1 // SPDX-License-Identifier: GPL-2.0-only
3 * mm/truncate.c - code for taking down pages from address_spaces
5 * Copyright (C) 2002, Linus Torvalds
7 * 10Sep2002 Andrew Morton
11 #include <linux/kernel.h>
12 #include <linux/backing-dev.h>
13 #include <linux/dax.h>
14 #include <linux/gfp.h>
16 #include <linux/swap.h>
17 #include <linux/export.h>
18 #include <linux/pagemap.h>
19 #include <linux/highmem.h>
20 #include <linux/pagevec.h>
21 #include <linux/task_io_accounting_ops.h>
22 #include <linux/buffer_head.h> /* grr. try_to_release_page,
24 #include <linux/shmem_fs.h>
25 #include <linux/cleancache.h>
26 #include <linux/rmap.h>
30 * Regular page slots are stabilized by the page lock even without the tree
31 * itself locked. These unlocked entries need verification under the tree
34 static inline void __clear_shadow_entry(struct address_space *mapping,
35 pgoff_t index, void *entry)
37 XA_STATE(xas, &mapping->i_pages, index);
39 xas_set_update(&xas, workingset_update_node);
40 if (xas_load(&xas) != entry)
42 xas_store(&xas, NULL);
45 static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
48 spin_lock(&mapping->host->i_lock);
49 xa_lock_irq(&mapping->i_pages);
50 __clear_shadow_entry(mapping, index, entry);
51 xa_unlock_irq(&mapping->i_pages);
52 if (mapping_shrinkable(mapping))
53 inode_add_lru(mapping->host);
54 spin_unlock(&mapping->host->i_lock);
58 * Unconditionally remove exceptional entries. Usually called from truncate
59 * path. Note that the pagevec may be altered by this function by removing
60 * exceptional entries similar to what pagevec_remove_exceptionals does.
62 static void truncate_exceptional_pvec_entries(struct address_space *mapping,
63 struct pagevec *pvec, pgoff_t *indices)
68 /* Handled by shmem itself */
69 if (shmem_mapping(mapping))
72 for (j = 0; j < pagevec_count(pvec); j++)
73 if (xa_is_value(pvec->pages[j]))
76 if (j == pagevec_count(pvec))
79 dax = dax_mapping(mapping);
81 spin_lock(&mapping->host->i_lock);
82 xa_lock_irq(&mapping->i_pages);
85 for (i = j; i < pagevec_count(pvec); i++) {
86 struct page *page = pvec->pages[i];
87 pgoff_t index = indices[i];
89 if (!xa_is_value(page)) {
90 pvec->pages[j++] = page;
95 dax_delete_mapping_entry(mapping, index);
99 __clear_shadow_entry(mapping, index, page);
103 xa_unlock_irq(&mapping->i_pages);
104 if (mapping_shrinkable(mapping))
105 inode_add_lru(mapping->host);
106 spin_unlock(&mapping->host->i_lock);
112 * Invalidate exceptional entry if easily possible. This handles exceptional
113 * entries for invalidate_inode_pages().
115 static int invalidate_exceptional_entry(struct address_space *mapping,
116 pgoff_t index, void *entry)
118 /* Handled by shmem itself, or for DAX we do nothing. */
119 if (shmem_mapping(mapping) || dax_mapping(mapping))
121 clear_shadow_entry(mapping, index, entry);
126 * Invalidate exceptional entry if clean. This handles exceptional entries for
127 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
129 static int invalidate_exceptional_entry2(struct address_space *mapping,
130 pgoff_t index, void *entry)
132 /* Handled by shmem itself */
133 if (shmem_mapping(mapping))
135 if (dax_mapping(mapping))
136 return dax_invalidate_mapping_entry_sync(mapping, index);
137 clear_shadow_entry(mapping, index, entry);
142 * do_invalidatepage - invalidate part or all of a page
143 * @page: the page which is affected
144 * @offset: start of the range to invalidate
145 * @length: length of the range to invalidate
147 * do_invalidatepage() is called when all or part of the page has become
148 * invalidated by a truncate operation.
150 * do_invalidatepage() does not have to release all buffers, but it must
151 * ensure that no dirty buffer is left outside @offset and that no I/O
152 * is underway against any of the blocks which are outside the truncation
153 * point. Because the caller is about to free (and possibly reuse) those
156 void do_invalidatepage(struct page *page, unsigned int offset,
159 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
161 invalidatepage = page->mapping->a_ops->invalidatepage;
164 invalidatepage = block_invalidatepage;
167 (*invalidatepage)(page, offset, length);
171 * If truncate cannot remove the fs-private metadata from the page, the page
172 * becomes orphaned. It will be left on the LRU and may even be mapped into
173 * user pagetables if we're racing with filemap_fault().
175 * We need to bail out if page->mapping is no longer equal to the original
176 * mapping. This happens a) when the VM reclaimed the page while we waited on
177 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
178 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
180 static void truncate_cleanup_page(struct page *page)
182 if (page_mapped(page))
183 unmap_mapping_page(page);
185 if (page_has_private(page))
186 do_invalidatepage(page, 0, thp_size(page));
189 * Some filesystems seem to re-dirty the page even after
190 * the VM has canceled the dirty bit (eg ext3 journaling).
191 * Hence dirty accounting check is placed after invalidation.
193 cancel_dirty_page(page);
194 ClearPageMappedToDisk(page);
198 * This is for invalidate_mapping_pages(). That function can be called at
199 * any time, and is not supposed to throw away dirty pages. But pages can
200 * be marked dirty at any time too, so use remove_mapping which safely
201 * discards clean, unused pages.
203 * Returns non-zero if the page was successfully invalidated.
206 invalidate_complete_page(struct address_space *mapping, struct page *page)
210 if (page->mapping != mapping)
213 if (page_has_private(page) && !try_to_release_page(page, 0))
216 ret = remove_mapping(mapping, page);
221 int truncate_inode_page(struct address_space *mapping, struct page *page)
223 VM_BUG_ON_PAGE(PageTail(page), page);
225 if (page->mapping != mapping)
228 truncate_cleanup_page(page);
229 delete_from_page_cache(page);
234 * Used to get rid of pages on hardware memory corruption.
236 int generic_error_remove_page(struct address_space *mapping, struct page *page)
241 * Only punch for normal data pages for now.
242 * Handling other types like directories would need more auditing.
244 if (!S_ISREG(mapping->host->i_mode))
246 return truncate_inode_page(mapping, page);
248 EXPORT_SYMBOL(generic_error_remove_page);
251 * Safely invalidate one page from its pagecache mapping.
252 * It only drops clean, unused pages. The page must be locked.
254 * Returns 1 if the page is successfully invalidated, otherwise 0.
256 int invalidate_inode_page(struct page *page)
258 struct address_space *mapping = page_mapping(page);
261 if (PageDirty(page) || PageWriteback(page))
263 if (page_mapped(page))
265 return invalidate_complete_page(mapping, page);
269 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
270 * @mapping: mapping to truncate
271 * @lstart: offset from which to truncate
272 * @lend: offset to which to truncate (inclusive)
274 * Truncate the page cache, removing the pages that are between
275 * specified offsets (and zeroing out partial pages
276 * if lstart or lend + 1 is not page aligned).
278 * Truncate takes two passes - the first pass is nonblocking. It will not
279 * block on page locks and it will not block on writeback. The second pass
280 * will wait. This is to prevent as much IO as possible in the affected region.
281 * The first pass will remove most pages, so the search cost of the second pass
284 * We pass down the cache-hot hint to the page freeing code. Even if the
285 * mapping is large, it is probably the case that the final pages are the most
286 * recently touched, and freeing happens in ascending file offset order.
288 * Note that since ->invalidatepage() accepts range to invalidate
289 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
290 * page aligned properly.
292 void truncate_inode_pages_range(struct address_space *mapping,
293 loff_t lstart, loff_t lend)
295 pgoff_t start; /* inclusive */
296 pgoff_t end; /* exclusive */
297 unsigned int partial_start; /* inclusive */
298 unsigned int partial_end; /* exclusive */
300 pgoff_t indices[PAGEVEC_SIZE];
304 if (mapping_empty(mapping))
307 /* Offsets within partial pages */
308 partial_start = lstart & (PAGE_SIZE - 1);
309 partial_end = (lend + 1) & (PAGE_SIZE - 1);
312 * 'start' and 'end' always covers the range of pages to be fully
313 * truncated. Partial pages are covered with 'partial_start' at the
314 * start of the range and 'partial_end' at the end of the range.
315 * Note that 'end' is exclusive while 'lend' is inclusive.
317 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
320 * lend == -1 indicates end-of-file so we have to set 'end'
321 * to the highest possible pgoff_t and since the type is
322 * unsigned we're using -1.
326 end = (lend + 1) >> PAGE_SHIFT;
330 while (index < end && find_lock_entries(mapping, index, end - 1,
332 index = indices[pagevec_count(&pvec) - 1] + 1;
333 truncate_exceptional_pvec_entries(mapping, &pvec, indices);
334 for (i = 0; i < pagevec_count(&pvec); i++)
335 truncate_cleanup_page(pvec.pages[i]);
336 delete_from_page_cache_batch(mapping, &pvec);
337 for (i = 0; i < pagevec_count(&pvec); i++)
338 unlock_page(pvec.pages[i]);
339 pagevec_release(&pvec);
344 struct page *page = find_lock_page(mapping, start - 1);
346 unsigned int top = PAGE_SIZE;
348 /* Truncation within a single page */
352 wait_on_page_writeback(page);
353 zero_user_segment(page, partial_start, top);
354 cleancache_invalidate_page(mapping, page);
355 if (page_has_private(page))
356 do_invalidatepage(page, partial_start,
357 top - partial_start);
363 struct page *page = find_lock_page(mapping, end);
365 wait_on_page_writeback(page);
366 zero_user_segment(page, 0, partial_end);
367 cleancache_invalidate_page(mapping, page);
368 if (page_has_private(page))
369 do_invalidatepage(page, 0,
376 * If the truncation happened within a single page no pages
377 * will be released, just zeroed, so we can bail out now.
385 if (!find_get_entries(mapping, index, end - 1, &pvec,
387 /* If all gone from start onwards, we're done */
390 /* Otherwise restart to make sure all gone */
395 for (i = 0; i < pagevec_count(&pvec); i++) {
396 struct page *page = pvec.pages[i];
398 /* We rely upon deletion not changing page->index */
401 if (xa_is_value(page))
405 WARN_ON(page_to_index(page) != index);
406 wait_on_page_writeback(page);
407 truncate_inode_page(mapping, page);
410 truncate_exceptional_pvec_entries(mapping, &pvec, indices);
411 pagevec_release(&pvec);
416 cleancache_invalidate_inode(mapping);
418 EXPORT_SYMBOL(truncate_inode_pages_range);
421 * truncate_inode_pages - truncate *all* the pages from an offset
422 * @mapping: mapping to truncate
423 * @lstart: offset from which to truncate
425 * Called under (and serialised by) inode->i_rwsem and
426 * mapping->invalidate_lock.
428 * Note: When this function returns, there can be a page in the process of
429 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
430 * mapping->nrpages can be non-zero when this function returns even after
431 * truncation of the whole mapping.
433 void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
435 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
437 EXPORT_SYMBOL(truncate_inode_pages);
440 * truncate_inode_pages_final - truncate *all* pages before inode dies
441 * @mapping: mapping to truncate
443 * Called under (and serialized by) inode->i_rwsem.
445 * Filesystems have to use this in the .evict_inode path to inform the
446 * VM that this is the final truncate and the inode is going away.
448 void truncate_inode_pages_final(struct address_space *mapping)
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);
459 if (!mapping_empty(mapping)) {
461 * As truncation uses a lockless tree lookup, cycle
462 * the tree lock to make sure any ongoing tree
463 * modification that does not see AS_EXITING is
464 * completed before starting the final truncate.
466 xa_lock_irq(&mapping->i_pages);
467 xa_unlock_irq(&mapping->i_pages);
471 * Cleancache needs notification even if there are no pages or shadow
474 truncate_inode_pages(mapping, 0);
476 EXPORT_SYMBOL(truncate_inode_pages_final);
478 static unsigned long __invalidate_mapping_pages(struct address_space *mapping,
479 pgoff_t start, pgoff_t end, unsigned long *nr_pagevec)
481 pgoff_t indices[PAGEVEC_SIZE];
483 pgoff_t index = start;
485 unsigned long count = 0;
489 while (find_lock_entries(mapping, index, end, &pvec, indices)) {
490 for (i = 0; i < pagevec_count(&pvec); i++) {
491 struct page *page = pvec.pages[i];
493 /* We rely upon deletion not changing page->index */
496 if (xa_is_value(page)) {
497 count += invalidate_exceptional_entry(mapping,
502 index += thp_nr_pages(page) - 1;
504 ret = invalidate_inode_page(page);
507 * Invalidation is a hint that the page is no longer
508 * of interest and try to speed up its reclaim.
511 deactivate_file_page(page);
512 /* It is likely on the pagevec of a remote CPU */
518 pagevec_remove_exceptionals(&pvec);
519 pagevec_release(&pvec);
527 * invalidate_mapping_pages - Invalidate all clean, unlocked cache of one inode
528 * @mapping: the address_space which holds the cache to invalidate
529 * @start: the offset 'from' which to invalidate
530 * @end: the offset 'to' which to invalidate (inclusive)
532 * This function removes pages that are clean, unmapped and unlocked,
533 * as well as shadow entries. It will not block on IO activity.
535 * If you want to remove all the pages of one inode, regardless of
536 * their use and writeback state, use truncate_inode_pages().
538 * Return: the number of the cache entries that were invalidated
540 unsigned long invalidate_mapping_pages(struct address_space *mapping,
541 pgoff_t start, pgoff_t end)
543 return __invalidate_mapping_pages(mapping, start, end, NULL);
545 EXPORT_SYMBOL(invalidate_mapping_pages);
548 * invalidate_mapping_pagevec - Invalidate all the unlocked pages of one inode
549 * @mapping: the address_space which holds the pages to invalidate
550 * @start: the offset 'from' which to invalidate
551 * @end: the offset 'to' which to invalidate (inclusive)
552 * @nr_pagevec: invalidate failed page number for caller
554 * This helper is similar to invalidate_mapping_pages(), except that it accounts
555 * for pages that are likely on a pagevec and counts them in @nr_pagevec, which
556 * will be used by the caller.
558 void invalidate_mapping_pagevec(struct address_space *mapping,
559 pgoff_t start, pgoff_t end, unsigned long *nr_pagevec)
561 __invalidate_mapping_pages(mapping, start, end, nr_pagevec);
565 * This is like invalidate_complete_page(), except it ignores the page's
566 * refcount. We do this because invalidate_inode_pages2() needs stronger
567 * invalidation guarantees, and cannot afford to leave pages behind because
568 * shrink_page_list() has a temp ref on them, or because they're transiently
569 * sitting in the lru_cache_add() pagevecs.
572 invalidate_complete_page2(struct address_space *mapping, struct page *page)
574 if (page->mapping != mapping)
577 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
580 spin_lock(&mapping->host->i_lock);
581 xa_lock_irq(&mapping->i_pages);
585 BUG_ON(page_has_private(page));
586 __delete_from_page_cache(page, NULL);
587 xa_unlock_irq(&mapping->i_pages);
588 if (mapping_shrinkable(mapping))
589 inode_add_lru(mapping->host);
590 spin_unlock(&mapping->host->i_lock);
592 if (mapping->a_ops->freepage)
593 mapping->a_ops->freepage(page);
595 put_page(page); /* pagecache ref */
598 xa_unlock_irq(&mapping->i_pages);
599 spin_unlock(&mapping->host->i_lock);
603 static int do_launder_page(struct address_space *mapping, struct page *page)
605 if (!PageDirty(page))
607 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
609 return mapping->a_ops->launder_page(page);
613 * invalidate_inode_pages2_range - remove range of pages from an address_space
614 * @mapping: the address_space
615 * @start: the page offset 'from' which to invalidate
616 * @end: the page offset 'to' which to invalidate (inclusive)
618 * Any pages which are found to be mapped into pagetables are unmapped prior to
621 * Return: -EBUSY if any pages could not be invalidated.
623 int invalidate_inode_pages2_range(struct address_space *mapping,
624 pgoff_t start, pgoff_t end)
626 pgoff_t indices[PAGEVEC_SIZE];
632 int did_range_unmap = 0;
634 if (mapping_empty(mapping))
639 while (find_get_entries(mapping, index, end, &pvec, indices)) {
640 for (i = 0; i < pagevec_count(&pvec); i++) {
641 struct page *page = pvec.pages[i];
643 /* We rely upon deletion not changing page->index */
646 if (xa_is_value(page)) {
647 if (!invalidate_exceptional_entry2(mapping,
653 if (!did_range_unmap && page_mapped(page)) {
655 * If page is mapped, before taking its lock,
656 * zap the rest of the file in one hit.
658 unmap_mapping_pages(mapping, index,
659 (1 + end - index), false);
664 WARN_ON(page_to_index(page) != index);
665 if (page->mapping != mapping) {
669 wait_on_page_writeback(page);
671 if (page_mapped(page))
672 unmap_mapping_page(page);
673 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);
690 * For DAX we invalidate page tables after invalidating page cache. We
691 * could invalidate page tables while invalidating each entry however
692 * that would be expensive. And doing range unmapping before doesn't
693 * work as we have no cheap way to find whether page cache entry didn't
694 * get remapped later.
696 if (dax_mapping(mapping)) {
697 unmap_mapping_pages(mapping, start, end - start + 1, false);
700 cleancache_invalidate_inode(mapping);
703 EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
706 * invalidate_inode_pages2 - remove all pages from an address_space
707 * @mapping: the address_space
709 * Any pages which are found to be mapped into pagetables are unmapped prior to
712 * Return: -EBUSY if any pages could not be invalidated.
714 int invalidate_inode_pages2(struct address_space *mapping)
716 return invalidate_inode_pages2_range(mapping, 0, -1);
718 EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
721 * truncate_pagecache - unmap and remove pagecache that has been truncated
723 * @newsize: new file size
725 * inode's new i_size must already be written before truncate_pagecache
728 * This function should typically be called before the filesystem
729 * releases resources associated with the freed range (eg. deallocates
730 * blocks). This way, pagecache will always stay logically coherent
731 * with on-disk format, and the filesystem would not have to deal with
732 * situations such as writepage being called for a page that has already
733 * had its underlying blocks deallocated.
735 void truncate_pagecache(struct inode *inode, loff_t newsize)
737 struct address_space *mapping = inode->i_mapping;
738 loff_t holebegin = round_up(newsize, PAGE_SIZE);
741 * unmap_mapping_range is called twice, first simply for
742 * efficiency so that truncate_inode_pages does fewer
743 * single-page unmaps. However after this first call, and
744 * before truncate_inode_pages finishes, it is possible for
745 * private pages to be COWed, which remain after
746 * truncate_inode_pages finishes, hence the second
747 * unmap_mapping_range call must be made for correctness.
749 unmap_mapping_range(mapping, holebegin, 0, 1);
750 truncate_inode_pages(mapping, newsize);
751 unmap_mapping_range(mapping, holebegin, 0, 1);
753 EXPORT_SYMBOL(truncate_pagecache);
756 * truncate_setsize - update inode and pagecache for a new file size
758 * @newsize: new file size
760 * truncate_setsize updates i_size and performs pagecache truncation (if
761 * necessary) to @newsize. It will be typically be called from the filesystem's
762 * setattr function when ATTR_SIZE is passed in.
764 * Must be called with a lock serializing truncates and writes (generally
765 * i_rwsem but e.g. xfs uses a different lock) and before all filesystem
766 * specific block truncation has been performed.
768 void truncate_setsize(struct inode *inode, loff_t newsize)
770 loff_t oldsize = inode->i_size;
772 i_size_write(inode, newsize);
773 if (newsize > oldsize)
774 pagecache_isize_extended(inode, oldsize, newsize);
775 truncate_pagecache(inode, newsize);
777 EXPORT_SYMBOL(truncate_setsize);
780 * pagecache_isize_extended - update pagecache after extension of i_size
781 * @inode: inode for which i_size was extended
782 * @from: original inode size
783 * @to: new inode size
785 * Handle extension of inode size either caused by extending truncate or by
786 * write starting after current i_size. We mark the page straddling current
787 * i_size RO so that page_mkwrite() is called on the nearest write access to
788 * the page. This way filesystem can be sure that page_mkwrite() is called on
789 * the page before user writes to the page via mmap after the i_size has been
792 * The function must be called after i_size is updated so that page fault
793 * coming after we unlock the page will already see the new i_size.
794 * The function must be called while we still hold i_rwsem - this not only
795 * makes sure i_size is stable but also that userspace cannot observe new
796 * i_size value before we are prepared to store mmap writes at new inode size.
798 void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
800 int bsize = i_blocksize(inode);
805 WARN_ON(to > inode->i_size);
807 if (from >= to || bsize == PAGE_SIZE)
809 /* Page straddling @from will not have any hole block created? */
810 rounded_from = round_up(from, bsize);
811 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
814 index = from >> PAGE_SHIFT;
815 page = find_lock_page(inode->i_mapping, index);
816 /* Page not cached? Nothing to do */
820 * See clear_page_dirty_for_io() for details why set_page_dirty()
823 if (page_mkclean(page))
824 set_page_dirty(page);
828 EXPORT_SYMBOL(pagecache_isize_extended);
831 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
833 * @lstart: offset of beginning of hole
834 * @lend: offset of last byte of hole
836 * This function should typically be called before the filesystem
837 * releases resources associated with the freed range (eg. deallocates
838 * blocks). This way, pagecache will always stay logically coherent
839 * with on-disk format, and the filesystem would not have to deal with
840 * situations such as writepage being called for a page that has already
841 * had its underlying blocks deallocated.
843 void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
845 struct address_space *mapping = inode->i_mapping;
846 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
847 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
849 * This rounding is currently just for example: unmap_mapping_range
850 * expands its hole outwards, whereas we want it to contract the hole
851 * inwards. However, existing callers of truncate_pagecache_range are
852 * doing their own page rounding first. Note that unmap_mapping_range
853 * allows holelen 0 for all, and we allow lend -1 for end of file.
857 * Unlike in truncate_pagecache, unmap_mapping_range is called only
858 * once (before truncating pagecache), and without "even_cows" flag:
859 * hole-punching should not remove private COWed pages from the hole.
861 if ((u64)unmap_end > (u64)unmap_start)
862 unmap_mapping_range(mapping, unmap_start,
863 1 + unmap_end - unmap_start, 0);
864 truncate_inode_pages_range(mapping, lstart, lend);
866 EXPORT_SYMBOL(truncate_pagecache_range);