truncate: new helpers
[linux-3.10.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/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.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/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
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  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  */
108
109 /*
110  * Remove a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold the mapping's tree_lock.
113  */
114 void __remove_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         radix_tree_delete(&mapping->page_tree, page->index);
119         page->mapping = NULL;
120         mapping->nrpages--;
121         __dec_zone_page_state(page, NR_FILE_PAGES);
122         if (PageSwapBacked(page))
123                 __dec_zone_page_state(page, NR_SHMEM);
124         BUG_ON(page_mapped(page));
125
126         /*
127          * Some filesystems seem to re-dirty the page even after
128          * the VM has canceled the dirty bit (eg ext3 journaling).
129          *
130          * Fix it up by doing a final dirty accounting check after
131          * having removed the page entirely.
132          */
133         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
134                 dec_zone_page_state(page, NR_FILE_DIRTY);
135                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
136         }
137 }
138
139 void remove_from_page_cache(struct page *page)
140 {
141         struct address_space *mapping = page->mapping;
142
143         BUG_ON(!PageLocked(page));
144
145         spin_lock_irq(&mapping->tree_lock);
146         __remove_from_page_cache(page);
147         spin_unlock_irq(&mapping->tree_lock);
148         mem_cgroup_uncharge_cache_page(page);
149 }
150
151 static int sync_page(void *word)
152 {
153         struct address_space *mapping;
154         struct page *page;
155
156         page = container_of((unsigned long *)word, struct page, flags);
157
158         /*
159          * page_mapping() is being called without PG_locked held.
160          * Some knowledge of the state and use of the page is used to
161          * reduce the requirements down to a memory barrier.
162          * The danger here is of a stale page_mapping() return value
163          * indicating a struct address_space different from the one it's
164          * associated with when it is associated with one.
165          * After smp_mb(), it's either the correct page_mapping() for
166          * the page, or an old page_mapping() and the page's own
167          * page_mapping() has gone NULL.
168          * The ->sync_page() address_space operation must tolerate
169          * page_mapping() going NULL. By an amazing coincidence,
170          * this comes about because none of the users of the page
171          * in the ->sync_page() methods make essential use of the
172          * page_mapping(), merely passing the page down to the backing
173          * device's unplug functions when it's non-NULL, which in turn
174          * ignore it for all cases but swap, where only page_private(page) is
175          * of interest. When page_mapping() does go NULL, the entire
176          * call stack gracefully ignores the page and returns.
177          * -- wli
178          */
179         smp_mb();
180         mapping = page_mapping(page);
181         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
182                 mapping->a_ops->sync_page(page);
183         io_schedule();
184         return 0;
185 }
186
187 static int sync_page_killable(void *word)
188 {
189         sync_page(word);
190         return fatal_signal_pending(current) ? -EINTR : 0;
191 }
192
193 /**
194  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
195  * @mapping:    address space structure to write
196  * @start:      offset in bytes where the range starts
197  * @end:        offset in bytes where the range ends (inclusive)
198  * @sync_mode:  enable synchronous operation
199  *
200  * Start writeback against all of a mapping's dirty pages that lie
201  * within the byte offsets <start, end> inclusive.
202  *
203  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
204  * opposed to a regular memory cleansing writeback.  The difference between
205  * these two operations is that if a dirty page/buffer is encountered, it must
206  * be waited upon, and not just skipped over.
207  */
208 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
209                                 loff_t end, int sync_mode)
210 {
211         int ret;
212         struct writeback_control wbc = {
213                 .sync_mode = sync_mode,
214                 .nr_to_write = LONG_MAX,
215                 .range_start = start,
216                 .range_end = end,
217         };
218
219         if (!mapping_cap_writeback_dirty(mapping))
220                 return 0;
221
222         ret = do_writepages(mapping, &wbc);
223         return ret;
224 }
225
226 static inline int __filemap_fdatawrite(struct address_space *mapping,
227         int sync_mode)
228 {
229         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
230 }
231
232 int filemap_fdatawrite(struct address_space *mapping)
233 {
234         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
235 }
236 EXPORT_SYMBOL(filemap_fdatawrite);
237
238 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
239                                 loff_t end)
240 {
241         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
242 }
243 EXPORT_SYMBOL(filemap_fdatawrite_range);
244
245 /**
246  * filemap_flush - mostly a non-blocking flush
247  * @mapping:    target address_space
248  *
249  * This is a mostly non-blocking flush.  Not suitable for data-integrity
250  * purposes - I/O may not be started against all dirty pages.
251  */
252 int filemap_flush(struct address_space *mapping)
253 {
254         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
255 }
256 EXPORT_SYMBOL(filemap_flush);
257
258 /**
259  * wait_on_page_writeback_range - wait for writeback to complete
260  * @mapping:    target address_space
261  * @start:      beginning page index
262  * @end:        ending page index
263  *
264  * Wait for writeback to complete against pages indexed by start->end
265  * inclusive
266  */
267 int wait_on_page_writeback_range(struct address_space *mapping,
268                                 pgoff_t start, pgoff_t end)
269 {
270         struct pagevec pvec;
271         int nr_pages;
272         int ret = 0;
273         pgoff_t index;
274
275         if (end < start)
276                 return 0;
277
278         pagevec_init(&pvec, 0);
279         index = start;
280         while ((index <= end) &&
281                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
282                         PAGECACHE_TAG_WRITEBACK,
283                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
284                 unsigned i;
285
286                 for (i = 0; i < nr_pages; i++) {
287                         struct page *page = pvec.pages[i];
288
289                         /* until radix tree lookup accepts end_index */
290                         if (page->index > end)
291                                 continue;
292
293                         wait_on_page_writeback(page);
294                         if (PageError(page))
295                                 ret = -EIO;
296                 }
297                 pagevec_release(&pvec);
298                 cond_resched();
299         }
300
301         /* Check for outstanding write errors */
302         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
303                 ret = -ENOSPC;
304         if (test_and_clear_bit(AS_EIO, &mapping->flags))
305                 ret = -EIO;
306
307         return ret;
308 }
309
310 /**
311  * filemap_fdatawait_range - wait for all under-writeback pages to complete in a given range
312  * @mapping: address space structure to wait for
313  * @start:      offset in bytes where the range starts
314  * @end:        offset in bytes where the range ends (inclusive)
315  *
316  * Walk the list of under-writeback pages of the given address space
317  * in the given range and wait for all of them.
318  *
319  * This is just a simple wrapper so that callers don't have to convert offsets
320  * to page indexes themselves
321  */
322 int filemap_fdatawait_range(struct address_space *mapping, loff_t start,
323                             loff_t end)
324 {
325         return wait_on_page_writeback_range(mapping, start >> PAGE_CACHE_SHIFT,
326                                             end >> PAGE_CACHE_SHIFT);
327 }
328 EXPORT_SYMBOL(filemap_fdatawait_range);
329
330 /**
331  * filemap_fdatawait - wait for all under-writeback pages to complete
332  * @mapping: address space structure to wait for
333  *
334  * Walk the list of under-writeback pages of the given address space
335  * and wait for all of them.
336  */
337 int filemap_fdatawait(struct address_space *mapping)
338 {
339         loff_t i_size = i_size_read(mapping->host);
340
341         if (i_size == 0)
342                 return 0;
343
344         return wait_on_page_writeback_range(mapping, 0,
345                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
346 }
347 EXPORT_SYMBOL(filemap_fdatawait);
348
349 int filemap_write_and_wait(struct address_space *mapping)
350 {
351         int err = 0;
352
353         if (mapping->nrpages) {
354                 err = filemap_fdatawrite(mapping);
355                 /*
356                  * Even if the above returned error, the pages may be
357                  * written partially (e.g. -ENOSPC), so we wait for it.
358                  * But the -EIO is special case, it may indicate the worst
359                  * thing (e.g. bug) happened, so we avoid waiting for it.
360                  */
361                 if (err != -EIO) {
362                         int err2 = filemap_fdatawait(mapping);
363                         if (!err)
364                                 err = err2;
365                 }
366         }
367         return err;
368 }
369 EXPORT_SYMBOL(filemap_write_and_wait);
370
371 /**
372  * filemap_write_and_wait_range - write out & wait on a file range
373  * @mapping:    the address_space for the pages
374  * @lstart:     offset in bytes where the range starts
375  * @lend:       offset in bytes where the range ends (inclusive)
376  *
377  * Write out and wait upon file offsets lstart->lend, inclusive.
378  *
379  * Note that `lend' is inclusive (describes the last byte to be written) so
380  * that this function can be used to write to the very end-of-file (end = -1).
381  */
382 int filemap_write_and_wait_range(struct address_space *mapping,
383                                  loff_t lstart, loff_t lend)
384 {
385         int err = 0;
386
387         if (mapping->nrpages) {
388                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
389                                                  WB_SYNC_ALL);
390                 /* See comment of filemap_write_and_wait() */
391                 if (err != -EIO) {
392                         int err2 = wait_on_page_writeback_range(mapping,
393                                                 lstart >> PAGE_CACHE_SHIFT,
394                                                 lend >> PAGE_CACHE_SHIFT);
395                         if (!err)
396                                 err = err2;
397                 }
398         }
399         return err;
400 }
401 EXPORT_SYMBOL(filemap_write_and_wait_range);
402
403 /**
404  * add_to_page_cache_locked - add a locked page to the pagecache
405  * @page:       page to add
406  * @mapping:    the page's address_space
407  * @offset:     page index
408  * @gfp_mask:   page allocation mode
409  *
410  * This function is used to add a page to the pagecache. It must be locked.
411  * This function does not add the page to the LRU.  The caller must do that.
412  */
413 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
414                 pgoff_t offset, gfp_t gfp_mask)
415 {
416         int error;
417
418         VM_BUG_ON(!PageLocked(page));
419
420         error = mem_cgroup_cache_charge(page, current->mm,
421                                         gfp_mask & GFP_RECLAIM_MASK);
422         if (error)
423                 goto out;
424
425         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
426         if (error == 0) {
427                 page_cache_get(page);
428                 page->mapping = mapping;
429                 page->index = offset;
430
431                 spin_lock_irq(&mapping->tree_lock);
432                 error = radix_tree_insert(&mapping->page_tree, offset, page);
433                 if (likely(!error)) {
434                         mapping->nrpages++;
435                         __inc_zone_page_state(page, NR_FILE_PAGES);
436                         if (PageSwapBacked(page))
437                                 __inc_zone_page_state(page, NR_SHMEM);
438                         spin_unlock_irq(&mapping->tree_lock);
439                 } else {
440                         page->mapping = NULL;
441                         spin_unlock_irq(&mapping->tree_lock);
442                         mem_cgroup_uncharge_cache_page(page);
443                         page_cache_release(page);
444                 }
445                 radix_tree_preload_end();
446         } else
447                 mem_cgroup_uncharge_cache_page(page);
448 out:
449         return error;
450 }
451 EXPORT_SYMBOL(add_to_page_cache_locked);
452
453 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
454                                 pgoff_t offset, gfp_t gfp_mask)
455 {
456         int ret;
457
458         /*
459          * Splice_read and readahead add shmem/tmpfs pages into the page cache
460          * before shmem_readpage has a chance to mark them as SwapBacked: they
461          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
462          * (called in add_to_page_cache) needs to know where they're going too.
463          */
464         if (mapping_cap_swap_backed(mapping))
465                 SetPageSwapBacked(page);
466
467         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
468         if (ret == 0) {
469                 if (page_is_file_cache(page))
470                         lru_cache_add_file(page);
471                 else
472                         lru_cache_add_active_anon(page);
473         }
474         return ret;
475 }
476 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
477
478 #ifdef CONFIG_NUMA
479 struct page *__page_cache_alloc(gfp_t gfp)
480 {
481         if (cpuset_do_page_mem_spread()) {
482                 int n = cpuset_mem_spread_node();
483                 return alloc_pages_exact_node(n, gfp, 0);
484         }
485         return alloc_pages(gfp, 0);
486 }
487 EXPORT_SYMBOL(__page_cache_alloc);
488 #endif
489
490 static int __sleep_on_page_lock(void *word)
491 {
492         io_schedule();
493         return 0;
494 }
495
496 /*
497  * In order to wait for pages to become available there must be
498  * waitqueues associated with pages. By using a hash table of
499  * waitqueues where the bucket discipline is to maintain all
500  * waiters on the same queue and wake all when any of the pages
501  * become available, and for the woken contexts to check to be
502  * sure the appropriate page became available, this saves space
503  * at a cost of "thundering herd" phenomena during rare hash
504  * collisions.
505  */
506 static wait_queue_head_t *page_waitqueue(struct page *page)
507 {
508         const struct zone *zone = page_zone(page);
509
510         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
511 }
512
513 static inline void wake_up_page(struct page *page, int bit)
514 {
515         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
516 }
517
518 void wait_on_page_bit(struct page *page, int bit_nr)
519 {
520         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
521
522         if (test_bit(bit_nr, &page->flags))
523                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
524                                                         TASK_UNINTERRUPTIBLE);
525 }
526 EXPORT_SYMBOL(wait_on_page_bit);
527
528 /**
529  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
530  * @page: Page defining the wait queue of interest
531  * @waiter: Waiter to add to the queue
532  *
533  * Add an arbitrary @waiter to the wait queue for the nominated @page.
534  */
535 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
536 {
537         wait_queue_head_t *q = page_waitqueue(page);
538         unsigned long flags;
539
540         spin_lock_irqsave(&q->lock, flags);
541         __add_wait_queue(q, waiter);
542         spin_unlock_irqrestore(&q->lock, flags);
543 }
544 EXPORT_SYMBOL_GPL(add_page_wait_queue);
545
546 /**
547  * unlock_page - unlock a locked page
548  * @page: the page
549  *
550  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
551  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
552  * mechananism between PageLocked pages and PageWriteback pages is shared.
553  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
554  *
555  * The mb is necessary to enforce ordering between the clear_bit and the read
556  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
557  */
558 void unlock_page(struct page *page)
559 {
560         VM_BUG_ON(!PageLocked(page));
561         clear_bit_unlock(PG_locked, &page->flags);
562         smp_mb__after_clear_bit();
563         wake_up_page(page, PG_locked);
564 }
565 EXPORT_SYMBOL(unlock_page);
566
567 /**
568  * end_page_writeback - end writeback against a page
569  * @page: the page
570  */
571 void end_page_writeback(struct page *page)
572 {
573         if (TestClearPageReclaim(page))
574                 rotate_reclaimable_page(page);
575
576         if (!test_clear_page_writeback(page))
577                 BUG();
578
579         smp_mb__after_clear_bit();
580         wake_up_page(page, PG_writeback);
581 }
582 EXPORT_SYMBOL(end_page_writeback);
583
584 /**
585  * __lock_page - get a lock on the page, assuming we need to sleep to get it
586  * @page: the page to lock
587  *
588  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
589  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
590  * chances are that on the second loop, the block layer's plug list is empty,
591  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
592  */
593 void __lock_page(struct page *page)
594 {
595         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
596
597         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
598                                                         TASK_UNINTERRUPTIBLE);
599 }
600 EXPORT_SYMBOL(__lock_page);
601
602 int __lock_page_killable(struct page *page)
603 {
604         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
605
606         return __wait_on_bit_lock(page_waitqueue(page), &wait,
607                                         sync_page_killable, TASK_KILLABLE);
608 }
609 EXPORT_SYMBOL_GPL(__lock_page_killable);
610
611 /**
612  * __lock_page_nosync - get a lock on the page, without calling sync_page()
613  * @page: the page to lock
614  *
615  * Variant of lock_page that does not require the caller to hold a reference
616  * on the page's mapping.
617  */
618 void __lock_page_nosync(struct page *page)
619 {
620         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
621         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
622                                                         TASK_UNINTERRUPTIBLE);
623 }
624
625 /**
626  * find_get_page - find and get a page reference
627  * @mapping: the address_space to search
628  * @offset: the page index
629  *
630  * Is there a pagecache struct page at the given (mapping, offset) tuple?
631  * If yes, increment its refcount and return it; if no, return NULL.
632  */
633 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
634 {
635         void **pagep;
636         struct page *page;
637
638         rcu_read_lock();
639 repeat:
640         page = NULL;
641         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
642         if (pagep) {
643                 page = radix_tree_deref_slot(pagep);
644                 if (unlikely(!page || page == RADIX_TREE_RETRY))
645                         goto repeat;
646
647                 if (!page_cache_get_speculative(page))
648                         goto repeat;
649
650                 /*
651                  * Has the page moved?
652                  * This is part of the lockless pagecache protocol. See
653                  * include/linux/pagemap.h for details.
654                  */
655                 if (unlikely(page != *pagep)) {
656                         page_cache_release(page);
657                         goto repeat;
658                 }
659         }
660         rcu_read_unlock();
661
662         return page;
663 }
664 EXPORT_SYMBOL(find_get_page);
665
666 /**
667  * find_lock_page - locate, pin and lock a pagecache page
668  * @mapping: the address_space to search
669  * @offset: the page index
670  *
671  * Locates the desired pagecache page, locks it, increments its reference
672  * count and returns its address.
673  *
674  * Returns zero if the page was not present. find_lock_page() may sleep.
675  */
676 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
677 {
678         struct page *page;
679
680 repeat:
681         page = find_get_page(mapping, offset);
682         if (page) {
683                 lock_page(page);
684                 /* Has the page been truncated? */
685                 if (unlikely(page->mapping != mapping)) {
686                         unlock_page(page);
687                         page_cache_release(page);
688                         goto repeat;
689                 }
690                 VM_BUG_ON(page->index != offset);
691         }
692         return page;
693 }
694 EXPORT_SYMBOL(find_lock_page);
695
696 /**
697  * find_or_create_page - locate or add a pagecache page
698  * @mapping: the page's address_space
699  * @index: the page's index into the mapping
700  * @gfp_mask: page allocation mode
701  *
702  * Locates a page in the pagecache.  If the page is not present, a new page
703  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
704  * LRU list.  The returned page is locked and has its reference count
705  * incremented.
706  *
707  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
708  * allocation!
709  *
710  * find_or_create_page() returns the desired page's address, or zero on
711  * memory exhaustion.
712  */
713 struct page *find_or_create_page(struct address_space *mapping,
714                 pgoff_t index, gfp_t gfp_mask)
715 {
716         struct page *page;
717         int err;
718 repeat:
719         page = find_lock_page(mapping, index);
720         if (!page) {
721                 page = __page_cache_alloc(gfp_mask);
722                 if (!page)
723                         return NULL;
724                 /*
725                  * We want a regular kernel memory (not highmem or DMA etc)
726                  * allocation for the radix tree nodes, but we need to honour
727                  * the context-specific requirements the caller has asked for.
728                  * GFP_RECLAIM_MASK collects those requirements.
729                  */
730                 err = add_to_page_cache_lru(page, mapping, index,
731                         (gfp_mask & GFP_RECLAIM_MASK));
732                 if (unlikely(err)) {
733                         page_cache_release(page);
734                         page = NULL;
735                         if (err == -EEXIST)
736                                 goto repeat;
737                 }
738         }
739         return page;
740 }
741 EXPORT_SYMBOL(find_or_create_page);
742
743 /**
744  * find_get_pages - gang pagecache lookup
745  * @mapping:    The address_space to search
746  * @start:      The starting page index
747  * @nr_pages:   The maximum number of pages
748  * @pages:      Where the resulting pages are placed
749  *
750  * find_get_pages() will search for and return a group of up to
751  * @nr_pages pages in the mapping.  The pages are placed at @pages.
752  * find_get_pages() takes a reference against the returned pages.
753  *
754  * The search returns a group of mapping-contiguous pages with ascending
755  * indexes.  There may be holes in the indices due to not-present pages.
756  *
757  * find_get_pages() returns the number of pages which were found.
758  */
759 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
760                             unsigned int nr_pages, struct page **pages)
761 {
762         unsigned int i;
763         unsigned int ret;
764         unsigned int nr_found;
765
766         rcu_read_lock();
767 restart:
768         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
769                                 (void ***)pages, start, nr_pages);
770         ret = 0;
771         for (i = 0; i < nr_found; i++) {
772                 struct page *page;
773 repeat:
774                 page = radix_tree_deref_slot((void **)pages[i]);
775                 if (unlikely(!page))
776                         continue;
777                 /*
778                  * this can only trigger if nr_found == 1, making livelock
779                  * a non issue.
780                  */
781                 if (unlikely(page == RADIX_TREE_RETRY))
782                         goto restart;
783
784                 if (!page_cache_get_speculative(page))
785                         goto repeat;
786
787                 /* Has the page moved? */
788                 if (unlikely(page != *((void **)pages[i]))) {
789                         page_cache_release(page);
790                         goto repeat;
791                 }
792
793                 pages[ret] = page;
794                 ret++;
795         }
796         rcu_read_unlock();
797         return ret;
798 }
799
800 /**
801  * find_get_pages_contig - gang contiguous pagecache lookup
802  * @mapping:    The address_space to search
803  * @index:      The starting page index
804  * @nr_pages:   The maximum number of pages
805  * @pages:      Where the resulting pages are placed
806  *
807  * find_get_pages_contig() works exactly like find_get_pages(), except
808  * that the returned number of pages are guaranteed to be contiguous.
809  *
810  * find_get_pages_contig() returns the number of pages which were found.
811  */
812 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
813                                unsigned int nr_pages, struct page **pages)
814 {
815         unsigned int i;
816         unsigned int ret;
817         unsigned int nr_found;
818
819         rcu_read_lock();
820 restart:
821         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
822                                 (void ***)pages, index, nr_pages);
823         ret = 0;
824         for (i = 0; i < nr_found; i++) {
825                 struct page *page;
826 repeat:
827                 page = radix_tree_deref_slot((void **)pages[i]);
828                 if (unlikely(!page))
829                         continue;
830                 /*
831                  * this can only trigger if nr_found == 1, making livelock
832                  * a non issue.
833                  */
834                 if (unlikely(page == RADIX_TREE_RETRY))
835                         goto restart;
836
837                 if (page->mapping == NULL || page->index != index)
838                         break;
839
840                 if (!page_cache_get_speculative(page))
841                         goto repeat;
842
843                 /* Has the page moved? */
844                 if (unlikely(page != *((void **)pages[i]))) {
845                         page_cache_release(page);
846                         goto repeat;
847                 }
848
849                 pages[ret] = page;
850                 ret++;
851                 index++;
852         }
853         rcu_read_unlock();
854         return ret;
855 }
856 EXPORT_SYMBOL(find_get_pages_contig);
857
858 /**
859  * find_get_pages_tag - find and return pages that match @tag
860  * @mapping:    the address_space to search
861  * @index:      the starting page index
862  * @tag:        the tag index
863  * @nr_pages:   the maximum number of pages
864  * @pages:      where the resulting pages are placed
865  *
866  * Like find_get_pages, except we only return pages which are tagged with
867  * @tag.   We update @index to index the next page for the traversal.
868  */
869 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
870                         int tag, unsigned int nr_pages, struct page **pages)
871 {
872         unsigned int i;
873         unsigned int ret;
874         unsigned int nr_found;
875
876         rcu_read_lock();
877 restart:
878         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
879                                 (void ***)pages, *index, nr_pages, tag);
880         ret = 0;
881         for (i = 0; i < nr_found; i++) {
882                 struct page *page;
883 repeat:
884                 page = radix_tree_deref_slot((void **)pages[i]);
885                 if (unlikely(!page))
886                         continue;
887                 /*
888                  * this can only trigger if nr_found == 1, making livelock
889                  * a non issue.
890                  */
891                 if (unlikely(page == RADIX_TREE_RETRY))
892                         goto restart;
893
894                 if (!page_cache_get_speculative(page))
895                         goto repeat;
896
897                 /* Has the page moved? */
898                 if (unlikely(page != *((void **)pages[i]))) {
899                         page_cache_release(page);
900                         goto repeat;
901                 }
902
903                 pages[ret] = page;
904                 ret++;
905         }
906         rcu_read_unlock();
907
908         if (ret)
909                 *index = pages[ret - 1]->index + 1;
910
911         return ret;
912 }
913 EXPORT_SYMBOL(find_get_pages_tag);
914
915 /**
916  * grab_cache_page_nowait - returns locked page at given index in given cache
917  * @mapping: target address_space
918  * @index: the page index
919  *
920  * Same as grab_cache_page(), but do not wait if the page is unavailable.
921  * This is intended for speculative data generators, where the data can
922  * be regenerated if the page couldn't be grabbed.  This routine should
923  * be safe to call while holding the lock for another page.
924  *
925  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
926  * and deadlock against the caller's locked page.
927  */
928 struct page *
929 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
930 {
931         struct page *page = find_get_page(mapping, index);
932
933         if (page) {
934                 if (trylock_page(page))
935                         return page;
936                 page_cache_release(page);
937                 return NULL;
938         }
939         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
940         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
941                 page_cache_release(page);
942                 page = NULL;
943         }
944         return page;
945 }
946 EXPORT_SYMBOL(grab_cache_page_nowait);
947
948 /*
949  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
950  * a _large_ part of the i/o request. Imagine the worst scenario:
951  *
952  *      ---R__________________________________________B__________
953  *         ^ reading here                             ^ bad block(assume 4k)
954  *
955  * read(R) => miss => readahead(R...B) => media error => frustrating retries
956  * => failing the whole request => read(R) => read(R+1) =>
957  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
958  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
959  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
960  *
961  * It is going insane. Fix it by quickly scaling down the readahead size.
962  */
963 static void shrink_readahead_size_eio(struct file *filp,
964                                         struct file_ra_state *ra)
965 {
966         ra->ra_pages /= 4;
967 }
968
969 /**
970  * do_generic_file_read - generic file read routine
971  * @filp:       the file to read
972  * @ppos:       current file position
973  * @desc:       read_descriptor
974  * @actor:      read method
975  *
976  * This is a generic file read routine, and uses the
977  * mapping->a_ops->readpage() function for the actual low-level stuff.
978  *
979  * This is really ugly. But the goto's actually try to clarify some
980  * of the logic when it comes to error handling etc.
981  */
982 static void do_generic_file_read(struct file *filp, loff_t *ppos,
983                 read_descriptor_t *desc, read_actor_t actor)
984 {
985         struct address_space *mapping = filp->f_mapping;
986         struct inode *inode = mapping->host;
987         struct file_ra_state *ra = &filp->f_ra;
988         pgoff_t index;
989         pgoff_t last_index;
990         pgoff_t prev_index;
991         unsigned long offset;      /* offset into pagecache page */
992         unsigned int prev_offset;
993         int error;
994
995         index = *ppos >> PAGE_CACHE_SHIFT;
996         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
997         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
998         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
999         offset = *ppos & ~PAGE_CACHE_MASK;
1000
1001         for (;;) {
1002                 struct page *page;
1003                 pgoff_t end_index;
1004                 loff_t isize;
1005                 unsigned long nr, ret;
1006
1007                 cond_resched();
1008 find_page:
1009                 page = find_get_page(mapping, index);
1010                 if (!page) {
1011                         page_cache_sync_readahead(mapping,
1012                                         ra, filp,
1013                                         index, last_index - index);
1014                         page = find_get_page(mapping, index);
1015                         if (unlikely(page == NULL))
1016                                 goto no_cached_page;
1017                 }
1018                 if (PageReadahead(page)) {
1019                         page_cache_async_readahead(mapping,
1020                                         ra, filp, page,
1021                                         index, last_index - index);
1022                 }
1023                 if (!PageUptodate(page)) {
1024                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1025                                         !mapping->a_ops->is_partially_uptodate)
1026                                 goto page_not_up_to_date;
1027                         if (!trylock_page(page))
1028                                 goto page_not_up_to_date;
1029                         if (!mapping->a_ops->is_partially_uptodate(page,
1030                                                                 desc, offset))
1031                                 goto page_not_up_to_date_locked;
1032                         unlock_page(page);
1033                 }
1034 page_ok:
1035                 /*
1036                  * i_size must be checked after we know the page is Uptodate.
1037                  *
1038                  * Checking i_size after the check allows us to calculate
1039                  * the correct value for "nr", which means the zero-filled
1040                  * part of the page is not copied back to userspace (unless
1041                  * another truncate extends the file - this is desired though).
1042                  */
1043
1044                 isize = i_size_read(inode);
1045                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1046                 if (unlikely(!isize || index > end_index)) {
1047                         page_cache_release(page);
1048                         goto out;
1049                 }
1050
1051                 /* nr is the maximum number of bytes to copy from this page */
1052                 nr = PAGE_CACHE_SIZE;
1053                 if (index == end_index) {
1054                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1055                         if (nr <= offset) {
1056                                 page_cache_release(page);
1057                                 goto out;
1058                         }
1059                 }
1060                 nr = nr - offset;
1061
1062                 /* If users can be writing to this page using arbitrary
1063                  * virtual addresses, take care about potential aliasing
1064                  * before reading the page on the kernel side.
1065                  */
1066                 if (mapping_writably_mapped(mapping))
1067                         flush_dcache_page(page);
1068
1069                 /*
1070                  * When a sequential read accesses a page several times,
1071                  * only mark it as accessed the first time.
1072                  */
1073                 if (prev_index != index || offset != prev_offset)
1074                         mark_page_accessed(page);
1075                 prev_index = index;
1076
1077                 /*
1078                  * Ok, we have the page, and it's up-to-date, so
1079                  * now we can copy it to user space...
1080                  *
1081                  * The actor routine returns how many bytes were actually used..
1082                  * NOTE! This may not be the same as how much of a user buffer
1083                  * we filled up (we may be padding etc), so we can only update
1084                  * "pos" here (the actor routine has to update the user buffer
1085                  * pointers and the remaining count).
1086                  */
1087                 ret = actor(desc, page, offset, nr);
1088                 offset += ret;
1089                 index += offset >> PAGE_CACHE_SHIFT;
1090                 offset &= ~PAGE_CACHE_MASK;
1091                 prev_offset = offset;
1092
1093                 page_cache_release(page);
1094                 if (ret == nr && desc->count)
1095                         continue;
1096                 goto out;
1097
1098 page_not_up_to_date:
1099                 /* Get exclusive access to the page ... */
1100                 error = lock_page_killable(page);
1101                 if (unlikely(error))
1102                         goto readpage_error;
1103
1104 page_not_up_to_date_locked:
1105                 /* Did it get truncated before we got the lock? */
1106                 if (!page->mapping) {
1107                         unlock_page(page);
1108                         page_cache_release(page);
1109                         continue;
1110                 }
1111
1112                 /* Did somebody else fill it already? */
1113                 if (PageUptodate(page)) {
1114                         unlock_page(page);
1115                         goto page_ok;
1116                 }
1117
1118 readpage:
1119                 /* Start the actual read. The read will unlock the page. */
1120                 error = mapping->a_ops->readpage(filp, page);
1121
1122                 if (unlikely(error)) {
1123                         if (error == AOP_TRUNCATED_PAGE) {
1124                                 page_cache_release(page);
1125                                 goto find_page;
1126                         }
1127                         goto readpage_error;
1128                 }
1129
1130                 if (!PageUptodate(page)) {
1131                         error = lock_page_killable(page);
1132                         if (unlikely(error))
1133                                 goto readpage_error;
1134                         if (!PageUptodate(page)) {
1135                                 if (page->mapping == NULL) {
1136                                         /*
1137                                          * invalidate_inode_pages got it
1138                                          */
1139                                         unlock_page(page);
1140                                         page_cache_release(page);
1141                                         goto find_page;
1142                                 }
1143                                 unlock_page(page);
1144                                 shrink_readahead_size_eio(filp, ra);
1145                                 error = -EIO;
1146                                 goto readpage_error;
1147                         }
1148                         unlock_page(page);
1149                 }
1150
1151                 goto page_ok;
1152
1153 readpage_error:
1154                 /* UHHUH! A synchronous read error occurred. Report it */
1155                 desc->error = error;
1156                 page_cache_release(page);
1157                 goto out;
1158
1159 no_cached_page:
1160                 /*
1161                  * Ok, it wasn't cached, so we need to create a new
1162                  * page..
1163                  */
1164                 page = page_cache_alloc_cold(mapping);
1165                 if (!page) {
1166                         desc->error = -ENOMEM;
1167                         goto out;
1168                 }
1169                 error = add_to_page_cache_lru(page, mapping,
1170                                                 index, GFP_KERNEL);
1171                 if (error) {
1172                         page_cache_release(page);
1173                         if (error == -EEXIST)
1174                                 goto find_page;
1175                         desc->error = error;
1176                         goto out;
1177                 }
1178                 goto readpage;
1179         }
1180
1181 out:
1182         ra->prev_pos = prev_index;
1183         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1184         ra->prev_pos |= prev_offset;
1185
1186         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1187         file_accessed(filp);
1188 }
1189
1190 int file_read_actor(read_descriptor_t *desc, struct page *page,
1191                         unsigned long offset, unsigned long size)
1192 {
1193         char *kaddr;
1194         unsigned long left, count = desc->count;
1195
1196         if (size > count)
1197                 size = count;
1198
1199         /*
1200          * Faults on the destination of a read are common, so do it before
1201          * taking the kmap.
1202          */
1203         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1204                 kaddr = kmap_atomic(page, KM_USER0);
1205                 left = __copy_to_user_inatomic(desc->arg.buf,
1206                                                 kaddr + offset, size);
1207                 kunmap_atomic(kaddr, KM_USER0);
1208                 if (left == 0)
1209                         goto success;
1210         }
1211
1212         /* Do it the slow way */
1213         kaddr = kmap(page);
1214         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1215         kunmap(page);
1216
1217         if (left) {
1218                 size -= left;
1219                 desc->error = -EFAULT;
1220         }
1221 success:
1222         desc->count = count - size;
1223         desc->written += size;
1224         desc->arg.buf += size;
1225         return size;
1226 }
1227
1228 /*
1229  * Performs necessary checks before doing a write
1230  * @iov:        io vector request
1231  * @nr_segs:    number of segments in the iovec
1232  * @count:      number of bytes to write
1233  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1234  *
1235  * Adjust number of segments and amount of bytes to write (nr_segs should be
1236  * properly initialized first). Returns appropriate error code that caller
1237  * should return or zero in case that write should be allowed.
1238  */
1239 int generic_segment_checks(const struct iovec *iov,
1240                         unsigned long *nr_segs, size_t *count, int access_flags)
1241 {
1242         unsigned long   seg;
1243         size_t cnt = 0;
1244         for (seg = 0; seg < *nr_segs; seg++) {
1245                 const struct iovec *iv = &iov[seg];
1246
1247                 /*
1248                  * If any segment has a negative length, or the cumulative
1249                  * length ever wraps negative then return -EINVAL.
1250                  */
1251                 cnt += iv->iov_len;
1252                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1253                         return -EINVAL;
1254                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1255                         continue;
1256                 if (seg == 0)
1257                         return -EFAULT;
1258                 *nr_segs = seg;
1259                 cnt -= iv->iov_len;     /* This segment is no good */
1260                 break;
1261         }
1262         *count = cnt;
1263         return 0;
1264 }
1265 EXPORT_SYMBOL(generic_segment_checks);
1266
1267 /**
1268  * generic_file_aio_read - generic filesystem read routine
1269  * @iocb:       kernel I/O control block
1270  * @iov:        io vector request
1271  * @nr_segs:    number of segments in the iovec
1272  * @pos:        current file position
1273  *
1274  * This is the "read()" routine for all filesystems
1275  * that can use the page cache directly.
1276  */
1277 ssize_t
1278 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1279                 unsigned long nr_segs, loff_t pos)
1280 {
1281         struct file *filp = iocb->ki_filp;
1282         ssize_t retval;
1283         unsigned long seg;
1284         size_t count;
1285         loff_t *ppos = &iocb->ki_pos;
1286
1287         count = 0;
1288         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1289         if (retval)
1290                 return retval;
1291
1292         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1293         if (filp->f_flags & O_DIRECT) {
1294                 loff_t size;
1295                 struct address_space *mapping;
1296                 struct inode *inode;
1297
1298                 mapping = filp->f_mapping;
1299                 inode = mapping->host;
1300                 if (!count)
1301                         goto out; /* skip atime */
1302                 size = i_size_read(inode);
1303                 if (pos < size) {
1304                         retval = filemap_write_and_wait_range(mapping, pos,
1305                                         pos + iov_length(iov, nr_segs) - 1);
1306                         if (!retval) {
1307                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1308                                                         iov, pos, nr_segs);
1309                         }
1310                         if (retval > 0)
1311                                 *ppos = pos + retval;
1312                         if (retval) {
1313                                 file_accessed(filp);
1314                                 goto out;
1315                         }
1316                 }
1317         }
1318
1319         for (seg = 0; seg < nr_segs; seg++) {
1320                 read_descriptor_t desc;
1321
1322                 desc.written = 0;
1323                 desc.arg.buf = iov[seg].iov_base;
1324                 desc.count = iov[seg].iov_len;
1325                 if (desc.count == 0)
1326                         continue;
1327                 desc.error = 0;
1328                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1329                 retval += desc.written;
1330                 if (desc.error) {
1331                         retval = retval ?: desc.error;
1332                         break;
1333                 }
1334                 if (desc.count > 0)
1335                         break;
1336         }
1337 out:
1338         return retval;
1339 }
1340 EXPORT_SYMBOL(generic_file_aio_read);
1341
1342 static ssize_t
1343 do_readahead(struct address_space *mapping, struct file *filp,
1344              pgoff_t index, unsigned long nr)
1345 {
1346         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1347                 return -EINVAL;
1348
1349         force_page_cache_readahead(mapping, filp, index, nr);
1350         return 0;
1351 }
1352
1353 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1354 {
1355         ssize_t ret;
1356         struct file *file;
1357
1358         ret = -EBADF;
1359         file = fget(fd);
1360         if (file) {
1361                 if (file->f_mode & FMODE_READ) {
1362                         struct address_space *mapping = file->f_mapping;
1363                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1364                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1365                         unsigned long len = end - start + 1;
1366                         ret = do_readahead(mapping, file, start, len);
1367                 }
1368                 fput(file);
1369         }
1370         return ret;
1371 }
1372 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1373 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1374 {
1375         return SYSC_readahead((int) fd, offset, (size_t) count);
1376 }
1377 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1378 #endif
1379
1380 #ifdef CONFIG_MMU
1381 /**
1382  * page_cache_read - adds requested page to the page cache if not already there
1383  * @file:       file to read
1384  * @offset:     page index
1385  *
1386  * This adds the requested page to the page cache if it isn't already there,
1387  * and schedules an I/O to read in its contents from disk.
1388  */
1389 static int page_cache_read(struct file *file, pgoff_t offset)
1390 {
1391         struct address_space *mapping = file->f_mapping;
1392         struct page *page; 
1393         int ret;
1394
1395         do {
1396                 page = page_cache_alloc_cold(mapping);
1397                 if (!page)
1398                         return -ENOMEM;
1399
1400                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1401                 if (ret == 0)
1402                         ret = mapping->a_ops->readpage(file, page);
1403                 else if (ret == -EEXIST)
1404                         ret = 0; /* losing race to add is OK */
1405
1406                 page_cache_release(page);
1407
1408         } while (ret == AOP_TRUNCATED_PAGE);
1409                 
1410         return ret;
1411 }
1412
1413 #define MMAP_LOTSAMISS  (100)
1414
1415 /*
1416  * Synchronous readahead happens when we don't even find
1417  * a page in the page cache at all.
1418  */
1419 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1420                                    struct file_ra_state *ra,
1421                                    struct file *file,
1422                                    pgoff_t offset)
1423 {
1424         unsigned long ra_pages;
1425         struct address_space *mapping = file->f_mapping;
1426
1427         /* If we don't want any read-ahead, don't bother */
1428         if (VM_RandomReadHint(vma))
1429                 return;
1430
1431         if (VM_SequentialReadHint(vma) ||
1432                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1433                 page_cache_sync_readahead(mapping, ra, file, offset,
1434                                           ra->ra_pages);
1435                 return;
1436         }
1437
1438         if (ra->mmap_miss < INT_MAX)
1439                 ra->mmap_miss++;
1440
1441         /*
1442          * Do we miss much more than hit in this file? If so,
1443          * stop bothering with read-ahead. It will only hurt.
1444          */
1445         if (ra->mmap_miss > MMAP_LOTSAMISS)
1446                 return;
1447
1448         /*
1449          * mmap read-around
1450          */
1451         ra_pages = max_sane_readahead(ra->ra_pages);
1452         if (ra_pages) {
1453                 ra->start = max_t(long, 0, offset - ra_pages/2);
1454                 ra->size = ra_pages;
1455                 ra->async_size = 0;
1456                 ra_submit(ra, mapping, file);
1457         }
1458 }
1459
1460 /*
1461  * Asynchronous readahead happens when we find the page and PG_readahead,
1462  * so we want to possibly extend the readahead further..
1463  */
1464 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1465                                     struct file_ra_state *ra,
1466                                     struct file *file,
1467                                     struct page *page,
1468                                     pgoff_t offset)
1469 {
1470         struct address_space *mapping = file->f_mapping;
1471
1472         /* If we don't want any read-ahead, don't bother */
1473         if (VM_RandomReadHint(vma))
1474                 return;
1475         if (ra->mmap_miss > 0)
1476                 ra->mmap_miss--;
1477         if (PageReadahead(page))
1478                 page_cache_async_readahead(mapping, ra, file,
1479                                            page, offset, ra->ra_pages);
1480 }
1481
1482 /**
1483  * filemap_fault - read in file data for page fault handling
1484  * @vma:        vma in which the fault was taken
1485  * @vmf:        struct vm_fault containing details of the fault
1486  *
1487  * filemap_fault() is invoked via the vma operations vector for a
1488  * mapped memory region to read in file data during a page fault.
1489  *
1490  * The goto's are kind of ugly, but this streamlines the normal case of having
1491  * it in the page cache, and handles the special cases reasonably without
1492  * having a lot of duplicated code.
1493  */
1494 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1495 {
1496         int error;
1497         struct file *file = vma->vm_file;
1498         struct address_space *mapping = file->f_mapping;
1499         struct file_ra_state *ra = &file->f_ra;
1500         struct inode *inode = mapping->host;
1501         pgoff_t offset = vmf->pgoff;
1502         struct page *page;
1503         pgoff_t size;
1504         int ret = 0;
1505
1506         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1507         if (offset >= size)
1508                 return VM_FAULT_SIGBUS;
1509
1510         /*
1511          * Do we have something in the page cache already?
1512          */
1513         page = find_get_page(mapping, offset);
1514         if (likely(page)) {
1515                 /*
1516                  * We found the page, so try async readahead before
1517                  * waiting for the lock.
1518                  */
1519                 do_async_mmap_readahead(vma, ra, file, page, offset);
1520                 lock_page(page);
1521
1522                 /* Did it get truncated? */
1523                 if (unlikely(page->mapping != mapping)) {
1524                         unlock_page(page);
1525                         put_page(page);
1526                         goto no_cached_page;
1527                 }
1528         } else {
1529                 /* No page in the page cache at all */
1530                 do_sync_mmap_readahead(vma, ra, file, offset);
1531                 count_vm_event(PGMAJFAULT);
1532                 ret = VM_FAULT_MAJOR;
1533 retry_find:
1534                 page = find_lock_page(mapping, offset);
1535                 if (!page)
1536                         goto no_cached_page;
1537         }
1538
1539         /*
1540          * We have a locked page in the page cache, now we need to check
1541          * that it's up-to-date. If not, it is going to be due to an error.
1542          */
1543         if (unlikely(!PageUptodate(page)))
1544                 goto page_not_uptodate;
1545
1546         /*
1547          * Found the page and have a reference on it.
1548          * We must recheck i_size under page lock.
1549          */
1550         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1551         if (unlikely(offset >= size)) {
1552                 unlock_page(page);
1553                 page_cache_release(page);
1554                 return VM_FAULT_SIGBUS;
1555         }
1556
1557         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1558         vmf->page = page;
1559         return ret | VM_FAULT_LOCKED;
1560
1561 no_cached_page:
1562         /*
1563          * We're only likely to ever get here if MADV_RANDOM is in
1564          * effect.
1565          */
1566         error = page_cache_read(file, offset);
1567
1568         /*
1569          * The page we want has now been added to the page cache.
1570          * In the unlikely event that someone removed it in the
1571          * meantime, we'll just come back here and read it again.
1572          */
1573         if (error >= 0)
1574                 goto retry_find;
1575
1576         /*
1577          * An error return from page_cache_read can result if the
1578          * system is low on memory, or a problem occurs while trying
1579          * to schedule I/O.
1580          */
1581         if (error == -ENOMEM)
1582                 return VM_FAULT_OOM;
1583         return VM_FAULT_SIGBUS;
1584
1585 page_not_uptodate:
1586         /*
1587          * Umm, take care of errors if the page isn't up-to-date.
1588          * Try to re-read it _once_. We do this synchronously,
1589          * because there really aren't any performance issues here
1590          * and we need to check for errors.
1591          */
1592         ClearPageError(page);
1593         error = mapping->a_ops->readpage(file, page);
1594         if (!error) {
1595                 wait_on_page_locked(page);
1596                 if (!PageUptodate(page))
1597                         error = -EIO;
1598         }
1599         page_cache_release(page);
1600
1601         if (!error || error == AOP_TRUNCATED_PAGE)
1602                 goto retry_find;
1603
1604         /* Things didn't work out. Return zero to tell the mm layer so. */
1605         shrink_readahead_size_eio(file, ra);
1606         return VM_FAULT_SIGBUS;
1607 }
1608 EXPORT_SYMBOL(filemap_fault);
1609
1610 struct vm_operations_struct generic_file_vm_ops = {
1611         .fault          = filemap_fault,
1612 };
1613
1614 /* This is used for a general mmap of a disk file */
1615
1616 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1617 {
1618         struct address_space *mapping = file->f_mapping;
1619
1620         if (!mapping->a_ops->readpage)
1621                 return -ENOEXEC;
1622         file_accessed(file);
1623         vma->vm_ops = &generic_file_vm_ops;
1624         vma->vm_flags |= VM_CAN_NONLINEAR;
1625         return 0;
1626 }
1627
1628 /*
1629  * This is for filesystems which do not implement ->writepage.
1630  */
1631 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1632 {
1633         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1634                 return -EINVAL;
1635         return generic_file_mmap(file, vma);
1636 }
1637 #else
1638 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1639 {
1640         return -ENOSYS;
1641 }
1642 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1643 {
1644         return -ENOSYS;
1645 }
1646 #endif /* CONFIG_MMU */
1647
1648 EXPORT_SYMBOL(generic_file_mmap);
1649 EXPORT_SYMBOL(generic_file_readonly_mmap);
1650
1651 static struct page *__read_cache_page(struct address_space *mapping,
1652                                 pgoff_t index,
1653                                 int (*filler)(void *,struct page*),
1654                                 void *data)
1655 {
1656         struct page *page;
1657         int err;
1658 repeat:
1659         page = find_get_page(mapping, index);
1660         if (!page) {
1661                 page = page_cache_alloc_cold(mapping);
1662                 if (!page)
1663                         return ERR_PTR(-ENOMEM);
1664                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1665                 if (unlikely(err)) {
1666                         page_cache_release(page);
1667                         if (err == -EEXIST)
1668                                 goto repeat;
1669                         /* Presumably ENOMEM for radix tree node */
1670                         return ERR_PTR(err);
1671                 }
1672                 err = filler(data, page);
1673                 if (err < 0) {
1674                         page_cache_release(page);
1675                         page = ERR_PTR(err);
1676                 }
1677         }
1678         return page;
1679 }
1680
1681 /**
1682  * read_cache_page_async - read into page cache, fill it if needed
1683  * @mapping:    the page's address_space
1684  * @index:      the page index
1685  * @filler:     function to perform the read
1686  * @data:       destination for read data
1687  *
1688  * Same as read_cache_page, but don't wait for page to become unlocked
1689  * after submitting it to the filler.
1690  *
1691  * Read into the page cache. If a page already exists, and PageUptodate() is
1692  * not set, try to fill the page but don't wait for it to become unlocked.
1693  *
1694  * If the page does not get brought uptodate, return -EIO.
1695  */
1696 struct page *read_cache_page_async(struct address_space *mapping,
1697                                 pgoff_t index,
1698                                 int (*filler)(void *,struct page*),
1699                                 void *data)
1700 {
1701         struct page *page;
1702         int err;
1703
1704 retry:
1705         page = __read_cache_page(mapping, index, filler, data);
1706         if (IS_ERR(page))
1707                 return page;
1708         if (PageUptodate(page))
1709                 goto out;
1710
1711         lock_page(page);
1712         if (!page->mapping) {
1713                 unlock_page(page);
1714                 page_cache_release(page);
1715                 goto retry;
1716         }
1717         if (PageUptodate(page)) {
1718                 unlock_page(page);
1719                 goto out;
1720         }
1721         err = filler(data, page);
1722         if (err < 0) {
1723                 page_cache_release(page);
1724                 return ERR_PTR(err);
1725         }
1726 out:
1727         mark_page_accessed(page);
1728         return page;
1729 }
1730 EXPORT_SYMBOL(read_cache_page_async);
1731
1732 /**
1733  * read_cache_page - read into page cache, fill it if needed
1734  * @mapping:    the page's address_space
1735  * @index:      the page index
1736  * @filler:     function to perform the read
1737  * @data:       destination for read data
1738  *
1739  * Read into the page cache. If a page already exists, and PageUptodate() is
1740  * not set, try to fill the page then wait for it to become unlocked.
1741  *
1742  * If the page does not get brought uptodate, return -EIO.
1743  */
1744 struct page *read_cache_page(struct address_space *mapping,
1745                                 pgoff_t index,
1746                                 int (*filler)(void *,struct page*),
1747                                 void *data)
1748 {
1749         struct page *page;
1750
1751         page = read_cache_page_async(mapping, index, filler, data);
1752         if (IS_ERR(page))
1753                 goto out;
1754         wait_on_page_locked(page);
1755         if (!PageUptodate(page)) {
1756                 page_cache_release(page);
1757                 page = ERR_PTR(-EIO);
1758         }
1759  out:
1760         return page;
1761 }
1762 EXPORT_SYMBOL(read_cache_page);
1763
1764 /*
1765  * The logic we want is
1766  *
1767  *      if suid or (sgid and xgrp)
1768  *              remove privs
1769  */
1770 int should_remove_suid(struct dentry *dentry)
1771 {
1772         mode_t mode = dentry->d_inode->i_mode;
1773         int kill = 0;
1774
1775         /* suid always must be killed */
1776         if (unlikely(mode & S_ISUID))
1777                 kill = ATTR_KILL_SUID;
1778
1779         /*
1780          * sgid without any exec bits is just a mandatory locking mark; leave
1781          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1782          */
1783         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1784                 kill |= ATTR_KILL_SGID;
1785
1786         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1787                 return kill;
1788
1789         return 0;
1790 }
1791 EXPORT_SYMBOL(should_remove_suid);
1792
1793 static int __remove_suid(struct dentry *dentry, int kill)
1794 {
1795         struct iattr newattrs;
1796
1797         newattrs.ia_valid = ATTR_FORCE | kill;
1798         return notify_change(dentry, &newattrs);
1799 }
1800
1801 int file_remove_suid(struct file *file)
1802 {
1803         struct dentry *dentry = file->f_path.dentry;
1804         int killsuid = should_remove_suid(dentry);
1805         int killpriv = security_inode_need_killpriv(dentry);
1806         int error = 0;
1807
1808         if (killpriv < 0)
1809                 return killpriv;
1810         if (killpriv)
1811                 error = security_inode_killpriv(dentry);
1812         if (!error && killsuid)
1813                 error = __remove_suid(dentry, killsuid);
1814
1815         return error;
1816 }
1817 EXPORT_SYMBOL(file_remove_suid);
1818
1819 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1820                         const struct iovec *iov, size_t base, size_t bytes)
1821 {
1822         size_t copied = 0, left = 0;
1823
1824         while (bytes) {
1825                 char __user *buf = iov->iov_base + base;
1826                 int copy = min(bytes, iov->iov_len - base);
1827
1828                 base = 0;
1829                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1830                 copied += copy;
1831                 bytes -= copy;
1832                 vaddr += copy;
1833                 iov++;
1834
1835                 if (unlikely(left))
1836                         break;
1837         }
1838         return copied - left;
1839 }
1840
1841 /*
1842  * Copy as much as we can into the page and return the number of bytes which
1843  * were sucessfully copied.  If a fault is encountered then return the number of
1844  * bytes which were copied.
1845  */
1846 size_t iov_iter_copy_from_user_atomic(struct page *page,
1847                 struct iov_iter *i, unsigned long offset, size_t bytes)
1848 {
1849         char *kaddr;
1850         size_t copied;
1851
1852         BUG_ON(!in_atomic());
1853         kaddr = kmap_atomic(page, KM_USER0);
1854         if (likely(i->nr_segs == 1)) {
1855                 int left;
1856                 char __user *buf = i->iov->iov_base + i->iov_offset;
1857                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1858                 copied = bytes - left;
1859         } else {
1860                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1861                                                 i->iov, i->iov_offset, bytes);
1862         }
1863         kunmap_atomic(kaddr, KM_USER0);
1864
1865         return copied;
1866 }
1867 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1868
1869 /*
1870  * This has the same sideeffects and return value as
1871  * iov_iter_copy_from_user_atomic().
1872  * The difference is that it attempts to resolve faults.
1873  * Page must not be locked.
1874  */
1875 size_t iov_iter_copy_from_user(struct page *page,
1876                 struct iov_iter *i, unsigned long offset, size_t bytes)
1877 {
1878         char *kaddr;
1879         size_t copied;
1880
1881         kaddr = kmap(page);
1882         if (likely(i->nr_segs == 1)) {
1883                 int left;
1884                 char __user *buf = i->iov->iov_base + i->iov_offset;
1885                 left = __copy_from_user(kaddr + offset, buf, bytes);
1886                 copied = bytes - left;
1887         } else {
1888                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1889                                                 i->iov, i->iov_offset, bytes);
1890         }
1891         kunmap(page);
1892         return copied;
1893 }
1894 EXPORT_SYMBOL(iov_iter_copy_from_user);
1895
1896 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1897 {
1898         BUG_ON(i->count < bytes);
1899
1900         if (likely(i->nr_segs == 1)) {
1901                 i->iov_offset += bytes;
1902                 i->count -= bytes;
1903         } else {
1904                 const struct iovec *iov = i->iov;
1905                 size_t base = i->iov_offset;
1906
1907                 /*
1908                  * The !iov->iov_len check ensures we skip over unlikely
1909                  * zero-length segments (without overruning the iovec).
1910                  */
1911                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1912                         int copy;
1913
1914                         copy = min(bytes, iov->iov_len - base);
1915                         BUG_ON(!i->count || i->count < copy);
1916                         i->count -= copy;
1917                         bytes -= copy;
1918                         base += copy;
1919                         if (iov->iov_len == base) {
1920                                 iov++;
1921                                 base = 0;
1922                         }
1923                 }
1924                 i->iov = iov;
1925                 i->iov_offset = base;
1926         }
1927 }
1928 EXPORT_SYMBOL(iov_iter_advance);
1929
1930 /*
1931  * Fault in the first iovec of the given iov_iter, to a maximum length
1932  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1933  * accessed (ie. because it is an invalid address).
1934  *
1935  * writev-intensive code may want this to prefault several iovecs -- that
1936  * would be possible (callers must not rely on the fact that _only_ the
1937  * first iovec will be faulted with the current implementation).
1938  */
1939 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1940 {
1941         char __user *buf = i->iov->iov_base + i->iov_offset;
1942         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1943         return fault_in_pages_readable(buf, bytes);
1944 }
1945 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1946
1947 /*
1948  * Return the count of just the current iov_iter segment.
1949  */
1950 size_t iov_iter_single_seg_count(struct iov_iter *i)
1951 {
1952         const struct iovec *iov = i->iov;
1953         if (i->nr_segs == 1)
1954                 return i->count;
1955         else
1956                 return min(i->count, iov->iov_len - i->iov_offset);
1957 }
1958 EXPORT_SYMBOL(iov_iter_single_seg_count);
1959
1960 /*
1961  * Performs necessary checks before doing a write
1962  *
1963  * Can adjust writing position or amount of bytes to write.
1964  * Returns appropriate error code that caller should return or
1965  * zero in case that write should be allowed.
1966  */
1967 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1968 {
1969         struct inode *inode = file->f_mapping->host;
1970         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1971
1972         if (unlikely(*pos < 0))
1973                 return -EINVAL;
1974
1975         if (!isblk) {
1976                 /* FIXME: this is for backwards compatibility with 2.4 */
1977                 if (file->f_flags & O_APPEND)
1978                         *pos = i_size_read(inode);
1979
1980                 if (limit != RLIM_INFINITY) {
1981                         if (*pos >= limit) {
1982                                 send_sig(SIGXFSZ, current, 0);
1983                                 return -EFBIG;
1984                         }
1985                         if (*count > limit - (typeof(limit))*pos) {
1986                                 *count = limit - (typeof(limit))*pos;
1987                         }
1988                 }
1989         }
1990
1991         /*
1992          * LFS rule
1993          */
1994         if (unlikely(*pos + *count > MAX_NON_LFS &&
1995                                 !(file->f_flags & O_LARGEFILE))) {
1996                 if (*pos >= MAX_NON_LFS) {
1997                         return -EFBIG;
1998                 }
1999                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2000                         *count = MAX_NON_LFS - (unsigned long)*pos;
2001                 }
2002         }
2003
2004         /*
2005          * Are we about to exceed the fs block limit ?
2006          *
2007          * If we have written data it becomes a short write.  If we have
2008          * exceeded without writing data we send a signal and return EFBIG.
2009          * Linus frestrict idea will clean these up nicely..
2010          */
2011         if (likely(!isblk)) {
2012                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2013                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2014                                 return -EFBIG;
2015                         }
2016                         /* zero-length writes at ->s_maxbytes are OK */
2017                 }
2018
2019                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2020                         *count = inode->i_sb->s_maxbytes - *pos;
2021         } else {
2022 #ifdef CONFIG_BLOCK
2023                 loff_t isize;
2024                 if (bdev_read_only(I_BDEV(inode)))
2025                         return -EPERM;
2026                 isize = i_size_read(inode);
2027                 if (*pos >= isize) {
2028                         if (*count || *pos > isize)
2029                                 return -ENOSPC;
2030                 }
2031
2032                 if (*pos + *count > isize)
2033                         *count = isize - *pos;
2034 #else
2035                 return -EPERM;
2036 #endif
2037         }
2038         return 0;
2039 }
2040 EXPORT_SYMBOL(generic_write_checks);
2041
2042 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2043                                 loff_t pos, unsigned len, unsigned flags,
2044                                 struct page **pagep, void **fsdata)
2045 {
2046         const struct address_space_operations *aops = mapping->a_ops;
2047
2048         return aops->write_begin(file, mapping, pos, len, flags,
2049                                                         pagep, fsdata);
2050 }
2051 EXPORT_SYMBOL(pagecache_write_begin);
2052
2053 int pagecache_write_end(struct file *file, struct address_space *mapping,
2054                                 loff_t pos, unsigned len, unsigned copied,
2055                                 struct page *page, void *fsdata)
2056 {
2057         const struct address_space_operations *aops = mapping->a_ops;
2058
2059         mark_page_accessed(page);
2060         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2061 }
2062 EXPORT_SYMBOL(pagecache_write_end);
2063
2064 ssize_t
2065 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2066                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2067                 size_t count, size_t ocount)
2068 {
2069         struct file     *file = iocb->ki_filp;
2070         struct address_space *mapping = file->f_mapping;
2071         struct inode    *inode = mapping->host;
2072         ssize_t         written;
2073         size_t          write_len;
2074         pgoff_t         end;
2075
2076         if (count != ocount)
2077                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2078
2079         write_len = iov_length(iov, *nr_segs);
2080         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2081
2082         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2083         if (written)
2084                 goto out;
2085
2086         /*
2087          * After a write we want buffered reads to be sure to go to disk to get
2088          * the new data.  We invalidate clean cached page from the region we're
2089          * about to write.  We do this *before* the write so that we can return
2090          * without clobbering -EIOCBQUEUED from ->direct_IO().
2091          */
2092         if (mapping->nrpages) {
2093                 written = invalidate_inode_pages2_range(mapping,
2094                                         pos >> PAGE_CACHE_SHIFT, end);
2095                 /*
2096                  * If a page can not be invalidated, return 0 to fall back
2097                  * to buffered write.
2098                  */
2099                 if (written) {
2100                         if (written == -EBUSY)
2101                                 return 0;
2102                         goto out;
2103                 }
2104         }
2105
2106         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2107
2108         /*
2109          * Finally, try again to invalidate clean pages which might have been
2110          * cached by non-direct readahead, or faulted in by get_user_pages()
2111          * if the source of the write was an mmap'ed region of the file
2112          * we're writing.  Either one is a pretty crazy thing to do,
2113          * so we don't support it 100%.  If this invalidation
2114          * fails, tough, the write still worked...
2115          */
2116         if (mapping->nrpages) {
2117                 invalidate_inode_pages2_range(mapping,
2118                                               pos >> PAGE_CACHE_SHIFT, end);
2119         }
2120
2121         if (written > 0) {
2122                 loff_t end = pos + written;
2123                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2124                         i_size_write(inode,  end);
2125                         mark_inode_dirty(inode);
2126                 }
2127                 *ppos = end;
2128         }
2129 out:
2130         return written;
2131 }
2132 EXPORT_SYMBOL(generic_file_direct_write);
2133
2134 /*
2135  * Find or create a page at the given pagecache position. Return the locked
2136  * page. This function is specifically for buffered writes.
2137  */
2138 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2139                                         pgoff_t index, unsigned flags)
2140 {
2141         int status;
2142         struct page *page;
2143         gfp_t gfp_notmask = 0;
2144         if (flags & AOP_FLAG_NOFS)
2145                 gfp_notmask = __GFP_FS;
2146 repeat:
2147         page = find_lock_page(mapping, index);
2148         if (likely(page))
2149                 return page;
2150
2151         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2152         if (!page)
2153                 return NULL;
2154         status = add_to_page_cache_lru(page, mapping, index,
2155                                                 GFP_KERNEL & ~gfp_notmask);
2156         if (unlikely(status)) {
2157                 page_cache_release(page);
2158                 if (status == -EEXIST)
2159                         goto repeat;
2160                 return NULL;
2161         }
2162         return page;
2163 }
2164 EXPORT_SYMBOL(grab_cache_page_write_begin);
2165
2166 static ssize_t generic_perform_write(struct file *file,
2167                                 struct iov_iter *i, loff_t pos)
2168 {
2169         struct address_space *mapping = file->f_mapping;
2170         const struct address_space_operations *a_ops = mapping->a_ops;
2171         long status = 0;
2172         ssize_t written = 0;
2173         unsigned int flags = 0;
2174
2175         /*
2176          * Copies from kernel address space cannot fail (NFSD is a big user).
2177          */
2178         if (segment_eq(get_fs(), KERNEL_DS))
2179                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2180
2181         do {
2182                 struct page *page;
2183                 pgoff_t index;          /* Pagecache index for current page */
2184                 unsigned long offset;   /* Offset into pagecache page */
2185                 unsigned long bytes;    /* Bytes to write to page */
2186                 size_t copied;          /* Bytes copied from user */
2187                 void *fsdata;
2188
2189                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2190                 index = pos >> PAGE_CACHE_SHIFT;
2191                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2192                                                 iov_iter_count(i));
2193
2194 again:
2195
2196                 /*
2197                  * Bring in the user page that we will copy from _first_.
2198                  * Otherwise there's a nasty deadlock on copying from the
2199                  * same page as we're writing to, without it being marked
2200                  * up-to-date.
2201                  *
2202                  * Not only is this an optimisation, but it is also required
2203                  * to check that the address is actually valid, when atomic
2204                  * usercopies are used, below.
2205                  */
2206                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2207                         status = -EFAULT;
2208                         break;
2209                 }
2210
2211                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2212                                                 &page, &fsdata);
2213                 if (unlikely(status))
2214                         break;
2215
2216                 pagefault_disable();
2217                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2218                 pagefault_enable();
2219                 flush_dcache_page(page);
2220
2221                 mark_page_accessed(page);
2222                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2223                                                 page, fsdata);
2224                 if (unlikely(status < 0))
2225                         break;
2226                 copied = status;
2227
2228                 cond_resched();
2229
2230                 iov_iter_advance(i, copied);
2231                 if (unlikely(copied == 0)) {
2232                         /*
2233                          * If we were unable to copy any data at all, we must
2234                          * fall back to a single segment length write.
2235                          *
2236                          * If we didn't fallback here, we could livelock
2237                          * because not all segments in the iov can be copied at
2238                          * once without a pagefault.
2239                          */
2240                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2241                                                 iov_iter_single_seg_count(i));
2242                         goto again;
2243                 }
2244                 pos += copied;
2245                 written += copied;
2246
2247                 balance_dirty_pages_ratelimited(mapping);
2248
2249         } while (iov_iter_count(i));
2250
2251         return written ? written : status;
2252 }
2253
2254 ssize_t
2255 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2256                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2257                 size_t count, ssize_t written)
2258 {
2259         struct file *file = iocb->ki_filp;
2260         struct address_space *mapping = file->f_mapping;
2261         ssize_t status;
2262         struct iov_iter i;
2263
2264         iov_iter_init(&i, iov, nr_segs, count, written);
2265         status = generic_perform_write(file, &i, pos);
2266
2267         if (likely(status >= 0)) {
2268                 written += status;
2269                 *ppos = pos + status;
2270         }
2271         
2272         /*
2273          * If we get here for O_DIRECT writes then we must have fallen through
2274          * to buffered writes (block instantiation inside i_size).  So we sync
2275          * the file data here, to try to honour O_DIRECT expectations.
2276          */
2277         if (unlikely(file->f_flags & O_DIRECT) && written)
2278                 status = filemap_write_and_wait_range(mapping,
2279                                         pos, pos + written - 1);
2280
2281         return written ? written : status;
2282 }
2283 EXPORT_SYMBOL(generic_file_buffered_write);
2284
2285 /**
2286  * __generic_file_aio_write - write data to a file
2287  * @iocb:       IO state structure (file, offset, etc.)
2288  * @iov:        vector with data to write
2289  * @nr_segs:    number of segments in the vector
2290  * @ppos:       position where to write
2291  *
2292  * This function does all the work needed for actually writing data to a
2293  * file. It does all basic checks, removes SUID from the file, updates
2294  * modification times and calls proper subroutines depending on whether we
2295  * do direct IO or a standard buffered write.
2296  *
2297  * It expects i_mutex to be grabbed unless we work on a block device or similar
2298  * object which does not need locking at all.
2299  *
2300  * This function does *not* take care of syncing data in case of O_SYNC write.
2301  * A caller has to handle it. This is mainly due to the fact that we want to
2302  * avoid syncing under i_mutex.
2303  */
2304 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2305                                  unsigned long nr_segs, loff_t *ppos)
2306 {
2307         struct file *file = iocb->ki_filp;
2308         struct address_space * mapping = file->f_mapping;
2309         size_t ocount;          /* original count */
2310         size_t count;           /* after file limit checks */
2311         struct inode    *inode = mapping->host;
2312         loff_t          pos;
2313         ssize_t         written;
2314         ssize_t         err;
2315
2316         ocount = 0;
2317         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2318         if (err)
2319                 return err;
2320
2321         count = ocount;
2322         pos = *ppos;
2323
2324         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2325
2326         /* We can write back this queue in page reclaim */
2327         current->backing_dev_info = mapping->backing_dev_info;
2328         written = 0;
2329
2330         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2331         if (err)
2332                 goto out;
2333
2334         if (count == 0)
2335                 goto out;
2336
2337         err = file_remove_suid(file);
2338         if (err)
2339                 goto out;
2340
2341         file_update_time(file);
2342
2343         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2344         if (unlikely(file->f_flags & O_DIRECT)) {
2345                 loff_t endbyte;
2346                 ssize_t written_buffered;
2347
2348                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2349                                                         ppos, count, ocount);
2350                 if (written < 0 || written == count)
2351                         goto out;
2352                 /*
2353                  * direct-io write to a hole: fall through to buffered I/O
2354                  * for completing the rest of the request.
2355                  */
2356                 pos += written;
2357                 count -= written;
2358                 written_buffered = generic_file_buffered_write(iocb, iov,
2359                                                 nr_segs, pos, ppos, count,
2360                                                 written);
2361                 /*
2362                  * If generic_file_buffered_write() retuned a synchronous error
2363                  * then we want to return the number of bytes which were
2364                  * direct-written, or the error code if that was zero.  Note
2365                  * that this differs from normal direct-io semantics, which
2366                  * will return -EFOO even if some bytes were written.
2367                  */
2368                 if (written_buffered < 0) {
2369                         err = written_buffered;
2370                         goto out;
2371                 }
2372
2373                 /*
2374                  * We need to ensure that the page cache pages are written to
2375                  * disk and invalidated to preserve the expected O_DIRECT
2376                  * semantics.
2377                  */
2378                 endbyte = pos + written_buffered - written - 1;
2379                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2380                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2381                                             SYNC_FILE_RANGE_WRITE|
2382                                             SYNC_FILE_RANGE_WAIT_AFTER);
2383                 if (err == 0) {
2384                         written = written_buffered;
2385                         invalidate_mapping_pages(mapping,
2386                                                  pos >> PAGE_CACHE_SHIFT,
2387                                                  endbyte >> PAGE_CACHE_SHIFT);
2388                 } else {
2389                         /*
2390                          * We don't know how much we wrote, so just return
2391                          * the number of bytes which were direct-written
2392                          */
2393                 }
2394         } else {
2395                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2396                                 pos, ppos, count, written);
2397         }
2398 out:
2399         current->backing_dev_info = NULL;
2400         return written ? written : err;
2401 }
2402 EXPORT_SYMBOL(__generic_file_aio_write);
2403
2404 /**
2405  * generic_file_aio_write - write data to a file
2406  * @iocb:       IO state structure
2407  * @iov:        vector with data to write
2408  * @nr_segs:    number of segments in the vector
2409  * @pos:        position in file where to write
2410  *
2411  * This is a wrapper around __generic_file_aio_write() to be used by most
2412  * filesystems. It takes care of syncing the file in case of O_SYNC file
2413  * and acquires i_mutex as needed.
2414  */
2415 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2416                 unsigned long nr_segs, loff_t pos)
2417 {
2418         struct file *file = iocb->ki_filp;
2419         struct inode *inode = file->f_mapping->host;
2420         ssize_t ret;
2421
2422         BUG_ON(iocb->ki_pos != pos);
2423
2424         mutex_lock(&inode->i_mutex);
2425         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2426         mutex_unlock(&inode->i_mutex);
2427
2428         if (ret > 0 || ret == -EIOCBQUEUED) {
2429                 ssize_t err;
2430
2431                 err = generic_write_sync(file, pos, ret);
2432                 if (err < 0 && ret > 0)
2433                         ret = err;
2434         }
2435         return ret;
2436 }
2437 EXPORT_SYMBOL(generic_file_aio_write);
2438
2439 /**
2440  * try_to_release_page() - release old fs-specific metadata on a page
2441  *
2442  * @page: the page which the kernel is trying to free
2443  * @gfp_mask: memory allocation flags (and I/O mode)
2444  *
2445  * The address_space is to try to release any data against the page
2446  * (presumably at page->private).  If the release was successful, return `1'.
2447  * Otherwise return zero.
2448  *
2449  * This may also be called if PG_fscache is set on a page, indicating that the
2450  * page is known to the local caching routines.
2451  *
2452  * The @gfp_mask argument specifies whether I/O may be performed to release
2453  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2454  *
2455  */
2456 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2457 {
2458         struct address_space * const mapping = page->mapping;
2459
2460         BUG_ON(!PageLocked(page));
2461         if (PageWriteback(page))
2462                 return 0;
2463
2464         if (mapping && mapping->a_ops->releasepage)
2465                 return mapping->a_ops->releasepage(page, gfp_mask);
2466         return try_to_free_buffers(page);
2467 }
2468
2469 EXPORT_SYMBOL(try_to_release_page);