Linux Audio

Check our new training course

Loading...
v4.6
 
  1/*
  2 * mm/truncate.c - code for taking down pages from address_spaces
  3 *
  4 * Copyright (C) 2002, Linus Torvalds
  5 *
  6 * 10Sep2002	Andrew Morton
  7 *		Initial version.
  8 */
  9
 10#include <linux/kernel.h>
 11#include <linux/backing-dev.h>
 12#include <linux/dax.h>
 13#include <linux/gfp.h>
 14#include <linux/mm.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,
 22				   do_invalidatepage */
 23#include <linux/cleancache.h>
 24#include <linux/rmap.h>
 25#include "internal.h"
 26
 27static void clear_exceptional_entry(struct address_space *mapping,
 28				    pgoff_t index, void *entry)
 
 
 
 
 
 29{
 30	struct radix_tree_node *node;
 31	void **slot;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 32
 33	/* Handled by shmem itself */
 34	if (shmem_mapping(mapping))
 35		return;
 36
 37	spin_lock_irq(&mapping->tree_lock);
 
 
 38
 39	if (dax_mapping(mapping)) {
 40		if (radix_tree_delete_item(&mapping->page_tree, index, entry))
 41			mapping->nrexceptional--;
 42	} else {
 43		/*
 44		 * Regular page slots are stabilized by the page lock even
 45		 * without the tree itself locked.  These unlocked entries
 46		 * need verification under the tree lock.
 47		 */
 48		if (!__radix_tree_lookup(&mapping->page_tree, index, &node,
 49					&slot))
 50			goto unlock;
 51		if (*slot != entry)
 52			goto unlock;
 53		radix_tree_replace_slot(slot, NULL);
 54		mapping->nrexceptional--;
 55		if (!node)
 56			goto unlock;
 57		workingset_node_shadows_dec(node);
 58		/*
 59		 * Don't track node without shadow entries.
 60		 *
 61		 * Avoid acquiring the list_lru lock if already untracked.
 62		 * The list_empty() test is safe as node->private_list is
 63		 * protected by mapping->tree_lock.
 64		 */
 65		if (!workingset_node_shadows(node) &&
 66		    !list_empty(&node->private_list))
 67			list_lru_del(&workingset_shadow_nodes,
 68					&node->private_list);
 69		__radix_tree_delete_node(&mapping->page_tree, node);
 70	}
 71unlock:
 72	spin_unlock_irq(&mapping->tree_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 73}
 74
 75/**
 76 * do_invalidatepage - invalidate part or all of a page
 77 * @page: the page which is affected
 78 * @offset: start of the range to invalidate
 79 * @length: length of the range to invalidate
 80 *
 81 * do_invalidatepage() is called when all or part of the page has become
 82 * invalidated by a truncate operation.
 83 *
 84 * do_invalidatepage() does not have to release all buffers, but it must
 85 * ensure that no dirty buffer is left outside @offset and that no I/O
 86 * is underway against any of the blocks which are outside the truncation
 87 * point.  Because the caller is about to free (and possibly reuse) those
 88 * blocks on-disk.
 89 */
 90void do_invalidatepage(struct page *page, unsigned int offset,
 91		       unsigned int length)
 92{
 93	void (*invalidatepage)(struct page *, unsigned int, unsigned int);
 94
 95	invalidatepage = page->mapping->a_ops->invalidatepage;
 96#ifdef CONFIG_BLOCK
 97	if (!invalidatepage)
 98		invalidatepage = block_invalidatepage;
 99#endif
100	if (invalidatepage)
101		(*invalidatepage)(page, offset, length);
102}
 
103
104/*
105 * If truncate cannot remove the fs-private metadata from the page, the page
106 * becomes orphaned.  It will be left on the LRU and may even be mapped into
107 * user pagetables if we're racing with filemap_fault().
108 *
109 * We need to bale out if page->mapping is no longer equal to the original
110 * mapping.  This happens a) when the VM reclaimed the page while we waited on
111 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
112 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
113 */
114static int
115truncate_complete_page(struct address_space *mapping, struct page *page)
116{
117	if (page->mapping != mapping)
118		return -EIO;
119
120	if (page_has_private(page))
121		do_invalidatepage(page, 0, PAGE_SIZE);
122
123	/*
124	 * Some filesystems seem to re-dirty the page even after
125	 * the VM has canceled the dirty bit (eg ext3 journaling).
126	 * Hence dirty accounting check is placed after invalidation.
127	 */
128	cancel_dirty_page(page);
129	ClearPageMappedToDisk(page);
130	delete_from_page_cache(page);
 
 
 
 
 
 
 
 
131	return 0;
132}
133
134/*
135 * This is for invalidate_mapping_pages().  That function can be called at
136 * any time, and is not supposed to throw away dirty pages.  But pages can
137 * be marked dirty at any time too, so use remove_mapping which safely
138 * discards clean, unused pages.
139 *
140 * Returns non-zero if the page was successfully invalidated.
 
 
 
141 */
142static int
143invalidate_complete_page(struct address_space *mapping, struct page *page)
144{
145	int ret;
146
147	if (page->mapping != mapping)
148		return 0;
149
150	if (page_has_private(page) && !try_to_release_page(page, 0))
151		return 0;
 
 
 
 
 
 
 
152
153	ret = remove_mapping(mapping, page);
 
 
 
 
154
155	return ret;
156}
 
 
 
 
157
158int truncate_inode_page(struct address_space *mapping, struct page *page)
159{
160	if (page_mapped(page)) {
161		unmap_mapping_range(mapping,
162				   (loff_t)page->index << PAGE_SHIFT,
163				   PAGE_SIZE, 0);
164	}
165	return truncate_complete_page(mapping, page);
 
 
166}
167
168/*
169 * Used to get rid of pages on hardware memory corruption.
170 */
171int generic_error_remove_page(struct address_space *mapping, struct page *page)
 
172{
173	if (!mapping)
174		return -EINVAL;
175	/*
176	 * Only punch for normal data pages for now.
177	 * Handling other types like directories would need more auditing.
178	 */
179	if (!S_ISREG(mapping->host->i_mode))
180		return -EIO;
181	return truncate_inode_page(mapping, page);
182}
183EXPORT_SYMBOL(generic_error_remove_page);
184
185/*
186 * Safely invalidate one page from its pagecache mapping.
187 * It only drops clean, unused pages. The page must be locked.
 
188 *
189 * Returns 1 if the page is successfully invalidated, otherwise 0.
 
 
 
 
190 */
191int invalidate_inode_page(struct page *page)
192{
193	struct address_space *mapping = page_mapping(page);
194	if (!mapping)
195		return 0;
196	if (PageDirty(page) || PageWriteback(page))
 
 
 
 
197		return 0;
198	if (page_mapped(page))
199		return 0;
200	return invalidate_complete_page(mapping, page);
 
201}
202
203/**
204 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
205 * @mapping: mapping to truncate
206 * @lstart: offset from which to truncate
207 * @lend: offset to which to truncate (inclusive)
208 *
209 * Truncate the page cache, removing the pages that are between
210 * specified offsets (and zeroing out partial pages
211 * if lstart or lend + 1 is not page aligned).
212 *
213 * Truncate takes two passes - the first pass is nonblocking.  It will not
214 * block on page locks and it will not block on writeback.  The second pass
215 * will wait.  This is to prevent as much IO as possible in the affected region.
216 * The first pass will remove most pages, so the search cost of the second pass
217 * is low.
218 *
219 * We pass down the cache-hot hint to the page freeing code.  Even if the
220 * mapping is large, it is probably the case that the final pages are the most
221 * recently touched, and freeing happens in ascending file offset order.
222 *
223 * Note that since ->invalidatepage() accepts range to invalidate
224 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
225 * page aligned properly.
226 */
227void truncate_inode_pages_range(struct address_space *mapping,
228				loff_t lstart, loff_t lend)
229{
230	pgoff_t		start;		/* inclusive */
231	pgoff_t		end;		/* exclusive */
232	unsigned int	partial_start;	/* inclusive */
233	unsigned int	partial_end;	/* exclusive */
234	struct pagevec	pvec;
235	pgoff_t		indices[PAGEVEC_SIZE];
236	pgoff_t		index;
237	int		i;
 
 
238
239	cleancache_invalidate_inode(mapping);
240	if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
241		return;
242
243	/* Offsets within partial pages */
244	partial_start = lstart & (PAGE_SIZE - 1);
245	partial_end = (lend + 1) & (PAGE_SIZE - 1);
246
247	/*
248	 * 'start' and 'end' always covers the range of pages to be fully
249	 * truncated. Partial pages are covered with 'partial_start' at the
250	 * start of the range and 'partial_end' at the end of the range.
251	 * Note that 'end' is exclusive while 'lend' is inclusive.
252	 */
253	start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
254	if (lend == -1)
255		/*
256		 * lend == -1 indicates end-of-file so we have to set 'end'
257		 * to the highest possible pgoff_t and since the type is
258		 * unsigned we're using -1.
259		 */
260		end = -1;
261	else
262		end = (lend + 1) >> PAGE_SHIFT;
263
264	pagevec_init(&pvec, 0);
265	index = start;
266	while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
267			min(end - index, (pgoff_t)PAGEVEC_SIZE),
268			indices)) {
269		for (i = 0; i < pagevec_count(&pvec); i++) {
270			struct page *page = pvec.pages[i];
271
272			/* We rely upon deletion not changing page->index */
273			index = indices[i];
274			if (index >= end)
275				break;
276
277			if (radix_tree_exceptional_entry(page)) {
278				clear_exceptional_entry(mapping, index, page);
279				continue;
280			}
281
282			if (!trylock_page(page))
283				continue;
284			WARN_ON(page->index != index);
285			if (PageWriteback(page)) {
286				unlock_page(page);
287				continue;
288			}
289			truncate_inode_page(mapping, page);
290			unlock_page(page);
291		}
292		pagevec_remove_exceptionals(&pvec);
293		pagevec_release(&pvec);
294		cond_resched();
295		index++;
296	}
297
298	if (partial_start) {
299		struct page *page = find_lock_page(mapping, start - 1);
300		if (page) {
301			unsigned int top = PAGE_SIZE;
302			if (start > end) {
303				/* Truncation within a single page */
304				top = partial_end;
305				partial_end = 0;
306			}
307			wait_on_page_writeback(page);
308			zero_user_segment(page, partial_start, top);
309			cleancache_invalidate_page(mapping, page);
310			if (page_has_private(page))
311				do_invalidatepage(page, partial_start,
312						  top - partial_start);
313			unlock_page(page);
314			put_page(page);
315		}
 
 
 
316	}
317	if (partial_end) {
318		struct page *page = find_lock_page(mapping, end);
319		if (page) {
320			wait_on_page_writeback(page);
321			zero_user_segment(page, 0, partial_end);
322			cleancache_invalidate_page(mapping, page);
323			if (page_has_private(page))
324				do_invalidatepage(page, 0,
325						  partial_end);
326			unlock_page(page);
327			put_page(page);
328		}
329	}
330	/*
331	 * If the truncation happened within a single page no pages
332	 * will be released, just zeroed, so we can bail out now.
333	 */
334	if (start >= end)
335		return;
336
337	index = start;
338	for ( ; ; ) {
339		cond_resched();
340		if (!pagevec_lookup_entries(&pvec, mapping, index,
341			min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
342			/* If all gone from start onwards, we're done */
343			if (index == start)
344				break;
345			/* Otherwise restart to make sure all gone */
346			index = start;
347			continue;
348		}
349		if (index == start && indices[0] >= end) {
350			/* All gone out of hole to be punched, we're done */
351			pagevec_remove_exceptionals(&pvec);
352			pagevec_release(&pvec);
353			break;
354		}
355		for (i = 0; i < pagevec_count(&pvec); i++) {
356			struct page *page = pvec.pages[i];
357
358			/* We rely upon deletion not changing page->index */
359			index = indices[i];
360			if (index >= end) {
361				/* Restart punch to make sure all gone */
362				index = start - 1;
363				break;
364			}
365
366			if (radix_tree_exceptional_entry(page)) {
367				clear_exceptional_entry(mapping, index, page);
368				continue;
369			}
370
371			lock_page(page);
372			WARN_ON(page->index != index);
373			wait_on_page_writeback(page);
374			truncate_inode_page(mapping, page);
375			unlock_page(page);
376		}
377		pagevec_remove_exceptionals(&pvec);
378		pagevec_release(&pvec);
379		index++;
380	}
381	cleancache_invalidate_inode(mapping);
382}
383EXPORT_SYMBOL(truncate_inode_pages_range);
384
385/**
386 * truncate_inode_pages - truncate *all* the pages from an offset
387 * @mapping: mapping to truncate
388 * @lstart: offset from which to truncate
389 *
390 * Called under (and serialised by) inode->i_mutex.
 
391 *
392 * Note: When this function returns, there can be a page in the process of
393 * deletion (inside __delete_from_page_cache()) in the specified range.  Thus
394 * mapping->nrpages can be non-zero when this function returns even after
395 * truncation of the whole mapping.
396 */
397void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
398{
399	truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
400}
401EXPORT_SYMBOL(truncate_inode_pages);
402
403/**
404 * truncate_inode_pages_final - truncate *all* pages before inode dies
405 * @mapping: mapping to truncate
406 *
407 * Called under (and serialized by) inode->i_mutex.
408 *
409 * Filesystems have to use this in the .evict_inode path to inform the
410 * VM that this is the final truncate and the inode is going away.
411 */
412void truncate_inode_pages_final(struct address_space *mapping)
413{
414	unsigned long nrexceptional;
415	unsigned long nrpages;
416
417	/*
418	 * Page reclaim can not participate in regular inode lifetime
419	 * management (can't call iput()) and thus can race with the
420	 * inode teardown.  Tell it when the address space is exiting,
421	 * so that it does not install eviction information after the
422	 * final truncate has begun.
423	 */
424	mapping_set_exiting(mapping);
425
426	/*
427	 * When reclaim installs eviction entries, it increases
428	 * nrexceptional first, then decreases nrpages.  Make sure we see
429	 * this in the right order or we might miss an entry.
430	 */
431	nrpages = mapping->nrpages;
432	smp_rmb();
433	nrexceptional = mapping->nrexceptional;
434
435	if (nrpages || nrexceptional) {
436		/*
437		 * As truncation uses a lockless tree lookup, cycle
438		 * the tree lock to make sure any ongoing tree
439		 * modification that does not see AS_EXITING is
440		 * completed before starting the final truncate.
441		 */
442		spin_lock_irq(&mapping->tree_lock);
443		spin_unlock_irq(&mapping->tree_lock);
444
445		truncate_inode_pages(mapping, 0);
446	}
 
 
447}
448EXPORT_SYMBOL(truncate_inode_pages_final);
449
450/**
451 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
452 * @mapping: the address_space which holds the pages to invalidate
453 * @start: the offset 'from' which to invalidate
454 * @end: the offset 'to' which to invalidate (inclusive)
 
455 *
456 * This function only removes the unlocked pages, if you want to
457 * remove all the pages of one inode, you must call truncate_inode_pages.
458 *
459 * invalidate_mapping_pages() will not block on IO activity. It will not
460 * invalidate pages which are dirty, locked, under writeback or mapped into
461 * pagetables.
462 */
463unsigned long invalidate_mapping_pages(struct address_space *mapping,
464		pgoff_t start, pgoff_t end)
465{
466	pgoff_t indices[PAGEVEC_SIZE];
467	struct pagevec pvec;
468	pgoff_t index = start;
469	unsigned long ret;
470	unsigned long count = 0;
471	int i;
472
473	pagevec_init(&pvec, 0);
474	while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
475			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
476			indices)) {
477		for (i = 0; i < pagevec_count(&pvec); i++) {
478			struct page *page = pvec.pages[i];
479
480			/* We rely upon deletion not changing page->index */
481			index = indices[i];
482			if (index > end)
483				break;
484
485			if (radix_tree_exceptional_entry(page)) {
486				clear_exceptional_entry(mapping, index, page);
487				continue;
488			}
489
490			if (!trylock_page(page))
491				continue;
492			WARN_ON(page->index != index);
493			ret = invalidate_inode_page(page);
494			unlock_page(page);
495			/*
496			 * Invalidation is a hint that the page is no longer
497			 * of interest and try to speed up its reclaim.
498			 */
499			if (!ret)
500				deactivate_file_page(page);
 
 
 
 
501			count += ret;
502		}
503		pagevec_remove_exceptionals(&pvec);
504		pagevec_release(&pvec);
505		cond_resched();
506		index++;
507	}
508	return count;
509}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
510EXPORT_SYMBOL(invalidate_mapping_pages);
511
512/*
513 * This is like invalidate_complete_page(), except it ignores the page's
514 * refcount.  We do this because invalidate_inode_pages2() needs stronger
515 * invalidation guarantees, and cannot afford to leave pages behind because
516 * shrink_page_list() has a temp ref on them, or because they're transiently
517 * sitting in the lru_cache_add() pagevecs.
518 */
519static int
520invalidate_complete_page2(struct address_space *mapping, struct page *page)
521{
522	unsigned long flags;
523
524	if (page->mapping != mapping)
525		return 0;
526
527	if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
528		return 0;
529
530	spin_lock_irqsave(&mapping->tree_lock, flags);
531	if (PageDirty(page))
 
532		goto failed;
533
534	BUG_ON(page_has_private(page));
535	__delete_from_page_cache(page, NULL);
536	spin_unlock_irqrestore(&mapping->tree_lock, flags);
537
538	if (mapping->a_ops->freepage)
539		mapping->a_ops->freepage(page);
540
541	put_page(page);	/* pagecache ref */
542	return 1;
543failed:
544	spin_unlock_irqrestore(&mapping->tree_lock, flags);
 
545	return 0;
546}
547
548static int do_launder_page(struct address_space *mapping, struct page *page)
549{
550	if (!PageDirty(page))
551		return 0;
552	if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
553		return 0;
554	return mapping->a_ops->launder_page(page);
555}
556
557/**
558 * invalidate_inode_pages2_range - remove range of pages from an address_space
559 * @mapping: the address_space
560 * @start: the page offset 'from' which to invalidate
561 * @end: the page offset 'to' which to invalidate (inclusive)
562 *
563 * Any pages which are found to be mapped into pagetables are unmapped prior to
564 * invalidation.
565 *
566 * Returns -EBUSY if any pages could not be invalidated.
567 */
568int invalidate_inode_pages2_range(struct address_space *mapping,
569				  pgoff_t start, pgoff_t end)
570{
571	pgoff_t indices[PAGEVEC_SIZE];
572	struct pagevec pvec;
573	pgoff_t index;
574	int i;
575	int ret = 0;
576	int ret2 = 0;
577	int did_range_unmap = 0;
578
579	cleancache_invalidate_inode(mapping);
580	pagevec_init(&pvec, 0);
581	index = start;
582	while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
583			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
584			indices)) {
585		for (i = 0; i < pagevec_count(&pvec); i++) {
586			struct page *page = pvec.pages[i];
587
588			/* We rely upon deletion not changing page->index */
589			index = indices[i];
590			if (index > end)
591				break;
592
593			if (radix_tree_exceptional_entry(page)) {
594				clear_exceptional_entry(mapping, index, page);
 
 
 
 
 
 
 
 
 
 
595				continue;
596			}
597
598			lock_page(page);
599			WARN_ON(page->index != index);
600			if (page->mapping != mapping) {
601				unlock_page(page);
602				continue;
 
 
 
603			}
604			wait_on_page_writeback(page);
605			if (page_mapped(page)) {
606				if (!did_range_unmap) {
607					/*
608					 * Zap the rest of the file in one hit.
609					 */
610					unmap_mapping_range(mapping,
611					   (loff_t)index << PAGE_SHIFT,
612					   (loff_t)(1 + end - index)
613							 << PAGE_SHIFT,
614							 0);
615					did_range_unmap = 1;
616				} else {
617					/*
618					 * Just zap this page
619					 */
620					unmap_mapping_range(mapping,
621					   (loff_t)index << PAGE_SHIFT,
622					   PAGE_SIZE, 0);
623				}
624			}
625			BUG_ON(page_mapped(page));
626			ret2 = do_launder_page(mapping, page);
 
 
 
 
 
 
627			if (ret2 == 0) {
628				if (!invalidate_complete_page2(mapping, page))
629					ret2 = -EBUSY;
630			}
631			if (ret2 < 0)
632				ret = ret2;
633			unlock_page(page);
634		}
635		pagevec_remove_exceptionals(&pvec);
636		pagevec_release(&pvec);
637		cond_resched();
638		index++;
639	}
640	cleancache_invalidate_inode(mapping);
 
 
 
 
 
 
 
 
 
641	return ret;
642}
643EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
644
645/**
646 * invalidate_inode_pages2 - remove all pages from an address_space
647 * @mapping: the address_space
648 *
649 * Any pages which are found to be mapped into pagetables are unmapped prior to
650 * invalidation.
651 *
652 * Returns -EBUSY if any pages could not be invalidated.
653 */
654int invalidate_inode_pages2(struct address_space *mapping)
655{
656	return invalidate_inode_pages2_range(mapping, 0, -1);
657}
658EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
659
660/**
661 * truncate_pagecache - unmap and remove pagecache that has been truncated
662 * @inode: inode
663 * @newsize: new file size
664 *
665 * inode's new i_size must already be written before truncate_pagecache
666 * is called.
667 *
668 * This function should typically be called before the filesystem
669 * releases resources associated with the freed range (eg. deallocates
670 * blocks). This way, pagecache will always stay logically coherent
671 * with on-disk format, and the filesystem would not have to deal with
672 * situations such as writepage being called for a page that has already
673 * had its underlying blocks deallocated.
674 */
675void truncate_pagecache(struct inode *inode, loff_t newsize)
676{
677	struct address_space *mapping = inode->i_mapping;
678	loff_t holebegin = round_up(newsize, PAGE_SIZE);
679
680	/*
681	 * unmap_mapping_range is called twice, first simply for
682	 * efficiency so that truncate_inode_pages does fewer
683	 * single-page unmaps.  However after this first call, and
684	 * before truncate_inode_pages finishes, it is possible for
685	 * private pages to be COWed, which remain after
686	 * truncate_inode_pages finishes, hence the second
687	 * unmap_mapping_range call must be made for correctness.
688	 */
689	unmap_mapping_range(mapping, holebegin, 0, 1);
690	truncate_inode_pages(mapping, newsize);
691	unmap_mapping_range(mapping, holebegin, 0, 1);
692}
693EXPORT_SYMBOL(truncate_pagecache);
694
695/**
696 * truncate_setsize - update inode and pagecache for a new file size
697 * @inode: inode
698 * @newsize: new file size
699 *
700 * truncate_setsize updates i_size and performs pagecache truncation (if
701 * necessary) to @newsize. It will be typically be called from the filesystem's
702 * setattr function when ATTR_SIZE is passed in.
703 *
704 * Must be called with a lock serializing truncates and writes (generally
705 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
706 * specific block truncation has been performed.
707 */
708void truncate_setsize(struct inode *inode, loff_t newsize)
709{
710	loff_t oldsize = inode->i_size;
711
712	i_size_write(inode, newsize);
713	if (newsize > oldsize)
714		pagecache_isize_extended(inode, oldsize, newsize);
715	truncate_pagecache(inode, newsize);
716}
717EXPORT_SYMBOL(truncate_setsize);
718
719/**
720 * pagecache_isize_extended - update pagecache after extension of i_size
721 * @inode:	inode for which i_size was extended
722 * @from:	original inode size
723 * @to:		new inode size
724 *
725 * Handle extension of inode size either caused by extending truncate or by
726 * write starting after current i_size. We mark the page straddling current
727 * i_size RO so that page_mkwrite() is called on the nearest write access to
728 * the page.  This way filesystem can be sure that page_mkwrite() is called on
729 * the page before user writes to the page via mmap after the i_size has been
730 * changed.
731 *
732 * The function must be called after i_size is updated so that page fault
733 * coming after we unlock the page will already see the new i_size.
734 * The function must be called while we still hold i_mutex - this not only
735 * makes sure i_size is stable but also that userspace cannot observe new
736 * i_size value before we are prepared to store mmap writes at new inode size.
737 */
738void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
739{
740	int bsize = 1 << inode->i_blkbits;
741	loff_t rounded_from;
742	struct page *page;
743	pgoff_t index;
744
745	WARN_ON(to > inode->i_size);
746
747	if (from >= to || bsize == PAGE_SIZE)
748		return;
749	/* Page straddling @from will not have any hole block created? */
750	rounded_from = round_up(from, bsize);
751	if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
752		return;
753
754	index = from >> PAGE_SHIFT;
755	page = find_lock_page(inode->i_mapping, index);
756	/* Page not cached? Nothing to do */
757	if (!page)
758		return;
759	/*
760	 * See clear_page_dirty_for_io() for details why set_page_dirty()
761	 * is needed.
762	 */
763	if (page_mkclean(page))
764		set_page_dirty(page);
765	unlock_page(page);
766	put_page(page);
767}
768EXPORT_SYMBOL(pagecache_isize_extended);
769
770/**
771 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
772 * @inode: inode
773 * @lstart: offset of beginning of hole
774 * @lend: offset of last byte of hole
775 *
776 * This function should typically be called before the filesystem
777 * releases resources associated with the freed range (eg. deallocates
778 * blocks). This way, pagecache will always stay logically coherent
779 * with on-disk format, and the filesystem would not have to deal with
780 * situations such as writepage being called for a page that has already
781 * had its underlying blocks deallocated.
782 */
783void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
784{
785	struct address_space *mapping = inode->i_mapping;
786	loff_t unmap_start = round_up(lstart, PAGE_SIZE);
787	loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
788	/*
789	 * This rounding is currently just for example: unmap_mapping_range
790	 * expands its hole outwards, whereas we want it to contract the hole
791	 * inwards.  However, existing callers of truncate_pagecache_range are
792	 * doing their own page rounding first.  Note that unmap_mapping_range
793	 * allows holelen 0 for all, and we allow lend -1 for end of file.
794	 */
795
796	/*
797	 * Unlike in truncate_pagecache, unmap_mapping_range is called only
798	 * once (before truncating pagecache), and without "even_cows" flag:
799	 * hole-punching should not remove private COWed pages from the hole.
800	 */
801	if ((u64)unmap_end > (u64)unmap_start)
802		unmap_mapping_range(mapping, unmap_start,
803				    1 + unmap_end - unmap_start, 0);
804	truncate_inode_pages_range(mapping, lstart, lend);
805}
806EXPORT_SYMBOL(truncate_pagecache_range);
v6.8
  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * mm/truncate.c - code for taking down pages from address_spaces
  4 *
  5 * Copyright (C) 2002, Linus Torvalds
  6 *
  7 * 10Sep2002	Andrew Morton
  8 *		Initial version.
  9 */
 10
 11#include <linux/kernel.h>
 12#include <linux/backing-dev.h>
 13#include <linux/dax.h>
 14#include <linux/gfp.h>
 15#include <linux/mm.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/shmem_fs.h>
 
 
 23#include <linux/rmap.h>
 24#include "internal.h"
 25
 26/*
 27 * Regular page slots are stabilized by the page lock even without the tree
 28 * itself locked.  These unlocked entries need verification under the tree
 29 * lock.
 30 */
 31static inline void __clear_shadow_entry(struct address_space *mapping,
 32				pgoff_t index, void *entry)
 33{
 34	XA_STATE(xas, &mapping->i_pages, index);
 35
 36	xas_set_update(&xas, workingset_update_node);
 37	if (xas_load(&xas) != entry)
 38		return;
 39	xas_store(&xas, NULL);
 40}
 41
 42static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
 43			       void *entry)
 44{
 45	spin_lock(&mapping->host->i_lock);
 46	xa_lock_irq(&mapping->i_pages);
 47	__clear_shadow_entry(mapping, index, entry);
 48	xa_unlock_irq(&mapping->i_pages);
 49	if (mapping_shrinkable(mapping))
 50		inode_add_lru(mapping->host);
 51	spin_unlock(&mapping->host->i_lock);
 52}
 53
 54/*
 55 * Unconditionally remove exceptional entries. Usually called from truncate
 56 * path. Note that the folio_batch may be altered by this function by removing
 57 * exceptional entries similar to what folio_batch_remove_exceptionals() does.
 58 */
 59static void truncate_folio_batch_exceptionals(struct address_space *mapping,
 60				struct folio_batch *fbatch, pgoff_t *indices)
 61{
 62	int i, j;
 63	bool dax;
 64
 65	/* Handled by shmem itself */
 66	if (shmem_mapping(mapping))
 67		return;
 68
 69	for (j = 0; j < folio_batch_count(fbatch); j++)
 70		if (xa_is_value(fbatch->folios[j]))
 71			break;
 72
 73	if (j == folio_batch_count(fbatch))
 74		return;
 75
 76	dax = dax_mapping(mapping);
 77	if (!dax) {
 78		spin_lock(&mapping->host->i_lock);
 79		xa_lock_irq(&mapping->i_pages);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 80	}
 81
 82	for (i = j; i < folio_batch_count(fbatch); i++) {
 83		struct folio *folio = fbatch->folios[i];
 84		pgoff_t index = indices[i];
 85
 86		if (!xa_is_value(folio)) {
 87			fbatch->folios[j++] = folio;
 88			continue;
 89		}
 90
 91		if (unlikely(dax)) {
 92			dax_delete_mapping_entry(mapping, index);
 93			continue;
 94		}
 95
 96		__clear_shadow_entry(mapping, index, folio);
 97	}
 98
 99	if (!dax) {
100		xa_unlock_irq(&mapping->i_pages);
101		if (mapping_shrinkable(mapping))
102			inode_add_lru(mapping->host);
103		spin_unlock(&mapping->host->i_lock);
104	}
105	fbatch->nr = j;
106}
107
108/*
109 * Invalidate exceptional entry if easily possible. This handles exceptional
110 * entries for invalidate_inode_pages().
111 */
112static int invalidate_exceptional_entry(struct address_space *mapping,
113					pgoff_t index, void *entry)
114{
115	/* Handled by shmem itself, or for DAX we do nothing. */
116	if (shmem_mapping(mapping) || dax_mapping(mapping))
117		return 1;
118	clear_shadow_entry(mapping, index, entry);
119	return 1;
120}
121
122/*
123 * Invalidate exceptional entry if clean. This handles exceptional entries for
124 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
125 */
126static int invalidate_exceptional_entry2(struct address_space *mapping,
127					 pgoff_t index, void *entry)
128{
129	/* Handled by shmem itself */
130	if (shmem_mapping(mapping))
131		return 1;
132	if (dax_mapping(mapping))
133		return dax_invalidate_mapping_entry_sync(mapping, index);
134	clear_shadow_entry(mapping, index, entry);
135	return 1;
136}
137
138/**
139 * folio_invalidate - Invalidate part or all of a folio.
140 * @folio: The folio which is affected.
141 * @offset: start of the range to invalidate
142 * @length: length of the range to invalidate
143 *
144 * folio_invalidate() is called when all or part of the folio has become
145 * invalidated by a truncate operation.
146 *
147 * folio_invalidate() does not have to release all buffers, but it must
148 * ensure that no dirty buffer is left outside @offset and that no I/O
149 * is underway against any of the blocks which are outside the truncation
150 * point.  Because the caller is about to free (and possibly reuse) those
151 * blocks on-disk.
152 */
153void folio_invalidate(struct folio *folio, size_t offset, size_t length)
 
154{
155	const struct address_space_operations *aops = folio->mapping->a_ops;
156
157	if (aops->invalidate_folio)
158		aops->invalidate_folio(folio, offset, length);
 
 
 
 
 
159}
160EXPORT_SYMBOL_GPL(folio_invalidate);
161
162/*
163 * If truncate cannot remove the fs-private metadata from the page, the page
164 * becomes orphaned.  It will be left on the LRU and may even be mapped into
165 * user pagetables if we're racing with filemap_fault().
166 *
167 * We need to bail out if page->mapping is no longer equal to the original
168 * mapping.  This happens a) when the VM reclaimed the page while we waited on
169 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
170 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
171 */
172static void truncate_cleanup_folio(struct folio *folio)
 
173{
174	if (folio_mapped(folio))
175		unmap_mapping_folio(folio);
176
177	if (folio_has_private(folio))
178		folio_invalidate(folio, 0, folio_size(folio));
179
180	/*
181	 * Some filesystems seem to re-dirty the page even after
182	 * the VM has canceled the dirty bit (eg ext3 journaling).
183	 * Hence dirty accounting check is placed after invalidation.
184	 */
185	folio_cancel_dirty(folio);
186	folio_clear_mappedtodisk(folio);
187}
188
189int truncate_inode_folio(struct address_space *mapping, struct folio *folio)
190{
191	if (folio->mapping != mapping)
192		return -EIO;
193
194	truncate_cleanup_folio(folio);
195	filemap_remove_folio(folio);
196	return 0;
197}
198
199/*
200 * Handle partial folios.  The folio may be entirely within the
201 * range if a split has raced with us.  If not, we zero the part of the
202 * folio that's within the [start, end] range, and then split the folio if
203 * it's large.  split_page_range() will discard pages which now lie beyond
204 * i_size, and we rely on the caller to discard pages which lie within a
205 * newly created hole.
206 *
207 * Returns false if splitting failed so the caller can avoid
208 * discarding the entire folio which is stubbornly unsplit.
209 */
210bool truncate_inode_partial_folio(struct folio *folio, loff_t start, loff_t end)
 
211{
212	loff_t pos = folio_pos(folio);
213	unsigned int offset, length;
 
 
214
215	if (pos < start)
216		offset = start - pos;
217	else
218		offset = 0;
219	length = folio_size(folio);
220	if (pos + length <= (u64)end)
221		length = length - offset;
222	else
223		length = end + 1 - pos - offset;
224
225	folio_wait_writeback(folio);
226	if (length == folio_size(folio)) {
227		truncate_inode_folio(folio->mapping, folio);
228		return true;
229	}
230
231	/*
232	 * We may be zeroing pages we're about to discard, but it avoids
233	 * doing a complex calculation here, and then doing the zeroing
234	 * anyway if the page split fails.
235	 */
236	folio_zero_range(folio, offset, length);
237
238	if (folio_has_private(folio))
239		folio_invalidate(folio, offset, length);
240	if (!folio_test_large(folio))
241		return true;
242	if (split_folio(folio) == 0)
243		return true;
244	if (folio_test_dirty(folio))
245		return false;
246	truncate_inode_folio(folio->mapping, folio);
247	return true;
248}
249
250/*
251 * Used to get rid of pages on hardware memory corruption.
252 */
253int generic_error_remove_folio(struct address_space *mapping,
254		struct folio *folio)
255{
256	if (!mapping)
257		return -EINVAL;
258	/*
259	 * Only punch for normal data pages for now.
260	 * Handling other types like directories would need more auditing.
261	 */
262	if (!S_ISREG(mapping->host->i_mode))
263		return -EIO;
264	return truncate_inode_folio(mapping, folio);
265}
266EXPORT_SYMBOL(generic_error_remove_folio);
267
268/**
269 * mapping_evict_folio() - Remove an unused folio from the page-cache.
270 * @mapping: The mapping this folio belongs to.
271 * @folio: The folio to remove.
272 *
273 * Safely remove one folio from the page cache.
274 * It only drops clean, unused folios.
275 *
276 * Context: Folio must be locked.
277 * Return: The number of pages successfully removed.
278 */
279long mapping_evict_folio(struct address_space *mapping, struct folio *folio)
280{
281	/* The page may have been truncated before it was locked */
282	if (!mapping)
283		return 0;
284	if (folio_test_dirty(folio) || folio_test_writeback(folio))
285		return 0;
286	/* The refcount will be elevated if any page in the folio is mapped */
287	if (folio_ref_count(folio) >
288			folio_nr_pages(folio) + folio_has_private(folio) + 1)
289		return 0;
290	if (!filemap_release_folio(folio, 0))
291		return 0;
292
293	return remove_mapping(mapping, folio);
294}
295
296/**
297 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
298 * @mapping: mapping to truncate
299 * @lstart: offset from which to truncate
300 * @lend: offset to which to truncate (inclusive)
301 *
302 * Truncate the page cache, removing the pages that are between
303 * specified offsets (and zeroing out partial pages
304 * if lstart or lend + 1 is not page aligned).
305 *
306 * Truncate takes two passes - the first pass is nonblocking.  It will not
307 * block on page locks and it will not block on writeback.  The second pass
308 * will wait.  This is to prevent as much IO as possible in the affected region.
309 * The first pass will remove most pages, so the search cost of the second pass
310 * is low.
311 *
312 * We pass down the cache-hot hint to the page freeing code.  Even if the
313 * mapping is large, it is probably the case that the final pages are the most
314 * recently touched, and freeing happens in ascending file offset order.
315 *
316 * Note that since ->invalidate_folio() accepts range to invalidate
317 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
318 * page aligned properly.
319 */
320void truncate_inode_pages_range(struct address_space *mapping,
321				loff_t lstart, loff_t lend)
322{
323	pgoff_t		start;		/* inclusive */
324	pgoff_t		end;		/* exclusive */
325	struct folio_batch fbatch;
 
 
326	pgoff_t		indices[PAGEVEC_SIZE];
327	pgoff_t		index;
328	int		i;
329	struct folio	*folio;
330	bool		same_folio;
331
332	if (mapping_empty(mapping))
 
333		return;
334
 
 
 
 
335	/*
336	 * 'start' and 'end' always covers the range of pages to be fully
337	 * truncated. Partial pages are covered with 'partial_start' at the
338	 * start of the range and 'partial_end' at the end of the range.
339	 * Note that 'end' is exclusive while 'lend' is inclusive.
340	 */
341	start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
342	if (lend == -1)
343		/*
344		 * lend == -1 indicates end-of-file so we have to set 'end'
345		 * to the highest possible pgoff_t and since the type is
346		 * unsigned we're using -1.
347		 */
348		end = -1;
349	else
350		end = (lend + 1) >> PAGE_SHIFT;
351
352	folio_batch_init(&fbatch);
353	index = start;
354	while (index < end && find_lock_entries(mapping, &index, end - 1,
355			&fbatch, indices)) {
356		truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
357		for (i = 0; i < folio_batch_count(&fbatch); i++)
358			truncate_cleanup_folio(fbatch.folios[i]);
359		delete_from_page_cache_batch(mapping, &fbatch);
360		for (i = 0; i < folio_batch_count(&fbatch); i++)
361			folio_unlock(fbatch.folios[i]);
362		folio_batch_release(&fbatch);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
363		cond_resched();
 
364	}
365
366	same_folio = (lstart >> PAGE_SHIFT) == (lend >> PAGE_SHIFT);
367	folio = __filemap_get_folio(mapping, lstart >> PAGE_SHIFT, FGP_LOCK, 0);
368	if (!IS_ERR(folio)) {
369		same_folio = lend < folio_pos(folio) + folio_size(folio);
370		if (!truncate_inode_partial_folio(folio, lstart, lend)) {
371			start = folio_next_index(folio);
372			if (same_folio)
373				end = folio->index;
 
 
 
 
 
 
 
 
 
374		}
375		folio_unlock(folio);
376		folio_put(folio);
377		folio = NULL;
378	}
379
380	if (!same_folio) {
381		folio = __filemap_get_folio(mapping, lend >> PAGE_SHIFT,
382						FGP_LOCK, 0);
383		if (!IS_ERR(folio)) {
384			if (!truncate_inode_partial_folio(folio, lstart, lend))
385				end = folio->index;
386			folio_unlock(folio);
387			folio_put(folio);
 
 
388		}
389	}
 
 
 
 
 
 
390
391	index = start;
392	while (index < end) {
393		cond_resched();
394		if (!find_get_entries(mapping, &index, end - 1, &fbatch,
395				indices)) {
396			/* If all gone from start onwards, we're done */
397			if (index == start)
398				break;
399			/* Otherwise restart to make sure all gone */
400			index = start;
401			continue;
402		}
403
404		for (i = 0; i < folio_batch_count(&fbatch); i++) {
405			struct folio *folio = fbatch.folios[i];
 
 
 
 
 
406
407			/* We rely upon deletion not changing page->index */
 
 
 
 
 
 
408
409			if (xa_is_value(folio))
 
410				continue;
 
411
412			folio_lock(folio);
413			VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
414			folio_wait_writeback(folio);
415			truncate_inode_folio(mapping, folio);
416			folio_unlock(folio);
417		}
418		truncate_folio_batch_exceptionals(mapping, &fbatch, indices);
419		folio_batch_release(&fbatch);
 
420	}
 
421}
422EXPORT_SYMBOL(truncate_inode_pages_range);
423
424/**
425 * truncate_inode_pages - truncate *all* the pages from an offset
426 * @mapping: mapping to truncate
427 * @lstart: offset from which to truncate
428 *
429 * Called under (and serialised by) inode->i_rwsem and
430 * mapping->invalidate_lock.
431 *
432 * Note: When this function returns, there can be a page in the process of
433 * deletion (inside __filemap_remove_folio()) in the specified range.  Thus
434 * mapping->nrpages can be non-zero when this function returns even after
435 * truncation of the whole mapping.
436 */
437void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
438{
439	truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
440}
441EXPORT_SYMBOL(truncate_inode_pages);
442
443/**
444 * truncate_inode_pages_final - truncate *all* pages before inode dies
445 * @mapping: mapping to truncate
446 *
447 * Called under (and serialized by) inode->i_rwsem.
448 *
449 * Filesystems have to use this in the .evict_inode path to inform the
450 * VM that this is the final truncate and the inode is going away.
451 */
452void truncate_inode_pages_final(struct address_space *mapping)
453{
 
 
 
454	/*
455	 * Page reclaim can not participate in regular inode lifetime
456	 * management (can't call iput()) and thus can race with the
457	 * inode teardown.  Tell it when the address space is exiting,
458	 * so that it does not install eviction information after the
459	 * final truncate has begun.
460	 */
461	mapping_set_exiting(mapping);
462
463	if (!mapping_empty(mapping)) {
 
 
 
 
 
 
 
 
 
464		/*
465		 * As truncation uses a lockless tree lookup, cycle
466		 * the tree lock to make sure any ongoing tree
467		 * modification that does not see AS_EXITING is
468		 * completed before starting the final truncate.
469		 */
470		xa_lock_irq(&mapping->i_pages);
471		xa_unlock_irq(&mapping->i_pages);
 
 
472	}
473
474	truncate_inode_pages(mapping, 0);
475}
476EXPORT_SYMBOL(truncate_inode_pages_final);
477
478/**
479 * mapping_try_invalidate - Invalidate all the evictable folios of one inode
480 * @mapping: the address_space which holds the folios to invalidate
481 * @start: the offset 'from' which to invalidate
482 * @end: the offset 'to' which to invalidate (inclusive)
483 * @nr_failed: How many folio invalidations failed
484 *
485 * This function is similar to invalidate_mapping_pages(), except that it
486 * returns the number of folios which could not be evicted in @nr_failed.
 
 
 
 
487 */
488unsigned long mapping_try_invalidate(struct address_space *mapping,
489		pgoff_t start, pgoff_t end, unsigned long *nr_failed)
490{
491	pgoff_t indices[PAGEVEC_SIZE];
492	struct folio_batch fbatch;
493	pgoff_t index = start;
494	unsigned long ret;
495	unsigned long count = 0;
496	int i;
497
498	folio_batch_init(&fbatch);
499	while (find_lock_entries(mapping, &index, end, &fbatch, indices)) {
500		for (i = 0; i < folio_batch_count(&fbatch); i++) {
501			struct folio *folio = fbatch.folios[i];
502
503			/* We rely upon deletion not changing folio->index */
504
505			if (xa_is_value(folio)) {
506				count += invalidate_exceptional_entry(mapping,
507							     indices[i], folio);
 
 
 
 
508				continue;
509			}
510
511			ret = mapping_evict_folio(mapping, folio);
512			folio_unlock(folio);
 
 
 
513			/*
514			 * Invalidation is a hint that the folio is no longer
515			 * of interest and try to speed up its reclaim.
516			 */
517			if (!ret) {
518				deactivate_file_folio(folio);
519				/* Likely in the lru cache of a remote CPU */
520				if (nr_failed)
521					(*nr_failed)++;
522			}
523			count += ret;
524		}
525		folio_batch_remove_exceptionals(&fbatch);
526		folio_batch_release(&fbatch);
527		cond_resched();
 
528	}
529	return count;
530}
531
532/**
533 * invalidate_mapping_pages - Invalidate all clean, unlocked cache of one inode
534 * @mapping: the address_space which holds the cache to invalidate
535 * @start: the offset 'from' which to invalidate
536 * @end: the offset 'to' which to invalidate (inclusive)
537 *
538 * This function removes pages that are clean, unmapped and unlocked,
539 * as well as shadow entries. It will not block on IO activity.
540 *
541 * If you want to remove all the pages of one inode, regardless of
542 * their use and writeback state, use truncate_inode_pages().
543 *
544 * Return: The number of indices that had their contents invalidated
545 */
546unsigned long invalidate_mapping_pages(struct address_space *mapping,
547		pgoff_t start, pgoff_t end)
548{
549	return mapping_try_invalidate(mapping, start, end, NULL);
550}
551EXPORT_SYMBOL(invalidate_mapping_pages);
552
553/*
554 * This is like mapping_evict_folio(), except it ignores the folio's
555 * refcount.  We do this because invalidate_inode_pages2() needs stronger
556 * invalidation guarantees, and cannot afford to leave folios behind because
557 * shrink_page_list() has a temp ref on them, or because they're transiently
558 * sitting in the folio_add_lru() caches.
559 */
560static int invalidate_complete_folio2(struct address_space *mapping,
561					struct folio *folio)
562{
563	if (folio->mapping != mapping)
 
 
564		return 0;
565
566	if (!filemap_release_folio(folio, GFP_KERNEL))
567		return 0;
568
569	spin_lock(&mapping->host->i_lock);
570	xa_lock_irq(&mapping->i_pages);
571	if (folio_test_dirty(folio))
572		goto failed;
573
574	BUG_ON(folio_has_private(folio));
575	__filemap_remove_folio(folio, NULL);
576	xa_unlock_irq(&mapping->i_pages);
577	if (mapping_shrinkable(mapping))
578		inode_add_lru(mapping->host);
579	spin_unlock(&mapping->host->i_lock);
580
581	filemap_free_folio(mapping, folio);
582	return 1;
583failed:
584	xa_unlock_irq(&mapping->i_pages);
585	spin_unlock(&mapping->host->i_lock);
586	return 0;
587}
588
589static int folio_launder(struct address_space *mapping, struct folio *folio)
590{
591	if (!folio_test_dirty(folio))
592		return 0;
593	if (folio->mapping != mapping || mapping->a_ops->launder_folio == NULL)
594		return 0;
595	return mapping->a_ops->launder_folio(folio);
596}
597
598/**
599 * invalidate_inode_pages2_range - remove range of pages from an address_space
600 * @mapping: the address_space
601 * @start: the page offset 'from' which to invalidate
602 * @end: the page offset 'to' which to invalidate (inclusive)
603 *
604 * Any pages which are found to be mapped into pagetables are unmapped prior to
605 * invalidation.
606 *
607 * Return: -EBUSY if any pages could not be invalidated.
608 */
609int invalidate_inode_pages2_range(struct address_space *mapping,
610				  pgoff_t start, pgoff_t end)
611{
612	pgoff_t indices[PAGEVEC_SIZE];
613	struct folio_batch fbatch;
614	pgoff_t index;
615	int i;
616	int ret = 0;
617	int ret2 = 0;
618	int did_range_unmap = 0;
619
620	if (mapping_empty(mapping))
621		return 0;
 
 
 
 
 
 
 
 
 
 
 
622
623	folio_batch_init(&fbatch);
624	index = start;
625	while (find_get_entries(mapping, &index, end, &fbatch, indices)) {
626		for (i = 0; i < folio_batch_count(&fbatch); i++) {
627			struct folio *folio = fbatch.folios[i];
628
629			/* We rely upon deletion not changing folio->index */
630
631			if (xa_is_value(folio)) {
632				if (!invalidate_exceptional_entry2(mapping,
633						indices[i], folio))
634					ret = -EBUSY;
635				continue;
636			}
637
638			if (!did_range_unmap && folio_mapped(folio)) {
639				/*
640				 * If folio is mapped, before taking its lock,
641				 * zap the rest of the file in one hit.
642				 */
643				unmap_mapping_pages(mapping, indices[i],
644						(1 + end - indices[i]), false);
645				did_range_unmap = 1;
646			}
647
648			folio_lock(folio);
649			if (unlikely(folio->mapping != mapping)) {
650				folio_unlock(folio);
651				continue;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
652			}
653			VM_BUG_ON_FOLIO(!folio_contains(folio, indices[i]), folio);
654			folio_wait_writeback(folio);
655
656			if (folio_mapped(folio))
657				unmap_mapping_folio(folio);
658			BUG_ON(folio_mapped(folio));
659
660			ret2 = folio_launder(mapping, folio);
661			if (ret2 == 0) {
662				if (!invalidate_complete_folio2(mapping, folio))
663					ret2 = -EBUSY;
664			}
665			if (ret2 < 0)
666				ret = ret2;
667			folio_unlock(folio);
668		}
669		folio_batch_remove_exceptionals(&fbatch);
670		folio_batch_release(&fbatch);
671		cond_resched();
 
672	}
673	/*
674	 * For DAX we invalidate page tables after invalidating page cache.  We
675	 * could invalidate page tables while invalidating each entry however
676	 * that would be expensive. And doing range unmapping before doesn't
677	 * work as we have no cheap way to find whether page cache entry didn't
678	 * get remapped later.
679	 */
680	if (dax_mapping(mapping)) {
681		unmap_mapping_pages(mapping, start, end - start + 1, false);
682	}
683	return ret;
684}
685EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
686
687/**
688 * invalidate_inode_pages2 - remove all pages from an address_space
689 * @mapping: the address_space
690 *
691 * Any pages which are found to be mapped into pagetables are unmapped prior to
692 * invalidation.
693 *
694 * Return: -EBUSY if any pages could not be invalidated.
695 */
696int invalidate_inode_pages2(struct address_space *mapping)
697{
698	return invalidate_inode_pages2_range(mapping, 0, -1);
699}
700EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
701
702/**
703 * truncate_pagecache - unmap and remove pagecache that has been truncated
704 * @inode: inode
705 * @newsize: new file size
706 *
707 * inode's new i_size must already be written before truncate_pagecache
708 * is called.
709 *
710 * This function should typically be called before the filesystem
711 * releases resources associated with the freed range (eg. deallocates
712 * blocks). This way, pagecache will always stay logically coherent
713 * with on-disk format, and the filesystem would not have to deal with
714 * situations such as writepage being called for a page that has already
715 * had its underlying blocks deallocated.
716 */
717void truncate_pagecache(struct inode *inode, loff_t newsize)
718{
719	struct address_space *mapping = inode->i_mapping;
720	loff_t holebegin = round_up(newsize, PAGE_SIZE);
721
722	/*
723	 * unmap_mapping_range is called twice, first simply for
724	 * efficiency so that truncate_inode_pages does fewer
725	 * single-page unmaps.  However after this first call, and
726	 * before truncate_inode_pages finishes, it is possible for
727	 * private pages to be COWed, which remain after
728	 * truncate_inode_pages finishes, hence the second
729	 * unmap_mapping_range call must be made for correctness.
730	 */
731	unmap_mapping_range(mapping, holebegin, 0, 1);
732	truncate_inode_pages(mapping, newsize);
733	unmap_mapping_range(mapping, holebegin, 0, 1);
734}
735EXPORT_SYMBOL(truncate_pagecache);
736
737/**
738 * truncate_setsize - update inode and pagecache for a new file size
739 * @inode: inode
740 * @newsize: new file size
741 *
742 * truncate_setsize updates i_size and performs pagecache truncation (if
743 * necessary) to @newsize. It will be typically be called from the filesystem's
744 * setattr function when ATTR_SIZE is passed in.
745 *
746 * Must be called with a lock serializing truncates and writes (generally
747 * i_rwsem but e.g. xfs uses a different lock) and before all filesystem
748 * specific block truncation has been performed.
749 */
750void truncate_setsize(struct inode *inode, loff_t newsize)
751{
752	loff_t oldsize = inode->i_size;
753
754	i_size_write(inode, newsize);
755	if (newsize > oldsize)
756		pagecache_isize_extended(inode, oldsize, newsize);
757	truncate_pagecache(inode, newsize);
758}
759EXPORT_SYMBOL(truncate_setsize);
760
761/**
762 * pagecache_isize_extended - update pagecache after extension of i_size
763 * @inode:	inode for which i_size was extended
764 * @from:	original inode size
765 * @to:		new inode size
766 *
767 * Handle extension of inode size either caused by extending truncate or by
768 * write starting after current i_size. We mark the page straddling current
769 * i_size RO so that page_mkwrite() is called on the nearest write access to
770 * the page.  This way filesystem can be sure that page_mkwrite() is called on
771 * the page before user writes to the page via mmap after the i_size has been
772 * changed.
773 *
774 * The function must be called after i_size is updated so that page fault
775 * coming after we unlock the page will already see the new i_size.
776 * The function must be called while we still hold i_rwsem - this not only
777 * makes sure i_size is stable but also that userspace cannot observe new
778 * i_size value before we are prepared to store mmap writes at new inode size.
779 */
780void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
781{
782	int bsize = i_blocksize(inode);
783	loff_t rounded_from;
784	struct page *page;
785	pgoff_t index;
786
787	WARN_ON(to > inode->i_size);
788
789	if (from >= to || bsize == PAGE_SIZE)
790		return;
791	/* Page straddling @from will not have any hole block created? */
792	rounded_from = round_up(from, bsize);
793	if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
794		return;
795
796	index = from >> PAGE_SHIFT;
797	page = find_lock_page(inode->i_mapping, index);
798	/* Page not cached? Nothing to do */
799	if (!page)
800		return;
801	/*
802	 * See clear_page_dirty_for_io() for details why set_page_dirty()
803	 * is needed.
804	 */
805	if (page_mkclean(page))
806		set_page_dirty(page);
807	unlock_page(page);
808	put_page(page);
809}
810EXPORT_SYMBOL(pagecache_isize_extended);
811
812/**
813 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
814 * @inode: inode
815 * @lstart: offset of beginning of hole
816 * @lend: offset of last byte of hole
817 *
818 * This function should typically be called before the filesystem
819 * releases resources associated with the freed range (eg. deallocates
820 * blocks). This way, pagecache will always stay logically coherent
821 * with on-disk format, and the filesystem would not have to deal with
822 * situations such as writepage being called for a page that has already
823 * had its underlying blocks deallocated.
824 */
825void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
826{
827	struct address_space *mapping = inode->i_mapping;
828	loff_t unmap_start = round_up(lstart, PAGE_SIZE);
829	loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
830	/*
831	 * This rounding is currently just for example: unmap_mapping_range
832	 * expands its hole outwards, whereas we want it to contract the hole
833	 * inwards.  However, existing callers of truncate_pagecache_range are
834	 * doing their own page rounding first.  Note that unmap_mapping_range
835	 * allows holelen 0 for all, and we allow lend -1 for end of file.
836	 */
837
838	/*
839	 * Unlike in truncate_pagecache, unmap_mapping_range is called only
840	 * once (before truncating pagecache), and without "even_cows" flag:
841	 * hole-punching should not remove private COWed pages from the hole.
842	 */
843	if ((u64)unmap_end > (u64)unmap_start)
844		unmap_mapping_range(mapping, unmap_start,
845				    1 + unmap_end - unmap_start, 0);
846	truncate_inode_pages_range(mapping, lstart, lend);
847}
848EXPORT_SYMBOL(truncate_pagecache_range);