dax: New fault locking
[sfrench/cifs-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/dax.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/cpuset.h>
33 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34 #include <linux/hugetlb.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cleancache.h>
37 #include <linux/rmap.h>
38 #include "internal.h"
39
40 #define CREATE_TRACE_POINTS
41 #include <trace/events/filemap.h>
42
43 /*
44  * FIXME: remove all knowledge of the buffer layer from the core VM
45  */
46 #include <linux/buffer_head.h> /* for try_to_free_buffers */
47
48 #include <asm/mman.h>
49
50 /*
51  * Shared mappings implemented 30.11.1994. It's not fully working yet,
52  * though.
53  *
54  * Shared mappings now work. 15.8.1995  Bruno.
55  *
56  * finished 'unifying' the page and buffer cache and SMP-threaded the
57  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
58  *
59  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
60  */
61
62 /*
63  * Lock ordering:
64  *
65  *  ->i_mmap_rwsem              (truncate_pagecache)
66  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
67  *      ->swap_lock             (exclusive_swap_page, others)
68  *        ->mapping->tree_lock
69  *
70  *  ->i_mutex
71  *    ->i_mmap_rwsem            (truncate->unmap_mapping_range)
72  *
73  *  ->mmap_sem
74  *    ->i_mmap_rwsem
75  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
76  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
77  *
78  *  ->mmap_sem
79  *    ->lock_page               (access_process_vm)
80  *
81  *  ->i_mutex                   (generic_perform_write)
82  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
83  *
84  *  bdi->wb.list_lock
85  *    sb_lock                   (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_rwsem
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
103  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
104  *    ->memcg->move_lock        (page_remove_rmap->lock_page_memcg)
105  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
106  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
107  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
108  *
109  * ->i_mmap_rwsem
110  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
111  */
112
113 static void page_cache_tree_delete(struct address_space *mapping,
114                                    struct page *page, void *shadow)
115 {
116         struct radix_tree_node *node;
117         unsigned long index;
118         unsigned int offset;
119         unsigned int tag;
120         void **slot;
121
122         VM_BUG_ON(!PageLocked(page));
123
124         __radix_tree_lookup(&mapping->page_tree, page->index, &node, &slot);
125
126         if (shadow) {
127                 mapping->nrexceptional++;
128                 /*
129                  * Make sure the nrexceptional update is committed before
130                  * the nrpages update so that final truncate racing
131                  * with reclaim does not see both counters 0 at the
132                  * same time and miss a shadow entry.
133                  */
134                 smp_wmb();
135         }
136         mapping->nrpages--;
137
138         if (!node) {
139                 /* Clear direct pointer tags in root node */
140                 mapping->page_tree.gfp_mask &= __GFP_BITS_MASK;
141                 radix_tree_replace_slot(slot, shadow);
142                 return;
143         }
144
145         /* Clear tree tags for the removed page */
146         index = page->index;
147         offset = index & RADIX_TREE_MAP_MASK;
148         for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
149                 if (test_bit(offset, node->tags[tag]))
150                         radix_tree_tag_clear(&mapping->page_tree, index, tag);
151         }
152
153         /* Delete page, swap shadow entry */
154         radix_tree_replace_slot(slot, shadow);
155         workingset_node_pages_dec(node);
156         if (shadow)
157                 workingset_node_shadows_inc(node);
158         else
159                 if (__radix_tree_delete_node(&mapping->page_tree, node))
160                         return;
161
162         /*
163          * Track node that only contains shadow entries. DAX mappings contain
164          * no shadow entries and may contain other exceptional entries so skip
165          * those.
166          *
167          * Avoid acquiring the list_lru lock if already tracked.  The
168          * list_empty() test is safe as node->private_list is
169          * protected by mapping->tree_lock.
170          */
171         if (!dax_mapping(mapping) && !workingset_node_pages(node) &&
172             list_empty(&node->private_list)) {
173                 node->private_data = mapping;
174                 list_lru_add(&workingset_shadow_nodes, &node->private_list);
175         }
176 }
177
178 /*
179  * Delete a page from the page cache and free it. Caller has to make
180  * sure the page is locked and that nobody else uses it - or that usage
181  * is safe.  The caller must hold the mapping's tree_lock.
182  */
183 void __delete_from_page_cache(struct page *page, void *shadow)
184 {
185         struct address_space *mapping = page->mapping;
186
187         trace_mm_filemap_delete_from_page_cache(page);
188         /*
189          * if we're uptodate, flush out into the cleancache, otherwise
190          * invalidate any existing cleancache entries.  We can't leave
191          * stale data around in the cleancache once our page is gone
192          */
193         if (PageUptodate(page) && PageMappedToDisk(page))
194                 cleancache_put_page(page);
195         else
196                 cleancache_invalidate_page(mapping, page);
197
198         VM_BUG_ON_PAGE(page_mapped(page), page);
199         if (!IS_ENABLED(CONFIG_DEBUG_VM) && unlikely(page_mapped(page))) {
200                 int mapcount;
201
202                 pr_alert("BUG: Bad page cache in process %s  pfn:%05lx\n",
203                          current->comm, page_to_pfn(page));
204                 dump_page(page, "still mapped when deleted");
205                 dump_stack();
206                 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
207
208                 mapcount = page_mapcount(page);
209                 if (mapping_exiting(mapping) &&
210                     page_count(page) >= mapcount + 2) {
211                         /*
212                          * All vmas have already been torn down, so it's
213                          * a good bet that actually the page is unmapped,
214                          * and we'd prefer not to leak it: if we're wrong,
215                          * some other bad page check should catch it later.
216                          */
217                         page_mapcount_reset(page);
218                         atomic_sub(mapcount, &page->_count);
219                 }
220         }
221
222         page_cache_tree_delete(mapping, page, shadow);
223
224         page->mapping = NULL;
225         /* Leave page->index set: truncation lookup relies upon it */
226
227         /* hugetlb pages do not participate in page cache accounting. */
228         if (!PageHuge(page))
229                 __dec_zone_page_state(page, NR_FILE_PAGES);
230         if (PageSwapBacked(page))
231                 __dec_zone_page_state(page, NR_SHMEM);
232
233         /*
234          * At this point page must be either written or cleaned by truncate.
235          * Dirty page here signals a bug and loss of unwritten data.
236          *
237          * This fixes dirty accounting after removing the page entirely but
238          * leaves PageDirty set: it has no effect for truncated page and
239          * anyway will be cleared before returning page into buddy allocator.
240          */
241         if (WARN_ON_ONCE(PageDirty(page)))
242                 account_page_cleaned(page, mapping, inode_to_wb(mapping->host));
243 }
244
245 /**
246  * delete_from_page_cache - delete page from page cache
247  * @page: the page which the kernel is trying to remove from page cache
248  *
249  * This must be called only on pages that have been verified to be in the page
250  * cache and locked.  It will never put the page into the free list, the caller
251  * has a reference on the page.
252  */
253 void delete_from_page_cache(struct page *page)
254 {
255         struct address_space *mapping = page->mapping;
256         unsigned long flags;
257
258         void (*freepage)(struct page *);
259
260         BUG_ON(!PageLocked(page));
261
262         freepage = mapping->a_ops->freepage;
263
264         spin_lock_irqsave(&mapping->tree_lock, flags);
265         __delete_from_page_cache(page, NULL);
266         spin_unlock_irqrestore(&mapping->tree_lock, flags);
267
268         if (freepage)
269                 freepage(page);
270         put_page(page);
271 }
272 EXPORT_SYMBOL(delete_from_page_cache);
273
274 static int filemap_check_errors(struct address_space *mapping)
275 {
276         int ret = 0;
277         /* Check for outstanding write errors */
278         if (test_bit(AS_ENOSPC, &mapping->flags) &&
279             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
280                 ret = -ENOSPC;
281         if (test_bit(AS_EIO, &mapping->flags) &&
282             test_and_clear_bit(AS_EIO, &mapping->flags))
283                 ret = -EIO;
284         return ret;
285 }
286
287 /**
288  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
289  * @mapping:    address space structure to write
290  * @start:      offset in bytes where the range starts
291  * @end:        offset in bytes where the range ends (inclusive)
292  * @sync_mode:  enable synchronous operation
293  *
294  * Start writeback against all of a mapping's dirty pages that lie
295  * within the byte offsets <start, end> inclusive.
296  *
297  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
298  * opposed to a regular memory cleansing writeback.  The difference between
299  * these two operations is that if a dirty page/buffer is encountered, it must
300  * be waited upon, and not just skipped over.
301  */
302 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
303                                 loff_t end, int sync_mode)
304 {
305         int ret;
306         struct writeback_control wbc = {
307                 .sync_mode = sync_mode,
308                 .nr_to_write = LONG_MAX,
309                 .range_start = start,
310                 .range_end = end,
311         };
312
313         if (!mapping_cap_writeback_dirty(mapping))
314                 return 0;
315
316         wbc_attach_fdatawrite_inode(&wbc, mapping->host);
317         ret = do_writepages(mapping, &wbc);
318         wbc_detach_inode(&wbc);
319         return ret;
320 }
321
322 static inline int __filemap_fdatawrite(struct address_space *mapping,
323         int sync_mode)
324 {
325         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
326 }
327
328 int filemap_fdatawrite(struct address_space *mapping)
329 {
330         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
331 }
332 EXPORT_SYMBOL(filemap_fdatawrite);
333
334 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
335                                 loff_t end)
336 {
337         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
338 }
339 EXPORT_SYMBOL(filemap_fdatawrite_range);
340
341 /**
342  * filemap_flush - mostly a non-blocking flush
343  * @mapping:    target address_space
344  *
345  * This is a mostly non-blocking flush.  Not suitable for data-integrity
346  * purposes - I/O may not be started against all dirty pages.
347  */
348 int filemap_flush(struct address_space *mapping)
349 {
350         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
351 }
352 EXPORT_SYMBOL(filemap_flush);
353
354 static int __filemap_fdatawait_range(struct address_space *mapping,
355                                      loff_t start_byte, loff_t end_byte)
356 {
357         pgoff_t index = start_byte >> PAGE_SHIFT;
358         pgoff_t end = end_byte >> PAGE_SHIFT;
359         struct pagevec pvec;
360         int nr_pages;
361         int ret = 0;
362
363         if (end_byte < start_byte)
364                 goto out;
365
366         pagevec_init(&pvec, 0);
367         while ((index <= end) &&
368                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
369                         PAGECACHE_TAG_WRITEBACK,
370                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
371                 unsigned i;
372
373                 for (i = 0; i < nr_pages; i++) {
374                         struct page *page = pvec.pages[i];
375
376                         /* until radix tree lookup accepts end_index */
377                         if (page->index > end)
378                                 continue;
379
380                         wait_on_page_writeback(page);
381                         if (TestClearPageError(page))
382                                 ret = -EIO;
383                 }
384                 pagevec_release(&pvec);
385                 cond_resched();
386         }
387 out:
388         return ret;
389 }
390
391 /**
392  * filemap_fdatawait_range - wait for writeback to complete
393  * @mapping:            address space structure to wait for
394  * @start_byte:         offset in bytes where the range starts
395  * @end_byte:           offset in bytes where the range ends (inclusive)
396  *
397  * Walk the list of under-writeback pages of the given address space
398  * in the given range and wait for all of them.  Check error status of
399  * the address space and return it.
400  *
401  * Since the error status of the address space is cleared by this function,
402  * callers are responsible for checking the return value and handling and/or
403  * reporting the error.
404  */
405 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
406                             loff_t end_byte)
407 {
408         int ret, ret2;
409
410         ret = __filemap_fdatawait_range(mapping, start_byte, end_byte);
411         ret2 = filemap_check_errors(mapping);
412         if (!ret)
413                 ret = ret2;
414
415         return ret;
416 }
417 EXPORT_SYMBOL(filemap_fdatawait_range);
418
419 /**
420  * filemap_fdatawait_keep_errors - wait for writeback without clearing errors
421  * @mapping: address space structure to wait for
422  *
423  * Walk the list of under-writeback pages of the given address space
424  * and wait for all of them.  Unlike filemap_fdatawait(), this function
425  * does not clear error status of the address space.
426  *
427  * Use this function if callers don't handle errors themselves.  Expected
428  * call sites are system-wide / filesystem-wide data flushers: e.g. sync(2),
429  * fsfreeze(8)
430  */
431 void filemap_fdatawait_keep_errors(struct address_space *mapping)
432 {
433         loff_t i_size = i_size_read(mapping->host);
434
435         if (i_size == 0)
436                 return;
437
438         __filemap_fdatawait_range(mapping, 0, i_size - 1);
439 }
440
441 /**
442  * filemap_fdatawait - wait for all under-writeback pages to complete
443  * @mapping: address space structure to wait for
444  *
445  * Walk the list of under-writeback pages of the given address space
446  * and wait for all of them.  Check error status of the address space
447  * and return it.
448  *
449  * Since the error status of the address space is cleared by this function,
450  * callers are responsible for checking the return value and handling and/or
451  * reporting the error.
452  */
453 int filemap_fdatawait(struct address_space *mapping)
454 {
455         loff_t i_size = i_size_read(mapping->host);
456
457         if (i_size == 0)
458                 return 0;
459
460         return filemap_fdatawait_range(mapping, 0, i_size - 1);
461 }
462 EXPORT_SYMBOL(filemap_fdatawait);
463
464 int filemap_write_and_wait(struct address_space *mapping)
465 {
466         int err = 0;
467
468         if ((!dax_mapping(mapping) && mapping->nrpages) ||
469             (dax_mapping(mapping) && mapping->nrexceptional)) {
470                 err = filemap_fdatawrite(mapping);
471                 /*
472                  * Even if the above returned error, the pages may be
473                  * written partially (e.g. -ENOSPC), so we wait for it.
474                  * But the -EIO is special case, it may indicate the worst
475                  * thing (e.g. bug) happened, so we avoid waiting for it.
476                  */
477                 if (err != -EIO) {
478                         int err2 = filemap_fdatawait(mapping);
479                         if (!err)
480                                 err = err2;
481                 }
482         } else {
483                 err = filemap_check_errors(mapping);
484         }
485         return err;
486 }
487 EXPORT_SYMBOL(filemap_write_and_wait);
488
489 /**
490  * filemap_write_and_wait_range - write out & wait on a file range
491  * @mapping:    the address_space for the pages
492  * @lstart:     offset in bytes where the range starts
493  * @lend:       offset in bytes where the range ends (inclusive)
494  *
495  * Write out and wait upon file offsets lstart->lend, inclusive.
496  *
497  * Note that `lend' is inclusive (describes the last byte to be written) so
498  * that this function can be used to write to the very end-of-file (end = -1).
499  */
500 int filemap_write_and_wait_range(struct address_space *mapping,
501                                  loff_t lstart, loff_t lend)
502 {
503         int err = 0;
504
505         if ((!dax_mapping(mapping) && mapping->nrpages) ||
506             (dax_mapping(mapping) && mapping->nrexceptional)) {
507                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
508                                                  WB_SYNC_ALL);
509                 /* See comment of filemap_write_and_wait() */
510                 if (err != -EIO) {
511                         int err2 = filemap_fdatawait_range(mapping,
512                                                 lstart, lend);
513                         if (!err)
514                                 err = err2;
515                 }
516         } else {
517                 err = filemap_check_errors(mapping);
518         }
519         return err;
520 }
521 EXPORT_SYMBOL(filemap_write_and_wait_range);
522
523 /**
524  * replace_page_cache_page - replace a pagecache page with a new one
525  * @old:        page to be replaced
526  * @new:        page to replace with
527  * @gfp_mask:   allocation mode
528  *
529  * This function replaces a page in the pagecache with a new one.  On
530  * success it acquires the pagecache reference for the new page and
531  * drops it for the old page.  Both the old and new pages must be
532  * locked.  This function does not add the new page to the LRU, the
533  * caller must do that.
534  *
535  * The remove + add is atomic.  The only way this function can fail is
536  * memory allocation failure.
537  */
538 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
539 {
540         int error;
541
542         VM_BUG_ON_PAGE(!PageLocked(old), old);
543         VM_BUG_ON_PAGE(!PageLocked(new), new);
544         VM_BUG_ON_PAGE(new->mapping, new);
545
546         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
547         if (!error) {
548                 struct address_space *mapping = old->mapping;
549                 void (*freepage)(struct page *);
550                 unsigned long flags;
551
552                 pgoff_t offset = old->index;
553                 freepage = mapping->a_ops->freepage;
554
555                 get_page(new);
556                 new->mapping = mapping;
557                 new->index = offset;
558
559                 spin_lock_irqsave(&mapping->tree_lock, flags);
560                 __delete_from_page_cache(old, NULL);
561                 error = radix_tree_insert(&mapping->page_tree, offset, new);
562                 BUG_ON(error);
563                 mapping->nrpages++;
564
565                 /*
566                  * hugetlb pages do not participate in page cache accounting.
567                  */
568                 if (!PageHuge(new))
569                         __inc_zone_page_state(new, NR_FILE_PAGES);
570                 if (PageSwapBacked(new))
571                         __inc_zone_page_state(new, NR_SHMEM);
572                 spin_unlock_irqrestore(&mapping->tree_lock, flags);
573                 mem_cgroup_migrate(old, new);
574                 radix_tree_preload_end();
575                 if (freepage)
576                         freepage(old);
577                 put_page(old);
578         }
579
580         return error;
581 }
582 EXPORT_SYMBOL_GPL(replace_page_cache_page);
583
584 static int page_cache_tree_insert(struct address_space *mapping,
585                                   struct page *page, void **shadowp)
586 {
587         struct radix_tree_node *node;
588         void **slot;
589         int error;
590
591         error = __radix_tree_create(&mapping->page_tree, page->index, 0,
592                                     &node, &slot);
593         if (error)
594                 return error;
595         if (*slot) {
596                 void *p;
597
598                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
599                 if (!radix_tree_exceptional_entry(p))
600                         return -EEXIST;
601
602                 mapping->nrexceptional--;
603                 if (!dax_mapping(mapping)) {
604                         if (shadowp)
605                                 *shadowp = p;
606                         if (node)
607                                 workingset_node_shadows_dec(node);
608                 } else {
609                         /* DAX can replace empty locked entry with a hole */
610                         WARN_ON_ONCE(p !=
611                                 (void *)(RADIX_TREE_EXCEPTIONAL_ENTRY |
612                                          RADIX_DAX_ENTRY_LOCK));
613                         /* DAX accounts exceptional entries as normal pages */
614                         if (node)
615                                 workingset_node_pages_dec(node);
616                         /* Wakeup waiters for exceptional entry lock */
617                         dax_wake_mapping_entry_waiter(mapping, page->index,
618                                                       false);
619                 }
620         }
621         radix_tree_replace_slot(slot, page);
622         mapping->nrpages++;
623         if (node) {
624                 workingset_node_pages_inc(node);
625                 /*
626                  * Don't track node that contains actual pages.
627                  *
628                  * Avoid acquiring the list_lru lock if already
629                  * untracked.  The list_empty() test is safe as
630                  * node->private_list is protected by
631                  * mapping->tree_lock.
632                  */
633                 if (!list_empty(&node->private_list))
634                         list_lru_del(&workingset_shadow_nodes,
635                                      &node->private_list);
636         }
637         return 0;
638 }
639
640 static int __add_to_page_cache_locked(struct page *page,
641                                       struct address_space *mapping,
642                                       pgoff_t offset, gfp_t gfp_mask,
643                                       void **shadowp)
644 {
645         int huge = PageHuge(page);
646         struct mem_cgroup *memcg;
647         int error;
648
649         VM_BUG_ON_PAGE(!PageLocked(page), page);
650         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
651
652         if (!huge) {
653                 error = mem_cgroup_try_charge(page, current->mm,
654                                               gfp_mask, &memcg, false);
655                 if (error)
656                         return error;
657         }
658
659         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
660         if (error) {
661                 if (!huge)
662                         mem_cgroup_cancel_charge(page, memcg, false);
663                 return error;
664         }
665
666         get_page(page);
667         page->mapping = mapping;
668         page->index = offset;
669
670         spin_lock_irq(&mapping->tree_lock);
671         error = page_cache_tree_insert(mapping, page, shadowp);
672         radix_tree_preload_end();
673         if (unlikely(error))
674                 goto err_insert;
675
676         /* hugetlb pages do not participate in page cache accounting. */
677         if (!huge)
678                 __inc_zone_page_state(page, NR_FILE_PAGES);
679         spin_unlock_irq(&mapping->tree_lock);
680         if (!huge)
681                 mem_cgroup_commit_charge(page, memcg, false, false);
682         trace_mm_filemap_add_to_page_cache(page);
683         return 0;
684 err_insert:
685         page->mapping = NULL;
686         /* Leave page->index set: truncation relies upon it */
687         spin_unlock_irq(&mapping->tree_lock);
688         if (!huge)
689                 mem_cgroup_cancel_charge(page, memcg, false);
690         put_page(page);
691         return error;
692 }
693
694 /**
695  * add_to_page_cache_locked - add a locked page to the pagecache
696  * @page:       page to add
697  * @mapping:    the page's address_space
698  * @offset:     page index
699  * @gfp_mask:   page allocation mode
700  *
701  * This function is used to add a page to the pagecache. It must be locked.
702  * This function does not add the page to the LRU.  The caller must do that.
703  */
704 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
705                 pgoff_t offset, gfp_t gfp_mask)
706 {
707         return __add_to_page_cache_locked(page, mapping, offset,
708                                           gfp_mask, NULL);
709 }
710 EXPORT_SYMBOL(add_to_page_cache_locked);
711
712 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
713                                 pgoff_t offset, gfp_t gfp_mask)
714 {
715         void *shadow = NULL;
716         int ret;
717
718         __SetPageLocked(page);
719         ret = __add_to_page_cache_locked(page, mapping, offset,
720                                          gfp_mask, &shadow);
721         if (unlikely(ret))
722                 __ClearPageLocked(page);
723         else {
724                 /*
725                  * The page might have been evicted from cache only
726                  * recently, in which case it should be activated like
727                  * any other repeatedly accessed page.
728                  */
729                 if (shadow && workingset_refault(shadow)) {
730                         SetPageActive(page);
731                         workingset_activation(page);
732                 } else
733                         ClearPageActive(page);
734                 lru_cache_add(page);
735         }
736         return ret;
737 }
738 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
739
740 #ifdef CONFIG_NUMA
741 struct page *__page_cache_alloc(gfp_t gfp)
742 {
743         int n;
744         struct page *page;
745
746         if (cpuset_do_page_mem_spread()) {
747                 unsigned int cpuset_mems_cookie;
748                 do {
749                         cpuset_mems_cookie = read_mems_allowed_begin();
750                         n = cpuset_mem_spread_node();
751                         page = __alloc_pages_node(n, gfp, 0);
752                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
753
754                 return page;
755         }
756         return alloc_pages(gfp, 0);
757 }
758 EXPORT_SYMBOL(__page_cache_alloc);
759 #endif
760
761 /*
762  * In order to wait for pages to become available there must be
763  * waitqueues associated with pages. By using a hash table of
764  * waitqueues where the bucket discipline is to maintain all
765  * waiters on the same queue and wake all when any of the pages
766  * become available, and for the woken contexts to check to be
767  * sure the appropriate page became available, this saves space
768  * at a cost of "thundering herd" phenomena during rare hash
769  * collisions.
770  */
771 wait_queue_head_t *page_waitqueue(struct page *page)
772 {
773         const struct zone *zone = page_zone(page);
774
775         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
776 }
777 EXPORT_SYMBOL(page_waitqueue);
778
779 void wait_on_page_bit(struct page *page, int bit_nr)
780 {
781         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
782
783         if (test_bit(bit_nr, &page->flags))
784                 __wait_on_bit(page_waitqueue(page), &wait, bit_wait_io,
785                                                         TASK_UNINTERRUPTIBLE);
786 }
787 EXPORT_SYMBOL(wait_on_page_bit);
788
789 int wait_on_page_bit_killable(struct page *page, int bit_nr)
790 {
791         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
792
793         if (!test_bit(bit_nr, &page->flags))
794                 return 0;
795
796         return __wait_on_bit(page_waitqueue(page), &wait,
797                              bit_wait_io, TASK_KILLABLE);
798 }
799
800 int wait_on_page_bit_killable_timeout(struct page *page,
801                                        int bit_nr, unsigned long timeout)
802 {
803         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
804
805         wait.key.timeout = jiffies + timeout;
806         if (!test_bit(bit_nr, &page->flags))
807                 return 0;
808         return __wait_on_bit(page_waitqueue(page), &wait,
809                              bit_wait_io_timeout, TASK_KILLABLE);
810 }
811 EXPORT_SYMBOL_GPL(wait_on_page_bit_killable_timeout);
812
813 /**
814  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
815  * @page: Page defining the wait queue of interest
816  * @waiter: Waiter to add to the queue
817  *
818  * Add an arbitrary @waiter to the wait queue for the nominated @page.
819  */
820 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
821 {
822         wait_queue_head_t *q = page_waitqueue(page);
823         unsigned long flags;
824
825         spin_lock_irqsave(&q->lock, flags);
826         __add_wait_queue(q, waiter);
827         spin_unlock_irqrestore(&q->lock, flags);
828 }
829 EXPORT_SYMBOL_GPL(add_page_wait_queue);
830
831 /**
832  * unlock_page - unlock a locked page
833  * @page: the page
834  *
835  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
836  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
837  * mechanism between PageLocked pages and PageWriteback pages is shared.
838  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
839  *
840  * The mb is necessary to enforce ordering between the clear_bit and the read
841  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
842  */
843 void unlock_page(struct page *page)
844 {
845         page = compound_head(page);
846         VM_BUG_ON_PAGE(!PageLocked(page), page);
847         clear_bit_unlock(PG_locked, &page->flags);
848         smp_mb__after_atomic();
849         wake_up_page(page, PG_locked);
850 }
851 EXPORT_SYMBOL(unlock_page);
852
853 /**
854  * end_page_writeback - end writeback against a page
855  * @page: the page
856  */
857 void end_page_writeback(struct page *page)
858 {
859         /*
860          * TestClearPageReclaim could be used here but it is an atomic
861          * operation and overkill in this particular case. Failing to
862          * shuffle a page marked for immediate reclaim is too mild to
863          * justify taking an atomic operation penalty at the end of
864          * ever page writeback.
865          */
866         if (PageReclaim(page)) {
867                 ClearPageReclaim(page);
868                 rotate_reclaimable_page(page);
869         }
870
871         if (!test_clear_page_writeback(page))
872                 BUG();
873
874         smp_mb__after_atomic();
875         wake_up_page(page, PG_writeback);
876 }
877 EXPORT_SYMBOL(end_page_writeback);
878
879 /*
880  * After completing I/O on a page, call this routine to update the page
881  * flags appropriately
882  */
883 void page_endio(struct page *page, int rw, int err)
884 {
885         if (rw == READ) {
886                 if (!err) {
887                         SetPageUptodate(page);
888                 } else {
889                         ClearPageUptodate(page);
890                         SetPageError(page);
891                 }
892                 unlock_page(page);
893         } else { /* rw == WRITE */
894                 if (err) {
895                         SetPageError(page);
896                         if (page->mapping)
897                                 mapping_set_error(page->mapping, err);
898                 }
899                 end_page_writeback(page);
900         }
901 }
902 EXPORT_SYMBOL_GPL(page_endio);
903
904 /**
905  * __lock_page - get a lock on the page, assuming we need to sleep to get it
906  * @page: the page to lock
907  */
908 void __lock_page(struct page *page)
909 {
910         struct page *page_head = compound_head(page);
911         DEFINE_WAIT_BIT(wait, &page_head->flags, PG_locked);
912
913         __wait_on_bit_lock(page_waitqueue(page_head), &wait, bit_wait_io,
914                                                         TASK_UNINTERRUPTIBLE);
915 }
916 EXPORT_SYMBOL(__lock_page);
917
918 int __lock_page_killable(struct page *page)
919 {
920         struct page *page_head = compound_head(page);
921         DEFINE_WAIT_BIT(wait, &page_head->flags, PG_locked);
922
923         return __wait_on_bit_lock(page_waitqueue(page_head), &wait,
924                                         bit_wait_io, TASK_KILLABLE);
925 }
926 EXPORT_SYMBOL_GPL(__lock_page_killable);
927
928 /*
929  * Return values:
930  * 1 - page is locked; mmap_sem is still held.
931  * 0 - page is not locked.
932  *     mmap_sem has been released (up_read()), unless flags had both
933  *     FAULT_FLAG_ALLOW_RETRY and FAULT_FLAG_RETRY_NOWAIT set, in
934  *     which case mmap_sem is still held.
935  *
936  * If neither ALLOW_RETRY nor KILLABLE are set, will always return 1
937  * with the page locked and the mmap_sem unperturbed.
938  */
939 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
940                          unsigned int flags)
941 {
942         if (flags & FAULT_FLAG_ALLOW_RETRY) {
943                 /*
944                  * CAUTION! In this case, mmap_sem is not released
945                  * even though return 0.
946                  */
947                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
948                         return 0;
949
950                 up_read(&mm->mmap_sem);
951                 if (flags & FAULT_FLAG_KILLABLE)
952                         wait_on_page_locked_killable(page);
953                 else
954                         wait_on_page_locked(page);
955                 return 0;
956         } else {
957                 if (flags & FAULT_FLAG_KILLABLE) {
958                         int ret;
959
960                         ret = __lock_page_killable(page);
961                         if (ret) {
962                                 up_read(&mm->mmap_sem);
963                                 return 0;
964                         }
965                 } else
966                         __lock_page(page);
967                 return 1;
968         }
969 }
970
971 /**
972  * page_cache_next_hole - find the next hole (not-present entry)
973  * @mapping: mapping
974  * @index: index
975  * @max_scan: maximum range to search
976  *
977  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
978  * lowest indexed hole.
979  *
980  * Returns: the index of the hole if found, otherwise returns an index
981  * outside of the set specified (in which case 'return - index >=
982  * max_scan' will be true). In rare cases of index wrap-around, 0 will
983  * be returned.
984  *
985  * page_cache_next_hole may be called under rcu_read_lock. However,
986  * like radix_tree_gang_lookup, this will not atomically search a
987  * snapshot of the tree at a single point in time. For example, if a
988  * hole is created at index 5, then subsequently a hole is created at
989  * index 10, page_cache_next_hole covering both indexes may return 10
990  * if called under rcu_read_lock.
991  */
992 pgoff_t page_cache_next_hole(struct address_space *mapping,
993                              pgoff_t index, unsigned long max_scan)
994 {
995         unsigned long i;
996
997         for (i = 0; i < max_scan; i++) {
998                 struct page *page;
999
1000                 page = radix_tree_lookup(&mapping->page_tree, index);
1001                 if (!page || radix_tree_exceptional_entry(page))
1002                         break;
1003                 index++;
1004                 if (index == 0)
1005                         break;
1006         }
1007
1008         return index;
1009 }
1010 EXPORT_SYMBOL(page_cache_next_hole);
1011
1012 /**
1013  * page_cache_prev_hole - find the prev hole (not-present entry)
1014  * @mapping: mapping
1015  * @index: index
1016  * @max_scan: maximum range to search
1017  *
1018  * Search backwards in the range [max(index-max_scan+1, 0), index] for
1019  * the first hole.
1020  *
1021  * Returns: the index of the hole if found, otherwise returns an index
1022  * outside of the set specified (in which case 'index - return >=
1023  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
1024  * will be returned.
1025  *
1026  * page_cache_prev_hole may be called under rcu_read_lock. However,
1027  * like radix_tree_gang_lookup, this will not atomically search a
1028  * snapshot of the tree at a single point in time. For example, if a
1029  * hole is created at index 10, then subsequently a hole is created at
1030  * index 5, page_cache_prev_hole covering both indexes may return 5 if
1031  * called under rcu_read_lock.
1032  */
1033 pgoff_t page_cache_prev_hole(struct address_space *mapping,
1034                              pgoff_t index, unsigned long max_scan)
1035 {
1036         unsigned long i;
1037
1038         for (i = 0; i < max_scan; i++) {
1039                 struct page *page;
1040
1041                 page = radix_tree_lookup(&mapping->page_tree, index);
1042                 if (!page || radix_tree_exceptional_entry(page))
1043                         break;
1044                 index--;
1045                 if (index == ULONG_MAX)
1046                         break;
1047         }
1048
1049         return index;
1050 }
1051 EXPORT_SYMBOL(page_cache_prev_hole);
1052
1053 /**
1054  * find_get_entry - find and get a page cache entry
1055  * @mapping: the address_space to search
1056  * @offset: the page cache index
1057  *
1058  * Looks up the page cache slot at @mapping & @offset.  If there is a
1059  * page cache page, it is returned with an increased refcount.
1060  *
1061  * If the slot holds a shadow entry of a previously evicted page, or a
1062  * swap entry from shmem/tmpfs, it is returned.
1063  *
1064  * Otherwise, %NULL is returned.
1065  */
1066 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
1067 {
1068         void **pagep;
1069         struct page *page;
1070
1071         rcu_read_lock();
1072 repeat:
1073         page = NULL;
1074         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
1075         if (pagep) {
1076                 page = radix_tree_deref_slot(pagep);
1077                 if (unlikely(!page))
1078                         goto out;
1079                 if (radix_tree_exception(page)) {
1080                         if (radix_tree_deref_retry(page))
1081                                 goto repeat;
1082                         /*
1083                          * A shadow entry of a recently evicted page,
1084                          * or a swap entry from shmem/tmpfs.  Return
1085                          * it without attempting to raise page count.
1086                          */
1087                         goto out;
1088                 }
1089                 if (!page_cache_get_speculative(page))
1090                         goto repeat;
1091
1092                 /*
1093                  * Has the page moved?
1094                  * This is part of the lockless pagecache protocol. See
1095                  * include/linux/pagemap.h for details.
1096                  */
1097                 if (unlikely(page != *pagep)) {
1098                         put_page(page);
1099                         goto repeat;
1100                 }
1101         }
1102 out:
1103         rcu_read_unlock();
1104
1105         return page;
1106 }
1107 EXPORT_SYMBOL(find_get_entry);
1108
1109 /**
1110  * find_lock_entry - locate, pin and lock a page cache entry
1111  * @mapping: the address_space to search
1112  * @offset: the page cache index
1113  *
1114  * Looks up the page cache slot at @mapping & @offset.  If there is a
1115  * page cache page, it is returned locked and with an increased
1116  * refcount.
1117  *
1118  * If the slot holds a shadow entry of a previously evicted page, or a
1119  * swap entry from shmem/tmpfs, it is returned.
1120  *
1121  * Otherwise, %NULL is returned.
1122  *
1123  * find_lock_entry() may sleep.
1124  */
1125 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
1126 {
1127         struct page *page;
1128
1129 repeat:
1130         page = find_get_entry(mapping, offset);
1131         if (page && !radix_tree_exception(page)) {
1132                 lock_page(page);
1133                 /* Has the page been truncated? */
1134                 if (unlikely(page->mapping != mapping)) {
1135                         unlock_page(page);
1136                         put_page(page);
1137                         goto repeat;
1138                 }
1139                 VM_BUG_ON_PAGE(page->index != offset, page);
1140         }
1141         return page;
1142 }
1143 EXPORT_SYMBOL(find_lock_entry);
1144
1145 /**
1146  * pagecache_get_page - find and get a page reference
1147  * @mapping: the address_space to search
1148  * @offset: the page index
1149  * @fgp_flags: PCG flags
1150  * @gfp_mask: gfp mask to use for the page cache data page allocation
1151  *
1152  * Looks up the page cache slot at @mapping & @offset.
1153  *
1154  * PCG flags modify how the page is returned.
1155  *
1156  * FGP_ACCESSED: the page will be marked accessed
1157  * FGP_LOCK: Page is return locked
1158  * FGP_CREAT: If page is not present then a new page is allocated using
1159  *              @gfp_mask and added to the page cache and the VM's LRU
1160  *              list. The page is returned locked and with an increased
1161  *              refcount. Otherwise, %NULL is returned.
1162  *
1163  * If FGP_LOCK or FGP_CREAT are specified then the function may sleep even
1164  * if the GFP flags specified for FGP_CREAT are atomic.
1165  *
1166  * If there is a page cache page, it is returned with an increased refcount.
1167  */
1168 struct page *pagecache_get_page(struct address_space *mapping, pgoff_t offset,
1169         int fgp_flags, gfp_t gfp_mask)
1170 {
1171         struct page *page;
1172
1173 repeat:
1174         page = find_get_entry(mapping, offset);
1175         if (radix_tree_exceptional_entry(page))
1176                 page = NULL;
1177         if (!page)
1178                 goto no_page;
1179
1180         if (fgp_flags & FGP_LOCK) {
1181                 if (fgp_flags & FGP_NOWAIT) {
1182                         if (!trylock_page(page)) {
1183                                 put_page(page);
1184                                 return NULL;
1185                         }
1186                 } else {
1187                         lock_page(page);
1188                 }
1189
1190                 /* Has the page been truncated? */
1191                 if (unlikely(page->mapping != mapping)) {
1192                         unlock_page(page);
1193                         put_page(page);
1194                         goto repeat;
1195                 }
1196                 VM_BUG_ON_PAGE(page->index != offset, page);
1197         }
1198
1199         if (page && (fgp_flags & FGP_ACCESSED))
1200                 mark_page_accessed(page);
1201
1202 no_page:
1203         if (!page && (fgp_flags & FGP_CREAT)) {
1204                 int err;
1205                 if ((fgp_flags & FGP_WRITE) && mapping_cap_account_dirty(mapping))
1206                         gfp_mask |= __GFP_WRITE;
1207                 if (fgp_flags & FGP_NOFS)
1208                         gfp_mask &= ~__GFP_FS;
1209
1210                 page = __page_cache_alloc(gfp_mask);
1211                 if (!page)
1212                         return NULL;
1213
1214                 if (WARN_ON_ONCE(!(fgp_flags & FGP_LOCK)))
1215                         fgp_flags |= FGP_LOCK;
1216
1217                 /* Init accessed so avoid atomic mark_page_accessed later */
1218                 if (fgp_flags & FGP_ACCESSED)
1219                         __SetPageReferenced(page);
1220
1221                 err = add_to_page_cache_lru(page, mapping, offset,
1222                                 gfp_mask & GFP_RECLAIM_MASK);
1223                 if (unlikely(err)) {
1224                         put_page(page);
1225                         page = NULL;
1226                         if (err == -EEXIST)
1227                                 goto repeat;
1228                 }
1229         }
1230
1231         return page;
1232 }
1233 EXPORT_SYMBOL(pagecache_get_page);
1234
1235 /**
1236  * find_get_entries - gang pagecache lookup
1237  * @mapping:    The address_space to search
1238  * @start:      The starting page cache index
1239  * @nr_entries: The maximum number of entries
1240  * @entries:    Where the resulting entries are placed
1241  * @indices:    The cache indices corresponding to the entries in @entries
1242  *
1243  * find_get_entries() will search for and return a group of up to
1244  * @nr_entries entries in the mapping.  The entries are placed at
1245  * @entries.  find_get_entries() takes a reference against any actual
1246  * pages it returns.
1247  *
1248  * The search returns a group of mapping-contiguous page cache entries
1249  * with ascending indexes.  There may be holes in the indices due to
1250  * not-present pages.
1251  *
1252  * Any shadow entries of evicted pages, or swap entries from
1253  * shmem/tmpfs, are included in the returned array.
1254  *
1255  * find_get_entries() returns the number of pages and shadow entries
1256  * which were found.
1257  */
1258 unsigned find_get_entries(struct address_space *mapping,
1259                           pgoff_t start, unsigned int nr_entries,
1260                           struct page **entries, pgoff_t *indices)
1261 {
1262         void **slot;
1263         unsigned int ret = 0;
1264         struct radix_tree_iter iter;
1265
1266         if (!nr_entries)
1267                 return 0;
1268
1269         rcu_read_lock();
1270         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1271                 struct page *page;
1272 repeat:
1273                 page = radix_tree_deref_slot(slot);
1274                 if (unlikely(!page))
1275                         continue;
1276                 if (radix_tree_exception(page)) {
1277                         if (radix_tree_deref_retry(page)) {
1278                                 slot = radix_tree_iter_retry(&iter);
1279                                 continue;
1280                         }
1281                         /*
1282                          * A shadow entry of a recently evicted page, a swap
1283                          * entry from shmem/tmpfs or a DAX entry.  Return it
1284                          * without attempting to raise page count.
1285                          */
1286                         goto export;
1287                 }
1288                 if (!page_cache_get_speculative(page))
1289                         goto repeat;
1290
1291                 /* Has the page moved? */
1292                 if (unlikely(page != *slot)) {
1293                         put_page(page);
1294                         goto repeat;
1295                 }
1296 export:
1297                 indices[ret] = iter.index;
1298                 entries[ret] = page;
1299                 if (++ret == nr_entries)
1300                         break;
1301         }
1302         rcu_read_unlock();
1303         return ret;
1304 }
1305
1306 /**
1307  * find_get_pages - gang pagecache lookup
1308  * @mapping:    The address_space to search
1309  * @start:      The starting page index
1310  * @nr_pages:   The maximum number of pages
1311  * @pages:      Where the resulting pages are placed
1312  *
1313  * find_get_pages() will search for and return a group of up to
1314  * @nr_pages pages in the mapping.  The pages are placed at @pages.
1315  * find_get_pages() takes a reference against the returned pages.
1316  *
1317  * The search returns a group of mapping-contiguous pages with ascending
1318  * indexes.  There may be holes in the indices due to not-present pages.
1319  *
1320  * find_get_pages() returns the number of pages which were found.
1321  */
1322 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
1323                             unsigned int nr_pages, struct page **pages)
1324 {
1325         struct radix_tree_iter iter;
1326         void **slot;
1327         unsigned ret = 0;
1328
1329         if (unlikely(!nr_pages))
1330                 return 0;
1331
1332         rcu_read_lock();
1333         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1334                 struct page *page;
1335 repeat:
1336                 page = radix_tree_deref_slot(slot);
1337                 if (unlikely(!page))
1338                         continue;
1339
1340                 if (radix_tree_exception(page)) {
1341                         if (radix_tree_deref_retry(page)) {
1342                                 slot = radix_tree_iter_retry(&iter);
1343                                 continue;
1344                         }
1345                         /*
1346                          * A shadow entry of a recently evicted page,
1347                          * or a swap entry from shmem/tmpfs.  Skip
1348                          * over it.
1349                          */
1350                         continue;
1351                 }
1352
1353                 if (!page_cache_get_speculative(page))
1354                         goto repeat;
1355
1356                 /* Has the page moved? */
1357                 if (unlikely(page != *slot)) {
1358                         put_page(page);
1359                         goto repeat;
1360                 }
1361
1362                 pages[ret] = page;
1363                 if (++ret == nr_pages)
1364                         break;
1365         }
1366
1367         rcu_read_unlock();
1368         return ret;
1369 }
1370
1371 /**
1372  * find_get_pages_contig - gang contiguous pagecache lookup
1373  * @mapping:    The address_space to search
1374  * @index:      The starting page index
1375  * @nr_pages:   The maximum number of pages
1376  * @pages:      Where the resulting pages are placed
1377  *
1378  * find_get_pages_contig() works exactly like find_get_pages(), except
1379  * that the returned number of pages are guaranteed to be contiguous.
1380  *
1381  * find_get_pages_contig() returns the number of pages which were found.
1382  */
1383 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1384                                unsigned int nr_pages, struct page **pages)
1385 {
1386         struct radix_tree_iter iter;
1387         void **slot;
1388         unsigned int ret = 0;
1389
1390         if (unlikely(!nr_pages))
1391                 return 0;
1392
1393         rcu_read_lock();
1394         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1395                 struct page *page;
1396 repeat:
1397                 page = radix_tree_deref_slot(slot);
1398                 /* The hole, there no reason to continue */
1399                 if (unlikely(!page))
1400                         break;
1401
1402                 if (radix_tree_exception(page)) {
1403                         if (radix_tree_deref_retry(page)) {
1404                                 slot = radix_tree_iter_retry(&iter);
1405                                 continue;
1406                         }
1407                         /*
1408                          * A shadow entry of a recently evicted page,
1409                          * or a swap entry from shmem/tmpfs.  Stop
1410                          * looking for contiguous pages.
1411                          */
1412                         break;
1413                 }
1414
1415                 if (!page_cache_get_speculative(page))
1416                         goto repeat;
1417
1418                 /* Has the page moved? */
1419                 if (unlikely(page != *slot)) {
1420                         put_page(page);
1421                         goto repeat;
1422                 }
1423
1424                 /*
1425                  * must check mapping and index after taking the ref.
1426                  * otherwise we can get both false positives and false
1427                  * negatives, which is just confusing to the caller.
1428                  */
1429                 if (page->mapping == NULL || page->index != iter.index) {
1430                         put_page(page);
1431                         break;
1432                 }
1433
1434                 pages[ret] = page;
1435                 if (++ret == nr_pages)
1436                         break;
1437         }
1438         rcu_read_unlock();
1439         return ret;
1440 }
1441 EXPORT_SYMBOL(find_get_pages_contig);
1442
1443 /**
1444  * find_get_pages_tag - find and return pages that match @tag
1445  * @mapping:    the address_space to search
1446  * @index:      the starting page index
1447  * @tag:        the tag index
1448  * @nr_pages:   the maximum number of pages
1449  * @pages:      where the resulting pages are placed
1450  *
1451  * Like find_get_pages, except we only return pages which are tagged with
1452  * @tag.   We update @index to index the next page for the traversal.
1453  */
1454 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1455                         int tag, unsigned int nr_pages, struct page **pages)
1456 {
1457         struct radix_tree_iter iter;
1458         void **slot;
1459         unsigned ret = 0;
1460
1461         if (unlikely(!nr_pages))
1462                 return 0;
1463
1464         rcu_read_lock();
1465         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1466                                    &iter, *index, tag) {
1467                 struct page *page;
1468 repeat:
1469                 page = radix_tree_deref_slot(slot);
1470                 if (unlikely(!page))
1471                         continue;
1472
1473                 if (radix_tree_exception(page)) {
1474                         if (radix_tree_deref_retry(page)) {
1475                                 slot = radix_tree_iter_retry(&iter);
1476                                 continue;
1477                         }
1478                         /*
1479                          * A shadow entry of a recently evicted page.
1480                          *
1481                          * Those entries should never be tagged, but
1482                          * this tree walk is lockless and the tags are
1483                          * looked up in bulk, one radix tree node at a
1484                          * time, so there is a sizable window for page
1485                          * reclaim to evict a page we saw tagged.
1486                          *
1487                          * Skip over it.
1488                          */
1489                         continue;
1490                 }
1491
1492                 if (!page_cache_get_speculative(page))
1493                         goto repeat;
1494
1495                 /* Has the page moved? */
1496                 if (unlikely(page != *slot)) {
1497                         put_page(page);
1498                         goto repeat;
1499                 }
1500
1501                 pages[ret] = page;
1502                 if (++ret == nr_pages)
1503                         break;
1504         }
1505
1506         rcu_read_unlock();
1507
1508         if (ret)
1509                 *index = pages[ret - 1]->index + 1;
1510
1511         return ret;
1512 }
1513 EXPORT_SYMBOL(find_get_pages_tag);
1514
1515 /**
1516  * find_get_entries_tag - find and return entries that match @tag
1517  * @mapping:    the address_space to search
1518  * @start:      the starting page cache index
1519  * @tag:        the tag index
1520  * @nr_entries: the maximum number of entries
1521  * @entries:    where the resulting entries are placed
1522  * @indices:    the cache indices corresponding to the entries in @entries
1523  *
1524  * Like find_get_entries, except we only return entries which are tagged with
1525  * @tag.
1526  */
1527 unsigned find_get_entries_tag(struct address_space *mapping, pgoff_t start,
1528                         int tag, unsigned int nr_entries,
1529                         struct page **entries, pgoff_t *indices)
1530 {
1531         void **slot;
1532         unsigned int ret = 0;
1533         struct radix_tree_iter iter;
1534
1535         if (!nr_entries)
1536                 return 0;
1537
1538         rcu_read_lock();
1539         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1540                                    &iter, start, tag) {
1541                 struct page *page;
1542 repeat:
1543                 page = radix_tree_deref_slot(slot);
1544                 if (unlikely(!page))
1545                         continue;
1546                 if (radix_tree_exception(page)) {
1547                         if (radix_tree_deref_retry(page)) {
1548                                 slot = radix_tree_iter_retry(&iter);
1549                                 continue;
1550                         }
1551
1552                         /*
1553                          * A shadow entry of a recently evicted page, a swap
1554                          * entry from shmem/tmpfs or a DAX entry.  Return it
1555                          * without attempting to raise page count.
1556                          */
1557                         goto export;
1558                 }
1559                 if (!page_cache_get_speculative(page))
1560                         goto repeat;
1561
1562                 /* Has the page moved? */
1563                 if (unlikely(page != *slot)) {
1564                         put_page(page);
1565                         goto repeat;
1566                 }
1567 export:
1568                 indices[ret] = iter.index;
1569                 entries[ret] = page;
1570                 if (++ret == nr_entries)
1571                         break;
1572         }
1573         rcu_read_unlock();
1574         return ret;
1575 }
1576 EXPORT_SYMBOL(find_get_entries_tag);
1577
1578 /*
1579  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1580  * a _large_ part of the i/o request. Imagine the worst scenario:
1581  *
1582  *      ---R__________________________________________B__________
1583  *         ^ reading here                             ^ bad block(assume 4k)
1584  *
1585  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1586  * => failing the whole request => read(R) => read(R+1) =>
1587  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1588  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1589  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1590  *
1591  * It is going insane. Fix it by quickly scaling down the readahead size.
1592  */
1593 static void shrink_readahead_size_eio(struct file *filp,
1594                                         struct file_ra_state *ra)
1595 {
1596         ra->ra_pages /= 4;
1597 }
1598
1599 /**
1600  * do_generic_file_read - generic file read routine
1601  * @filp:       the file to read
1602  * @ppos:       current file position
1603  * @iter:       data destination
1604  * @written:    already copied
1605  *
1606  * This is a generic file read routine, and uses the
1607  * mapping->a_ops->readpage() function for the actual low-level stuff.
1608  *
1609  * This is really ugly. But the goto's actually try to clarify some
1610  * of the logic when it comes to error handling etc.
1611  */
1612 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1613                 struct iov_iter *iter, ssize_t written)
1614 {
1615         struct address_space *mapping = filp->f_mapping;
1616         struct inode *inode = mapping->host;
1617         struct file_ra_state *ra = &filp->f_ra;
1618         pgoff_t index;
1619         pgoff_t last_index;
1620         pgoff_t prev_index;
1621         unsigned long offset;      /* offset into pagecache page */
1622         unsigned int prev_offset;
1623         int error = 0;
1624
1625         index = *ppos >> PAGE_SHIFT;
1626         prev_index = ra->prev_pos >> PAGE_SHIFT;
1627         prev_offset = ra->prev_pos & (PAGE_SIZE-1);
1628         last_index = (*ppos + iter->count + PAGE_SIZE-1) >> PAGE_SHIFT;
1629         offset = *ppos & ~PAGE_MASK;
1630
1631         for (;;) {
1632                 struct page *page;
1633                 pgoff_t end_index;
1634                 loff_t isize;
1635                 unsigned long nr, ret;
1636
1637                 cond_resched();
1638 find_page:
1639                 page = find_get_page(mapping, index);
1640                 if (!page) {
1641                         page_cache_sync_readahead(mapping,
1642                                         ra, filp,
1643                                         index, last_index - index);
1644                         page = find_get_page(mapping, index);
1645                         if (unlikely(page == NULL))
1646                                 goto no_cached_page;
1647                 }
1648                 if (PageReadahead(page)) {
1649                         page_cache_async_readahead(mapping,
1650                                         ra, filp, page,
1651                                         index, last_index - index);
1652                 }
1653                 if (!PageUptodate(page)) {
1654                         /*
1655                          * See comment in do_read_cache_page on why
1656                          * wait_on_page_locked is used to avoid unnecessarily
1657                          * serialisations and why it's safe.
1658                          */
1659                         wait_on_page_locked_killable(page);
1660                         if (PageUptodate(page))
1661                                 goto page_ok;
1662
1663                         if (inode->i_blkbits == PAGE_SHIFT ||
1664                                         !mapping->a_ops->is_partially_uptodate)
1665                                 goto page_not_up_to_date;
1666                         if (!trylock_page(page))
1667                                 goto page_not_up_to_date;
1668                         /* Did it get truncated before we got the lock? */
1669                         if (!page->mapping)
1670                                 goto page_not_up_to_date_locked;
1671                         if (!mapping->a_ops->is_partially_uptodate(page,
1672                                                         offset, iter->count))
1673                                 goto page_not_up_to_date_locked;
1674                         unlock_page(page);
1675                 }
1676 page_ok:
1677                 /*
1678                  * i_size must be checked after we know the page is Uptodate.
1679                  *
1680                  * Checking i_size after the check allows us to calculate
1681                  * the correct value for "nr", which means the zero-filled
1682                  * part of the page is not copied back to userspace (unless
1683                  * another truncate extends the file - this is desired though).
1684                  */
1685
1686                 isize = i_size_read(inode);
1687                 end_index = (isize - 1) >> PAGE_SHIFT;
1688                 if (unlikely(!isize || index > end_index)) {
1689                         put_page(page);
1690                         goto out;
1691                 }
1692
1693                 /* nr is the maximum number of bytes to copy from this page */
1694                 nr = PAGE_SIZE;
1695                 if (index == end_index) {
1696                         nr = ((isize - 1) & ~PAGE_MASK) + 1;
1697                         if (nr <= offset) {
1698                                 put_page(page);
1699                                 goto out;
1700                         }
1701                 }
1702                 nr = nr - offset;
1703
1704                 /* If users can be writing to this page using arbitrary
1705                  * virtual addresses, take care about potential aliasing
1706                  * before reading the page on the kernel side.
1707                  */
1708                 if (mapping_writably_mapped(mapping))
1709                         flush_dcache_page(page);
1710
1711                 /*
1712                  * When a sequential read accesses a page several times,
1713                  * only mark it as accessed the first time.
1714                  */
1715                 if (prev_index != index || offset != prev_offset)
1716                         mark_page_accessed(page);
1717                 prev_index = index;
1718
1719                 /*
1720                  * Ok, we have the page, and it's up-to-date, so
1721                  * now we can copy it to user space...
1722                  */
1723
1724                 ret = copy_page_to_iter(page, offset, nr, iter);
1725                 offset += ret;
1726                 index += offset >> PAGE_SHIFT;
1727                 offset &= ~PAGE_MASK;
1728                 prev_offset = offset;
1729
1730                 put_page(page);
1731                 written += ret;
1732                 if (!iov_iter_count(iter))
1733                         goto out;
1734                 if (ret < nr) {
1735                         error = -EFAULT;
1736                         goto out;
1737                 }
1738                 continue;
1739
1740 page_not_up_to_date:
1741                 /* Get exclusive access to the page ... */
1742                 error = lock_page_killable(page);
1743                 if (unlikely(error))
1744                         goto readpage_error;
1745
1746 page_not_up_to_date_locked:
1747                 /* Did it get truncated before we got the lock? */
1748                 if (!page->mapping) {
1749                         unlock_page(page);
1750                         put_page(page);
1751                         continue;
1752                 }
1753
1754                 /* Did somebody else fill it already? */
1755                 if (PageUptodate(page)) {
1756                         unlock_page(page);
1757                         goto page_ok;
1758                 }
1759
1760 readpage:
1761                 /*
1762                  * A previous I/O error may have been due to temporary
1763                  * failures, eg. multipath errors.
1764                  * PG_error will be set again if readpage fails.
1765                  */
1766                 ClearPageError(page);
1767                 /* Start the actual read. The read will unlock the page. */
1768                 error = mapping->a_ops->readpage(filp, page);
1769
1770                 if (unlikely(error)) {
1771                         if (error == AOP_TRUNCATED_PAGE) {
1772                                 put_page(page);
1773                                 error = 0;
1774                                 goto find_page;
1775                         }
1776                         goto readpage_error;
1777                 }
1778
1779                 if (!PageUptodate(page)) {
1780                         error = lock_page_killable(page);
1781                         if (unlikely(error))
1782                                 goto readpage_error;
1783                         if (!PageUptodate(page)) {
1784                                 if (page->mapping == NULL) {
1785                                         /*
1786                                          * invalidate_mapping_pages got it
1787                                          */
1788                                         unlock_page(page);
1789                                         put_page(page);
1790                                         goto find_page;
1791                                 }
1792                                 unlock_page(page);
1793                                 shrink_readahead_size_eio(filp, ra);
1794                                 error = -EIO;
1795                                 goto readpage_error;
1796                         }
1797                         unlock_page(page);
1798                 }
1799
1800                 goto page_ok;
1801
1802 readpage_error:
1803                 /* UHHUH! A synchronous read error occurred. Report it */
1804                 put_page(page);
1805                 goto out;
1806
1807 no_cached_page:
1808                 /*
1809                  * Ok, it wasn't cached, so we need to create a new
1810                  * page..
1811                  */
1812                 page = page_cache_alloc_cold(mapping);
1813                 if (!page) {
1814                         error = -ENOMEM;
1815                         goto out;
1816                 }
1817                 error = add_to_page_cache_lru(page, mapping, index,
1818                                 mapping_gfp_constraint(mapping, GFP_KERNEL));
1819                 if (error) {
1820                         put_page(page);
1821                         if (error == -EEXIST) {
1822                                 error = 0;
1823                                 goto find_page;
1824                         }
1825                         goto out;
1826                 }
1827                 goto readpage;
1828         }
1829
1830 out:
1831         ra->prev_pos = prev_index;
1832         ra->prev_pos <<= PAGE_SHIFT;
1833         ra->prev_pos |= prev_offset;
1834
1835         *ppos = ((loff_t)index << PAGE_SHIFT) + offset;
1836         file_accessed(filp);
1837         return written ? written : error;
1838 }
1839
1840 /**
1841  * generic_file_read_iter - generic filesystem read routine
1842  * @iocb:       kernel I/O control block
1843  * @iter:       destination for the data read
1844  *
1845  * This is the "read_iter()" routine for all filesystems
1846  * that can use the page cache directly.
1847  */
1848 ssize_t
1849 generic_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
1850 {
1851         struct file *file = iocb->ki_filp;
1852         ssize_t retval = 0;
1853         loff_t *ppos = &iocb->ki_pos;
1854         loff_t pos = *ppos;
1855         size_t count = iov_iter_count(iter);
1856
1857         if (!count)
1858                 goto out; /* skip atime */
1859
1860         if (iocb->ki_flags & IOCB_DIRECT) {
1861                 struct address_space *mapping = file->f_mapping;
1862                 struct inode *inode = mapping->host;
1863                 loff_t size;
1864
1865                 size = i_size_read(inode);
1866                 retval = filemap_write_and_wait_range(mapping, pos,
1867                                         pos + count - 1);
1868                 if (!retval) {
1869                         struct iov_iter data = *iter;
1870                         retval = mapping->a_ops->direct_IO(iocb, &data, pos);
1871                 }
1872
1873                 if (retval > 0) {
1874                         *ppos = pos + retval;
1875                         iov_iter_advance(iter, retval);
1876                 }
1877
1878                 /*
1879                  * Btrfs can have a short DIO read if we encounter
1880                  * compressed extents, so if there was an error, or if
1881                  * we've already read everything we wanted to, or if
1882                  * there was a short read because we hit EOF, go ahead
1883                  * and return.  Otherwise fallthrough to buffered io for
1884                  * the rest of the read.  Buffered reads will not work for
1885                  * DAX files, so don't bother trying.
1886                  */
1887                 if (retval < 0 || !iov_iter_count(iter) || *ppos >= size ||
1888                     IS_DAX(inode)) {
1889                         file_accessed(file);
1890                         goto out;
1891                 }
1892         }
1893
1894         retval = do_generic_file_read(file, ppos, iter, retval);
1895 out:
1896         return retval;
1897 }
1898 EXPORT_SYMBOL(generic_file_read_iter);
1899
1900 #ifdef CONFIG_MMU
1901 /**
1902  * page_cache_read - adds requested page to the page cache if not already there
1903  * @file:       file to read
1904  * @offset:     page index
1905  * @gfp_mask:   memory allocation flags
1906  *
1907  * This adds the requested page to the page cache if it isn't already there,
1908  * and schedules an I/O to read in its contents from disk.
1909  */
1910 static int page_cache_read(struct file *file, pgoff_t offset, gfp_t gfp_mask)
1911 {
1912         struct address_space *mapping = file->f_mapping;
1913         struct page *page;
1914         int ret;
1915
1916         do {
1917                 page = __page_cache_alloc(gfp_mask|__GFP_COLD);
1918                 if (!page)
1919                         return -ENOMEM;
1920
1921                 ret = add_to_page_cache_lru(page, mapping, offset, gfp_mask & GFP_KERNEL);
1922                 if (ret == 0)
1923                         ret = mapping->a_ops->readpage(file, page);
1924                 else if (ret == -EEXIST)
1925                         ret = 0; /* losing race to add is OK */
1926
1927                 put_page(page);
1928
1929         } while (ret == AOP_TRUNCATED_PAGE);
1930
1931         return ret;
1932 }
1933
1934 #define MMAP_LOTSAMISS  (100)
1935
1936 /*
1937  * Synchronous readahead happens when we don't even find
1938  * a page in the page cache at all.
1939  */
1940 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1941                                    struct file_ra_state *ra,
1942                                    struct file *file,
1943                                    pgoff_t offset)
1944 {
1945         struct address_space *mapping = file->f_mapping;
1946
1947         /* If we don't want any read-ahead, don't bother */
1948         if (vma->vm_flags & VM_RAND_READ)
1949                 return;
1950         if (!ra->ra_pages)
1951                 return;
1952
1953         if (vma->vm_flags & VM_SEQ_READ) {
1954                 page_cache_sync_readahead(mapping, ra, file, offset,
1955                                           ra->ra_pages);
1956                 return;
1957         }
1958
1959         /* Avoid banging the cache line if not needed */
1960         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1961                 ra->mmap_miss++;
1962
1963         /*
1964          * Do we miss much more than hit in this file? If so,
1965          * stop bothering with read-ahead. It will only hurt.
1966          */
1967         if (ra->mmap_miss > MMAP_LOTSAMISS)
1968                 return;
1969
1970         /*
1971          * mmap read-around
1972          */
1973         ra->start = max_t(long, 0, offset - ra->ra_pages / 2);
1974         ra->size = ra->ra_pages;
1975         ra->async_size = ra->ra_pages / 4;
1976         ra_submit(ra, mapping, file);
1977 }
1978
1979 /*
1980  * Asynchronous readahead happens when we find the page and PG_readahead,
1981  * so we want to possibly extend the readahead further..
1982  */
1983 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1984                                     struct file_ra_state *ra,
1985                                     struct file *file,
1986                                     struct page *page,
1987                                     pgoff_t offset)
1988 {
1989         struct address_space *mapping = file->f_mapping;
1990
1991         /* If we don't want any read-ahead, don't bother */
1992         if (vma->vm_flags & VM_RAND_READ)
1993                 return;
1994         if (ra->mmap_miss > 0)
1995                 ra->mmap_miss--;
1996         if (PageReadahead(page))
1997                 page_cache_async_readahead(mapping, ra, file,
1998                                            page, offset, ra->ra_pages);
1999 }
2000
2001 /**
2002  * filemap_fault - read in file data for page fault handling
2003  * @vma:        vma in which the fault was taken
2004  * @vmf:        struct vm_fault containing details of the fault
2005  *
2006  * filemap_fault() is invoked via the vma operations vector for a
2007  * mapped memory region to read in file data during a page fault.
2008  *
2009  * The goto's are kind of ugly, but this streamlines the normal case of having
2010  * it in the page cache, and handles the special cases reasonably without
2011  * having a lot of duplicated code.
2012  *
2013  * vma->vm_mm->mmap_sem must be held on entry.
2014  *
2015  * If our return value has VM_FAULT_RETRY set, it's because
2016  * lock_page_or_retry() returned 0.
2017  * The mmap_sem has usually been released in this case.
2018  * See __lock_page_or_retry() for the exception.
2019  *
2020  * If our return value does not have VM_FAULT_RETRY set, the mmap_sem
2021  * has not been released.
2022  *
2023  * We never return with VM_FAULT_RETRY and a bit from VM_FAULT_ERROR set.
2024  */
2025 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
2026 {
2027         int error;
2028         struct file *file = vma->vm_file;
2029         struct address_space *mapping = file->f_mapping;
2030         struct file_ra_state *ra = &file->f_ra;
2031         struct inode *inode = mapping->host;
2032         pgoff_t offset = vmf->pgoff;
2033         struct page *page;
2034         loff_t size;
2035         int ret = 0;
2036
2037         size = round_up(i_size_read(inode), PAGE_SIZE);
2038         if (offset >= size >> PAGE_SHIFT)
2039                 return VM_FAULT_SIGBUS;
2040
2041         /*
2042          * Do we have something in the page cache already?
2043          */
2044         page = find_get_page(mapping, offset);
2045         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
2046                 /*
2047                  * We found the page, so try async readahead before
2048                  * waiting for the lock.
2049                  */
2050                 do_async_mmap_readahead(vma, ra, file, page, offset);
2051         } else if (!page) {
2052                 /* No page in the page cache at all */
2053                 do_sync_mmap_readahead(vma, ra, file, offset);
2054                 count_vm_event(PGMAJFAULT);
2055                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2056                 ret = VM_FAULT_MAJOR;
2057 retry_find:
2058                 page = find_get_page(mapping, offset);
2059                 if (!page)
2060                         goto no_cached_page;
2061         }
2062
2063         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
2064                 put_page(page);
2065                 return ret | VM_FAULT_RETRY;
2066         }
2067
2068         /* Did it get truncated? */
2069         if (unlikely(page->mapping != mapping)) {
2070                 unlock_page(page);
2071                 put_page(page);
2072                 goto retry_find;
2073         }
2074         VM_BUG_ON_PAGE(page->index != offset, page);
2075
2076         /*
2077          * We have a locked page in the page cache, now we need to check
2078          * that it's up-to-date. If not, it is going to be due to an error.
2079          */
2080         if (unlikely(!PageUptodate(page)))
2081                 goto page_not_uptodate;
2082
2083         /*
2084          * Found the page and have a reference on it.
2085          * We must recheck i_size under page lock.
2086          */
2087         size = round_up(i_size_read(inode), PAGE_SIZE);
2088         if (unlikely(offset >= size >> PAGE_SHIFT)) {
2089                 unlock_page(page);
2090                 put_page(page);
2091                 return VM_FAULT_SIGBUS;
2092         }
2093
2094         vmf->page = page;
2095         return ret | VM_FAULT_LOCKED;
2096
2097 no_cached_page:
2098         /*
2099          * We're only likely to ever get here if MADV_RANDOM is in
2100          * effect.
2101          */
2102         error = page_cache_read(file, offset, vmf->gfp_mask);
2103
2104         /*
2105          * The page we want has now been added to the page cache.
2106          * In the unlikely event that someone removed it in the
2107          * meantime, we'll just come back here and read it again.
2108          */
2109         if (error >= 0)
2110                 goto retry_find;
2111
2112         /*
2113          * An error return from page_cache_read can result if the
2114          * system is low on memory, or a problem occurs while trying
2115          * to schedule I/O.
2116          */
2117         if (error == -ENOMEM)
2118                 return VM_FAULT_OOM;
2119         return VM_FAULT_SIGBUS;
2120
2121 page_not_uptodate:
2122         /*
2123          * Umm, take care of errors if the page isn't up-to-date.
2124          * Try to re-read it _once_. We do this synchronously,
2125          * because there really aren't any performance issues here
2126          * and we need to check for errors.
2127          */
2128         ClearPageError(page);
2129         error = mapping->a_ops->readpage(file, page);
2130         if (!error) {
2131                 wait_on_page_locked(page);
2132                 if (!PageUptodate(page))
2133                         error = -EIO;
2134         }
2135         put_page(page);
2136
2137         if (!error || error == AOP_TRUNCATED_PAGE)
2138                 goto retry_find;
2139
2140         /* Things didn't work out. Return zero to tell the mm layer so. */
2141         shrink_readahead_size_eio(file, ra);
2142         return VM_FAULT_SIGBUS;
2143 }
2144 EXPORT_SYMBOL(filemap_fault);
2145
2146 void filemap_map_pages(struct vm_area_struct *vma, struct vm_fault *vmf)
2147 {
2148         struct radix_tree_iter iter;
2149         void **slot;
2150         struct file *file = vma->vm_file;
2151         struct address_space *mapping = file->f_mapping;
2152         loff_t size;
2153         struct page *page;
2154         unsigned long address = (unsigned long) vmf->virtual_address;
2155         unsigned long addr;
2156         pte_t *pte;
2157
2158         rcu_read_lock();
2159         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, vmf->pgoff) {
2160                 if (iter.index > vmf->max_pgoff)
2161                         break;
2162 repeat:
2163                 page = radix_tree_deref_slot(slot);
2164                 if (unlikely(!page))
2165                         goto next;
2166                 if (radix_tree_exception(page)) {
2167                         if (radix_tree_deref_retry(page)) {
2168                                 slot = radix_tree_iter_retry(&iter);
2169                                 continue;
2170                         }
2171                         goto next;
2172                 }
2173
2174                 if (!page_cache_get_speculative(page))
2175                         goto repeat;
2176
2177                 /* Has the page moved? */
2178                 if (unlikely(page != *slot)) {
2179                         put_page(page);
2180                         goto repeat;
2181                 }
2182
2183                 if (!PageUptodate(page) ||
2184                                 PageReadahead(page) ||
2185                                 PageHWPoison(page))
2186                         goto skip;
2187                 if (!trylock_page(page))
2188                         goto skip;
2189
2190                 if (page->mapping != mapping || !PageUptodate(page))
2191                         goto unlock;
2192
2193                 size = round_up(i_size_read(mapping->host), PAGE_SIZE);
2194                 if (page->index >= size >> PAGE_SHIFT)
2195                         goto unlock;
2196
2197                 pte = vmf->pte + page->index - vmf->pgoff;
2198                 if (!pte_none(*pte))
2199                         goto unlock;
2200
2201                 if (file->f_ra.mmap_miss > 0)
2202                         file->f_ra.mmap_miss--;
2203                 addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
2204                 do_set_pte(vma, addr, page, pte, false, false);
2205                 unlock_page(page);
2206                 goto next;
2207 unlock:
2208                 unlock_page(page);
2209 skip:
2210                 put_page(page);
2211 next:
2212                 if (iter.index == vmf->max_pgoff)
2213                         break;
2214         }
2215         rcu_read_unlock();
2216 }
2217 EXPORT_SYMBOL(filemap_map_pages);
2218
2219 int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2220 {
2221         struct page *page = vmf->page;
2222         struct inode *inode = file_inode(vma->vm_file);
2223         int ret = VM_FAULT_LOCKED;
2224
2225         sb_start_pagefault(inode->i_sb);
2226         file_update_time(vma->vm_file);
2227         lock_page(page);
2228         if (page->mapping != inode->i_mapping) {
2229                 unlock_page(page);
2230                 ret = VM_FAULT_NOPAGE;
2231                 goto out;
2232         }
2233         /*
2234          * We mark the page dirty already here so that when freeze is in
2235          * progress, we are guaranteed that writeback during freezing will
2236          * see the dirty page and writeprotect it again.
2237          */
2238         set_page_dirty(page);
2239         wait_for_stable_page(page);
2240 out:
2241         sb_end_pagefault(inode->i_sb);
2242         return ret;
2243 }
2244 EXPORT_SYMBOL(filemap_page_mkwrite);
2245
2246 const struct vm_operations_struct generic_file_vm_ops = {
2247         .fault          = filemap_fault,
2248         .map_pages      = filemap_map_pages,
2249         .page_mkwrite   = filemap_page_mkwrite,
2250 };
2251
2252 /* This is used for a general mmap of a disk file */
2253
2254 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2255 {
2256         struct address_space *mapping = file->f_mapping;
2257
2258         if (!mapping->a_ops->readpage)
2259                 return -ENOEXEC;
2260         file_accessed(file);
2261         vma->vm_ops = &generic_file_vm_ops;
2262         return 0;
2263 }
2264
2265 /*
2266  * This is for filesystems which do not implement ->writepage.
2267  */
2268 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2269 {
2270         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2271                 return -EINVAL;
2272         return generic_file_mmap(file, vma);
2273 }
2274 #else
2275 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2276 {
2277         return -ENOSYS;
2278 }
2279 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2280 {
2281         return -ENOSYS;
2282 }
2283 #endif /* CONFIG_MMU */
2284
2285 EXPORT_SYMBOL(generic_file_mmap);
2286 EXPORT_SYMBOL(generic_file_readonly_mmap);
2287
2288 static struct page *wait_on_page_read(struct page *page)
2289 {
2290         if (!IS_ERR(page)) {
2291                 wait_on_page_locked(page);
2292                 if (!PageUptodate(page)) {
2293                         put_page(page);
2294                         page = ERR_PTR(-EIO);
2295                 }
2296         }
2297         return page;
2298 }
2299
2300 static struct page *do_read_cache_page(struct address_space *mapping,
2301                                 pgoff_t index,
2302                                 int (*filler)(void *, struct page *),
2303                                 void *data,
2304                                 gfp_t gfp)
2305 {
2306         struct page *page;
2307         int err;
2308 repeat:
2309         page = find_get_page(mapping, index);
2310         if (!page) {
2311                 page = __page_cache_alloc(gfp | __GFP_COLD);
2312                 if (!page)
2313                         return ERR_PTR(-ENOMEM);
2314                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2315                 if (unlikely(err)) {
2316                         put_page(page);
2317                         if (err == -EEXIST)
2318                                 goto repeat;
2319                         /* Presumably ENOMEM for radix tree node */
2320                         return ERR_PTR(err);
2321                 }
2322
2323 filler:
2324                 err = filler(data, page);
2325                 if (err < 0) {
2326                         put_page(page);
2327                         return ERR_PTR(err);
2328                 }
2329
2330                 page = wait_on_page_read(page);
2331                 if (IS_ERR(page))
2332                         return page;
2333                 goto out;
2334         }
2335         if (PageUptodate(page))
2336                 goto out;
2337
2338         /*
2339          * Page is not up to date and may be locked due one of the following
2340          * case a: Page is being filled and the page lock is held
2341          * case b: Read/write error clearing the page uptodate status
2342          * case c: Truncation in progress (page locked)
2343          * case d: Reclaim in progress
2344          *
2345          * Case a, the page will be up to date when the page is unlocked.
2346          *    There is no need to serialise on the page lock here as the page
2347          *    is pinned so the lock gives no additional protection. Even if the
2348          *    the page is truncated, the data is still valid if PageUptodate as
2349          *    it's a race vs truncate race.
2350          * Case b, the page will not be up to date
2351          * Case c, the page may be truncated but in itself, the data may still
2352          *    be valid after IO completes as it's a read vs truncate race. The
2353          *    operation must restart if the page is not uptodate on unlock but
2354          *    otherwise serialising on page lock to stabilise the mapping gives
2355          *    no additional guarantees to the caller as the page lock is
2356          *    released before return.
2357          * Case d, similar to truncation. If reclaim holds the page lock, it
2358          *    will be a race with remove_mapping that determines if the mapping
2359          *    is valid on unlock but otherwise the data is valid and there is
2360          *    no need to serialise with page lock.
2361          *
2362          * As the page lock gives no additional guarantee, we optimistically
2363          * wait on the page to be unlocked and check if it's up to date and
2364          * use the page if it is. Otherwise, the page lock is required to
2365          * distinguish between the different cases. The motivation is that we
2366          * avoid spurious serialisations and wakeups when multiple processes
2367          * wait on the same page for IO to complete.
2368          */
2369         wait_on_page_locked(page);
2370         if (PageUptodate(page))
2371                 goto out;
2372
2373         /* Distinguish between all the cases under the safety of the lock */
2374         lock_page(page);
2375
2376         /* Case c or d, restart the operation */
2377         if (!page->mapping) {
2378                 unlock_page(page);
2379                 put_page(page);
2380                 goto repeat;
2381         }
2382
2383         /* Someone else locked and filled the page in a very small window */
2384         if (PageUptodate(page)) {
2385                 unlock_page(page);
2386                 goto out;
2387         }
2388         goto filler;
2389
2390 out:
2391         mark_page_accessed(page);
2392         return page;
2393 }
2394
2395 /**
2396  * read_cache_page - read into page cache, fill it if needed
2397  * @mapping:    the page's address_space
2398  * @index:      the page index
2399  * @filler:     function to perform the read
2400  * @data:       first arg to filler(data, page) function, often left as NULL
2401  *
2402  * Read into the page cache. If a page already exists, and PageUptodate() is
2403  * not set, try to fill the page and wait for it to become unlocked.
2404  *
2405  * If the page does not get brought uptodate, return -EIO.
2406  */
2407 struct page *read_cache_page(struct address_space *mapping,
2408                                 pgoff_t index,
2409                                 int (*filler)(void *, struct page *),
2410                                 void *data)
2411 {
2412         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2413 }
2414 EXPORT_SYMBOL(read_cache_page);
2415
2416 /**
2417  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2418  * @mapping:    the page's address_space
2419  * @index:      the page index
2420  * @gfp:        the page allocator flags to use if allocating
2421  *
2422  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2423  * any new page allocations done using the specified allocation flags.
2424  *
2425  * If the page does not get brought uptodate, return -EIO.
2426  */
2427 struct page *read_cache_page_gfp(struct address_space *mapping,
2428                                 pgoff_t index,
2429                                 gfp_t gfp)
2430 {
2431         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2432
2433         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2434 }
2435 EXPORT_SYMBOL(read_cache_page_gfp);
2436
2437 /*
2438  * Performs necessary checks before doing a write
2439  *
2440  * Can adjust writing position or amount of bytes to write.
2441  * Returns appropriate error code that caller should return or
2442  * zero in case that write should be allowed.
2443  */
2444 inline ssize_t generic_write_checks(struct kiocb *iocb, struct iov_iter *from)
2445 {
2446         struct file *file = iocb->ki_filp;
2447         struct inode *inode = file->f_mapping->host;
2448         unsigned long limit = rlimit(RLIMIT_FSIZE);
2449         loff_t pos;
2450
2451         if (!iov_iter_count(from))
2452                 return 0;
2453
2454         /* FIXME: this is for backwards compatibility with 2.4 */
2455         if (iocb->ki_flags & IOCB_APPEND)
2456                 iocb->ki_pos = i_size_read(inode);
2457
2458         pos = iocb->ki_pos;
2459
2460         if (limit != RLIM_INFINITY) {
2461                 if (iocb->ki_pos >= limit) {
2462                         send_sig(SIGXFSZ, current, 0);
2463                         return -EFBIG;
2464                 }
2465                 iov_iter_truncate(from, limit - (unsigned long)pos);
2466         }
2467
2468         /*
2469          * LFS rule
2470          */
2471         if (unlikely(pos + iov_iter_count(from) > MAX_NON_LFS &&
2472                                 !(file->f_flags & O_LARGEFILE))) {
2473                 if (pos >= MAX_NON_LFS)
2474                         return -EFBIG;
2475                 iov_iter_truncate(from, MAX_NON_LFS - (unsigned long)pos);
2476         }
2477
2478         /*
2479          * Are we about to exceed the fs block limit ?
2480          *
2481          * If we have written data it becomes a short write.  If we have
2482          * exceeded without writing data we send a signal and return EFBIG.
2483          * Linus frestrict idea will clean these up nicely..
2484          */
2485         if (unlikely(pos >= inode->i_sb->s_maxbytes))
2486                 return -EFBIG;
2487
2488         iov_iter_truncate(from, inode->i_sb->s_maxbytes - pos);
2489         return iov_iter_count(from);
2490 }
2491 EXPORT_SYMBOL(generic_write_checks);
2492
2493 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2494                                 loff_t pos, unsigned len, unsigned flags,
2495                                 struct page **pagep, void **fsdata)
2496 {
2497         const struct address_space_operations *aops = mapping->a_ops;
2498
2499         return aops->write_begin(file, mapping, pos, len, flags,
2500                                                         pagep, fsdata);
2501 }
2502 EXPORT_SYMBOL(pagecache_write_begin);
2503
2504 int pagecache_write_end(struct file *file, struct address_space *mapping,
2505                                 loff_t pos, unsigned len, unsigned copied,
2506                                 struct page *page, void *fsdata)
2507 {
2508         const struct address_space_operations *aops = mapping->a_ops;
2509
2510         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2511 }
2512 EXPORT_SYMBOL(pagecache_write_end);
2513
2514 ssize_t
2515 generic_file_direct_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos)
2516 {
2517         struct file     *file = iocb->ki_filp;
2518         struct address_space *mapping = file->f_mapping;
2519         struct inode    *inode = mapping->host;
2520         ssize_t         written;
2521         size_t          write_len;
2522         pgoff_t         end;
2523         struct iov_iter data;
2524
2525         write_len = iov_iter_count(from);
2526         end = (pos + write_len - 1) >> PAGE_SHIFT;
2527
2528         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2529         if (written)
2530                 goto out;
2531
2532         /*
2533          * After a write we want buffered reads to be sure to go to disk to get
2534          * the new data.  We invalidate clean cached page from the region we're
2535          * about to write.  We do this *before* the write so that we can return
2536          * without clobbering -EIOCBQUEUED from ->direct_IO().
2537          */
2538         if (mapping->nrpages) {
2539                 written = invalidate_inode_pages2_range(mapping,
2540                                         pos >> PAGE_SHIFT, end);
2541                 /*
2542                  * If a page can not be invalidated, return 0 to fall back
2543                  * to buffered write.
2544                  */
2545                 if (written) {
2546                         if (written == -EBUSY)
2547                                 return 0;
2548                         goto out;
2549                 }
2550         }
2551
2552         data = *from;
2553         written = mapping->a_ops->direct_IO(iocb, &data, pos);
2554
2555         /*
2556          * Finally, try again to invalidate clean pages which might have been
2557          * cached by non-direct readahead, or faulted in by get_user_pages()
2558          * if the source of the write was an mmap'ed region of the file
2559          * we're writing.  Either one is a pretty crazy thing to do,
2560          * so we don't support it 100%.  If this invalidation
2561          * fails, tough, the write still worked...
2562          */
2563         if (mapping->nrpages) {
2564                 invalidate_inode_pages2_range(mapping,
2565                                               pos >> PAGE_SHIFT, end);
2566         }
2567
2568         if (written > 0) {
2569                 pos += written;
2570                 iov_iter_advance(from, written);
2571                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2572                         i_size_write(inode, pos);
2573                         mark_inode_dirty(inode);
2574                 }
2575                 iocb->ki_pos = pos;
2576         }
2577 out:
2578         return written;
2579 }
2580 EXPORT_SYMBOL(generic_file_direct_write);
2581
2582 /*
2583  * Find or create a page at the given pagecache position. Return the locked
2584  * page. This function is specifically for buffered writes.
2585  */
2586 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2587                                         pgoff_t index, unsigned flags)
2588 {
2589         struct page *page;
2590         int fgp_flags = FGP_LOCK|FGP_ACCESSED|FGP_WRITE|FGP_CREAT;
2591
2592         if (flags & AOP_FLAG_NOFS)
2593                 fgp_flags |= FGP_NOFS;
2594
2595         page = pagecache_get_page(mapping, index, fgp_flags,
2596                         mapping_gfp_mask(mapping));
2597         if (page)
2598                 wait_for_stable_page(page);
2599
2600         return page;
2601 }
2602 EXPORT_SYMBOL(grab_cache_page_write_begin);
2603
2604 ssize_t generic_perform_write(struct file *file,
2605                                 struct iov_iter *i, loff_t pos)
2606 {
2607         struct address_space *mapping = file->f_mapping;
2608         const struct address_space_operations *a_ops = mapping->a_ops;
2609         long status = 0;
2610         ssize_t written = 0;
2611         unsigned int flags = 0;
2612
2613         /*
2614          * Copies from kernel address space cannot fail (NFSD is a big user).
2615          */
2616         if (!iter_is_iovec(i))
2617                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2618
2619         do {
2620                 struct page *page;
2621                 unsigned long offset;   /* Offset into pagecache page */
2622                 unsigned long bytes;    /* Bytes to write to page */
2623                 size_t copied;          /* Bytes copied from user */
2624                 void *fsdata;
2625
2626                 offset = (pos & (PAGE_SIZE - 1));
2627                 bytes = min_t(unsigned long, PAGE_SIZE - offset,
2628                                                 iov_iter_count(i));
2629
2630 again:
2631                 /*
2632                  * Bring in the user page that we will copy from _first_.
2633                  * Otherwise there's a nasty deadlock on copying from the
2634                  * same page as we're writing to, without it being marked
2635                  * up-to-date.
2636                  *
2637                  * Not only is this an optimisation, but it is also required
2638                  * to check that the address is actually valid, when atomic
2639                  * usercopies are used, below.
2640                  */
2641                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2642                         status = -EFAULT;
2643                         break;
2644                 }
2645
2646                 if (fatal_signal_pending(current)) {
2647                         status = -EINTR;
2648                         break;
2649                 }
2650
2651                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2652                                                 &page, &fsdata);
2653                 if (unlikely(status < 0))
2654                         break;
2655
2656                 if (mapping_writably_mapped(mapping))
2657                         flush_dcache_page(page);
2658
2659                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2660                 flush_dcache_page(page);
2661
2662                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2663                                                 page, fsdata);
2664                 if (unlikely(status < 0))
2665                         break;
2666                 copied = status;
2667
2668                 cond_resched();
2669
2670                 iov_iter_advance(i, copied);
2671                 if (unlikely(copied == 0)) {
2672                         /*
2673                          * If we were unable to copy any data at all, we must
2674                          * fall back to a single segment length write.
2675                          *
2676                          * If we didn't fallback here, we could livelock
2677                          * because not all segments in the iov can be copied at
2678                          * once without a pagefault.
2679                          */
2680                         bytes = min_t(unsigned long, PAGE_SIZE - offset,
2681                                                 iov_iter_single_seg_count(i));
2682                         goto again;
2683                 }
2684                 pos += copied;
2685                 written += copied;
2686
2687                 balance_dirty_pages_ratelimited(mapping);
2688         } while (iov_iter_count(i));
2689
2690         return written ? written : status;
2691 }
2692 EXPORT_SYMBOL(generic_perform_write);
2693
2694 /**
2695  * __generic_file_write_iter - write data to a file
2696  * @iocb:       IO state structure (file, offset, etc.)
2697  * @from:       iov_iter with data to write
2698  *
2699  * This function does all the work needed for actually writing data to a
2700  * file. It does all basic checks, removes SUID from the file, updates
2701  * modification times and calls proper subroutines depending on whether we
2702  * do direct IO or a standard buffered write.
2703  *
2704  * It expects i_mutex to be grabbed unless we work on a block device or similar
2705  * object which does not need locking at all.
2706  *
2707  * This function does *not* take care of syncing data in case of O_SYNC write.
2708  * A caller has to handle it. This is mainly due to the fact that we want to
2709  * avoid syncing under i_mutex.
2710  */
2711 ssize_t __generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2712 {
2713         struct file *file = iocb->ki_filp;
2714         struct address_space * mapping = file->f_mapping;
2715         struct inode    *inode = mapping->host;
2716         ssize_t         written = 0;
2717         ssize_t         err;
2718         ssize_t         status;
2719
2720         /* We can write back this queue in page reclaim */
2721         current->backing_dev_info = inode_to_bdi(inode);
2722         err = file_remove_privs(file);
2723         if (err)
2724                 goto out;
2725
2726         err = file_update_time(file);
2727         if (err)
2728                 goto out;
2729
2730         if (iocb->ki_flags & IOCB_DIRECT) {
2731                 loff_t pos, endbyte;
2732
2733                 written = generic_file_direct_write(iocb, from, iocb->ki_pos);
2734                 /*
2735                  * If the write stopped short of completing, fall back to
2736                  * buffered writes.  Some filesystems do this for writes to
2737                  * holes, for example.  For DAX files, a buffered write will
2738                  * not succeed (even if it did, DAX does not handle dirty
2739                  * page-cache pages correctly).
2740                  */
2741                 if (written < 0 || !iov_iter_count(from) || IS_DAX(inode))
2742                         goto out;
2743
2744                 status = generic_perform_write(file, from, pos = iocb->ki_pos);
2745                 /*
2746                  * If generic_perform_write() returned a synchronous error
2747                  * then we want to return the number of bytes which were
2748                  * direct-written, or the error code if that was zero.  Note
2749                  * that this differs from normal direct-io semantics, which
2750                  * will return -EFOO even if some bytes were written.
2751                  */
2752                 if (unlikely(status < 0)) {
2753                         err = status;
2754                         goto out;
2755                 }
2756                 /*
2757                  * We need to ensure that the page cache pages are written to
2758                  * disk and invalidated to preserve the expected O_DIRECT
2759                  * semantics.
2760                  */
2761                 endbyte = pos + status - 1;
2762                 err = filemap_write_and_wait_range(mapping, pos, endbyte);
2763                 if (err == 0) {
2764                         iocb->ki_pos = endbyte + 1;
2765                         written += status;
2766                         invalidate_mapping_pages(mapping,
2767                                                  pos >> PAGE_SHIFT,
2768                                                  endbyte >> PAGE_SHIFT);
2769                 } else {
2770                         /*
2771                          * We don't know how much we wrote, so just return
2772                          * the number of bytes which were direct-written
2773                          */
2774                 }
2775         } else {
2776                 written = generic_perform_write(file, from, iocb->ki_pos);
2777                 if (likely(written > 0))
2778                         iocb->ki_pos += written;
2779         }
2780 out:
2781         current->backing_dev_info = NULL;
2782         return written ? written : err;
2783 }
2784 EXPORT_SYMBOL(__generic_file_write_iter);
2785
2786 /**
2787  * generic_file_write_iter - write data to a file
2788  * @iocb:       IO state structure
2789  * @from:       iov_iter with data to write
2790  *
2791  * This is a wrapper around __generic_file_write_iter() to be used by most
2792  * filesystems. It takes care of syncing the file in case of O_SYNC file
2793  * and acquires i_mutex as needed.
2794  */
2795 ssize_t generic_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
2796 {
2797         struct file *file = iocb->ki_filp;
2798         struct inode *inode = file->f_mapping->host;
2799         ssize_t ret;
2800
2801         inode_lock(inode);
2802         ret = generic_write_checks(iocb, from);
2803         if (ret > 0)
2804                 ret = __generic_file_write_iter(iocb, from);
2805         inode_unlock(inode);
2806
2807         if (ret > 0) {
2808                 ssize_t err;
2809
2810                 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
2811                 if (err < 0)
2812                         ret = err;
2813         }
2814         return ret;
2815 }
2816 EXPORT_SYMBOL(generic_file_write_iter);
2817
2818 /**
2819  * try_to_release_page() - release old fs-specific metadata on a page
2820  *
2821  * @page: the page which the kernel is trying to free
2822  * @gfp_mask: memory allocation flags (and I/O mode)
2823  *
2824  * The address_space is to try to release any data against the page
2825  * (presumably at page->private).  If the release was successful, return `1'.
2826  * Otherwise return zero.
2827  *
2828  * This may also be called if PG_fscache is set on a page, indicating that the
2829  * page is known to the local caching routines.
2830  *
2831  * The @gfp_mask argument specifies whether I/O may be performed to release
2832  * this page (__GFP_IO), and whether the call may block (__GFP_RECLAIM & __GFP_FS).
2833  *
2834  */
2835 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2836 {
2837         struct address_space * const mapping = page->mapping;
2838
2839         BUG_ON(!PageLocked(page));
2840         if (PageWriteback(page))
2841                 return 0;
2842
2843         if (mapping && mapping->a_ops->releasepage)
2844                 return mapping->a_ops->releasepage(page, gfp_mask);
2845         return try_to_free_buffers(page);
2846 }
2847
2848 EXPORT_SYMBOL(try_to_release_page);