mapping_range(mapping,
619 (loff_t)index << PAGE_CACHE_SHIFT,
623 BUG_ON(page_mapped(page));
624 ret2 = do_launder_page(mapping, page);
626 if (!invalidate_complete_page2(mapping, page))
633 pagevec_remove_exceptionals(&pvec);
634 pagevec_release(&pvec);
638 cleancache_invalidate_inode(mapping);
641 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
644 * invalidate_inode_pages2 - remove all pages from an address_space
645 * @mapping: the address_space
647 * Any pages which are found to be mapped into pagetables are unmapped prior to
650 * Returns -EBUSY if any pages could not be invalidated.
652 int invalidate_inode_pages2(struct address_space *mapping)
654 return invalidate_inode_pages2_range(mapping, 0, -1);
656 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
659 * truncate_pagecache - unmap and remove pagecache that has been truncated
661 * @newsize: new file size
663 * inode's new i_size must already be written before truncate_pagecache
666 * This function should typically be called before the filesystem
667 * releases resources associated with the freed range (eg. deallocates
668 * blocks). This way, pagecache will always stay logically coherent
669 * with on-disk format, and the filesystem would not have to deal with
670 * situations such as writepage being called for a page that has already
671 * had its underlying blocks deallocated.
673 void truncate_pagecache(struct inode *inode, loff_t newsize)
675 struct address_space *mapping = inode->i_mapping;
676 loff_t holebegin = round_up(newsize, PAGE_SIZE);
679 * unmap_mapping_range is called twice, first simply for
680 * efficiency so that truncate_inode_pages does fewer
681 * single-page unmaps. However after this first call, and
682 * before truncate_inode_pages finishes, it is possible for
683 * private pages to be COWed, which remain after
684 * truncate_inode_pages finishes, hence the second
685 * unmap_mapping_range call must be made for correctness.
687 unmap_mapping_range(mapping, holebegin, 0, 1);
688 truncate_inode_pages(mapping, newsize);
689 unmap_mapping_range(mapping, holebegin, 0, 1);
691 EXPORT_SYMBOL(truncate_pagecache);
694 * truncate_setsize - update inode and pagecache for a new file size
696 * @newsize: new file size
698 * truncate_setsize updates i_size and performs pagecache truncation (if
699 * necessary) to @newsize. It will be typically be called from the filesystem's
700 * setattr function when ATTR_SIZE is passed in.
702 * Must be called with a lock serializing truncates and writes (generally
703 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
704 * specific block truncation has been performed.
706 void truncate_setsize(struct inode *inode, loff_t newsize)
708 loff_t oldsize = inode->i_size;
710 i_size_write(inode, newsize);
711 if (newsize > oldsize)
712 pagecache_isize_extended(inode, oldsize, newsize);
713 truncate_pagecache(inode, newsize);
715 EXPORT_SYMBOL(truncate_setsize);
718 * pagecache_isize_extended - update pagecache after extension of i_size
719 * @inode: inode for which i_size was extended
720 * @from: original inode size
721 * @to: new inode size
723 * Handle extension of inode size either caused by extending truncate or by
724 * write starting after current i_size. We mark the page straddling current
725 * i_size RO so that page_mkwrite() is called on the nearest write access to
726 * the page. This way filesystem can be sure that page_mkwrite() is called on
727 * the page before user writes to the page via mmap after the i_size has been
730 * The function must be called after i_size is updated so that page fault
731 * coming after we unlock the page will already see the new i_size.
732 * The function must be called while we still hold i_mutex - this not only
733 * makes sure i_size is stable but also that userspace cannot observe new
734 * i_size value before we are prepared to store mmap writes at new inode size.
736 void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
738 int bsize = 1 << inode->i_blkbits;
743 WARN_ON(to > inode->i_size);
745 if (from >= to || bsize == PAGE_CACHE_SIZE)
747 /* Page straddling @from will not have any hole block created? */
748 rounded_from = round_up(from, bsize);
749 if (to <= rounded_from || !(rounded_from & (PAGE_CACHE_SIZE - 1)))
752 index = from >> PAGE_CACHE_SHIFT;
753 page = find_lock_page(inode->i_mapping, index);
754 /* Page not cached? Nothing to do */
758 * See clear_page_dirty_for_io() for details why set_page_dirty()
761 if (page_mkclean(page))
762 set_page_dirty(page);
764 page_cache_release(page);
766 EXPORT_SYMBOL(pagecache_isize_extended);
769 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
771 * @lstart: offset of beginning of hole
772 * @lend: offset of last byte of hole
774 * This function should typically be called before the filesystem
775 * releases resources associated with the freed range (eg. deallocates
776 * blocks). This way, pagecache will always stay logically coherent
777 * with on-disk format, and the filesystem would not have to deal with
778 * situations such as writepage being called for a page that has already
779 * had its underlying blocks deallocated.
781 void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
783 struct address_space *mapping = inode->i_mapping;
784 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
785 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
787 * This rounding is currently just for example: unmap_mapping_range
788 * expands its hole outwards, whereas we want it to contract the hole
789 * inwards. However, existing callers of truncate_pagecache_range are
790 * doing their own page rounding first. Note that unmap_mapping_range
791 * allows holelen 0 for all, and we allow lend -1 for end of file.
795 * Unlike in truncate_pagecache, unmap_mapping_range is called only
796 * once (before truncating pagecache), and without "even_cows" flag:
797 * hole-punching should not remove private COWed pages from the hole.
799 if ((u64)unmap_end > (u64)unmap_start)
800 unmap_mapping_range(mapping, unmap_start,
801 1 + unmap_end - unmap_start, 0);
802 truncate_inode_pages_range(mapping, lstart, lend);
804 EXPORT_SYMBOL(truncate_pagecache_range);