5dc18d76e7032f390aa2b1686f3ee1d9664d1e2c
[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, unsigned long 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                                 unsigned long offset)
621 {
622         struct page *page;
623
624         read_lock_irq(&mapping->tree_lock);
625 repeat:
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                         read_lock_irq(&mapping->tree_lock);
633
634                         /* Has the page been truncated while we slept? */
635                         if (unlikely(page->mapping != mapping ||
636                                      page->index != offset)) {
637                                 unlock_page(page);
638                                 page_cache_release(page);
639                                 goto repeat;
640                         }
641                 }
642         }
643         read_unlock_irq(&mapping->tree_lock);
644         return page;
645 }
646 EXPORT_SYMBOL(find_lock_page);
647
648 /**
649  * find_or_create_page - locate or add a pagecache page
650  * @mapping: the page's address_space
651  * @index: the page's index into the mapping
652  * @gfp_mask: page allocation mode
653  *
654  * Locates a page in the pagecache.  If the page is not present, a new page
655  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
656  * LRU list.  The returned page is locked and has its reference count
657  * incremented.
658  *
659  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
660  * allocation!
661  *
662  * find_or_create_page() returns the desired page's address, or zero on
663  * memory exhaustion.
664  */
665 struct page *find_or_create_page(struct address_space *mapping,
666                 unsigned long index, gfp_t gfp_mask)
667 {
668         struct page *page, *cached_page = NULL;
669         int err;
670 repeat:
671         page = find_lock_page(mapping, index);
672         if (!page) {
673                 if (!cached_page) {
674                         cached_page =
675                                 __page_cache_alloc(gfp_mask);
676                         if (!cached_page)
677                                 return NULL;
678                 }
679                 err = add_to_page_cache_lru(cached_page, mapping,
680                                         index, gfp_mask);
681                 if (!err) {
682                         page = cached_page;
683                         cached_page = NULL;
684                 } else if (err == -EEXIST)
685                         goto repeat;
686         }
687         if (cached_page)
688                 page_cache_release(cached_page);
689         return page;
690 }
691 EXPORT_SYMBOL(find_or_create_page);
692
693 /**
694  * find_get_pages - gang pagecache lookup
695  * @mapping:    The address_space to search
696  * @start:      The starting page index
697  * @nr_pages:   The maximum number of pages
698  * @pages:      Where the resulting pages are placed
699  *
700  * find_get_pages() will search for and return a group of up to
701  * @nr_pages pages in the mapping.  The pages are placed at @pages.
702  * find_get_pages() takes a reference against the returned pages.
703  *
704  * The search returns a group of mapping-contiguous pages with ascending
705  * indexes.  There may be holes in the indices due to not-present pages.
706  *
707  * find_get_pages() returns the number of pages which were found.
708  */
709 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
710                             unsigned int nr_pages, struct page **pages)
711 {
712         unsigned int i;
713         unsigned int ret;
714
715         read_lock_irq(&mapping->tree_lock);
716         ret = radix_tree_gang_lookup(&mapping->page_tree,
717                                 (void **)pages, start, nr_pages);
718         for (i = 0; i < ret; i++)
719                 page_cache_get(pages[i]);
720         read_unlock_irq(&mapping->tree_lock);
721         return ret;
722 }
723
724 /**
725  * find_get_pages_contig - gang contiguous pagecache lookup
726  * @mapping:    The address_space to search
727  * @index:      The starting page index
728  * @nr_pages:   The maximum number of pages
729  * @pages:      Where the resulting pages are placed
730  *
731  * find_get_pages_contig() works exactly like find_get_pages(), except
732  * that the returned number of pages are guaranteed to be contiguous.
733  *
734  * find_get_pages_contig() returns the number of pages which were found.
735  */
736 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
737                                unsigned int nr_pages, struct page **pages)
738 {
739         unsigned int i;
740         unsigned int ret;
741
742         read_lock_irq(&mapping->tree_lock);
743         ret = radix_tree_gang_lookup(&mapping->page_tree,
744                                 (void **)pages, index, nr_pages);
745         for (i = 0; i < ret; i++) {
746                 if (pages[i]->mapping == NULL || pages[i]->index != index)
747                         break;
748
749                 page_cache_get(pages[i]);
750                 index++;
751         }
752         read_unlock_irq(&mapping->tree_lock);
753         return i;
754 }
755 EXPORT_SYMBOL(find_get_pages_contig);
756
757 /**
758  * find_get_pages_tag - find and return pages that match @tag
759  * @mapping:    the address_space to search
760  * @index:      the starting page index
761  * @tag:        the tag index
762  * @nr_pages:   the maximum number of pages
763  * @pages:      where the resulting pages are placed
764  *
765  * Like find_get_pages, except we only return pages which are tagged with
766  * @tag.   We update @index to index the next page for the traversal.
767  */
768 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
769                         int tag, unsigned int nr_pages, struct page **pages)
770 {
771         unsigned int i;
772         unsigned int ret;
773
774         read_lock_irq(&mapping->tree_lock);
775         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
776                                 (void **)pages, *index, nr_pages, tag);
777         for (i = 0; i < ret; i++)
778                 page_cache_get(pages[i]);
779         if (ret)
780                 *index = pages[ret - 1]->index + 1;
781         read_unlock_irq(&mapping->tree_lock);
782         return ret;
783 }
784 EXPORT_SYMBOL(find_get_pages_tag);
785
786 /**
787  * grab_cache_page_nowait - returns locked page at given index in given cache
788  * @mapping: target address_space
789  * @index: the page index
790  *
791  * Same as grab_cache_page(), but do not wait if the page is unavailable.
792  * This is intended for speculative data generators, where the data can
793  * be regenerated if the page couldn't be grabbed.  This routine should
794  * be safe to call while holding the lock for another page.
795  *
796  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
797  * and deadlock against the caller's locked page.
798  */
799 struct page *
800 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
801 {
802         struct page *page = find_get_page(mapping, index);
803
804         if (page) {
805                 if (!TestSetPageLocked(page))
806                         return page;
807                 page_cache_release(page);
808                 return NULL;
809         }
810         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
811         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
812                 page_cache_release(page);
813                 page = NULL;
814         }
815         return page;
816 }
817 EXPORT_SYMBOL(grab_cache_page_nowait);
818
819 /*
820  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
821  * a _large_ part of the i/o request. Imagine the worst scenario:
822  *
823  *      ---R__________________________________________B__________
824  *         ^ reading here                             ^ bad block(assume 4k)
825  *
826  * read(R) => miss => readahead(R...B) => media error => frustrating retries
827  * => failing the whole request => read(R) => read(R+1) =>
828  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
829  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
830  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
831  *
832  * It is going insane. Fix it by quickly scaling down the readahead size.
833  */
834 static void shrink_readahead_size_eio(struct file *filp,
835                                         struct file_ra_state *ra)
836 {
837         if (!ra->ra_pages)
838                 return;
839
840         ra->ra_pages /= 4;
841 }
842
843 /**
844  * do_generic_mapping_read - generic file read routine
845  * @mapping:    address_space to be read
846  * @_ra:        file's readahead state
847  * @filp:       the file to read
848  * @ppos:       current file position
849  * @desc:       read_descriptor
850  * @actor:      read method
851  *
852  * This is a generic file read routine, and uses the
853  * mapping->a_ops->readpage() function for the actual low-level stuff.
854  *
855  * This is really ugly. But the goto's actually try to clarify some
856  * of the logic when it comes to error handling etc.
857  *
858  * Note the struct file* is only passed for the use of readpage.
859  * It may be NULL.
860  */
861 void do_generic_mapping_read(struct address_space *mapping,
862                              struct file_ra_state *_ra,
863                              struct file *filp,
864                              loff_t *ppos,
865                              read_descriptor_t *desc,
866                              read_actor_t actor)
867 {
868         struct inode *inode = mapping->host;
869         unsigned long index;
870         unsigned long offset;
871         unsigned long last_index;
872         unsigned long next_index;
873         unsigned long prev_index;
874         unsigned int prev_offset;
875         struct page *cached_page;
876         int error;
877         struct file_ra_state ra = *_ra;
878
879         cached_page = NULL;
880         index = *ppos >> PAGE_CACHE_SHIFT;
881         next_index = index;
882         prev_index = ra.prev_index;
883         prev_offset = ra.prev_offset;
884         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
885         offset = *ppos & ~PAGE_CACHE_MASK;
886
887         for (;;) {
888                 struct page *page;
889                 unsigned long end_index;
890                 loff_t isize;
891                 unsigned long nr, ret;
892
893                 cond_resched();
894 find_page:
895                 page = find_get_page(mapping, index);
896                 if (!page) {
897                         page_cache_sync_readahead(mapping,
898                                         &ra, filp,
899                                         index, last_index - index);
900                         page = find_get_page(mapping, index);
901                         if (unlikely(page == NULL))
902                                 goto no_cached_page;
903                 }
904                 if (PageReadahead(page)) {
905                         page_cache_async_readahead(mapping,
906                                         &ra, filp, page,
907                                         index, last_index - index);
908                 }
909                 if (!PageUptodate(page))
910                         goto page_not_up_to_date;
911 page_ok:
912                 /*
913                  * i_size must be checked after we know the page is Uptodate.
914                  *
915                  * Checking i_size after the check allows us to calculate
916                  * the correct value for "nr", which means the zero-filled
917                  * part of the page is not copied back to userspace (unless
918                  * another truncate extends the file - this is desired though).
919                  */
920
921                 isize = i_size_read(inode);
922                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
923                 if (unlikely(!isize || index > end_index)) {
924                         page_cache_release(page);
925                         goto out;
926                 }
927
928                 /* nr is the maximum number of bytes to copy from this page */
929                 nr = PAGE_CACHE_SIZE;
930                 if (index == end_index) {
931                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
932                         if (nr <= offset) {
933                                 page_cache_release(page);
934                                 goto out;
935                         }
936                 }
937                 nr = nr - offset;
938
939                 /* If users can be writing to this page using arbitrary
940                  * virtual addresses, take care about potential aliasing
941                  * before reading the page on the kernel side.
942                  */
943                 if (mapping_writably_mapped(mapping))
944                         flush_dcache_page(page);
945
946                 /*
947                  * When a sequential read accesses a page several times,
948                  * only mark it as accessed the first time.
949                  */
950                 if (prev_index != index || offset != prev_offset)
951                         mark_page_accessed(page);
952                 prev_index = index;
953
954                 /*
955                  * Ok, we have the page, and it's up-to-date, so
956                  * now we can copy it to user space...
957                  *
958                  * The actor routine returns how many bytes were actually used..
959                  * NOTE! This may not be the same as how much of a user buffer
960                  * we filled up (we may be padding etc), so we can only update
961                  * "pos" here (the actor routine has to update the user buffer
962                  * pointers and the remaining count).
963                  */
964                 ret = actor(desc, page, offset, nr);
965                 offset += ret;
966                 index += offset >> PAGE_CACHE_SHIFT;
967                 offset &= ~PAGE_CACHE_MASK;
968                 prev_offset = offset;
969                 ra.prev_offset = offset;
970
971                 page_cache_release(page);
972                 if (ret == nr && desc->count)
973                         continue;
974                 goto out;
975
976 page_not_up_to_date:
977                 /* Get exclusive access to the page ... */
978                 lock_page(page);
979
980                 /* Did it get truncated before we got the lock? */
981                 if (!page->mapping) {
982                         unlock_page(page);
983                         page_cache_release(page);
984                         continue;
985                 }
986
987                 /* Did somebody else fill it already? */
988                 if (PageUptodate(page)) {
989                         unlock_page(page);
990                         goto page_ok;
991                 }
992
993 readpage:
994                 /* Start the actual read. The read will unlock the page. */
995                 error = mapping->a_ops->readpage(filp, page);
996
997                 if (unlikely(error)) {
998                         if (error == AOP_TRUNCATED_PAGE) {
999                                 page_cache_release(page);
1000                                 goto find_page;
1001                         }
1002                         goto readpage_error;
1003                 }
1004
1005                 if (!PageUptodate(page)) {
1006                         lock_page(page);
1007                         if (!PageUptodate(page)) {
1008                                 if (page->mapping == NULL) {
1009                                         /*
1010                                          * invalidate_inode_pages got it
1011                                          */
1012                                         unlock_page(page);
1013                                         page_cache_release(page);
1014                                         goto find_page;
1015                                 }
1016                                 unlock_page(page);
1017                                 error = -EIO;
1018                                 shrink_readahead_size_eio(filp, &ra);
1019                                 goto readpage_error;
1020                         }
1021                         unlock_page(page);
1022                 }
1023
1024                 goto page_ok;
1025
1026 readpage_error:
1027                 /* UHHUH! A synchronous read error occurred. Report it */
1028                 desc->error = error;
1029                 page_cache_release(page);
1030                 goto out;
1031
1032 no_cached_page:
1033                 /*
1034                  * Ok, it wasn't cached, so we need to create a new
1035                  * page..
1036                  */
1037                 if (!cached_page) {
1038                         cached_page = page_cache_alloc_cold(mapping);
1039                         if (!cached_page) {
1040                                 desc->error = -ENOMEM;
1041                                 goto out;
1042                         }
1043                 }
1044                 error = add_to_page_cache_lru(cached_page, mapping,
1045                                                 index, GFP_KERNEL);
1046                 if (error) {
1047                         if (error == -EEXIST)
1048                                 goto find_page;
1049                         desc->error = error;
1050                         goto out;
1051                 }
1052                 page = cached_page;
1053                 cached_page = NULL;
1054                 goto readpage;
1055         }
1056
1057 out:
1058         *_ra = ra;
1059         _ra->prev_index = prev_index;
1060
1061         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1062         if (cached_page)
1063                 page_cache_release(cached_page);
1064         if (filp)
1065                 file_accessed(filp);
1066 }
1067 EXPORT_SYMBOL(do_generic_mapping_read);
1068
1069 int file_read_actor(read_descriptor_t *desc, struct page *page,
1070                         unsigned long offset, unsigned long size)
1071 {
1072         char *kaddr;
1073         unsigned long left, count = desc->count;
1074
1075         if (size > count)
1076                 size = count;
1077
1078         /*
1079          * Faults on the destination of a read are common, so do it before
1080          * taking the kmap.
1081          */
1082         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1083                 kaddr = kmap_atomic(page, KM_USER0);
1084                 left = __copy_to_user_inatomic(desc->arg.buf,
1085                                                 kaddr + offset, size);
1086                 kunmap_atomic(kaddr, KM_USER0);
1087                 if (left == 0)
1088                         goto success;
1089         }
1090
1091         /* Do it the slow way */
1092         kaddr = kmap(page);
1093         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1094         kunmap(page);
1095
1096         if (left) {
1097                 size -= left;
1098                 desc->error = -EFAULT;
1099         }
1100 success:
1101         desc->count = count - size;
1102         desc->written += size;
1103         desc->arg.buf += size;
1104         return size;
1105 }
1106
1107 /*
1108  * Performs necessary checks before doing a write
1109  * @iov:        io vector request
1110  * @nr_segs:    number of segments in the iovec
1111  * @count:      number of bytes to write
1112  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1113  *
1114  * Adjust number of segments and amount of bytes to write (nr_segs should be
1115  * properly initialized first). Returns appropriate error code that caller
1116  * should return or zero in case that write should be allowed.
1117  */
1118 int generic_segment_checks(const struct iovec *iov,
1119                         unsigned long *nr_segs, size_t *count, int access_flags)
1120 {
1121         unsigned long   seg;
1122         size_t cnt = 0;
1123         for (seg = 0; seg < *nr_segs; seg++) {
1124                 const struct iovec *iv = &iov[seg];
1125
1126                 /*
1127                  * If any segment has a negative length, or the cumulative
1128                  * length ever wraps negative then return -EINVAL.
1129                  */
1130                 cnt += iv->iov_len;
1131                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1132                         return -EINVAL;
1133                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1134                         continue;
1135                 if (seg == 0)
1136                         return -EFAULT;
1137                 *nr_segs = seg;
1138                 cnt -= iv->iov_len;     /* This segment is no good */
1139                 break;
1140         }
1141         *count = cnt;
1142         return 0;
1143 }
1144 EXPORT_SYMBOL(generic_segment_checks);
1145
1146 /**
1147  * generic_file_aio_read - generic filesystem read routine
1148  * @iocb:       kernel I/O control block
1149  * @iov:        io vector request
1150  * @nr_segs:    number of segments in the iovec
1151  * @pos:        current file position
1152  *
1153  * This is the "read()" routine for all filesystems
1154  * that can use the page cache directly.
1155  */
1156 ssize_t
1157 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1158                 unsigned long nr_segs, loff_t pos)
1159 {
1160         struct file *filp = iocb->ki_filp;
1161         ssize_t retval;
1162         unsigned long seg;
1163         size_t count;
1164         loff_t *ppos = &iocb->ki_pos;
1165
1166         count = 0;
1167         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1168         if (retval)
1169                 return retval;
1170
1171         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1172         if (filp->f_flags & O_DIRECT) {
1173                 loff_t size;
1174                 struct address_space *mapping;
1175                 struct inode *inode;
1176
1177                 mapping = filp->f_mapping;
1178                 inode = mapping->host;
1179                 retval = 0;
1180                 if (!count)
1181                         goto out; /* skip atime */
1182                 size = i_size_read(inode);
1183                 if (pos < size) {
1184                         retval = generic_file_direct_IO(READ, iocb,
1185                                                 iov, pos, nr_segs);
1186                         if (retval > 0)
1187                                 *ppos = pos + retval;
1188                 }
1189                 if (likely(retval != 0)) {
1190                         file_accessed(filp);
1191                         goto out;
1192                 }
1193         }
1194
1195         retval = 0;
1196         if (count) {
1197                 for (seg = 0; seg < nr_segs; seg++) {
1198                         read_descriptor_t desc;
1199
1200                         desc.written = 0;
1201                         desc.arg.buf = iov[seg].iov_base;
1202                         desc.count = iov[seg].iov_len;
1203                         if (desc.count == 0)
1204                                 continue;
1205                         desc.error = 0;
1206                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1207                         retval += desc.written;
1208                         if (desc.error) {
1209                                 retval = retval ?: desc.error;
1210                                 break;
1211                         }
1212                         if (desc.count > 0)
1213                                 break;
1214                 }
1215         }
1216 out:
1217         return retval;
1218 }
1219 EXPORT_SYMBOL(generic_file_aio_read);
1220
1221 static ssize_t
1222 do_readahead(struct address_space *mapping, struct file *filp,
1223              unsigned long index, unsigned long nr)
1224 {
1225         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1226                 return -EINVAL;
1227
1228         force_page_cache_readahead(mapping, filp, index,
1229                                         max_sane_readahead(nr));
1230         return 0;
1231 }
1232
1233 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1234 {
1235         ssize_t ret;
1236         struct file *file;
1237
1238         ret = -EBADF;
1239         file = fget(fd);
1240         if (file) {
1241                 if (file->f_mode & FMODE_READ) {
1242                         struct address_space *mapping = file->f_mapping;
1243                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1244                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1245                         unsigned long len = end - start + 1;
1246                         ret = do_readahead(mapping, file, start, len);
1247                 }
1248                 fput(file);
1249         }
1250         return ret;
1251 }
1252
1253 #ifdef CONFIG_MMU
1254 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1255 /**
1256  * page_cache_read - adds requested page to the page cache if not already there
1257  * @file:       file to read
1258  * @offset:     page index
1259  *
1260  * This adds the requested page to the page cache if it isn't already there,
1261  * and schedules an I/O to read in its contents from disk.
1262  */
1263 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1264 {
1265         struct address_space *mapping = file->f_mapping;
1266         struct page *page; 
1267         int ret;
1268
1269         do {
1270                 page = page_cache_alloc_cold(mapping);
1271                 if (!page)
1272                         return -ENOMEM;
1273
1274                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1275                 if (ret == 0)
1276                         ret = mapping->a_ops->readpage(file, page);
1277                 else if (ret == -EEXIST)
1278                         ret = 0; /* losing race to add is OK */
1279
1280                 page_cache_release(page);
1281
1282         } while (ret == AOP_TRUNCATED_PAGE);
1283                 
1284         return ret;
1285 }
1286
1287 #define MMAP_LOTSAMISS  (100)
1288
1289 /**
1290  * filemap_fault - read in file data for page fault handling
1291  * @vma:        vma in which the fault was taken
1292  * @vmf:        struct vm_fault containing details of the fault
1293  *
1294  * filemap_fault() is invoked via the vma operations vector for a
1295  * mapped memory region to read in file data during a page fault.
1296  *
1297  * The goto's are kind of ugly, but this streamlines the normal case of having
1298  * it in the page cache, and handles the special cases reasonably without
1299  * having a lot of duplicated code.
1300  */
1301 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1302 {
1303         int error;
1304         struct file *file = vma->vm_file;
1305         struct address_space *mapping = file->f_mapping;
1306         struct file_ra_state *ra = &file->f_ra;
1307         struct inode *inode = mapping->host;
1308         struct page *page;
1309         unsigned long size;
1310         int did_readaround = 0;
1311         int ret = 0;
1312
1313         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1314         if (vmf->pgoff >= size)
1315                 goto outside_data_content;
1316
1317         /* If we don't want any read-ahead, don't bother */
1318         if (VM_RandomReadHint(vma))
1319                 goto no_cached_page;
1320
1321         /*
1322          * Do we have something in the page cache already?
1323          */
1324 retry_find:
1325         page = find_lock_page(mapping, vmf->pgoff);
1326         /*
1327          * For sequential accesses, we use the generic readahead logic.
1328          */
1329         if (VM_SequentialReadHint(vma)) {
1330                 if (!page) {
1331                         page_cache_sync_readahead(mapping, ra, file,
1332                                                            vmf->pgoff, 1);
1333                         page = find_lock_page(mapping, vmf->pgoff);
1334                         if (!page)
1335                                 goto no_cached_page;
1336                 }
1337                 if (PageReadahead(page)) {
1338                         page_cache_async_readahead(mapping, ra, file, page,
1339                                                            vmf->pgoff, 1);
1340                 }
1341         }
1342
1343         if (!page) {
1344                 unsigned long ra_pages;
1345
1346                 ra->mmap_miss++;
1347
1348                 /*
1349                  * Do we miss much more than hit in this file? If so,
1350                  * stop bothering with read-ahead. It will only hurt.
1351                  */
1352                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1353                         goto no_cached_page;
1354
1355                 /*
1356                  * To keep the pgmajfault counter straight, we need to
1357                  * check did_readaround, as this is an inner loop.
1358                  */
1359                 if (!did_readaround) {
1360                         ret = VM_FAULT_MAJOR;
1361                         count_vm_event(PGMAJFAULT);
1362                 }
1363                 did_readaround = 1;
1364                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1365                 if (ra_pages) {
1366                         pgoff_t start = 0;
1367
1368                         if (vmf->pgoff > ra_pages / 2)
1369                                 start = vmf->pgoff - ra_pages / 2;
1370                         do_page_cache_readahead(mapping, file, start, ra_pages);
1371                 }
1372                 page = find_lock_page(mapping, vmf->pgoff);
1373                 if (!page)
1374                         goto no_cached_page;
1375         }
1376
1377         if (!did_readaround)
1378                 ra->mmap_miss--;
1379
1380         /*
1381          * We have a locked page in the page cache, now we need to check
1382          * that it's up-to-date. If not, it is going to be due to an error.
1383          */
1384         if (unlikely(!PageUptodate(page)))
1385                 goto page_not_uptodate;
1386
1387         /* Must recheck i_size under page lock */
1388         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1389         if (unlikely(vmf->pgoff >= size)) {
1390                 unlock_page(page);
1391                 page_cache_release(page);
1392                 goto outside_data_content;
1393         }
1394
1395         /*
1396          * Found the page and have a reference on it.
1397          */
1398         mark_page_accessed(page);
1399         ra->prev_index = page->index;
1400         vmf->page = page;
1401         return ret | VM_FAULT_LOCKED;
1402
1403 outside_data_content:
1404         /*
1405          * An external ptracer can access pages that normally aren't
1406          * accessible..
1407          */
1408         if (vma->vm_mm == current->mm)
1409                 return VM_FAULT_SIGBUS;
1410
1411         /* Fall through to the non-read-ahead case */
1412 no_cached_page:
1413         /*
1414          * We're only likely to ever get here if MADV_RANDOM is in
1415          * effect.
1416          */
1417         error = page_cache_read(file, vmf->pgoff);
1418
1419         /*
1420          * The page we want has now been added to the page cache.
1421          * In the unlikely event that someone removed it in the
1422          * meantime, we'll just come back here and read it again.
1423          */
1424         if (error >= 0)
1425                 goto retry_find;
1426
1427         /*
1428          * An error return from page_cache_read can result if the
1429          * system is low on memory, or a problem occurs while trying
1430          * to schedule I/O.
1431          */
1432         if (error == -ENOMEM)
1433                 return VM_FAULT_OOM;
1434         return VM_FAULT_SIGBUS;
1435
1436 page_not_uptodate:
1437         /* IO error path */
1438         if (!did_readaround) {
1439                 ret = VM_FAULT_MAJOR;
1440                 count_vm_event(PGMAJFAULT);
1441         }
1442
1443         /*
1444          * Umm, take care of errors if the page isn't up-to-date.
1445          * Try to re-read it _once_. We do this synchronously,
1446          * because there really aren't any performance issues here
1447          * and we need to check for errors.
1448          */
1449         ClearPageError(page);
1450         error = mapping->a_ops->readpage(file, page);
1451         page_cache_release(page);
1452
1453         if (!error || error == AOP_TRUNCATED_PAGE)
1454                 goto retry_find;
1455
1456         /* Things didn't work out. Return zero to tell the mm layer so. */
1457         shrink_readahead_size_eio(file, ra);
1458         return VM_FAULT_SIGBUS;
1459 }
1460 EXPORT_SYMBOL(filemap_fault);
1461
1462 struct vm_operations_struct generic_file_vm_ops = {
1463         .fault          = filemap_fault,
1464 };
1465
1466 /* This is used for a general mmap of a disk file */
1467
1468 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1469 {
1470         struct address_space *mapping = file->f_mapping;
1471
1472         if (!mapping->a_ops->readpage)
1473                 return -ENOEXEC;
1474         file_accessed(file);
1475         vma->vm_ops = &generic_file_vm_ops;
1476         vma->vm_flags |= VM_CAN_NONLINEAR;
1477         return 0;
1478 }
1479
1480 /*
1481  * This is for filesystems which do not implement ->writepage.
1482  */
1483 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1484 {
1485         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1486                 return -EINVAL;
1487         return generic_file_mmap(file, vma);
1488 }
1489 #else
1490 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1491 {
1492         return -ENOSYS;
1493 }
1494 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1495 {
1496         return -ENOSYS;
1497 }
1498 #endif /* CONFIG_MMU */
1499
1500 EXPORT_SYMBOL(generic_file_mmap);
1501 EXPORT_SYMBOL(generic_file_readonly_mmap);
1502
1503 static struct page *__read_cache_page(struct address_space *mapping,
1504                                 unsigned long index,
1505                                 int (*filler)(void *,struct page*),
1506                                 void *data)
1507 {
1508         struct page *page, *cached_page = NULL;
1509         int err;
1510 repeat:
1511         page = find_get_page(mapping, index);
1512         if (!page) {
1513                 if (!cached_page) {
1514                         cached_page = page_cache_alloc_cold(mapping);
1515                         if (!cached_page)
1516                                 return ERR_PTR(-ENOMEM);
1517                 }
1518                 err = add_to_page_cache_lru(cached_page, mapping,
1519                                         index, GFP_KERNEL);
1520                 if (err == -EEXIST)
1521                         goto repeat;
1522                 if (err < 0) {
1523                         /* Presumably ENOMEM for radix tree node */
1524                         page_cache_release(cached_page);
1525                         return ERR_PTR(err);
1526                 }
1527                 page = cached_page;
1528                 cached_page = NULL;
1529                 err = filler(data, page);
1530                 if (err < 0) {
1531                         page_cache_release(page);
1532                         page = ERR_PTR(err);
1533                 }
1534         }
1535         if (cached_page)
1536                 page_cache_release(cached_page);
1537         return page;
1538 }
1539
1540 /*
1541  * Same as read_cache_page, but don't wait for page to become unlocked
1542  * after submitting it to the filler.
1543  */
1544 struct page *read_cache_page_async(struct address_space *mapping,
1545                                 unsigned long index,
1546                                 int (*filler)(void *,struct page*),
1547                                 void *data)
1548 {
1549         struct page *page;
1550         int err;
1551
1552 retry:
1553         page = __read_cache_page(mapping, index, filler, data);
1554         if (IS_ERR(page))
1555                 return page;
1556         if (PageUptodate(page))
1557                 goto out;
1558
1559         lock_page(page);
1560         if (!page->mapping) {
1561                 unlock_page(page);
1562                 page_cache_release(page);
1563                 goto retry;
1564         }
1565         if (PageUptodate(page)) {
1566                 unlock_page(page);
1567                 goto out;
1568         }
1569         err = filler(data, page);
1570         if (err < 0) {
1571                 page_cache_release(page);
1572                 return ERR_PTR(err);
1573         }
1574 out:
1575         mark_page_accessed(page);
1576         return page;
1577 }
1578 EXPORT_SYMBOL(read_cache_page_async);
1579
1580 /**
1581  * read_cache_page - read into page cache, fill it if needed
1582  * @mapping:    the page's address_space
1583  * @index:      the page index
1584  * @filler:     function to perform the read
1585  * @data:       destination for read data
1586  *
1587  * Read into the page cache. If a page already exists, and PageUptodate() is
1588  * not set, try to fill the page then wait for it to become unlocked.
1589  *
1590  * If the page does not get brought uptodate, return -EIO.
1591  */
1592 struct page *read_cache_page(struct address_space *mapping,
1593                                 unsigned long index,
1594                                 int (*filler)(void *,struct page*),
1595                                 void *data)
1596 {
1597         struct page *page;
1598
1599         page = read_cache_page_async(mapping, index, filler, data);
1600         if (IS_ERR(page))
1601                 goto out;
1602         wait_on_page_locked(page);
1603         if (!PageUptodate(page)) {
1604                 page_cache_release(page);
1605                 page = ERR_PTR(-EIO);
1606         }
1607  out:
1608         return page;
1609 }
1610 EXPORT_SYMBOL(read_cache_page);
1611
1612 /*
1613  * If the page was newly created, increment its refcount and add it to the
1614  * caller's lru-buffering pagevec.  This function is specifically for
1615  * generic_file_write().
1616  */
1617 static inline struct page *
1618 __grab_cache_page(struct address_space *mapping, unsigned long index,
1619                         struct page **cached_page, struct pagevec *lru_pvec)
1620 {
1621         int err;
1622         struct page *page;
1623 repeat:
1624         page = find_lock_page(mapping, index);
1625         if (!page) {
1626                 if (!*cached_page) {
1627                         *cached_page = page_cache_alloc(mapping);
1628                         if (!*cached_page)
1629                                 return NULL;
1630                 }
1631                 err = add_to_page_cache(*cached_page, mapping,
1632                                         index, GFP_KERNEL);
1633                 if (err == -EEXIST)
1634                         goto repeat;
1635                 if (err == 0) {
1636                         page = *cached_page;
1637                         page_cache_get(page);
1638                         if (!pagevec_add(lru_pvec, page))
1639                                 __pagevec_lru_add(lru_pvec);
1640                         *cached_page = NULL;
1641                 }
1642         }
1643         return page;
1644 }
1645
1646 /*
1647  * The logic we want is
1648  *
1649  *      if suid or (sgid and xgrp)
1650  *              remove privs
1651  */
1652 int should_remove_suid(struct dentry *dentry)
1653 {
1654         mode_t mode = dentry->d_inode->i_mode;
1655         int kill = 0;
1656
1657         /* suid always must be killed */
1658         if (unlikely(mode & S_ISUID))
1659                 kill = ATTR_KILL_SUID;
1660
1661         /*
1662          * sgid without any exec bits is just a mandatory locking mark; leave
1663          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1664          */
1665         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1666                 kill |= ATTR_KILL_SGID;
1667
1668         if (unlikely(kill && !capable(CAP_FSETID)))
1669                 return kill;
1670
1671         return 0;
1672 }
1673 EXPORT_SYMBOL(should_remove_suid);
1674
1675 int __remove_suid(struct dentry *dentry, int kill)
1676 {
1677         struct iattr newattrs;
1678
1679         newattrs.ia_valid = ATTR_FORCE | kill;
1680         return notify_change(dentry, &newattrs);
1681 }
1682
1683 int remove_suid(struct dentry *dentry)
1684 {
1685         int kill = should_remove_suid(dentry);
1686
1687         if (unlikely(kill))
1688                 return __remove_suid(dentry, kill);
1689
1690         return 0;
1691 }
1692 EXPORT_SYMBOL(remove_suid);
1693
1694 size_t
1695 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1696                         const struct iovec *iov, size_t base, size_t bytes)
1697 {
1698         size_t copied = 0, left = 0;
1699
1700         while (bytes) {
1701                 char __user *buf = iov->iov_base + base;
1702                 int copy = min(bytes, iov->iov_len - base);
1703
1704                 base = 0;
1705                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1706                 copied += copy;
1707                 bytes -= copy;
1708                 vaddr += copy;
1709                 iov++;
1710
1711                 if (unlikely(left))
1712                         break;
1713         }
1714         return copied - left;
1715 }
1716
1717 /*
1718  * Performs necessary checks before doing a write
1719  *
1720  * Can adjust writing position or amount of bytes to write.
1721  * Returns appropriate error code that caller should return or
1722  * zero in case that write should be allowed.
1723  */
1724 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1725 {
1726         struct inode *inode = file->f_mapping->host;
1727         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1728
1729         if (unlikely(*pos < 0))
1730                 return -EINVAL;
1731
1732         if (!isblk) {
1733                 /* FIXME: this is for backwards compatibility with 2.4 */
1734                 if (file->f_flags & O_APPEND)
1735                         *pos = i_size_read(inode);
1736
1737                 if (limit != RLIM_INFINITY) {
1738                         if (*pos >= limit) {
1739                                 send_sig(SIGXFSZ, current, 0);
1740                                 return -EFBIG;
1741                         }
1742                         if (*count > limit - (typeof(limit))*pos) {
1743                                 *count = limit - (typeof(limit))*pos;
1744                         }
1745                 }
1746         }
1747
1748         /*
1749          * LFS rule
1750          */
1751         if (unlikely(*pos + *count > MAX_NON_LFS &&
1752                                 !(file->f_flags & O_LARGEFILE))) {
1753                 if (*pos >= MAX_NON_LFS) {
1754                         return -EFBIG;
1755                 }
1756                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1757                         *count = MAX_NON_LFS - (unsigned long)*pos;
1758                 }
1759         }
1760
1761         /*
1762          * Are we about to exceed the fs block limit ?
1763          *
1764          * If we have written data it becomes a short write.  If we have
1765          * exceeded without writing data we send a signal and return EFBIG.
1766          * Linus frestrict idea will clean these up nicely..
1767          */
1768         if (likely(!isblk)) {
1769                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1770                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1771                                 return -EFBIG;
1772                         }
1773                         /* zero-length writes at ->s_maxbytes are OK */
1774                 }
1775
1776                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1777                         *count = inode->i_sb->s_maxbytes - *pos;
1778         } else {
1779 #ifdef CONFIG_BLOCK
1780                 loff_t isize;
1781                 if (bdev_read_only(I_BDEV(inode)))
1782                         return -EPERM;
1783                 isize = i_size_read(inode);
1784                 if (*pos >= isize) {
1785                         if (*count || *pos > isize)
1786                                 return -ENOSPC;
1787                 }
1788
1789                 if (*pos + *count > isize)
1790                         *count = isize - *pos;
1791 #else
1792                 return -EPERM;
1793 #endif
1794         }
1795         return 0;
1796 }
1797 EXPORT_SYMBOL(generic_write_checks);
1798
1799 ssize_t
1800 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1801                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1802                 size_t count, size_t ocount)
1803 {
1804         struct file     *file = iocb->ki_filp;
1805         struct address_space *mapping = file->f_mapping;
1806         struct inode    *inode = mapping->host;
1807         ssize_t         written;
1808
1809         if (count != ocount)
1810                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1811
1812         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1813         if (written > 0) {
1814                 loff_t end = pos + written;
1815                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1816                         i_size_write(inode,  end);
1817                         mark_inode_dirty(inode);
1818                 }
1819                 *ppos = end;
1820         }
1821
1822         /*
1823          * Sync the fs metadata but not the minor inode changes and
1824          * of course not the data as we did direct DMA for the IO.
1825          * i_mutex is held, which protects generic_osync_inode() from
1826          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
1827          */
1828         if ((written >= 0 || written == -EIOCBQUEUED) &&
1829             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1830                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1831                 if (err < 0)
1832                         written = err;
1833         }
1834         return written;
1835 }
1836 EXPORT_SYMBOL(generic_file_direct_write);
1837
1838 ssize_t
1839 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1840                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1841                 size_t count, ssize_t written)
1842 {
1843         struct file *file = iocb->ki_filp;
1844         struct address_space * mapping = file->f_mapping;
1845         const struct address_space_operations *a_ops = mapping->a_ops;
1846         struct inode    *inode = mapping->host;
1847         long            status = 0;
1848         struct page     *page;
1849         struct page     *cached_page = NULL;
1850         size_t          bytes;
1851         struct pagevec  lru_pvec;
1852         const struct iovec *cur_iov = iov; /* current iovec */
1853         size_t          iov_base = 0;      /* offset in the current iovec */
1854         char __user     *buf;
1855
1856         pagevec_init(&lru_pvec, 0);
1857
1858         /*
1859          * handle partial DIO write.  Adjust cur_iov if needed.
1860          */
1861         if (likely(nr_segs == 1))
1862                 buf = iov->iov_base + written;
1863         else {
1864                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1865                 buf = cur_iov->iov_base + iov_base;
1866         }
1867
1868         do {
1869                 unsigned long index;
1870                 unsigned long offset;
1871                 size_t copied;
1872
1873                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1874                 index = pos >> PAGE_CACHE_SHIFT;
1875                 bytes = PAGE_CACHE_SIZE - offset;
1876
1877                 /* Limit the size of the copy to the caller's write size */
1878                 bytes = min(bytes, count);
1879
1880                 /* We only need to worry about prefaulting when writes are from
1881                  * user-space.  NFSd uses vfs_writev with several non-aligned
1882                  * segments in the vector, and limiting to one segment a time is
1883                  * a noticeable performance for re-write
1884                  */
1885                 if (!segment_eq(get_fs(), KERNEL_DS)) {
1886                         /*
1887                          * Limit the size of the copy to that of the current
1888                          * segment, because fault_in_pages_readable() doesn't
1889                          * know how to walk segments.
1890                          */
1891                         bytes = min(bytes, cur_iov->iov_len - iov_base);
1892
1893                         /*
1894                          * Bring in the user page that we will copy from
1895                          * _first_.  Otherwise there's a nasty deadlock on
1896                          * copying from the same page as we're writing to,
1897                          * without it being marked up-to-date.
1898                          */
1899                         fault_in_pages_readable(buf, bytes);
1900                 }
1901                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1902                 if (!page) {
1903                         status = -ENOMEM;
1904                         break;
1905                 }
1906
1907                 if (unlikely(bytes == 0)) {
1908                         status = 0;
1909                         copied = 0;
1910                         goto zero_length_segment;
1911                 }
1912
1913                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
1914                 if (unlikely(status)) {
1915                         loff_t isize = i_size_read(inode);
1916
1917                         if (status != AOP_TRUNCATED_PAGE)
1918                                 unlock_page(page);
1919                         page_cache_release(page);
1920                         if (status == AOP_TRUNCATED_PAGE)
1921                                 continue;
1922                         /*
1923                          * prepare_write() may have instantiated a few blocks
1924                          * outside i_size.  Trim these off again.
1925                          */
1926                         if (pos + bytes > isize)
1927                                 vmtruncate(inode, isize);
1928                         break;
1929                 }
1930                 if (likely(nr_segs == 1))
1931                         copied = filemap_copy_from_user(page, offset,
1932                                                         buf, bytes);
1933                 else
1934                         copied = filemap_copy_from_user_iovec(page, offset,
1935                                                 cur_iov, iov_base, bytes);
1936                 flush_dcache_page(page);
1937                 status = a_ops->commit_write(file, page, offset, offset+bytes);
1938                 if (status == AOP_TRUNCATED_PAGE) {
1939                         page_cache_release(page);
1940                         continue;
1941                 }
1942 zero_length_segment:
1943                 if (likely(copied >= 0)) {
1944                         if (!status)
1945                                 status = copied;
1946
1947                         if (status >= 0) {
1948                                 written += status;
1949                                 count -= status;
1950                                 pos += status;
1951                                 buf += status;
1952                                 if (unlikely(nr_segs > 1)) {
1953                                         filemap_set_next_iovec(&cur_iov,
1954                                                         &iov_base, status);
1955                                         if (count)
1956                                                 buf = cur_iov->iov_base +
1957                                                         iov_base;
1958                                 } else {
1959                                         iov_base += status;
1960                                 }
1961                         }
1962                 }
1963                 if (unlikely(copied != bytes))
1964                         if (status >= 0)
1965                                 status = -EFAULT;
1966                 unlock_page(page);
1967                 mark_page_accessed(page);
1968                 page_cache_release(page);
1969                 if (status < 0)
1970                         break;
1971                 balance_dirty_pages_ratelimited(mapping);
1972                 cond_resched();
1973         } while (count);
1974         *ppos = pos;
1975
1976         if (cached_page)
1977                 page_cache_release(cached_page);
1978
1979         /*
1980          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
1981          */
1982         if (likely(status >= 0)) {
1983                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1984                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
1985                                 status = generic_osync_inode(inode, mapping,
1986                                                 OSYNC_METADATA|OSYNC_DATA);
1987                 }
1988         }
1989         
1990         /*
1991          * If we get here for O_DIRECT writes then we must have fallen through
1992          * to buffered writes (block instantiation inside i_size).  So we sync
1993          * the file data here, to try to honour O_DIRECT expectations.
1994          */
1995         if (unlikely(file->f_flags & O_DIRECT) && written)
1996                 status = filemap_write_and_wait(mapping);
1997
1998         pagevec_lru_add(&lru_pvec);
1999         return written ? written : status;
2000 }
2001 EXPORT_SYMBOL(generic_file_buffered_write);
2002
2003 static ssize_t
2004 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2005                                 unsigned long nr_segs, loff_t *ppos)
2006 {
2007         struct file *file = iocb->ki_filp;
2008         struct address_space * mapping = file->f_mapping;
2009         size_t ocount;          /* original count */
2010         size_t count;           /* after file limit checks */
2011         struct inode    *inode = mapping->host;
2012         loff_t          pos;
2013         ssize_t         written;
2014         ssize_t         err;
2015
2016         ocount = 0;
2017         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2018         if (err)
2019                 return err;
2020
2021         count = ocount;
2022         pos = *ppos;
2023
2024         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2025
2026         /* We can write back this queue in page reclaim */
2027         current->backing_dev_info = mapping->backing_dev_info;
2028         written = 0;
2029
2030         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2031         if (err)
2032                 goto out;
2033
2034         if (count == 0)
2035                 goto out;
2036
2037         err = remove_suid(file->f_path.dentry);
2038         if (err)
2039                 goto out;
2040
2041         file_update_time(file);
2042
2043         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2044         if (unlikely(file->f_flags & O_DIRECT)) {
2045                 loff_t endbyte;
2046                 ssize_t written_buffered;
2047
2048                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2049                                                         ppos, count, ocount);
2050                 if (written < 0 || written == count)
2051                         goto out;
2052                 /*
2053                  * direct-io write to a hole: fall through to buffered I/O
2054                  * for completing the rest of the request.
2055                  */
2056                 pos += written;
2057                 count -= written;
2058                 written_buffered = generic_file_buffered_write(iocb, iov,
2059                                                 nr_segs, pos, ppos, count,
2060                                                 written);
2061                 /*
2062                  * If generic_file_buffered_write() retuned a synchronous error
2063                  * then we want to return the number of bytes which were
2064                  * direct-written, or the error code if that was zero.  Note
2065                  * that this differs from normal direct-io semantics, which
2066                  * will return -EFOO even if some bytes were written.
2067                  */
2068                 if (written_buffered < 0) {
2069                         err = written_buffered;
2070                         goto out;
2071                 }
2072
2073                 /*
2074                  * We need to ensure that the page cache pages are written to
2075                  * disk and invalidated to preserve the expected O_DIRECT
2076                  * semantics.
2077                  */
2078                 endbyte = pos + written_buffered - written - 1;
2079                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2080                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2081                                             SYNC_FILE_RANGE_WRITE|
2082                                             SYNC_FILE_RANGE_WAIT_AFTER);
2083                 if (err == 0) {
2084                         written = written_buffered;
2085                         invalidate_mapping_pages(mapping,
2086                                                  pos >> PAGE_CACHE_SHIFT,
2087                                                  endbyte >> PAGE_CACHE_SHIFT);
2088                 } else {
2089                         /*
2090                          * We don't know how much we wrote, so just return
2091                          * the number of bytes which were direct-written
2092                          */
2093                 }
2094         } else {
2095                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2096                                 pos, ppos, count, written);
2097         }
2098 out:
2099         current->backing_dev_info = NULL;
2100         return written ? written : err;
2101 }
2102
2103 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2104                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2105 {
2106         struct file *file = iocb->ki_filp;
2107         struct address_space *mapping = file->f_mapping;
2108         struct inode *inode = mapping->host;
2109         ssize_t ret;
2110
2111         BUG_ON(iocb->ki_pos != pos);
2112
2113         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2114                         &iocb->ki_pos);
2115
2116         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2117                 ssize_t err;
2118
2119                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2120                 if (err < 0)
2121                         ret = err;
2122         }
2123         return ret;
2124 }
2125 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2126
2127 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2128                 unsigned long nr_segs, loff_t pos)
2129 {
2130         struct file *file = iocb->ki_filp;
2131         struct address_space *mapping = file->f_mapping;
2132         struct inode *inode = mapping->host;
2133         ssize_t ret;
2134
2135         BUG_ON(iocb->ki_pos != pos);
2136
2137         mutex_lock(&inode->i_mutex);
2138         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2139                         &iocb->ki_pos);
2140         mutex_unlock(&inode->i_mutex);
2141
2142         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2143                 ssize_t err;
2144
2145                 err = sync_page_range(inode, mapping, pos, ret);
2146                 if (err < 0)
2147                         ret = err;
2148         }
2149         return ret;
2150 }
2151 EXPORT_SYMBOL(generic_file_aio_write);
2152
2153 /*
2154  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2155  * went wrong during pagecache shootdown.
2156  */
2157 static ssize_t
2158 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2159         loff_t offset, unsigned long nr_segs)
2160 {
2161         struct file *file = iocb->ki_filp;
2162         struct address_space *mapping = file->f_mapping;
2163         ssize_t retval;
2164         size_t write_len;
2165         pgoff_t end = 0; /* silence gcc */
2166
2167         /*
2168          * If it's a write, unmap all mmappings of the file up-front.  This
2169          * will cause any pte dirty bits to be propagated into the pageframes
2170          * for the subsequent filemap_write_and_wait().
2171          */
2172         if (rw == WRITE) {
2173                 write_len = iov_length(iov, nr_segs);
2174                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2175                 if (mapping_mapped(mapping))
2176                         unmap_mapping_range(mapping, offset, write_len, 0);
2177         }
2178
2179         retval = filemap_write_and_wait(mapping);
2180         if (retval)
2181                 goto out;
2182
2183         /*
2184          * After a write we want buffered reads to be sure to go to disk to get
2185          * the new data.  We invalidate clean cached page from the region we're
2186          * about to write.  We do this *before* the write so that we can return
2187          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2188          */
2189         if (rw == WRITE && mapping->nrpages) {
2190                 retval = invalidate_inode_pages2_range(mapping,
2191                                         offset >> PAGE_CACHE_SHIFT, end);
2192                 if (retval)
2193                         goto out;
2194         }
2195
2196         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2197         if (retval)
2198                 goto out;
2199
2200         /*
2201          * Finally, try again to invalidate clean pages which might have been
2202          * faulted in by get_user_pages() if the source of the write was an
2203          * mmap()ed region of the file we're writing.  That's a pretty crazy
2204          * thing to do, so we don't support it 100%.  If this invalidation
2205          * fails and we have -EIOCBQUEUED we ignore the failure.
2206          */
2207         if (rw == WRITE && mapping->nrpages) {
2208                 int err = invalidate_inode_pages2_range(mapping,
2209                                               offset >> PAGE_CACHE_SHIFT, end);
2210                 if (err && retval >= 0)
2211                         retval = err;
2212         }
2213 out:
2214         return retval;
2215 }
2216
2217 /**
2218  * try_to_release_page() - release old fs-specific metadata on a page
2219  *
2220  * @page: the page which the kernel is trying to free
2221  * @gfp_mask: memory allocation flags (and I/O mode)
2222  *
2223  * The address_space is to try to release any data against the page
2224  * (presumably at page->private).  If the release was successful, return `1'.
2225  * Otherwise return zero.
2226  *
2227  * The @gfp_mask argument specifies whether I/O may be performed to release
2228  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2229  *
2230  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2231  */
2232 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2233 {
2234         struct address_space * const mapping = page->mapping;
2235
2236         BUG_ON(!PageLocked(page));
2237         if (PageWriteback(page))
2238                 return 0;
2239
2240         if (mapping && mapping->a_ops->releasepage)
2241                 return mapping->a_ops->releasepage(page, gfp_mask);
2242         return try_to_free_buffers(page);
2243 }
2244
2245 EXPORT_SYMBOL(try_to_release_page);