e006c57bda5431325b511e25996b3f910d7b4b8a
[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 end_index;
871         unsigned long offset;
872         unsigned long last_index;
873         unsigned long next_index;
874         unsigned long prev_index;
875         unsigned int prev_offset;
876         loff_t isize;
877         struct page *cached_page;
878         int error;
879         struct file_ra_state ra = *_ra;
880
881         cached_page = NULL;
882         index = *ppos >> PAGE_CACHE_SHIFT;
883         next_index = index;
884         prev_index = ra.prev_index;
885         prev_offset = ra.prev_offset;
886         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
887         offset = *ppos & ~PAGE_CACHE_MASK;
888
889         isize = i_size_read(inode);
890         if (!isize)
891                 goto out;
892
893         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
894         for (;;) {
895                 struct page *page;
896                 unsigned long nr, ret;
897
898                 /* nr is the maximum number of bytes to copy from this page */
899                 nr = PAGE_CACHE_SIZE;
900                 if (index >= end_index) {
901                         if (index > end_index)
902                                 goto out;
903                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
904                         if (nr <= offset) {
905                                 goto out;
906                         }
907                 }
908                 nr = nr - offset;
909
910                 cond_resched();
911                 if (index == next_index)
912                         next_index = page_cache_readahead(mapping, &ra, filp,
913                                         index, last_index - index);
914
915 find_page:
916                 page = find_get_page(mapping, index);
917                 if (unlikely(page == NULL)) {
918                         handle_ra_miss(mapping, &ra, index);
919                         goto no_cached_page;
920                 }
921                 if (!PageUptodate(page))
922                         goto page_not_up_to_date;
923 page_ok:
924
925                 /* If users can be writing to this page using arbitrary
926                  * virtual addresses, take care about potential aliasing
927                  * before reading the page on the kernel side.
928                  */
929                 if (mapping_writably_mapped(mapping))
930                         flush_dcache_page(page);
931
932                 /*
933                  * When a sequential read accesses a page several times,
934                  * only mark it as accessed the first time.
935                  */
936                 if (prev_index != index || offset != prev_offset)
937                         mark_page_accessed(page);
938                 prev_index = index;
939
940                 /*
941                  * Ok, we have the page, and it's up-to-date, so
942                  * now we can copy it to user space...
943                  *
944                  * The actor routine returns how many bytes were actually used..
945                  * NOTE! This may not be the same as how much of a user buffer
946                  * we filled up (we may be padding etc), so we can only update
947                  * "pos" here (the actor routine has to update the user buffer
948                  * pointers and the remaining count).
949                  */
950                 ret = actor(desc, page, offset, nr);
951                 offset += ret;
952                 index += offset >> PAGE_CACHE_SHIFT;
953                 offset &= ~PAGE_CACHE_MASK;
954                 prev_offset = offset;
955                 ra.prev_offset = offset;
956
957                 page_cache_release(page);
958                 if (ret == nr && desc->count)
959                         continue;
960                 goto out;
961
962 page_not_up_to_date:
963                 /* Get exclusive access to the page ... */
964                 lock_page(page);
965
966                 /* Did it get truncated before we got the lock? */
967                 if (!page->mapping) {
968                         unlock_page(page);
969                         page_cache_release(page);
970                         continue;
971                 }
972
973                 /* Did somebody else fill it already? */
974                 if (PageUptodate(page)) {
975                         unlock_page(page);
976                         goto page_ok;
977                 }
978
979 readpage:
980                 /* Start the actual read. The read will unlock the page. */
981                 error = mapping->a_ops->readpage(filp, page);
982
983                 if (unlikely(error)) {
984                         if (error == AOP_TRUNCATED_PAGE) {
985                                 page_cache_release(page);
986                                 goto find_page;
987                         }
988                         goto readpage_error;
989                 }
990
991                 if (!PageUptodate(page)) {
992                         lock_page(page);
993                         if (!PageUptodate(page)) {
994                                 if (page->mapping == NULL) {
995                                         /*
996                                          * invalidate_inode_pages got it
997                                          */
998                                         unlock_page(page);
999                                         page_cache_release(page);
1000                                         goto find_page;
1001                                 }
1002                                 unlock_page(page);
1003                                 error = -EIO;
1004                                 shrink_readahead_size_eio(filp, &ra);
1005                                 goto readpage_error;
1006                         }
1007                         unlock_page(page);
1008                 }
1009
1010                 /*
1011                  * i_size must be checked after we have done ->readpage.
1012                  *
1013                  * Checking i_size after the readpage allows us to calculate
1014                  * the correct value for "nr", which means the zero-filled
1015                  * part of the page is not copied back to userspace (unless
1016                  * another truncate extends the file - this is desired though).
1017                  */
1018                 isize = i_size_read(inode);
1019                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1020                 if (unlikely(!isize || index > end_index)) {
1021                         page_cache_release(page);
1022                         goto out;
1023                 }
1024
1025                 /* nr is the maximum number of bytes to copy from this page */
1026                 nr = PAGE_CACHE_SIZE;
1027                 if (index == end_index) {
1028                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1029                         if (nr <= offset) {
1030                                 page_cache_release(page);
1031                                 goto out;
1032                         }
1033                 }
1034                 nr = nr - offset;
1035                 goto page_ok;
1036
1037 readpage_error:
1038                 /* UHHUH! A synchronous read error occurred. Report it */
1039                 desc->error = error;
1040                 page_cache_release(page);
1041                 goto out;
1042
1043 no_cached_page:
1044                 /*
1045                  * Ok, it wasn't cached, so we need to create a new
1046                  * page..
1047                  */
1048                 if (!cached_page) {
1049                         cached_page = page_cache_alloc_cold(mapping);
1050                         if (!cached_page) {
1051                                 desc->error = -ENOMEM;
1052                                 goto out;
1053                         }
1054                 }
1055                 error = add_to_page_cache_lru(cached_page, mapping,
1056                                                 index, GFP_KERNEL);
1057                 if (error) {
1058                         if (error == -EEXIST)
1059                                 goto find_page;
1060                         desc->error = error;
1061                         goto out;
1062                 }
1063                 page = cached_page;
1064                 cached_page = NULL;
1065                 goto readpage;
1066         }
1067
1068 out:
1069         *_ra = ra;
1070
1071         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1072         if (cached_page)
1073                 page_cache_release(cached_page);
1074         if (filp)
1075                 file_accessed(filp);
1076 }
1077 EXPORT_SYMBOL(do_generic_mapping_read);
1078
1079 int file_read_actor(read_descriptor_t *desc, struct page *page,
1080                         unsigned long offset, unsigned long size)
1081 {
1082         char *kaddr;
1083         unsigned long left, count = desc->count;
1084
1085         if (size > count)
1086                 size = count;
1087
1088         /*
1089          * Faults on the destination of a read are common, so do it before
1090          * taking the kmap.
1091          */
1092         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1093                 kaddr = kmap_atomic(page, KM_USER0);
1094                 left = __copy_to_user_inatomic(desc->arg.buf,
1095                                                 kaddr + offset, size);
1096                 kunmap_atomic(kaddr, KM_USER0);
1097                 if (left == 0)
1098                         goto success;
1099         }
1100
1101         /* Do it the slow way */
1102         kaddr = kmap(page);
1103         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1104         kunmap(page);
1105
1106         if (left) {
1107                 size -= left;
1108                 desc->error = -EFAULT;
1109         }
1110 success:
1111         desc->count = count - size;
1112         desc->written += size;
1113         desc->arg.buf += size;
1114         return size;
1115 }
1116
1117 /*
1118  * Performs necessary checks before doing a write
1119  * @iov:        io vector request
1120  * @nr_segs:    number of segments in the iovec
1121  * @count:      number of bytes to write
1122  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1123  *
1124  * Adjust number of segments and amount of bytes to write (nr_segs should be
1125  * properly initialized first). Returns appropriate error code that caller
1126  * should return or zero in case that write should be allowed.
1127  */
1128 int generic_segment_checks(const struct iovec *iov,
1129                         unsigned long *nr_segs, size_t *count, int access_flags)
1130 {
1131         unsigned long   seg;
1132         size_t cnt = 0;
1133         for (seg = 0; seg < *nr_segs; seg++) {
1134                 const struct iovec *iv = &iov[seg];
1135
1136                 /*
1137                  * If any segment has a negative length, or the cumulative
1138                  * length ever wraps negative then return -EINVAL.
1139                  */
1140                 cnt += iv->iov_len;
1141                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1142                         return -EINVAL;
1143                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1144                         continue;
1145                 if (seg == 0)
1146                         return -EFAULT;
1147                 *nr_segs = seg;
1148                 cnt -= iv->iov_len;     /* This segment is no good */
1149                 break;
1150         }
1151         *count = cnt;
1152         return 0;
1153 }
1154 EXPORT_SYMBOL(generic_segment_checks);
1155
1156 /**
1157  * generic_file_aio_read - generic filesystem read routine
1158  * @iocb:       kernel I/O control block
1159  * @iov:        io vector request
1160  * @nr_segs:    number of segments in the iovec
1161  * @pos:        current file position
1162  *
1163  * This is the "read()" routine for all filesystems
1164  * that can use the page cache directly.
1165  */
1166 ssize_t
1167 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1168                 unsigned long nr_segs, loff_t pos)
1169 {
1170         struct file *filp = iocb->ki_filp;
1171         ssize_t retval;
1172         unsigned long seg;
1173         size_t count;
1174         loff_t *ppos = &iocb->ki_pos;
1175
1176         count = 0;
1177         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1178         if (retval)
1179                 return retval;
1180
1181         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1182         if (filp->f_flags & O_DIRECT) {
1183                 loff_t size;
1184                 struct address_space *mapping;
1185                 struct inode *inode;
1186
1187                 mapping = filp->f_mapping;
1188                 inode = mapping->host;
1189                 retval = 0;
1190                 if (!count)
1191                         goto out; /* skip atime */
1192                 size = i_size_read(inode);
1193                 if (pos < size) {
1194                         retval = generic_file_direct_IO(READ, iocb,
1195                                                 iov, pos, nr_segs);
1196                         if (retval > 0)
1197                                 *ppos = pos + retval;
1198                 }
1199                 if (likely(retval != 0)) {
1200                         file_accessed(filp);
1201                         goto out;
1202                 }
1203         }
1204
1205         retval = 0;
1206         if (count) {
1207                 for (seg = 0; seg < nr_segs; seg++) {
1208                         read_descriptor_t desc;
1209
1210                         desc.written = 0;
1211                         desc.arg.buf = iov[seg].iov_base;
1212                         desc.count = iov[seg].iov_len;
1213                         if (desc.count == 0)
1214                                 continue;
1215                         desc.error = 0;
1216                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1217                         retval += desc.written;
1218                         if (desc.error) {
1219                                 retval = retval ?: desc.error;
1220                                 break;
1221                         }
1222                 }
1223         }
1224 out:
1225         return retval;
1226 }
1227 EXPORT_SYMBOL(generic_file_aio_read);
1228
1229 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1230 {
1231         ssize_t written;
1232         unsigned long count = desc->count;
1233         struct file *file = desc->arg.data;
1234
1235         if (size > count)
1236                 size = count;
1237
1238         written = file->f_op->sendpage(file, page, offset,
1239                                        size, &file->f_pos, size<count);
1240         if (written < 0) {
1241                 desc->error = written;
1242                 written = 0;
1243         }
1244         desc->count = count - written;
1245         desc->written += written;
1246         return written;
1247 }
1248
1249 static ssize_t
1250 do_readahead(struct address_space *mapping, struct file *filp,
1251              unsigned long index, unsigned long nr)
1252 {
1253         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1254                 return -EINVAL;
1255
1256         force_page_cache_readahead(mapping, filp, index,
1257                                         max_sane_readahead(nr));
1258         return 0;
1259 }
1260
1261 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1262 {
1263         ssize_t ret;
1264         struct file *file;
1265
1266         ret = -EBADF;
1267         file = fget(fd);
1268         if (file) {
1269                 if (file->f_mode & FMODE_READ) {
1270                         struct address_space *mapping = file->f_mapping;
1271                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1272                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1273                         unsigned long len = end - start + 1;
1274                         ret = do_readahead(mapping, file, start, len);
1275                 }
1276                 fput(file);
1277         }
1278         return ret;
1279 }
1280
1281 #ifdef CONFIG_MMU
1282 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1283 /**
1284  * page_cache_read - adds requested page to the page cache if not already there
1285  * @file:       file to read
1286  * @offset:     page index
1287  *
1288  * This adds the requested page to the page cache if it isn't already there,
1289  * and schedules an I/O to read in its contents from disk.
1290  */
1291 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1292 {
1293         struct address_space *mapping = file->f_mapping;
1294         struct page *page; 
1295         int ret;
1296
1297         do {
1298                 page = page_cache_alloc_cold(mapping);
1299                 if (!page)
1300                         return -ENOMEM;
1301
1302                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1303                 if (ret == 0)
1304                         ret = mapping->a_ops->readpage(file, page);
1305                 else if (ret == -EEXIST)
1306                         ret = 0; /* losing race to add is OK */
1307
1308                 page_cache_release(page);
1309
1310         } while (ret == AOP_TRUNCATED_PAGE);
1311                 
1312         return ret;
1313 }
1314
1315 #define MMAP_LOTSAMISS  (100)
1316
1317 /**
1318  * filemap_nopage - read in file data for page fault handling
1319  * @area:       the applicable vm_area
1320  * @address:    target address to read in
1321  * @type:       returned with VM_FAULT_{MINOR,MAJOR} if not %NULL
1322  *
1323  * filemap_nopage() is invoked via the vma operations vector for a
1324  * mapped memory region to read in file data during a page fault.
1325  *
1326  * The goto's are kind of ugly, but this streamlines the normal case of having
1327  * it in the page cache, and handles the special cases reasonably without
1328  * having a lot of duplicated code.
1329  */
1330 struct page *filemap_nopage(struct vm_area_struct *area,
1331                                 unsigned long address, int *type)
1332 {
1333         int error;
1334         struct file *file = area->vm_file;
1335         struct address_space *mapping = file->f_mapping;
1336         struct file_ra_state *ra = &file->f_ra;
1337         struct inode *inode = mapping->host;
1338         struct page *page;
1339         unsigned long size, pgoff;
1340         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1341
1342         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1343
1344 retry_all:
1345         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1346         if (pgoff >= size)
1347                 goto outside_data_content;
1348
1349         /* If we don't want any read-ahead, don't bother */
1350         if (VM_RandomReadHint(area))
1351                 goto no_cached_page;
1352
1353         /*
1354          * The readahead code wants to be told about each and every page
1355          * so it can build and shrink its windows appropriately
1356          *
1357          * For sequential accesses, we use the generic readahead logic.
1358          */
1359         if (VM_SequentialReadHint(area))
1360                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1361
1362         /*
1363          * Do we have something in the page cache already?
1364          */
1365 retry_find:
1366         page = find_get_page(mapping, pgoff);
1367         if (!page) {
1368                 unsigned long ra_pages;
1369
1370                 if (VM_SequentialReadHint(area)) {
1371                         handle_ra_miss(mapping, ra, pgoff);
1372                         goto no_cached_page;
1373                 }
1374                 ra->mmap_miss++;
1375
1376                 /*
1377                  * Do we miss much more than hit in this file? If so,
1378                  * stop bothering with read-ahead. It will only hurt.
1379                  */
1380                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1381                         goto no_cached_page;
1382
1383                 /*
1384                  * To keep the pgmajfault counter straight, we need to
1385                  * check did_readaround, as this is an inner loop.
1386                  */
1387                 if (!did_readaround) {
1388                         majmin = VM_FAULT_MAJOR;
1389                         count_vm_event(PGMAJFAULT);
1390                 }
1391                 did_readaround = 1;
1392                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1393                 if (ra_pages) {
1394                         pgoff_t start = 0;
1395
1396                         if (pgoff > ra_pages / 2)
1397                                 start = pgoff - ra_pages / 2;
1398                         do_page_cache_readahead(mapping, file, start, ra_pages);
1399                 }
1400                 page = find_get_page(mapping, pgoff);
1401                 if (!page)
1402                         goto no_cached_page;
1403         }
1404
1405         if (!did_readaround)
1406                 ra->mmap_hit++;
1407
1408         /*
1409          * Ok, found a page in the page cache, now we need to check
1410          * that it's up-to-date.
1411          */
1412         if (!PageUptodate(page))
1413                 goto page_not_uptodate;
1414
1415 success:
1416         /*
1417          * Found the page and have a reference on it.
1418          */
1419         mark_page_accessed(page);
1420         if (type)
1421                 *type = majmin;
1422         return page;
1423
1424 outside_data_content:
1425         /*
1426          * An external ptracer can access pages that normally aren't
1427          * accessible..
1428          */
1429         if (area->vm_mm == current->mm)
1430                 return NOPAGE_SIGBUS;
1431         /* Fall through to the non-read-ahead case */
1432 no_cached_page:
1433         /*
1434          * We're only likely to ever get here if MADV_RANDOM is in
1435          * effect.
1436          */
1437         error = page_cache_read(file, pgoff);
1438
1439         /*
1440          * The page we want has now been added to the page cache.
1441          * In the unlikely event that someone removed it in the
1442          * meantime, we'll just come back here and read it again.
1443          */
1444         if (error >= 0)
1445                 goto retry_find;
1446
1447         /*
1448          * An error return from page_cache_read can result if the
1449          * system is low on memory, or a problem occurs while trying
1450          * to schedule I/O.
1451          */
1452         if (error == -ENOMEM)
1453                 return NOPAGE_OOM;
1454         return NOPAGE_SIGBUS;
1455
1456 page_not_uptodate:
1457         if (!did_readaround) {
1458                 majmin = VM_FAULT_MAJOR;
1459                 count_vm_event(PGMAJFAULT);
1460         }
1461
1462         /*
1463          * Umm, take care of errors if the page isn't up-to-date.
1464          * Try to re-read it _once_. We do this synchronously,
1465          * because there really aren't any performance issues here
1466          * and we need to check for errors.
1467          */
1468         lock_page(page);
1469
1470         /* Somebody truncated the page on us? */
1471         if (!page->mapping) {
1472                 unlock_page(page);
1473                 page_cache_release(page);
1474                 goto retry_all;
1475         }
1476
1477         /* Somebody else successfully read it in? */
1478         if (PageUptodate(page)) {
1479                 unlock_page(page);
1480                 goto success;
1481         }
1482         ClearPageError(page);
1483         error = mapping->a_ops->readpage(file, page);
1484         if (!error) {
1485                 wait_on_page_locked(page);
1486                 if (PageUptodate(page))
1487                         goto success;
1488         } else if (error == AOP_TRUNCATED_PAGE) {
1489                 page_cache_release(page);
1490                 goto retry_find;
1491         }
1492
1493         /*
1494          * Things didn't work out. Return zero to tell the
1495          * mm layer so, possibly freeing the page cache page first.
1496          */
1497         shrink_readahead_size_eio(file, ra);
1498         page_cache_release(page);
1499         return NOPAGE_SIGBUS;
1500 }
1501 EXPORT_SYMBOL(filemap_nopage);
1502
1503 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1504                                         int nonblock)
1505 {
1506         struct address_space *mapping = file->f_mapping;
1507         struct page *page;
1508         int error;
1509
1510         /*
1511          * Do we have something in the page cache already?
1512          */
1513 retry_find:
1514         page = find_get_page(mapping, pgoff);
1515         if (!page) {
1516                 if (nonblock)
1517                         return NULL;
1518                 goto no_cached_page;
1519         }
1520
1521         /*
1522          * Ok, found a page in the page cache, now we need to check
1523          * that it's up-to-date.
1524          */
1525         if (!PageUptodate(page)) {
1526                 if (nonblock) {
1527                         page_cache_release(page);
1528                         return NULL;
1529                 }
1530                 goto page_not_uptodate;
1531         }
1532
1533 success:
1534         /*
1535          * Found the page and have a reference on it.
1536          */
1537         mark_page_accessed(page);
1538         return page;
1539
1540 no_cached_page:
1541         error = page_cache_read(file, pgoff);
1542
1543         /*
1544          * The page we want has now been added to the page cache.
1545          * In the unlikely event that someone removed it in the
1546          * meantime, we'll just come back here and read it again.
1547          */
1548         if (error >= 0)
1549                 goto retry_find;
1550
1551         /*
1552          * An error return from page_cache_read can result if the
1553          * system is low on memory, or a problem occurs while trying
1554          * to schedule I/O.
1555          */
1556         return NULL;
1557
1558 page_not_uptodate:
1559         lock_page(page);
1560
1561         /* Did it get truncated while we waited for it? */
1562         if (!page->mapping) {
1563                 unlock_page(page);
1564                 goto err;
1565         }
1566
1567         /* Did somebody else get it up-to-date? */
1568         if (PageUptodate(page)) {
1569                 unlock_page(page);
1570                 goto success;
1571         }
1572
1573         error = mapping->a_ops->readpage(file, page);
1574         if (!error) {
1575                 wait_on_page_locked(page);
1576                 if (PageUptodate(page))
1577                         goto success;
1578         } else if (error == AOP_TRUNCATED_PAGE) {
1579                 page_cache_release(page);
1580                 goto retry_find;
1581         }
1582
1583         /*
1584          * Umm, take care of errors if the page isn't up-to-date.
1585          * Try to re-read it _once_. We do this synchronously,
1586          * because there really aren't any performance issues here
1587          * and we need to check for errors.
1588          */
1589         lock_page(page);
1590
1591         /* Somebody truncated the page on us? */
1592         if (!page->mapping) {
1593                 unlock_page(page);
1594                 goto err;
1595         }
1596         /* Somebody else successfully read it in? */
1597         if (PageUptodate(page)) {
1598                 unlock_page(page);
1599                 goto success;
1600         }
1601
1602         ClearPageError(page);
1603         error = mapping->a_ops->readpage(file, page);
1604         if (!error) {
1605                 wait_on_page_locked(page);
1606                 if (PageUptodate(page))
1607                         goto success;
1608         } else if (error == AOP_TRUNCATED_PAGE) {
1609                 page_cache_release(page);
1610                 goto retry_find;
1611         }
1612
1613         /*
1614          * Things didn't work out. Return zero to tell the
1615          * mm layer so, possibly freeing the page cache page first.
1616          */
1617 err:
1618         page_cache_release(page);
1619
1620         return NULL;
1621 }
1622
1623 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1624                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1625                 int nonblock)
1626 {
1627         struct file *file = vma->vm_file;
1628         struct address_space *mapping = file->f_mapping;
1629         struct inode *inode = mapping->host;
1630         unsigned long size;
1631         struct mm_struct *mm = vma->vm_mm;
1632         struct page *page;
1633         int err;
1634
1635         if (!nonblock)
1636                 force_page_cache_readahead(mapping, vma->vm_file,
1637                                         pgoff, len >> PAGE_CACHE_SHIFT);
1638
1639 repeat:
1640         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1641         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1642                 return -EINVAL;
1643
1644         page = filemap_getpage(file, pgoff, nonblock);
1645
1646         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1647          * done in shmem_populate calling shmem_getpage */
1648         if (!page && !nonblock)
1649                 return -ENOMEM;
1650
1651         if (page) {
1652                 err = install_page(mm, vma, addr, page, prot);
1653                 if (err) {
1654                         page_cache_release(page);
1655                         return err;
1656                 }
1657         } else if (vma->vm_flags & VM_NONLINEAR) {
1658                 /* No page was found just because we can't read it in now (being
1659                  * here implies nonblock != 0), but the page may exist, so set
1660                  * the PTE to fault it in later. */
1661                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1662                 if (err)
1663                         return err;
1664         }
1665
1666         len -= PAGE_SIZE;
1667         addr += PAGE_SIZE;
1668         pgoff++;
1669         if (len)
1670                 goto repeat;
1671
1672         return 0;
1673 }
1674 EXPORT_SYMBOL(filemap_populate);
1675
1676 struct vm_operations_struct generic_file_vm_ops = {
1677         .nopage         = filemap_nopage,
1678         .populate       = filemap_populate,
1679 };
1680
1681 /* This is used for a general mmap of a disk file */
1682
1683 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1684 {
1685         struct address_space *mapping = file->f_mapping;
1686
1687         if (!mapping->a_ops->readpage)
1688                 return -ENOEXEC;
1689         file_accessed(file);
1690         vma->vm_ops = &generic_file_vm_ops;
1691         return 0;
1692 }
1693
1694 /*
1695  * This is for filesystems which do not implement ->writepage.
1696  */
1697 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1698 {
1699         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1700                 return -EINVAL;
1701         return generic_file_mmap(file, vma);
1702 }
1703 #else
1704 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1705 {
1706         return -ENOSYS;
1707 }
1708 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1709 {
1710         return -ENOSYS;
1711 }
1712 #endif /* CONFIG_MMU */
1713
1714 EXPORT_SYMBOL(generic_file_mmap);
1715 EXPORT_SYMBOL(generic_file_readonly_mmap);
1716
1717 static struct page *__read_cache_page(struct address_space *mapping,
1718                                 unsigned long index,
1719                                 int (*filler)(void *,struct page*),
1720                                 void *data)
1721 {
1722         struct page *page, *cached_page = NULL;
1723         int err;
1724 repeat:
1725         page = find_get_page(mapping, index);
1726         if (!page) {
1727                 if (!cached_page) {
1728                         cached_page = page_cache_alloc_cold(mapping);
1729                         if (!cached_page)
1730                                 return ERR_PTR(-ENOMEM);
1731                 }
1732                 err = add_to_page_cache_lru(cached_page, mapping,
1733                                         index, GFP_KERNEL);
1734                 if (err == -EEXIST)
1735                         goto repeat;
1736                 if (err < 0) {
1737                         /* Presumably ENOMEM for radix tree node */
1738                         page_cache_release(cached_page);
1739                         return ERR_PTR(err);
1740                 }
1741                 page = cached_page;
1742                 cached_page = NULL;
1743                 err = filler(data, page);
1744                 if (err < 0) {
1745                         page_cache_release(page);
1746                         page = ERR_PTR(err);
1747                 }
1748         }
1749         if (cached_page)
1750                 page_cache_release(cached_page);
1751         return page;
1752 }
1753
1754 /*
1755  * Same as read_cache_page, but don't wait for page to become unlocked
1756  * after submitting it to the filler.
1757  */
1758 struct page *read_cache_page_async(struct address_space *mapping,
1759                                 unsigned long index,
1760                                 int (*filler)(void *,struct page*),
1761                                 void *data)
1762 {
1763         struct page *page;
1764         int err;
1765
1766 retry:
1767         page = __read_cache_page(mapping, index, filler, data);
1768         if (IS_ERR(page))
1769                 return page;
1770         if (PageUptodate(page))
1771                 goto out;
1772
1773         lock_page(page);
1774         if (!page->mapping) {
1775                 unlock_page(page);
1776                 page_cache_release(page);
1777                 goto retry;
1778         }
1779         if (PageUptodate(page)) {
1780                 unlock_page(page);
1781                 goto out;
1782         }
1783         err = filler(data, page);
1784         if (err < 0) {
1785                 page_cache_release(page);
1786                 return ERR_PTR(err);
1787         }
1788 out:
1789         mark_page_accessed(page);
1790         return page;
1791 }
1792 EXPORT_SYMBOL(read_cache_page_async);
1793
1794 /**
1795  * read_cache_page - read into page cache, fill it if needed
1796  * @mapping:    the page's address_space
1797  * @index:      the page index
1798  * @filler:     function to perform the read
1799  * @data:       destination for read data
1800  *
1801  * Read into the page cache. If a page already exists, and PageUptodate() is
1802  * not set, try to fill the page then wait for it to become unlocked.
1803  *
1804  * If the page does not get brought uptodate, return -EIO.
1805  */
1806 struct page *read_cache_page(struct address_space *mapping,
1807                                 unsigned long index,
1808                                 int (*filler)(void *,struct page*),
1809                                 void *data)
1810 {
1811         struct page *page;
1812
1813         page = read_cache_page_async(mapping, index, filler, data);
1814         if (IS_ERR(page))
1815                 goto out;
1816         wait_on_page_locked(page);
1817         if (!PageUptodate(page)) {
1818                 page_cache_release(page);
1819                 page = ERR_PTR(-EIO);
1820         }
1821  out:
1822         return page;
1823 }
1824 EXPORT_SYMBOL(read_cache_page);
1825
1826 /*
1827  * If the page was newly created, increment its refcount and add it to the
1828  * caller's lru-buffering pagevec.  This function is specifically for
1829  * generic_file_write().
1830  */
1831 static inline struct page *
1832 __grab_cache_page(struct address_space *mapping, unsigned long index,
1833                         struct page **cached_page, struct pagevec *lru_pvec)
1834 {
1835         int err;
1836         struct page *page;
1837 repeat:
1838         page = find_lock_page(mapping, index);
1839         if (!page) {
1840                 if (!*cached_page) {
1841                         *cached_page = page_cache_alloc(mapping);
1842                         if (!*cached_page)
1843                                 return NULL;
1844                 }
1845                 err = add_to_page_cache(*cached_page, mapping,
1846                                         index, GFP_KERNEL);
1847                 if (err == -EEXIST)
1848                         goto repeat;
1849                 if (err == 0) {
1850                         page = *cached_page;
1851                         page_cache_get(page);
1852                         if (!pagevec_add(lru_pvec, page))
1853                                 __pagevec_lru_add(lru_pvec);
1854                         *cached_page = NULL;
1855                 }
1856         }
1857         return page;
1858 }
1859
1860 /*
1861  * The logic we want is
1862  *
1863  *      if suid or (sgid and xgrp)
1864  *              remove privs
1865  */
1866 int should_remove_suid(struct dentry *dentry)
1867 {
1868         mode_t mode = dentry->d_inode->i_mode;
1869         int kill = 0;
1870
1871         /* suid always must be killed */
1872         if (unlikely(mode & S_ISUID))
1873                 kill = ATTR_KILL_SUID;
1874
1875         /*
1876          * sgid without any exec bits is just a mandatory locking mark; leave
1877          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1878          */
1879         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1880                 kill |= ATTR_KILL_SGID;
1881
1882         if (unlikely(kill && !capable(CAP_FSETID)))
1883                 return kill;
1884
1885         return 0;
1886 }
1887 EXPORT_SYMBOL(should_remove_suid);
1888
1889 int __remove_suid(struct dentry *dentry, int kill)
1890 {
1891         struct iattr newattrs;
1892
1893         newattrs.ia_valid = ATTR_FORCE | kill;
1894         return notify_change(dentry, &newattrs);
1895 }
1896
1897 int remove_suid(struct dentry *dentry)
1898 {
1899         int kill = should_remove_suid(dentry);
1900
1901         if (unlikely(kill))
1902                 return __remove_suid(dentry, kill);
1903
1904         return 0;
1905 }
1906 EXPORT_SYMBOL(remove_suid);
1907
1908 size_t
1909 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1910                         const struct iovec *iov, size_t base, size_t bytes)
1911 {
1912         size_t copied = 0, left = 0;
1913
1914         while (bytes) {
1915                 char __user *buf = iov->iov_base + base;
1916                 int copy = min(bytes, iov->iov_len - base);
1917
1918                 base = 0;
1919                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1920                 copied += copy;
1921                 bytes -= copy;
1922                 vaddr += copy;
1923                 iov++;
1924
1925                 if (unlikely(left))
1926                         break;
1927         }
1928         return copied - left;
1929 }
1930
1931 /*
1932  * Performs necessary checks before doing a write
1933  *
1934  * Can adjust writing position or amount of bytes to write.
1935  * Returns appropriate error code that caller should return or
1936  * zero in case that write should be allowed.
1937  */
1938 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1939 {
1940         struct inode *inode = file->f_mapping->host;
1941         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1942
1943         if (unlikely(*pos < 0))
1944                 return -EINVAL;
1945
1946         if (!isblk) {
1947                 /* FIXME: this is for backwards compatibility with 2.4 */
1948                 if (file->f_flags & O_APPEND)
1949                         *pos = i_size_read(inode);
1950
1951                 if (limit != RLIM_INFINITY) {
1952                         if (*pos >= limit) {
1953                                 send_sig(SIGXFSZ, current, 0);
1954                                 return -EFBIG;
1955                         }
1956                         if (*count > limit - (typeof(limit))*pos) {
1957                                 *count = limit - (typeof(limit))*pos;
1958                         }
1959                 }
1960         }
1961
1962         /*
1963          * LFS rule
1964          */
1965         if (unlikely(*pos + *count > MAX_NON_LFS &&
1966                                 !(file->f_flags & O_LARGEFILE))) {
1967                 if (*pos >= MAX_NON_LFS) {
1968                         send_sig(SIGXFSZ, current, 0);
1969                         return -EFBIG;
1970                 }
1971                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1972                         *count = MAX_NON_LFS - (unsigned long)*pos;
1973                 }
1974         }
1975
1976         /*
1977          * Are we about to exceed the fs block limit ?
1978          *
1979          * If we have written data it becomes a short write.  If we have
1980          * exceeded without writing data we send a signal and return EFBIG.
1981          * Linus frestrict idea will clean these up nicely..
1982          */
1983         if (likely(!isblk)) {
1984                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1985                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1986                                 send_sig(SIGXFSZ, current, 0);
1987                                 return -EFBIG;
1988                         }
1989                         /* zero-length writes at ->s_maxbytes are OK */
1990                 }
1991
1992                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1993                         *count = inode->i_sb->s_maxbytes - *pos;
1994         } else {
1995 #ifdef CONFIG_BLOCK
1996                 loff_t isize;
1997                 if (bdev_read_only(I_BDEV(inode)))
1998                         return -EPERM;
1999                 isize = i_size_read(inode);
2000                 if (*pos >= isize) {
2001                         if (*count || *pos > isize)
2002                                 return -ENOSPC;
2003                 }
2004
2005                 if (*pos + *count > isize)
2006                         *count = isize - *pos;
2007 #else
2008                 return -EPERM;
2009 #endif
2010         }
2011         return 0;
2012 }
2013 EXPORT_SYMBOL(generic_write_checks);
2014
2015 ssize_t
2016 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2017                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2018                 size_t count, size_t ocount)
2019 {
2020         struct file     *file = iocb->ki_filp;
2021         struct address_space *mapping = file->f_mapping;
2022         struct inode    *inode = mapping->host;
2023         ssize_t         written;
2024
2025         if (count != ocount)
2026                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2027
2028         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2029         if (written > 0) {
2030                 loff_t end = pos + written;
2031                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2032                         i_size_write(inode,  end);
2033                         mark_inode_dirty(inode);
2034                 }
2035                 *ppos = end;
2036         }
2037
2038         /*
2039          * Sync the fs metadata but not the minor inode changes and
2040          * of course not the data as we did direct DMA for the IO.
2041          * i_mutex is held, which protects generic_osync_inode() from
2042          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2043          */
2044         if ((written >= 0 || written == -EIOCBQUEUED) &&
2045             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2046                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2047                 if (err < 0)
2048                         written = err;
2049         }
2050         return written;
2051 }
2052 EXPORT_SYMBOL(generic_file_direct_write);
2053
2054 ssize_t
2055 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2056                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2057                 size_t count, ssize_t written)
2058 {
2059         struct file *file = iocb->ki_filp;
2060         struct address_space * mapping = file->f_mapping;
2061         const struct address_space_operations *a_ops = mapping->a_ops;
2062         struct inode    *inode = mapping->host;
2063         long            status = 0;
2064         struct page     *page;
2065         struct page     *cached_page = NULL;
2066         size_t          bytes;
2067         struct pagevec  lru_pvec;
2068         const struct iovec *cur_iov = iov; /* current iovec */
2069         size_t          iov_base = 0;      /* offset in the current iovec */
2070         char __user     *buf;
2071
2072         pagevec_init(&lru_pvec, 0);
2073
2074         /*
2075          * handle partial DIO write.  Adjust cur_iov if needed.
2076          */
2077         if (likely(nr_segs == 1))
2078                 buf = iov->iov_base + written;
2079         else {
2080                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
2081                 buf = cur_iov->iov_base + iov_base;
2082         }
2083
2084         do {
2085                 unsigned long index;
2086                 unsigned long offset;
2087                 size_t copied;
2088
2089                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
2090                 index = pos >> PAGE_CACHE_SHIFT;
2091                 bytes = PAGE_CACHE_SIZE - offset;
2092
2093                 /* Limit the size of the copy to the caller's write size */
2094                 bytes = min(bytes, count);
2095
2096                 /* We only need to worry about prefaulting when writes are from
2097                  * user-space.  NFSd uses vfs_writev with several non-aligned
2098                  * segments in the vector, and limiting to one segment a time is
2099                  * a noticeable performance for re-write
2100                  */
2101                 if (!segment_eq(get_fs(), KERNEL_DS)) {
2102                         /*
2103                          * Limit the size of the copy to that of the current
2104                          * segment, because fault_in_pages_readable() doesn't
2105                          * know how to walk segments.
2106                          */
2107                         bytes = min(bytes, cur_iov->iov_len - iov_base);
2108
2109                         /*
2110                          * Bring in the user page that we will copy from
2111                          * _first_.  Otherwise there's a nasty deadlock on
2112                          * copying from the same page as we're writing to,
2113                          * without it being marked up-to-date.
2114                          */
2115                         fault_in_pages_readable(buf, bytes);
2116                 }
2117                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
2118                 if (!page) {
2119                         status = -ENOMEM;
2120                         break;
2121                 }
2122
2123                 if (unlikely(bytes == 0)) {
2124                         status = 0;
2125                         copied = 0;
2126                         goto zero_length_segment;
2127                 }
2128
2129                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2130                 if (unlikely(status)) {
2131                         loff_t isize = i_size_read(inode);
2132
2133                         if (status != AOP_TRUNCATED_PAGE)
2134                                 unlock_page(page);
2135                         page_cache_release(page);
2136                         if (status == AOP_TRUNCATED_PAGE)
2137                                 continue;
2138                         /*
2139                          * prepare_write() may have instantiated a few blocks
2140                          * outside i_size.  Trim these off again.
2141                          */
2142                         if (pos + bytes > isize)
2143                                 vmtruncate(inode, isize);
2144                         break;
2145                 }
2146                 if (likely(nr_segs == 1))
2147                         copied = filemap_copy_from_user(page, offset,
2148                                                         buf, bytes);
2149                 else
2150                         copied = filemap_copy_from_user_iovec(page, offset,
2151                                                 cur_iov, iov_base, bytes);
2152                 flush_dcache_page(page);
2153                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2154                 if (status == AOP_TRUNCATED_PAGE) {
2155                         page_cache_release(page);
2156                         continue;
2157                 }
2158 zero_length_segment:
2159                 if (likely(copied >= 0)) {
2160                         if (!status)
2161                                 status = copied;
2162
2163                         if (status >= 0) {
2164                                 written += status;
2165                                 count -= status;
2166                                 pos += status;
2167                                 buf += status;
2168                                 if (unlikely(nr_segs > 1)) {
2169                                         filemap_set_next_iovec(&cur_iov,
2170                                                         &iov_base, status);
2171                                         if (count)
2172                                                 buf = cur_iov->iov_base +
2173                                                         iov_base;
2174                                 } else {
2175                                         iov_base += status;
2176                                 }
2177                         }
2178                 }
2179                 if (unlikely(copied != bytes))
2180                         if (status >= 0)
2181                                 status = -EFAULT;
2182                 unlock_page(page);
2183                 mark_page_accessed(page);
2184                 page_cache_release(page);
2185                 if (status < 0)
2186                         break;
2187                 balance_dirty_pages_ratelimited(mapping);
2188                 cond_resched();
2189         } while (count);
2190         *ppos = pos;
2191
2192         if (cached_page)
2193                 page_cache_release(cached_page);
2194
2195         /*
2196          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2197          */
2198         if (likely(status >= 0)) {
2199                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2200                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2201                                 status = generic_osync_inode(inode, mapping,
2202                                                 OSYNC_METADATA|OSYNC_DATA);
2203                 }
2204         }
2205         
2206         /*
2207          * If we get here for O_DIRECT writes then we must have fallen through
2208          * to buffered writes (block instantiation inside i_size).  So we sync
2209          * the file data here, to try to honour O_DIRECT expectations.
2210          */
2211         if (unlikely(file->f_flags & O_DIRECT) && written)
2212                 status = filemap_write_and_wait(mapping);
2213
2214         pagevec_lru_add(&lru_pvec);
2215         return written ? written : status;
2216 }
2217 EXPORT_SYMBOL(generic_file_buffered_write);
2218
2219 static ssize_t
2220 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2221                                 unsigned long nr_segs, loff_t *ppos)
2222 {
2223         struct file *file = iocb->ki_filp;
2224         struct address_space * mapping = file->f_mapping;
2225         size_t ocount;          /* original count */
2226         size_t count;           /* after file limit checks */
2227         struct inode    *inode = mapping->host;
2228         loff_t          pos;
2229         ssize_t         written;
2230         ssize_t         err;
2231
2232         ocount = 0;
2233         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2234         if (err)
2235                 return err;
2236
2237         count = ocount;
2238         pos = *ppos;
2239
2240         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2241
2242         /* We can write back this queue in page reclaim */
2243         current->backing_dev_info = mapping->backing_dev_info;
2244         written = 0;
2245
2246         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2247         if (err)
2248                 goto out;
2249
2250         if (count == 0)
2251                 goto out;
2252
2253         err = remove_suid(file->f_path.dentry);
2254         if (err)
2255                 goto out;
2256
2257         file_update_time(file);
2258
2259         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2260         if (unlikely(file->f_flags & O_DIRECT)) {
2261                 loff_t endbyte;
2262                 ssize_t written_buffered;
2263
2264                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2265                                                         ppos, count, ocount);
2266                 if (written < 0 || written == count)
2267                         goto out;
2268                 /*
2269                  * direct-io write to a hole: fall through to buffered I/O
2270                  * for completing the rest of the request.
2271                  */
2272                 pos += written;
2273                 count -= written;
2274                 written_buffered = generic_file_buffered_write(iocb, iov,
2275                                                 nr_segs, pos, ppos, count,
2276                                                 written);
2277                 /*
2278                  * If generic_file_buffered_write() retuned a synchronous error
2279                  * then we want to return the number of bytes which were
2280                  * direct-written, or the error code if that was zero.  Note
2281                  * that this differs from normal direct-io semantics, which
2282                  * will return -EFOO even if some bytes were written.
2283                  */
2284                 if (written_buffered < 0) {
2285                         err = written_buffered;
2286                         goto out;
2287                 }
2288
2289                 /*
2290                  * We need to ensure that the page cache pages are written to
2291                  * disk and invalidated to preserve the expected O_DIRECT
2292                  * semantics.
2293                  */
2294                 endbyte = pos + written_buffered - written - 1;
2295                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2296                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2297                                             SYNC_FILE_RANGE_WRITE|
2298                                             SYNC_FILE_RANGE_WAIT_AFTER);
2299                 if (err == 0) {
2300                         written = written_buffered;
2301                         invalidate_mapping_pages(mapping,
2302                                                  pos >> PAGE_CACHE_SHIFT,
2303                                                  endbyte >> PAGE_CACHE_SHIFT);
2304                 } else {
2305                         /*
2306                          * We don't know how much we wrote, so just return
2307                          * the number of bytes which were direct-written
2308                          */
2309                 }
2310         } else {
2311                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2312                                 pos, ppos, count, written);
2313         }
2314 out:
2315         current->backing_dev_info = NULL;
2316         return written ? written : err;
2317 }
2318
2319 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2320                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2321 {
2322         struct file *file = iocb->ki_filp;
2323         struct address_space *mapping = file->f_mapping;
2324         struct inode *inode = mapping->host;
2325         ssize_t ret;
2326
2327         BUG_ON(iocb->ki_pos != pos);
2328
2329         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2330                         &iocb->ki_pos);
2331
2332         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2333                 ssize_t err;
2334
2335                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2336                 if (err < 0)
2337                         ret = err;
2338         }
2339         return ret;
2340 }
2341 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2342
2343 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2344                 unsigned long nr_segs, loff_t pos)
2345 {
2346         struct file *file = iocb->ki_filp;
2347         struct address_space *mapping = file->f_mapping;
2348         struct inode *inode = mapping->host;
2349         ssize_t ret;
2350
2351         BUG_ON(iocb->ki_pos != pos);
2352
2353         mutex_lock(&inode->i_mutex);
2354         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2355                         &iocb->ki_pos);
2356         mutex_unlock(&inode->i_mutex);
2357
2358         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2359                 ssize_t err;
2360
2361                 err = sync_page_range(inode, mapping, pos, ret);
2362                 if (err < 0)
2363                         ret = err;
2364         }
2365         return ret;
2366 }
2367 EXPORT_SYMBOL(generic_file_aio_write);
2368
2369 /*
2370  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2371  * went wrong during pagecache shootdown.
2372  */
2373 static ssize_t
2374 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2375         loff_t offset, unsigned long nr_segs)
2376 {
2377         struct file *file = iocb->ki_filp;
2378         struct address_space *mapping = file->f_mapping;
2379         ssize_t retval;
2380         size_t write_len;
2381         pgoff_t end = 0; /* silence gcc */
2382
2383         /*
2384          * If it's a write, unmap all mmappings of the file up-front.  This
2385          * will cause any pte dirty bits to be propagated into the pageframes
2386          * for the subsequent filemap_write_and_wait().
2387          */
2388         if (rw == WRITE) {
2389                 write_len = iov_length(iov, nr_segs);
2390                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2391                 if (mapping_mapped(mapping))
2392                         unmap_mapping_range(mapping, offset, write_len, 0);
2393         }
2394
2395         retval = filemap_write_and_wait(mapping);
2396         if (retval)
2397                 goto out;
2398
2399         /*
2400          * After a write we want buffered reads to be sure to go to disk to get
2401          * the new data.  We invalidate clean cached page from the region we're
2402          * about to write.  We do this *before* the write so that we can return
2403          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2404          */
2405         if (rw == WRITE && mapping->nrpages) {
2406                 retval = invalidate_inode_pages2_range(mapping,
2407                                         offset >> PAGE_CACHE_SHIFT, end);
2408                 if (retval)
2409                         goto out;
2410         }
2411
2412         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2413         if (retval)
2414                 goto out;
2415
2416         /*
2417          * Finally, try again to invalidate clean pages which might have been
2418          * faulted in by get_user_pages() if the source of the write was an
2419          * mmap()ed region of the file we're writing.  That's a pretty crazy
2420          * thing to do, so we don't support it 100%.  If this invalidation
2421          * fails and we have -EIOCBQUEUED we ignore the failure.
2422          */
2423         if (rw == WRITE && mapping->nrpages) {
2424                 int err = invalidate_inode_pages2_range(mapping,
2425                                               offset >> PAGE_CACHE_SHIFT, end);
2426                 if (err && retval >= 0)
2427                         retval = err;
2428         }
2429 out:
2430         return retval;
2431 }
2432
2433 /**
2434  * try_to_release_page() - release old fs-specific metadata on a page
2435  *
2436  * @page: the page which the kernel is trying to free
2437  * @gfp_mask: memory allocation flags (and I/O mode)
2438  *
2439  * The address_space is to try to release any data against the page
2440  * (presumably at page->private).  If the release was successful, return `1'.
2441  * Otherwise return zero.
2442  *
2443  * The @gfp_mask argument specifies whether I/O may be performed to release
2444  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2445  *
2446  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2447  */
2448 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2449 {
2450         struct address_space * const mapping = page->mapping;
2451
2452         BUG_ON(!PageLocked(page));
2453         if (PageWriteback(page))
2454                 return 0;
2455
2456         if (mapping && mapping->a_ops->releasepage)
2457                 return mapping->a_ops->releasepage(page, gfp_mask);
2458         return try_to_free_buffers(page);
2459 }
2460
2461 EXPORT_SYMBOL(try_to_release_page);