Loading...
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/gfp.h>
13#include <linux/mm.h>
14#include <linux/swap.h>
15#include <linux/module.h>
16#include <linux/pagemap.h>
17#include <linux/highmem.h>
18#include <linux/pagevec.h>
19#include <linux/task_io_accounting_ops.h>
20#include <linux/buffer_head.h> /* grr. try_to_release_page,
21 do_invalidatepage */
22#include <linux/cleancache.h>
23#include "internal.h"
24
25
26/**
27 * do_invalidatepage - invalidate part or all of a page
28 * @page: the page which is affected
29 * @offset: the index of the truncation point
30 *
31 * do_invalidatepage() is called when all or part of the page has become
32 * invalidated by a truncate operation.
33 *
34 * do_invalidatepage() does not have to release all buffers, but it must
35 * ensure that no dirty buffer is left outside @offset and that no I/O
36 * is underway against any of the blocks which are outside the truncation
37 * point. Because the caller is about to free (and possibly reuse) those
38 * blocks on-disk.
39 */
40void do_invalidatepage(struct page *page, unsigned long offset)
41{
42 void (*invalidatepage)(struct page *, unsigned long);
43 invalidatepage = page->mapping->a_ops->invalidatepage;
44#ifdef CONFIG_BLOCK
45 if (!invalidatepage)
46 invalidatepage = block_invalidatepage;
47#endif
48 if (invalidatepage)
49 (*invalidatepage)(page, offset);
50}
51
52static inline void truncate_partial_page(struct page *page, unsigned partial)
53{
54 zero_user_segment(page, partial, PAGE_CACHE_SIZE);
55 cleancache_flush_page(page->mapping, page);
56 if (page_has_private(page))
57 do_invalidatepage(page, partial);
58}
59
60/*
61 * This cancels just the dirty bit on the kernel page itself, it
62 * does NOT actually remove dirty bits on any mmap's that may be
63 * around. It also leaves the page tagged dirty, so any sync
64 * activity will still find it on the dirty lists, and in particular,
65 * clear_page_dirty_for_io() will still look at the dirty bits in
66 * the VM.
67 *
68 * Doing this should *normally* only ever be done when a page
69 * is truncated, and is not actually mapped anywhere at all. However,
70 * fs/buffer.c does this when it notices that somebody has cleaned
71 * out all the buffers on a page without actually doing it through
72 * the VM. Can you say "ext3 is horribly ugly"? Tought you could.
73 */
74void cancel_dirty_page(struct page *page, unsigned int account_size)
75{
76 if (TestClearPageDirty(page)) {
77 struct address_space *mapping = page->mapping;
78 if (mapping && mapping_cap_account_dirty(mapping)) {
79 dec_zone_page_state(page, NR_FILE_DIRTY);
80 dec_bdi_stat(mapping->backing_dev_info,
81 BDI_RECLAIMABLE);
82 if (account_size)
83 task_io_account_cancelled_write(account_size);
84 }
85 }
86}
87EXPORT_SYMBOL(cancel_dirty_page);
88
89/*
90 * If truncate cannot remove the fs-private metadata from the page, the page
91 * becomes orphaned. It will be left on the LRU and may even be mapped into
92 * user pagetables if we're racing with filemap_fault().
93 *
94 * We need to bale out if page->mapping is no longer equal to the original
95 * mapping. This happens a) when the VM reclaimed the page while we waited on
96 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
97 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
98 */
99static int
100truncate_complete_page(struct address_space *mapping, struct page *page)
101{
102 if (page->mapping != mapping)
103 return -EIO;
104
105 if (page_has_private(page))
106 do_invalidatepage(page, 0);
107
108 cancel_dirty_page(page, PAGE_CACHE_SIZE);
109
110 clear_page_mlock(page);
111 ClearPageMappedToDisk(page);
112 delete_from_page_cache(page);
113 return 0;
114}
115
116/*
117 * This is for invalidate_mapping_pages(). That function can be called at
118 * any time, and is not supposed to throw away dirty pages. But pages can
119 * be marked dirty at any time too, so use remove_mapping which safely
120 * discards clean, unused pages.
121 *
122 * Returns non-zero if the page was successfully invalidated.
123 */
124static int
125invalidate_complete_page(struct address_space *mapping, struct page *page)
126{
127 int ret;
128
129 if (page->mapping != mapping)
130 return 0;
131
132 if (page_has_private(page) && !try_to_release_page(page, 0))
133 return 0;
134
135 clear_page_mlock(page);
136 ret = remove_mapping(mapping, page);
137
138 return ret;
139}
140
141int truncate_inode_page(struct address_space *mapping, struct page *page)
142{
143 if (page_mapped(page)) {
144 unmap_mapping_range(mapping,
145 (loff_t)page->index << PAGE_CACHE_SHIFT,
146 PAGE_CACHE_SIZE, 0);
147 }
148 return truncate_complete_page(mapping, page);
149}
150
151/*
152 * Used to get rid of pages on hardware memory corruption.
153 */
154int generic_error_remove_page(struct address_space *mapping, struct page *page)
155{
156 if (!mapping)
157 return -EINVAL;
158 /*
159 * Only punch for normal data pages for now.
160 * Handling other types like directories would need more auditing.
161 */
162 if (!S_ISREG(mapping->host->i_mode))
163 return -EIO;
164 return truncate_inode_page(mapping, page);
165}
166EXPORT_SYMBOL(generic_error_remove_page);
167
168/*
169 * Safely invalidate one page from its pagecache mapping.
170 * It only drops clean, unused pages. The page must be locked.
171 *
172 * Returns 1 if the page is successfully invalidated, otherwise 0.
173 */
174int invalidate_inode_page(struct page *page)
175{
176 struct address_space *mapping = page_mapping(page);
177 if (!mapping)
178 return 0;
179 if (PageDirty(page) || PageWriteback(page))
180 return 0;
181 if (page_mapped(page))
182 return 0;
183 return invalidate_complete_page(mapping, page);
184}
185
186/**
187 * truncate_inode_pages - truncate range of pages specified by start & end byte offsets
188 * @mapping: mapping to truncate
189 * @lstart: offset from which to truncate
190 * @lend: offset to which to truncate
191 *
192 * Truncate the page cache, removing the pages that are between
193 * specified offsets (and zeroing out partial page
194 * (if lstart is not page aligned)).
195 *
196 * Truncate takes two passes - the first pass is nonblocking. It will not
197 * block on page locks and it will not block on writeback. The second pass
198 * will wait. This is to prevent as much IO as possible in the affected region.
199 * The first pass will remove most pages, so the search cost of the second pass
200 * is low.
201 *
202 * We pass down the cache-hot hint to the page freeing code. Even if the
203 * mapping is large, it is probably the case that the final pages are the most
204 * recently touched, and freeing happens in ascending file offset order.
205 */
206void truncate_inode_pages_range(struct address_space *mapping,
207 loff_t lstart, loff_t lend)
208{
209 const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
210 const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
211 struct pagevec pvec;
212 pgoff_t index;
213 pgoff_t end;
214 int i;
215
216 cleancache_flush_inode(mapping);
217 if (mapping->nrpages == 0)
218 return;
219
220 BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
221 end = (lend >> PAGE_CACHE_SHIFT);
222
223 pagevec_init(&pvec, 0);
224 index = start;
225 while (index <= end && pagevec_lookup(&pvec, mapping, index,
226 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
227 mem_cgroup_uncharge_start();
228 for (i = 0; i < pagevec_count(&pvec); i++) {
229 struct page *page = pvec.pages[i];
230
231 /* We rely upon deletion not changing page->index */
232 index = page->index;
233 if (index > end)
234 break;
235
236 if (!trylock_page(page))
237 continue;
238 WARN_ON(page->index != index);
239 if (PageWriteback(page)) {
240 unlock_page(page);
241 continue;
242 }
243 truncate_inode_page(mapping, page);
244 unlock_page(page);
245 }
246 pagevec_release(&pvec);
247 mem_cgroup_uncharge_end();
248 cond_resched();
249 index++;
250 }
251
252 if (partial) {
253 struct page *page = find_lock_page(mapping, start - 1);
254 if (page) {
255 wait_on_page_writeback(page);
256 truncate_partial_page(page, partial);
257 unlock_page(page);
258 page_cache_release(page);
259 }
260 }
261
262 index = start;
263 for ( ; ; ) {
264 cond_resched();
265 if (!pagevec_lookup(&pvec, mapping, index,
266 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
267 if (index == start)
268 break;
269 index = start;
270 continue;
271 }
272 if (index == start && pvec.pages[0]->index > end) {
273 pagevec_release(&pvec);
274 break;
275 }
276 mem_cgroup_uncharge_start();
277 for (i = 0; i < pagevec_count(&pvec); i++) {
278 struct page *page = pvec.pages[i];
279
280 /* We rely upon deletion not changing page->index */
281 index = page->index;
282 if (index > end)
283 break;
284
285 lock_page(page);
286 WARN_ON(page->index != index);
287 wait_on_page_writeback(page);
288 truncate_inode_page(mapping, page);
289 unlock_page(page);
290 }
291 pagevec_release(&pvec);
292 mem_cgroup_uncharge_end();
293 index++;
294 }
295 cleancache_flush_inode(mapping);
296}
297EXPORT_SYMBOL(truncate_inode_pages_range);
298
299/**
300 * truncate_inode_pages - truncate *all* the pages from an offset
301 * @mapping: mapping to truncate
302 * @lstart: offset from which to truncate
303 *
304 * Called under (and serialised by) inode->i_mutex.
305 *
306 * Note: When this function returns, there can be a page in the process of
307 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
308 * mapping->nrpages can be non-zero when this function returns even after
309 * truncation of the whole mapping.
310 */
311void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
312{
313 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
314}
315EXPORT_SYMBOL(truncate_inode_pages);
316
317/**
318 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
319 * @mapping: the address_space which holds the pages to invalidate
320 * @start: the offset 'from' which to invalidate
321 * @end: the offset 'to' which to invalidate (inclusive)
322 *
323 * This function only removes the unlocked pages, if you want to
324 * remove all the pages of one inode, you must call truncate_inode_pages.
325 *
326 * invalidate_mapping_pages() will not block on IO activity. It will not
327 * invalidate pages which are dirty, locked, under writeback or mapped into
328 * pagetables.
329 */
330unsigned long invalidate_mapping_pages(struct address_space *mapping,
331 pgoff_t start, pgoff_t end)
332{
333 struct pagevec pvec;
334 pgoff_t index = start;
335 unsigned long ret;
336 unsigned long count = 0;
337 int i;
338
339 /*
340 * Note: this function may get called on a shmem/tmpfs mapping:
341 * pagevec_lookup() might then return 0 prematurely (because it
342 * got a gangful of swap entries); but it's hardly worth worrying
343 * about - it can rarely have anything to free from such a mapping
344 * (most pages are dirty), and already skips over any difficulties.
345 */
346
347 pagevec_init(&pvec, 0);
348 while (index <= end && pagevec_lookup(&pvec, mapping, index,
349 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
350 mem_cgroup_uncharge_start();
351 for (i = 0; i < pagevec_count(&pvec); i++) {
352 struct page *page = pvec.pages[i];
353
354 /* We rely upon deletion not changing page->index */
355 index = page->index;
356 if (index > end)
357 break;
358
359 if (!trylock_page(page))
360 continue;
361 WARN_ON(page->index != index);
362 ret = invalidate_inode_page(page);
363 unlock_page(page);
364 /*
365 * Invalidation is a hint that the page is no longer
366 * of interest and try to speed up its reclaim.
367 */
368 if (!ret)
369 deactivate_page(page);
370 count += ret;
371 }
372 pagevec_release(&pvec);
373 mem_cgroup_uncharge_end();
374 cond_resched();
375 index++;
376 }
377 return count;
378}
379EXPORT_SYMBOL(invalidate_mapping_pages);
380
381/*
382 * This is like invalidate_complete_page(), except it ignores the page's
383 * refcount. We do this because invalidate_inode_pages2() needs stronger
384 * invalidation guarantees, and cannot afford to leave pages behind because
385 * shrink_page_list() has a temp ref on them, or because they're transiently
386 * sitting in the lru_cache_add() pagevecs.
387 */
388static int
389invalidate_complete_page2(struct address_space *mapping, struct page *page)
390{
391 if (page->mapping != mapping)
392 return 0;
393
394 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
395 return 0;
396
397 spin_lock_irq(&mapping->tree_lock);
398 if (PageDirty(page))
399 goto failed;
400
401 clear_page_mlock(page);
402 BUG_ON(page_has_private(page));
403 __delete_from_page_cache(page);
404 spin_unlock_irq(&mapping->tree_lock);
405 mem_cgroup_uncharge_cache_page(page);
406
407 if (mapping->a_ops->freepage)
408 mapping->a_ops->freepage(page);
409
410 page_cache_release(page); /* pagecache ref */
411 return 1;
412failed:
413 spin_unlock_irq(&mapping->tree_lock);
414 return 0;
415}
416
417static int do_launder_page(struct address_space *mapping, struct page *page)
418{
419 if (!PageDirty(page))
420 return 0;
421 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
422 return 0;
423 return mapping->a_ops->launder_page(page);
424}
425
426/**
427 * invalidate_inode_pages2_range - remove range of pages from an address_space
428 * @mapping: the address_space
429 * @start: the page offset 'from' which to invalidate
430 * @end: the page offset 'to' which to invalidate (inclusive)
431 *
432 * Any pages which are found to be mapped into pagetables are unmapped prior to
433 * invalidation.
434 *
435 * Returns -EBUSY if any pages could not be invalidated.
436 */
437int invalidate_inode_pages2_range(struct address_space *mapping,
438 pgoff_t start, pgoff_t end)
439{
440 struct pagevec pvec;
441 pgoff_t index;
442 int i;
443 int ret = 0;
444 int ret2 = 0;
445 int did_range_unmap = 0;
446
447 cleancache_flush_inode(mapping);
448 pagevec_init(&pvec, 0);
449 index = start;
450 while (index <= end && pagevec_lookup(&pvec, mapping, index,
451 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
452 mem_cgroup_uncharge_start();
453 for (i = 0; i < pagevec_count(&pvec); i++) {
454 struct page *page = pvec.pages[i];
455
456 /* We rely upon deletion not changing page->index */
457 index = page->index;
458 if (index > end)
459 break;
460
461 lock_page(page);
462 WARN_ON(page->index != index);
463 if (page->mapping != mapping) {
464 unlock_page(page);
465 continue;
466 }
467 wait_on_page_writeback(page);
468 if (page_mapped(page)) {
469 if (!did_range_unmap) {
470 /*
471 * Zap the rest of the file in one hit.
472 */
473 unmap_mapping_range(mapping,
474 (loff_t)index << PAGE_CACHE_SHIFT,
475 (loff_t)(1 + end - index)
476 << PAGE_CACHE_SHIFT,
477 0);
478 did_range_unmap = 1;
479 } else {
480 /*
481 * Just zap this page
482 */
483 unmap_mapping_range(mapping,
484 (loff_t)index << PAGE_CACHE_SHIFT,
485 PAGE_CACHE_SIZE, 0);
486 }
487 }
488 BUG_ON(page_mapped(page));
489 ret2 = do_launder_page(mapping, page);
490 if (ret2 == 0) {
491 if (!invalidate_complete_page2(mapping, page))
492 ret2 = -EBUSY;
493 }
494 if (ret2 < 0)
495 ret = ret2;
496 unlock_page(page);
497 }
498 pagevec_release(&pvec);
499 mem_cgroup_uncharge_end();
500 cond_resched();
501 index++;
502 }
503 cleancache_flush_inode(mapping);
504 return ret;
505}
506EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
507
508/**
509 * invalidate_inode_pages2 - remove all pages from an address_space
510 * @mapping: the address_space
511 *
512 * Any pages which are found to be mapped into pagetables are unmapped prior to
513 * invalidation.
514 *
515 * Returns -EBUSY if any pages could not be invalidated.
516 */
517int invalidate_inode_pages2(struct address_space *mapping)
518{
519 return invalidate_inode_pages2_range(mapping, 0, -1);
520}
521EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
522
523/**
524 * truncate_pagecache - unmap and remove pagecache that has been truncated
525 * @inode: inode
526 * @oldsize: old file size
527 * @newsize: new file size
528 *
529 * inode's new i_size must already be written before truncate_pagecache
530 * is called.
531 *
532 * This function should typically be called before the filesystem
533 * releases resources associated with the freed range (eg. deallocates
534 * blocks). This way, pagecache will always stay logically coherent
535 * with on-disk format, and the filesystem would not have to deal with
536 * situations such as writepage being called for a page that has already
537 * had its underlying blocks deallocated.
538 */
539void truncate_pagecache(struct inode *inode, loff_t oldsize, loff_t newsize)
540{
541 struct address_space *mapping = inode->i_mapping;
542 loff_t holebegin = round_up(newsize, PAGE_SIZE);
543
544 /*
545 * unmap_mapping_range is called twice, first simply for
546 * efficiency so that truncate_inode_pages does fewer
547 * single-page unmaps. However after this first call, and
548 * before truncate_inode_pages finishes, it is possible for
549 * private pages to be COWed, which remain after
550 * truncate_inode_pages finishes, hence the second
551 * unmap_mapping_range call must be made for correctness.
552 */
553 unmap_mapping_range(mapping, holebegin, 0, 1);
554 truncate_inode_pages(mapping, newsize);
555 unmap_mapping_range(mapping, holebegin, 0, 1);
556}
557EXPORT_SYMBOL(truncate_pagecache);
558
559/**
560 * truncate_setsize - update inode and pagecache for a new file size
561 * @inode: inode
562 * @newsize: new file size
563 *
564 * truncate_setsize updates i_size and performs pagecache truncation (if
565 * necessary) to @newsize. It will be typically be called from the filesystem's
566 * setattr function when ATTR_SIZE is passed in.
567 *
568 * Must be called with inode_mutex held and before all filesystem specific
569 * block truncation has been performed.
570 */
571void truncate_setsize(struct inode *inode, loff_t newsize)
572{
573 loff_t oldsize;
574
575 oldsize = inode->i_size;
576 i_size_write(inode, newsize);
577
578 truncate_pagecache(inode, oldsize, newsize);
579}
580EXPORT_SYMBOL(truncate_setsize);
581
582/**
583 * vmtruncate - unmap mappings "freed" by truncate() syscall
584 * @inode: inode of the file used
585 * @newsize: file offset to start truncating
586 *
587 * This function is deprecated and truncate_setsize or truncate_pagecache
588 * should be used instead, together with filesystem specific block truncation.
589 */
590int vmtruncate(struct inode *inode, loff_t newsize)
591{
592 int error;
593
594 error = inode_newsize_ok(inode, newsize);
595 if (error)
596 return error;
597
598 truncate_setsize(inode, newsize);
599 if (inode->i_op->truncate)
600 inode->i_op->truncate(inode);
601 return 0;
602}
603EXPORT_SYMBOL(vmtruncate);
604
605int vmtruncate_range(struct inode *inode, loff_t lstart, loff_t lend)
606{
607 struct address_space *mapping = inode->i_mapping;
608 loff_t holebegin = round_up(lstart, PAGE_SIZE);
609 loff_t holelen = 1 + lend - holebegin;
610
611 /*
612 * If the underlying filesystem is not going to provide
613 * a way to truncate a range of blocks (punch a hole) -
614 * we should return failure right now.
615 */
616 if (!inode->i_op->truncate_range)
617 return -ENOSYS;
618
619 mutex_lock(&inode->i_mutex);
620 inode_dio_wait(inode);
621 unmap_mapping_range(mapping, holebegin, holelen, 1);
622 inode->i_op->truncate_range(inode, lstart, lend);
623 /* unmap again to remove racily COWed private pages */
624 unmap_mapping_range(mapping, holebegin, holelen, 1);
625 mutex_unlock(&inode->i_mutex);
626
627 return 0;
628}
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/gfp.h>
13#include <linux/mm.h>
14#include <linux/swap.h>
15#include <linux/export.h>
16#include <linux/pagemap.h>
17#include <linux/highmem.h>
18#include <linux/pagevec.h>
19#include <linux/task_io_accounting_ops.h>
20#include <linux/buffer_head.h> /* grr. try_to_release_page,
21 do_invalidatepage */
22#include <linux/cleancache.h>
23#include "internal.h"
24
25static void clear_exceptional_entry(struct address_space *mapping,
26 pgoff_t index, void *entry)
27{
28 struct radix_tree_node *node;
29 void **slot;
30
31 /* Handled by shmem itself */
32 if (shmem_mapping(mapping))
33 return;
34
35 spin_lock_irq(&mapping->tree_lock);
36 /*
37 * Regular page slots are stabilized by the page lock even
38 * without the tree itself locked. These unlocked entries
39 * need verification under the tree lock.
40 */
41 if (!__radix_tree_lookup(&mapping->page_tree, index, &node, &slot))
42 goto unlock;
43 if (*slot != entry)
44 goto unlock;
45 radix_tree_replace_slot(slot, NULL);
46 mapping->nrshadows--;
47 if (!node)
48 goto unlock;
49 workingset_node_shadows_dec(node);
50 /*
51 * Don't track node without shadow entries.
52 *
53 * Avoid acquiring the list_lru lock if already untracked.
54 * The list_empty() test is safe as node->private_list is
55 * protected by mapping->tree_lock.
56 */
57 if (!workingset_node_shadows(node) &&
58 !list_empty(&node->private_list))
59 list_lru_del(&workingset_shadow_nodes, &node->private_list);
60 __radix_tree_delete_node(&mapping->page_tree, node);
61unlock:
62 spin_unlock_irq(&mapping->tree_lock);
63}
64
65/**
66 * do_invalidatepage - invalidate part or all of a page
67 * @page: the page which is affected
68 * @offset: start of the range to invalidate
69 * @length: length of the range to invalidate
70 *
71 * do_invalidatepage() is called when all or part of the page has become
72 * invalidated by a truncate operation.
73 *
74 * do_invalidatepage() does not have to release all buffers, but it must
75 * ensure that no dirty buffer is left outside @offset and that no I/O
76 * is underway against any of the blocks which are outside the truncation
77 * point. Because the caller is about to free (and possibly reuse) those
78 * blocks on-disk.
79 */
80void do_invalidatepage(struct page *page, unsigned int offset,
81 unsigned int length)
82{
83 void (*invalidatepage)(struct page *, unsigned int, unsigned int);
84
85 invalidatepage = page->mapping->a_ops->invalidatepage;
86#ifdef CONFIG_BLOCK
87 if (!invalidatepage)
88 invalidatepage = block_invalidatepage;
89#endif
90 if (invalidatepage)
91 (*invalidatepage)(page, offset, length);
92}
93
94/*
95 * This cancels just the dirty bit on the kernel page itself, it
96 * does NOT actually remove dirty bits on any mmap's that may be
97 * around. It also leaves the page tagged dirty, so any sync
98 * activity will still find it on the dirty lists, and in particular,
99 * clear_page_dirty_for_io() will still look at the dirty bits in
100 * the VM.
101 *
102 * Doing this should *normally* only ever be done when a page
103 * is truncated, and is not actually mapped anywhere at all. However,
104 * fs/buffer.c does this when it notices that somebody has cleaned
105 * out all the buffers on a page without actually doing it through
106 * the VM. Can you say "ext3 is horribly ugly"? Tought you could.
107 */
108void cancel_dirty_page(struct page *page, unsigned int account_size)
109{
110 if (TestClearPageDirty(page)) {
111 struct address_space *mapping = page->mapping;
112 if (mapping && mapping_cap_account_dirty(mapping)) {
113 dec_zone_page_state(page, NR_FILE_DIRTY);
114 dec_bdi_stat(mapping->backing_dev_info,
115 BDI_RECLAIMABLE);
116 if (account_size)
117 task_io_account_cancelled_write(account_size);
118 }
119 }
120}
121EXPORT_SYMBOL(cancel_dirty_page);
122
123/*
124 * If truncate cannot remove the fs-private metadata from the page, the page
125 * becomes orphaned. It will be left on the LRU and may even be mapped into
126 * user pagetables if we're racing with filemap_fault().
127 *
128 * We need to bale out if page->mapping is no longer equal to the original
129 * mapping. This happens a) when the VM reclaimed the page while we waited on
130 * its lock, b) when a concurrent invalidate_mapping_pages got there first and
131 * c) when tmpfs swizzles a page between a tmpfs inode and swapper_space.
132 */
133static int
134truncate_complete_page(struct address_space *mapping, struct page *page)
135{
136 if (page->mapping != mapping)
137 return -EIO;
138
139 if (page_has_private(page))
140 do_invalidatepage(page, 0, PAGE_CACHE_SIZE);
141
142 cancel_dirty_page(page, PAGE_CACHE_SIZE);
143
144 ClearPageMappedToDisk(page);
145 delete_from_page_cache(page);
146 return 0;
147}
148
149/*
150 * This is for invalidate_mapping_pages(). That function can be called at
151 * any time, and is not supposed to throw away dirty pages. But pages can
152 * be marked dirty at any time too, so use remove_mapping which safely
153 * discards clean, unused pages.
154 *
155 * Returns non-zero if the page was successfully invalidated.
156 */
157static int
158invalidate_complete_page(struct address_space *mapping, struct page *page)
159{
160 int ret;
161
162 if (page->mapping != mapping)
163 return 0;
164
165 if (page_has_private(page) && !try_to_release_page(page, 0))
166 return 0;
167
168 ret = remove_mapping(mapping, page);
169
170 return ret;
171}
172
173int truncate_inode_page(struct address_space *mapping, struct page *page)
174{
175 if (page_mapped(page)) {
176 unmap_mapping_range(mapping,
177 (loff_t)page->index << PAGE_CACHE_SHIFT,
178 PAGE_CACHE_SIZE, 0);
179 }
180 return truncate_complete_page(mapping, page);
181}
182
183/*
184 * Used to get rid of pages on hardware memory corruption.
185 */
186int generic_error_remove_page(struct address_space *mapping, struct page *page)
187{
188 if (!mapping)
189 return -EINVAL;
190 /*
191 * Only punch for normal data pages for now.
192 * Handling other types like directories would need more auditing.
193 */
194 if (!S_ISREG(mapping->host->i_mode))
195 return -EIO;
196 return truncate_inode_page(mapping, page);
197}
198EXPORT_SYMBOL(generic_error_remove_page);
199
200/*
201 * Safely invalidate one page from its pagecache mapping.
202 * It only drops clean, unused pages. The page must be locked.
203 *
204 * Returns 1 if the page is successfully invalidated, otherwise 0.
205 */
206int invalidate_inode_page(struct page *page)
207{
208 struct address_space *mapping = page_mapping(page);
209 if (!mapping)
210 return 0;
211 if (PageDirty(page) || PageWriteback(page))
212 return 0;
213 if (page_mapped(page))
214 return 0;
215 return invalidate_complete_page(mapping, page);
216}
217
218/**
219 * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
220 * @mapping: mapping to truncate
221 * @lstart: offset from which to truncate
222 * @lend: offset to which to truncate (inclusive)
223 *
224 * Truncate the page cache, removing the pages that are between
225 * specified offsets (and zeroing out partial pages
226 * if lstart or lend + 1 is not page aligned).
227 *
228 * Truncate takes two passes - the first pass is nonblocking. It will not
229 * block on page locks and it will not block on writeback. The second pass
230 * will wait. This is to prevent as much IO as possible in the affected region.
231 * The first pass will remove most pages, so the search cost of the second pass
232 * is low.
233 *
234 * We pass down the cache-hot hint to the page freeing code. Even if the
235 * mapping is large, it is probably the case that the final pages are the most
236 * recently touched, and freeing happens in ascending file offset order.
237 *
238 * Note that since ->invalidatepage() accepts range to invalidate
239 * truncate_inode_pages_range is able to handle cases where lend + 1 is not
240 * page aligned properly.
241 */
242void truncate_inode_pages_range(struct address_space *mapping,
243 loff_t lstart, loff_t lend)
244{
245 pgoff_t start; /* inclusive */
246 pgoff_t end; /* exclusive */
247 unsigned int partial_start; /* inclusive */
248 unsigned int partial_end; /* exclusive */
249 struct pagevec pvec;
250 pgoff_t indices[PAGEVEC_SIZE];
251 pgoff_t index;
252 int i;
253
254 cleancache_invalidate_inode(mapping);
255 if (mapping->nrpages == 0 && mapping->nrshadows == 0)
256 return;
257
258 /* Offsets within partial pages */
259 partial_start = lstart & (PAGE_CACHE_SIZE - 1);
260 partial_end = (lend + 1) & (PAGE_CACHE_SIZE - 1);
261
262 /*
263 * 'start' and 'end' always covers the range of pages to be fully
264 * truncated. Partial pages are covered with 'partial_start' at the
265 * start of the range and 'partial_end' at the end of the range.
266 * Note that 'end' is exclusive while 'lend' is inclusive.
267 */
268 start = (lstart + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
269 if (lend == -1)
270 /*
271 * lend == -1 indicates end-of-file so we have to set 'end'
272 * to the highest possible pgoff_t and since the type is
273 * unsigned we're using -1.
274 */
275 end = -1;
276 else
277 end = (lend + 1) >> PAGE_CACHE_SHIFT;
278
279 pagevec_init(&pvec, 0);
280 index = start;
281 while (index < end && pagevec_lookup_entries(&pvec, mapping, index,
282 min(end - index, (pgoff_t)PAGEVEC_SIZE),
283 indices)) {
284 mem_cgroup_uncharge_start();
285 for (i = 0; i < pagevec_count(&pvec); i++) {
286 struct page *page = pvec.pages[i];
287
288 /* We rely upon deletion not changing page->index */
289 index = indices[i];
290 if (index >= end)
291 break;
292
293 if (radix_tree_exceptional_entry(page)) {
294 clear_exceptional_entry(mapping, index, page);
295 continue;
296 }
297
298 if (!trylock_page(page))
299 continue;
300 WARN_ON(page->index != index);
301 if (PageWriteback(page)) {
302 unlock_page(page);
303 continue;
304 }
305 truncate_inode_page(mapping, page);
306 unlock_page(page);
307 }
308 pagevec_remove_exceptionals(&pvec);
309 pagevec_release(&pvec);
310 mem_cgroup_uncharge_end();
311 cond_resched();
312 index++;
313 }
314
315 if (partial_start) {
316 struct page *page = find_lock_page(mapping, start - 1);
317 if (page) {
318 unsigned int top = PAGE_CACHE_SIZE;
319 if (start > end) {
320 /* Truncation within a single page */
321 top = partial_end;
322 partial_end = 0;
323 }
324 wait_on_page_writeback(page);
325 zero_user_segment(page, partial_start, top);
326 cleancache_invalidate_page(mapping, page);
327 if (page_has_private(page))
328 do_invalidatepage(page, partial_start,
329 top - partial_start);
330 unlock_page(page);
331 page_cache_release(page);
332 }
333 }
334 if (partial_end) {
335 struct page *page = find_lock_page(mapping, end);
336 if (page) {
337 wait_on_page_writeback(page);
338 zero_user_segment(page, 0, partial_end);
339 cleancache_invalidate_page(mapping, page);
340 if (page_has_private(page))
341 do_invalidatepage(page, 0,
342 partial_end);
343 unlock_page(page);
344 page_cache_release(page);
345 }
346 }
347 /*
348 * If the truncation happened within a single page no pages
349 * will be released, just zeroed, so we can bail out now.
350 */
351 if (start >= end)
352 return;
353
354 index = start;
355 for ( ; ; ) {
356 cond_resched();
357 if (!pagevec_lookup_entries(&pvec, mapping, index,
358 min(end - index, (pgoff_t)PAGEVEC_SIZE),
359 indices)) {
360 if (index == start)
361 break;
362 index = start;
363 continue;
364 }
365 if (index == start && indices[0] >= end) {
366 pagevec_remove_exceptionals(&pvec);
367 pagevec_release(&pvec);
368 break;
369 }
370 mem_cgroup_uncharge_start();
371 for (i = 0; i < pagevec_count(&pvec); i++) {
372 struct page *page = pvec.pages[i];
373
374 /* We rely upon deletion not changing page->index */
375 index = indices[i];
376 if (index >= end)
377 break;
378
379 if (radix_tree_exceptional_entry(page)) {
380 clear_exceptional_entry(mapping, index, page);
381 continue;
382 }
383
384 lock_page(page);
385 WARN_ON(page->index != index);
386 wait_on_page_writeback(page);
387 truncate_inode_page(mapping, page);
388 unlock_page(page);
389 }
390 pagevec_remove_exceptionals(&pvec);
391 pagevec_release(&pvec);
392 mem_cgroup_uncharge_end();
393 index++;
394 }
395 cleancache_invalidate_inode(mapping);
396}
397EXPORT_SYMBOL(truncate_inode_pages_range);
398
399/**
400 * truncate_inode_pages - truncate *all* the pages from an offset
401 * @mapping: mapping to truncate
402 * @lstart: offset from which to truncate
403 *
404 * Called under (and serialised by) inode->i_mutex.
405 *
406 * Note: When this function returns, there can be a page in the process of
407 * deletion (inside __delete_from_page_cache()) in the specified range. Thus
408 * mapping->nrpages can be non-zero when this function returns even after
409 * truncation of the whole mapping.
410 */
411void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
412{
413 truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
414}
415EXPORT_SYMBOL(truncate_inode_pages);
416
417/**
418 * truncate_inode_pages_final - truncate *all* pages before inode dies
419 * @mapping: mapping to truncate
420 *
421 * Called under (and serialized by) inode->i_mutex.
422 *
423 * Filesystems have to use this in the .evict_inode path to inform the
424 * VM that this is the final truncate and the inode is going away.
425 */
426void truncate_inode_pages_final(struct address_space *mapping)
427{
428 unsigned long nrshadows;
429 unsigned long nrpages;
430
431 /*
432 * Page reclaim can not participate in regular inode lifetime
433 * management (can't call iput()) and thus can race with the
434 * inode teardown. Tell it when the address space is exiting,
435 * so that it does not install eviction information after the
436 * final truncate has begun.
437 */
438 mapping_set_exiting(mapping);
439
440 /*
441 * When reclaim installs eviction entries, it increases
442 * nrshadows first, then decreases nrpages. Make sure we see
443 * this in the right order or we might miss an entry.
444 */
445 nrpages = mapping->nrpages;
446 smp_rmb();
447 nrshadows = mapping->nrshadows;
448
449 if (nrpages || nrshadows) {
450 /*
451 * As truncation uses a lockless tree lookup, cycle
452 * the tree lock to make sure any ongoing tree
453 * modification that does not see AS_EXITING is
454 * completed before starting the final truncate.
455 */
456 spin_lock_irq(&mapping->tree_lock);
457 spin_unlock_irq(&mapping->tree_lock);
458
459 truncate_inode_pages(mapping, 0);
460 }
461}
462EXPORT_SYMBOL(truncate_inode_pages_final);
463
464/**
465 * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
466 * @mapping: the address_space which holds the pages to invalidate
467 * @start: the offset 'from' which to invalidate
468 * @end: the offset 'to' which to invalidate (inclusive)
469 *
470 * This function only removes the unlocked pages, if you want to
471 * remove all the pages of one inode, you must call truncate_inode_pages.
472 *
473 * invalidate_mapping_pages() will not block on IO activity. It will not
474 * invalidate pages which are dirty, locked, under writeback or mapped into
475 * pagetables.
476 */
477unsigned long invalidate_mapping_pages(struct address_space *mapping,
478 pgoff_t start, pgoff_t end)
479{
480 pgoff_t indices[PAGEVEC_SIZE];
481 struct pagevec pvec;
482 pgoff_t index = start;
483 unsigned long ret;
484 unsigned long count = 0;
485 int i;
486
487 pagevec_init(&pvec, 0);
488 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
489 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
490 indices)) {
491 mem_cgroup_uncharge_start();
492 for (i = 0; i < pagevec_count(&pvec); i++) {
493 struct page *page = pvec.pages[i];
494
495 /* We rely upon deletion not changing page->index */
496 index = indices[i];
497 if (index > end)
498 break;
499
500 if (radix_tree_exceptional_entry(page)) {
501 clear_exceptional_entry(mapping, index, page);
502 continue;
503 }
504
505 if (!trylock_page(page))
506 continue;
507 WARN_ON(page->index != index);
508 ret = invalidate_inode_page(page);
509 unlock_page(page);
510 /*
511 * Invalidation is a hint that the page is no longer
512 * of interest and try to speed up its reclaim.
513 */
514 if (!ret)
515 deactivate_page(page);
516 count += ret;
517 }
518 pagevec_remove_exceptionals(&pvec);
519 pagevec_release(&pvec);
520 mem_cgroup_uncharge_end();
521 cond_resched();
522 index++;
523 }
524 return count;
525}
526EXPORT_SYMBOL(invalidate_mapping_pages);
527
528/*
529 * This is like invalidate_complete_page(), except it ignores the page's
530 * refcount. We do this because invalidate_inode_pages2() needs stronger
531 * invalidation guarantees, and cannot afford to leave pages behind because
532 * shrink_page_list() has a temp ref on them, or because they're transiently
533 * sitting in the lru_cache_add() pagevecs.
534 */
535static int
536invalidate_complete_page2(struct address_space *mapping, struct page *page)
537{
538 if (page->mapping != mapping)
539 return 0;
540
541 if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
542 return 0;
543
544 spin_lock_irq(&mapping->tree_lock);
545 if (PageDirty(page))
546 goto failed;
547
548 BUG_ON(page_has_private(page));
549 __delete_from_page_cache(page, NULL);
550 spin_unlock_irq(&mapping->tree_lock);
551 mem_cgroup_uncharge_cache_page(page);
552
553 if (mapping->a_ops->freepage)
554 mapping->a_ops->freepage(page);
555
556 page_cache_release(page); /* pagecache ref */
557 return 1;
558failed:
559 spin_unlock_irq(&mapping->tree_lock);
560 return 0;
561}
562
563static int do_launder_page(struct address_space *mapping, struct page *page)
564{
565 if (!PageDirty(page))
566 return 0;
567 if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
568 return 0;
569 return mapping->a_ops->launder_page(page);
570}
571
572/**
573 * invalidate_inode_pages2_range - remove range of pages from an address_space
574 * @mapping: the address_space
575 * @start: the page offset 'from' which to invalidate
576 * @end: the page offset 'to' which to invalidate (inclusive)
577 *
578 * Any pages which are found to be mapped into pagetables are unmapped prior to
579 * invalidation.
580 *
581 * Returns -EBUSY if any pages could not be invalidated.
582 */
583int invalidate_inode_pages2_range(struct address_space *mapping,
584 pgoff_t start, pgoff_t end)
585{
586 pgoff_t indices[PAGEVEC_SIZE];
587 struct pagevec pvec;
588 pgoff_t index;
589 int i;
590 int ret = 0;
591 int ret2 = 0;
592 int did_range_unmap = 0;
593
594 cleancache_invalidate_inode(mapping);
595 pagevec_init(&pvec, 0);
596 index = start;
597 while (index <= end && pagevec_lookup_entries(&pvec, mapping, index,
598 min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1,
599 indices)) {
600 mem_cgroup_uncharge_start();
601 for (i = 0; i < pagevec_count(&pvec); i++) {
602 struct page *page = pvec.pages[i];
603
604 /* We rely upon deletion not changing page->index */
605 index = indices[i];
606 if (index > end)
607 break;
608
609 if (radix_tree_exceptional_entry(page)) {
610 clear_exceptional_entry(mapping, index, page);
611 continue;
612 }
613
614 lock_page(page);
615 WARN_ON(page->index != index);
616 if (page->mapping != mapping) {
617 unlock_page(page);
618 continue;
619 }
620 wait_on_page_writeback(page);
621 if (page_mapped(page)) {
622 if (!did_range_unmap) {
623 /*
624 * Zap the rest of the file in one hit.
625 */
626 unmap_mapping_range(mapping,
627 (loff_t)index << PAGE_CACHE_SHIFT,
628 (loff_t)(1 + end - index)
629 << PAGE_CACHE_SHIFT,
630 0);
631 did_range_unmap = 1;
632 } else {
633 /*
634 * Just zap this page
635 */
636 unmap_mapping_range(mapping,
637 (loff_t)index << PAGE_CACHE_SHIFT,
638 PAGE_CACHE_SIZE, 0);
639 }
640 }
641 BUG_ON(page_mapped(page));
642 ret2 = do_launder_page(mapping, page);
643 if (ret2 == 0) {
644 if (!invalidate_complete_page2(mapping, page))
645 ret2 = -EBUSY;
646 }
647 if (ret2 < 0)
648 ret = ret2;
649 unlock_page(page);
650 }
651 pagevec_remove_exceptionals(&pvec);
652 pagevec_release(&pvec);
653 mem_cgroup_uncharge_end();
654 cond_resched();
655 index++;
656 }
657 cleancache_invalidate_inode(mapping);
658 return ret;
659}
660EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
661
662/**
663 * invalidate_inode_pages2 - remove all pages from an address_space
664 * @mapping: the address_space
665 *
666 * Any pages which are found to be mapped into pagetables are unmapped prior to
667 * invalidation.
668 *
669 * Returns -EBUSY if any pages could not be invalidated.
670 */
671int invalidate_inode_pages2(struct address_space *mapping)
672{
673 return invalidate_inode_pages2_range(mapping, 0, -1);
674}
675EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
676
677/**
678 * truncate_pagecache - unmap and remove pagecache that has been truncated
679 * @inode: inode
680 * @newsize: new file size
681 *
682 * inode's new i_size must already be written before truncate_pagecache
683 * is called.
684 *
685 * This function should typically be called before the filesystem
686 * releases resources associated with the freed range (eg. deallocates
687 * blocks). This way, pagecache will always stay logically coherent
688 * with on-disk format, and the filesystem would not have to deal with
689 * situations such as writepage being called for a page that has already
690 * had its underlying blocks deallocated.
691 */
692void truncate_pagecache(struct inode *inode, loff_t newsize)
693{
694 struct address_space *mapping = inode->i_mapping;
695 loff_t holebegin = round_up(newsize, PAGE_SIZE);
696
697 /*
698 * unmap_mapping_range is called twice, first simply for
699 * efficiency so that truncate_inode_pages does fewer
700 * single-page unmaps. However after this first call, and
701 * before truncate_inode_pages finishes, it is possible for
702 * private pages to be COWed, which remain after
703 * truncate_inode_pages finishes, hence the second
704 * unmap_mapping_range call must be made for correctness.
705 */
706 unmap_mapping_range(mapping, holebegin, 0, 1);
707 truncate_inode_pages(mapping, newsize);
708 unmap_mapping_range(mapping, holebegin, 0, 1);
709}
710EXPORT_SYMBOL(truncate_pagecache);
711
712/**
713 * truncate_setsize - update inode and pagecache for a new file size
714 * @inode: inode
715 * @newsize: new file size
716 *
717 * truncate_setsize updates i_size and performs pagecache truncation (if
718 * necessary) to @newsize. It will be typically be called from the filesystem's
719 * setattr function when ATTR_SIZE is passed in.
720 *
721 * Must be called with inode_mutex held and before all filesystem specific
722 * block truncation has been performed.
723 */
724void truncate_setsize(struct inode *inode, loff_t newsize)
725{
726 i_size_write(inode, newsize);
727 truncate_pagecache(inode, newsize);
728}
729EXPORT_SYMBOL(truncate_setsize);
730
731/**
732 * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
733 * @inode: inode
734 * @lstart: offset of beginning of hole
735 * @lend: offset of last byte of hole
736 *
737 * This function should typically be called before the filesystem
738 * releases resources associated with the freed range (eg. deallocates
739 * blocks). This way, pagecache will always stay logically coherent
740 * with on-disk format, and the filesystem would not have to deal with
741 * situations such as writepage being called for a page that has already
742 * had its underlying blocks deallocated.
743 */
744void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
745{
746 struct address_space *mapping = inode->i_mapping;
747 loff_t unmap_start = round_up(lstart, PAGE_SIZE);
748 loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
749 /*
750 * This rounding is currently just for example: unmap_mapping_range
751 * expands its hole outwards, whereas we want it to contract the hole
752 * inwards. However, existing callers of truncate_pagecache_range are
753 * doing their own page rounding first. Note that unmap_mapping_range
754 * allows holelen 0 for all, and we allow lend -1 for end of file.
755 */
756
757 /*
758 * Unlike in truncate_pagecache, unmap_mapping_range is called only
759 * once (before truncating pagecache), and without "even_cows" flag:
760 * hole-punching should not remove private COWed pages from the hole.
761 */
762 if ((u64)unmap_end > (u64)unmap_start)
763 unmap_mapping_range(mapping, unmap_start,
764 1 + unmap_end - unmap_start, 0);
765 truncate_inode_pages_range(mapping, lstart, lend);
766}
767EXPORT_SYMBOL(truncate_pagecache_range);