xref: /openbmc/linux/mm/truncate.c (revision f35e839a)
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  
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   */
40  void 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  
52  static inline void truncate_partial_page(struct page *page, unsigned partial)
53  {
54  	zero_user_segment(page, partial, PAGE_CACHE_SIZE);
55  	cleancache_invalidate_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   */
74  void 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  }
87  EXPORT_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   */
99  static int
100  truncate_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  	ClearPageMappedToDisk(page);
111  	delete_from_page_cache(page);
112  	return 0;
113  }
114  
115  /*
116   * This is for invalidate_mapping_pages().  That function can be called at
117   * any time, and is not supposed to throw away dirty pages.  But pages can
118   * be marked dirty at any time too, so use remove_mapping which safely
119   * discards clean, unused pages.
120   *
121   * Returns non-zero if the page was successfully invalidated.
122   */
123  static int
124  invalidate_complete_page(struct address_space *mapping, struct page *page)
125  {
126  	int ret;
127  
128  	if (page->mapping != mapping)
129  		return 0;
130  
131  	if (page_has_private(page) && !try_to_release_page(page, 0))
132  		return 0;
133  
134  	ret = remove_mapping(mapping, page);
135  
136  	return ret;
137  }
138  
139  int truncate_inode_page(struct address_space *mapping, struct page *page)
140  {
141  	if (page_mapped(page)) {
142  		unmap_mapping_range(mapping,
143  				   (loff_t)page->index << PAGE_CACHE_SHIFT,
144  				   PAGE_CACHE_SIZE, 0);
145  	}
146  	return truncate_complete_page(mapping, page);
147  }
148  
149  /*
150   * Used to get rid of pages on hardware memory corruption.
151   */
152  int generic_error_remove_page(struct address_space *mapping, struct page *page)
153  {
154  	if (!mapping)
155  		return -EINVAL;
156  	/*
157  	 * Only punch for normal data pages for now.
158  	 * Handling other types like directories would need more auditing.
159  	 */
160  	if (!S_ISREG(mapping->host->i_mode))
161  		return -EIO;
162  	return truncate_inode_page(mapping, page);
163  }
164  EXPORT_SYMBOL(generic_error_remove_page);
165  
166  /*
167   * Safely invalidate one page from its pagecache mapping.
168   * It only drops clean, unused pages. The page must be locked.
169   *
170   * Returns 1 if the page is successfully invalidated, otherwise 0.
171   */
172  int invalidate_inode_page(struct page *page)
173  {
174  	struct address_space *mapping = page_mapping(page);
175  	if (!mapping)
176  		return 0;
177  	if (PageDirty(page) || PageWriteback(page))
178  		return 0;
179  	if (page_mapped(page))
180  		return 0;
181  	return invalidate_complete_page(mapping, page);
182  }
183  
184  /**
185   * truncate_inode_pages_range - truncate range of pages specified by start & end byte offsets
186   * @mapping: mapping to truncate
187   * @lstart: offset from which to truncate
188   * @lend: offset to which to truncate
189   *
190   * Truncate the page cache, removing the pages that are between
191   * specified offsets (and zeroing out partial page
192   * (if lstart is not page aligned)).
193   *
194   * Truncate takes two passes - the first pass is nonblocking.  It will not
195   * block on page locks and it will not block on writeback.  The second pass
196   * will wait.  This is to prevent as much IO as possible in the affected region.
197   * The first pass will remove most pages, so the search cost of the second pass
198   * is low.
199   *
200   * We pass down the cache-hot hint to the page freeing code.  Even if the
201   * mapping is large, it is probably the case that the final pages are the most
202   * recently touched, and freeing happens in ascending file offset order.
203   */
204  void truncate_inode_pages_range(struct address_space *mapping,
205  				loff_t lstart, loff_t lend)
206  {
207  	const pgoff_t start = (lstart + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
208  	const unsigned partial = lstart & (PAGE_CACHE_SIZE - 1);
209  	struct pagevec pvec;
210  	pgoff_t index;
211  	pgoff_t end;
212  	int i;
213  
214  	cleancache_invalidate_inode(mapping);
215  	if (mapping->nrpages == 0)
216  		return;
217  
218  	BUG_ON((lend & (PAGE_CACHE_SIZE - 1)) != (PAGE_CACHE_SIZE - 1));
219  	end = (lend >> PAGE_CACHE_SHIFT);
220  
221  	pagevec_init(&pvec, 0);
222  	index = start;
223  	while (index <= end && pagevec_lookup(&pvec, mapping, index,
224  			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
225  		mem_cgroup_uncharge_start();
226  		for (i = 0; i < pagevec_count(&pvec); i++) {
227  			struct page *page = pvec.pages[i];
228  
229  			/* We rely upon deletion not changing page->index */
230  			index = page->index;
231  			if (index > end)
232  				break;
233  
234  			if (!trylock_page(page))
235  				continue;
236  			WARN_ON(page->index != index);
237  			if (PageWriteback(page)) {
238  				unlock_page(page);
239  				continue;
240  			}
241  			truncate_inode_page(mapping, page);
242  			unlock_page(page);
243  		}
244  		pagevec_release(&pvec);
245  		mem_cgroup_uncharge_end();
246  		cond_resched();
247  		index++;
248  	}
249  
250  	if (partial) {
251  		struct page *page = find_lock_page(mapping, start - 1);
252  		if (page) {
253  			wait_on_page_writeback(page);
254  			truncate_partial_page(page, partial);
255  			unlock_page(page);
256  			page_cache_release(page);
257  		}
258  	}
259  
260  	index = start;
261  	for ( ; ; ) {
262  		cond_resched();
263  		if (!pagevec_lookup(&pvec, mapping, index,
264  			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
265  			if (index == start)
266  				break;
267  			index = start;
268  			continue;
269  		}
270  		if (index == start && pvec.pages[0]->index > end) {
271  			pagevec_release(&pvec);
272  			break;
273  		}
274  		mem_cgroup_uncharge_start();
275  		for (i = 0; i < pagevec_count(&pvec); i++) {
276  			struct page *page = pvec.pages[i];
277  
278  			/* We rely upon deletion not changing page->index */
279  			index = page->index;
280  			if (index > end)
281  				break;
282  
283  			lock_page(page);
284  			WARN_ON(page->index != index);
285  			wait_on_page_writeback(page);
286  			truncate_inode_page(mapping, page);
287  			unlock_page(page);
288  		}
289  		pagevec_release(&pvec);
290  		mem_cgroup_uncharge_end();
291  		index++;
292  	}
293  	cleancache_invalidate_inode(mapping);
294  }
295  EXPORT_SYMBOL(truncate_inode_pages_range);
296  
297  /**
298   * truncate_inode_pages - truncate *all* the pages from an offset
299   * @mapping: mapping to truncate
300   * @lstart: offset from which to truncate
301   *
302   * Called under (and serialised by) inode->i_mutex.
303   *
304   * Note: When this function returns, there can be a page in the process of
305   * deletion (inside __delete_from_page_cache()) in the specified range.  Thus
306   * mapping->nrpages can be non-zero when this function returns even after
307   * truncation of the whole mapping.
308   */
309  void truncate_inode_pages(struct address_space *mapping, loff_t lstart)
310  {
311  	truncate_inode_pages_range(mapping, lstart, (loff_t)-1);
312  }
313  EXPORT_SYMBOL(truncate_inode_pages);
314  
315  /**
316   * invalidate_mapping_pages - Invalidate all the unlocked pages of one inode
317   * @mapping: the address_space which holds the pages to invalidate
318   * @start: the offset 'from' which to invalidate
319   * @end: the offset 'to' which to invalidate (inclusive)
320   *
321   * This function only removes the unlocked pages, if you want to
322   * remove all the pages of one inode, you must call truncate_inode_pages.
323   *
324   * invalidate_mapping_pages() will not block on IO activity. It will not
325   * invalidate pages which are dirty, locked, under writeback or mapped into
326   * pagetables.
327   */
328  unsigned long invalidate_mapping_pages(struct address_space *mapping,
329  		pgoff_t start, pgoff_t end)
330  {
331  	struct pagevec pvec;
332  	pgoff_t index = start;
333  	unsigned long ret;
334  	unsigned long count = 0;
335  	int i;
336  
337  	/*
338  	 * Note: this function may get called on a shmem/tmpfs mapping:
339  	 * pagevec_lookup() might then return 0 prematurely (because it
340  	 * got a gangful of swap entries); but it's hardly worth worrying
341  	 * about - it can rarely have anything to free from such a mapping
342  	 * (most pages are dirty), and already skips over any difficulties.
343  	 */
344  
345  	pagevec_init(&pvec, 0);
346  	while (index <= end && pagevec_lookup(&pvec, mapping, index,
347  			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
348  		mem_cgroup_uncharge_start();
349  		for (i = 0; i < pagevec_count(&pvec); i++) {
350  			struct page *page = pvec.pages[i];
351  
352  			/* We rely upon deletion not changing page->index */
353  			index = page->index;
354  			if (index > end)
355  				break;
356  
357  			if (!trylock_page(page))
358  				continue;
359  			WARN_ON(page->index != index);
360  			ret = invalidate_inode_page(page);
361  			unlock_page(page);
362  			/*
363  			 * Invalidation is a hint that the page is no longer
364  			 * of interest and try to speed up its reclaim.
365  			 */
366  			if (!ret)
367  				deactivate_page(page);
368  			count += ret;
369  		}
370  		pagevec_release(&pvec);
371  		mem_cgroup_uncharge_end();
372  		cond_resched();
373  		index++;
374  	}
375  	return count;
376  }
377  EXPORT_SYMBOL(invalidate_mapping_pages);
378  
379  /*
380   * This is like invalidate_complete_page(), except it ignores the page's
381   * refcount.  We do this because invalidate_inode_pages2() needs stronger
382   * invalidation guarantees, and cannot afford to leave pages behind because
383   * shrink_page_list() has a temp ref on them, or because they're transiently
384   * sitting in the lru_cache_add() pagevecs.
385   */
386  static int
387  invalidate_complete_page2(struct address_space *mapping, struct page *page)
388  {
389  	if (page->mapping != mapping)
390  		return 0;
391  
392  	if (page_has_private(page) && !try_to_release_page(page, GFP_KERNEL))
393  		return 0;
394  
395  	spin_lock_irq(&mapping->tree_lock);
396  	if (PageDirty(page))
397  		goto failed;
398  
399  	BUG_ON(page_has_private(page));
400  	__delete_from_page_cache(page);
401  	spin_unlock_irq(&mapping->tree_lock);
402  	mem_cgroup_uncharge_cache_page(page);
403  
404  	if (mapping->a_ops->freepage)
405  		mapping->a_ops->freepage(page);
406  
407  	page_cache_release(page);	/* pagecache ref */
408  	return 1;
409  failed:
410  	spin_unlock_irq(&mapping->tree_lock);
411  	return 0;
412  }
413  
414  static int do_launder_page(struct address_space *mapping, struct page *page)
415  {
416  	if (!PageDirty(page))
417  		return 0;
418  	if (page->mapping != mapping || mapping->a_ops->launder_page == NULL)
419  		return 0;
420  	return mapping->a_ops->launder_page(page);
421  }
422  
423  /**
424   * invalidate_inode_pages2_range - remove range of pages from an address_space
425   * @mapping: the address_space
426   * @start: the page offset 'from' which to invalidate
427   * @end: the page offset 'to' which to invalidate (inclusive)
428   *
429   * Any pages which are found to be mapped into pagetables are unmapped prior to
430   * invalidation.
431   *
432   * Returns -EBUSY if any pages could not be invalidated.
433   */
434  int invalidate_inode_pages2_range(struct address_space *mapping,
435  				  pgoff_t start, pgoff_t end)
436  {
437  	struct pagevec pvec;
438  	pgoff_t index;
439  	int i;
440  	int ret = 0;
441  	int ret2 = 0;
442  	int did_range_unmap = 0;
443  
444  	cleancache_invalidate_inode(mapping);
445  	pagevec_init(&pvec, 0);
446  	index = start;
447  	while (index <= end && pagevec_lookup(&pvec, mapping, index,
448  			min(end - index, (pgoff_t)PAGEVEC_SIZE - 1) + 1)) {
449  		mem_cgroup_uncharge_start();
450  		for (i = 0; i < pagevec_count(&pvec); i++) {
451  			struct page *page = pvec.pages[i];
452  
453  			/* We rely upon deletion not changing page->index */
454  			index = page->index;
455  			if (index > end)
456  				break;
457  
458  			lock_page(page);
459  			WARN_ON(page->index != index);
460  			if (page->mapping != mapping) {
461  				unlock_page(page);
462  				continue;
463  			}
464  			wait_on_page_writeback(page);
465  			if (page_mapped(page)) {
466  				if (!did_range_unmap) {
467  					/*
468  					 * Zap the rest of the file in one hit.
469  					 */
470  					unmap_mapping_range(mapping,
471  					   (loff_t)index << PAGE_CACHE_SHIFT,
472  					   (loff_t)(1 + end - index)
473  							 << PAGE_CACHE_SHIFT,
474  					    0);
475  					did_range_unmap = 1;
476  				} else {
477  					/*
478  					 * Just zap this page
479  					 */
480  					unmap_mapping_range(mapping,
481  					   (loff_t)index << PAGE_CACHE_SHIFT,
482  					   PAGE_CACHE_SIZE, 0);
483  				}
484  			}
485  			BUG_ON(page_mapped(page));
486  			ret2 = do_launder_page(mapping, page);
487  			if (ret2 == 0) {
488  				if (!invalidate_complete_page2(mapping, page))
489  					ret2 = -EBUSY;
490  			}
491  			if (ret2 < 0)
492  				ret = ret2;
493  			unlock_page(page);
494  		}
495  		pagevec_release(&pvec);
496  		mem_cgroup_uncharge_end();
497  		cond_resched();
498  		index++;
499  	}
500  	cleancache_invalidate_inode(mapping);
501  	return ret;
502  }
503  EXPORT_SYMBOL_GPL(invalidate_inode_pages2_range);
504  
505  /**
506   * invalidate_inode_pages2 - remove all pages from an address_space
507   * @mapping: the address_space
508   *
509   * Any pages which are found to be mapped into pagetables are unmapped prior to
510   * invalidation.
511   *
512   * Returns -EBUSY if any pages could not be invalidated.
513   */
514  int invalidate_inode_pages2(struct address_space *mapping)
515  {
516  	return invalidate_inode_pages2_range(mapping, 0, -1);
517  }
518  EXPORT_SYMBOL_GPL(invalidate_inode_pages2);
519  
520  /**
521   * truncate_pagecache - unmap and remove pagecache that has been truncated
522   * @inode: inode
523   * @oldsize: old file size
524   * @newsize: new file size
525   *
526   * inode's new i_size must already be written before truncate_pagecache
527   * is called.
528   *
529   * This function should typically be called before the filesystem
530   * releases resources associated with the freed range (eg. deallocates
531   * blocks). This way, pagecache will always stay logically coherent
532   * with on-disk format, and the filesystem would not have to deal with
533   * situations such as writepage being called for a page that has already
534   * had its underlying blocks deallocated.
535   */
536  void truncate_pagecache(struct inode *inode, loff_t oldsize, loff_t newsize)
537  {
538  	struct address_space *mapping = inode->i_mapping;
539  	loff_t holebegin = round_up(newsize, PAGE_SIZE);
540  
541  	/*
542  	 * unmap_mapping_range is called twice, first simply for
543  	 * efficiency so that truncate_inode_pages does fewer
544  	 * single-page unmaps.  However after this first call, and
545  	 * before truncate_inode_pages finishes, it is possible for
546  	 * private pages to be COWed, which remain after
547  	 * truncate_inode_pages finishes, hence the second
548  	 * unmap_mapping_range call must be made for correctness.
549  	 */
550  	unmap_mapping_range(mapping, holebegin, 0, 1);
551  	truncate_inode_pages(mapping, newsize);
552  	unmap_mapping_range(mapping, holebegin, 0, 1);
553  }
554  EXPORT_SYMBOL(truncate_pagecache);
555  
556  /**
557   * truncate_setsize - update inode and pagecache for a new file size
558   * @inode: inode
559   * @newsize: new file size
560   *
561   * truncate_setsize updates i_size and performs pagecache truncation (if
562   * necessary) to @newsize. It will be typically be called from the filesystem's
563   * setattr function when ATTR_SIZE is passed in.
564   *
565   * Must be called with inode_mutex held and before all filesystem specific
566   * block truncation has been performed.
567   */
568  void truncate_setsize(struct inode *inode, loff_t newsize)
569  {
570  	loff_t oldsize;
571  
572  	oldsize = inode->i_size;
573  	i_size_write(inode, newsize);
574  
575  	truncate_pagecache(inode, oldsize, newsize);
576  }
577  EXPORT_SYMBOL(truncate_setsize);
578  
579  /**
580   * truncate_pagecache_range - unmap and remove pagecache that is hole-punched
581   * @inode: inode
582   * @lstart: offset of beginning of hole
583   * @lend: offset of last byte of hole
584   *
585   * This function should typically be called before the filesystem
586   * releases resources associated with the freed range (eg. deallocates
587   * blocks). This way, pagecache will always stay logically coherent
588   * with on-disk format, and the filesystem would not have to deal with
589   * situations such as writepage being called for a page that has already
590   * had its underlying blocks deallocated.
591   */
592  void truncate_pagecache_range(struct inode *inode, loff_t lstart, loff_t lend)
593  {
594  	struct address_space *mapping = inode->i_mapping;
595  	loff_t unmap_start = round_up(lstart, PAGE_SIZE);
596  	loff_t unmap_end = round_down(1 + lend, PAGE_SIZE) - 1;
597  	/*
598  	 * This rounding is currently just for example: unmap_mapping_range
599  	 * expands its hole outwards, whereas we want it to contract the hole
600  	 * inwards.  However, existing callers of truncate_pagecache_range are
601  	 * doing their own page rounding first; and truncate_inode_pages_range
602  	 * currently BUGs if lend is not pagealigned-1 (it handles partial
603  	 * page at start of hole, but not partial page at end of hole).  Note
604  	 * unmap_mapping_range allows holelen 0 for all, and we allow lend -1.
605  	 */
606  
607  	/*
608  	 * Unlike in truncate_pagecache, unmap_mapping_range is called only
609  	 * once (before truncating pagecache), and without "even_cows" flag:
610  	 * hole-punching should not remove private COWed pages from the hole.
611  	 */
612  	if ((u64)unmap_end > (u64)unmap_start)
613  		unmap_mapping_range(mapping, unmap_start,
614  				    1 + unmap_end - unmap_start, 0);
615  	truncate_inode_pages_range(mapping, lstart, lend);
616  }
617  EXPORT_SYMBOL(truncate_pagecache_range);
618