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/cleancache.h>
24 #include <linux/rmap.h>
27 static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
30 struct radix_tree_node *node;
33 spin_lock_irq(&mapping->tree_lock);
35 * Regular page slots are stabilized by the page lock even
36 * without the tree itself locked. These unlocked entries
37 * need verification under the tree lock.
39 if (!__radix_tree_lookup(&mapping->page_tree, index, &node, &slot))
43 __radix_tree_replace(&mapping->page_tree, node, slot, NULL,
44 workingset_update_node, mapping);
45 mapping->nrexceptional--;
47 spin_unlock_irq(&mapping->tree_lock);
51 * Unconditionally remove exceptional entry. Usually called from truncate path.
53 static void truncate_exceptional_entry(struct address_space *mapping,
54 pgoff_t index, void *entry)
56 /* Handled by shmem itself */
57 if (shmem_mapping(mapping))
60 if (dax_mapping(mapping)) {
61 dax_delete_mapping_entry(mapping, index);
64 clear_shadow_entry(mapping, index, entry);
68 * Invalidate exceptional entry if easily possible. This handles exceptional
69 * entries for invalidate_inode_pages() so for DAX it evicts only unlocked and
72 static int invalidate_exceptional_entry(struct address_space *mapping,
73 pgoff_t index, void *entry)
75 /* Handled by shmem itself */
76 if (shmem_mapping(mapping))
78 if (dax_mapping(mapping))
79 return dax_invalidate_mapping_entry(mapping, index);
80 clear_shadow_entry(mapping, index, entry);
85 * Invalidate exceptional entry if clean. This handles exceptional entries for
86 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
88 static int invalidate_exceptional_entry2(struct address_space *mapping,
89 pgoff_t index, void *entry)
91 /* Handled by shmem itself */
92 if (shmem_mapping(mapping))
94 if (dax_mapping(mapping))
95 return dax_invalidate_mapping_entry_sync(mapping, index);
96 clear_shadow_entry(mapping, index, entry);
101 * do_invalidatepage - invalidate part or all of a page
102 * @page: the page which is affected
103 * @offset: start of the range to invalidate
104 * @length: length of the range to invalidate
106 * do_invalidatepage() is called when all or part of the page has become
107 * invalidated by a truncate operation.
109 * do_invalidatepage() does not have to release all buffers, but it must
110 * ensure that no dirty buffer is left outside @offset and that no I/O
111 * is underway against any of the blocks which are outside the truncation
112 * point. Because the caller is about to free (and possibly reuse) those
115 void do_invalidatepage(struct page *page, unsigned int offset,
118 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
120 invalidatepage = page->mapping->a_ops->invalidatepage;
123 invalidatepage = block_invalidatepage;
126 (*invalidatepage)(page, offset, length);
130 * If truncate cannot remove the fs-private metadata from the page, the page
131 * becomes orphaned. It will be left on the LRU and may even be mapped into
132 * user pagetables if we're racing with filemap_fault().
134 * We need to bale out if page->mapping is no longer equal to the original
135 * mapping. This happens a) when the VM reclaimed the page while we waited on
136 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
137 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
140 truncate_complete_page(struct address_space *mapping, struct page *page)
142 if (page->mapping != mapping)
145 if (page_has_private(page))
146 do_invalidatepage(page, 0, PAGE_SIZE);
149 * Some filesystems seem to re-dirty the page even after
150 * the VM has canceled the dirty bit (eg ext3 journaling).
151 * Hence dirty accounting check is placed after invalidation.
153 cancel_dirty_page(page);
154 ClearPageMappedToDisk(page);
155 delete_from_page_cache(page);
160 * This is for invalidate_mapping_pages(). That function can be called at
161 * any time, and is not supposed to throw away dirty pages. But pages can
162 * be marked dirty at any time too, so use remove_mapping which safely
163 * discards clean, unused pages.
165 * Returns non-zero if the page was successfully invalidated.
168 invalidate_complete_page(struct address_space *mapping, struct page *page)
172 if (page->mapping != mapping)
175 if (page_has_private(page) && !try_to_release_page(page, 0))
178 ret = remove_mapping(mapping, page);
183 int truncate_inode_page(struct address_space *mapping, struct page *page)
186 VM_BUG_ON_PAGE(PageTail(page), page);
188 holelen = PageTransHuge(page) ? HPAGE_PMD_SIZE : PAGE_SIZE;
189 if (page_mapped(page)) {
190 unmap_mapping_range(mapping,
191 (loff_t)page->index << PAGE_SHIFT,
194 return truncate_complete_page(mapping, page);
198 * Used to get rid of pages on hardware memory corruption.
200 int generic_error_remove_page(struct address_space *mapping, struct page *page)
205 * Only punch for normal data pages for now.
206 * Handling other types like directories would need more auditing.
208 if (!S_ISREG(mapping->host->i_mode))
210 return truncate_inode_page(mapping, page);
212 EXPORT_SYMBOL(generic_error_remove_page);
215 * Safely invalidate one page from its pagecache mapping.
216 * It only drops clean, unused pages. The page must be locked.
218 * Returns 1 if the page is successfully invalidated, otherwise 0.
220 int invalidate_inode_page(struct page *page)
222 struct address_space *mapping = page_mapping(page);
225 if (PageDirty(page) || PageWriteback(page))
227 if (page_mapped(page))
229 return invalidate_complete_page(mapping, page);
233 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
234 * @mapping: mapping to truncate
235 * @lstart: offset from which to truncate
236 * @lend: offset to which to truncate (inclusive)
238 * Truncate the page cache, removing the pages that are between
239 * specified offsets (and zeroing out partial pages
240 * if lstart or lend + 1 is not page aligned).
242 * Truncate takes two passes - the first pass is nonblocking. It will not
243 * block on page locks and it will not block on writeback. The second pass
244 * will wait. This is to prevent as much IO as possible in the affected region.
245 * The first pass will remove most pages, so the search cost of the second pass
248 * We pass down the cache-hot hint to the page freeing code. Even if the
249 * mapping is large, it is probably the case that the final pages are the most
250 * recently touched, and freeing happens in ascending file offset order.
252 * Note that since ->invalidatepage() accepts range to invalidate
253 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
254 * page aligned properly.
256 void truncate_inode_pages_range(struct address_space *mapping,
257 loff_t lstart, loff_t lend)
259 pgoff_t start; /* inclusive */
260 pgoff_t end; /* exclusive */
261 unsigned int partial_start; /* inclusive */
262 unsigned int partial_end; /* exclusive */
264 pgoff_t indices[PAGEVEC_SIZE];
268 cleancache_invalidate_inode(mapping);
269 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
272 /* Offsets within partial pages */
273 partial_start = lstart & (PAGE_SIZE - 1);
274 partial_end = (lend + 1) & (PAGE_SIZE - 1);
277 * 'start' and 'end' always covers the range of pages to be fully
278 * truncated. Partial pages are covered with 'partial_start' at the
279 * start of the range and 'partial_end' at the end of the range.
280 * Note that 'end' is exclusive while 'lend' is inclusive.
282 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
285 * lend == -1 indicates end-of-file so we have to set 'end'
286 * to the highest possible pgoff_t and since the type is
287 * unsigned we're using -1.
291 end = (lend + 1) >> PAGE_SHIFT;
293 pagevec_init(&pvec, 0);
295 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
296 min(end - index, (pgoff_t)PAGEVEC_SIZE),
298 for (i = 0; i < pagevec_count(&pvec); i++) {
299 struct page *page = pvec.pages[i];
301 /* We rely upon deletion not changing page->index */
306 if (radix_tree_exceptional_entry(page)) {
307 truncate_exceptional_entry(mapping, index,
312 if (!trylock_page(page))
314 WARN_ON(page_to_index(page) != index);
315 if (PageWriteback(page)) {
319 truncate_inode_page(mapping, page);
322 pagevec_remove_exceptionals(&pvec);
323 pagevec_release(&pvec);
329 struct page *page = find_lock_page(mapping, start - 1);
331 unsigned int top = PAGE_SIZE;
333 /* Truncation within a single page */
337 wait_on_page_writeback(page);
338 zero_user_segment(page, partial_start, top);
339 cleancache_invalidate_page(mapping, page);
340 if (page_has_private(page))
341 do_invalidatepage(page, partial_start,
342 top - partial_start);
348 struct page *page = find_lock_page(mapping, end);
350 wait_on_page_writeback(page);
351 zero_user_segment(page, 0, partial_end);
352 cleancache_invalidate_page(mapping, page);
353 if (page_has_private(page))
354 do_invalidatepage(page, 0,
361 * If the truncation happened within a single page no pages
362 * will be released, just zeroed, so we can bail out now.
370 if (!pagevec_lookup_entries(&pvec, mapping, index,
371 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
372 /* If all gone from start onwards, we're done */
375 /* Otherwise restart to make sure all gone */
379 if (index == start && indices[0] >= end) {
380 /* All gone out of hole to be punched, we're done */
381 pagevec_remove_exceptionals(&pvec);
382 pagevec_release(&pvec);
385 for (i = 0; i < pagevec_count(&pvec); i++) {
386 struct page *page = pvec.pages[i];
388 /* We rely upon deletion not changing page->index */
391 /* Restart punch to make sure all gone */
396 if (radix_tree_exceptional_entry(page)) {
397 truncate_exceptional_entry(mapping, index,
403 WARN_ON(page_to_index(page) != index);
404 wait_on_page_writeback(page);
405 truncate_inode_page(mapping, page);
408 pagevec_remove_exceptionals(&pvec);
409 pagevec_release(&pvec);
412 cleancache_invalidate_inode(mapping);
414 EXPORT_SYMBOL(truncate_inode_pages_range);
417 * truncate_inode_pages - truncate *all* the pages from an offset
418 * @mapping: mapping to truncate
419 * @lstart: offset from which to truncate
421 * Called under (and serialised by) inode->i_mutex.
423 * Note: When this function returns, there can be a page in the process of
424 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
425 * mapping->nrpages can be non-zero when this function returns even after
426 * truncation of the whole mapping.
428 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
430 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
432 EXPORT_SYMBOL(truncate_inode_pages);
435 * truncate_inode_pages_final - truncate *all* pages before inode dies
436 * @mapping: mapping to truncate
438 * Called under (and serialized by) inode->i_mutex.
440 * Filesystems have to use this in the .evict_inode path to inform the
441 * VM that this is the final truncate and the inode is going away.
443 void truncate_inode_pages_final(struct address_space *mapping)
445 unsigned long nrexceptional;
446 unsigned long nrpages;
449 * Page reclaim can not participate in regular inode lifetime
450 * management (can't call iput()) and thus can race with the
451 * inode teardown. Tell it when the address space is exiting,
452 * so that it does not install eviction information after the
453 * final truncate has begun.
455 mapping_set_exiting(mapping);
458 * When reclaim installs eviction entries, it increases
459 * nrexceptional first, then decreases nrpages. Make sure we see
460 * this in the right order or we might miss an entry.
462 nrpages = mapping->nrpages;
464 nrexceptional = mapping->nrexceptional;
466 if (nrpages || nrexceptional) {
468 * As truncation uses a lockless tree lookup, cycle
469 * the tree lock to make sure any ongoing tree
470 * modification that does not see AS_EXITING is
471 * completed before starting the final truncate.
473 spin_lock_irq(&mapping->tree_lock);
474 spin_unlock_irq(&mapping->tree_lock);
476 truncate_inode_pages(mapping, 0);
479 EXPORT_SYMBOL(truncate_inode_pages_final);
482 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
483 * @mapping: the address_space which holds the pages to invalidate
484 * @start: the offset 'from' which to invalidate
485 * @end: the offset 'to' which to invalidate (inclusive)
487 * This function only removes the unlocked pages, if you want to
488 * remove all the pages of one inode, you must call truncate_inode_pages.
490 * invalidate_mapping_pages() will not block on IO activity. It will not
491 * invalidate pages which are dirty, locked, under writeback or mapped into
494 unsigned long invalidate_mapping_pages(struct address_space *mapping,
495 pgoff_t start, pgoff_t end)
497 pgoff_t indices[PAGEVEC_SIZE];
499 pgoff_t index = start;
501 unsigned long count = 0;
504 pagevec_init(&pvec, 0);
505 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
506 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
508 for (i = 0; i < pagevec_count(&pvec); i++) {
509 struct page *page = pvec.pages[i];
511 /* We rely upon deletion not changing page->index */
516 if (radix_tree_exceptional_entry(page)) {
517 invalidate_exceptional_entry(mapping, index,
522 if (!trylock_page(page))
525 WARN_ON(page_to_index(page) != index);
527 /* Middle of THP: skip */
528 if (PageTransTail(page)) {
531 } else if (PageTransHuge(page)) {
532 index += HPAGE_PMD_NR - 1;
533 i += HPAGE_PMD_NR - 1;
534 /* 'end' is in the middle of THP */
535 if (index == round_down(end, HPAGE_PMD_NR))
539 ret = invalidate_inode_page(page);
542 * Invalidation is a hint that the page is no longer
543 * of interest and try to speed up its reclaim.
546 deactivate_file_page(page);
549 pagevec_remove_exceptionals(&pvec);
550 pagevec_release(&pvec);
556 EXPORT_SYMBOL(invalidate_mapping_pages);
559 * This is like invalidate_complete_page(), except it ignores the page's
560 * refcount. We do this because invalidate_inode_pages2() needs stronger
561 * invalidation guarantees, and cannot afford to leave pages behind because
562 * shrink_page_list() has a temp ref on them, or because they're transiently
563 * sitting in the lru_cache_add() pagevecs.
566 invalidate_complete_page2(struct address_space *mapping, struct page *page)
570 if (page->mapping != mapping)
573 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
576 spin_lock_irqsave(&mapping->tree_lock, flags);
580 BUG_ON(page_has_private(page));
581 __delete_from_page_cache(page, NULL);
582 spin_unlock_irqrestore(&mapping->tree_lock, flags);
584 if (mapping->a_ops->freepage)
585 mapping->a_ops->freepage(page);
587 put_page(page); /* pagecache ref */
590 spin_unlock_irqrestore(&mapping->tree_lock, flags);
594 static int do_launder_page(struct address_space *mapping, struct page *page)
596 if (!PageDirty(page))
598 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
600 return mapping->a_ops->launder_page(page);
604 * invalidate_inode_pages2_range - remove range of pages from an address_space
605 * @mapping: the address_space
606 * @start: the page offset 'from' which to invalidate
607 * @end: the page offset 'to' which to invalidate (inclusive)
609 * Any pages which are found to be mapped into pagetables are unmapped prior to
612 * Returns -EBUSY if any pages could not be invalidated.
614 int invalidate_inode_pages2_range(struct address_space *mapping,
615 pgoff_t start, pgoff_t end)
617 pgoff_t indices[PAGEVEC_SIZE];
623 int did_range_unmap = 0;
625 cleancache_invalidate_inode(mapping);
626 pagevec_init(&pvec, 0);
628 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
629 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
631 for (i = 0; i < pagevec_count(&pvec); i++) {
632 struct page *page = pvec.pages[i];
634 /* We rely upon deletion not changing page->index */
639 if (radix_tree_exceptional_entry(page)) {
640 if (!invalidate_exceptional_entry2(mapping,
647 WARN_ON(page_to_index(page) != index);
648 if (page->mapping != mapping) {
652 wait_on_page_writeback(page);
653 if (page_mapped(page)) {
654 if (!did_range_unmap) {
656 * Zap the rest of the file in one hit.
658 unmap_mapping_range(mapping,
659 (loff_t)index << PAGE_SHIFT,
660 (loff_t)(1 + end - index)
668 unmap_mapping_range(mapping,
669 (loff_t)index << PAGE_SHIFT,
673 BUG_ON(page_mapped(page));
674 ret2 = do_launder_page(mapping, page);
676 if (!invalidate_complete_page2(mapping, page))
683 pagevec_remove_exceptionals(&pvec);
684 pagevec_release(&pvec);
688 cleancache_invalidate_inode(mapping);
691 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
694 * invalidate_inode_pages2 - remove all pages from an address_space
695 * @mapping: the address_space
697 * Any pages which are found to be mapped into pagetables are unmapped prior to
700 * Returns -EBUSY if any pages could not be invalidated.
702 int invalidate_inode_pages2(struct address_space *mapping)
704 return invalidate_inode_pages2_range(mapping, 0, -1);
706 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
709 * truncate_pagecache - unmap and remove pagecache that has been truncated
711 * @newsize: new file size
713 * inode's new i_size must already be written before truncate_pagecache
716 * This function should typically be called before the filesystem
717 * releases resources associated with the freed range (eg. deallocates
718 * blocks). This way, pagecache will always stay logically coherent
719 * with on-disk format, and the filesystem would not have to deal with
720 * situations such as writepage being called for a page that has already
721 * had its underlying blocks deallocated.
723 void truncate_pagecache(struct inode *inode, loff_t newsize)
725 struct address_space *mapping = inode->i_mapping;
726 loff_t holebegin = round_up(newsize, PAGE_SIZE);
729 * unmap_mapping_range is called twice, first simply for
730 * efficiency so that truncate_inode_pages does fewer
731 * single-page unmaps. However after this first call, and
732 * before truncate_inode_pages finishes, it is possible for
733 * private pages to be COWed, which remain after
734 * truncate_inode_pages finishes, hence the second
735 * unmap_mapping_range call must be made for correctness.
737 unmap_mapping_range(mapping, holebegin, 0, 1);
738 truncate_inode_pages(mapping, newsize);
739 unmap_mapping_range(mapping, holebegin, 0, 1);
741 EXPORT_SYMBOL(truncate_pagecache);
744 * truncate_setsize - update inode and pagecache for a new file size
746 * @newsize: new file size
748 * truncate_setsize updates i_size and performs pagecache truncation (if
749 * necessary) to @newsize. It will be typically be called from the filesystem's
750 * setattr function when ATTR_SIZE is passed in.
752 * Must be called with a lock serializing truncates and writes (generally
753 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
754 * specific block truncation has been performed.
756 void truncate_setsize(struct inode *inode, loff_t newsize)
758 loff_t oldsize = inode->i_size;
760 i_size_write(inode, newsize);
761 if (newsize > oldsize)
762 pagecache_isize_extended(inode, oldsize, newsize);
763 truncate_pagecache(inode, newsize);
765 EXPORT_SYMBOL(truncate_setsize);
768 * pagecache_isize_extended - update pagecache after extension of i_size
769 * @inode: inode for which i_size was extended
770 * @from: original inode size
771 * @to: new inode size
773 * Handle extension of inode size either caused by extending truncate or by
774 * write starting after current i_size. We mark the page straddling current
775 * i_size RO so that page_mkwrite() is called on the nearest write access to
776 * the page. This way filesystem can be sure that page_mkwrite() is called on
777 * the page before user writes to the page via mmap after the i_size has been
780 * The function must be called after i_size is updated so that page fault
781 * coming after we unlock the page will already see the new i_size.
782 * The function must be called while we still hold i_mutex - this not only
783 * makes sure i_size is stable but also that userspace cannot observe new
784 * i_size value before we are prepared to store mmap writes at new inode size.
786 void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
788 int bsize = 1 << inode->i_blkbits;
793 WARN_ON(to > inode->i_size);
795 if (from >= to || bsize == PAGE_SIZE)
797 /* Page straddling @from will not have any hole block created? */
798 rounded_from = round_up(from, bsize);
799 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
802 index = from >> PAGE_SHIFT;
803 page = find_lock_page(inode->i_mapping, index);
804 /* Page not cached? Nothing to do */
808 * See clear_page_dirty_for_io() for details why set_page_dirty()
811 if (page_mkclean(page))
812 set_page_dirty(page);
816 EXPORT_SYMBOL(pagecache_isize_extended);
819 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
821 * @lstart: offset of beginning of hole
822 * @lend: offset of last byte of hole
824 * This function should typically be called before the filesystem
825 * releases resources associated with the freed range (eg. deallocates
826 * blocks). This way, pagecache will always stay logically coherent
827 * with on-disk format, and the filesystem would not have to deal with
828 * situations such as writepage being called for a page that has already
829 * had its underlying blocks deallocated.
831 void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
833 struct address_space *mapping = inode->i_mapping;
834 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
835 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
837 * This rounding is currently just for example: unmap_mapping_range
838 * expands its hole outwards, whereas we want it to contract the hole
839 * inwards. However, existing callers of truncate_pagecache_range are
840 * doing their own page rounding first. Note that unmap_mapping_range
841 * allows holelen 0 for all, and we allow lend -1 for end of file.
845 * Unlike in truncate_pagecache, unmap_mapping_range is called only
846 * once (before truncating pagecache), and without "even_cows" flag:
847 * hole-punching should not remove private COWed pages from the hole.
849 if ((u64)unmap_end > (u64)unmap_start)
850 unmap_mapping_range(mapping, unmap_start,
851 1 + unmap_end - unmap_start, 0);
852 truncate_inode_pages_range(mapping, lstart, lend);
854 EXPORT_SYMBOL(truncate_pagecache_range);