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/dax.h>
13 #include <linux/gfp.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,
23 #include <linux/shmem_fs.h>
24 #include <linux/cleancache.h>
25 #include <linux/rmap.h>
28 static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
31 struct radix_tree_node *node;
34 spin_lock_irq(&mapping->tree_lock);
36 * Regular page slots are stabilized by the page lock even
37 * without the tree itself locked. These unlocked entries
38 * need verification under the tree lock.
40 if (!__radix_tree_lookup(&mapping->page_tree, index, &node, &slot))
44 __radix_tree_replace(&mapping->page_tree, node, slot, NULL,
45 workingset_update_node, mapping);
46 mapping->nrexceptional--;
48 spin_unlock_irq(&mapping->tree_lock);
52 * Unconditionally remove exceptional entry. Usually called from truncate path.
54 static void truncate_exceptional_entry(struct address_space *mapping,
55 pgoff_t index, void *entry)
57 /* Handled by shmem itself */
58 if (shmem_mapping(mapping))
61 if (dax_mapping(mapping)) {
62 dax_delete_mapping_entry(mapping, index);
65 clear_shadow_entry(mapping, index, entry);
69 * Invalidate exceptional entry if easily possible. This handles exceptional
70 * entries for invalidate_inode_pages() so for DAX it evicts only unlocked and
73 static int invalidate_exceptional_entry(struct address_space *mapping,
74 pgoff_t index, void *entry)
76 /* Handled by shmem itself */
77 if (shmem_mapping(mapping))
79 if (dax_mapping(mapping))
80 return dax_invalidate_mapping_entry(mapping, index);
81 clear_shadow_entry(mapping, index, entry);
86 * Invalidate exceptional entry if clean. This handles exceptional entries for
87 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
89 static int invalidate_exceptional_entry2(struct address_space *mapping,
90 pgoff_t index, void *entry)
92 /* Handled by shmem itself */
93 if (shmem_mapping(mapping))
95 if (dax_mapping(mapping))
96 return dax_invalidate_mapping_entry_sync(mapping, index);
97 clear_shadow_entry(mapping, index, entry);
102 * do_invalidatepage - invalidate part or all of a page
103 * @page: the page which is affected
104 * @offset: start of the range to invalidate
105 * @length: length of the range to invalidate
107 * do_invalidatepage() is called when all or part of the page has become
108 * invalidated by a truncate operation.
110 * do_invalidatepage() does not have to release all buffers, but it must
111 * ensure that no dirty buffer is left outside @offset and that no I/O
112 * is underway against any of the blocks which are outside the truncation
113 * point. Because the caller is about to free (and possibly reuse) those
116 void do_invalidatepage(struct page *page, unsigned int offset,
119 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
121 invalidatepage = page->mapping->a_ops->invalidatepage;
124 invalidatepage = block_invalidatepage;
127 (*invalidatepage)(page, offset, length);
131 * If truncate cannot remove the fs-private metadata from the page, the page
132 * becomes orphaned. It will be left on the LRU and may even be mapped into
133 * user pagetables if we're racing with filemap_fault().
135 * We need to bale out if page->mapping is no longer equal to the original
136 * mapping. This happens a) when the VM reclaimed the page while we waited on
137 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
138 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
141 truncate_complete_page(struct address_space *mapping, struct page *page)
143 if (page->mapping != mapping)
146 if (page_has_private(page))
147 do_invalidatepage(page, 0, PAGE_SIZE);
150 * Some filesystems seem to re-dirty the page even after
151 * the VM has canceled the dirty bit (eg ext3 journaling).
152 * Hence dirty accounting check is placed after invalidation.
154 cancel_dirty_page(page);
155 ClearPageMappedToDisk(page);
156 delete_from_page_cache(page);
161 * This is for invalidate_mapping_pages(). That function can be called at
162 * any time, and is not supposed to throw away dirty pages. But pages can
163 * be marked dirty at any time too, so use remove_mapping which safely
164 * discards clean, unused pages.
166 * Returns non-zero if the page was successfully invalidated.
169 invalidate_complete_page(struct address_space *mapping, struct page *page)
173 if (page->mapping != mapping)
176 if (page_has_private(page) && !try_to_release_page(page, 0))
179 ret = remove_mapping(mapping, page);
184 int truncate_inode_page(struct address_space *mapping, struct page *page)
187 VM_BUG_ON_PAGE(PageTail(page), page);
189 holelen = PageTransHuge(page) ? HPAGE_PMD_SIZE : PAGE_SIZE;
190 if (page_mapped(page)) {
191 unmap_mapping_range(mapping,
192 (loff_t)page->index << PAGE_SHIFT,
195 return truncate_complete_page(mapping, page);
199 * Used to get rid of pages on hardware memory corruption.
201 int generic_error_remove_page(struct address_space *mapping, struct page *page)
206 * Only punch for normal data pages for now.
207 * Handling other types like directories would need more auditing.
209 if (!S_ISREG(mapping->host->i_mode))
211 return truncate_inode_page(mapping, page);
213 EXPORT_SYMBOL(generic_error_remove_page);
216 * Safely invalidate one page from its pagecache mapping.
217 * It only drops clean, unused pages. The page must be locked.
219 * Returns 1 if the page is successfully invalidated, otherwise 0.
221 int invalidate_inode_page(struct page *page)
223 struct address_space *mapping = page_mapping(page);
226 if (PageDirty(page) || PageWriteback(page))
228 if (page_mapped(page))
230 return invalidate_complete_page(mapping, page);
234 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
235 * @mapping: mapping to truncate
236 * @lstart: offset from which to truncate
237 * @lend: offset to which to truncate (inclusive)
239 * Truncate the page cache, removing the pages that are between
240 * specified offsets (and zeroing out partial pages
241 * if lstart or lend + 1 is not page aligned).
243 * Truncate takes two passes - the first pass is nonblocking. It will not
244 * block on page locks and it will not block on writeback. The second pass
245 * will wait. This is to prevent as much IO as possible in the affected region.
246 * The first pass will remove most pages, so the search cost of the second pass
249 * We pass down the cache-hot hint to the page freeing code. Even if the
250 * mapping is large, it is probably the case that the final pages are the most
251 * recently touched, and freeing happens in ascending file offset order.
253 * Note that since ->invalidatepage() accepts range to invalidate
254 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
255 * page aligned properly.
257 void truncate_inode_pages_range(struct address_space *mapping,
258 loff_t lstart, loff_t lend)
260 pgoff_t start; /* inclusive */
261 pgoff_t end; /* exclusive */
262 unsigned int partial_start; /* inclusive */
263 unsigned int partial_end; /* exclusive */
265 pgoff_t indices[PAGEVEC_SIZE];
269 cleancache_invalidate_inode(mapping);
270 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
273 /* Offsets within partial pages */
274 partial_start = lstart & (PAGE_SIZE - 1);
275 partial_end = (lend + 1) & (PAGE_SIZE - 1);
278 * 'start' and 'end' always covers the range of pages to be fully
279 * truncated. Partial pages are covered with 'partial_start' at the
280 * start of the range and 'partial_end' at the end of the range.
281 * Note that 'end' is exclusive while 'lend' is inclusive.
283 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
286 * lend == -1 indicates end-of-file so we have to set 'end'
287 * to the highest possible pgoff_t and since the type is
288 * unsigned we're using -1.
292 end = (lend + 1) >> PAGE_SHIFT;
294 pagevec_init(&pvec, 0);
296 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
297 min(end - index, (pgoff_t)PAGEVEC_SIZE),
299 for (i = 0; i < pagevec_count(&pvec); i++) {
300 struct page *page = pvec.pages[i];
302 /* We rely upon deletion not changing page->index */
307 if (radix_tree_exceptional_entry(page)) {
308 truncate_exceptional_entry(mapping, index,
313 if (!trylock_page(page))
315 WARN_ON(page_to_index(page) != index);
316 if (PageWriteback(page)) {
320 truncate_inode_page(mapping, page);
323 pagevec_remove_exceptionals(&pvec);
324 pagevec_release(&pvec);
330 struct page *page = find_lock_page(mapping, start - 1);
332 unsigned int top = PAGE_SIZE;
334 /* Truncation within a single page */
338 wait_on_page_writeback(page);
339 zero_user_segment(page, partial_start, top);
340 cleancache_invalidate_page(mapping, page);
341 if (page_has_private(page))
342 do_invalidatepage(page, partial_start,
343 top - partial_start);
349 struct page *page = find_lock_page(mapping, end);
351 wait_on_page_writeback(page);
352 zero_user_segment(page, 0, partial_end);
353 cleancache_invalidate_page(mapping, page);
354 if (page_has_private(page))
355 do_invalidatepage(page, 0,
362 * If the truncation happened within a single page no pages
363 * will be released, just zeroed, so we can bail out now.
371 if (!pagevec_lookup_entries(&pvec, mapping, index,
372 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
373 /* If all gone from start onwards, we're done */
376 /* Otherwise restart to make sure all gone */
380 if (index == start && indices[0] >= end) {
381 /* All gone out of hole to be punched, we're done */
382 pagevec_remove_exceptionals(&pvec);
383 pagevec_release(&pvec);
386 for (i = 0; i < pagevec_count(&pvec); i++) {
387 struct page *page = pvec.pages[i];
389 /* We rely upon deletion not changing page->index */
392 /* Restart punch to make sure all gone */
397 if (radix_tree_exceptional_entry(page)) {
398 truncate_exceptional_entry(mapping, index,
404 WARN_ON(page_to_index(page) != index);
405 wait_on_page_writeback(page);
406 truncate_inode_page(mapping, page);
409 pagevec_remove_exceptionals(&pvec);
410 pagevec_release(&pvec);
413 cleancache_invalidate_inode(mapping);
415 EXPORT_SYMBOL(truncate_inode_pages_range);
418 * truncate_inode_pages - truncate *all* the pages from an offset
419 * @mapping: mapping to truncate
420 * @lstart: offset from which to truncate
422 * Called under (and serialised by) inode->i_mutex.
424 * Note: When this function returns, there can be a page in the process of
425 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
426 * mapping->nrpages can be non-zero when this function returns even after
427 * truncation of the whole mapping.
429 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
431 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
433 EXPORT_SYMBOL(truncate_inode_pages);
436 * truncate_inode_pages_final - truncate *all* pages before inode dies
437 * @mapping: mapping to truncate
439 * Called under (and serialized by) inode->i_mutex.
441 * Filesystems have to use this in the .evict_inode path to inform the
442 * VM that this is the final truncate and the inode is going away.
444 void truncate_inode_pages_final(struct address_space *mapping)
446 unsigned long nrexceptional;
447 unsigned long nrpages;
450 * Page reclaim can not participate in regular inode lifetime
451 * management (can't call iput()) and thus can race with the
452 * inode teardown. Tell it when the address space is exiting,
453 * so that it does not install eviction information after the
454 * final truncate has begun.
456 mapping_set_exiting(mapping);
459 * When reclaim installs eviction entries, it increases
460 * nrexceptional first, then decreases nrpages. Make sure we see
461 * this in the right order or we might miss an entry.
463 nrpages = mapping->nrpages;
465 nrexceptional = mapping->nrexceptional;
467 if (nrpages || nrexceptional) {
469 * As truncation uses a lockless tree lookup, cycle
470 * the tree lock to make sure any ongoing tree
471 * modification that does not see AS_EXITING is
472 * completed before starting the final truncate.
474 spin_lock_irq(&mapping->tree_lock);
475 spin_unlock_irq(&mapping->tree_lock);
477 truncate_inode_pages(mapping, 0);
480 EXPORT_SYMBOL(truncate_inode_pages_final);
483 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
484 * @mapping: the address_space which holds the pages to invalidate
485 * @start: the offset 'from' which to invalidate
486 * @end: the offset 'to' which to invalidate (inclusive)
488 * This function only removes the unlocked pages, if you want to
489 * remove all the pages of one inode, you must call truncate_inode_pages.
491 * invalidate_mapping_pages() will not block on IO activity. It will not
492 * invalidate pages which are dirty, locked, under writeback or mapped into
495 unsigned long invalidate_mapping_pages(struct address_space *mapping,
496 pgoff_t start, pgoff_t end)
498 pgoff_t indices[PAGEVEC_SIZE];
500 pgoff_t index = start;
502 unsigned long count = 0;
505 pagevec_init(&pvec, 0);
506 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
507 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
509 for (i = 0; i < pagevec_count(&pvec); i++) {
510 struct page *page = pvec.pages[i];
512 /* We rely upon deletion not changing page->index */
517 if (radix_tree_exceptional_entry(page)) {
518 invalidate_exceptional_entry(mapping, index,
523 if (!trylock_page(page))
526 WARN_ON(page_to_index(page) != index);
528 /* Middle of THP: skip */
529 if (PageTransTail(page)) {
532 } else if (PageTransHuge(page)) {
533 index += HPAGE_PMD_NR - 1;
534 i += HPAGE_PMD_NR - 1;
535 /* 'end' is in the middle of THP */
536 if (index == round_down(end, HPAGE_PMD_NR))
540 ret = invalidate_inode_page(page);
543 * Invalidation is a hint that the page is no longer
544 * of interest and try to speed up its reclaim.
547 deactivate_file_page(page);
550 pagevec_remove_exceptionals(&pvec);
551 pagevec_release(&pvec);
557 EXPORT_SYMBOL(invalidate_mapping_pages);
560 * This is like invalidate_complete_page(), except it ignores the page's
561 * refcount. We do this because invalidate_inode_pages2() needs stronger
562 * invalidation guarantees, and cannot afford to leave pages behind because
563 * shrink_page_list() has a temp ref on them, or because they're transiently
564 * sitting in the lru_cache_add() pagevecs.
567 invalidate_complete_page2(struct address_space *mapping, struct page *page)
571 if (page->mapping != mapping)
574 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
577 spin_lock_irqsave(&mapping->tree_lock, flags);
581 BUG_ON(page_has_private(page));
582 __delete_from_page_cache(page, NULL);
583 spin_unlock_irqrestore(&mapping->tree_lock, flags);
585 if (mapping->a_ops->freepage)
586 mapping->a_ops->freepage(page);
588 put_page(page); /* pagecache ref */
591 spin_unlock_irqrestore(&mapping->tree_lock, flags);
595 static int do_launder_page(struct address_space *mapping, struct page *page)
597 if (!PageDirty(page))
599 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
601 return mapping->a_ops->launder_page(page);
605 * invalidate_inode_pages2_range - remove range of pages from an address_space
606 * @mapping: the address_space
607 * @start: the page offset 'from' which to invalidate
608 * @end: the page offset 'to' which to invalidate (inclusive)
610 * Any pages which are found to be mapped into pagetables are unmapped prior to
613 * Returns -EBUSY if any pages could not be invalidated.
615 int invalidate_inode_pages2_range(struct address_space *mapping,
616 pgoff_t start, pgoff_t end)
618 pgoff_t indices[PAGEVEC_SIZE];
624 int did_range_unmap = 0;
626 cleancache_invalidate_inode(mapping);
627 pagevec_init(&pvec, 0);
629 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
630 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
632 for (i = 0; i < pagevec_count(&pvec); i++) {
633 struct page *page = pvec.pages[i];
635 /* We rely upon deletion not changing page->index */
640 if (radix_tree_exceptional_entry(page)) {
641 if (!invalidate_exceptional_entry2(mapping,
648 WARN_ON(page_to_index(page) != index);
649 if (page->mapping != mapping) {
653 wait_on_page_writeback(page);
654 if (page_mapped(page)) {
655 if (!did_range_unmap) {
657 * Zap the rest of the file in one hit.
659 unmap_mapping_range(mapping,
660 (loff_t)index << PAGE_SHIFT,
661 (loff_t)(1 + end - index)
669 unmap_mapping_range(mapping,
670 (loff_t)index << PAGE_SHIFT,
674 BUG_ON(page_mapped(page));
675 ret2 = do_launder_page(mapping, page);
677 if (!invalidate_complete_page2(mapping, page))
684 pagevec_remove_exceptionals(&pvec);
685 pagevec_release(&pvec);
689 cleancache_invalidate_inode(mapping);
692 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
695 * invalidate_inode_pages2 - remove all pages from an address_space
696 * @mapping: the address_space
698 * Any pages which are found to be mapped into pagetables are unmapped prior to
701 * Returns -EBUSY if any pages could not be invalidated.
703 int invalidate_inode_pages2(struct address_space *mapping)
705 return invalidate_inode_pages2_range(mapping, 0, -1);
707 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
710 * truncate_pagecache - unmap and remove pagecache that has been truncated
712 * @newsize: new file size
714 * inode's new i_size must already be written before truncate_pagecache
717 * This function should typically be called before the filesystem
718 * releases resources associated with the freed range (eg. deallocates
719 * blocks). This way, pagecache will always stay logically coherent
720 * with on-disk format, and the filesystem would not have to deal with
721 * situations such as writepage being called for a page that has already
722 * had its underlying blocks deallocated.
724 void truncate_pagecache(struct inode *inode, loff_t newsize)
726 struct address_space *mapping = inode->i_mapping;
727 loff_t holebegin = round_up(newsize, PAGE_SIZE);
730 * unmap_mapping_range is called twice, first simply for
731 * efficiency so that truncate_inode_pages does fewer
732 * single-page unmaps. However after this first call, and
733 * before truncate_inode_pages finishes, it is possible for
734 * private pages to be COWed, which remain after
735 * truncate_inode_pages finishes, hence the second
736 * unmap_mapping_range call must be made for correctness.
738 unmap_mapping_range(mapping, holebegin, 0, 1);
739 truncate_inode_pages(mapping, newsize);
740 unmap_mapping_range(mapping, holebegin, 0, 1);
742 EXPORT_SYMBOL(truncate_pagecache);
745 * truncate_setsize - update inode and pagecache for a new file size
747 * @newsize: new file size
749 * truncate_setsize updates i_size and performs pagecache truncation (if
750 * necessary) to @newsize. It will be typically be called from the filesystem's
751 * setattr function when ATTR_SIZE is passed in.
753 * Must be called with a lock serializing truncates and writes (generally
754 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
755 * specific block truncation has been performed.
757 void truncate_setsize(struct inode *inode, loff_t newsize)
759 loff_t oldsize = inode->i_size;
761 i_size_write(inode, newsize);
762 if (newsize > oldsize)
763 pagecache_isize_extended(inode, oldsize, newsize);
764 truncate_pagecache(inode, newsize);
766 EXPORT_SYMBOL(truncate_setsize);
769 * pagecache_isize_extended - update pagecache after extension of i_size
770 * @inode: inode for which i_size was extended
771 * @from: original inode size
772 * @to: new inode size
774 * Handle extension of inode size either caused by extending truncate or by
775 * write starting after current i_size. We mark the page straddling current
776 * i_size RO so that page_mkwrite() is called on the nearest write access to
777 * the page. This way filesystem can be sure that page_mkwrite() is called on
778 * the page before user writes to the page via mmap after the i_size has been
781 * The function must be called after i_size is updated so that page fault
782 * coming after we unlock the page will already see the new i_size.
783 * The function must be called while we still hold i_mutex - this not only
784 * makes sure i_size is stable but also that userspace cannot observe new
785 * i_size value before we are prepared to store mmap writes at new inode size.
787 void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
789 int bsize = i_blocksize(inode);
794 WARN_ON(to > inode->i_size);
796 if (from >= to || bsize == PAGE_SIZE)
798 /* Page straddling @from will not have any hole block created? */
799 rounded_from = round_up(from, bsize);
800 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
803 index = from >> PAGE_SHIFT;
804 page = find_lock_page(inode->i_mapping, index);
805 /* Page not cached? Nothing to do */
809 * See clear_page_dirty_for_io() for details why set_page_dirty()
812 if (page_mkclean(page))
813 set_page_dirty(page);
817 EXPORT_SYMBOL(pagecache_isize_extended);
820 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
822 * @lstart: offset of beginning of hole
823 * @lend: offset of last byte of hole
825 * This function should typically be called before the filesystem
826 * releases resources associated with the freed range (eg. deallocates
827 * blocks). This way, pagecache will always stay logically coherent
828 * with on-disk format, and the filesystem would not have to deal with
829 * situations such as writepage being called for a page that has already
830 * had its underlying blocks deallocated.
832 void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
834 struct address_space *mapping = inode->i_mapping;
835 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
836 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
838 * This rounding is currently just for example: unmap_mapping_range
839 * expands its hole outwards, whereas we want it to contract the hole
840 * inwards. However, existing callers of truncate_pagecache_range are
841 * doing their own page rounding first. Note that unmap_mapping_range
842 * allows holelen 0 for all, and we allow lend -1 for end of file.
846 * Unlike in truncate_pagecache, unmap_mapping_range is called only
847 * once (before truncating pagecache), and without "even_cows" flag:
848 * hole-punching should not remove private COWed pages from the hole.
850 if ((u64)unmap_end > (u64)unmap_start)
851 unmap_mapping_range(mapping, unmap_start,
852 1 + unmap_end - unmap_start, 0);
853 truncate_inode_pages_range(mapping, lstart, lend);
855 EXPORT_SYMBOL(truncate_pagecache_range);