Loading...
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);
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/shmem_fs.h>
24#include <linux/cleancache.h>
25#include <linux/rmap.h>
26#include "internal.h"
27
28/*
29 * Regular page slots are stabilized by the page lock even without the tree
30 * itself locked. These unlocked entries need verification under the tree
31 * lock.
32 */
33static inline void __clear_shadow_entry(struct address_space *mapping,
34 pgoff_t index, void *entry)
35{
36 struct radix_tree_node *node;
37 void **slot;
38
39 if (!__radix_tree_lookup(&mapping->i_pages, index, &node, &slot))
40 return;
41 if (*slot != entry)
42 return;
43 __radix_tree_replace(&mapping->i_pages, node, slot, NULL,
44 workingset_update_node);
45 mapping->nrexceptional--;
46}
47
48static void clear_shadow_entry(struct address_space *mapping, pgoff_t index,
49 void *entry)
50{
51 xa_lock_irq(&mapping->i_pages);
52 __clear_shadow_entry(mapping, index, entry);
53 xa_unlock_irq(&mapping->i_pages);
54}
55
56/*
57 * Unconditionally remove exceptional entries. Usually called from truncate
58 * path. Note that the pagevec may be altered by this function by removing
59 * exceptional entries similar to what pagevec_remove_exceptionals does.
60 */
61static void truncate_exceptional_pvec_entries(struct address_space *mapping,
62 struct pagevec *pvec, pgoff_t *indices,
63 pgoff_t end)
64{
65 int i, j;
66 bool dax, lock;
67
68 /* Handled by shmem itself */
69 if (shmem_mapping(mapping))
70 return;
71
72 for (j = 0; j < pagevec_count(pvec); j++)
73 if (radix_tree_exceptional_entry(pvec->pages[j]))
74 break;
75
76 if (j == pagevec_count(pvec))
77 return;
78
79 dax = dax_mapping(mapping);
80 lock = !dax && indices[j] < end;
81 if (lock)
82 xa_lock_irq(&mapping->i_pages);
83
84 for (i = j; i < pagevec_count(pvec); i++) {
85 struct page *page = pvec->pages[i];
86 pgoff_t index = indices[i];
87
88 if (!radix_tree_exceptional_entry(page)) {
89 pvec->pages[j++] = page;
90 continue;
91 }
92
93 if (index >= end)
94 continue;
95
96 if (unlikely(dax)) {
97 dax_delete_mapping_entry(mapping, index);
98 continue;
99 }
100
101 __clear_shadow_entry(mapping, index, page);
102 }
103
104 if (lock)
105 xa_unlock_irq(&mapping->i_pages);
106 pvec->nr = j;
107}
108
109/*
110 * Invalidate exceptional entry if easily possible. This handles exceptional
111 * entries for invalidate_inode_pages().
112 */
113static int invalidate_exceptional_entry(struct address_space *mapping,
114 pgoff_t index, void *entry)
115{
116 /* Handled by shmem itself, or for DAX we do nothing. */
117 if (shmem_mapping(mapping) || dax_mapping(mapping))
118 return 1;
119 clear_shadow_entry(mapping, index, entry);
120 return 1;
121}
122
123/*
124 * Invalidate exceptional entry if clean. This handles exceptional entries for
125 * invalidate_inode_pages2() so for DAX it evicts only clean entries.
126 */
127static int invalidate_exceptional_entry2(struct address_space *mapping,
128 pgoff_t index, void *entry)
129{
130 /* Handled by shmem itself */
131 if (shmem_mapping(mapping))
132 return 1;
133 if (dax_mapping(mapping))
134 return dax_invalidate_mapping_entry_sync(mapping, index);
135 clear_shadow_entry(mapping, index, entry);
136 return 1;
137}
138
139/**
140 * do_invalidatepage - invalidate part or all of a page
141 * @page: the page which is affected
142 * @offset: start of the range to invalidate
143 * @length: length of the range to invalidate
144 *
145 * do_invalidatepage() is called when all or part of the page has become
146 * invalidated by a truncate operation.
147 *
148 * do_invalidatepage() does not have to release all buffers, but it must
149 * ensure that no dirty buffer is left outside @offset and that no I/O
150 * is underway against any of the blocks which are outside the truncation
151 * point. Because the caller is about to free (and possibly reuse) those
152 * blocks on-disk.
153 */
154void do_invalidatepage(struct page *page, unsigned int offset,
155 unsigned int length)
156{
157 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
158
159 invalidatepage = page->mapping->a_ops->invalidatepage;
160#ifdef CONFIG_BLOCK
161 if (!invalidatepage)
162 invalidatepage = block_invalidatepage;
163#endif
164 if (invalidatepage)
165 (*invalidatepage)(page, offset, length);
166}
167
168/*
169 * If truncate cannot remove the fs-private metadata from the page, the page
170 * becomes orphaned. It will be left on the LRU and may even be mapped into
171 * user pagetables if we're racing with filemap_fault().
172 *
173 * We need to bale out if page->mapping is no longer equal to the original
174 * mapping. This happens a) when the VM reclaimed the page while we waited on
175 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
176 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
177 */
178static void
179truncate_cleanup_page(struct address_space *mapping, struct page *page)
180{
181 if (page_mapped(page)) {
182 pgoff_t nr = PageTransHuge(page) ? HPAGE_PMD_NR : 1;
183 unmap_mapping_pages(mapping, page->index, nr, false);
184 }
185
186 if (page_has_private(page))
187 do_invalidatepage(page, 0, PAGE_SIZE);
188
189 /*
190 * Some filesystems seem to re-dirty the page even after
191 * the VM has canceled the dirty bit (eg ext3 journaling).
192 * Hence dirty accounting check is placed after invalidation.
193 */
194 cancel_dirty_page(page);
195 ClearPageMappedToDisk(page);
196}
197
198/*
199 * This is for invalidate_mapping_pages(). That function can be called at
200 * any time, and is not supposed to throw away dirty pages. But pages can
201 * be marked dirty at any time too, so use remove_mapping which safely
202 * discards clean, unused pages.
203 *
204 * Returns non-zero if the page was successfully invalidated.
205 */
206static int
207invalidate_complete_page(struct address_space *mapping, struct page *page)
208{
209 int ret;
210
211 if (page->mapping != mapping)
212 return 0;
213
214 if (page_has_private(page) && !try_to_release_page(page, 0))
215 return 0;
216
217 ret = remove_mapping(mapping, page);
218
219 return ret;
220}
221
222int truncate_inode_page(struct address_space *mapping, struct page *page)
223{
224 VM_BUG_ON_PAGE(PageTail(page), page);
225
226 if (page->mapping != mapping)
227 return -EIO;
228
229 truncate_cleanup_page(mapping, page);
230 delete_from_page_cache(page);
231 return 0;
232}
233
234/*
235 * Used to get rid of pages on hardware memory corruption.
236 */
237int generic_error_remove_page(struct address_space *mapping, struct page *page)
238{
239 if (!mapping)
240 return -EINVAL;
241 /*
242 * Only punch for normal data pages for now.
243 * Handling other types like directories would need more auditing.
244 */
245 if (!S_ISREG(mapping->host->i_mode))
246 return -EIO;
247 return truncate_inode_page(mapping, page);
248}
249EXPORT_SYMBOL(generic_error_remove_page);
250
251/*
252 * Safely invalidate one page from its pagecache mapping.
253 * It only drops clean, unused pages. The page must be locked.
254 *
255 * Returns 1 if the page is successfully invalidated, otherwise 0.
256 */
257int invalidate_inode_page(struct page *page)
258{
259 struct address_space *mapping = page_mapping(page);
260 if (!mapping)
261 return 0;
262 if (PageDirty(page) || PageWriteback(page))
263 return 0;
264 if (page_mapped(page))
265 return 0;
266 return invalidate_complete_page(mapping, page);
267}
268
269/**
270 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
271 * @mapping: mapping to truncate
272 * @lstart: offset from which to truncate
273 * @lend: offset to which to truncate (inclusive)
274 *
275 * Truncate the page cache, removing the pages that are between
276 * specified offsets (and zeroing out partial pages
277 * if lstart or lend + 1 is not page aligned).
278 *
279 * Truncate takes two passes - the first pass is nonblocking. It will not
280 * block on page locks and it will not block on writeback. The second pass
281 * will wait. This is to prevent as much IO as possible in the affected region.
282 * The first pass will remove most pages, so the search cost of the second pass
283 * is low.
284 *
285 * We pass down the cache-hot hint to the page freeing code. Even if the
286 * mapping is large, it is probably the case that the final pages are the most
287 * recently touched, and freeing happens in ascending file offset order.
288 *
289 * Note that since ->invalidatepage() accepts range to invalidate
290 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
291 * page aligned properly.
292 */
293void truncate_inode_pages_range(struct address_space *mapping,
294 loff_t lstart, loff_t lend)
295{
296 pgoff_t start; /* inclusive */
297 pgoff_t end; /* exclusive */
298 unsigned int partial_start; /* inclusive */
299 unsigned int partial_end; /* exclusive */
300 struct pagevec pvec;
301 pgoff_t indices[PAGEVEC_SIZE];
302 pgoff_t index;
303 int i;
304
305 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
306 goto out;
307
308 /* Offsets within partial pages */
309 partial_start = lstart & (PAGE_SIZE - 1);
310 partial_end = (lend + 1) & (PAGE_SIZE - 1);
311
312 /*
313 * 'start' and 'end' always covers the range of pages to be fully
314 * truncated. Partial pages are covered with 'partial_start' at the
315 * start of the range and 'partial_end' at the end of the range.
316 * Note that 'end' is exclusive while 'lend' is inclusive.
317 */
318 start = (lstart + PAGE_SIZE - 1) >> PAGE_SHIFT;
319 if (lend == -1)
320 /*
321 * lend == -1 indicates end-of-file so we have to set 'end'
322 * to the highest possible pgoff_t and since the type is
323 * unsigned we're using -1.
324 */
325 end = -1;
326 else
327 end = (lend + 1) >> PAGE_SHIFT;
328
329 pagevec_init(&pvec);
330 index = start;
331 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
332 min(end - index, (pgoff_t)PAGEVEC_SIZE),
333 indices)) {
334 /*
335 * Pagevec array has exceptional entries and we may also fail
336 * to lock some pages. So we store pages that can be deleted
337 * in a new pagevec.
338 */
339 struct pagevec locked_pvec;
340
341 pagevec_init(&locked_pvec);
342 for (i = 0; i < pagevec_count(&pvec); i++) {
343 struct page *page = pvec.pages[i];
344
345 /* We rely upon deletion not changing page->index */
346 index = indices[i];
347 if (index >= end)
348 break;
349
350 if (radix_tree_exceptional_entry(page))
351 continue;
352
353 if (!trylock_page(page))
354 continue;
355 WARN_ON(page_to_index(page) != index);
356 if (PageWriteback(page)) {
357 unlock_page(page);
358 continue;
359 }
360 if (page->mapping != mapping) {
361 unlock_page(page);
362 continue;
363 }
364 pagevec_add(&locked_pvec, page);
365 }
366 for (i = 0; i < pagevec_count(&locked_pvec); i++)
367 truncate_cleanup_page(mapping, locked_pvec.pages[i]);
368 delete_from_page_cache_batch(mapping, &locked_pvec);
369 for (i = 0; i < pagevec_count(&locked_pvec); i++)
370 unlock_page(locked_pvec.pages[i]);
371 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
372 pagevec_release(&pvec);
373 cond_resched();
374 index++;
375 }
376 if (partial_start) {
377 struct page *page = find_lock_page(mapping, start - 1);
378 if (page) {
379 unsigned int top = PAGE_SIZE;
380 if (start > end) {
381 /* Truncation within a single page */
382 top = partial_end;
383 partial_end = 0;
384 }
385 wait_on_page_writeback(page);
386 zero_user_segment(page, partial_start, top);
387 cleancache_invalidate_page(mapping, page);
388 if (page_has_private(page))
389 do_invalidatepage(page, partial_start,
390 top - partial_start);
391 unlock_page(page);
392 put_page(page);
393 }
394 }
395 if (partial_end) {
396 struct page *page = find_lock_page(mapping, end);
397 if (page) {
398 wait_on_page_writeback(page);
399 zero_user_segment(page, 0, partial_end);
400 cleancache_invalidate_page(mapping, page);
401 if (page_has_private(page))
402 do_invalidatepage(page, 0,
403 partial_end);
404 unlock_page(page);
405 put_page(page);
406 }
407 }
408 /*
409 * If the truncation happened within a single page no pages
410 * will be released, just zeroed, so we can bail out now.
411 */
412 if (start >= end)
413 goto out;
414
415 index = start;
416 for ( ; ; ) {
417 cond_resched();
418 if (!pagevec_lookup_entries(&pvec, mapping, index,
419 min(end - index, (pgoff_t)PAGEVEC_SIZE), indices)) {
420 /* If all gone from start onwards, we're done */
421 if (index == start)
422 break;
423 /* Otherwise restart to make sure all gone */
424 index = start;
425 continue;
426 }
427 if (index == start && indices[0] >= end) {
428 /* All gone out of hole to be punched, we're done */
429 pagevec_remove_exceptionals(&pvec);
430 pagevec_release(&pvec);
431 break;
432 }
433
434 for (i = 0; i < pagevec_count(&pvec); i++) {
435 struct page *page = pvec.pages[i];
436
437 /* We rely upon deletion not changing page->index */
438 index = indices[i];
439 if (index >= end) {
440 /* Restart punch to make sure all gone */
441 index = start - 1;
442 break;
443 }
444
445 if (radix_tree_exceptional_entry(page))
446 continue;
447
448 lock_page(page);
449 WARN_ON(page_to_index(page) != index);
450 wait_on_page_writeback(page);
451 truncate_inode_page(mapping, page);
452 unlock_page(page);
453 }
454 truncate_exceptional_pvec_entries(mapping, &pvec, indices, end);
455 pagevec_release(&pvec);
456 index++;
457 }
458
459out:
460 cleancache_invalidate_inode(mapping);
461}
462EXPORT_SYMBOL(truncate_inode_pages_range);
463
464/**
465 * truncate_inode_pages - truncate *all* the pages from an offset
466 * @mapping: mapping to truncate
467 * @lstart: offset from which to truncate
468 *
469 * Called under (and serialised by) inode->i_mutex.
470 *
471 * Note: When this function returns, there can be a page in the process of
472 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
473 * mapping->nrpages can be non-zero when this function returns even after
474 * truncation of the whole mapping.
475 */
476void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
477{
478 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
479}
480EXPORT_SYMBOL(truncate_inode_pages);
481
482/**
483 * truncate_inode_pages_final - truncate *all* pages before inode dies
484 * @mapping: mapping to truncate
485 *
486 * Called under (and serialized by) inode->i_mutex.
487 *
488 * Filesystems have to use this in the .evict_inode path to inform the
489 * VM that this is the final truncate and the inode is going away.
490 */
491void truncate_inode_pages_final(struct address_space *mapping)
492{
493 unsigned long nrexceptional;
494 unsigned long nrpages;
495
496 /*
497 * Page reclaim can not participate in regular inode lifetime
498 * management (can't call iput()) and thus can race with the
499 * inode teardown. Tell it when the address space is exiting,
500 * so that it does not install eviction information after the
501 * final truncate has begun.
502 */
503 mapping_set_exiting(mapping);
504
505 /*
506 * When reclaim installs eviction entries, it increases
507 * nrexceptional first, then decreases nrpages. Make sure we see
508 * this in the right order or we might miss an entry.
509 */
510 nrpages = mapping->nrpages;
511 smp_rmb();
512 nrexceptional = mapping->nrexceptional;
513
514 if (nrpages || nrexceptional) {
515 /*
516 * As truncation uses a lockless tree lookup, cycle
517 * the tree lock to make sure any ongoing tree
518 * modification that does not see AS_EXITING is
519 * completed before starting the final truncate.
520 */
521 xa_lock_irq(&mapping->i_pages);
522 xa_unlock_irq(&mapping->i_pages);
523
524 truncate_inode_pages(mapping, 0);
525 }
526}
527EXPORT_SYMBOL(truncate_inode_pages_final);
528
529/**
530 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
531 * @mapping: the address_space which holds the pages to invalidate
532 * @start: the offset 'from' which to invalidate
533 * @end: the offset 'to' which to invalidate (inclusive)
534 *
535 * This function only removes the unlocked pages, if you want to
536 * remove all the pages of one inode, you must call truncate_inode_pages.
537 *
538 * invalidate_mapping_pages() will not block on IO activity. It will not
539 * invalidate pages which are dirty, locked, under writeback or mapped into
540 * pagetables.
541 */
542unsigned long invalidate_mapping_pages(struct address_space *mapping,
543 pgoff_t start, pgoff_t end)
544{
545 pgoff_t indices[PAGEVEC_SIZE];
546 struct pagevec pvec;
547 pgoff_t index = start;
548 unsigned long ret;
549 unsigned long count = 0;
550 int i;
551
552 pagevec_init(&pvec);
553 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
554 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
555 indices)) {
556 for (i = 0; i < pagevec_count(&pvec); i++) {
557 struct page *page = pvec.pages[i];
558
559 /* We rely upon deletion not changing page->index */
560 index = indices[i];
561 if (index > end)
562 break;
563
564 if (radix_tree_exceptional_entry(page)) {
565 invalidate_exceptional_entry(mapping, index,
566 page);
567 continue;
568 }
569
570 if (!trylock_page(page))
571 continue;
572
573 WARN_ON(page_to_index(page) != index);
574
575 /* Middle of THP: skip */
576 if (PageTransTail(page)) {
577 unlock_page(page);
578 continue;
579 } else if (PageTransHuge(page)) {
580 index += HPAGE_PMD_NR - 1;
581 i += HPAGE_PMD_NR - 1;
582 /*
583 * 'end' is in the middle of THP. Don't
584 * invalidate the page as the part outside of
585 * 'end' could be still useful.
586 */
587 if (index > end) {
588 unlock_page(page);
589 continue;
590 }
591 }
592
593 ret = invalidate_inode_page(page);
594 unlock_page(page);
595 /*
596 * Invalidation is a hint that the page is no longer
597 * of interest and try to speed up its reclaim.
598 */
599 if (!ret)
600 deactivate_file_page(page);
601 count += ret;
602 }
603 pagevec_remove_exceptionals(&pvec);
604 pagevec_release(&pvec);
605 cond_resched();
606 index++;
607 }
608 return count;
609}
610EXPORT_SYMBOL(invalidate_mapping_pages);
611
612/*
613 * This is like invalidate_complete_page(), except it ignores the page's
614 * refcount. We do this because invalidate_inode_pages2() needs stronger
615 * invalidation guarantees, and cannot afford to leave pages behind because
616 * shrink_page_list() has a temp ref on them, or because they're transiently
617 * sitting in the lru_cache_add() pagevecs.
618 */
619static int
620invalidate_complete_page2(struct address_space *mapping, struct page *page)
621{
622 unsigned long flags;
623
624 if (page->mapping != mapping)
625 return 0;
626
627 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
628 return 0;
629
630 xa_lock_irqsave(&mapping->i_pages, flags);
631 if (PageDirty(page))
632 goto failed;
633
634 BUG_ON(page_has_private(page));
635 __delete_from_page_cache(page, NULL);
636 xa_unlock_irqrestore(&mapping->i_pages, flags);
637
638 if (mapping->a_ops->freepage)
639 mapping->a_ops->freepage(page);
640
641 put_page(page); /* pagecache ref */
642 return 1;
643failed:
644 xa_unlock_irqrestore(&mapping->i_pages, flags);
645 return 0;
646}
647
648static int do_launder_page(struct address_space *mapping, struct page *page)
649{
650 if (!PageDirty(page))
651 return 0;
652 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
653 return 0;
654 return mapping->a_ops->launder_page(page);
655}
656
657/**
658 * invalidate_inode_pages2_range - remove range of pages from an address_space
659 * @mapping: the address_space
660 * @start: the page offset 'from' which to invalidate
661 * @end: the page offset 'to' which to invalidate (inclusive)
662 *
663 * Any pages which are found to be mapped into pagetables are unmapped prior to
664 * invalidation.
665 *
666 * Returns -EBUSY if any pages could not be invalidated.
667 */
668int invalidate_inode_pages2_range(struct address_space *mapping,
669 pgoff_t start, pgoff_t end)
670{
671 pgoff_t indices[PAGEVEC_SIZE];
672 struct pagevec pvec;
673 pgoff_t index;
674 int i;
675 int ret = 0;
676 int ret2 = 0;
677 int did_range_unmap = 0;
678
679 if (mapping->nrpages == 0 && mapping->nrexceptional == 0)
680 goto out;
681
682 pagevec_init(&pvec);
683 index = start;
684 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
685 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
686 indices)) {
687 for (i = 0; i < pagevec_count(&pvec); i++) {
688 struct page *page = pvec.pages[i];
689
690 /* We rely upon deletion not changing page->index */
691 index = indices[i];
692 if (index > end)
693 break;
694
695 if (radix_tree_exceptional_entry(page)) {
696 if (!invalidate_exceptional_entry2(mapping,
697 index, page))
698 ret = -EBUSY;
699 continue;
700 }
701
702 lock_page(page);
703 WARN_ON(page_to_index(page) != index);
704 if (page->mapping != mapping) {
705 unlock_page(page);
706 continue;
707 }
708 wait_on_page_writeback(page);
709 if (page_mapped(page)) {
710 if (!did_range_unmap) {
711 /*
712 * Zap the rest of the file in one hit.
713 */
714 unmap_mapping_pages(mapping, index,
715 (1 + end - index), false);
716 did_range_unmap = 1;
717 } else {
718 /*
719 * Just zap this page
720 */
721 unmap_mapping_pages(mapping, index,
722 1, false);
723 }
724 }
725 BUG_ON(page_mapped(page));
726 ret2 = do_launder_page(mapping, page);
727 if (ret2 == 0) {
728 if (!invalidate_complete_page2(mapping, page))
729 ret2 = -EBUSY;
730 }
731 if (ret2 < 0)
732 ret = ret2;
733 unlock_page(page);
734 }
735 pagevec_remove_exceptionals(&pvec);
736 pagevec_release(&pvec);
737 cond_resched();
738 index++;
739 }
740 /*
741 * For DAX we invalidate page tables after invalidating radix tree. We
742 * could invalidate page tables while invalidating each entry however
743 * that would be expensive. And doing range unmapping before doesn't
744 * work as we have no cheap way to find whether radix tree entry didn't
745 * get remapped later.
746 */
747 if (dax_mapping(mapping)) {
748 unmap_mapping_pages(mapping, start, end - start + 1, false);
749 }
750out:
751 cleancache_invalidate_inode(mapping);
752 return ret;
753}
754EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
755
756/**
757 * invalidate_inode_pages2 - remove all pages from an address_space
758 * @mapping: the address_space
759 *
760 * Any pages which are found to be mapped into pagetables are unmapped prior to
761 * invalidation.
762 *
763 * Returns -EBUSY if any pages could not be invalidated.
764 */
765int invalidate_inode_pages2(struct address_space *mapping)
766{
767 return invalidate_inode_pages2_range(mapping, 0, -1);
768}
769EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
770
771/**
772 * truncate_pagecache - unmap and remove pagecache that has been truncated
773 * @inode: inode
774 * @newsize: new file size
775 *
776 * inode's new i_size must already be written before truncate_pagecache
777 * is called.
778 *
779 * This function should typically be called before the filesystem
780 * releases resources associated with the freed range (eg. deallocates
781 * blocks). This way, pagecache will always stay logically coherent
782 * with on-disk format, and the filesystem would not have to deal with
783 * situations such as writepage being called for a page that has already
784 * had its underlying blocks deallocated.
785 */
786void truncate_pagecache(struct inode *inode, loff_t newsize)
787{
788 struct address_space *mapping = inode->i_mapping;
789 loff_t holebegin = round_up(newsize, PAGE_SIZE);
790
791 /*
792 * unmap_mapping_range is called twice, first simply for
793 * efficiency so that truncate_inode_pages does fewer
794 * single-page unmaps. However after this first call, and
795 * before truncate_inode_pages finishes, it is possible for
796 * private pages to be COWed, which remain after
797 * truncate_inode_pages finishes, hence the second
798 * unmap_mapping_range call must be made for correctness.
799 */
800 unmap_mapping_range(mapping, holebegin, 0, 1);
801 truncate_inode_pages(mapping, newsize);
802 unmap_mapping_range(mapping, holebegin, 0, 1);
803}
804EXPORT_SYMBOL(truncate_pagecache);
805
806/**
807 * truncate_setsize - update inode and pagecache for a new file size
808 * @inode: inode
809 * @newsize: new file size
810 *
811 * truncate_setsize updates i_size and performs pagecache truncation (if
812 * necessary) to @newsize. It will be typically be called from the filesystem's
813 * setattr function when ATTR_SIZE is passed in.
814 *
815 * Must be called with a lock serializing truncates and writes (generally
816 * i_mutex but e.g. xfs uses a different lock) and before all filesystem
817 * specific block truncation has been performed.
818 */
819void truncate_setsize(struct inode *inode, loff_t newsize)
820{
821 loff_t oldsize = inode->i_size;
822
823 i_size_write(inode, newsize);
824 if (newsize > oldsize)
825 pagecache_isize_extended(inode, oldsize, newsize);
826 truncate_pagecache(inode, newsize);
827}
828EXPORT_SYMBOL(truncate_setsize);
829
830/**
831 * pagecache_isize_extended - update pagecache after extension of i_size
832 * @inode: inode for which i_size was extended
833 * @from: original inode size
834 * @to: new inode size
835 *
836 * Handle extension of inode size either caused by extending truncate or by
837 * write starting after current i_size. We mark the page straddling current
838 * i_size RO so that page_mkwrite() is called on the nearest write access to
839 * the page. This way filesystem can be sure that page_mkwrite() is called on
840 * the page before user writes to the page via mmap after the i_size has been
841 * changed.
842 *
843 * The function must be called after i_size is updated so that page fault
844 * coming after we unlock the page will already see the new i_size.
845 * The function must be called while we still hold i_mutex - this not only
846 * makes sure i_size is stable but also that userspace cannot observe new
847 * i_size value before we are prepared to store mmap writes at new inode size.
848 */
849void pagecache_isize_extended(struct inode *inode, loff_t from, loff_t to)
850{
851 int bsize = i_blocksize(inode);
852 loff_t rounded_from;
853 struct page *page;
854 pgoff_t index;
855
856 WARN_ON(to > inode->i_size);
857
858 if (from >= to || bsize == PAGE_SIZE)
859 return;
860 /* Page straddling @from will not have any hole block created? */
861 rounded_from = round_up(from, bsize);
862 if (to <= rounded_from || !(rounded_from & (PAGE_SIZE - 1)))
863 return;
864
865 index = from >> PAGE_SHIFT;
866 page = find_lock_page(inode->i_mapping, index);
867 /* Page not cached? Nothing to do */
868 if (!page)
869 return;
870 /*
871 * See clear_page_dirty_for_io() for details why set_page_dirty()
872 * is needed.
873 */
874 if (page_mkclean(page))
875 set_page_dirty(page);
876 unlock_page(page);
877 put_page(page);
878}
879EXPORT_SYMBOL(pagecache_isize_extended);
880
881/**
882 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
883 * @inode: inode
884 * @lstart: offset of beginning of hole
885 * @lend: offset of last byte of hole
886 *
887 * This function should typically be called before the filesystem
888 * releases resources associated with the freed range (eg. deallocates
889 * blocks). This way, pagecache will always stay logically coherent
890 * with on-disk format, and the filesystem would not have to deal with
891 * situations such as writepage being called for a page that has already
892 * had its underlying blocks deallocated.
893 */
894void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
895{
896 struct address_space *mapping = inode->i_mapping;
897 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
898 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
899 /*
900 * This rounding is currently just for example: unmap_mapping_range
901 * expands its hole outwards, whereas we want it to contract the hole
902 * inwards. However, existing callers of truncate_pagecache_range are
903 * doing their own page rounding first. Note that unmap_mapping_range
904 * allows holelen 0 for all, and we allow lend -1 for end of file.
905 */
906
907 /*
908 * Unlike in truncate_pagecache, unmap_mapping_range is called only
909 * once (before truncating pagecache), and without "even_cows" flag:
910 * hole-punching should not remove private COWed pages from the hole.
911 */
912 if ((u64)unmap_end > (u64)unmap_start)
913 unmap_mapping_range(mapping, unmap_start,
914 1 + unmap_end - unmap_start, 0);
915 truncate_inode_pages_range(mapping, lstart, lend);
916}
917EXPORT_SYMBOL(truncate_pagecache_range);