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