[PATCH] grab swap token reordered
[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_trylock_page - find and lock a page
610  * @mapping: the address_space to search
611  * @offset: the page index
612  *
613  * Same as find_get_page(), but trylock it instead of incrementing the count.
614  */
615 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
616 {
617         struct page *page;
618
619         read_lock_irq(&mapping->tree_lock);
620         page = radix_tree_lookup(&mapping->page_tree, offset);
621         if (page && TestSetPageLocked(page))
622                 page = NULL;
623         read_unlock_irq(&mapping->tree_lock);
624         return page;
625 }
626 EXPORT_SYMBOL(find_trylock_page);
627
628 /**
629  * find_lock_page - locate, pin and lock a pagecache page
630  * @mapping: the address_space to search
631  * @offset: the page index
632  *
633  * Locates the desired pagecache page, locks it, increments its reference
634  * count and returns its address.
635  *
636  * Returns zero if the page was not present. find_lock_page() may sleep.
637  */
638 struct page *find_lock_page(struct address_space *mapping,
639                                 unsigned long offset)
640 {
641         struct page *page;
642
643         read_lock_irq(&mapping->tree_lock);
644 repeat:
645         page = radix_tree_lookup(&mapping->page_tree, offset);
646         if (page) {
647                 page_cache_get(page);
648                 if (TestSetPageLocked(page)) {
649                         read_unlock_irq(&mapping->tree_lock);
650                         __lock_page(page);
651                         read_lock_irq(&mapping->tree_lock);
652
653                         /* Has the page been truncated while we slept? */
654                         if (unlikely(page->mapping != mapping ||
655                                      page->index != offset)) {
656                                 unlock_page(page);
657                                 page_cache_release(page);
658                                 goto repeat;
659                         }
660                 }
661         }
662         read_unlock_irq(&mapping->tree_lock);
663         return page;
664 }
665 EXPORT_SYMBOL(find_lock_page);
666
667 /**
668  * find_or_create_page - locate or add a pagecache page
669  * @mapping: the page's address_space
670  * @index: the page's index into the mapping
671  * @gfp_mask: page allocation mode
672  *
673  * Locates a page in the pagecache.  If the page is not present, a new page
674  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
675  * LRU list.  The returned page is locked and has its reference count
676  * incremented.
677  *
678  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
679  * allocation!
680  *
681  * find_or_create_page() returns the desired page's address, or zero on
682  * memory exhaustion.
683  */
684 struct page *find_or_create_page(struct address_space *mapping,
685                 unsigned long index, gfp_t gfp_mask)
686 {
687         struct page *page, *cached_page = NULL;
688         int err;
689 repeat:
690         page = find_lock_page(mapping, index);
691         if (!page) {
692                 if (!cached_page) {
693                         cached_page = alloc_page(gfp_mask);
694                         if (!cached_page)
695                                 return NULL;
696                 }
697                 err = add_to_page_cache_lru(cached_page, mapping,
698                                         index, gfp_mask);
699                 if (!err) {
700                         page = cached_page;
701                         cached_page = NULL;
702                 } else if (err == -EEXIST)
703                         goto repeat;
704         }
705         if (cached_page)
706                 page_cache_release(cached_page);
707         return page;
708 }
709 EXPORT_SYMBOL(find_or_create_page);
710
711 /**
712  * find_get_pages - gang pagecache lookup
713  * @mapping:    The address_space to search
714  * @start:      The starting page index
715  * @nr_pages:   The maximum number of pages
716  * @pages:      Where the resulting pages are placed
717  *
718  * find_get_pages() will search for and return a group of up to
719  * @nr_pages pages in the mapping.  The pages are placed at @pages.
720  * find_get_pages() takes a reference against the returned pages.
721  *
722  * The search returns a group of mapping-contiguous pages with ascending
723  * indexes.  There may be holes in the indices due to not-present pages.
724  *
725  * find_get_pages() returns the number of pages which were found.
726  */
727 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
728                             unsigned int nr_pages, struct page **pages)
729 {
730         unsigned int i;
731         unsigned int ret;
732
733         read_lock_irq(&mapping->tree_lock);
734         ret = radix_tree_gang_lookup(&mapping->page_tree,
735                                 (void **)pages, start, nr_pages);
736         for (i = 0; i < ret; i++)
737                 page_cache_get(pages[i]);
738         read_unlock_irq(&mapping->tree_lock);
739         return ret;
740 }
741
742 /**
743  * find_get_pages_contig - gang contiguous pagecache lookup
744  * @mapping:    The address_space to search
745  * @index:      The starting page index
746  * @nr_pages:   The maximum number of pages
747  * @pages:      Where the resulting pages are placed
748  *
749  * find_get_pages_contig() works exactly like find_get_pages(), except
750  * that the returned number of pages are guaranteed to be contiguous.
751  *
752  * find_get_pages_contig() returns the number of pages which were found.
753  */
754 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
755                                unsigned int nr_pages, struct page **pages)
756 {
757         unsigned int i;
758         unsigned int ret;
759
760         read_lock_irq(&mapping->tree_lock);
761         ret = radix_tree_gang_lookup(&mapping->page_tree,
762                                 (void **)pages, index, nr_pages);
763         for (i = 0; i < ret; i++) {
764                 if (pages[i]->mapping == NULL || pages[i]->index != index)
765                         break;
766
767                 page_cache_get(pages[i]);
768                 index++;
769         }
770         read_unlock_irq(&mapping->tree_lock);
771         return i;
772 }
773
774 /**
775  * find_get_pages_tag - find and return pages that match @tag
776  * @mapping:    the address_space to search
777  * @index:      the starting page index
778  * @tag:        the tag index
779  * @nr_pages:   the maximum number of pages
780  * @pages:      where the resulting pages are placed
781  *
782  * Like find_get_pages, except we only return pages which are tagged with
783  * @tag.   We update @index to index the next page for the traversal.
784  */
785 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
786                         int tag, unsigned int nr_pages, struct page **pages)
787 {
788         unsigned int i;
789         unsigned int ret;
790
791         read_lock_irq(&mapping->tree_lock);
792         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
793                                 (void **)pages, *index, nr_pages, tag);
794         for (i = 0; i < ret; i++)
795                 page_cache_get(pages[i]);
796         if (ret)
797                 *index = pages[ret - 1]->index + 1;
798         read_unlock_irq(&mapping->tree_lock);
799         return ret;
800 }
801
802 /**
803  * grab_cache_page_nowait - returns locked page at given index in given cache
804  * @mapping: target address_space
805  * @index: the page index
806  *
807  * Same as grab_cache_page, but do not wait if the page is unavailable.
808  * This is intended for speculative data generators, where the data can
809  * be regenerated if the page couldn't be grabbed.  This routine should
810  * be safe to call while holding the lock for another page.
811  *
812  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
813  * and deadlock against the caller's locked page.
814  */
815 struct page *
816 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
817 {
818         struct page *page = find_get_page(mapping, index);
819
820         if (page) {
821                 if (!TestSetPageLocked(page))
822                         return page;
823                 page_cache_release(page);
824                 return NULL;
825         }
826         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
827         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
828                 page_cache_release(page);
829                 page = NULL;
830         }
831         return page;
832 }
833 EXPORT_SYMBOL(grab_cache_page_nowait);
834
835 /*
836  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
837  * a _large_ part of the i/o request. Imagine the worst scenario:
838  *
839  *      ---R__________________________________________B__________
840  *         ^ reading here                             ^ bad block(assume 4k)
841  *
842  * read(R) => miss => readahead(R...B) => media error => frustrating retries
843  * => failing the whole request => read(R) => read(R+1) =>
844  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
845  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
846  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
847  *
848  * It is going insane. Fix it by quickly scaling down the readahead size.
849  */
850 static void shrink_readahead_size_eio(struct file *filp,
851                                         struct file_ra_state *ra)
852 {
853         if (!ra->ra_pages)
854                 return;
855
856         ra->ra_pages /= 4;
857 }
858
859 /**
860  * do_generic_mapping_read - generic file read routine
861  * @mapping:    address_space to be read
862  * @_ra:        file's readahead state
863  * @filp:       the file to read
864  * @ppos:       current file position
865  * @desc:       read_descriptor
866  * @actor:      read method
867  *
868  * This is a generic file read routine, and uses the
869  * mapping->a_ops->readpage() function for the actual low-level stuff.
870  *
871  * This is really ugly. But the goto's actually try to clarify some
872  * of the logic when it comes to error handling etc.
873  *
874  * Note the struct file* is only passed for the use of readpage.
875  * It may be NULL.
876  */
877 void do_generic_mapping_read(struct address_space *mapping,
878                              struct file_ra_state *_ra,
879                              struct file *filp,
880                              loff_t *ppos,
881                              read_descriptor_t *desc,
882                              read_actor_t actor)
883 {
884         struct inode *inode = mapping->host;
885         unsigned long index;
886         unsigned long end_index;
887         unsigned long offset;
888         unsigned long last_index;
889         unsigned long next_index;
890         unsigned long prev_index;
891         loff_t isize;
892         struct page *cached_page;
893         int error;
894         struct file_ra_state ra = *_ra;
895
896         cached_page = NULL;
897         index = *ppos >> PAGE_CACHE_SHIFT;
898         next_index = index;
899         prev_index = ra.prev_page;
900         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
901         offset = *ppos & ~PAGE_CACHE_MASK;
902
903         isize = i_size_read(inode);
904         if (!isize)
905                 goto out;
906
907         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
908         for (;;) {
909                 struct page *page;
910                 unsigned long nr, ret;
911
912                 /* nr is the maximum number of bytes to copy from this page */
913                 nr = PAGE_CACHE_SIZE;
914                 if (index >= end_index) {
915                         if (index > end_index)
916                                 goto out;
917                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
918                         if (nr <= offset) {
919                                 goto out;
920                         }
921                 }
922                 nr = nr - offset;
923
924                 cond_resched();
925                 if (index == next_index)
926                         next_index = page_cache_readahead(mapping, &ra, filp,
927                                         index, last_index - index);
928
929 find_page:
930                 page = find_get_page(mapping, index);
931                 if (unlikely(page == NULL)) {
932                         handle_ra_miss(mapping, &ra, index);
933                         goto no_cached_page;
934                 }
935                 if (!PageUptodate(page))
936                         goto page_not_up_to_date;
937 page_ok:
938
939                 /* If users can be writing to this page using arbitrary
940                  * virtual addresses, take care about potential aliasing
941                  * before reading the page on the kernel side.
942                  */
943                 if (mapping_writably_mapped(mapping))
944                         flush_dcache_page(page);
945
946                 /*
947                  * When (part of) the same page is read multiple times
948                  * in succession, only mark it as accessed the first time.
949                  */
950                 if (prev_index != index)
951                         mark_page_accessed(page);
952                 prev_index = index;
953
954                 /*
955                  * Ok, we have the page, and it's up-to-date, so
956                  * now we can copy it to user space...
957                  *
958                  * The actor routine returns how many bytes were actually used..
959                  * NOTE! This may not be the same as how much of a user buffer
960                  * we filled up (we may be padding etc), so we can only update
961                  * "pos" here (the actor routine has to update the user buffer
962                  * pointers and the remaining count).
963                  */
964                 ret = actor(desc, page, offset, nr);
965                 offset += ret;
966                 index += offset >> PAGE_CACHE_SHIFT;
967                 offset &= ~PAGE_CACHE_MASK;
968
969                 page_cache_release(page);
970                 if (ret == nr && desc->count)
971                         continue;
972                 goto out;
973
974 page_not_up_to_date:
975                 /* Get exclusive access to the page ... */
976                 lock_page(page);
977
978                 /* Did it get truncated before we got the lock? */
979                 if (!page->mapping) {
980                         unlock_page(page);
981                         page_cache_release(page);
982                         continue;
983                 }
984
985                 /* Did somebody else fill it already? */
986                 if (PageUptodate(page)) {
987                         unlock_page(page);
988                         goto page_ok;
989                 }
990
991 readpage:
992                 /* Start the actual read. The read will unlock the page. */
993                 error = mapping->a_ops->readpage(filp, page);
994
995                 if (unlikely(error)) {
996                         if (error == AOP_TRUNCATED_PAGE) {
997                                 page_cache_release(page);
998                                 goto find_page;
999                         }
1000                         goto readpage_error;
1001                 }
1002
1003                 if (!PageUptodate(page)) {
1004                         lock_page(page);
1005                         if (!PageUptodate(page)) {
1006                                 if (page->mapping == NULL) {
1007                                         /*
1008                                          * invalidate_inode_pages got it
1009                                          */
1010                                         unlock_page(page);
1011                                         page_cache_release(page);
1012                                         goto find_page;
1013                                 }
1014                                 unlock_page(page);
1015                                 error = -EIO;
1016                                 shrink_readahead_size_eio(filp, &ra);
1017                                 goto readpage_error;
1018                         }
1019                         unlock_page(page);
1020                 }
1021
1022                 /*
1023                  * i_size must be checked after we have done ->readpage.
1024                  *
1025                  * Checking i_size after the readpage allows us to calculate
1026                  * the correct value for "nr", which means the zero-filled
1027                  * part of the page is not copied back to userspace (unless
1028                  * another truncate extends the file - this is desired though).
1029                  */
1030                 isize = i_size_read(inode);
1031                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1032                 if (unlikely(!isize || index > end_index)) {
1033                         page_cache_release(page);
1034                         goto out;
1035                 }
1036
1037                 /* nr is the maximum number of bytes to copy from this page */
1038                 nr = PAGE_CACHE_SIZE;
1039                 if (index == end_index) {
1040                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1041                         if (nr <= offset) {
1042                                 page_cache_release(page);
1043                                 goto out;
1044                         }
1045                 }
1046                 nr = nr - offset;
1047                 goto page_ok;
1048
1049 readpage_error:
1050                 /* UHHUH! A synchronous read error occurred. Report it */
1051                 desc->error = error;
1052                 page_cache_release(page);
1053                 goto out;
1054
1055 no_cached_page:
1056                 /*
1057                  * Ok, it wasn't cached, so we need to create a new
1058                  * page..
1059                  */
1060                 if (!cached_page) {
1061                         cached_page = page_cache_alloc_cold(mapping);
1062                         if (!cached_page) {
1063                                 desc->error = -ENOMEM;
1064                                 goto out;
1065                         }
1066                 }
1067                 error = add_to_page_cache_lru(cached_page, mapping,
1068                                                 index, GFP_KERNEL);
1069                 if (error) {
1070                         if (error == -EEXIST)
1071                                 goto find_page;
1072                         desc->error = error;
1073                         goto out;
1074                 }
1075                 page = cached_page;
1076                 cached_page = NULL;
1077                 goto readpage;
1078         }
1079
1080 out:
1081         *_ra = ra;
1082
1083         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1084         if (cached_page)
1085                 page_cache_release(cached_page);
1086         if (filp)
1087                 file_accessed(filp);
1088 }
1089 EXPORT_SYMBOL(do_generic_mapping_read);
1090
1091 int file_read_actor(read_descriptor_t *desc, struct page *page,
1092                         unsigned long offset, unsigned long size)
1093 {
1094         char *kaddr;
1095         unsigned long left, count = desc->count;
1096
1097         if (size > count)
1098                 size = count;
1099
1100         /*
1101          * Faults on the destination of a read are common, so do it before
1102          * taking the kmap.
1103          */
1104         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1105                 kaddr = kmap_atomic(page, KM_USER0);
1106                 left = __copy_to_user_inatomic(desc->arg.buf,
1107                                                 kaddr + offset, size);
1108                 kunmap_atomic(kaddr, KM_USER0);
1109                 if (left == 0)
1110                         goto success;
1111         }
1112
1113         /* Do it the slow way */
1114         kaddr = kmap(page);
1115         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1116         kunmap(page);
1117
1118         if (left) {
1119                 size -= left;
1120                 desc->error = -EFAULT;
1121         }
1122 success:
1123         desc->count = count - size;
1124         desc->written += size;
1125         desc->arg.buf += size;
1126         return size;
1127 }
1128
1129 /**
1130  * generic_file_aio_read - generic filesystem read routine
1131  * @iocb:       kernel I/O control block
1132  * @iov:        io vector request
1133  * @nr_segs:    number of segments in the iovec
1134  * @pos:        current file position
1135  *
1136  * This is the "read()" routine for all filesystems
1137  * that can use the page cache directly.
1138  */
1139 ssize_t
1140 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1141                 unsigned long nr_segs, loff_t pos)
1142 {
1143         struct file *filp = iocb->ki_filp;
1144         ssize_t retval;
1145         unsigned long seg;
1146         size_t count;
1147         loff_t *ppos = &iocb->ki_pos;
1148
1149         count = 0;
1150         for (seg = 0; seg < nr_segs; seg++) {
1151                 const struct iovec *iv = &iov[seg];
1152
1153                 /*
1154                  * If any segment has a negative length, or the cumulative
1155                  * length ever wraps negative then return -EINVAL.
1156                  */
1157                 count += iv->iov_len;
1158                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1159                         return -EINVAL;
1160                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1161                         continue;
1162                 if (seg == 0)
1163                         return -EFAULT;
1164                 nr_segs = seg;
1165                 count -= iv->iov_len;   /* This segment is no good */
1166                 break;
1167         }
1168
1169         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1170         if (filp->f_flags & O_DIRECT) {
1171                 loff_t size;
1172                 struct address_space *mapping;
1173                 struct inode *inode;
1174
1175                 mapping = filp->f_mapping;
1176                 inode = mapping->host;
1177                 retval = 0;
1178                 if (!count)
1179                         goto out; /* skip atime */
1180                 size = i_size_read(inode);
1181                 if (pos < size) {
1182                         retval = generic_file_direct_IO(READ, iocb,
1183                                                 iov, pos, nr_segs);
1184                         if (retval > 0 && !is_sync_kiocb(iocb))
1185                                 retval = -EIOCBQUEUED;
1186                         if (retval > 0)
1187                                 *ppos = pos + retval;
1188                 }
1189                 if (likely(retval != 0)) {
1190                         file_accessed(filp);
1191                         goto out;
1192                 }
1193         }
1194
1195         retval = 0;
1196         if (count) {
1197                 for (seg = 0; seg < nr_segs; seg++) {
1198                         read_descriptor_t desc;
1199
1200                         desc.written = 0;
1201                         desc.arg.buf = iov[seg].iov_base;
1202                         desc.count = iov[seg].iov_len;
1203                         if (desc.count == 0)
1204                                 continue;
1205                         desc.error = 0;
1206                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1207                         retval += desc.written;
1208                         if (desc.error) {
1209                                 retval = retval ?: desc.error;
1210                                 break;
1211                         }
1212                 }
1213         }
1214 out:
1215         return retval;
1216 }
1217 EXPORT_SYMBOL(generic_file_aio_read);
1218
1219 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1220 {
1221         ssize_t written;
1222         unsigned long count = desc->count;
1223         struct file *file = desc->arg.data;
1224
1225         if (size > count)
1226                 size = count;
1227
1228         written = file->f_op->sendpage(file, page, offset,
1229                                        size, &file->f_pos, size<count);
1230         if (written < 0) {
1231                 desc->error = written;
1232                 written = 0;
1233         }
1234         desc->count = count - written;
1235         desc->written += written;
1236         return written;
1237 }
1238
1239 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1240                          size_t count, read_actor_t actor, void *target)
1241 {
1242         read_descriptor_t desc;
1243
1244         if (!count)
1245                 return 0;
1246
1247         desc.written = 0;
1248         desc.count = count;
1249         desc.arg.data = target;
1250         desc.error = 0;
1251
1252         do_generic_file_read(in_file, ppos, &desc, actor);
1253         if (desc.written)
1254                 return desc.written;
1255         return desc.error;
1256 }
1257 EXPORT_SYMBOL(generic_file_sendfile);
1258
1259 static ssize_t
1260 do_readahead(struct address_space *mapping, struct file *filp,
1261              unsigned long index, unsigned long nr)
1262 {
1263         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1264                 return -EINVAL;
1265
1266         force_page_cache_readahead(mapping, filp, index,
1267                                         max_sane_readahead(nr));
1268         return 0;
1269 }
1270
1271 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1272 {
1273         ssize_t ret;
1274         struct file *file;
1275
1276         ret = -EBADF;
1277         file = fget(fd);
1278         if (file) {
1279                 if (file->f_mode & FMODE_READ) {
1280                         struct address_space *mapping = file->f_mapping;
1281                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1282                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1283                         unsigned long len = end - start + 1;
1284                         ret = do_readahead(mapping, file, start, len);
1285                 }
1286                 fput(file);
1287         }
1288         return ret;
1289 }
1290
1291 #ifdef CONFIG_MMU
1292 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1293 /**
1294  * page_cache_read - adds requested page to the page cache if not already there
1295  * @file:       file to read
1296  * @offset:     page index
1297  *
1298  * This adds the requested page to the page cache if it isn't already there,
1299  * and schedules an I/O to read in its contents from disk.
1300  */
1301 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1302 {
1303         struct address_space *mapping = file->f_mapping;
1304         struct page *page; 
1305         int ret;
1306
1307         do {
1308                 page = page_cache_alloc_cold(mapping);
1309                 if (!page)
1310                         return -ENOMEM;
1311
1312                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1313                 if (ret == 0)
1314                         ret = mapping->a_ops->readpage(file, page);
1315                 else if (ret == -EEXIST)
1316                         ret = 0; /* losing race to add is OK */
1317
1318                 page_cache_release(page);
1319
1320         } while (ret == AOP_TRUNCATED_PAGE);
1321                 
1322         return ret;
1323 }
1324
1325 #define MMAP_LOTSAMISS  (100)
1326
1327 /**
1328  * filemap_nopage - read in file data for page fault handling
1329  * @area:       the applicable vm_area
1330  * @address:    target address to read in
1331  * @type:       returned with VM_FAULT_{MINOR,MAJOR} if not %NULL
1332  *
1333  * filemap_nopage() is invoked via the vma operations vector for a
1334  * mapped memory region to read in file data during a page fault.
1335  *
1336  * The goto's are kind of ugly, but this streamlines the normal case of having
1337  * it in the page cache, and handles the special cases reasonably without
1338  * having a lot of duplicated code.
1339  */
1340 struct page *filemap_nopage(struct vm_area_struct *area,
1341                                 unsigned long address, int *type)
1342 {
1343         int error;
1344         struct file *file = area->vm_file;
1345         struct address_space *mapping = file->f_mapping;
1346         struct file_ra_state *ra = &file->f_ra;
1347         struct inode *inode = mapping->host;
1348         struct page *page;
1349         unsigned long size, pgoff;
1350         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1351
1352         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1353
1354 retry_all:
1355         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1356         if (pgoff >= size)
1357                 goto outside_data_content;
1358
1359         /* If we don't want any read-ahead, don't bother */
1360         if (VM_RandomReadHint(area))
1361                 goto no_cached_page;
1362
1363         /*
1364          * The readahead code wants to be told about each and every page
1365          * so it can build and shrink its windows appropriately
1366          *
1367          * For sequential accesses, we use the generic readahead logic.
1368          */
1369         if (VM_SequentialReadHint(area))
1370                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1371
1372         /*
1373          * Do we have something in the page cache already?
1374          */
1375 retry_find:
1376         page = find_get_page(mapping, pgoff);
1377         if (!page) {
1378                 unsigned long ra_pages;
1379
1380                 if (VM_SequentialReadHint(area)) {
1381                         handle_ra_miss(mapping, ra, pgoff);
1382                         goto no_cached_page;
1383                 }
1384                 ra->mmap_miss++;
1385
1386                 /*
1387                  * Do we miss much more than hit in this file? If so,
1388                  * stop bothering with read-ahead. It will only hurt.
1389                  */
1390                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1391                         goto no_cached_page;
1392
1393                 /*
1394                  * To keep the pgmajfault counter straight, we need to
1395                  * check did_readaround, as this is an inner loop.
1396                  */
1397                 if (!did_readaround) {
1398                         majmin = VM_FAULT_MAJOR;
1399                         count_vm_event(PGMAJFAULT);
1400                 }
1401                 did_readaround = 1;
1402                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1403                 if (ra_pages) {
1404                         pgoff_t start = 0;
1405
1406                         if (pgoff > ra_pages / 2)
1407                                 start = pgoff - ra_pages / 2;
1408                         do_page_cache_readahead(mapping, file, start, ra_pages);
1409                 }
1410                 page = find_get_page(mapping, pgoff);
1411                 if (!page)
1412                         goto no_cached_page;
1413         }
1414
1415         if (!did_readaround)
1416                 ra->mmap_hit++;
1417
1418         /*
1419          * Ok, found a page in the page cache, now we need to check
1420          * that it's up-to-date.
1421          */
1422         if (!PageUptodate(page))
1423                 goto page_not_uptodate;
1424
1425 success:
1426         /*
1427          * Found the page and have a reference on it.
1428          */
1429         mark_page_accessed(page);
1430         if (type)
1431                 *type = majmin;
1432         return page;
1433
1434 outside_data_content:
1435         /*
1436          * An external ptracer can access pages that normally aren't
1437          * accessible..
1438          */
1439         if (area->vm_mm == current->mm)
1440                 return NOPAGE_SIGBUS;
1441         /* Fall through to the non-read-ahead case */
1442 no_cached_page:
1443         /*
1444          * We're only likely to ever get here if MADV_RANDOM is in
1445          * effect.
1446          */
1447         error = page_cache_read(file, pgoff);
1448
1449         /*
1450          * The page we want has now been added to the page cache.
1451          * In the unlikely event that someone removed it in the
1452          * meantime, we'll just come back here and read it again.
1453          */
1454         if (error >= 0)
1455                 goto retry_find;
1456
1457         /*
1458          * An error return from page_cache_read can result if the
1459          * system is low on memory, or a problem occurs while trying
1460          * to schedule I/O.
1461          */
1462         if (error == -ENOMEM)
1463                 return NOPAGE_OOM;
1464         return NOPAGE_SIGBUS;
1465
1466 page_not_uptodate:
1467         if (!did_readaround) {
1468                 majmin = VM_FAULT_MAJOR;
1469                 count_vm_event(PGMAJFAULT);
1470         }
1471         lock_page(page);
1472
1473         /* Did it get unhashed while we waited for it? */
1474         if (!page->mapping) {
1475                 unlock_page(page);
1476                 page_cache_release(page);
1477                 goto retry_all;
1478         }
1479
1480         /* Did somebody else get it up-to-date? */
1481         if (PageUptodate(page)) {
1482                 unlock_page(page);
1483                 goto success;
1484         }
1485
1486         error = mapping->a_ops->readpage(file, page);
1487         if (!error) {
1488                 wait_on_page_locked(page);
1489                 if (PageUptodate(page))
1490                         goto success;
1491         } else if (error == AOP_TRUNCATED_PAGE) {
1492                 page_cache_release(page);
1493                 goto retry_find;
1494         }
1495
1496         /*
1497          * Umm, take care of errors if the page isn't up-to-date.
1498          * Try to re-read it _once_. We do this synchronously,
1499          * because there really aren't any performance issues here
1500          * and we need to check for errors.
1501          */
1502         lock_page(page);
1503
1504         /* Somebody truncated the page on us? */
1505         if (!page->mapping) {
1506                 unlock_page(page);
1507                 page_cache_release(page);
1508                 goto retry_all;
1509         }
1510
1511         /* Somebody else successfully read it in? */
1512         if (PageUptodate(page)) {
1513                 unlock_page(page);
1514                 goto success;
1515         }
1516         ClearPageError(page);
1517         error = mapping->a_ops->readpage(file, page);
1518         if (!error) {
1519                 wait_on_page_locked(page);
1520                 if (PageUptodate(page))
1521                         goto success;
1522         } else if (error == AOP_TRUNCATED_PAGE) {
1523                 page_cache_release(page);
1524                 goto retry_find;
1525         }
1526
1527         /*
1528          * Things didn't work out. Return zero to tell the
1529          * mm layer so, possibly freeing the page cache page first.
1530          */
1531         shrink_readahead_size_eio(file, ra);
1532         page_cache_release(page);
1533         return NOPAGE_SIGBUS;
1534 }
1535 EXPORT_SYMBOL(filemap_nopage);
1536
1537 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1538                                         int nonblock)
1539 {
1540         struct address_space *mapping = file->f_mapping;
1541         struct page *page;
1542         int error;
1543
1544         /*
1545          * Do we have something in the page cache already?
1546          */
1547 retry_find:
1548         page = find_get_page(mapping, pgoff);
1549         if (!page) {
1550                 if (nonblock)
1551                         return NULL;
1552                 goto no_cached_page;
1553         }
1554
1555         /*
1556          * Ok, found a page in the page cache, now we need to check
1557          * that it's up-to-date.
1558          */
1559         if (!PageUptodate(page)) {
1560                 if (nonblock) {
1561                         page_cache_release(page);
1562                         return NULL;
1563                 }
1564                 goto page_not_uptodate;
1565         }
1566
1567 success:
1568         /*
1569          * Found the page and have a reference on it.
1570          */
1571         mark_page_accessed(page);
1572         return page;
1573
1574 no_cached_page:
1575         error = page_cache_read(file, pgoff);
1576
1577         /*
1578          * The page we want has now been added to the page cache.
1579          * In the unlikely event that someone removed it in the
1580          * meantime, we'll just come back here and read it again.
1581          */
1582         if (error >= 0)
1583                 goto retry_find;
1584
1585         /*
1586          * An error return from page_cache_read can result if the
1587          * system is low on memory, or a problem occurs while trying
1588          * to schedule I/O.
1589          */
1590         return NULL;
1591
1592 page_not_uptodate:
1593         lock_page(page);
1594
1595         /* Did it get truncated while we waited for it? */
1596         if (!page->mapping) {
1597                 unlock_page(page);
1598                 goto err;
1599         }
1600
1601         /* Did somebody else get it up-to-date? */
1602         if (PageUptodate(page)) {
1603                 unlock_page(page);
1604                 goto success;
1605         }
1606
1607         error = mapping->a_ops->readpage(file, page);
1608         if (!error) {
1609                 wait_on_page_locked(page);
1610                 if (PageUptodate(page))
1611                         goto success;
1612         } else if (error == AOP_TRUNCATED_PAGE) {
1613                 page_cache_release(page);
1614                 goto retry_find;
1615         }
1616
1617         /*
1618          * Umm, take care of errors if the page isn't up-to-date.
1619          * Try to re-read it _once_. We do this synchronously,
1620          * because there really aren't any performance issues here
1621          * and we need to check for errors.
1622          */
1623         lock_page(page);
1624
1625         /* Somebody truncated the page on us? */
1626         if (!page->mapping) {
1627                 unlock_page(page);
1628                 goto err;
1629         }
1630         /* Somebody else successfully read it in? */
1631         if (PageUptodate(page)) {
1632                 unlock_page(page);
1633                 goto success;
1634         }
1635
1636         ClearPageError(page);
1637         error = mapping->a_ops->readpage(file, page);
1638         if (!error) {
1639                 wait_on_page_locked(page);
1640                 if (PageUptodate(page))
1641                         goto success;
1642         } else if (error == AOP_TRUNCATED_PAGE) {
1643                 page_cache_release(page);
1644                 goto retry_find;
1645         }
1646
1647         /*
1648          * Things didn't work out. Return zero to tell the
1649          * mm layer so, possibly freeing the page cache page first.
1650          */
1651 err:
1652         page_cache_release(page);
1653
1654         return NULL;
1655 }
1656
1657 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1658                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1659                 int nonblock)
1660 {
1661         struct file *file = vma->vm_file;
1662         struct address_space *mapping = file->f_mapping;
1663         struct inode *inode = mapping->host;
1664         unsigned long size;
1665         struct mm_struct *mm = vma->vm_mm;
1666         struct page *page;
1667         int err;
1668
1669         if (!nonblock)
1670                 force_page_cache_readahead(mapping, vma->vm_file,
1671                                         pgoff, len >> PAGE_CACHE_SHIFT);
1672
1673 repeat:
1674         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1675         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1676                 return -EINVAL;
1677
1678         page = filemap_getpage(file, pgoff, nonblock);
1679
1680         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1681          * done in shmem_populate calling shmem_getpage */
1682         if (!page && !nonblock)
1683                 return -ENOMEM;
1684
1685         if (page) {
1686                 err = install_page(mm, vma, addr, page, prot);
1687                 if (err) {
1688                         page_cache_release(page);
1689                         return err;
1690                 }
1691         } else if (vma->vm_flags & VM_NONLINEAR) {
1692                 /* No page was found just because we can't read it in now (being
1693                  * here implies nonblock != 0), but the page may exist, so set
1694                  * the PTE to fault it in later. */
1695                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1696                 if (err)
1697                         return err;
1698         }
1699
1700         len -= PAGE_SIZE;
1701         addr += PAGE_SIZE;
1702         pgoff++;
1703         if (len)
1704                 goto repeat;
1705
1706         return 0;
1707 }
1708 EXPORT_SYMBOL(filemap_populate);
1709
1710 struct vm_operations_struct generic_file_vm_ops = {
1711         .nopage         = filemap_nopage,
1712         .populate       = filemap_populate,
1713 };
1714
1715 /* This is used for a general mmap of a disk file */
1716
1717 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1718 {
1719         struct address_space *mapping = file->f_mapping;
1720
1721         if (!mapping->a_ops->readpage)
1722                 return -ENOEXEC;
1723         file_accessed(file);
1724         vma->vm_ops = &generic_file_vm_ops;
1725         return 0;
1726 }
1727
1728 /*
1729  * This is for filesystems which do not implement ->writepage.
1730  */
1731 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1732 {
1733         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1734                 return -EINVAL;
1735         return generic_file_mmap(file, vma);
1736 }
1737 #else
1738 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1739 {
1740         return -ENOSYS;
1741 }
1742 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1743 {
1744         return -ENOSYS;
1745 }
1746 #endif /* CONFIG_MMU */
1747
1748 EXPORT_SYMBOL(generic_file_mmap);
1749 EXPORT_SYMBOL(generic_file_readonly_mmap);
1750
1751 static inline struct page *__read_cache_page(struct address_space *mapping,
1752                                 unsigned long index,
1753                                 int (*filler)(void *,struct page*),
1754                                 void *data)
1755 {
1756         struct page *page, *cached_page = NULL;
1757         int err;
1758 repeat:
1759         page = find_get_page(mapping, index);
1760         if (!page) {
1761                 if (!cached_page) {
1762                         cached_page = page_cache_alloc_cold(mapping);
1763                         if (!cached_page)
1764                                 return ERR_PTR(-ENOMEM);
1765                 }
1766                 err = add_to_page_cache_lru(cached_page, mapping,
1767                                         index, GFP_KERNEL);
1768                 if (err == -EEXIST)
1769                         goto repeat;
1770                 if (err < 0) {
1771                         /* Presumably ENOMEM for radix tree node */
1772                         page_cache_release(cached_page);
1773                         return ERR_PTR(err);
1774                 }
1775                 page = cached_page;
1776                 cached_page = NULL;
1777                 err = filler(data, page);
1778                 if (err < 0) {
1779                         page_cache_release(page);
1780                         page = ERR_PTR(err);
1781                 }
1782         }
1783         if (cached_page)
1784                 page_cache_release(cached_page);
1785         return page;
1786 }
1787
1788 /**
1789  * read_cache_page - read into page cache, fill it if needed
1790  * @mapping:    the page's address_space
1791  * @index:      the page index
1792  * @filler:     function to perform the read
1793  * @data:       destination for read data
1794  *
1795  * Read into the page cache. If a page already exists,
1796  * and PageUptodate() is not set, try to fill the page.
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         int err;
1805
1806 retry:
1807         page = __read_cache_page(mapping, index, filler, data);
1808         if (IS_ERR(page))
1809                 goto out;
1810         mark_page_accessed(page);
1811         if (PageUptodate(page))
1812                 goto out;
1813
1814         lock_page(page);
1815         if (!page->mapping) {
1816                 unlock_page(page);
1817                 page_cache_release(page);
1818                 goto retry;
1819         }
1820         if (PageUptodate(page)) {
1821                 unlock_page(page);
1822                 goto out;
1823         }
1824         err = filler(data, page);
1825         if (err < 0) {
1826                 page_cache_release(page);
1827                 page = ERR_PTR(err);
1828         }
1829  out:
1830         return page;
1831 }
1832 EXPORT_SYMBOL(read_cache_page);
1833
1834 /*
1835  * If the page was newly created, increment its refcount and add it to the
1836  * caller's lru-buffering pagevec.  This function is specifically for
1837  * generic_file_write().
1838  */
1839 static inline struct page *
1840 __grab_cache_page(struct address_space *mapping, unsigned long index,
1841                         struct page **cached_page, struct pagevec *lru_pvec)
1842 {
1843         int err;
1844         struct page *page;
1845 repeat:
1846         page = find_lock_page(mapping, index);
1847         if (!page) {
1848                 if (!*cached_page) {
1849                         *cached_page = page_cache_alloc(mapping);
1850                         if (!*cached_page)
1851                                 return NULL;
1852                 }
1853                 err = add_to_page_cache(*cached_page, mapping,
1854                                         index, GFP_KERNEL);
1855                 if (err == -EEXIST)
1856                         goto repeat;
1857                 if (err == 0) {
1858                         page = *cached_page;
1859                         page_cache_get(page);
1860                         if (!pagevec_add(lru_pvec, page))
1861                                 __pagevec_lru_add(lru_pvec);
1862                         *cached_page = NULL;
1863                 }
1864         }
1865         return page;
1866 }
1867
1868 /*
1869  * The logic we want is
1870  *
1871  *      if suid or (sgid and xgrp)
1872  *              remove privs
1873  */
1874 int should_remove_suid(struct dentry *dentry)
1875 {
1876         mode_t mode = dentry->d_inode->i_mode;
1877         int kill = 0;
1878
1879         /* suid always must be killed */
1880         if (unlikely(mode & S_ISUID))
1881                 kill = ATTR_KILL_SUID;
1882
1883         /*
1884          * sgid without any exec bits is just a mandatory locking mark; leave
1885          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1886          */
1887         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1888                 kill |= ATTR_KILL_SGID;
1889
1890         if (unlikely(kill && !capable(CAP_FSETID)))
1891                 return kill;
1892
1893         return 0;
1894 }
1895 EXPORT_SYMBOL(should_remove_suid);
1896
1897 int __remove_suid(struct dentry *dentry, int kill)
1898 {
1899         struct iattr newattrs;
1900
1901         newattrs.ia_valid = ATTR_FORCE | kill;
1902         return notify_change(dentry, &newattrs);
1903 }
1904
1905 int remove_suid(struct dentry *dentry)
1906 {
1907         int kill = should_remove_suid(dentry);
1908
1909         if (unlikely(kill))
1910                 return __remove_suid(dentry, kill);
1911
1912         return 0;
1913 }
1914 EXPORT_SYMBOL(remove_suid);
1915
1916 size_t
1917 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1918                         const struct iovec *iov, size_t base, size_t bytes)
1919 {
1920         size_t copied = 0, left = 0;
1921
1922         while (bytes) {
1923                 char __user *buf = iov->iov_base + base;
1924                 int copy = min(bytes, iov->iov_len - base);
1925
1926                 base = 0;
1927                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1928                 copied += copy;
1929                 bytes -= copy;
1930                 vaddr += copy;
1931                 iov++;
1932
1933                 if (unlikely(left))
1934                         break;
1935         }
1936         return copied - left;
1937 }
1938
1939 /*
1940  * Performs necessary checks before doing a write
1941  *
1942  * Can adjust writing position or amount of bytes to write.
1943  * Returns appropriate error code that caller should return or
1944  * zero in case that write should be allowed.
1945  */
1946 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1947 {
1948         struct inode *inode = file->f_mapping->host;
1949         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1950
1951         if (unlikely(*pos < 0))
1952                 return -EINVAL;
1953
1954         if (!isblk) {
1955                 /* FIXME: this is for backwards compatibility with 2.4 */
1956                 if (file->f_flags & O_APPEND)
1957                         *pos = i_size_read(inode);
1958
1959                 if (limit != RLIM_INFINITY) {
1960                         if (*pos >= limit) {
1961                                 send_sig(SIGXFSZ, current, 0);
1962                                 return -EFBIG;
1963                         }
1964                         if (*count > limit - (typeof(limit))*pos) {
1965                                 *count = limit - (typeof(limit))*pos;
1966                         }
1967                 }
1968         }
1969
1970         /*
1971          * LFS rule
1972          */
1973         if (unlikely(*pos + *count > MAX_NON_LFS &&
1974                                 !(file->f_flags & O_LARGEFILE))) {
1975                 if (*pos >= MAX_NON_LFS) {
1976                         send_sig(SIGXFSZ, current, 0);
1977                         return -EFBIG;
1978                 }
1979                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1980                         *count = MAX_NON_LFS - (unsigned long)*pos;
1981                 }
1982         }
1983
1984         /*
1985          * Are we about to exceed the fs block limit ?
1986          *
1987          * If we have written data it becomes a short write.  If we have
1988          * exceeded without writing data we send a signal and return EFBIG.
1989          * Linus frestrict idea will clean these up nicely..
1990          */
1991         if (likely(!isblk)) {
1992                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1993                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1994                                 send_sig(SIGXFSZ, current, 0);
1995                                 return -EFBIG;
1996                         }
1997                         /* zero-length writes at ->s_maxbytes are OK */
1998                 }
1999
2000                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2001                         *count = inode->i_sb->s_maxbytes - *pos;
2002         } else {
2003 #ifdef CONFIG_BLOCK
2004                 loff_t isize;
2005                 if (bdev_read_only(I_BDEV(inode)))
2006                         return -EPERM;
2007                 isize = i_size_read(inode);
2008                 if (*pos >= isize) {
2009                         if (*count || *pos > isize)
2010                                 return -ENOSPC;
2011                 }
2012
2013                 if (*pos + *count > isize)
2014                         *count = isize - *pos;
2015 #else
2016                 return -EPERM;
2017 #endif
2018         }
2019         return 0;
2020 }
2021 EXPORT_SYMBOL(generic_write_checks);
2022
2023 ssize_t
2024 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2025                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2026                 size_t count, size_t ocount)
2027 {
2028         struct file     *file = iocb->ki_filp;
2029         struct address_space *mapping = file->f_mapping;
2030         struct inode    *inode = mapping->host;
2031         ssize_t         written;
2032
2033         if (count != ocount)
2034                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2035
2036         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2037         if (written > 0) {
2038                 loff_t end = pos + written;
2039                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2040                         i_size_write(inode,  end);
2041                         mark_inode_dirty(inode);
2042                 }
2043                 *ppos = end;
2044         }
2045
2046         /*
2047          * Sync the fs metadata but not the minor inode changes and
2048          * of course not the data as we did direct DMA for the IO.
2049          * i_mutex is held, which protects generic_osync_inode() from
2050          * livelocking.
2051          */
2052         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2053                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2054                 if (err < 0)
2055                         written = err;
2056         }
2057         if (written == count && !is_sync_kiocb(iocb))
2058                 written = -EIOCBQUEUED;
2059         return written;
2060 }
2061 EXPORT_SYMBOL(generic_file_direct_write);
2062
2063 ssize_t
2064 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2065                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2066                 size_t count, ssize_t written)
2067 {
2068         struct file *file = iocb->ki_filp;
2069         struct address_space * mapping = file->f_mapping;
2070         const struct address_space_operations *a_ops = mapping->a_ops;
2071         struct inode    *inode = mapping->host;
2072         long            status = 0;
2073         struct page     *page;
2074         struct page     *cached_page = NULL;
2075         size_t          bytes;
2076         struct pagevec  lru_pvec;
2077         const struct iovec *cur_iov = iov; /* current iovec */
2078         size_t          iov_base = 0;      /* offset in the current iovec */
2079         char __user     *buf;
2080
2081         pagevec_init(&lru_pvec, 0);
2082
2083         /*
2084          * handle partial DIO write.  Adjust cur_iov if needed.
2085          */
2086         if (likely(nr_segs == 1))
2087                 buf = iov->iov_base + written;
2088         else {
2089                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
2090                 buf = cur_iov->iov_base + iov_base;
2091         }
2092
2093         do {
2094                 unsigned long index;
2095                 unsigned long offset;
2096                 size_t copied;
2097
2098                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
2099                 index = pos >> PAGE_CACHE_SHIFT;
2100                 bytes = PAGE_CACHE_SIZE - offset;
2101
2102                 /* Limit the size of the copy to the caller's write size */
2103                 bytes = min(bytes, count);
2104
2105                 /*
2106                  * Limit the size of the copy to that of the current segment,
2107                  * because fault_in_pages_readable() doesn't know how to walk
2108                  * segments.
2109                  */
2110                 bytes = min(bytes, cur_iov->iov_len - iov_base);
2111
2112                 /*
2113                  * Bring in the user page that we will copy from _first_.
2114                  * Otherwise there's a nasty deadlock on copying from the
2115                  * same page as we're writing to, without it being marked
2116                  * up-to-date.
2117                  */
2118                 fault_in_pages_readable(buf, bytes);
2119
2120                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
2121                 if (!page) {
2122                         status = -ENOMEM;
2123                         break;
2124                 }
2125
2126                 if (unlikely(bytes == 0)) {
2127                         status = 0;
2128                         copied = 0;
2129                         goto zero_length_segment;
2130                 }
2131
2132                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2133                 if (unlikely(status)) {
2134                         loff_t isize = i_size_read(inode);
2135
2136                         if (status != AOP_TRUNCATED_PAGE)
2137                                 unlock_page(page);
2138                         page_cache_release(page);
2139                         if (status == AOP_TRUNCATED_PAGE)
2140                                 continue;
2141                         /*
2142                          * prepare_write() may have instantiated a few blocks
2143                          * outside i_size.  Trim these off again.
2144                          */
2145                         if (pos + bytes > isize)
2146                                 vmtruncate(inode, isize);
2147                         break;
2148                 }
2149                 if (likely(nr_segs == 1))
2150                         copied = filemap_copy_from_user(page, offset,
2151                                                         buf, bytes);
2152                 else
2153                         copied = filemap_copy_from_user_iovec(page, offset,
2154                                                 cur_iov, iov_base, bytes);
2155                 flush_dcache_page(page);
2156                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2157                 if (status == AOP_TRUNCATED_PAGE) {
2158                         page_cache_release(page);
2159                         continue;
2160                 }
2161 zero_length_segment:
2162                 if (likely(copied >= 0)) {
2163                         if (!status)
2164                                 status = copied;
2165
2166                         if (status >= 0) {
2167                                 written += status;
2168                                 count -= status;
2169                                 pos += status;
2170                                 buf += status;
2171                                 if (unlikely(nr_segs > 1)) {
2172                                         filemap_set_next_iovec(&cur_iov,
2173                                                         &iov_base, status);
2174                                         if (count)
2175                                                 buf = cur_iov->iov_base +
2176                                                         iov_base;
2177                                 } else {
2178                                         iov_base += status;
2179                                 }
2180                         }
2181                 }
2182                 if (unlikely(copied != bytes))
2183                         if (status >= 0)
2184                                 status = -EFAULT;
2185                 unlock_page(page);
2186                 mark_page_accessed(page);
2187                 page_cache_release(page);
2188                 if (status < 0)
2189                         break;
2190                 balance_dirty_pages_ratelimited(mapping);
2191                 cond_resched();
2192         } while (count);
2193         *ppos = pos;
2194
2195         if (cached_page)
2196                 page_cache_release(cached_page);
2197
2198         /*
2199          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2200          */
2201         if (likely(status >= 0)) {
2202                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2203                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2204                                 status = generic_osync_inode(inode, mapping,
2205                                                 OSYNC_METADATA|OSYNC_DATA);
2206                 }
2207         }
2208         
2209         /*
2210          * If we get here for O_DIRECT writes then we must have fallen through
2211          * to buffered writes (block instantiation inside i_size).  So we sync
2212          * the file data here, to try to honour O_DIRECT expectations.
2213          */
2214         if (unlikely(file->f_flags & O_DIRECT) && written)
2215                 status = filemap_write_and_wait(mapping);
2216
2217         pagevec_lru_add(&lru_pvec);
2218         return written ? written : status;
2219 }
2220 EXPORT_SYMBOL(generic_file_buffered_write);
2221
2222 static ssize_t
2223 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2224                                 unsigned long nr_segs, loff_t *ppos)
2225 {
2226         struct file *file = iocb->ki_filp;
2227         struct address_space * mapping = file->f_mapping;
2228         size_t ocount;          /* original count */
2229         size_t count;           /* after file limit checks */
2230         struct inode    *inode = mapping->host;
2231         unsigned long   seg;
2232         loff_t          pos;
2233         ssize_t         written;
2234         ssize_t         err;
2235
2236         ocount = 0;
2237         for (seg = 0; seg < nr_segs; seg++) {
2238                 const struct iovec *iv = &iov[seg];
2239
2240                 /*
2241                  * If any segment has a negative length, or the cumulative
2242                  * length ever wraps negative then return -EINVAL.
2243                  */
2244                 ocount += iv->iov_len;
2245                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2246                         return -EINVAL;
2247                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2248                         continue;
2249                 if (seg == 0)
2250                         return -EFAULT;
2251                 nr_segs = seg;
2252                 ocount -= iv->iov_len;  /* This segment is no good */
2253                 break;
2254         }
2255
2256         count = ocount;
2257         pos = *ppos;
2258
2259         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2260
2261         /* We can write back this queue in page reclaim */
2262         current->backing_dev_info = mapping->backing_dev_info;
2263         written = 0;
2264
2265         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2266         if (err)
2267                 goto out;
2268
2269         if (count == 0)
2270                 goto out;
2271
2272         err = remove_suid(file->f_dentry);
2273         if (err)
2274                 goto out;
2275
2276         file_update_time(file);
2277
2278         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2279         if (unlikely(file->f_flags & O_DIRECT)) {
2280                 loff_t endbyte;
2281                 ssize_t written_buffered;
2282
2283                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2284                                                         ppos, count, ocount);
2285                 if (written < 0 || written == count)
2286                         goto out;
2287                 /*
2288                  * direct-io write to a hole: fall through to buffered I/O
2289                  * for completing the rest of the request.
2290                  */
2291                 pos += written;
2292                 count -= written;
2293                 written_buffered = generic_file_buffered_write(iocb, iov,
2294                                                 nr_segs, pos, ppos, count,
2295                                                 written);
2296                 /*
2297                  * If generic_file_buffered_write() retuned a synchronous error
2298                  * then we want to return the number of bytes which were
2299                  * direct-written, or the error code if that was zero.  Note
2300                  * that this differs from normal direct-io semantics, which
2301                  * will return -EFOO even if some bytes were written.
2302                  */
2303                 if (written_buffered < 0) {
2304                         err = written_buffered;
2305                         goto out;
2306                 }
2307
2308                 /*
2309                  * We need to ensure that the page cache pages are written to
2310                  * disk and invalidated to preserve the expected O_DIRECT
2311                  * semantics.
2312                  */
2313                 endbyte = pos + written_buffered - written - 1;
2314                 err = do_sync_file_range(file, pos, endbyte,
2315                                          SYNC_FILE_RANGE_WAIT_BEFORE|
2316                                          SYNC_FILE_RANGE_WRITE|
2317                                          SYNC_FILE_RANGE_WAIT_AFTER);
2318                 if (err == 0) {
2319                         written = written_buffered;
2320                         invalidate_mapping_pages(mapping,
2321                                                  pos >> PAGE_CACHE_SHIFT,
2322                                                  endbyte >> PAGE_CACHE_SHIFT);
2323                 } else {
2324                         /*
2325                          * We don't know how much we wrote, so just return
2326                          * the number of bytes which were direct-written
2327                          */
2328                 }
2329         } else {
2330                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2331                                 pos, ppos, count, written);
2332         }
2333 out:
2334         current->backing_dev_info = NULL;
2335         return written ? written : err;
2336 }
2337
2338 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2339                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2340 {
2341         struct file *file = iocb->ki_filp;
2342         struct address_space *mapping = file->f_mapping;
2343         struct inode *inode = mapping->host;
2344         ssize_t ret;
2345
2346         BUG_ON(iocb->ki_pos != pos);
2347
2348         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2349                         &iocb->ki_pos);
2350
2351         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2352                 ssize_t err;
2353
2354                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2355                 if (err < 0)
2356                         ret = err;
2357         }
2358         return ret;
2359 }
2360 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2361
2362 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2363                 unsigned long nr_segs, loff_t pos)
2364 {
2365         struct file *file = iocb->ki_filp;
2366         struct address_space *mapping = file->f_mapping;
2367         struct inode *inode = mapping->host;
2368         ssize_t ret;
2369
2370         BUG_ON(iocb->ki_pos != pos);
2371
2372         mutex_lock(&inode->i_mutex);
2373         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2374                         &iocb->ki_pos);
2375         mutex_unlock(&inode->i_mutex);
2376
2377         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2378                 ssize_t err;
2379
2380                 err = sync_page_range(inode, mapping, pos, ret);
2381                 if (err < 0)
2382                         ret = err;
2383         }
2384         return ret;
2385 }
2386 EXPORT_SYMBOL(generic_file_aio_write);
2387
2388 /*
2389  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2390  * went wrong during pagecache shootdown.
2391  */
2392 static ssize_t
2393 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2394         loff_t offset, unsigned long nr_segs)
2395 {
2396         struct file *file = iocb->ki_filp;
2397         struct address_space *mapping = file->f_mapping;
2398         ssize_t retval;
2399         size_t write_len = 0;
2400
2401         /*
2402          * If it's a write, unmap all mmappings of the file up-front.  This
2403          * will cause any pte dirty bits to be propagated into the pageframes
2404          * for the subsequent filemap_write_and_wait().
2405          */
2406         if (rw == WRITE) {
2407                 write_len = iov_length(iov, nr_segs);
2408                 if (mapping_mapped(mapping))
2409                         unmap_mapping_range(mapping, offset, write_len, 0);
2410         }
2411
2412         retval = filemap_write_and_wait(mapping);
2413         if (retval == 0) {
2414                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2415                                                 offset, nr_segs);
2416                 if (rw == WRITE && mapping->nrpages) {
2417                         pgoff_t end = (offset + write_len - 1)
2418                                                 >> PAGE_CACHE_SHIFT;
2419                         int err = invalidate_inode_pages2_range(mapping,
2420                                         offset >> PAGE_CACHE_SHIFT, end);
2421                         if (err)
2422                                 retval = err;
2423                 }
2424         }
2425         return retval;
2426 }
2427
2428 /**
2429  * try_to_release_page() - release old fs-specific metadata on a page
2430  *
2431  * @page: the page which the kernel is trying to free
2432  * @gfp_mask: memory allocation flags (and I/O mode)
2433  *
2434  * The address_space is to try to release any data against the page
2435  * (presumably at page->private).  If the release was successful, return `1'.
2436  * Otherwise return zero.
2437  *
2438  * The @gfp_mask argument specifies whether I/O may be performed to release
2439  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2440  *
2441  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2442  */
2443 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2444 {
2445         struct address_space * const mapping = page->mapping;
2446
2447         BUG_ON(!PageLocked(page));
2448         if (PageWriteback(page))
2449                 return 0;
2450
2451         if (mapping && mapping->a_ops->releasepage)
2452                 return mapping->a_ops->releasepage(page, gfp_mask);
2453         return try_to_free_buffers(page);
2454 }
2455
2456 EXPORT_SYMBOL(try_to_release_page);