caaaa7adfdf99fd0413c34561d9bd9f93e0adedb
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/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/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/security.h>
31 #include <linux/syscalls.h>
32 #include <linux/cpuset.h>
33 #include "filemap.h"
34 #include "internal.h"
35
36 /*
37  * FIXME: remove all knowledge of the buffer layer from the core VM
38  */
39 #include <linux/buffer_head.h> /* for generic_osync_inode */
40
41 #include <asm/mman.h>
42
43 static ssize_t
44 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
45         loff_t offset, unsigned long nr_segs);
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_lock               (vmtruncate)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_lock
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  ->i_mutex
82  *    ->i_alloc_sem             (various)
83  *
84  *  ->inode_lock
85  *    ->sb_lock                 (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_lock
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
103  *    ->inode_lock              (zap_pte_range->set_page_dirty)
104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
105  *
106  *  ->task->proc_lock
107  *    ->dcache_lock             (proc_pid_lookup)
108  */
109
110 /*
111  * Remove a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
114  */
115 void __remove_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         BUG_ON(page_mapped(page));
124 }
125
126 void remove_from_page_cache(struct page *page)
127 {
128         struct address_space *mapping = page->mapping;
129
130         BUG_ON(!PageLocked(page));
131
132         write_lock_irq(&mapping->tree_lock);
133         __remove_from_page_cache(page);
134         write_unlock_irq(&mapping->tree_lock);
135 }
136
137 static int sync_page(void *word)
138 {
139         struct address_space *mapping;
140         struct page *page;
141
142         page = container_of((unsigned long *)word, struct page, flags);
143
144         /*
145          * page_mapping() is being called without PG_locked held.
146          * Some knowledge of the state and use of the page is used to
147          * reduce the requirements down to a memory barrier.
148          * The danger here is of a stale page_mapping() return value
149          * indicating a struct address_space different from the one it's
150          * associated with when it is associated with one.
151          * After smp_mb(), it's either the correct page_mapping() for
152          * the page, or an old page_mapping() and the page's own
153          * page_mapping() has gone NULL.
154          * The ->sync_page() address_space operation must tolerate
155          * page_mapping() going NULL. By an amazing coincidence,
156          * this comes about because none of the users of the page
157          * in the ->sync_page() methods make essential use of the
158          * page_mapping(), merely passing the page down to the backing
159          * device's unplug functions when it's non-NULL, which in turn
160          * ignore it for all cases but swap, where only page_private(page) is
161          * of interest. When page_mapping() does go NULL, the entire
162          * call stack gracefully ignores the page and returns.
163          * -- wli
164          */
165         smp_mb();
166         mapping = page_mapping(page);
167         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
168                 mapping->a_ops->sync_page(page);
169         io_schedule();
170         return 0;
171 }
172
173 /**
174  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
175  * @mapping:    address space structure to write
176  * @start:      offset in bytes where the range starts
177  * @end:        offset in bytes where the range ends (inclusive)
178  * @sync_mode:  enable synchronous operation
179  *
180  * Start writeback against all of a mapping's dirty pages that lie
181  * within the byte offsets <start, end> inclusive.
182  *
183  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
184  * opposed to a regular memory cleansing writeback.  The difference between
185  * these two operations is that if a dirty page/buffer is encountered, it must
186  * be waited upon, and not just skipped over.
187  */
188 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
189                                 loff_t end, int sync_mode)
190 {
191         int ret;
192         struct writeback_control wbc = {
193                 .sync_mode = sync_mode,
194                 .nr_to_write = mapping->nrpages * 2,
195                 .range_start = start,
196                 .range_end = end,
197         };
198
199         if (!mapping_cap_writeback_dirty(mapping))
200                 return 0;
201
202         ret = do_writepages(mapping, &wbc);
203         return ret;
204 }
205
206 static inline int __filemap_fdatawrite(struct address_space *mapping,
207         int sync_mode)
208 {
209         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
210 }
211
212 int filemap_fdatawrite(struct address_space *mapping)
213 {
214         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
215 }
216 EXPORT_SYMBOL(filemap_fdatawrite);
217
218 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
219                                 loff_t end)
220 {
221         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
222 }
223
224 /**
225  * filemap_flush - mostly a non-blocking flush
226  * @mapping:    target address_space
227  *
228  * This is a mostly non-blocking flush.  Not suitable for data-integrity
229  * purposes - I/O may not be started against all dirty pages.
230  */
231 int filemap_flush(struct address_space *mapping)
232 {
233         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
234 }
235 EXPORT_SYMBOL(filemap_flush);
236
237 /**
238  * wait_on_page_writeback_range - wait for writeback to complete
239  * @mapping:    target address_space
240  * @start:      beginning page index
241  * @end:        ending page index
242  *
243  * Wait for writeback to complete against pages indexed by start->end
244  * inclusive
245  */
246 int wait_on_page_writeback_range(struct address_space *mapping,
247                                 pgoff_t start, pgoff_t end)
248 {
249         struct pagevec pvec;
250         int nr_pages;
251         int ret = 0;
252         pgoff_t index;
253
254         if (end < start)
255                 return 0;
256
257         pagevec_init(&pvec, 0);
258         index = start;
259         while ((index <= end) &&
260                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
261                         PAGECACHE_TAG_WRITEBACK,
262                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
263                 unsigned i;
264
265                 for (i = 0; i < nr_pages; i++) {
266                         struct page *page = pvec.pages[i];
267
268                         /* until radix tree lookup accepts end_index */
269                         if (page->index > end)
270                                 continue;
271
272                         wait_on_page_writeback(page);
273                         if (PageError(page))
274                                 ret = -EIO;
275                 }
276                 pagevec_release(&pvec);
277                 cond_resched();
278         }
279
280         /* Check for outstanding write errors */
281         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
282                 ret = -ENOSPC;
283         if (test_and_clear_bit(AS_EIO, &mapping->flags))
284                 ret = -EIO;
285
286         return ret;
287 }
288
289 /**
290  * sync_page_range - write and wait on all pages in the passed range
291  * @inode:      target inode
292  * @mapping:    target address_space
293  * @pos:        beginning offset in pages to write
294  * @count:      number of bytes to write
295  *
296  * Write and wait upon all the pages in the passed range.  This is a "data
297  * integrity" operation.  It waits upon in-flight writeout before starting and
298  * waiting upon new writeout.  If there was an IO error, return it.
299  *
300  * We need to re-take i_mutex during the generic_osync_inode list walk because
301  * it is otherwise livelockable.
302  */
303 int sync_page_range(struct inode *inode, struct address_space *mapping,
304                         loff_t pos, loff_t count)
305 {
306         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
307         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
308         int ret;
309
310         if (!mapping_cap_writeback_dirty(mapping) || !count)
311                 return 0;
312         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
313         if (ret == 0) {
314                 mutex_lock(&inode->i_mutex);
315                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
316                 mutex_unlock(&inode->i_mutex);
317         }
318         if (ret == 0)
319                 ret = wait_on_page_writeback_range(mapping, start, end);
320         return ret;
321 }
322 EXPORT_SYMBOL(sync_page_range);
323
324 /**
325  * sync_page_range_nolock
326  * @inode:      target inode
327  * @mapping:    target address_space
328  * @pos:        beginning offset in pages to write
329  * @count:      number of bytes to write
330  *
331  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
332  * as it forces O_SYNC writers to different parts of the same file
333  * to be serialised right until io completion.
334  */
335 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
336                            loff_t pos, loff_t count)
337 {
338         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
339         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
340         int ret;
341
342         if (!mapping_cap_writeback_dirty(mapping) || !count)
343                 return 0;
344         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
345         if (ret == 0)
346                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
347         if (ret == 0)
348                 ret = wait_on_page_writeback_range(mapping, start, end);
349         return ret;
350 }
351 EXPORT_SYMBOL(sync_page_range_nolock);
352
353 /**
354  * filemap_fdatawait - wait for all under-writeback pages to complete
355  * @mapping: address space structure to wait for
356  *
357  * Walk the list of under-writeback pages of the given address space
358  * and wait for all of them.
359  */
360 int filemap_fdatawait(struct address_space *mapping)
361 {
362         loff_t i_size = i_size_read(mapping->host);
363
364         if (i_size == 0)
365                 return 0;
366
367         return wait_on_page_writeback_range(mapping, 0,
368                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
369 }
370 EXPORT_SYMBOL(filemap_fdatawait);
371
372 int filemap_write_and_wait(struct address_space *mapping)
373 {
374         int err = 0;
375
376         if (mapping->nrpages) {
377                 err = filemap_fdatawrite(mapping);
378                 /*
379                  * Even if the above returned error, the pages may be
380                  * written partially (e.g. -ENOSPC), so we wait for it.
381                  * But the -EIO is special case, it may indicate the worst
382                  * thing (e.g. bug) happened, so we avoid waiting for it.
383                  */
384                 if (err != -EIO) {
385                         int err2 = filemap_fdatawait(mapping);
386                         if (!err)
387                                 err = err2;
388                 }
389         }
390         return err;
391 }
392 EXPORT_SYMBOL(filemap_write_and_wait);
393
394 /**
395  * filemap_write_and_wait_range - write out & wait on a file range
396  * @mapping:    the address_space for the pages
397  * @lstart:     offset in bytes where the range starts
398  * @lend:       offset in bytes where the range ends (inclusive)
399  *
400  * Write out and wait upon file offsets lstart->lend, inclusive.
401  *
402  * Note that `lend' is inclusive (describes the last byte to be written) so
403  * that this function can be used to write to the very end-of-file (end = -1).
404  */
405 int filemap_write_and_wait_range(struct address_space *mapping,
406                                  loff_t lstart, loff_t lend)
407 {
408         int err = 0;
409
410         if (mapping->nrpages) {
411                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
412                                                  WB_SYNC_ALL);
413                 /* See comment of filemap_write_and_wait() */
414                 if (err != -EIO) {
415                         int err2 = wait_on_page_writeback_range(mapping,
416                                                 lstart >> PAGE_CACHE_SHIFT,
417                                                 lend >> PAGE_CACHE_SHIFT);
418                         if (!err)
419                                 err = err2;
420                 }
421         }
422         return err;
423 }
424
425 /**
426  * add_to_page_cache - add newly allocated pagecache pages
427  * @page:       page to add
428  * @mapping:    the page's address_space
429  * @offset:     page index
430  * @gfp_mask:   page allocation mode
431  *
432  * This function is used to add newly allocated pagecache pages;
433  * the page is new, so we can just run SetPageLocked() against it.
434  * The other page state flags were set by rmqueue().
435  *
436  * This function does not add the page to the LRU.  The caller must do that.
437  */
438 int add_to_page_cache(struct page *page, struct address_space *mapping,
439                 pgoff_t offset, gfp_t gfp_mask)
440 {
441         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
442
443         if (error == 0) {
444                 write_lock_irq(&mapping->tree_lock);
445                 error = radix_tree_insert(&mapping->page_tree, offset, page);
446                 if (!error) {
447                         page_cache_get(page);
448                         SetPageLocked(page);
449                         page->mapping = mapping;
450                         page->index = offset;
451                         mapping->nrpages++;
452                         __inc_zone_page_state(page, NR_FILE_PAGES);
453                 }
454                 write_unlock_irq(&mapping->tree_lock);
455                 radix_tree_preload_end();
456         }
457         return error;
458 }
459 EXPORT_SYMBOL(add_to_page_cache);
460
461 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
462                                 pgoff_t offset, gfp_t gfp_mask)
463 {
464         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
465         if (ret == 0)
466                 lru_cache_add(page);
467         return ret;
468 }
469
470 #ifdef CONFIG_NUMA
471 struct page *__page_cache_alloc(gfp_t gfp)
472 {
473         if (cpuset_do_page_mem_spread()) {
474                 int n = cpuset_mem_spread_node();
475                 return alloc_pages_node(n, gfp, 0);
476         }
477         return alloc_pages(gfp, 0);
478 }
479 EXPORT_SYMBOL(__page_cache_alloc);
480 #endif
481
482 static int __sleep_on_page_lock(void *word)
483 {
484         io_schedule();
485         return 0;
486 }
487
488 /*
489  * In order to wait for pages to become available there must be
490  * waitqueues associated with pages. By using a hash table of
491  * waitqueues where the bucket discipline is to maintain all
492  * waiters on the same queue and wake all when any of the pages
493  * become available, and for the woken contexts to check to be
494  * sure the appropriate page became available, this saves space
495  * at a cost of "thundering herd" phenomena during rare hash
496  * collisions.
497  */
498 static wait_queue_head_t *page_waitqueue(struct page *page)
499 {
500         const struct zone *zone = page_zone(page);
501
502         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
503 }
504
505 static inline void wake_up_page(struct page *page, int bit)
506 {
507         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
508 }
509
510 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
511 {
512         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
513
514         if (test_bit(bit_nr, &page->flags))
515                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
516                                                         TASK_UNINTERRUPTIBLE);
517 }
518 EXPORT_SYMBOL(wait_on_page_bit);
519
520 /**
521  * unlock_page - unlock a locked page
522  * @page: the page
523  *
524  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
525  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
526  * mechananism between PageLocked pages and PageWriteback pages is shared.
527  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
528  *
529  * The first mb is necessary to safely close the critical section opened by the
530  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
531  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
532  * parallel wait_on_page_locked()).
533  */
534 void fastcall unlock_page(struct page *page)
535 {
536         smp_mb__before_clear_bit();
537         if (!TestClearPageLocked(page))
538                 BUG();
539         smp_mb__after_clear_bit(); 
540         wake_up_page(page, PG_locked);
541 }
542 EXPORT_SYMBOL(unlock_page);
543
544 /**
545  * end_page_writeback - end writeback against a page
546  * @page: the page
547  */
548 void end_page_writeback(struct page *page)
549 {
550         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
551                 if (!test_clear_page_writeback(page))
552                         BUG();
553         }
554         smp_mb__after_clear_bit();
555         wake_up_page(page, PG_writeback);
556 }
557 EXPORT_SYMBOL(end_page_writeback);
558
559 /**
560  * __lock_page - get a lock on the page, assuming we need to sleep to get it
561  * @page: the page to lock
562  *
563  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
564  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
565  * chances are that on the second loop, the block layer's plug list is empty,
566  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
567  */
568 void fastcall __lock_page(struct page *page)
569 {
570         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
571
572         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
573                                                         TASK_UNINTERRUPTIBLE);
574 }
575 EXPORT_SYMBOL(__lock_page);
576
577 /*
578  * Variant of lock_page that does not require the caller to hold a reference
579  * on the page's mapping.
580  */
581 void fastcall __lock_page_nosync(struct page *page)
582 {
583         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
584         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
585                                                         TASK_UNINTERRUPTIBLE);
586 }
587
588 /**
589  * find_get_page - find and get a page reference
590  * @mapping: the address_space to search
591  * @offset: the page index
592  *
593  * Is there a pagecache struct page at the given (mapping, offset) tuple?
594  * If yes, increment its refcount and return it; if no, return NULL.
595  */
596 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
597 {
598         struct page *page;
599
600         read_lock_irq(&mapping->tree_lock);
601         page = radix_tree_lookup(&mapping->page_tree, offset);
602         if (page)
603                 page_cache_get(page);
604         read_unlock_irq(&mapping->tree_lock);
605         return page;
606 }
607 EXPORT_SYMBOL(find_get_page);
608
609 /**
610  * find_lock_page - locate, pin and lock a pagecache page
611  * @mapping: the address_space to search
612  * @offset: the page index
613  *
614  * Locates the desired pagecache page, locks it, increments its reference
615  * count and returns its address.
616  *
617  * Returns zero if the page was not present. find_lock_page() may sleep.
618  */
619 struct page *find_lock_page(struct address_space *mapping,
620                                 pgoff_t offset)
621 {
622         struct page *page;
623
624 repeat:
625         read_lock_irq(&mapping->tree_lock);
626         page = radix_tree_lookup(&mapping->page_tree, offset);
627         if (page) {
628                 page_cache_get(page);
629                 if (TestSetPageLocked(page)) {
630                         read_unlock_irq(&mapping->tree_lock);
631                         __lock_page(page);
632
633                         /* Has the page been truncated while we slept? */
634                         if (unlikely(page->mapping != mapping)) {
635                                 unlock_page(page);
636                                 page_cache_release(page);
637                                 goto repeat;
638                         }
639                         VM_BUG_ON(page->index != offset);
640                         goto out;
641                 }
642         }
643         read_unlock_irq(&mapping->tree_lock);
644 out:
645         return page;
646 }
647 EXPORT_SYMBOL(find_lock_page);
648
649 /**
650  * find_or_create_page - locate or add a pagecache page
651  * @mapping: the page's address_space
652  * @index: the page's index into the mapping
653  * @gfp_mask: page allocation mode
654  *
655  * Locates a page in the pagecache.  If the page is not present, a new page
656  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
657  * LRU list.  The returned page is locked and has its reference count
658  * incremented.
659  *
660  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
661  * allocation!
662  *
663  * find_or_create_page() returns the desired page's address, or zero on
664  * memory exhaustion.
665  */
666 struct page *find_or_create_page(struct address_space *mapping,
667                 pgoff_t index, gfp_t gfp_mask)
668 {
669         struct page *page, *cached_page = NULL;
670         int err;
671 repeat:
672         page = find_lock_page(mapping, index);
673         if (!page) {
674                 if (!cached_page) {
675                         cached_page =
676                                 __page_cache_alloc(gfp_mask);
677                         if (!cached_page)
678                                 return NULL;
679                 }
680                 err = add_to_page_cache_lru(cached_page, mapping,
681                                         index, gfp_mask);
682                 if (!err) {
683                         page = cached_page;
684                         cached_page = NULL;
685                 } else if (err == -EEXIST)
686                         goto repeat;
687         }
688         if (cached_page)
689                 page_cache_release(cached_page);
690         return page;
691 }
692 EXPORT_SYMBOL(find_or_create_page);
693
694 /**
695  * find_get_pages - gang pagecache lookup
696  * @mapping:    The address_space to search
697  * @start:      The starting page index
698  * @nr_pages:   The maximum number of pages
699  * @pages:      Where the resulting pages are placed
700  *
701  * find_get_pages() will search for and return a group of up to
702  * @nr_pages pages in the mapping.  The pages are placed at @pages.
703  * find_get_pages() takes a reference against the returned pages.
704  *
705  * The search returns a group of mapping-contiguous pages with ascending
706  * indexes.  There may be holes in the indices due to not-present pages.
707  *
708  * find_get_pages() returns the number of pages which were found.
709  */
710 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
711                             unsigned int nr_pages, struct page **pages)
712 {
713         unsigned int i;
714         unsigned int ret;
715
716         read_lock_irq(&mapping->tree_lock);
717         ret = radix_tree_gang_lookup(&mapping->page_tree,
718                                 (void **)pages, start, nr_pages);
719         for (i = 0; i < ret; i++)
720                 page_cache_get(pages[i]);
721         read_unlock_irq(&mapping->tree_lock);
722         return ret;
723 }
724
725 /**
726  * find_get_pages_contig - gang contiguous pagecache lookup
727  * @mapping:    The address_space to search
728  * @index:      The starting page index
729  * @nr_pages:   The maximum number of pages
730  * @pages:      Where the resulting pages are placed
731  *
732  * find_get_pages_contig() works exactly like find_get_pages(), except
733  * that the returned number of pages are guaranteed to be contiguous.
734  *
735  * find_get_pages_contig() returns the number of pages which were found.
736  */
737 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
738                                unsigned int nr_pages, struct page **pages)
739 {
740         unsigned int i;
741         unsigned int ret;
742
743         read_lock_irq(&mapping->tree_lock);
744         ret = radix_tree_gang_lookup(&mapping->page_tree,
745                                 (void **)pages, index, nr_pages);
746         for (i = 0; i < ret; i++) {
747                 if (pages[i]->mapping == NULL || pages[i]->index != index)
748                         break;
749
750                 page_cache_get(pages[i]);
751                 index++;
752         }
753         read_unlock_irq(&mapping->tree_lock);
754         return i;
755 }
756 EXPORT_SYMBOL(find_get_pages_contig);
757
758 /**
759  * find_get_pages_tag - find and return pages that match @tag
760  * @mapping:    the address_space to search
761  * @index:      the starting page index
762  * @tag:        the tag index
763  * @nr_pages:   the maximum number of pages
764  * @pages:      where the resulting pages are placed
765  *
766  * Like find_get_pages, except we only return pages which are tagged with
767  * @tag.   We update @index to index the next page for the traversal.
768  */
769 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
770                         int tag, unsigned int nr_pages, struct page **pages)
771 {
772         unsigned int i;
773         unsigned int ret;
774
775         read_lock_irq(&mapping->tree_lock);
776         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
777                                 (void **)pages, *index, nr_pages, tag);
778         for (i = 0; i < ret; i++)
779                 page_cache_get(pages[i]);
780         if (ret)
781                 *index = pages[ret - 1]->index + 1;
782         read_unlock_irq(&mapping->tree_lock);
783         return ret;
784 }
785 EXPORT_SYMBOL(find_get_pages_tag);
786
787 /**
788  * grab_cache_page_nowait - returns locked page at given index in given cache
789  * @mapping: target address_space
790  * @index: the page index
791  *
792  * Same as grab_cache_page(), but do not wait if the page is unavailable.
793  * This is intended for speculative data generators, where the data can
794  * be regenerated if the page couldn't be grabbed.  This routine should
795  * be safe to call while holding the lock for another page.
796  *
797  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
798  * and deadlock against the caller's locked page.
799  */
800 struct page *
801 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
802 {
803         struct page *page = find_get_page(mapping, index);
804
805         if (page) {
806                 if (!TestSetPageLocked(page))
807                         return page;
808                 page_cache_release(page);
809                 return NULL;
810         }
811         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
812         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
813                 page_cache_release(page);
814                 page = NULL;
815         }
816         return page;
817 }
818 EXPORT_SYMBOL(grab_cache_page_nowait);
819
820 /*
821  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
822  * a _large_ part of the i/o request. Imagine the worst scenario:
823  *
824  *      ---R__________________________________________B__________
825  *         ^ reading here                             ^ bad block(assume 4k)
826  *
827  * read(R) => miss => readahead(R...B) => media error => frustrating retries
828  * => failing the whole request => read(R) => read(R+1) =>
829  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
830  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
831  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
832  *
833  * It is going insane. Fix it by quickly scaling down the readahead size.
834  */
835 static void shrink_readahead_size_eio(struct file *filp,
836                                         struct file_ra_state *ra)
837 {
838         if (!ra->ra_pages)
839                 return;
840
841         ra->ra_pages /= 4;
842 }
843
844 /**
845  * do_generic_mapping_read - generic file read routine
846  * @mapping:    address_space to be read
847  * @_ra:        file's readahead state
848  * @filp:       the file to read
849  * @ppos:       current file position
850  * @desc:       read_descriptor
851  * @actor:      read method
852  *
853  * This is a generic file read routine, and uses the
854  * mapping->a_ops->readpage() function for the actual low-level stuff.
855  *
856  * This is really ugly. But the goto's actually try to clarify some
857  * of the logic when it comes to error handling etc.
858  *
859  * Note the struct file* is only passed for the use of readpage.
860  * It may be NULL.
861  */
862 void do_generic_mapping_read(struct address_space *mapping,
863                              struct file_ra_state *ra,
864                              struct file *filp,
865                              loff_t *ppos,
866                              read_descriptor_t *desc,
867                              read_actor_t actor)
868 {
869         struct inode *inode = mapping->host;
870         pgoff_t index;
871         pgoff_t last_index;
872         pgoff_t prev_index;
873         unsigned long offset;      /* offset into pagecache page */
874         unsigned int prev_offset;
875         struct page *cached_page;
876         int error;
877
878         cached_page = NULL;
879         index = *ppos >> PAGE_CACHE_SHIFT;
880         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
881         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
882         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
883         offset = *ppos & ~PAGE_CACHE_MASK;
884
885         for (;;) {
886                 struct page *page;
887                 pgoff_t end_index;
888                 loff_t isize;
889                 unsigned long nr, ret;
890
891                 cond_resched();
892 find_page:
893                 page = find_get_page(mapping, index);
894                 if (!page) {
895                         page_cache_sync_readahead(mapping,
896                                         ra, filp,
897                                         index, last_index - index);
898                         page = find_get_page(mapping, index);
899                         if (unlikely(page == NULL))
900                                 goto no_cached_page;
901                 }
902                 if (PageReadahead(page)) {
903                         page_cache_async_readahead(mapping,
904                                         ra, filp, page,
905                                         index, last_index - index);
906                 }
907                 if (!PageUptodate(page))
908                         goto page_not_up_to_date;
909 page_ok:
910                 /*
911                  * i_size must be checked after we know the page is Uptodate.
912                  *
913                  * Checking i_size after the check allows us to calculate
914                  * the correct value for "nr", which means the zero-filled
915                  * part of the page is not copied back to userspace (unless
916                  * another truncate extends the file - this is desired though).
917                  */
918
919                 isize = i_size_read(inode);
920                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
921                 if (unlikely(!isize || index > end_index)) {
922                         page_cache_release(page);
923                         goto out;
924                 }
925
926                 /* nr is the maximum number of bytes to copy from this page */
927                 nr = PAGE_CACHE_SIZE;
928                 if (index == end_index) {
929                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
930                         if (nr <= offset) {
931                                 page_cache_release(page);
932                                 goto out;
933                         }
934                 }
935                 nr = nr - offset;
936
937                 /* If users can be writing to this page using arbitrary
938                  * virtual addresses, take care about potential aliasing
939                  * before reading the page on the kernel side.
940                  */
941                 if (mapping_writably_mapped(mapping))
942                         flush_dcache_page(page);
943
944                 /*
945                  * When a sequential read accesses a page several times,
946                  * only mark it as accessed the first time.
947                  */
948                 if (prev_index != index || offset != prev_offset)
949                         mark_page_accessed(page);
950                 prev_index = index;
951
952                 /*
953                  * Ok, we have the page, and it's up-to-date, so
954                  * now we can copy it to user space...
955                  *
956                  * The actor routine returns how many bytes were actually used..
957                  * NOTE! This may not be the same as how much of a user buffer
958                  * we filled up (we may be padding etc), so we can only update
959                  * "pos" here (the actor routine has to update the user buffer
960                  * pointers and the remaining count).
961                  */
962                 ret = actor(desc, page, offset, nr);
963                 offset += ret;
964                 index += offset >> PAGE_CACHE_SHIFT;
965                 offset &= ~PAGE_CACHE_MASK;
966                 prev_offset = offset;
967
968                 page_cache_release(page);
969                 if (ret == nr && desc->count)
970                         continue;
971                 goto out;
972
973 page_not_up_to_date:
974                 /* Get exclusive access to the page ... */
975                 lock_page(page);
976
977                 /* Did it get truncated before we got the lock? */
978                 if (!page->mapping) {
979                         unlock_page(page);
980                         page_cache_release(page);
981                         continue;
982                 }
983
984                 /* Did somebody else fill it already? */
985                 if (PageUptodate(page)) {
986                         unlock_page(page);
987                         goto page_ok;
988                 }
989
990 readpage:
991                 /* Start the actual read. The read will unlock the page. */
992                 error = mapping->a_ops->readpage(filp, page);
993
994                 if (unlikely(error)) {
995                         if (error == AOP_TRUNCATED_PAGE) {
996                                 page_cache_release(page);
997                                 goto find_page;
998                         }
999                         goto readpage_error;
1000                 }
1001
1002                 if (!PageUptodate(page)) {
1003                         lock_page(page);
1004                         if (!PageUptodate(page)) {
1005                                 if (page->mapping == NULL) {
1006                                         /*
1007                                          * invalidate_inode_pages got it
1008                                          */
1009                                         unlock_page(page);
1010                                         page_cache_release(page);
1011                                         goto find_page;
1012                                 }
1013                                 unlock_page(page);
1014                                 error = -EIO;
1015                                 shrink_readahead_size_eio(filp, ra);
1016                                 goto readpage_error;
1017                         }
1018                         unlock_page(page);
1019                 }
1020
1021                 goto page_ok;
1022
1023 readpage_error:
1024                 /* UHHUH! A synchronous read error occurred. Report it */
1025                 desc->error = error;
1026                 page_cache_release(page);
1027                 goto out;
1028
1029 no_cached_page:
1030                 /*
1031                  * Ok, it wasn't cached, so we need to create a new
1032                  * page..
1033                  */
1034                 if (!cached_page) {
1035                         cached_page = page_cache_alloc_cold(mapping);
1036                         if (!cached_page) {
1037                                 desc->error = -ENOMEM;
1038                                 goto out;
1039                         }
1040                 }
1041                 error = add_to_page_cache_lru(cached_page, mapping,
1042                                                 index, GFP_KERNEL);
1043                 if (error) {
1044                         if (error == -EEXIST)
1045                                 goto find_page;
1046                         desc->error = error;
1047                         goto out;
1048                 }
1049                 page = cached_page;
1050                 cached_page = NULL;
1051                 goto readpage;
1052         }
1053
1054 out:
1055         ra->prev_pos = prev_index;
1056         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1057         ra->prev_pos |= prev_offset;
1058
1059         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1060         if (cached_page)
1061                 page_cache_release(cached_page);
1062         if (filp)
1063                 file_accessed(filp);
1064 }
1065 EXPORT_SYMBOL(do_generic_mapping_read);
1066
1067 int file_read_actor(read_descriptor_t *desc, struct page *page,
1068                         unsigned long offset, unsigned long size)
1069 {
1070         char *kaddr;
1071         unsigned long left, count = desc->count;
1072
1073         if (size > count)
1074                 size = count;
1075
1076         /*
1077          * Faults on the destination of a read are common, so do it before
1078          * taking the kmap.
1079          */
1080         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1081                 kaddr = kmap_atomic(page, KM_USER0);
1082                 left = __copy_to_user_inatomic(desc->arg.buf,
1083                                                 kaddr + offset, size);
1084                 kunmap_atomic(kaddr, KM_USER0);
1085                 if (left == 0)
1086                         goto success;
1087         }
1088
1089         /* Do it the slow way */
1090         kaddr = kmap(page);
1091         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1092         kunmap(page);
1093
1094         if (left) {
1095                 size -= left;
1096                 desc->error = -EFAULT;
1097         }
1098 success:
1099         desc->count = count - size;
1100         desc->written += size;
1101         desc->arg.buf += size;
1102         return size;
1103 }
1104
1105 /*
1106  * Performs necessary checks before doing a write
1107  * @iov:        io vector request
1108  * @nr_segs:    number of segments in the iovec
1109  * @count:      number of bytes to write
1110  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1111  *
1112  * Adjust number of segments and amount of bytes to write (nr_segs should be
1113  * properly initialized first). Returns appropriate error code that caller
1114  * should return or zero in case that write should be allowed.
1115  */
1116 int generic_segment_checks(const struct iovec *iov,
1117                         unsigned long *nr_segs, size_t *count, int access_flags)
1118 {
1119         unsigned long   seg;
1120         size_t cnt = 0;
1121         for (seg = 0; seg < *nr_segs; seg++) {
1122                 const struct iovec *iv = &iov[seg];
1123
1124                 /*
1125                  * If any segment has a negative length, or the cumulative
1126                  * length ever wraps negative then return -EINVAL.
1127                  */
1128                 cnt += iv->iov_len;
1129                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1130                         return -EINVAL;
1131                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1132                         continue;
1133                 if (seg == 0)
1134                         return -EFAULT;
1135                 *nr_segs = seg;
1136                 cnt -= iv->iov_len;     /* This segment is no good */
1137                 break;
1138         }
1139         *count = cnt;
1140         return 0;
1141 }
1142 EXPORT_SYMBOL(generic_segment_checks);
1143
1144 /**
1145  * generic_file_aio_read - generic filesystem read routine
1146  * @iocb:       kernel I/O control block
1147  * @iov:        io vector request
1148  * @nr_segs:    number of segments in the iovec
1149  * @pos:        current file position
1150  *
1151  * This is the "read()" routine for all filesystems
1152  * that can use the page cache directly.
1153  */
1154 ssize_t
1155 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1156                 unsigned long nr_segs, loff_t pos)
1157 {
1158         struct file *filp = iocb->ki_filp;
1159         ssize_t retval;
1160         unsigned long seg;
1161         size_t count;
1162         loff_t *ppos = &iocb->ki_pos;
1163
1164         count = 0;
1165         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1166         if (retval)
1167                 return retval;
1168
1169         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1170         if (filp->f_flags & O_DIRECT) {
1171                 loff_t size;
1172                 struct address_space *mapping;
1173                 struct inode *inode;
1174
1175                 mapping = filp->f_mapping;
1176                 inode = mapping->host;
1177                 retval = 0;
1178                 if (!count)
1179                         goto out; /* skip atime */
1180                 size = i_size_read(inode);
1181                 if (pos < size) {
1182                         retval = generic_file_direct_IO(READ, iocb,
1183                                                 iov, pos, nr_segs);
1184                         if (retval > 0)
1185                                 *ppos = pos + retval;
1186                 }
1187                 if (likely(retval != 0)) {
1188                         file_accessed(filp);
1189                         goto out;
1190                 }
1191         }
1192
1193         retval = 0;
1194         if (count) {
1195                 for (seg = 0; seg < nr_segs; seg++) {
1196                         read_descriptor_t desc;
1197
1198                         desc.written = 0;
1199                         desc.arg.buf = iov[seg].iov_base;
1200                         desc.count = iov[seg].iov_len;
1201                         if (desc.count == 0)
1202                                 continue;
1203                         desc.error = 0;
1204                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1205                         retval += desc.written;
1206                         if (desc.error) {
1207                                 retval = retval ?: desc.error;
1208                                 break;
1209                         }
1210                         if (desc.count > 0)
1211                                 break;
1212                 }
1213         }
1214 out:
1215         return retval;
1216 }
1217 EXPORT_SYMBOL(generic_file_aio_read);
1218
1219 static ssize_t
1220 do_readahead(struct address_space *mapping, struct file *filp,
1221              pgoff_t index, unsigned long nr)
1222 {
1223         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1224                 return -EINVAL;
1225
1226         force_page_cache_readahead(mapping, filp, index,
1227                                         max_sane_readahead(nr));
1228         return 0;
1229 }
1230
1231 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1232 {
1233         ssize_t ret;
1234         struct file *file;
1235
1236         ret = -EBADF;
1237         file = fget(fd);
1238         if (file) {
1239                 if (file->f_mode & FMODE_READ) {
1240                         struct address_space *mapping = file->f_mapping;
1241                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1242                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1243                         unsigned long len = end - start + 1;
1244                         ret = do_readahead(mapping, file, start, len);
1245                 }
1246                 fput(file);
1247         }
1248         return ret;
1249 }
1250
1251 #ifdef CONFIG_MMU
1252 /**
1253  * page_cache_read - adds requested page to the page cache if not already there
1254  * @file:       file to read
1255  * @offset:     page index
1256  *
1257  * This adds the requested page to the page cache if it isn't already there,
1258  * and schedules an I/O to read in its contents from disk.
1259  */
1260 static int fastcall page_cache_read(struct file * file, pgoff_t offset)
1261 {
1262         struct address_space *mapping = file->f_mapping;
1263         struct page *page; 
1264         int ret;
1265
1266         do {
1267                 page = page_cache_alloc_cold(mapping);
1268                 if (!page)
1269                         return -ENOMEM;
1270
1271                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1272                 if (ret == 0)
1273                         ret = mapping->a_ops->readpage(file, page);
1274                 else if (ret == -EEXIST)
1275                         ret = 0; /* losing race to add is OK */
1276
1277                 page_cache_release(page);
1278
1279         } while (ret == AOP_TRUNCATED_PAGE);
1280                 
1281         return ret;
1282 }
1283
1284 #define MMAP_LOTSAMISS  (100)
1285
1286 /**
1287  * filemap_fault - read in file data for page fault handling
1288  * @vma:        vma in which the fault was taken
1289  * @vmf:        struct vm_fault containing details of the fault
1290  *
1291  * filemap_fault() is invoked via the vma operations vector for a
1292  * mapped memory region to read in file data during a page fault.
1293  *
1294  * The goto's are kind of ugly, but this streamlines the normal case of having
1295  * it in the page cache, and handles the special cases reasonably without
1296  * having a lot of duplicated code.
1297  */
1298 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1299 {
1300         int error;
1301         struct file *file = vma->vm_file;
1302         struct address_space *mapping = file->f_mapping;
1303         struct file_ra_state *ra = &file->f_ra;
1304         struct inode *inode = mapping->host;
1305         struct page *page;
1306         unsigned long size;
1307         int did_readaround = 0;
1308         int ret = 0;
1309
1310         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1311         if (vmf->pgoff >= size)
1312                 goto outside_data_content;
1313
1314         /* If we don't want any read-ahead, don't bother */
1315         if (VM_RandomReadHint(vma))
1316                 goto no_cached_page;
1317
1318         /*
1319          * Do we have something in the page cache already?
1320          */
1321 retry_find:
1322         page = find_lock_page(mapping, vmf->pgoff);
1323         /*
1324          * For sequential accesses, we use the generic readahead logic.
1325          */
1326         if (VM_SequentialReadHint(vma)) {
1327                 if (!page) {
1328                         page_cache_sync_readahead(mapping, ra, file,
1329                                                            vmf->pgoff, 1);
1330                         page = find_lock_page(mapping, vmf->pgoff);
1331                         if (!page)
1332                                 goto no_cached_page;
1333                 }
1334                 if (PageReadahead(page)) {
1335                         page_cache_async_readahead(mapping, ra, file, page,
1336                                                            vmf->pgoff, 1);
1337                 }
1338         }
1339
1340         if (!page) {
1341                 unsigned long ra_pages;
1342
1343                 ra->mmap_miss++;
1344
1345                 /*
1346                  * Do we miss much more than hit in this file? If so,
1347                  * stop bothering with read-ahead. It will only hurt.
1348                  */
1349                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1350                         goto no_cached_page;
1351
1352                 /*
1353                  * To keep the pgmajfault counter straight, we need to
1354                  * check did_readaround, as this is an inner loop.
1355                  */
1356                 if (!did_readaround) {
1357                         ret = VM_FAULT_MAJOR;
1358                         count_vm_event(PGMAJFAULT);
1359                 }
1360                 did_readaround = 1;
1361                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1362                 if (ra_pages) {
1363                         pgoff_t start = 0;
1364
1365                         if (vmf->pgoff > ra_pages / 2)
1366                                 start = vmf->pgoff - ra_pages / 2;
1367                         do_page_cache_readahead(mapping, file, start, ra_pages);
1368                 }
1369                 page = find_lock_page(mapping, vmf->pgoff);
1370                 if (!page)
1371                         goto no_cached_page;
1372         }
1373
1374         if (!did_readaround)
1375                 ra->mmap_miss--;
1376
1377         /*
1378          * We have a locked page in the page cache, now we need to check
1379          * that it's up-to-date. If not, it is going to be due to an error.
1380          */
1381         if (unlikely(!PageUptodate(page)))
1382                 goto page_not_uptodate;
1383
1384         /* Must recheck i_size under page lock */
1385         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1386         if (unlikely(vmf->pgoff >= size)) {
1387                 unlock_page(page);
1388                 page_cache_release(page);
1389                 goto outside_data_content;
1390         }
1391
1392         /*
1393          * Found the page and have a reference on it.
1394          */
1395         mark_page_accessed(page);
1396         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1397         vmf->page = page;
1398         return ret | VM_FAULT_LOCKED;
1399
1400 outside_data_content:
1401         /*
1402          * An external ptracer can access pages that normally aren't
1403          * accessible..
1404          */
1405         if (vma->vm_mm == current->mm)
1406                 return VM_FAULT_SIGBUS;
1407
1408         /* Fall through to the non-read-ahead case */
1409 no_cached_page:
1410         /*
1411          * We're only likely to ever get here if MADV_RANDOM is in
1412          * effect.
1413          */
1414         error = page_cache_read(file, vmf->pgoff);
1415
1416         /*
1417          * The page we want has now been added to the page cache.
1418          * In the unlikely event that someone removed it in the
1419          * meantime, we'll just come back here and read it again.
1420          */
1421         if (error >= 0)
1422                 goto retry_find;
1423
1424         /*
1425          * An error return from page_cache_read can result if the
1426          * system is low on memory, or a problem occurs while trying
1427          * to schedule I/O.
1428          */
1429         if (error == -ENOMEM)
1430                 return VM_FAULT_OOM;
1431         return VM_FAULT_SIGBUS;
1432
1433 page_not_uptodate:
1434         /* IO error path */
1435         if (!did_readaround) {
1436                 ret = VM_FAULT_MAJOR;
1437                 count_vm_event(PGMAJFAULT);
1438         }
1439
1440         /*
1441          * Umm, take care of errors if the page isn't up-to-date.
1442          * Try to re-read it _once_. We do this synchronously,
1443          * because there really aren't any performance issues here
1444          * and we need to check for errors.
1445          */
1446         ClearPageError(page);
1447         error = mapping->a_ops->readpage(file, page);
1448         page_cache_release(page);
1449
1450         if (!error || error == AOP_TRUNCATED_PAGE)
1451                 goto retry_find;
1452
1453         /* Things didn't work out. Return zero to tell the mm layer so. */
1454         shrink_readahead_size_eio(file, ra);
1455         return VM_FAULT_SIGBUS;
1456 }
1457 EXPORT_SYMBOL(filemap_fault);
1458
1459 struct vm_operations_struct generic_file_vm_ops = {
1460         .fault          = filemap_fault,
1461 };
1462
1463 /* This is used for a general mmap of a disk file */
1464
1465 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1466 {
1467         struct address_space *mapping = file->f_mapping;
1468
1469         if (!mapping->a_ops->readpage)
1470                 return -ENOEXEC;
1471         file_accessed(file);
1472         vma->vm_ops = &generic_file_vm_ops;
1473         vma->vm_flags |= VM_CAN_NONLINEAR;
1474         return 0;
1475 }
1476
1477 /*
1478  * This is for filesystems which do not implement ->writepage.
1479  */
1480 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1481 {
1482         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1483                 return -EINVAL;
1484         return generic_file_mmap(file, vma);
1485 }
1486 #else
1487 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1488 {
1489         return -ENOSYS;
1490 }
1491 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1492 {
1493         return -ENOSYS;
1494 }
1495 #endif /* CONFIG_MMU */
1496
1497 EXPORT_SYMBOL(generic_file_mmap);
1498 EXPORT_SYMBOL(generic_file_readonly_mmap);
1499
1500 static struct page *__read_cache_page(struct address_space *mapping,
1501                                 pgoff_t index,
1502                                 int (*filler)(void *,struct page*),
1503                                 void *data)
1504 {
1505         struct page *page, *cached_page = NULL;
1506         int err;
1507 repeat:
1508         page = find_get_page(mapping, index);
1509         if (!page) {
1510                 if (!cached_page) {
1511                         cached_page = page_cache_alloc_cold(mapping);
1512                         if (!cached_page)
1513                                 return ERR_PTR(-ENOMEM);
1514                 }
1515                 err = add_to_page_cache_lru(cached_page, mapping,
1516                                         index, GFP_KERNEL);
1517                 if (err == -EEXIST)
1518                         goto repeat;
1519                 if (err < 0) {
1520                         /* Presumably ENOMEM for radix tree node */
1521                         page_cache_release(cached_page);
1522                         return ERR_PTR(err);
1523                 }
1524                 page = cached_page;
1525                 cached_page = NULL;
1526                 err = filler(data, page);
1527                 if (err < 0) {
1528                         page_cache_release(page);
1529                         page = ERR_PTR(err);
1530                 }
1531         }
1532         if (cached_page)
1533                 page_cache_release(cached_page);
1534         return page;
1535 }
1536
1537 /*
1538  * Same as read_cache_page, but don't wait for page to become unlocked
1539  * after submitting it to the filler.
1540  */
1541 struct page *read_cache_page_async(struct address_space *mapping,
1542                                 pgoff_t index,
1543                                 int (*filler)(void *,struct page*),
1544                                 void *data)
1545 {
1546         struct page *page;
1547         int err;
1548
1549 retry:
1550         page = __read_cache_page(mapping, index, filler, data);
1551         if (IS_ERR(page))
1552                 return page;
1553         if (PageUptodate(page))
1554                 goto out;
1555
1556         lock_page(page);
1557         if (!page->mapping) {
1558                 unlock_page(page);
1559                 page_cache_release(page);
1560                 goto retry;
1561         }
1562         if (PageUptodate(page)) {
1563                 unlock_page(page);
1564                 goto out;
1565         }
1566         err = filler(data, page);
1567         if (err < 0) {
1568                 page_cache_release(page);
1569                 return ERR_PTR(err);
1570         }
1571 out:
1572         mark_page_accessed(page);
1573         return page;
1574 }
1575 EXPORT_SYMBOL(read_cache_page_async);
1576
1577 /**
1578  * read_cache_page - read into page cache, fill it if needed
1579  * @mapping:    the page's address_space
1580  * @index:      the page index
1581  * @filler:     function to perform the read
1582  * @data:       destination for read data
1583  *
1584  * Read into the page cache. If a page already exists, and PageUptodate() is
1585  * not set, try to fill the page then wait for it to become unlocked.
1586  *
1587  * If the page does not get brought uptodate, return -EIO.
1588  */
1589 struct page *read_cache_page(struct address_space *mapping,
1590                                 pgoff_t index,
1591                                 int (*filler)(void *,struct page*),
1592                                 void *data)
1593 {
1594         struct page *page;
1595
1596         page = read_cache_page_async(mapping, index, filler, data);
1597         if (IS_ERR(page))
1598                 goto out;
1599         wait_on_page_locked(page);
1600         if (!PageUptodate(page)) {
1601                 page_cache_release(page);
1602                 page = ERR_PTR(-EIO);
1603         }
1604  out:
1605         return page;
1606 }
1607 EXPORT_SYMBOL(read_cache_page);
1608
1609 /*
1610  * If the page was newly created, increment its refcount and add it to the
1611  * caller's lru-buffering pagevec.  This function is specifically for
1612  * generic_file_write().
1613  */
1614 static inline struct page *
1615 __grab_cache_page(struct address_space *mapping, unsigned long index,
1616                         struct page **cached_page, struct pagevec *lru_pvec)
1617 {
1618         int err;
1619         struct page *page;
1620 repeat:
1621         page = find_lock_page(mapping, index);
1622         if (!page) {
1623                 if (!*cached_page) {
1624                         *cached_page = page_cache_alloc(mapping);
1625                         if (!*cached_page)
1626                                 return NULL;
1627                 }
1628                 err = add_to_page_cache(*cached_page, mapping,
1629                                         index, GFP_KERNEL);
1630                 if (err == -EEXIST)
1631                         goto repeat;
1632                 if (err == 0) {
1633                         page = *cached_page;
1634                         page_cache_get(page);
1635                         if (!pagevec_add(lru_pvec, page))
1636                                 __pagevec_lru_add(lru_pvec);
1637                         *cached_page = NULL;
1638                 }
1639         }
1640         return page;
1641 }
1642
1643 /*
1644  * The logic we want is
1645  *
1646  *      if suid or (sgid and xgrp)
1647  *              remove privs
1648  */
1649 int should_remove_suid(struct dentry *dentry)
1650 {
1651         mode_t mode = dentry->d_inode->i_mode;
1652         int kill = 0;
1653
1654         /* suid always must be killed */
1655         if (unlikely(mode & S_ISUID))
1656                 kill = ATTR_KILL_SUID;
1657
1658         /*
1659          * sgid without any exec bits is just a mandatory locking mark; leave
1660          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1661          */
1662         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1663                 kill |= ATTR_KILL_SGID;
1664
1665         if (unlikely(kill && !capable(CAP_FSETID)))
1666                 return kill;
1667
1668         return 0;
1669 }
1670 EXPORT_SYMBOL(should_remove_suid);
1671
1672 int __remove_suid(struct dentry *dentry, int kill)
1673 {
1674         struct iattr newattrs;
1675
1676         newattrs.ia_valid = ATTR_FORCE | kill;
1677         return notify_change(dentry, &newattrs);
1678 }
1679
1680 int remove_suid(struct dentry *dentry)
1681 {
1682         int kill = should_remove_suid(dentry);
1683
1684         if (unlikely(kill))
1685                 return __remove_suid(dentry, kill);
1686
1687         return 0;
1688 }
1689 EXPORT_SYMBOL(remove_suid);
1690
1691 size_t
1692 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1693                         const struct iovec *iov, size_t base, size_t bytes)
1694 {
1695         size_t copied = 0, left = 0;
1696
1697         while (bytes) {
1698                 char __user *buf = iov->iov_base + base;
1699                 int copy = min(bytes, iov->iov_len - base);
1700
1701                 base = 0;
1702                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1703                 copied += copy;
1704                 bytes -= copy;
1705                 vaddr += copy;
1706                 iov++;
1707
1708                 if (unlikely(left))
1709                         break;
1710         }
1711         return copied - left;
1712 }
1713
1714 /*
1715  * Performs necessary checks before doing a write
1716  *
1717  * Can adjust writing position or amount of bytes to write.
1718  * Returns appropriate error code that caller should return or
1719  * zero in case that write should be allowed.
1720  */
1721 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1722 {
1723         struct inode *inode = file->f_mapping->host;
1724         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1725
1726         if (unlikely(*pos < 0))
1727                 return -EINVAL;
1728
1729         if (!isblk) {
1730                 /* FIXME: this is for backwards compatibility with 2.4 */
1731                 if (file->f_flags & O_APPEND)
1732                         *pos = i_size_read(inode);
1733
1734                 if (limit != RLIM_INFINITY) {
1735                         if (*pos >= limit) {
1736                                 send_sig(SIGXFSZ, current, 0);
1737                                 return -EFBIG;
1738                         }
1739                         if (*count > limit - (typeof(limit))*pos) {
1740                                 *count = limit - (typeof(limit))*pos;
1741                         }
1742                 }
1743         }
1744
1745         /*
1746          * LFS rule
1747          */
1748         if (unlikely(*pos + *count > MAX_NON_LFS &&
1749                                 !(file->f_flags & O_LARGEFILE))) {
1750                 if (*pos >= MAX_NON_LFS) {
1751                         return -EFBIG;
1752                 }
1753                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1754                         *count = MAX_NON_LFS - (unsigned long)*pos;
1755                 }
1756         }
1757
1758         /*
1759          * Are we about to exceed the fs block limit ?
1760          *
1761          * If we have written data it becomes a short write.  If we have
1762          * exceeded without writing data we send a signal and return EFBIG.
1763          * Linus frestrict idea will clean these up nicely..
1764          */
1765         if (likely(!isblk)) {
1766                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1767                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1768                                 return -EFBIG;
1769                         }
1770                         /* zero-length writes at ->s_maxbytes are OK */
1771                 }
1772
1773                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1774                         *count = inode->i_sb->s_maxbytes - *pos;
1775         } else {
1776 #ifdef CONFIG_BLOCK
1777                 loff_t isize;
1778                 if (bdev_read_only(I_BDEV(inode)))
1779                         return -EPERM;
1780                 isize = i_size_read(inode);
1781                 if (*pos >= isize) {
1782                         if (*count || *pos > isize)
1783                                 return -ENOSPC;
1784                 }
1785
1786                 if (*pos + *count > isize)
1787                         *count = isize - *pos;
1788 #else
1789                 return -EPERM;
1790 #endif
1791         }
1792         return 0;
1793 }
1794 EXPORT_SYMBOL(generic_write_checks);
1795
1796 ssize_t
1797 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1798                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1799                 size_t count, size_t ocount)
1800 {
1801         struct file     *file = iocb->ki_filp;
1802         struct address_space *mapping = file->f_mapping;
1803         struct inode    *inode = mapping->host;
1804         ssize_t         written;
1805
1806         if (count != ocount)
1807                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1808
1809         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1810         if (written > 0) {
1811                 loff_t end = pos + written;
1812                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1813                         i_size_write(inode,  end);
1814                         mark_inode_dirty(inode);
1815                 }
1816                 *ppos = end;
1817         }
1818
1819         /*
1820          * Sync the fs metadata but not the minor inode changes and
1821          * of course not the data as we did direct DMA for the IO.
1822          * i_mutex is held, which protects generic_osync_inode() from
1823          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
1824          */
1825         if ((written >= 0 || written == -EIOCBQUEUED) &&
1826             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1827                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1828                 if (err < 0)
1829                         written = err;
1830         }
1831         return written;
1832 }
1833 EXPORT_SYMBOL(generic_file_direct_write);
1834
1835 ssize_t
1836 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1837                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1838                 size_t count, ssize_t written)
1839 {
1840         struct file *file = iocb->ki_filp;
1841         struct address_space * mapping = file->f_mapping;
1842         const struct address_space_operations *a_ops = mapping->a_ops;
1843         struct inode    *inode = mapping->host;
1844         long            status = 0;
1845         struct page     *page;
1846         struct page     *cached_page = NULL;
1847         size_t          bytes;
1848         struct pagevec  lru_pvec;
1849         const struct iovec *cur_iov = iov; /* current iovec */
1850         size_t          iov_base = 0;      /* offset in the current iovec */
1851         char __user     *buf;
1852
1853         pagevec_init(&lru_pvec, 0);
1854
1855         /*
1856          * handle partial DIO write.  Adjust cur_iov if needed.
1857          */
1858         if (likely(nr_segs == 1))
1859                 buf = iov->iov_base + written;
1860         else {
1861                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1862                 buf = cur_iov->iov_base + iov_base;
1863         }
1864
1865         do {
1866                 unsigned long index;
1867                 unsigned long offset;
1868                 size_t copied;
1869
1870                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1871                 index = pos >> PAGE_CACHE_SHIFT;
1872                 bytes = PAGE_CACHE_SIZE - offset;
1873
1874                 /* Limit the size of the copy to the caller's write size */
1875                 bytes = min(bytes, count);
1876
1877                 /*
1878                  * Limit the size of the copy to that of the current segment,
1879                  * because fault_in_pages_readable() doesn't know how to walk
1880                  * segments.
1881                  */
1882                 bytes = min(bytes, cur_iov->iov_len - iov_base);
1883
1884                 /*
1885                  * Bring in the user page that we will copy from _first_.
1886                  * Otherwise there's a nasty deadlock on copying from the
1887                  * same page as we're writing to, without it being marked
1888                  * up-to-date.
1889                  */
1890                 fault_in_pages_readable(buf, bytes);
1891
1892                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1893                 if (!page) {
1894                         status = -ENOMEM;
1895                         break;
1896                 }
1897
1898                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1899                 if (unlikely(status)) {
1900                         loff_t isize = i_size_read(inode);
1901
1902                         if (status != AOP_TRUNCATED_PAGE)
1903                                 unlock_page(page);
1904                         page_cache_release(page);
1905                         if (status == AOP_TRUNCATED_PAGE)
1906                                 continue;
1907                         /*
1908                          * prepare_write() may have instantiated a few blocks
1909                          * outside i_size.  Trim these off again.
1910                          */
1911                         if (pos + bytes > isize)
1912                                 vmtruncate(inode, isize);
1913                         break;
1914                 }
1915                 if (likely(nr_segs == 1))
1916                         copied = filemap_copy_from_user(page, offset,
1917                                                         buf, bytes);
1918                 else
1919                         copied = filemap_copy_from_user_iovec(page, offset,
1920                                                 cur_iov, iov_base, bytes);
1921                 flush_dcache_page(page);
1922                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1923                 if (status == AOP_TRUNCATED_PAGE) {
1924                         page_cache_release(page);
1925                         continue;
1926                 }
1927                 if (likely(copied > 0)) {
1928                         if (!status)
1929                                 status = copied;
1930
1931                         if (status >= 0) {
1932                                 written += status;
1933                                 count -= status;
1934                                 pos += status;
1935                                 buf += status;
1936                                 if (unlikely(nr_segs > 1)) {
1937                                         filemap_set_next_iovec(&cur_iov,
1938                                                         &iov_base, status);
1939                                         if (count)
1940                                                 buf = cur_iov->iov_base +
1941                                                         iov_base;
1942                                 } else {
1943                                         iov_base += status;
1944                                 }
1945                         }
1946                 }
1947                 if (unlikely(copied != bytes))
1948                         if (status >= 0)
1949                                 status = -EFAULT;
1950                 unlock_page(page);
1951                 mark_page_accessed(page);
1952                 page_cache_release(page);
1953                 if (status < 0)
1954                         break;
1955                 balance_dirty_pages_ratelimited(mapping);
1956                 cond_resched();
1957         } while (count);
1958         *ppos = pos;
1959
1960         if (cached_page)
1961                 page_cache_release(cached_page);
1962
1963         /*
1964          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
1965          */
1966         if (likely(status >= 0)) {
1967                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1968                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
1969                                 status = generic_osync_inode(inode, mapping,
1970                                                 OSYNC_METADATA|OSYNC_DATA);
1971                 }
1972         }
1973         
1974         /*
1975          * If we get here for O_DIRECT writes then we must have fallen through
1976          * to buffered writes (block instantiation inside i_size).  So we sync
1977          * the file data here, to try to honour O_DIRECT expectations.
1978          */
1979         if (unlikely(file->f_flags & O_DIRECT) && written)
1980                 status = filemap_write_and_wait(mapping);
1981
1982         pagevec_lru_add(&lru_pvec);
1983         return written ? written : status;
1984 }
1985 EXPORT_SYMBOL(generic_file_buffered_write);
1986
1987 static ssize_t
1988 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
1989                                 unsigned long nr_segs, loff_t *ppos)
1990 {
1991         struct file *file = iocb->ki_filp;
1992         struct address_space * mapping = file->f_mapping;
1993         size_t ocount;          /* original count */
1994         size_t count;           /* after file limit checks */
1995         struct inode    *inode = mapping->host;
1996         loff_t          pos;
1997         ssize_t         written;
1998         ssize_t         err;
1999
2000         ocount = 0;
2001         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2002         if (err)
2003                 return err;
2004
2005         count = ocount;
2006         pos = *ppos;
2007
2008         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2009
2010         /* We can write back this queue in page reclaim */
2011         current->backing_dev_info = mapping->backing_dev_info;
2012         written = 0;
2013
2014         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2015         if (err)
2016                 goto out;
2017
2018         if (count == 0)
2019                 goto out;
2020
2021         err = remove_suid(file->f_path.dentry);
2022         if (err)
2023                 goto out;
2024
2025         file_update_time(file);
2026
2027         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2028         if (unlikely(file->f_flags & O_DIRECT)) {
2029                 loff_t endbyte;
2030                 ssize_t written_buffered;
2031
2032                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2033                                                         ppos, count, ocount);
2034                 if (written < 0 || written == count)
2035                         goto out;
2036                 /*
2037                  * direct-io write to a hole: fall through to buffered I/O
2038                  * for completing the rest of the request.
2039                  */
2040                 pos += written;
2041                 count -= written;
2042                 written_buffered = generic_file_buffered_write(iocb, iov,
2043                                                 nr_segs, pos, ppos, count,
2044                                                 written);
2045                 /*
2046                  * If generic_file_buffered_write() retuned a synchronous error
2047                  * then we want to return the number of bytes which were
2048                  * direct-written, or the error code if that was zero.  Note
2049                  * that this differs from normal direct-io semantics, which
2050                  * will return -EFOO even if some bytes were written.
2051                  */
2052                 if (written_buffered < 0) {
2053                         err = written_buffered;
2054                         goto out;
2055                 }
2056
2057                 /*
2058                  * We need to ensure that the page cache pages are written to
2059                  * disk and invalidated to preserve the expected O_DIRECT
2060                  * semantics.
2061                  */
2062                 endbyte = pos + written_buffered - written - 1;
2063                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2064                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2065                                             SYNC_FILE_RANGE_WRITE|
2066                                             SYNC_FILE_RANGE_WAIT_AFTER);
2067                 if (err == 0) {
2068                         written = written_buffered;
2069                         invalidate_mapping_pages(mapping,
2070                                                  pos >> PAGE_CACHE_SHIFT,
2071                                                  endbyte >> PAGE_CACHE_SHIFT);
2072                 } else {
2073                         /*
2074                          * We don't know how much we wrote, so just return
2075                          * the number of bytes which were direct-written
2076                          */
2077                 }
2078         } else {
2079                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2080                                 pos, ppos, count, written);
2081         }
2082 out:
2083         current->backing_dev_info = NULL;
2084         return written ? written : err;
2085 }
2086
2087 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2088                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2089 {
2090         struct file *file = iocb->ki_filp;
2091         struct address_space *mapping = file->f_mapping;
2092         struct inode *inode = mapping->host;
2093         ssize_t ret;
2094
2095         BUG_ON(iocb->ki_pos != pos);
2096
2097         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2098                         &iocb->ki_pos);
2099
2100         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2101                 ssize_t err;
2102
2103                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2104                 if (err < 0)
2105                         ret = err;
2106         }
2107         return ret;
2108 }
2109 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2110
2111 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2112                 unsigned long nr_segs, loff_t pos)
2113 {
2114         struct file *file = iocb->ki_filp;
2115         struct address_space *mapping = file->f_mapping;
2116         struct inode *inode = mapping->host;
2117         ssize_t ret;
2118
2119         BUG_ON(iocb->ki_pos != pos);
2120
2121         mutex_lock(&inode->i_mutex);
2122         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2123                         &iocb->ki_pos);
2124         mutex_unlock(&inode->i_mutex);
2125
2126         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2127                 ssize_t err;
2128
2129                 err = sync_page_range(inode, mapping, pos, ret);
2130                 if (err < 0)
2131                         ret = err;
2132         }
2133         return ret;
2134 }
2135 EXPORT_SYMBOL(generic_file_aio_write);
2136
2137 /*
2138  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2139  * went wrong during pagecache shootdown.
2140  */
2141 static ssize_t
2142 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2143         loff_t offset, unsigned long nr_segs)
2144 {
2145         struct file *file = iocb->ki_filp;
2146         struct address_space *mapping = file->f_mapping;
2147         ssize_t retval;
2148         size_t write_len;
2149         pgoff_t end = 0; /* silence gcc */
2150
2151         /*
2152          * If it's a write, unmap all mmappings of the file up-front.  This
2153          * will cause any pte dirty bits to be propagated into the pageframes
2154          * for the subsequent filemap_write_and_wait().
2155          */
2156         if (rw == WRITE) {
2157                 write_len = iov_length(iov, nr_segs);
2158                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2159                 if (mapping_mapped(mapping))
2160                         unmap_mapping_range(mapping, offset, write_len, 0);
2161         }
2162
2163         retval = filemap_write_and_wait(mapping);
2164         if (retval)
2165                 goto out;
2166
2167         /*
2168          * After a write we want buffered reads to be sure to go to disk to get
2169          * the new data.  We invalidate clean cached page from the region we're
2170          * about to write.  We do this *before* the write so that we can return
2171          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2172          */
2173         if (rw == WRITE && mapping->nrpages) {
2174                 retval = invalidate_inode_pages2_range(mapping,
2175                                         offset >> PAGE_CACHE_SHIFT, end);
2176                 if (retval)
2177                         goto out;
2178         }
2179
2180         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2181         if (retval)
2182                 goto out;
2183
2184         /*
2185          * Finally, try again to invalidate clean pages which might have been
2186          * faulted in by get_user_pages() if the source of the write was an
2187          * mmap()ed region of the file we're writing.  That's a pretty crazy
2188          * thing to do, so we don't support it 100%.  If this invalidation
2189          * fails and we have -EIOCBQUEUED we ignore the failure.
2190          */
2191         if (rw == WRITE && mapping->nrpages) {
2192                 int err = invalidate_inode_pages2_range(mapping,
2193                                               offset >> PAGE_CACHE_SHIFT, end);
2194                 if (err && retval >= 0)
2195                         retval = err;
2196         }
2197 out:
2198         return retval;
2199 }
2200
2201 /**
2202  * try_to_release_page() - release old fs-specific metadata on a page
2203  *
2204  * @page: the page which the kernel is trying to free
2205  * @gfp_mask: memory allocation flags (and I/O mode)
2206  *
2207  * The address_space is to try to release any data against the page
2208  * (presumably at page->private).  If the release was successful, return `1'.
2209  * Otherwise return zero.
2210  *
2211  * The @gfp_mask argument specifies whether I/O may be performed to release
2212  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2213  *
2214  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2215  */
2216 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2217 {
2218         struct address_space * const mapping = page->mapping;
2219
2220         BUG_ON(!PageLocked(page));
2221         if (PageWriteback(page))
2222                 return 0;
2223
2224         if (mapping && mapping->a_ops->releasepage)
2225                 return mapping->a_ops->releasepage(page, gfp_mask);
2226         return try_to_free_buffers(page);
2227 }
2228
2229 EXPORT_SYMBOL(try_to_release_page);