Merge branch 'pm-sleep'
[linux-2.6-block.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/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/cpuset.h>
33 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34 #include <linux/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_mutex              (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_mutex            (truncate->unmap_mapping_range)
71  *
72  *  ->mmap_sem
73  *    ->i_mmap_mutex
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_mutex
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  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
104  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
105  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
106  *
107  * ->i_mmap_mutex
108  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
109  */
110
111 static void page_cache_tree_delete(struct address_space *mapping,
112                                    struct page *page, void *shadow)
113 {
114         struct radix_tree_node *node;
115         unsigned long index;
116         unsigned int offset;
117         unsigned int tag;
118         void **slot;
119
120         VM_BUG_ON(!PageLocked(page));
121
122         __radix_tree_lookup(&mapping->page_tree, page->index, &node, &slot);
123
124         if (shadow) {
125                 mapping->nrshadows++;
126                 /*
127                  * Make sure the nrshadows update is committed before
128                  * the nrpages update so that final truncate racing
129                  * with reclaim does not see both counters 0 at the
130                  * same time and miss a shadow entry.
131                  */
132                 smp_wmb();
133         }
134         mapping->nrpages--;
135
136         if (!node) {
137                 /* Clear direct pointer tags in root node */
138                 mapping->page_tree.gfp_mask &= __GFP_BITS_MASK;
139                 radix_tree_replace_slot(slot, shadow);
140                 return;
141         }
142
143         /* Clear tree tags for the removed page */
144         index = page->index;
145         offset = index & RADIX_TREE_MAP_MASK;
146         for (tag = 0; tag < RADIX_TREE_MAX_TAGS; tag++) {
147                 if (test_bit(offset, node->tags[tag]))
148                         radix_tree_tag_clear(&mapping->page_tree, index, tag);
149         }
150
151         /* Delete page, swap shadow entry */
152         radix_tree_replace_slot(slot, shadow);
153         workingset_node_pages_dec(node);
154         if (shadow)
155                 workingset_node_shadows_inc(node);
156         else
157                 if (__radix_tree_delete_node(&mapping->page_tree, node))
158                         return;
159
160         /*
161          * Track node that only contains shadow entries.
162          *
163          * Avoid acquiring the list_lru lock if already tracked.  The
164          * list_empty() test is safe as node->private_list is
165          * protected by mapping->tree_lock.
166          */
167         if (!workingset_node_pages(node) &&
168             list_empty(&node->private_list)) {
169                 node->private_data = mapping;
170                 list_lru_add(&workingset_shadow_nodes, &node->private_list);
171         }
172 }
173
174 /*
175  * Delete a page from the page cache and free it. Caller has to make
176  * sure the page is locked and that nobody else uses it - or that usage
177  * is safe.  The caller must hold the mapping's tree_lock.
178  */
179 void __delete_from_page_cache(struct page *page, void *shadow)
180 {
181         struct address_space *mapping = page->mapping;
182
183         trace_mm_filemap_delete_from_page_cache(page);
184         /*
185          * if we're uptodate, flush out into the cleancache, otherwise
186          * invalidate any existing cleancache entries.  We can't leave
187          * stale data around in the cleancache once our page is gone
188          */
189         if (PageUptodate(page) && PageMappedToDisk(page))
190                 cleancache_put_page(page);
191         else
192                 cleancache_invalidate_page(mapping, page);
193
194         page_cache_tree_delete(mapping, page, shadow);
195
196         page->mapping = NULL;
197         /* Leave page->index set: truncation lookup relies upon it */
198
199         __dec_zone_page_state(page, NR_FILE_PAGES);
200         if (PageSwapBacked(page))
201                 __dec_zone_page_state(page, NR_SHMEM);
202         BUG_ON(page_mapped(page));
203
204         /*
205          * Some filesystems seem to re-dirty the page even after
206          * the VM has canceled the dirty bit (eg ext3 journaling).
207          *
208          * Fix it up by doing a final dirty accounting check after
209          * having removed the page entirely.
210          */
211         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
212                 dec_zone_page_state(page, NR_FILE_DIRTY);
213                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
214         }
215 }
216
217 /**
218  * delete_from_page_cache - delete page from page cache
219  * @page: the page which the kernel is trying to remove from page cache
220  *
221  * This must be called only on pages that have been verified to be in the page
222  * cache and locked.  It will never put the page into the free list, the caller
223  * has a reference on the page.
224  */
225 void delete_from_page_cache(struct page *page)
226 {
227         struct address_space *mapping = page->mapping;
228         void (*freepage)(struct page *);
229
230         BUG_ON(!PageLocked(page));
231
232         freepage = mapping->a_ops->freepage;
233         spin_lock_irq(&mapping->tree_lock);
234         __delete_from_page_cache(page, NULL);
235         spin_unlock_irq(&mapping->tree_lock);
236         mem_cgroup_uncharge_cache_page(page);
237
238         if (freepage)
239                 freepage(page);
240         page_cache_release(page);
241 }
242 EXPORT_SYMBOL(delete_from_page_cache);
243
244 static int sleep_on_page(void *word)
245 {
246         io_schedule();
247         return 0;
248 }
249
250 static int sleep_on_page_killable(void *word)
251 {
252         sleep_on_page(word);
253         return fatal_signal_pending(current) ? -EINTR : 0;
254 }
255
256 static int filemap_check_errors(struct address_space *mapping)
257 {
258         int ret = 0;
259         /* Check for outstanding write errors */
260         if (test_bit(AS_ENOSPC, &mapping->flags) &&
261             test_and_clear_bit(AS_ENOSPC, &mapping->flags))
262                 ret = -ENOSPC;
263         if (test_bit(AS_EIO, &mapping->flags) &&
264             test_and_clear_bit(AS_EIO, &mapping->flags))
265                 ret = -EIO;
266         return ret;
267 }
268
269 /**
270  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
271  * @mapping:    address space structure to write
272  * @start:      offset in bytes where the range starts
273  * @end:        offset in bytes where the range ends (inclusive)
274  * @sync_mode:  enable synchronous operation
275  *
276  * Start writeback against all of a mapping's dirty pages that lie
277  * within the byte offsets <start, end> inclusive.
278  *
279  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
280  * opposed to a regular memory cleansing writeback.  The difference between
281  * these two operations is that if a dirty page/buffer is encountered, it must
282  * be waited upon, and not just skipped over.
283  */
284 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
285                                 loff_t end, int sync_mode)
286 {
287         int ret;
288         struct writeback_control wbc = {
289                 .sync_mode = sync_mode,
290                 .nr_to_write = LONG_MAX,
291                 .range_start = start,
292                 .range_end = end,
293         };
294
295         if (!mapping_cap_writeback_dirty(mapping))
296                 return 0;
297
298         ret = do_writepages(mapping, &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
488                 pgoff_t offset = old->index;
489                 freepage = mapping->a_ops->freepage;
490
491                 page_cache_get(new);
492                 new->mapping = mapping;
493                 new->index = offset;
494
495                 spin_lock_irq(&mapping->tree_lock);
496                 __delete_from_page_cache(old, NULL);
497                 error = radix_tree_insert(&mapping->page_tree, offset, new);
498                 BUG_ON(error);
499                 mapping->nrpages++;
500                 __inc_zone_page_state(new, NR_FILE_PAGES);
501                 if (PageSwapBacked(new))
502                         __inc_zone_page_state(new, NR_SHMEM);
503                 spin_unlock_irq(&mapping->tree_lock);
504                 /* mem_cgroup codes must not be called under tree_lock */
505                 mem_cgroup_replace_page_cache(old, new);
506                 radix_tree_preload_end();
507                 if (freepage)
508                         freepage(old);
509                 page_cache_release(old);
510         }
511
512         return error;
513 }
514 EXPORT_SYMBOL_GPL(replace_page_cache_page);
515
516 static int page_cache_tree_insert(struct address_space *mapping,
517                                   struct page *page, void **shadowp)
518 {
519         struct radix_tree_node *node;
520         void **slot;
521         int error;
522
523         error = __radix_tree_create(&mapping->page_tree, page->index,
524                                     &node, &slot);
525         if (error)
526                 return error;
527         if (*slot) {
528                 void *p;
529
530                 p = radix_tree_deref_slot_protected(slot, &mapping->tree_lock);
531                 if (!radix_tree_exceptional_entry(p))
532                         return -EEXIST;
533                 if (shadowp)
534                         *shadowp = p;
535                 mapping->nrshadows--;
536                 if (node)
537                         workingset_node_shadows_dec(node);
538         }
539         radix_tree_replace_slot(slot, page);
540         mapping->nrpages++;
541         if (node) {
542                 workingset_node_pages_inc(node);
543                 /*
544                  * Don't track node that contains actual pages.
545                  *
546                  * Avoid acquiring the list_lru lock if already
547                  * untracked.  The list_empty() test is safe as
548                  * node->private_list is protected by
549                  * mapping->tree_lock.
550                  */
551                 if (!list_empty(&node->private_list))
552                         list_lru_del(&workingset_shadow_nodes,
553                                      &node->private_list);
554         }
555         return 0;
556 }
557
558 static int __add_to_page_cache_locked(struct page *page,
559                                       struct address_space *mapping,
560                                       pgoff_t offset, gfp_t gfp_mask,
561                                       void **shadowp)
562 {
563         int error;
564
565         VM_BUG_ON_PAGE(!PageLocked(page), page);
566         VM_BUG_ON_PAGE(PageSwapBacked(page), page);
567
568         error = mem_cgroup_charge_file(page, current->mm,
569                                         gfp_mask & GFP_RECLAIM_MASK);
570         if (error)
571                 return error;
572
573         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
574         if (error) {
575                 mem_cgroup_uncharge_cache_page(page);
576                 return error;
577         }
578
579         page_cache_get(page);
580         page->mapping = mapping;
581         page->index = offset;
582
583         spin_lock_irq(&mapping->tree_lock);
584         error = page_cache_tree_insert(mapping, page, shadowp);
585         radix_tree_preload_end();
586         if (unlikely(error))
587                 goto err_insert;
588         __inc_zone_page_state(page, NR_FILE_PAGES);
589         spin_unlock_irq(&mapping->tree_lock);
590         trace_mm_filemap_add_to_page_cache(page);
591         return 0;
592 err_insert:
593         page->mapping = NULL;
594         /* Leave page->index set: truncation relies upon it */
595         spin_unlock_irq(&mapping->tree_lock);
596         mem_cgroup_uncharge_cache_page(page);
597         page_cache_release(page);
598         return error;
599 }
600
601 /**
602  * add_to_page_cache_locked - add a locked page to the pagecache
603  * @page:       page to add
604  * @mapping:    the page's address_space
605  * @offset:     page index
606  * @gfp_mask:   page allocation mode
607  *
608  * This function is used to add a page to the pagecache. It must be locked.
609  * This function does not add the page to the LRU.  The caller must do that.
610  */
611 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
612                 pgoff_t offset, gfp_t gfp_mask)
613 {
614         return __add_to_page_cache_locked(page, mapping, offset,
615                                           gfp_mask, NULL);
616 }
617 EXPORT_SYMBOL(add_to_page_cache_locked);
618
619 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
620                                 pgoff_t offset, gfp_t gfp_mask)
621 {
622         void *shadow = NULL;
623         int ret;
624
625         __set_page_locked(page);
626         ret = __add_to_page_cache_locked(page, mapping, offset,
627                                          gfp_mask, &shadow);
628         if (unlikely(ret))
629                 __clear_page_locked(page);
630         else {
631                 /*
632                  * The page might have been evicted from cache only
633                  * recently, in which case it should be activated like
634                  * any other repeatedly accessed page.
635                  */
636                 if (shadow && workingset_refault(shadow)) {
637                         SetPageActive(page);
638                         workingset_activation(page);
639                 } else
640                         ClearPageActive(page);
641                 lru_cache_add(page);
642         }
643         return ret;
644 }
645 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
646
647 #ifdef CONFIG_NUMA
648 struct page *__page_cache_alloc(gfp_t gfp)
649 {
650         int n;
651         struct page *page;
652
653         if (cpuset_do_page_mem_spread()) {
654                 unsigned int cpuset_mems_cookie;
655                 do {
656                         cpuset_mems_cookie = read_mems_allowed_begin();
657                         n = cpuset_mem_spread_node();
658                         page = alloc_pages_exact_node(n, gfp, 0);
659                 } while (!page && read_mems_allowed_retry(cpuset_mems_cookie));
660
661                 return page;
662         }
663         return alloc_pages(gfp, 0);
664 }
665 EXPORT_SYMBOL(__page_cache_alloc);
666 #endif
667
668 /*
669  * In order to wait for pages to become available there must be
670  * waitqueues associated with pages. By using a hash table of
671  * waitqueues where the bucket discipline is to maintain all
672  * waiters on the same queue and wake all when any of the pages
673  * become available, and for the woken contexts to check to be
674  * sure the appropriate page became available, this saves space
675  * at a cost of "thundering herd" phenomena during rare hash
676  * collisions.
677  */
678 static wait_queue_head_t *page_waitqueue(struct page *page)
679 {
680         const struct zone *zone = page_zone(page);
681
682         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
683 }
684
685 static inline void wake_up_page(struct page *page, int bit)
686 {
687         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
688 }
689
690 void wait_on_page_bit(struct page *page, int bit_nr)
691 {
692         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
693
694         if (test_bit(bit_nr, &page->flags))
695                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
696                                                         TASK_UNINTERRUPTIBLE);
697 }
698 EXPORT_SYMBOL(wait_on_page_bit);
699
700 int wait_on_page_bit_killable(struct page *page, int bit_nr)
701 {
702         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
703
704         if (!test_bit(bit_nr, &page->flags))
705                 return 0;
706
707         return __wait_on_bit(page_waitqueue(page), &wait,
708                              sleep_on_page_killable, TASK_KILLABLE);
709 }
710
711 /**
712  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
713  * @page: Page defining the wait queue of interest
714  * @waiter: Waiter to add to the queue
715  *
716  * Add an arbitrary @waiter to the wait queue for the nominated @page.
717  */
718 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
719 {
720         wait_queue_head_t *q = page_waitqueue(page);
721         unsigned long flags;
722
723         spin_lock_irqsave(&q->lock, flags);
724         __add_wait_queue(q, waiter);
725         spin_unlock_irqrestore(&q->lock, flags);
726 }
727 EXPORT_SYMBOL_GPL(add_page_wait_queue);
728
729 /**
730  * unlock_page - unlock a locked page
731  * @page: the page
732  *
733  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
734  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
735  * mechananism between PageLocked pages and PageWriteback pages is shared.
736  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
737  *
738  * The mb is necessary to enforce ordering between the clear_bit and the read
739  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
740  */
741 void unlock_page(struct page *page)
742 {
743         VM_BUG_ON_PAGE(!PageLocked(page), page);
744         clear_bit_unlock(PG_locked, &page->flags);
745         smp_mb__after_clear_bit();
746         wake_up_page(page, PG_locked);
747 }
748 EXPORT_SYMBOL(unlock_page);
749
750 /**
751  * end_page_writeback - end writeback against a page
752  * @page: the page
753  */
754 void end_page_writeback(struct page *page)
755 {
756         if (TestClearPageReclaim(page))
757                 rotate_reclaimable_page(page);
758
759         if (!test_clear_page_writeback(page))
760                 BUG();
761
762         smp_mb__after_clear_bit();
763         wake_up_page(page, PG_writeback);
764 }
765 EXPORT_SYMBOL(end_page_writeback);
766
767 /**
768  * __lock_page - get a lock on the page, assuming we need to sleep to get it
769  * @page: the page to lock
770  */
771 void __lock_page(struct page *page)
772 {
773         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
774
775         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
776                                                         TASK_UNINTERRUPTIBLE);
777 }
778 EXPORT_SYMBOL(__lock_page);
779
780 int __lock_page_killable(struct page *page)
781 {
782         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
783
784         return __wait_on_bit_lock(page_waitqueue(page), &wait,
785                                         sleep_on_page_killable, TASK_KILLABLE);
786 }
787 EXPORT_SYMBOL_GPL(__lock_page_killable);
788
789 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
790                          unsigned int flags)
791 {
792         if (flags & FAULT_FLAG_ALLOW_RETRY) {
793                 /*
794                  * CAUTION! In this case, mmap_sem is not released
795                  * even though return 0.
796                  */
797                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
798                         return 0;
799
800                 up_read(&mm->mmap_sem);
801                 if (flags & FAULT_FLAG_KILLABLE)
802                         wait_on_page_locked_killable(page);
803                 else
804                         wait_on_page_locked(page);
805                 return 0;
806         } else {
807                 if (flags & FAULT_FLAG_KILLABLE) {
808                         int ret;
809
810                         ret = __lock_page_killable(page);
811                         if (ret) {
812                                 up_read(&mm->mmap_sem);
813                                 return 0;
814                         }
815                 } else
816                         __lock_page(page);
817                 return 1;
818         }
819 }
820
821 /**
822  * page_cache_next_hole - find the next hole (not-present entry)
823  * @mapping: mapping
824  * @index: index
825  * @max_scan: maximum range to search
826  *
827  * Search the set [index, min(index+max_scan-1, MAX_INDEX)] for the
828  * lowest indexed hole.
829  *
830  * Returns: the index of the hole if found, otherwise returns an index
831  * outside of the set specified (in which case 'return - index >=
832  * max_scan' will be true). In rare cases of index wrap-around, 0 will
833  * be returned.
834  *
835  * page_cache_next_hole may be called under rcu_read_lock. However,
836  * like radix_tree_gang_lookup, this will not atomically search a
837  * snapshot of the tree at a single point in time. For example, if a
838  * hole is created at index 5, then subsequently a hole is created at
839  * index 10, page_cache_next_hole covering both indexes may return 10
840  * if called under rcu_read_lock.
841  */
842 pgoff_t page_cache_next_hole(struct address_space *mapping,
843                              pgoff_t index, unsigned long max_scan)
844 {
845         unsigned long i;
846
847         for (i = 0; i < max_scan; i++) {
848                 struct page *page;
849
850                 page = radix_tree_lookup(&mapping->page_tree, index);
851                 if (!page || radix_tree_exceptional_entry(page))
852                         break;
853                 index++;
854                 if (index == 0)
855                         break;
856         }
857
858         return index;
859 }
860 EXPORT_SYMBOL(page_cache_next_hole);
861
862 /**
863  * page_cache_prev_hole - find the prev hole (not-present entry)
864  * @mapping: mapping
865  * @index: index
866  * @max_scan: maximum range to search
867  *
868  * Search backwards in the range [max(index-max_scan+1, 0), index] for
869  * the first hole.
870  *
871  * Returns: the index of the hole if found, otherwise returns an index
872  * outside of the set specified (in which case 'index - return >=
873  * max_scan' will be true). In rare cases of wrap-around, ULONG_MAX
874  * will be returned.
875  *
876  * page_cache_prev_hole may be called under rcu_read_lock. However,
877  * like radix_tree_gang_lookup, this will not atomically search a
878  * snapshot of the tree at a single point in time. For example, if a
879  * hole is created at index 10, then subsequently a hole is created at
880  * index 5, page_cache_prev_hole covering both indexes may return 5 if
881  * called under rcu_read_lock.
882  */
883 pgoff_t page_cache_prev_hole(struct address_space *mapping,
884                              pgoff_t index, unsigned long max_scan)
885 {
886         unsigned long i;
887
888         for (i = 0; i < max_scan; i++) {
889                 struct page *page;
890
891                 page = radix_tree_lookup(&mapping->page_tree, index);
892                 if (!page || radix_tree_exceptional_entry(page))
893                         break;
894                 index--;
895                 if (index == ULONG_MAX)
896                         break;
897         }
898
899         return index;
900 }
901 EXPORT_SYMBOL(page_cache_prev_hole);
902
903 /**
904  * find_get_entry - find and get a page cache entry
905  * @mapping: the address_space to search
906  * @offset: the page cache index
907  *
908  * Looks up the page cache slot at @mapping & @offset.  If there is a
909  * page cache page, it is returned with an increased refcount.
910  *
911  * If the slot holds a shadow entry of a previously evicted page, or a
912  * swap entry from shmem/tmpfs, it is returned.
913  *
914  * Otherwise, %NULL is returned.
915  */
916 struct page *find_get_entry(struct address_space *mapping, pgoff_t offset)
917 {
918         void **pagep;
919         struct page *page;
920
921         rcu_read_lock();
922 repeat:
923         page = NULL;
924         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
925         if (pagep) {
926                 page = radix_tree_deref_slot(pagep);
927                 if (unlikely(!page))
928                         goto out;
929                 if (radix_tree_exception(page)) {
930                         if (radix_tree_deref_retry(page))
931                                 goto repeat;
932                         /*
933                          * A shadow entry of a recently evicted page,
934                          * or a swap entry from shmem/tmpfs.  Return
935                          * it without attempting to raise page count.
936                          */
937                         goto out;
938                 }
939                 if (!page_cache_get_speculative(page))
940                         goto repeat;
941
942                 /*
943                  * Has the page moved?
944                  * This is part of the lockless pagecache protocol. See
945                  * include/linux/pagemap.h for details.
946                  */
947                 if (unlikely(page != *pagep)) {
948                         page_cache_release(page);
949                         goto repeat;
950                 }
951         }
952 out:
953         rcu_read_unlock();
954
955         return page;
956 }
957 EXPORT_SYMBOL(find_get_entry);
958
959 /**
960  * find_get_page - find and get a page reference
961  * @mapping: the address_space to search
962  * @offset: the page index
963  *
964  * Looks up the page cache slot at @mapping & @offset.  If there is a
965  * page cache page, it is returned with an increased refcount.
966  *
967  * Otherwise, %NULL is returned.
968  */
969 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
970 {
971         struct page *page = find_get_entry(mapping, offset);
972
973         if (radix_tree_exceptional_entry(page))
974                 page = NULL;
975         return page;
976 }
977 EXPORT_SYMBOL(find_get_page);
978
979 /**
980  * find_lock_entry - locate, pin and lock a page cache entry
981  * @mapping: the address_space to search
982  * @offset: the page cache index
983  *
984  * Looks up the page cache slot at @mapping & @offset.  If there is a
985  * page cache page, it is returned locked and with an increased
986  * refcount.
987  *
988  * If the slot holds a shadow entry of a previously evicted page, or a
989  * swap entry from shmem/tmpfs, it is returned.
990  *
991  * Otherwise, %NULL is returned.
992  *
993  * find_lock_entry() may sleep.
994  */
995 struct page *find_lock_entry(struct address_space *mapping, pgoff_t offset)
996 {
997         struct page *page;
998
999 repeat:
1000         page = find_get_entry(mapping, offset);
1001         if (page && !radix_tree_exception(page)) {
1002                 lock_page(page);
1003                 /* Has the page been truncated? */
1004                 if (unlikely(page->mapping != mapping)) {
1005                         unlock_page(page);
1006                         page_cache_release(page);
1007                         goto repeat;
1008                 }
1009                 VM_BUG_ON_PAGE(page->index != offset, page);
1010         }
1011         return page;
1012 }
1013 EXPORT_SYMBOL(find_lock_entry);
1014
1015 /**
1016  * find_lock_page - locate, pin and lock a pagecache page
1017  * @mapping: the address_space to search
1018  * @offset: the page index
1019  *
1020  * Looks up the page cache slot at @mapping & @offset.  If there is a
1021  * page cache page, it is returned locked and with an increased
1022  * refcount.
1023  *
1024  * Otherwise, %NULL is returned.
1025  *
1026  * find_lock_page() may sleep.
1027  */
1028 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
1029 {
1030         struct page *page = find_lock_entry(mapping, offset);
1031
1032         if (radix_tree_exceptional_entry(page))
1033                 page = NULL;
1034         return page;
1035 }
1036 EXPORT_SYMBOL(find_lock_page);
1037
1038 /**
1039  * find_or_create_page - locate or add a pagecache page
1040  * @mapping: the page's address_space
1041  * @index: the page's index into the mapping
1042  * @gfp_mask: page allocation mode
1043  *
1044  * Looks up the page cache slot at @mapping & @offset.  If there is a
1045  * page cache page, it is returned locked and with an increased
1046  * refcount.
1047  *
1048  * If the page is not present, a new page is allocated using @gfp_mask
1049  * and added to the page cache and the VM's LRU list.  The page is
1050  * returned locked and with an increased refcount.
1051  *
1052  * On memory exhaustion, %NULL is returned.
1053  *
1054  * find_or_create_page() may sleep, even if @gfp_flags specifies an
1055  * atomic allocation!
1056  */
1057 struct page *find_or_create_page(struct address_space *mapping,
1058                 pgoff_t index, gfp_t gfp_mask)
1059 {
1060         struct page *page;
1061         int err;
1062 repeat:
1063         page = find_lock_page(mapping, index);
1064         if (!page) {
1065                 page = __page_cache_alloc(gfp_mask);
1066                 if (!page)
1067                         return NULL;
1068                 /*
1069                  * We want a regular kernel memory (not highmem or DMA etc)
1070                  * allocation for the radix tree nodes, but we need to honour
1071                  * the context-specific requirements the caller has asked for.
1072                  * GFP_RECLAIM_MASK collects those requirements.
1073                  */
1074                 err = add_to_page_cache_lru(page, mapping, index,
1075                         (gfp_mask & GFP_RECLAIM_MASK));
1076                 if (unlikely(err)) {
1077                         page_cache_release(page);
1078                         page = NULL;
1079                         if (err == -EEXIST)
1080                                 goto repeat;
1081                 }
1082         }
1083         return page;
1084 }
1085 EXPORT_SYMBOL(find_or_create_page);
1086
1087 /**
1088  * find_get_entries - gang pagecache lookup
1089  * @mapping:    The address_space to search
1090  * @start:      The starting page cache index
1091  * @nr_entries: The maximum number of entries
1092  * @entries:    Where the resulting entries are placed
1093  * @indices:    The cache indices corresponding to the entries in @entries
1094  *
1095  * find_get_entries() will search for and return a group of up to
1096  * @nr_entries entries in the mapping.  The entries are placed at
1097  * @entries.  find_get_entries() takes a reference against any actual
1098  * pages it returns.
1099  *
1100  * The search returns a group of mapping-contiguous page cache entries
1101  * with ascending indexes.  There may be holes in the indices due to
1102  * not-present pages.
1103  *
1104  * Any shadow entries of evicted pages, or swap entries from
1105  * shmem/tmpfs, are included in the returned array.
1106  *
1107  * find_get_entries() returns the number of pages and shadow entries
1108  * which were found.
1109  */
1110 unsigned find_get_entries(struct address_space *mapping,
1111                           pgoff_t start, unsigned int nr_entries,
1112                           struct page **entries, pgoff_t *indices)
1113 {
1114         void **slot;
1115         unsigned int ret = 0;
1116         struct radix_tree_iter iter;
1117
1118         if (!nr_entries)
1119                 return 0;
1120
1121         rcu_read_lock();
1122 restart:
1123         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1124                 struct page *page;
1125 repeat:
1126                 page = radix_tree_deref_slot(slot);
1127                 if (unlikely(!page))
1128                         continue;
1129                 if (radix_tree_exception(page)) {
1130                         if (radix_tree_deref_retry(page))
1131                                 goto restart;
1132                         /*
1133                          * A shadow entry of a recently evicted page,
1134                          * or a swap entry from shmem/tmpfs.  Return
1135                          * it without attempting to raise page count.
1136                          */
1137                         goto export;
1138                 }
1139                 if (!page_cache_get_speculative(page))
1140                         goto repeat;
1141
1142                 /* Has the page moved? */
1143                 if (unlikely(page != *slot)) {
1144                         page_cache_release(page);
1145                         goto repeat;
1146                 }
1147 export:
1148                 indices[ret] = iter.index;
1149                 entries[ret] = page;
1150                 if (++ret == nr_entries)
1151                         break;
1152         }
1153         rcu_read_unlock();
1154         return ret;
1155 }
1156
1157 /**
1158  * find_get_pages - gang pagecache lookup
1159  * @mapping:    The address_space to search
1160  * @start:      The starting page index
1161  * @nr_pages:   The maximum number of pages
1162  * @pages:      Where the resulting pages are placed
1163  *
1164  * find_get_pages() will search for and return a group of up to
1165  * @nr_pages pages in the mapping.  The pages are placed at @pages.
1166  * find_get_pages() takes a reference against the returned pages.
1167  *
1168  * The search returns a group of mapping-contiguous pages with ascending
1169  * indexes.  There may be holes in the indices due to not-present pages.
1170  *
1171  * find_get_pages() returns the number of pages which were found.
1172  */
1173 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
1174                             unsigned int nr_pages, struct page **pages)
1175 {
1176         struct radix_tree_iter iter;
1177         void **slot;
1178         unsigned ret = 0;
1179
1180         if (unlikely(!nr_pages))
1181                 return 0;
1182
1183         rcu_read_lock();
1184 restart:
1185         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1186                 struct page *page;
1187 repeat:
1188                 page = radix_tree_deref_slot(slot);
1189                 if (unlikely(!page))
1190                         continue;
1191
1192                 if (radix_tree_exception(page)) {
1193                         if (radix_tree_deref_retry(page)) {
1194                                 /*
1195                                  * Transient condition which can only trigger
1196                                  * when entry at index 0 moves out of or back
1197                                  * to root: none yet gotten, safe to restart.
1198                                  */
1199                                 WARN_ON(iter.index);
1200                                 goto restart;
1201                         }
1202                         /*
1203                          * A shadow entry of a recently evicted page,
1204                          * or a swap entry from shmem/tmpfs.  Skip
1205                          * over it.
1206                          */
1207                         continue;
1208                 }
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
1219                 pages[ret] = page;
1220                 if (++ret == nr_pages)
1221                         break;
1222         }
1223
1224         rcu_read_unlock();
1225         return ret;
1226 }
1227
1228 /**
1229  * find_get_pages_contig - gang contiguous pagecache lookup
1230  * @mapping:    The address_space to search
1231  * @index:      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_contig() works exactly like find_get_pages(), except
1236  * that the returned number of pages are guaranteed to be contiguous.
1237  *
1238  * find_get_pages_contig() returns the number of pages which were found.
1239  */
1240 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
1241                                unsigned int nr_pages, struct page **pages)
1242 {
1243         struct radix_tree_iter iter;
1244         void **slot;
1245         unsigned int ret = 0;
1246
1247         if (unlikely(!nr_pages))
1248                 return 0;
1249
1250         rcu_read_lock();
1251 restart:
1252         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
1253                 struct page *page;
1254 repeat:
1255                 page = radix_tree_deref_slot(slot);
1256                 /* The hole, there no reason to continue */
1257                 if (unlikely(!page))
1258                         break;
1259
1260                 if (radix_tree_exception(page)) {
1261                         if (radix_tree_deref_retry(page)) {
1262                                 /*
1263                                  * Transient condition which can only trigger
1264                                  * when entry at index 0 moves out of or back
1265                                  * to root: none yet gotten, safe to restart.
1266                                  */
1267                                 goto restart;
1268                         }
1269                         /*
1270                          * A shadow entry of a recently evicted page,
1271                          * or a swap entry from shmem/tmpfs.  Stop
1272                          * looking for contiguous pages.
1273                          */
1274                         break;
1275                 }
1276
1277                 if (!page_cache_get_speculative(page))
1278                         goto repeat;
1279
1280                 /* Has the page moved? */
1281                 if (unlikely(page != *slot)) {
1282                         page_cache_release(page);
1283                         goto repeat;
1284                 }
1285
1286                 /*
1287                  * must check mapping and index after taking the ref.
1288                  * otherwise we can get both false positives and false
1289                  * negatives, which is just confusing to the caller.
1290                  */
1291                 if (page->mapping == NULL || page->index != iter.index) {
1292                         page_cache_release(page);
1293                         break;
1294                 }
1295
1296                 pages[ret] = page;
1297                 if (++ret == nr_pages)
1298                         break;
1299         }
1300         rcu_read_unlock();
1301         return ret;
1302 }
1303 EXPORT_SYMBOL(find_get_pages_contig);
1304
1305 /**
1306  * find_get_pages_tag - find and return pages that match @tag
1307  * @mapping:    the address_space to search
1308  * @index:      the starting page index
1309  * @tag:        the tag index
1310  * @nr_pages:   the maximum number of pages
1311  * @pages:      where the resulting pages are placed
1312  *
1313  * Like find_get_pages, except we only return pages which are tagged with
1314  * @tag.   We update @index to index the next page for the traversal.
1315  */
1316 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
1317                         int tag, unsigned int nr_pages, struct page **pages)
1318 {
1319         struct radix_tree_iter iter;
1320         void **slot;
1321         unsigned ret = 0;
1322
1323         if (unlikely(!nr_pages))
1324                 return 0;
1325
1326         rcu_read_lock();
1327 restart:
1328         radix_tree_for_each_tagged(slot, &mapping->page_tree,
1329                                    &iter, *index, tag) {
1330                 struct page *page;
1331 repeat:
1332                 page = radix_tree_deref_slot(slot);
1333                 if (unlikely(!page))
1334                         continue;
1335
1336                 if (radix_tree_exception(page)) {
1337                         if (radix_tree_deref_retry(page)) {
1338                                 /*
1339                                  * Transient condition which can only trigger
1340                                  * when entry at index 0 moves out of or back
1341                                  * to root: none yet gotten, safe to restart.
1342                                  */
1343                                 goto restart;
1344                         }
1345                         /*
1346                          * A shadow entry of a recently evicted page.
1347                          *
1348                          * Those entries should never be tagged, but
1349                          * this tree walk is lockless and the tags are
1350                          * looked up in bulk, one radix tree node at a
1351                          * time, so there is a sizable window for page
1352                          * reclaim to evict a page we saw tagged.
1353                          *
1354                          * Skip over it.
1355                          */
1356                         continue;
1357                 }
1358
1359                 if (!page_cache_get_speculative(page))
1360                         goto repeat;
1361
1362                 /* Has the page moved? */
1363                 if (unlikely(page != *slot)) {
1364                         page_cache_release(page);
1365                         goto repeat;
1366                 }
1367
1368                 pages[ret] = page;
1369                 if (++ret == nr_pages)
1370                         break;
1371         }
1372
1373         rcu_read_unlock();
1374
1375         if (ret)
1376                 *index = pages[ret - 1]->index + 1;
1377
1378         return ret;
1379 }
1380 EXPORT_SYMBOL(find_get_pages_tag);
1381
1382 /**
1383  * grab_cache_page_nowait - returns locked page at given index in given cache
1384  * @mapping: target address_space
1385  * @index: the page index
1386  *
1387  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1388  * This is intended for speculative data generators, where the data can
1389  * be regenerated if the page couldn't be grabbed.  This routine should
1390  * be safe to call while holding the lock for another page.
1391  *
1392  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1393  * and deadlock against the caller's locked page.
1394  */
1395 struct page *
1396 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1397 {
1398         struct page *page = find_get_page(mapping, index);
1399
1400         if (page) {
1401                 if (trylock_page(page))
1402                         return page;
1403                 page_cache_release(page);
1404                 return NULL;
1405         }
1406         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1407         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1408                 page_cache_release(page);
1409                 page = NULL;
1410         }
1411         return page;
1412 }
1413 EXPORT_SYMBOL(grab_cache_page_nowait);
1414
1415 /*
1416  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1417  * a _large_ part of the i/o request. Imagine the worst scenario:
1418  *
1419  *      ---R__________________________________________B__________
1420  *         ^ reading here                             ^ bad block(assume 4k)
1421  *
1422  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1423  * => failing the whole request => read(R) => read(R+1) =>
1424  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1425  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1426  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1427  *
1428  * It is going insane. Fix it by quickly scaling down the readahead size.
1429  */
1430 static void shrink_readahead_size_eio(struct file *filp,
1431                                         struct file_ra_state *ra)
1432 {
1433         ra->ra_pages /= 4;
1434 }
1435
1436 /**
1437  * do_generic_file_read - generic file read routine
1438  * @filp:       the file to read
1439  * @ppos:       current file position
1440  * @iter:       data destination
1441  * @written:    already copied
1442  *
1443  * This is a generic file read routine, and uses the
1444  * mapping->a_ops->readpage() function for the actual low-level stuff.
1445  *
1446  * This is really ugly. But the goto's actually try to clarify some
1447  * of the logic when it comes to error handling etc.
1448  */
1449 static ssize_t do_generic_file_read(struct file *filp, loff_t *ppos,
1450                 struct iov_iter *iter, ssize_t written)
1451 {
1452         struct address_space *mapping = filp->f_mapping;
1453         struct inode *inode = mapping->host;
1454         struct file_ra_state *ra = &filp->f_ra;
1455         pgoff_t index;
1456         pgoff_t last_index;
1457         pgoff_t prev_index;
1458         unsigned long offset;      /* offset into pagecache page */
1459         unsigned int prev_offset;
1460         int error = 0;
1461
1462         index = *ppos >> PAGE_CACHE_SHIFT;
1463         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1464         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1465         last_index = (*ppos + iter->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1466         offset = *ppos & ~PAGE_CACHE_MASK;
1467
1468         for (;;) {
1469                 struct page *page;
1470                 pgoff_t end_index;
1471                 loff_t isize;
1472                 unsigned long nr, ret;
1473
1474                 cond_resched();
1475 find_page:
1476                 page = find_get_page(mapping, index);
1477                 if (!page) {
1478                         page_cache_sync_readahead(mapping,
1479                                         ra, filp,
1480                                         index, last_index - index);
1481                         page = find_get_page(mapping, index);
1482                         if (unlikely(page == NULL))
1483                                 goto no_cached_page;
1484                 }
1485                 if (PageReadahead(page)) {
1486                         page_cache_async_readahead(mapping,
1487                                         ra, filp, page,
1488                                         index, last_index - index);
1489                 }
1490                 if (!PageUptodate(page)) {
1491                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1492                                         !mapping->a_ops->is_partially_uptodate)
1493                                 goto page_not_up_to_date;
1494                         if (!trylock_page(page))
1495                                 goto page_not_up_to_date;
1496                         /* Did it get truncated before we got the lock? */
1497                         if (!page->mapping)
1498                                 goto page_not_up_to_date_locked;
1499                         if (!mapping->a_ops->is_partially_uptodate(page,
1500                                                         offset, iter->count))
1501                                 goto page_not_up_to_date_locked;
1502                         unlock_page(page);
1503                 }
1504 page_ok:
1505                 /*
1506                  * i_size must be checked after we know the page is Uptodate.
1507                  *
1508                  * Checking i_size after the check allows us to calculate
1509                  * the correct value for "nr", which means the zero-filled
1510                  * part of the page is not copied back to userspace (unless
1511                  * another truncate extends the file - this is desired though).
1512                  */
1513
1514                 isize = i_size_read(inode);
1515                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1516                 if (unlikely(!isize || index > end_index)) {
1517                         page_cache_release(page);
1518                         goto out;
1519                 }
1520
1521                 /* nr is the maximum number of bytes to copy from this page */
1522                 nr = PAGE_CACHE_SIZE;
1523                 if (index == end_index) {
1524                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1525                         if (nr <= offset) {
1526                                 page_cache_release(page);
1527                                 goto out;
1528                         }
1529                 }
1530                 nr = nr - offset;
1531
1532                 /* If users can be writing to this page using arbitrary
1533                  * virtual addresses, take care about potential aliasing
1534                  * before reading the page on the kernel side.
1535                  */
1536                 if (mapping_writably_mapped(mapping))
1537                         flush_dcache_page(page);
1538
1539                 /*
1540                  * When a sequential read accesses a page several times,
1541                  * only mark it as accessed the first time.
1542                  */
1543                 if (prev_index != index || offset != prev_offset)
1544                         mark_page_accessed(page);
1545                 prev_index = index;
1546
1547                 /*
1548                  * Ok, we have the page, and it's up-to-date, so
1549                  * now we can copy it to user space...
1550                  */
1551
1552                 ret = copy_page_to_iter(page, offset, nr, iter);
1553                 offset += ret;
1554                 index += offset >> PAGE_CACHE_SHIFT;
1555                 offset &= ~PAGE_CACHE_MASK;
1556                 prev_offset = offset;
1557
1558                 page_cache_release(page);
1559                 written += ret;
1560                 if (!iov_iter_count(iter))
1561                         goto out;
1562                 if (ret < nr) {
1563                         error = -EFAULT;
1564                         goto out;
1565                 }
1566                 continue;
1567
1568 page_not_up_to_date:
1569                 /* Get exclusive access to the page ... */
1570                 error = lock_page_killable(page);
1571                 if (unlikely(error))
1572                         goto readpage_error;
1573
1574 page_not_up_to_date_locked:
1575                 /* Did it get truncated before we got the lock? */
1576                 if (!page->mapping) {
1577                         unlock_page(page);
1578                         page_cache_release(page);
1579                         continue;
1580                 }
1581
1582                 /* Did somebody else fill it already? */
1583                 if (PageUptodate(page)) {
1584                         unlock_page(page);
1585                         goto page_ok;
1586                 }
1587
1588 readpage:
1589                 /*
1590                  * A previous I/O error may have been due to temporary
1591                  * failures, eg. multipath errors.
1592                  * PG_error will be set again if readpage fails.
1593                  */
1594                 ClearPageError(page);
1595                 /* Start the actual read. The read will unlock the page. */
1596                 error = mapping->a_ops->readpage(filp, page);
1597
1598                 if (unlikely(error)) {
1599                         if (error == AOP_TRUNCATED_PAGE) {
1600                                 page_cache_release(page);
1601                                 error = 0;
1602                                 goto find_page;
1603                         }
1604                         goto readpage_error;
1605                 }
1606
1607                 if (!PageUptodate(page)) {
1608                         error = lock_page_killable(page);
1609                         if (unlikely(error))
1610                                 goto readpage_error;
1611                         if (!PageUptodate(page)) {
1612                                 if (page->mapping == NULL) {
1613                                         /*
1614                                          * invalidate_mapping_pages got it
1615                                          */
1616                                         unlock_page(page);
1617                                         page_cache_release(page);
1618                                         goto find_page;
1619                                 }
1620                                 unlock_page(page);
1621                                 shrink_readahead_size_eio(filp, ra);
1622                                 error = -EIO;
1623                                 goto readpage_error;
1624                         }
1625                         unlock_page(page);
1626                 }
1627
1628                 goto page_ok;
1629
1630 readpage_error:
1631                 /* UHHUH! A synchronous read error occurred. Report it */
1632                 page_cache_release(page);
1633                 goto out;
1634
1635 no_cached_page:
1636                 /*
1637                  * Ok, it wasn't cached, so we need to create a new
1638                  * page..
1639                  */
1640                 page = page_cache_alloc_cold(mapping);
1641                 if (!page) {
1642                         error = -ENOMEM;
1643                         goto out;
1644                 }
1645                 error = add_to_page_cache_lru(page, mapping,
1646                                                 index, GFP_KERNEL);
1647                 if (error) {
1648                         page_cache_release(page);
1649                         if (error == -EEXIST) {
1650                                 error = 0;
1651                                 goto find_page;
1652                         }
1653                         goto out;
1654                 }
1655                 goto readpage;
1656         }
1657
1658 out:
1659         ra->prev_pos = prev_index;
1660         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1661         ra->prev_pos |= prev_offset;
1662
1663         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1664         file_accessed(filp);
1665         return written ? written : error;
1666 }
1667
1668 /*
1669  * Performs necessary checks before doing a write
1670  * @iov:        io vector request
1671  * @nr_segs:    number of segments in the iovec
1672  * @count:      number of bytes to write
1673  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1674  *
1675  * Adjust number of segments and amount of bytes to write (nr_segs should be
1676  * properly initialized first). Returns appropriate error code that caller
1677  * should return or zero in case that write should be allowed.
1678  */
1679 int generic_segment_checks(const struct iovec *iov,
1680                         unsigned long *nr_segs, size_t *count, int access_flags)
1681 {
1682         unsigned long   seg;
1683         size_t cnt = 0;
1684         for (seg = 0; seg < *nr_segs; seg++) {
1685                 const struct iovec *iv = &iov[seg];
1686
1687                 /*
1688                  * If any segment has a negative length, or the cumulative
1689                  * length ever wraps negative then return -EINVAL.
1690                  */
1691                 cnt += iv->iov_len;
1692                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1693                         return -EINVAL;
1694                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1695                         continue;
1696                 if (seg == 0)
1697                         return -EFAULT;
1698                 *nr_segs = seg;
1699                 cnt -= iv->iov_len;     /* This segment is no good */
1700                 break;
1701         }
1702         *count = cnt;
1703         return 0;
1704 }
1705 EXPORT_SYMBOL(generic_segment_checks);
1706
1707 /**
1708  * generic_file_aio_read - generic filesystem read routine
1709  * @iocb:       kernel I/O control block
1710  * @iov:        io vector request
1711  * @nr_segs:    number of segments in the iovec
1712  * @pos:        current file position
1713  *
1714  * This is the "read()" routine for all filesystems
1715  * that can use the page cache directly.
1716  */
1717 ssize_t
1718 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1719                 unsigned long nr_segs, loff_t pos)
1720 {
1721         struct file *filp = iocb->ki_filp;
1722         ssize_t retval;
1723         size_t count;
1724         loff_t *ppos = &iocb->ki_pos;
1725         struct iov_iter i;
1726
1727         count = 0;
1728         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1729         if (retval)
1730                 return retval;
1731         iov_iter_init(&i, iov, nr_segs, count, 0);
1732
1733         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1734         if (filp->f_flags & O_DIRECT) {
1735                 loff_t size;
1736                 struct address_space *mapping;
1737                 struct inode *inode;
1738
1739                 mapping = filp->f_mapping;
1740                 inode = mapping->host;
1741                 if (!count)
1742                         goto out; /* skip atime */
1743                 size = i_size_read(inode);
1744                 retval = filemap_write_and_wait_range(mapping, pos,
1745                                         pos + iov_length(iov, nr_segs) - 1);
1746                 if (!retval) {
1747                         retval = mapping->a_ops->direct_IO(READ, iocb,
1748                                                            iov, pos, nr_segs);
1749                 }
1750                 if (retval > 0) {
1751                         *ppos = pos + retval;
1752                         count -= retval;
1753                         /*
1754                          * If we did a short DIO read we need to skip the
1755                          * section of the iov that we've already read data into.
1756                          */
1757                         iov_iter_advance(&i, retval);
1758                 }
1759
1760                 /*
1761                  * Btrfs can have a short DIO read if we encounter
1762                  * compressed extents, so if there was an error, or if
1763                  * we've already read everything we wanted to, or if
1764                  * there was a short read because we hit EOF, go ahead
1765                  * and return.  Otherwise fallthrough to buffered io for
1766                  * the rest of the read.
1767                  */
1768                 if (retval < 0 || !count || *ppos >= size) {
1769                         file_accessed(filp);
1770                         goto out;
1771                 }
1772         }
1773
1774         retval = do_generic_file_read(filp, ppos, &i, retval);
1775 out:
1776         return retval;
1777 }
1778 EXPORT_SYMBOL(generic_file_aio_read);
1779
1780 #ifdef CONFIG_MMU
1781 /**
1782  * page_cache_read - adds requested page to the page cache if not already there
1783  * @file:       file to read
1784  * @offset:     page index
1785  *
1786  * This adds the requested page to the page cache if it isn't already there,
1787  * and schedules an I/O to read in its contents from disk.
1788  */
1789 static int page_cache_read(struct file *file, pgoff_t offset)
1790 {
1791         struct address_space *mapping = file->f_mapping;
1792         struct page *page; 
1793         int ret;
1794
1795         do {
1796                 page = page_cache_alloc_cold(mapping);
1797                 if (!page)
1798                         return -ENOMEM;
1799
1800                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1801                 if (ret == 0)
1802                         ret = mapping->a_ops->readpage(file, page);
1803                 else if (ret == -EEXIST)
1804                         ret = 0; /* losing race to add is OK */
1805
1806                 page_cache_release(page);
1807
1808         } while (ret == AOP_TRUNCATED_PAGE);
1809                 
1810         return ret;
1811 }
1812
1813 #define MMAP_LOTSAMISS  (100)
1814
1815 /*
1816  * Synchronous readahead happens when we don't even find
1817  * a page in the page cache at all.
1818  */
1819 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1820                                    struct file_ra_state *ra,
1821                                    struct file *file,
1822                                    pgoff_t offset)
1823 {
1824         unsigned long ra_pages;
1825         struct address_space *mapping = file->f_mapping;
1826
1827         /* If we don't want any read-ahead, don't bother */
1828         if (vma->vm_flags & VM_RAND_READ)
1829                 return;
1830         if (!ra->ra_pages)
1831                 return;
1832
1833         if (vma->vm_flags & VM_SEQ_READ) {
1834                 page_cache_sync_readahead(mapping, ra, file, offset,
1835                                           ra->ra_pages);
1836                 return;
1837         }
1838
1839         /* Avoid banging the cache line if not needed */
1840         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1841                 ra->mmap_miss++;
1842
1843         /*
1844          * Do we miss much more than hit in this file? If so,
1845          * stop bothering with read-ahead. It will only hurt.
1846          */
1847         if (ra->mmap_miss > MMAP_LOTSAMISS)
1848                 return;
1849
1850         /*
1851          * mmap read-around
1852          */
1853         ra_pages = max_sane_readahead(ra->ra_pages);
1854         ra->start = max_t(long, 0, offset - ra_pages / 2);
1855         ra->size = ra_pages;
1856         ra->async_size = ra_pages / 4;
1857         ra_submit(ra, mapping, file);
1858 }
1859
1860 /*
1861  * Asynchronous readahead happens when we find the page and PG_readahead,
1862  * so we want to possibly extend the readahead further..
1863  */
1864 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1865                                     struct file_ra_state *ra,
1866                                     struct file *file,
1867                                     struct page *page,
1868                                     pgoff_t offset)
1869 {
1870         struct address_space *mapping = file->f_mapping;
1871
1872         /* If we don't want any read-ahead, don't bother */
1873         if (vma->vm_flags & VM_RAND_READ)
1874                 return;
1875         if (ra->mmap_miss > 0)
1876                 ra->mmap_miss--;
1877         if (PageReadahead(page))
1878                 page_cache_async_readahead(mapping, ra, file,
1879                                            page, offset, ra->ra_pages);
1880 }
1881
1882 /**
1883  * filemap_fault - read in file data for page fault handling
1884  * @vma:        vma in which the fault was taken
1885  * @vmf:        struct vm_fault containing details of the fault
1886  *
1887  * filemap_fault() is invoked via the vma operations vector for a
1888  * mapped memory region to read in file data during a page fault.
1889  *
1890  * The goto's are kind of ugly, but this streamlines the normal case of having
1891  * it in the page cache, and handles the special cases reasonably without
1892  * having a lot of duplicated code.
1893  */
1894 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1895 {
1896         int error;
1897         struct file *file = vma->vm_file;
1898         struct address_space *mapping = file->f_mapping;
1899         struct file_ra_state *ra = &file->f_ra;
1900         struct inode *inode = mapping->host;
1901         pgoff_t offset = vmf->pgoff;
1902         struct page *page;
1903         loff_t size;
1904         int ret = 0;
1905
1906         size = round_up(i_size_read(inode), PAGE_CACHE_SIZE);
1907         if (offset >= size >> PAGE_CACHE_SHIFT)
1908                 return VM_FAULT_SIGBUS;
1909
1910         /*
1911          * Do we have something in the page cache already?
1912          */
1913         page = find_get_page(mapping, offset);
1914         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
1915                 /*
1916                  * We found the page, so try async readahead before
1917                  * waiting for the lock.
1918                  */
1919                 do_async_mmap_readahead(vma, ra, file, page, offset);
1920         } else if (!page) {
1921                 /* No page in the page cache at all */
1922                 do_sync_mmap_readahead(vma, ra, file, offset);
1923                 count_vm_event(PGMAJFAULT);
1924                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1925                 ret = VM_FAULT_MAJOR;
1926 retry_find:
1927                 page = find_get_page(mapping, offset);
1928                 if (!page)
1929                         goto no_cached_page;
1930         }
1931
1932         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1933                 page_cache_release(page);
1934                 return ret | VM_FAULT_RETRY;
1935         }
1936
1937         /* Did it get truncated? */
1938         if (unlikely(page->mapping != mapping)) {
1939                 unlock_page(page);
1940                 put_page(page);
1941                 goto retry_find;
1942         }
1943         VM_BUG_ON_PAGE(page->index != offset, page);
1944
1945         /*
1946          * We have a locked page in the page cache, now we need to check
1947          * that it's up-to-date. If not, it is going to be due to an error.
1948          */
1949         if (unlikely(!PageUptodate(page)))
1950                 goto page_not_uptodate;
1951
1952         /*
1953          * Found the page and have a reference on it.
1954          * We must recheck i_size under page lock.
1955          */
1956         size = round_up(i_size_read(inode), PAGE_CACHE_SIZE);
1957         if (unlikely(offset >= size >> PAGE_CACHE_SHIFT)) {
1958                 unlock_page(page);
1959                 page_cache_release(page);
1960                 return VM_FAULT_SIGBUS;
1961         }
1962
1963         vmf->page = page;
1964         return ret | VM_FAULT_LOCKED;
1965
1966 no_cached_page:
1967         /*
1968          * We're only likely to ever get here if MADV_RANDOM is in
1969          * effect.
1970          */
1971         error = page_cache_read(file, offset);
1972
1973         /*
1974          * The page we want has now been added to the page cache.
1975          * In the unlikely event that someone removed it in the
1976          * meantime, we'll just come back here and read it again.
1977          */
1978         if (error >= 0)
1979                 goto retry_find;
1980
1981         /*
1982          * An error return from page_cache_read can result if the
1983          * system is low on memory, or a problem occurs while trying
1984          * to schedule I/O.
1985          */
1986         if (error == -ENOMEM)
1987                 return VM_FAULT_OOM;
1988         return VM_FAULT_SIGBUS;
1989
1990 page_not_uptodate:
1991         /*
1992          * Umm, take care of errors if the page isn't up-to-date.
1993          * Try to re-read it _once_. We do this synchronously,
1994          * because there really aren't any performance issues here
1995          * and we need to check for errors.
1996          */
1997         ClearPageError(page);
1998         error = mapping->a_ops->readpage(file, page);
1999         if (!error) {
2000                 wait_on_page_locked(page);
2001                 if (!PageUptodate(page))
2002                         error = -EIO;
2003         }
2004         page_cache_release(page);
2005
2006         if (!error || error == AOP_TRUNCATED_PAGE)
2007                 goto retry_find;
2008
2009         /* Things didn't work out. Return zero to tell the mm layer so. */
2010         shrink_readahead_size_eio(file, ra);
2011         return VM_FAULT_SIGBUS;
2012 }
2013 EXPORT_SYMBOL(filemap_fault);
2014
2015 void filemap_map_pages(struct vm_area_struct *vma, struct vm_fault *vmf)
2016 {
2017         struct radix_tree_iter iter;
2018         void **slot;
2019         struct file *file = vma->vm_file;
2020         struct address_space *mapping = file->f_mapping;
2021         loff_t size;
2022         struct page *page;
2023         unsigned long address = (unsigned long) vmf->virtual_address;
2024         unsigned long addr;
2025         pte_t *pte;
2026
2027         rcu_read_lock();
2028         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, vmf->pgoff) {
2029                 if (iter.index > vmf->max_pgoff)
2030                         break;
2031 repeat:
2032                 page = radix_tree_deref_slot(slot);
2033                 if (unlikely(!page))
2034                         goto next;
2035                 if (radix_tree_exception(page)) {
2036                         if (radix_tree_deref_retry(page))
2037                                 break;
2038                         else
2039                                 goto next;
2040                 }
2041
2042                 if (!page_cache_get_speculative(page))
2043                         goto repeat;
2044
2045                 /* Has the page moved? */
2046                 if (unlikely(page != *slot)) {
2047                         page_cache_release(page);
2048                         goto repeat;
2049                 }
2050
2051                 if (!PageUptodate(page) ||
2052                                 PageReadahead(page) ||
2053                                 PageHWPoison(page))
2054                         goto skip;
2055                 if (!trylock_page(page))
2056                         goto skip;
2057
2058                 if (page->mapping != mapping || !PageUptodate(page))
2059                         goto unlock;
2060
2061                 size = round_up(i_size_read(mapping->host), PAGE_CACHE_SIZE);
2062                 if (page->index >= size >> PAGE_CACHE_SHIFT)
2063                         goto unlock;
2064
2065                 pte = vmf->pte + page->index - vmf->pgoff;
2066                 if (!pte_none(*pte))
2067                         goto unlock;
2068
2069                 if (file->f_ra.mmap_miss > 0)
2070                         file->f_ra.mmap_miss--;
2071                 addr = address + (page->index - vmf->pgoff) * PAGE_SIZE;
2072                 do_set_pte(vma, addr, page, pte, false, false);
2073                 unlock_page(page);
2074                 goto next;
2075 unlock:
2076                 unlock_page(page);
2077 skip:
2078                 page_cache_release(page);
2079 next:
2080                 if (iter.index == vmf->max_pgoff)
2081                         break;
2082         }
2083         rcu_read_unlock();
2084 }
2085 EXPORT_SYMBOL(filemap_map_pages);
2086
2087 int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
2088 {
2089         struct page *page = vmf->page;
2090         struct inode *inode = file_inode(vma->vm_file);
2091         int ret = VM_FAULT_LOCKED;
2092
2093         sb_start_pagefault(inode->i_sb);
2094         file_update_time(vma->vm_file);
2095         lock_page(page);
2096         if (page->mapping != inode->i_mapping) {
2097                 unlock_page(page);
2098                 ret = VM_FAULT_NOPAGE;
2099                 goto out;
2100         }
2101         /*
2102          * We mark the page dirty already here so that when freeze is in
2103          * progress, we are guaranteed that writeback during freezing will
2104          * see the dirty page and writeprotect it again.
2105          */
2106         set_page_dirty(page);
2107         wait_for_stable_page(page);
2108 out:
2109         sb_end_pagefault(inode->i_sb);
2110         return ret;
2111 }
2112 EXPORT_SYMBOL(filemap_page_mkwrite);
2113
2114 const struct vm_operations_struct generic_file_vm_ops = {
2115         .fault          = filemap_fault,
2116         .map_pages      = filemap_map_pages,
2117         .page_mkwrite   = filemap_page_mkwrite,
2118         .remap_pages    = generic_file_remap_pages,
2119 };
2120
2121 /* This is used for a general mmap of a disk file */
2122
2123 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2124 {
2125         struct address_space *mapping = file->f_mapping;
2126
2127         if (!mapping->a_ops->readpage)
2128                 return -ENOEXEC;
2129         file_accessed(file);
2130         vma->vm_ops = &generic_file_vm_ops;
2131         return 0;
2132 }
2133
2134 /*
2135  * This is for filesystems which do not implement ->writepage.
2136  */
2137 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
2138 {
2139         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
2140                 return -EINVAL;
2141         return generic_file_mmap(file, vma);
2142 }
2143 #else
2144 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
2145 {
2146         return -ENOSYS;
2147 }
2148 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
2149 {
2150         return -ENOSYS;
2151 }
2152 #endif /* CONFIG_MMU */
2153
2154 EXPORT_SYMBOL(generic_file_mmap);
2155 EXPORT_SYMBOL(generic_file_readonly_mmap);
2156
2157 static struct page *wait_on_page_read(struct page *page)
2158 {
2159         if (!IS_ERR(page)) {
2160                 wait_on_page_locked(page);
2161                 if (!PageUptodate(page)) {
2162                         page_cache_release(page);
2163                         page = ERR_PTR(-EIO);
2164                 }
2165         }
2166         return page;
2167 }
2168
2169 static struct page *__read_cache_page(struct address_space *mapping,
2170                                 pgoff_t index,
2171                                 int (*filler)(void *, struct page *),
2172                                 void *data,
2173                                 gfp_t gfp)
2174 {
2175         struct page *page;
2176         int err;
2177 repeat:
2178         page = find_get_page(mapping, index);
2179         if (!page) {
2180                 page = __page_cache_alloc(gfp | __GFP_COLD);
2181                 if (!page)
2182                         return ERR_PTR(-ENOMEM);
2183                 err = add_to_page_cache_lru(page, mapping, index, gfp);
2184                 if (unlikely(err)) {
2185                         page_cache_release(page);
2186                         if (err == -EEXIST)
2187                                 goto repeat;
2188                         /* Presumably ENOMEM for radix tree node */
2189                         return ERR_PTR(err);
2190                 }
2191                 err = filler(data, page);
2192                 if (err < 0) {
2193                         page_cache_release(page);
2194                         page = ERR_PTR(err);
2195                 } else {
2196                         page = wait_on_page_read(page);
2197                 }
2198         }
2199         return page;
2200 }
2201
2202 static struct page *do_read_cache_page(struct address_space *mapping,
2203                                 pgoff_t index,
2204                                 int (*filler)(void *, struct page *),
2205                                 void *data,
2206                                 gfp_t gfp)
2207
2208 {
2209         struct page *page;
2210         int err;
2211
2212 retry:
2213         page = __read_cache_page(mapping, index, filler, data, gfp);
2214         if (IS_ERR(page))
2215                 return page;
2216         if (PageUptodate(page))
2217                 goto out;
2218
2219         lock_page(page);
2220         if (!page->mapping) {
2221                 unlock_page(page);
2222                 page_cache_release(page);
2223                 goto retry;
2224         }
2225         if (PageUptodate(page)) {
2226                 unlock_page(page);
2227                 goto out;
2228         }
2229         err = filler(data, page);
2230         if (err < 0) {
2231                 page_cache_release(page);
2232                 return ERR_PTR(err);
2233         } else {
2234                 page = wait_on_page_read(page);
2235                 if (IS_ERR(page))
2236                         return page;
2237         }
2238 out:
2239         mark_page_accessed(page);
2240         return page;
2241 }
2242
2243 /**
2244  * read_cache_page - read into page cache, fill it if needed
2245  * @mapping:    the page's address_space
2246  * @index:      the page index
2247  * @filler:     function to perform the read
2248  * @data:       first arg to filler(data, page) function, often left as NULL
2249  *
2250  * Read into the page cache. If a page already exists, and PageUptodate() is
2251  * not set, try to fill the page and wait for it to become unlocked.
2252  *
2253  * If the page does not get brought uptodate, return -EIO.
2254  */
2255 struct page *read_cache_page(struct address_space *mapping,
2256                                 pgoff_t index,
2257                                 int (*filler)(void *, struct page *),
2258                                 void *data)
2259 {
2260         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
2261 }
2262 EXPORT_SYMBOL(read_cache_page);
2263
2264 /**
2265  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
2266  * @mapping:    the page's address_space
2267  * @index:      the page index
2268  * @gfp:        the page allocator flags to use if allocating
2269  *
2270  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
2271  * any new page allocations done using the specified allocation flags.
2272  *
2273  * If the page does not get brought uptodate, return -EIO.
2274  */
2275 struct page *read_cache_page_gfp(struct address_space *mapping,
2276                                 pgoff_t index,
2277                                 gfp_t gfp)
2278 {
2279         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
2280
2281         return do_read_cache_page(mapping, index, filler, NULL, gfp);
2282 }
2283 EXPORT_SYMBOL(read_cache_page_gfp);
2284
2285 /*
2286  * Performs necessary checks before doing a write
2287  *
2288  * Can adjust writing position or amount of bytes to write.
2289  * Returns appropriate error code that caller should return or
2290  * zero in case that write should be allowed.
2291  */
2292 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2293 {
2294         struct inode *inode = file->f_mapping->host;
2295         unsigned long limit = rlimit(RLIMIT_FSIZE);
2296
2297         if (unlikely(*pos < 0))
2298                 return -EINVAL;
2299
2300         if (!isblk) {
2301                 /* FIXME: this is for backwards compatibility with 2.4 */
2302                 if (file->f_flags & O_APPEND)
2303                         *pos = i_size_read(inode);
2304
2305                 if (limit != RLIM_INFINITY) {
2306                         if (*pos >= limit) {
2307                                 send_sig(SIGXFSZ, current, 0);
2308                                 return -EFBIG;
2309                         }
2310                         if (*count > limit - (typeof(limit))*pos) {
2311                                 *count = limit - (typeof(limit))*pos;
2312                         }
2313                 }
2314         }
2315
2316         /*
2317          * LFS rule
2318          */
2319         if (unlikely(*pos + *count > MAX_NON_LFS &&
2320                                 !(file->f_flags & O_LARGEFILE))) {
2321                 if (*pos >= MAX_NON_LFS) {
2322                         return -EFBIG;
2323                 }
2324                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2325                         *count = MAX_NON_LFS - (unsigned long)*pos;
2326                 }
2327         }
2328
2329         /*
2330          * Are we about to exceed the fs block limit ?
2331          *
2332          * If we have written data it becomes a short write.  If we have
2333          * exceeded without writing data we send a signal and return EFBIG.
2334          * Linus frestrict idea will clean these up nicely..
2335          */
2336         if (likely(!isblk)) {
2337                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2338                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2339                                 return -EFBIG;
2340                         }
2341                         /* zero-length writes at ->s_maxbytes are OK */
2342                 }
2343
2344                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2345                         *count = inode->i_sb->s_maxbytes - *pos;
2346         } else {
2347 #ifdef CONFIG_BLOCK
2348                 loff_t isize;
2349                 if (bdev_read_only(I_BDEV(inode)))
2350                         return -EPERM;
2351                 isize = i_size_read(inode);
2352                 if (*pos >= isize) {
2353                         if (*count || *pos > isize)
2354                                 return -ENOSPC;
2355                 }
2356
2357                 if (*pos + *count > isize)
2358                         *count = isize - *pos;
2359 #else
2360                 return -EPERM;
2361 #endif
2362         }
2363         return 0;
2364 }
2365 EXPORT_SYMBOL(generic_write_checks);
2366
2367 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2368                                 loff_t pos, unsigned len, unsigned flags,
2369                                 struct page **pagep, void **fsdata)
2370 {
2371         const struct address_space_operations *aops = mapping->a_ops;
2372
2373         return aops->write_begin(file, mapping, pos, len, flags,
2374                                                         pagep, fsdata);
2375 }
2376 EXPORT_SYMBOL(pagecache_write_begin);
2377
2378 int pagecache_write_end(struct file *file, struct address_space *mapping,
2379                                 loff_t pos, unsigned len, unsigned copied,
2380                                 struct page *page, void *fsdata)
2381 {
2382         const struct address_space_operations *aops = mapping->a_ops;
2383
2384         mark_page_accessed(page);
2385         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2386 }
2387 EXPORT_SYMBOL(pagecache_write_end);
2388
2389 ssize_t
2390 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2391                 unsigned long *nr_segs, loff_t pos,
2392                 size_t count, size_t ocount)
2393 {
2394         struct file     *file = iocb->ki_filp;
2395         struct address_space *mapping = file->f_mapping;
2396         struct inode    *inode = mapping->host;
2397         ssize_t         written;
2398         size_t          write_len;
2399         pgoff_t         end;
2400
2401         if (count != ocount)
2402                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2403
2404         write_len = iov_length(iov, *nr_segs);
2405         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2406
2407         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2408         if (written)
2409                 goto out;
2410
2411         /*
2412          * After a write we want buffered reads to be sure to go to disk to get
2413          * the new data.  We invalidate clean cached page from the region we're
2414          * about to write.  We do this *before* the write so that we can return
2415          * without clobbering -EIOCBQUEUED from ->direct_IO().
2416          */
2417         if (mapping->nrpages) {
2418                 written = invalidate_inode_pages2_range(mapping,
2419                                         pos >> PAGE_CACHE_SHIFT, end);
2420                 /*
2421                  * If a page can not be invalidated, return 0 to fall back
2422                  * to buffered write.
2423                  */
2424                 if (written) {
2425                         if (written == -EBUSY)
2426                                 return 0;
2427                         goto out;
2428                 }
2429         }
2430
2431         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2432
2433         /*
2434          * Finally, try again to invalidate clean pages which might have been
2435          * cached by non-direct readahead, or faulted in by get_user_pages()
2436          * if the source of the write was an mmap'ed region of the file
2437          * we're writing.  Either one is a pretty crazy thing to do,
2438          * so we don't support it 100%.  If this invalidation
2439          * fails, tough, the write still worked...
2440          */
2441         if (mapping->nrpages) {
2442                 invalidate_inode_pages2_range(mapping,
2443                                               pos >> PAGE_CACHE_SHIFT, end);
2444         }
2445
2446         if (written > 0) {
2447                 pos += written;
2448                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2449                         i_size_write(inode, pos);
2450                         mark_inode_dirty(inode);
2451                 }
2452                 iocb->ki_pos = pos;
2453         }
2454 out:
2455         return written;
2456 }
2457 EXPORT_SYMBOL(generic_file_direct_write);
2458
2459 /*
2460  * Find or create a page at the given pagecache position. Return the locked
2461  * page. This function is specifically for buffered writes.
2462  */
2463 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2464                                         pgoff_t index, unsigned flags)
2465 {
2466         int status;
2467         gfp_t gfp_mask;
2468         struct page *page;
2469         gfp_t gfp_notmask = 0;
2470
2471         gfp_mask = mapping_gfp_mask(mapping);
2472         if (mapping_cap_account_dirty(mapping))
2473                 gfp_mask |= __GFP_WRITE;
2474         if (flags & AOP_FLAG_NOFS)
2475                 gfp_notmask = __GFP_FS;
2476 repeat:
2477         page = find_lock_page(mapping, index);
2478         if (page)
2479                 goto found;
2480
2481         page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2482         if (!page)
2483                 return NULL;
2484         status = add_to_page_cache_lru(page, mapping, index,
2485                                                 GFP_KERNEL & ~gfp_notmask);
2486         if (unlikely(status)) {
2487                 page_cache_release(page);
2488                 if (status == -EEXIST)
2489                         goto repeat;
2490                 return NULL;
2491         }
2492 found:
2493         wait_for_stable_page(page);
2494         return page;
2495 }
2496 EXPORT_SYMBOL(grab_cache_page_write_begin);
2497
2498 ssize_t generic_perform_write(struct file *file,
2499                                 struct iov_iter *i, loff_t pos)
2500 {
2501         struct address_space *mapping = file->f_mapping;
2502         const struct address_space_operations *a_ops = mapping->a_ops;
2503         long status = 0;
2504         ssize_t written = 0;
2505         unsigned int flags = 0;
2506
2507         /*
2508          * Copies from kernel address space cannot fail (NFSD is a big user).
2509          */
2510         if (segment_eq(get_fs(), KERNEL_DS))
2511                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2512
2513         do {
2514                 struct page *page;
2515                 unsigned long offset;   /* Offset into pagecache page */
2516                 unsigned long bytes;    /* Bytes to write to page */
2517                 size_t copied;          /* Bytes copied from user */
2518                 void *fsdata;
2519
2520                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2521                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2522                                                 iov_iter_count(i));
2523
2524 again:
2525                 /*
2526                  * Bring in the user page that we will copy from _first_.
2527                  * Otherwise there's a nasty deadlock on copying from the
2528                  * same page as we're writing to, without it being marked
2529                  * up-to-date.
2530                  *
2531                  * Not only is this an optimisation, but it is also required
2532                  * to check that the address is actually valid, when atomic
2533                  * usercopies are used, below.
2534                  */
2535                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2536                         status = -EFAULT;
2537                         break;
2538                 }
2539
2540                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2541                                                 &page, &fsdata);
2542                 if (unlikely(status))
2543                         break;
2544
2545                 if (mapping_writably_mapped(mapping))
2546                         flush_dcache_page(page);
2547
2548                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2549                 flush_dcache_page(page);
2550
2551                 mark_page_accessed(page);
2552                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2553                                                 page, fsdata);
2554                 if (unlikely(status < 0))
2555                         break;
2556                 copied = status;
2557
2558                 cond_resched();
2559
2560                 iov_iter_advance(i, copied);
2561                 if (unlikely(copied == 0)) {
2562                         /*
2563                          * If we were unable to copy any data at all, we must
2564                          * fall back to a single segment length write.
2565                          *
2566                          * If we didn't fallback here, we could livelock
2567                          * because not all segments in the iov can be copied at
2568                          * once without a pagefault.
2569                          */
2570                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2571                                                 iov_iter_single_seg_count(i));
2572                         goto again;
2573                 }
2574                 pos += copied;
2575                 written += copied;
2576
2577                 balance_dirty_pages_ratelimited(mapping);
2578                 if (fatal_signal_pending(current)) {
2579                         status = -EINTR;
2580                         break;
2581                 }
2582         } while (iov_iter_count(i));
2583
2584         return written ? written : status;
2585 }
2586 EXPORT_SYMBOL(generic_perform_write);
2587
2588 /**
2589  * __generic_file_aio_write - write data to a file
2590  * @iocb:       IO state structure (file, offset, etc.)
2591  * @iov:        vector with data to write
2592  * @nr_segs:    number of segments in the vector
2593  *
2594  * This function does all the work needed for actually writing data to a
2595  * file. It does all basic checks, removes SUID from the file, updates
2596  * modification times and calls proper subroutines depending on whether we
2597  * do direct IO or a standard buffered write.
2598  *
2599  * It expects i_mutex to be grabbed unless we work on a block device or similar
2600  * object which does not need locking at all.
2601  *
2602  * This function does *not* take care of syncing data in case of O_SYNC write.
2603  * A caller has to handle it. This is mainly due to the fact that we want to
2604  * avoid syncing under i_mutex.
2605  */
2606 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2607                                  unsigned long nr_segs)
2608 {
2609         struct file *file = iocb->ki_filp;
2610         struct address_space * mapping = file->f_mapping;
2611         size_t ocount;          /* original count */
2612         size_t count;           /* after file limit checks */
2613         struct inode    *inode = mapping->host;
2614         loff_t          pos = iocb->ki_pos;
2615         ssize_t         written = 0;
2616         ssize_t         err;
2617         ssize_t         status;
2618         struct iov_iter from;
2619
2620         ocount = 0;
2621         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2622         if (err)
2623                 return err;
2624
2625         count = ocount;
2626
2627         /* We can write back this queue in page reclaim */
2628         current->backing_dev_info = mapping->backing_dev_info;
2629         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2630         if (err)
2631                 goto out;
2632
2633         if (count == 0)
2634                 goto out;
2635
2636         err = file_remove_suid(file);
2637         if (err)
2638                 goto out;
2639
2640         err = file_update_time(file);
2641         if (err)
2642                 goto out;
2643
2644         iov_iter_init(&from, iov, nr_segs, count, 0);
2645
2646         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2647         if (unlikely(file->f_flags & O_DIRECT)) {
2648                 loff_t endbyte;
2649
2650                 written = generic_file_direct_write(iocb, iov, &from.nr_segs, pos,
2651                                                         count, ocount);
2652                 if (written < 0 || written == count)
2653                         goto out;
2654                 iov_iter_advance(&from, written);
2655
2656                 /*
2657                  * direct-io write to a hole: fall through to buffered I/O
2658                  * for completing the rest of the request.
2659                  */
2660                 pos += written;
2661                 count -= written;
2662
2663                 status = generic_perform_write(file, &from, pos);
2664                 /*
2665                  * If generic_perform_write() returned a synchronous error
2666                  * then we want to return the number of bytes which were
2667                  * direct-written, or the error code if that was zero.  Note
2668                  * that this differs from normal direct-io semantics, which
2669                  * will return -EFOO even if some bytes were written.
2670                  */
2671                 if (unlikely(status < 0) && !written) {
2672                         err = status;
2673                         goto out;
2674                 }
2675                 iocb->ki_pos = pos + status;
2676                 /*
2677                  * We need to ensure that the page cache pages are written to
2678                  * disk and invalidated to preserve the expected O_DIRECT
2679                  * semantics.
2680                  */
2681                 endbyte = pos + status - 1;
2682                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2683                 if (err == 0) {
2684                         written += status;
2685                         invalidate_mapping_pages(mapping,
2686                                                  pos >> PAGE_CACHE_SHIFT,
2687                                                  endbyte >> PAGE_CACHE_SHIFT);
2688                 } else {
2689                         /*
2690                          * We don't know how much we wrote, so just return
2691                          * the number of bytes which were direct-written
2692                          */
2693                 }
2694         } else {
2695                 written = generic_perform_write(file, &from, pos);
2696                 if (likely(written >= 0))
2697                         iocb->ki_pos = pos + written;
2698         }
2699 out:
2700         current->backing_dev_info = NULL;
2701         return written ? written : err;
2702 }
2703 EXPORT_SYMBOL(__generic_file_aio_write);
2704
2705 /**
2706  * generic_file_aio_write - write data to a file
2707  * @iocb:       IO state structure
2708  * @iov:        vector with data to write
2709  * @nr_segs:    number of segments in the vector
2710  * @pos:        position in file where to write
2711  *
2712  * This is a wrapper around __generic_file_aio_write() to be used by most
2713  * filesystems. It takes care of syncing the file in case of O_SYNC file
2714  * and acquires i_mutex as needed.
2715  */
2716 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2717                 unsigned long nr_segs, loff_t pos)
2718 {
2719         struct file *file = iocb->ki_filp;
2720         struct inode *inode = file->f_mapping->host;
2721         ssize_t ret;
2722
2723         BUG_ON(iocb->ki_pos != pos);
2724
2725         mutex_lock(&inode->i_mutex);
2726         ret = __generic_file_aio_write(iocb, iov, nr_segs);
2727         mutex_unlock(&inode->i_mutex);
2728
2729         if (ret > 0) {
2730                 ssize_t err;
2731
2732                 err = generic_write_sync(file, iocb->ki_pos - ret, ret);
2733                 if (err < 0)
2734                         ret = err;
2735         }
2736         return ret;
2737 }
2738 EXPORT_SYMBOL(generic_file_aio_write);
2739
2740 /**
2741  * try_to_release_page() - release old fs-specific metadata on a page
2742  *
2743  * @page: the page which the kernel is trying to free
2744  * @gfp_mask: memory allocation flags (and I/O mode)
2745  *
2746  * The address_space is to try to release any data against the page
2747  * (presumably at page->private).  If the release was successful, return `1'.
2748  * Otherwise return zero.
2749  *
2750  * This may also be called if PG_fscache is set on a page, indicating that the
2751  * page is known to the local caching routines.
2752  *
2753  * The @gfp_mask argument specifies whether I/O may be performed to release
2754  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2755  *
2756  */
2757 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2758 {
2759         struct address_space * const mapping = page->mapping;
2760
2761         BUG_ON(!PageLocked(page));
2762         if (PageWriteback(page))
2763                 return 0;
2764
2765         if (mapping && mapping->a_ops->releasepage)
2766                 return mapping->a_ops->releasepage(page, gfp_mask);
2767         return try_to_free_buffers(page);
2768 }
2769
2770 EXPORT_SYMBOL(try_to_release_page);