Linux Audio

Check our new training course

Loading...
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);
v5.14.15
  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/buffer_head.h>	/* grr. try_to_release_page,
 23				   do_invalidatepage */
 24#include <linux/shmem_fs.h>
 25#include <linux/cleancache.h>
 26#include <linux/rmap.h>
 27#include "internal.h"
 28
 29/*
 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
 32 * lock.
 33 */
 34static inline void __clear_shadow_entry(struct address_space *mapping,
 35				pgoff_t index, void *entry)
 36{
 37	XA_STATE(xas, &mapping->i_pages, index);
 38
 39	xas_set_update(&xas, workingset_update_node);
 40	if (xas_load(&xas) != entry)
 41		return;
 42	xas_store(&xas, NULL);
 43}
 44
 45static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
 46			       void *entry)
 47{
 
 48	xa_lock_irq(&mapping->i_pages);
 49	__clear_shadow_entry(mapping, index, entry);
 50	xa_unlock_irq(&mapping->i_pages);
 
 
 
 51}
 52
 53/*
 54 * Unconditionally remove exceptional entries. Usually called from truncate
 55 * path. Note that the pagevec may be altered by this function by removing
 56 * exceptional entries similar to what pagevec_remove_exceptionals does.
 57 */
 58static void truncate_exceptional_pvec_entries(struct address_space *mapping,
 59				struct pagevec *pvec, pgoff_t *indices)
 60{
 61	int i, j;
 62	bool dax;
 63
 64	/* Handled by shmem itself */
 65	if (shmem_mapping(mapping))
 66		return;
 67
 68	for (j = 0; j < pagevec_count(pvec); j++)
 69		if (xa_is_value(pvec->pages[j]))
 70			break;
 71
 72	if (j == pagevec_count(pvec))
 73		return;
 74
 75	dax = dax_mapping(mapping);
 76	if (!dax)
 
 77		xa_lock_irq(&mapping->i_pages);
 
 78
 79	for (i = j; i < pagevec_count(pvec); i++) {
 80		struct page *page = pvec->pages[i];
 81		pgoff_t index = indices[i];
 82
 83		if (!xa_is_value(page)) {
 84			pvec->pages[j++] = page;
 85			continue;
 86		}
 87
 88		if (unlikely(dax)) {
 89			dax_delete_mapping_entry(mapping, index);
 90			continue;
 91		}
 92
 93		__clear_shadow_entry(mapping, index, page);
 94	}
 95
 96	if (!dax)
 97		xa_unlock_irq(&mapping->i_pages);
 98	pvec->nr = j;
 
 
 
 
 99}
100
101/*
102 * Invalidate exceptional entry if easily possible. This handles exceptional
103 * entries for invalidate_inode_pages().
104 */
105static int invalidate_exceptional_entry(struct address_space *mapping,
106					pgoff_t index, void *entry)
107{
108	/* Handled by shmem itself, or for DAX we do nothing. */
109	if (shmem_mapping(mapping) || dax_mapping(mapping))
110		return 1;
111	clear_shadow_entry(mapping, index, entry);
112	return 1;
113}
114
115/*
116 * Invalidate exceptional entry if clean. This handles exceptional entries for
117 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
118 */
119static int invalidate_exceptional_entry2(struct address_space *mapping,
120					 pgoff_t index, void *entry)
121{
122	/* Handled by shmem itself */
123	if (shmem_mapping(mapping))
124		return 1;
125	if (dax_mapping(mapping))
126		return dax_invalidate_mapping_entry_sync(mapping, index);
127	clear_shadow_entry(mapping, index, entry);
128	return 1;
129}
130
131/**
132 * do_invalidatepage - invalidate part or all of a page
133 * @page: the page which is affected
134 * @offset: start of the range to invalidate
135 * @length: length of the range to invalidate
136 *
137 * do_invalidatepage() is called when all or part of the page has become
138 * invalidated by a truncate operation.
139 *
140 * do_invalidatepage() does not have to release all buffers, but it must
141 * ensure that no dirty buffer is left outside @offset and that no I/O
142 * is underway against any of the blocks which are outside the truncation
143 * point.  Because the caller is about to free (and possibly reuse) those
144 * blocks on-disk.
145 */
146void do_invalidatepage(struct page *page, unsigned int offset,
147		       unsigned int length)
148{
149	void (*invalidatepage)(struct page *, unsigned int, unsigned int);
150
151	invalidatepage = page->mapping->a_ops->invalidatepage;
152#ifdef CONFIG_BLOCK
153	if (!invalidatepage)
154		invalidatepage = block_invalidatepage;
155#endif
156	if (invalidatepage)
157		(*invalidatepage)(page, offset, length);
158}
 
159
160/*
161 * If truncate cannot remove the fs-private metadata from the page, the page
162 * becomes orphaned.  It will be left on the LRU and may even be mapped into
163 * user pagetables if we're racing with filemap_fault().
164 *
165 * We need to bail out if page->mapping is no longer equal to the original
166 * mapping.  This happens a) when the VM reclaimed the page while we waited on
167 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
168 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
169 */
170static void truncate_cleanup_page(struct page *page)
171{
172	if (page_mapped(page))
173		unmap_mapping_page(page);
174
175	if (page_has_private(page))
176		do_invalidatepage(page, 0, thp_size(page));
177
178	/*
179	 * Some filesystems seem to re-dirty the page even after
180	 * the VM has canceled the dirty bit (eg ext3 journaling).
181	 * Hence dirty accounting check is placed after invalidation.
182	 */
183	cancel_dirty_page(page);
184	ClearPageMappedToDisk(page);
 
 
 
 
 
 
 
 
 
 
185}
186
187/*
188 * This is for invalidate_mapping_pages().  That function can be called at
189 * any time, and is not supposed to throw away dirty pages.  But pages can
190 * be marked dirty at any time too, so use remove_mapping which safely
191 * discards clean, unused pages.
192 *
193 * Returns non-zero if the page was successfully invalidated.
 
 
 
194 */
195static int
196invalidate_complete_page(struct address_space *mapping, struct page *page)
197{
198	int ret;
 
199
200	if (page->mapping != mapping)
201		return 0;
 
 
 
 
 
 
 
202
203	if (page_has_private(page) && !try_to_release_page(page, 0))
204		return 0;
 
 
 
205
206	ret = remove_mapping(mapping, page);
 
 
 
 
 
207
208	return ret;
209}
210
211int truncate_inode_page(struct address_space *mapping, struct page *page)
212{
213	VM_BUG_ON_PAGE(PageTail(page), page);
214
215	if (page->mapping != mapping)
216		return -EIO;
217
218	truncate_cleanup_page(page);
219	delete_from_page_cache(page);
220	return 0;
221}
222
223/*
224 * Used to get rid of pages on hardware memory corruption.
225 */
226int generic_error_remove_page(struct address_space *mapping, struct page *page)
 
227{
228	if (!mapping)
229		return -EINVAL;
230	/*
231	 * Only punch for normal data pages for now.
232	 * Handling other types like directories would need more auditing.
233	 */
234	if (!S_ISREG(mapping->host->i_mode))
235		return -EIO;
236	return truncate_inode_page(mapping, page);
237}
238EXPORT_SYMBOL(generic_error_remove_page);
239
240/*
241 * Safely invalidate one page from its pagecache mapping.
242 * It only drops clean, unused pages. The page must be locked.
 
 
 
 
243 *
244 * Returns 1 if the page is successfully invalidated, otherwise 0.
 
245 */
246int invalidate_inode_page(struct page *page)
247{
248	struct address_space *mapping = page_mapping(page);
249	if (!mapping)
250		return 0;
251	if (PageDirty(page) || PageWriteback(page))
 
 
 
 
252		return 0;
253	if (page_mapped(page))
254		return 0;
255	return invalidate_complete_page(mapping, page);
 
256}
257
258/**
259 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
260 * @mapping: mapping to truncate
261 * @lstart: offset from which to truncate
262 * @lend: offset to which to truncate (inclusive)
263 *
264 * Truncate the page cache, removing the pages that are between
265 * specified offsets (and zeroing out partial pages
266 * if lstart or lend + 1 is not page aligned).
267 *
268 * Truncate takes two passes - the first pass is nonblocking.  It will not
269 * block on page locks and it will not block on writeback.  The second pass
270 * will wait.  This is to prevent as much IO as possible in the affected region.
271 * The first pass will remove most pages, so the search cost of the second pass
272 * is low.
273 *
274 * We pass down the cache-hot hint to the page freeing code.  Even if the
275 * mapping is large, it is probably the case that the final pages are the most
276 * recently touched, and freeing happens in ascending file offset order.
277 *
278 * Note that since ->invalidatepage() accepts range to invalidate
279 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
280 * page aligned properly.
281 */
282void truncate_inode_pages_range(struct address_space *mapping,
283				loff_t lstart, loff_t lend)
284{
285	pgoff_t		start;		/* inclusive */
286	pgoff_t		end;		/* exclusive */
287	unsigned int	partial_start;	/* inclusive */
288	unsigned int	partial_end;	/* exclusive */
289	struct pagevec	pvec;
290	pgoff_t		indices[PAGEVEC_SIZE];
291	pgoff_t		index;
292	int		i;
 
 
293
294	if (mapping_empty(mapping))
295		goto out;
296
297	/* Offsets within partial pages */
298	partial_start = lstart & (PAGE_SIZE - 1);
299	partial_end = (lend + 1) & (PAGE_SIZE - 1);
300
301	/*
302	 * 'start' and 'end' always covers the range of pages to be fully
303	 * truncated. Partial pages are covered with 'partial_start' at the
304	 * start of the range and 'partial_end' at the end of the range.
305	 * Note that 'end' is exclusive while 'lend' is inclusive.
306	 */
307	start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
308	if (lend == -1)
309		/*
310		 * lend == -1 indicates end-of-file so we have to set 'end'
311		 * to the highest possible pgoff_t and since the type is
312		 * unsigned we're using -1.
313		 */
314		end = -1;
315	else
316		end = (lend + 1) >> PAGE_SHIFT;
317
318	pagevec_init(&pvec);
319	index = start;
320	while (index < end && find_lock_entries(mapping, index, end - 1,
321			&pvec, indices)) {
322		index = indices[pagevec_count(&pvec) - 1] + 1;
323		truncate_exceptional_pvec_entries(mapping, &pvec, indices);
324		for (i = 0; i < pagevec_count(&pvec); i++)
325			truncate_cleanup_page(pvec.pages[i]);
326		delete_from_page_cache_batch(mapping, &pvec);
327		for (i = 0; i < pagevec_count(&pvec); i++)
328			unlock_page(pvec.pages[i]);
329		pagevec_release(&pvec);
330		cond_resched();
331	}
332
333	if (partial_start) {
334		struct page *page = find_lock_page(mapping, start - 1);
335		if (page) {
336			unsigned int top = PAGE_SIZE;
337			if (start > end) {
338				/* Truncation within a single page */
339				top = partial_end;
340				partial_end = 0;
341			}
342			wait_on_page_writeback(page);
343			zero_user_segment(page, partial_start, top);
344			cleancache_invalidate_page(mapping, page);
345			if (page_has_private(page))
346				do_invalidatepage(page, partial_start,
347						  top - partial_start);
348			unlock_page(page);
349			put_page(page);
350		}
 
 
 
351	}
352	if (partial_end) {
353		struct page *page = find_lock_page(mapping, end);
354		if (page) {
355			wait_on_page_writeback(page);
356			zero_user_segment(page, 0, partial_end);
357			cleancache_invalidate_page(mapping, page);
358			if (page_has_private(page))
359				do_invalidatepage(page, 0,
360						  partial_end);
361			unlock_page(page);
362			put_page(page);
363		}
364	}
365	/*
366	 * If the truncation happened within a single page no pages
367	 * will be released, just zeroed, so we can bail out now.
368	 */
369	if (start >= end)
370		goto out;
371
372	index = start;
373	for ( ; ; ) {
374		cond_resched();
375		if (!find_get_entries(mapping, index, end - 1, &pvec,
376				indices)) {
377			/* If all gone from start onwards, we're done */
378			if (index == start)
379				break;
380			/* Otherwise restart to make sure all gone */
381			index = start;
382			continue;
383		}
384
385		for (i = 0; i < pagevec_count(&pvec); i++) {
386			struct page *page = pvec.pages[i];
387
388			/* We rely upon deletion not changing page->index */
389			index = indices[i];
390
391			if (xa_is_value(page))
392				continue;
393
394			lock_page(page);
395			WARN_ON(page_to_index(page) != index);
396			wait_on_page_writeback(page);
397			truncate_inode_page(mapping, page);
398			unlock_page(page);
399		}
400		truncate_exceptional_pvec_entries(mapping, &pvec, indices);
401		pagevec_release(&pvec);
402		index++;
403	}
404
405out:
406	cleancache_invalidate_inode(mapping);
407}
408EXPORT_SYMBOL(truncate_inode_pages_range);
409
410/**
411 * truncate_inode_pages - truncate *all* the pages from an offset
412 * @mapping: mapping to truncate
413 * @lstart: offset from which to truncate
414 *
415 * Called under (and serialised by) inode->i_mutex.
 
416 *
417 * Note: When this function returns, there can be a page in the process of
418 * deletion (inside __delete_from_page_cache()) in the specified range.  Thus
419 * mapping->nrpages can be non-zero when this function returns even after
420 * truncation of the whole mapping.
421 */
422void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
423{
424	truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
425}
426EXPORT_SYMBOL(truncate_inode_pages);
427
428/**
429 * truncate_inode_pages_final - truncate *all* pages before inode dies
430 * @mapping: mapping to truncate
431 *
432 * Called under (and serialized by) inode->i_mutex.
433 *
434 * Filesystems have to use this in the .evict_inode path to inform the
435 * VM that this is the final truncate and the inode is going away.
436 */
437void truncate_inode_pages_final(struct address_space *mapping)
438{
439	/*
440	 * Page reclaim can not participate in regular inode lifetime
441	 * management (can't call iput()) and thus can race with the
442	 * inode teardown.  Tell it when the address space is exiting,
443	 * so that it does not install eviction information after the
444	 * final truncate has begun.
445	 */
446	mapping_set_exiting(mapping);
447
448	if (!mapping_empty(mapping)) {
449		/*
450		 * As truncation uses a lockless tree lookup, cycle
451		 * the tree lock to make sure any ongoing tree
452		 * modification that does not see AS_EXITING is
453		 * completed before starting the final truncate.
454		 */
455		xa_lock_irq(&mapping->i_pages);
456		xa_unlock_irq(&mapping->i_pages);
457	}
458
459	/*
460	 * Cleancache needs notification even if there are no pages or shadow
461	 * entries.
462	 */
463	truncate_inode_pages(mapping, 0);
464}
465EXPORT_SYMBOL(truncate_inode_pages_final);
466
467static unsigned long __invalidate_mapping_pages(struct address_space *mapping,
468		pgoff_t start, pgoff_t end, unsigned long *nr_pagevec)
 
 
 
 
 
 
 
 
 
 
469{
470	pgoff_t indices[PAGEVEC_SIZE];
471	struct pagevec pvec;
472	pgoff_t index = start;
473	unsigned long ret;
474	unsigned long count = 0;
475	int i;
476
477	pagevec_init(&pvec);
478	while (find_lock_entries(mapping, index, end, &pvec, indices)) {
479		for (i = 0; i < pagevec_count(&pvec); i++) {
480			struct page *page = pvec.pages[i];
481
482			/* We rely upon deletion not changing page->index */
483			index = indices[i];
484
485			if (xa_is_value(page)) {
486				invalidate_exceptional_entry(mapping, index,
487							     page);
488				continue;
489			}
490			index += thp_nr_pages(page) - 1;
491
492			ret = invalidate_inode_page(page);
493			unlock_page(page);
494			/*
495			 * Invalidation is a hint that the page is no longer
496			 * of interest and try to speed up its reclaim.
497			 */
498			if (!ret) {
499				deactivate_file_page(page);
500				/* It is likely on the pagevec of a remote CPU */
501				if (nr_pagevec)
502					(*nr_pagevec)++;
503			}
504			count += ret;
505		}
506		pagevec_remove_exceptionals(&pvec);
507		pagevec_release(&pvec);
508		cond_resched();
509		index++;
510	}
511	return count;
512}
513
514/**
515 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
516 * @mapping: the address_space which holds the pages to invalidate
517 * @start: the offset 'from' which to invalidate
518 * @end: the offset 'to' which to invalidate (inclusive)
519 *
520 * This function only removes the unlocked pages, if you want to
521 * remove all the pages of one inode, you must call truncate_inode_pages.
522 *
523 * invalidate_mapping_pages() will not block on IO activity. It will not
524 * invalidate pages which are dirty, locked, under writeback or mapped into
525 * pagetables.
526 *
527 * Return: the number of the pages that were invalidated
528 */
529unsigned long invalidate_mapping_pages(struct address_space *mapping,
530		pgoff_t start, pgoff_t end)
531{
532	return __invalidate_mapping_pages(mapping, start, end, NULL);
533}
534EXPORT_SYMBOL(invalidate_mapping_pages);
535
536/**
537 * invalidate_mapping_pagevec - Invalidate all the unlocked pages of one inode
538 * @mapping: the address_space which holds the pages to invalidate
539 * @start: the offset 'from' which to invalidate
540 * @end: the offset 'to' which to invalidate (inclusive)
541 * @nr_pagevec: invalidate failed page number for caller
542 *
543 * This helper is similar to invalidate_mapping_pages(), except that it accounts
544 * for pages that are likely on a pagevec and counts them in @nr_pagevec, which
545 * will be used by the caller.
546 */
547void invalidate_mapping_pagevec(struct address_space *mapping,
548		pgoff_t start, pgoff_t end, unsigned long *nr_pagevec)
549{
550	__invalidate_mapping_pages(mapping, start, end, nr_pagevec);
551}
552
553/*
554 * This is like invalidate_complete_page(), except it ignores the page's
555 * refcount.  We do this because invalidate_inode_pages2() needs stronger
556 * invalidation guarantees, and cannot afford to leave pages behind because
557 * shrink_page_list() has a temp ref on them, or because they're transiently
558 * sitting in the lru_cache_add() pagevecs.
559 */
560static int
561invalidate_complete_page2(struct address_space *mapping, struct page *page)
562{
563	unsigned long flags;
564
565	if (page->mapping != mapping)
566		return 0;
567
568	if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
569		return 0;
570
571	xa_lock_irqsave(&mapping->i_pages, flags);
572	if (PageDirty(page))
 
573		goto failed;
574
575	BUG_ON(page_has_private(page));
576	__delete_from_page_cache(page, NULL);
577	xa_unlock_irqrestore(&mapping->i_pages, flags);
578
579	if (mapping->a_ops->freepage)
580		mapping->a_ops->freepage(page);
581
582	put_page(page);	/* pagecache ref */
583	return 1;
584failed:
585	xa_unlock_irqrestore(&mapping->i_pages, flags);
 
586	return 0;
587}
588
589static int do_launder_page(struct address_space *mapping, struct page *page)
590{
591	if (!PageDirty(page))
592		return 0;
593	if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
594		return 0;
595	return mapping->a_ops->launder_page(page);
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 pagevec pvec;
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		goto out;
622
623	pagevec_init(&pvec);
624	index = start;
625	while (find_get_entries(mapping, index, end, &pvec, indices)) {
626		for (i = 0; i < pagevec_count(&pvec); i++) {
627			struct page *page = pvec.pages[i];
628
629			/* We rely upon deletion not changing page->index */
630			index = indices[i];
631
632			if (xa_is_value(page)) {
633				if (!invalidate_exceptional_entry2(mapping,
634								   index, page))
635					ret = -EBUSY;
636				continue;
637			}
638
639			if (!did_range_unmap && page_mapped(page)) {
640				/*
641				 * If page is mapped, before taking its lock,
642				 * zap the rest of the file in one hit.
643				 */
644				unmap_mapping_pages(mapping, index,
645						(1 + end - index), false);
646				did_range_unmap = 1;
647			}
648
649			lock_page(page);
650			WARN_ON(page_to_index(page) != index);
651			if (page->mapping != mapping) {
652				unlock_page(page);
653				continue;
654			}
655			wait_on_page_writeback(page);
 
656
657			if (page_mapped(page))
658				unmap_mapping_page(page);
659			BUG_ON(page_mapped(page));
660
661			ret2 = do_launder_page(mapping, page);
662			if (ret2 == 0) {
663				if (!invalidate_complete_page2(mapping, page))
664					ret2 = -EBUSY;
665			}
666			if (ret2 < 0)
667				ret = ret2;
668			unlock_page(page);
669		}
670		pagevec_remove_exceptionals(&pvec);
671		pagevec_release(&pvec);
672		cond_resched();
673		index++;
674	}
675	/*
676	 * For DAX we invalidate page tables after invalidating page cache.  We
677	 * could invalidate page tables while invalidating each entry however
678	 * that would be expensive. And doing range unmapping before doesn't
679	 * work as we have no cheap way to find whether page cache entry didn't
680	 * get remapped later.
681	 */
682	if (dax_mapping(mapping)) {
683		unmap_mapping_pages(mapping, start, end - start + 1, false);
684	}
685out:
686	cleancache_invalidate_inode(mapping);
687	return ret;
688}
689EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
690
691/**
692 * invalidate_inode_pages2 - remove all pages from an address_space
693 * @mapping: the address_space
694 *
695 * Any pages which are found to be mapped into pagetables are unmapped prior to
696 * invalidation.
697 *
698 * Return: -EBUSY if any pages could not be invalidated.
699 */
700int invalidate_inode_pages2(struct address_space *mapping)
701{
702	return invalidate_inode_pages2_range(mapping, 0, -1);
703}
704EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
705
706/**
707 * truncate_pagecache - unmap and remove pagecache that has been truncated
708 * @inode: inode
709 * @newsize: new file size
710 *
711 * inode's new i_size must already be written before truncate_pagecache
712 * is called.
713 *
714 * This function should typically be called before the filesystem
715 * releases resources associated with the freed range (eg. deallocates
716 * blocks). This way, pagecache will always stay logically coherent
717 * with on-disk format, and the filesystem would not have to deal with
718 * situations such as writepage being called for a page that has already
719 * had its underlying blocks deallocated.
720 */
721void truncate_pagecache(struct inode *inode, loff_t newsize)
722{
723	struct address_space *mapping = inode->i_mapping;
724	loff_t holebegin = round_up(newsize, PAGE_SIZE);
725
726	/*
727	 * unmap_mapping_range is called twice, first simply for
728	 * efficiency so that truncate_inode_pages does fewer
729	 * single-page unmaps.  However after this first call, and
730	 * before truncate_inode_pages finishes, it is possible for
731	 * private pages to be COWed, which remain after
732	 * truncate_inode_pages finishes, hence the second
733	 * unmap_mapping_range call must be made for correctness.
734	 */
735	unmap_mapping_range(mapping, holebegin, 0, 1);
736	truncate_inode_pages(mapping, newsize);
737	unmap_mapping_range(mapping, holebegin, 0, 1);
738}
739EXPORT_SYMBOL(truncate_pagecache);
740
741/**
742 * truncate_setsize - update inode and pagecache for a new file size
743 * @inode: inode
744 * @newsize: new file size
745 *
746 * truncate_setsize updates i_size and performs pagecache truncation (if
747 * necessary) to @newsize. It will be typically be called from the filesystem's
748 * setattr function when ATTR_SIZE is passed in.
749 *
750 * Must be called with a lock serializing truncates and writes (generally
751 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
752 * specific block truncation has been performed.
753 */
754void truncate_setsize(struct inode *inode, loff_t newsize)
755{
756	loff_t oldsize = inode->i_size;
757
758	i_size_write(inode, newsize);
759	if (newsize > oldsize)
760		pagecache_isize_extended(inode, oldsize, newsize);
761	truncate_pagecache(inode, newsize);
762}
763EXPORT_SYMBOL(truncate_setsize);
764
765/**
766 * pagecache_isize_extended - update pagecache after extension of i_size
767 * @inode:	inode for which i_size was extended
768 * @from:	original inode size
769 * @to:		new inode size
770 *
771 * Handle extension of inode size either caused by extending truncate or by
772 * write starting after current i_size. We mark the page straddling current
773 * i_size RO so that page_mkwrite() is called on the nearest write access to
774 * the page.  This way filesystem can be sure that page_mkwrite() is called on
775 * the page before user writes to the page via mmap after the i_size has been
776 * changed.
777 *
778 * The function must be called after i_size is updated so that page fault
779 * coming after we unlock the page will already see the new i_size.
780 * The function must be called while we still hold i_mutex - this not only
781 * makes sure i_size is stable but also that userspace cannot observe new
782 * i_size value before we are prepared to store mmap writes at new inode size.
783 */
784void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
785{
786	int bsize = i_blocksize(inode);
787	loff_t rounded_from;
788	struct page *page;
789	pgoff_t index;
790
791	WARN_ON(to > inode->i_size);
792
793	if (from >= to || bsize == PAGE_SIZE)
794		return;
795	/* Page straddling @from will not have any hole block created? */
796	rounded_from = round_up(from, bsize);
797	if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
798		return;
799
800	index = from >> PAGE_SHIFT;
801	page = find_lock_page(inode->i_mapping, index);
802	/* Page not cached? Nothing to do */
803	if (!page)
804		return;
805	/*
806	 * See clear_page_dirty_for_io() for details why set_page_dirty()
807	 * is needed.
808	 */
809	if (page_mkclean(page))
810		set_page_dirty(page);
811	unlock_page(page);
812	put_page(page);
813}
814EXPORT_SYMBOL(pagecache_isize_extended);
815
816/**
817 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
818 * @inode: inode
819 * @lstart: offset of beginning of hole
820 * @lend: offset of last byte of hole
821 *
822 * This function should typically be called before the filesystem
823 * releases resources associated with the freed range (eg. deallocates
824 * blocks). This way, pagecache will always stay logically coherent
825 * with on-disk format, and the filesystem would not have to deal with
826 * situations such as writepage being called for a page that has already
827 * had its underlying blocks deallocated.
828 */
829void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
830{
831	struct address_space *mapping = inode->i_mapping;
832	loff_t unmap_start = round_up(lstart, PAGE_SIZE);
833	loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
834	/*
835	 * This rounding is currently just for example: unmap_mapping_range
836	 * expands its hole outwards, whereas we want it to contract the hole
837	 * inwards.  However, existing callers of truncate_pagecache_range are
838	 * doing their own page rounding first.  Note that unmap_mapping_range
839	 * allows holelen 0 for all, and we allow lend -1 for end of file.
840	 */
841
842	/*
843	 * Unlike in truncate_pagecache, unmap_mapping_range is called only
844	 * once (before truncating pagecache), and without "even_cows" flag:
845	 * hole-punching should not remove private COWed pages from the hole.
846	 */
847	if ((u64)unmap_end > (u64)unmap_start)
848		unmap_mapping_range(mapping, unmap_start,
849				    1 + unmap_end - unmap_start, 0);
850	truncate_inode_pages_range(mapping, lstart, lend);
851}
852EXPORT_SYMBOL(truncate_pagecache_range);