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 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);
414 cleancache_invalidate_inode(mapping);
416 EXPORT_SYMBOL(truncate_inode_pages_range);
419 * truncate_inode_pages - truncate *all* the pages from an offset
420 * @mapping: mapping to truncate
421 * @lstart: offset from which to truncate
423 * Called under (and serialised by) inode->i_mutex.
425 * Note: When this function returns, there can be a page in the process of
426 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
427 * mapping->nrpages can be non-zero when this function returns even after
428 * truncation of the whole mapping.
430 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
432 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
434 EXPORT_SYMBOL(truncate_inode_pages);
437 * truncate_inode_pages_final - truncate *all* pages before inode dies
438 * @mapping: mapping to truncate
440 * Called under (and serialized by) inode->i_mutex.
442 * Filesystems have to use this in the .evict_inode path to inform the
443 * VM that this is the final truncate and the inode is going away.
445 void truncate_inode_pages_final(struct address_space *mapping)
447 unsigned long nrexceptional;
448 unsigned long nrpages;
451 * Page reclaim can not participate in regular inode lifetime
452 * management (can't call iput()) and thus can race with the
453 * inode teardown. Tell it when the address space is exiting,
454 * so that it does not install eviction information after the
455 * final truncate has begun.
457 mapping_set_exiting(mapping);
460 * When reclaim installs eviction entries, it increases
461 * nrexceptional first, then decreases nrpages. Make sure we see
462 * this in the right order or we might miss an entry.
464 nrpages = mapping->nrpages;
466 nrexceptional = mapping->nrexceptional;
468 if (nrpages || nrexceptional) {
470 * As truncation uses a lockless tree lookup, cycle
471 * the tree lock to make sure any ongoing tree
472 * modification that does not see AS_EXITING is
473 * completed before starting the final truncate.
475 spin_lock_irq(&mapping->tree_lock);
476 spin_unlock_irq(&mapping->tree_lock);
478 truncate_inode_pages(mapping, 0);
481 EXPORT_SYMBOL(truncate_inode_pages_final);
484 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
485 * @mapping: the address_space which holds the pages to invalidate
486 * @start: the offset 'from' which to invalidate
487 * @end: the offset 'to' which to invalidate (inclusive)
489 * This function only removes the unlocked pages, if you want to
490 * remove all the pages of one inode, you must call truncate_inode_pages.
492 * invalidate_mapping_pages() will not block on IO activity. It will not
493 * invalidate pages which are dirty, locked, under writeback or mapped into
496 unsigned long invalidate_mapping_pages(struct address_space *mapping,
497 pgoff_t start, pgoff_t end)
499 pgoff_t indices[PAGEVEC_SIZE];
501 pgoff_t index = start;
503 unsigned long count = 0;
506 pagevec_init(&pvec, 0);
507 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
508 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
510 for (i = 0; i < pagevec_count(&pvec); i++) {
511 struct page *page = pvec.pages[i];
513 /* We rely upon deletion not changing page->index */
518 if (radix_tree_exceptional_entry(page)) {
519 invalidate_exceptional_entry(mapping, index,
524 if (!trylock_page(page))
527 WARN_ON(page_to_index(page) != index);
529 /* Middle of THP: skip */
530 if (PageTransTail(page)) {
533 } else if (PageTransHuge(page)) {
534 index += HPAGE_PMD_NR - 1;
535 i += HPAGE_PMD_NR - 1;
536 /* 'end' is in the middle of THP */
537 if (index == round_down(end, HPAGE_PMD_NR))
541 ret = invalidate_inode_page(page);
544 * Invalidation is a hint that the page is no longer
545 * of interest and try to speed up its reclaim.
548 deactivate_file_page(page);
551 pagevec_remove_exceptionals(&pvec);
552 pagevec_release(&pvec);
558 EXPORT_SYMBOL(invalidate_mapping_pages);
561 * This is like invalidate_complete_page(), except it ignores the page's
562 * refcount. We do this because invalidate_inode_pages2() needs stronger
563 * invalidation guarantees, and cannot afford to leave pages behind because
564 * shrink_page_list() has a temp ref on them, or because they're transiently
565 * sitting in the lru_cache_add() pagevecs.
568 invalidate_complete_page2(struct address_space *mapping, struct page *page)
572 if (page->mapping != mapping)
575 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
578 spin_lock_irqsave(&mapping->tree_lock, flags);
582 BUG_ON(page_has_private(page));
583 __delete_from_page_cache(page, NULL);
584 spin_unlock_irqrestore(&mapping->tree_lock, flags);
586 if (mapping->a_ops->freepage)
587 mapping->a_ops->freepage(page);
589 put_page(page); /* pagecache ref */
592 spin_unlock_irqrestore(&mapping->tree_lock, flags);
596 static int do_launder_page(struct address_space *mapping, struct page *page)
598 if (!PageDirty(page))
600 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
602 return mapping->a_ops->launder_page(page);
606 * invalidate_inode_pages2_range - remove range of pages from an address_space
607 * @mapping: the address_space
608 * @start: the page offset 'from' which to invalidate
609 * @end: the page offset 'to' which to invalidate (inclusive)
611 * Any pages which are found to be mapped into pagetables are unmapped prior to
614 * Returns -EBUSY if any pages could not be invalidated.
616 int invalidate_inode_pages2_range(struct address_space *mapping,
617 pgoff_t start, pgoff_t end)
619 pgoff_t indices[PAGEVEC_SIZE];
625 int did_range_unmap = 0;
627 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
630 pagevec_init(&pvec, 0);
632 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
633 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
635 for (i = 0; i < pagevec_count(&pvec); i++) {
636 struct page *page = pvec.pages[i];
638 /* We rely upon deletion not changing page->index */
643 if (radix_tree_exceptional_entry(page)) {
644 if (!invalidate_exceptional_entry2(mapping,
651 WARN_ON(page_to_index(page) != index);
652 if (page->mapping != mapping) {
656 wait_on_page_writeback(page);
657 if (page_mapped(page)) {
658 if (!did_range_unmap) {
660 * Zap the rest of the file in one hit.
662 unmap_mapping_range(mapping,
663 (loff_t)index << PAGE_SHIFT,
664 (loff_t)(1 + end - index)
672 unmap_mapping_range(mapping,
673 (loff_t)index << PAGE_SHIFT,
677 BUG_ON(page_mapped(page));
678 ret2 = do_launder_page(mapping, page);
680 if (!invalidate_complete_page2(mapping, page))
687 pagevec_remove_exceptionals(&pvec);
688 pagevec_release(&pvec);
694 cleancache_invalidate_inode(mapping);
697 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
700 * invalidate_inode_pages2 - remove all pages from an address_space
701 * @mapping: the address_space
703 * Any pages which are found to be mapped into pagetables are unmapped prior to
706 * Returns -EBUSY if any pages could not be invalidated.
708 int invalidate_inode_pages2(struct address_space *mapping)
710 return invalidate_inode_pages2_range(mapping, 0, -1);
712 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
715 * truncate_pagecache - unmap and remove pagecache that has been truncated
717 * @newsize: new file size
719 * inode's new i_size must already be written before truncate_pagecache
722 * This function should typically be called before the filesystem
723 * releases resources associated with the freed range (eg. deallocates
724 * blocks). This way, pagecache will always stay logically coherent
725 * with on-disk format, and the filesystem would not have to deal with
726 * situations such as writepage being called for a page that has already
727 * had its underlying blocks deallocated.
729 void truncate_pagecache(struct inode *inode, loff_t newsize)
731 struct address_space *mapping = inode->i_mapping;
732 loff_t holebegin = round_up(newsize, PAGE_SIZE);
735 * unmap_mapping_range is called twice, first simply for
736 * efficiency so that truncate_inode_pages does fewer
737 * single-page unmaps. However after this first call, and
738 * before truncate_inode_pages finishes, it is possible for
739 * private pages to be COWed, which remain after
740 * truncate_inode_pages finishes, hence the second
741 * unmap_mapping_range call must be made for correctness.
743 unmap_mapping_range(mapping, holebegin, 0, 1);
744 truncate_inode_pages(mapping, newsize);
745 unmap_mapping_range(mapping, holebegin, 0, 1);
747 EXPORT_SYMBOL(truncate_pagecache);
750 * truncate_setsize - update inode and pagecache for a new file size
752 * @newsize: new file size
754 * truncate_setsize updates i_size and performs pagecache truncation (if
755 * necessary) to @newsize. It will be typically be called from the filesystem's
756 * setattr function when ATTR_SIZE is passed in.
758 * Must be called with a lock serializing truncates and writes (generally
759 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
760 * specific block truncation has been performed.
762 void truncate_setsize(struct inode *inode, loff_t newsize)
764 loff_t oldsize = inode->i_size;
766 i_size_write(inode, newsize);
767 if (newsize > oldsize)
768 pagecache_isize_extended(inode, oldsize, newsize);
769 truncate_pagecache(inode, newsize);
771 EXPORT_SYMBOL(truncate_setsize);
774 * pagecache_isize_extended - update pagecache after extension of i_size
775 * @inode: inode for which i_size was extended
776 * @from: original inode size
777 * @to: new inode size
779 * Handle extension of inode size either caused by extending truncate or by
780 * write starting after current i_size. We mark the page straddling current
781 * i_size RO so that page_mkwrite() is called on the nearest write access to
782 * the page. This way filesystem can be sure that page_mkwrite() is called on
783 * the page before user writes to the page via mmap after the i_size has been
786 * The function must be called after i_size is updated so that page fault
787 * coming after we unlock the page will already see the new i_size.
788 * The function must be called while we still hold i_mutex - this not only
789 * makes sure i_size is stable but also that userspace cannot observe new
790 * i_size value before we are prepared to store mmap writes at new inode size.
792 void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
794 int bsize = i_blocksize(inode);
799 WARN_ON(to > inode->i_size);
801 if (from >= to || bsize == PAGE_SIZE)
803 /* Page straddling @from will not have any hole block created? */
804 rounded_from = round_up(from, bsize);
805 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
808 index = from >> PAGE_SHIFT;
809 page = find_lock_page(inode->i_mapping, index);
810 /* Page not cached? Nothing to do */
814 * See clear_page_dirty_for_io() for details why set_page_dirty()
817 if (page_mkclean(page))
818 set_page_dirty(page);
822 EXPORT_SYMBOL(pagecache_isize_extended);
825 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
827 * @lstart: offset of beginning of hole
828 * @lend: offset of last byte of hole
830 * This function should typically be called before the filesystem
831 * releases resources associated with the freed range (eg. deallocates
832 * blocks). This way, pagecache will always stay logically coherent
833 * with on-disk format, and the filesystem would not have to deal with
834 * situations such as writepage being called for a page that has already
835 * had its underlying blocks deallocated.
837 void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
839 struct address_space *mapping = inode->i_mapping;
840 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
841 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
843 * This rounding is currently just for example: unmap_mapping_range
844 * expands its hole outwards, whereas we want it to contract the hole
845 * inwards. However, existing callers of truncate_pagecache_range are
846 * doing their own page rounding first. Note that unmap_mapping_range
847 * allows holelen 0 for all, and we allow lend -1 for end of file.
851 * Unlike in truncate_pagecache, unmap_mapping_range is called only
852 * once (before truncating pagecache), and without "even_cows" flag:
853 * hole-punching should not remove private COWed pages from the hole.
855 if ((u64)unmap_end > (u64)unmap_start)
856 unmap_mapping_range(mapping, unmap_start,
857 1 + unmap_end - unmap_start, 0);
858 truncate_inode_pages_range(mapping, lstart, lend);
860 EXPORT_SYMBOL(truncate_pagecache_range);