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