Merge git://git.kernel.org/pub/scm/linux/kernel/git/paulus/powerpc
[linux-2.6] / 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  *  ->mmap_sem
79  *    ->i_mutex                 (msync)
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(struct address_space *x)
471 {
472         if (cpuset_do_page_mem_spread()) {
473                 int n = cpuset_mem_spread_node();
474                 return alloc_pages_node(n, mapping_gfp_mask(x), 0);
475         }
476         return alloc_pages(mapping_gfp_mask(x), 0);
477 }
478 EXPORT_SYMBOL(page_cache_alloc);
479
480 struct page *page_cache_alloc_cold(struct address_space *x)
481 {
482         if (cpuset_do_page_mem_spread()) {
483                 int n = cpuset_mem_spread_node();
484                 return alloc_pages_node(n, mapping_gfp_mask(x)|__GFP_COLD, 0);
485         }
486         return alloc_pages(mapping_gfp_mask(x)|__GFP_COLD, 0);
487 }
488 EXPORT_SYMBOL(page_cache_alloc_cold);
489 #endif
490
491 /*
492  * In order to wait for pages to become available there must be
493  * waitqueues associated with pages. By using a hash table of
494  * waitqueues where the bucket discipline is to maintain all
495  * waiters on the same queue and wake all when any of the pages
496  * become available, and for the woken contexts to check to be
497  * sure the appropriate page became available, this saves space
498  * at a cost of "thundering herd" phenomena during rare hash
499  * collisions.
500  */
501 static wait_queue_head_t *page_waitqueue(struct page *page)
502 {
503         const struct zone *zone = page_zone(page);
504
505         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
506 }
507
508 static inline void wake_up_page(struct page *page, int bit)
509 {
510         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
511 }
512
513 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
514 {
515         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
516
517         if (test_bit(bit_nr, &page->flags))
518                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
519                                                         TASK_UNINTERRUPTIBLE);
520 }
521 EXPORT_SYMBOL(wait_on_page_bit);
522
523 /**
524  * unlock_page - unlock a locked page
525  * @page: the page
526  *
527  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
528  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
529  * mechananism between PageLocked pages and PageWriteback pages is shared.
530  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
531  *
532  * The first mb is necessary to safely close the critical section opened by the
533  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
534  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
535  * parallel wait_on_page_locked()).
536  */
537 void fastcall unlock_page(struct page *page)
538 {
539         smp_mb__before_clear_bit();
540         if (!TestClearPageLocked(page))
541                 BUG();
542         smp_mb__after_clear_bit(); 
543         wake_up_page(page, PG_locked);
544 }
545 EXPORT_SYMBOL(unlock_page);
546
547 /**
548  * end_page_writeback - end writeback against a page
549  * @page: the page
550  */
551 void end_page_writeback(struct page *page)
552 {
553         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
554                 if (!test_clear_page_writeback(page))
555                         BUG();
556         }
557         smp_mb__after_clear_bit();
558         wake_up_page(page, PG_writeback);
559 }
560 EXPORT_SYMBOL(end_page_writeback);
561
562 /**
563  * __lock_page - get a lock on the page, assuming we need to sleep to get it
564  * @page: the page to lock
565  *
566  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
567  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
568  * chances are that on the second loop, the block layer's plug list is empty,
569  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
570  */
571 void fastcall __lock_page(struct page *page)
572 {
573         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
574
575         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
576                                                         TASK_UNINTERRUPTIBLE);
577 }
578 EXPORT_SYMBOL(__lock_page);
579
580 /**
581  * find_get_page - find and get a page reference
582  * @mapping: the address_space to search
583  * @offset: the page index
584  *
585  * A rather lightweight function, finding and getting a reference to a
586  * hashed page atomically.
587  */
588 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
589 {
590         struct page *page;
591
592         read_lock_irq(&mapping->tree_lock);
593         page = radix_tree_lookup(&mapping->page_tree, offset);
594         if (page)
595                 page_cache_get(page);
596         read_unlock_irq(&mapping->tree_lock);
597         return page;
598 }
599 EXPORT_SYMBOL(find_get_page);
600
601 /**
602  * find_trylock_page - find and lock a page
603  * @mapping: the address_space to search
604  * @offset: the page index
605  *
606  * Same as find_get_page(), but trylock it instead of incrementing the count.
607  */
608 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
609 {
610         struct page *page;
611
612         read_lock_irq(&mapping->tree_lock);
613         page = radix_tree_lookup(&mapping->page_tree, offset);
614         if (page && TestSetPageLocked(page))
615                 page = NULL;
616         read_unlock_irq(&mapping->tree_lock);
617         return page;
618 }
619 EXPORT_SYMBOL(find_trylock_page);
620
621 /**
622  * find_lock_page - locate, pin and lock a pagecache page
623  * @mapping: the address_space to search
624  * @offset: the page index
625  *
626  * Locates the desired pagecache page, locks it, increments its reference
627  * count and returns its address.
628  *
629  * Returns zero if the page was not present. find_lock_page() may sleep.
630  */
631 struct page *find_lock_page(struct address_space *mapping,
632                                 unsigned long offset)
633 {
634         struct page *page;
635
636         read_lock_irq(&mapping->tree_lock);
637 repeat:
638         page = radix_tree_lookup(&mapping->page_tree, offset);
639         if (page) {
640                 page_cache_get(page);
641                 if (TestSetPageLocked(page)) {
642                         read_unlock_irq(&mapping->tree_lock);
643                         __lock_page(page);
644                         read_lock_irq(&mapping->tree_lock);
645
646                         /* Has the page been truncated while we slept? */
647                         if (unlikely(page->mapping != mapping ||
648                                      page->index != offset)) {
649                                 unlock_page(page);
650                                 page_cache_release(page);
651                                 goto repeat;
652                         }
653                 }
654         }
655         read_unlock_irq(&mapping->tree_lock);
656         return page;
657 }
658 EXPORT_SYMBOL(find_lock_page);
659
660 /**
661  * find_or_create_page - locate or add a pagecache page
662  * @mapping: the page's address_space
663  * @index: the page's index into the mapping
664  * @gfp_mask: page allocation mode
665  *
666  * Locates a page in the pagecache.  If the page is not present, a new page
667  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
668  * LRU list.  The returned page is locked and has its reference count
669  * incremented.
670  *
671  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
672  * allocation!
673  *
674  * find_or_create_page() returns the desired page's address, or zero on
675  * memory exhaustion.
676  */
677 struct page *find_or_create_page(struct address_space *mapping,
678                 unsigned long index, gfp_t gfp_mask)
679 {
680         struct page *page, *cached_page = NULL;
681         int err;
682 repeat:
683         page = find_lock_page(mapping, index);
684         if (!page) {
685                 if (!cached_page) {
686                         cached_page = alloc_page(gfp_mask);
687                         if (!cached_page)
688                                 return NULL;
689                 }
690                 err = add_to_page_cache_lru(cached_page, mapping,
691                                         index, gfp_mask);
692                 if (!err) {
693                         page = cached_page;
694                         cached_page = NULL;
695                 } else if (err == -EEXIST)
696                         goto repeat;
697         }
698         if (cached_page)
699                 page_cache_release(cached_page);
700         return page;
701 }
702 EXPORT_SYMBOL(find_or_create_page);
703
704 /**
705  * find_get_pages - gang pagecache lookup
706  * @mapping:    The address_space to search
707  * @start:      The starting page index
708  * @nr_pages:   The maximum number of pages
709  * @pages:      Where the resulting pages are placed
710  *
711  * find_get_pages() will search for and return a group of up to
712  * @nr_pages pages in the mapping.  The pages are placed at @pages.
713  * find_get_pages() takes a reference against the returned pages.
714  *
715  * The search returns a group of mapping-contiguous pages with ascending
716  * indexes.  There may be holes in the indices due to not-present pages.
717  *
718  * find_get_pages() returns the number of pages which were found.
719  */
720 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
721                             unsigned int nr_pages, struct page **pages)
722 {
723         unsigned int i;
724         unsigned int ret;
725
726         read_lock_irq(&mapping->tree_lock);
727         ret = radix_tree_gang_lookup(&mapping->page_tree,
728                                 (void **)pages, start, nr_pages);
729         for (i = 0; i < ret; i++)
730                 page_cache_get(pages[i]);
731         read_unlock_irq(&mapping->tree_lock);
732         return ret;
733 }
734
735 /**
736  * find_get_pages_contig - gang contiguous pagecache lookup
737  * @mapping:    The address_space to search
738  * @index:      The starting page index
739  * @nr_pages:   The maximum number of pages
740  * @pages:      Where the resulting pages are placed
741  *
742  * find_get_pages_contig() works exactly like find_get_pages(), except
743  * that the returned number of pages are guaranteed to be contiguous.
744  *
745  * find_get_pages_contig() returns the number of pages which were found.
746  */
747 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
748                                unsigned int nr_pages, struct page **pages)
749 {
750         unsigned int i;
751         unsigned int ret;
752
753         read_lock_irq(&mapping->tree_lock);
754         ret = radix_tree_gang_lookup(&mapping->page_tree,
755                                 (void **)pages, index, nr_pages);
756         for (i = 0; i < ret; i++) {
757                 if (pages[i]->mapping == NULL || pages[i]->index != index)
758                         break;
759
760                 page_cache_get(pages[i]);
761                 index++;
762         }
763         read_unlock_irq(&mapping->tree_lock);
764         return i;
765 }
766
767 /**
768  * find_get_pages_tag - find and return pages that match @tag
769  * @mapping:    the address_space to search
770  * @index:      the starting page index
771  * @tag:        the tag index
772  * @nr_pages:   the maximum number of pages
773  * @pages:      where the resulting pages are placed
774  *
775  * Like find_get_pages, except we only return pages which are tagged with
776  * @tag.   We update @index to index the next page for the traversal.
777  */
778 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
779                         int tag, unsigned int nr_pages, struct page **pages)
780 {
781         unsigned int i;
782         unsigned int ret;
783
784         read_lock_irq(&mapping->tree_lock);
785         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
786                                 (void **)pages, *index, nr_pages, tag);
787         for (i = 0; i < ret; i++)
788                 page_cache_get(pages[i]);
789         if (ret)
790                 *index = pages[ret - 1]->index + 1;
791         read_unlock_irq(&mapping->tree_lock);
792         return ret;
793 }
794
795 /**
796  * grab_cache_page_nowait - returns locked page at given index in given cache
797  * @mapping: target address_space
798  * @index: the page index
799  *
800  * Same as grab_cache_page, but do not wait if the page is unavailable.
801  * This is intended for speculative data generators, where the data can
802  * be regenerated if the page couldn't be grabbed.  This routine should
803  * be safe to call while holding the lock for another page.
804  *
805  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
806  * and deadlock against the caller's locked page.
807  */
808 struct page *
809 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
810 {
811         struct page *page = find_get_page(mapping, index);
812         gfp_t gfp_mask;
813
814         if (page) {
815                 if (!TestSetPageLocked(page))
816                         return page;
817                 page_cache_release(page);
818                 return NULL;
819         }
820         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
821         page = alloc_pages(gfp_mask, 0);
822         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
823                 page_cache_release(page);
824                 page = NULL;
825         }
826         return page;
827 }
828 EXPORT_SYMBOL(grab_cache_page_nowait);
829
830 /*
831  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
832  * a _large_ part of the i/o request. Imagine the worst scenario:
833  *
834  *      ---R__________________________________________B__________
835  *         ^ reading here                             ^ bad block(assume 4k)
836  *
837  * read(R) => miss => readahead(R...B) => media error => frustrating retries
838  * => failing the whole request => read(R) => read(R+1) =>
839  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
840  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
841  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
842  *
843  * It is going insane. Fix it by quickly scaling down the readahead size.
844  */
845 static void shrink_readahead_size_eio(struct file *filp,
846                                         struct file_ra_state *ra)
847 {
848         if (!ra->ra_pages)
849                 return;
850
851         ra->ra_pages /= 4;
852         printk(KERN_WARNING "Reducing readahead size to %luK\n",
853                         ra->ra_pages << (PAGE_CACHE_SHIFT - 10));
854 }
855
856 /**
857  * do_generic_mapping_read - generic file read routine
858  * @mapping:    address_space to be read
859  * @_ra:        file's readahead state
860  * @filp:       the file to read
861  * @ppos:       current file position
862  * @desc:       read_descriptor
863  * @actor:      read method
864  *
865  * This is a generic file read routine, and uses the
866  * mapping->a_ops->readpage() function for the actual low-level stuff.
867  *
868  * This is really ugly. But the goto's actually try to clarify some
869  * of the logic when it comes to error handling etc.
870  *
871  * Note the struct file* is only passed for the use of readpage.
872  * It may be NULL.
873  */
874 void do_generic_mapping_read(struct address_space *mapping,
875                              struct file_ra_state *_ra,
876                              struct file *filp,
877                              loff_t *ppos,
878                              read_descriptor_t *desc,
879                              read_actor_t actor)
880 {
881         struct inode *inode = mapping->host;
882         unsigned long index;
883         unsigned long end_index;
884         unsigned long offset;
885         unsigned long last_index;
886         unsigned long next_index;
887         unsigned long prev_index;
888         loff_t isize;
889         struct page *cached_page;
890         int error;
891         struct file_ra_state ra = *_ra;
892
893         cached_page = NULL;
894         index = *ppos >> PAGE_CACHE_SHIFT;
895         next_index = index;
896         prev_index = ra.prev_page;
897         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
898         offset = *ppos & ~PAGE_CACHE_MASK;
899
900         isize = i_size_read(inode);
901         if (!isize)
902                 goto out;
903
904         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
905         for (;;) {
906                 struct page *page;
907                 unsigned long nr, ret;
908
909                 /* nr is the maximum number of bytes to copy from this page */
910                 nr = PAGE_CACHE_SIZE;
911                 if (index >= end_index) {
912                         if (index > end_index)
913                                 goto out;
914                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
915                         if (nr <= offset) {
916                                 goto out;
917                         }
918                 }
919                 nr = nr - offset;
920
921                 cond_resched();
922                 if (index == next_index)
923                         next_index = page_cache_readahead(mapping, &ra, filp,
924                                         index, last_index - index);
925
926 find_page:
927                 page = find_get_page(mapping, index);
928                 if (unlikely(page == NULL)) {
929                         handle_ra_miss(mapping, &ra, index);
930                         goto no_cached_page;
931                 }
932                 if (!PageUptodate(page))
933                         goto page_not_up_to_date;
934 page_ok:
935
936                 /* If users can be writing to this page using arbitrary
937                  * virtual addresses, take care about potential aliasing
938                  * before reading the page on the kernel side.
939                  */
940                 if (mapping_writably_mapped(mapping))
941                         flush_dcache_page(page);
942
943                 /*
944                  * When (part of) the same page is read multiple times
945                  * in succession, only mark it as accessed the first time.
946                  */
947                 if (prev_index != index)
948                         mark_page_accessed(page);
949                 prev_index = index;
950
951                 /*
952                  * Ok, we have the page, and it's up-to-date, so
953                  * now we can copy it to user space...
954                  *
955                  * The actor routine returns how many bytes were actually used..
956                  * NOTE! This may not be the same as how much of a user buffer
957                  * we filled up (we may be padding etc), so we can only update
958                  * "pos" here (the actor routine has to update the user buffer
959                  * pointers and the remaining count).
960                  */
961                 ret = actor(desc, page, offset, nr);
962                 offset += ret;
963                 index += offset >> PAGE_CACHE_SHIFT;
964                 offset &= ~PAGE_CACHE_MASK;
965
966                 page_cache_release(page);
967                 if (ret == nr && desc->count)
968                         continue;
969                 goto out;
970
971 page_not_up_to_date:
972                 /* Get exclusive access to the page ... */
973                 lock_page(page);
974
975                 /* Did it get unhashed before we got the lock? */
976                 if (!page->mapping) {
977                         unlock_page(page);
978                         page_cache_release(page);
979                         continue;
980                 }
981
982                 /* Did somebody else fill it already? */
983                 if (PageUptodate(page)) {
984                         unlock_page(page);
985                         goto page_ok;
986                 }
987
988 readpage:
989                 /* Start the actual read. The read will unlock the page. */
990                 error = mapping->a_ops->readpage(filp, page);
991
992                 if (unlikely(error)) {
993                         if (error == AOP_TRUNCATED_PAGE) {
994                                 page_cache_release(page);
995                                 goto find_page;
996                         }
997                         goto readpage_error;
998                 }
999
1000                 if (!PageUptodate(page)) {
1001                         lock_page(page);
1002                         if (!PageUptodate(page)) {
1003                                 if (page->mapping == NULL) {
1004                                         /*
1005                                          * invalidate_inode_pages got it
1006                                          */
1007                                         unlock_page(page);
1008                                         page_cache_release(page);
1009                                         goto find_page;
1010                                 }
1011                                 unlock_page(page);
1012                                 error = -EIO;
1013                                 shrink_readahead_size_eio(filp, &ra);
1014                                 goto readpage_error;
1015                         }
1016                         unlock_page(page);
1017                 }
1018
1019                 /*
1020                  * i_size must be checked after we have done ->readpage.
1021                  *
1022                  * Checking i_size after the readpage allows us to calculate
1023                  * the correct value for "nr", which means the zero-filled
1024                  * part of the page is not copied back to userspace (unless
1025                  * another truncate extends the file - this is desired though).
1026                  */
1027                 isize = i_size_read(inode);
1028                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1029                 if (unlikely(!isize || index > end_index)) {
1030                         page_cache_release(page);
1031                         goto out;
1032                 }
1033
1034                 /* nr is the maximum number of bytes to copy from this page */
1035                 nr = PAGE_CACHE_SIZE;
1036                 if (index == end_index) {
1037                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1038                         if (nr <= offset) {
1039                                 page_cache_release(page);
1040                                 goto out;
1041                         }
1042                 }
1043                 nr = nr - offset;
1044                 goto page_ok;
1045
1046 readpage_error:
1047                 /* UHHUH! A synchronous read error occurred. Report it */
1048                 desc->error = error;
1049                 page_cache_release(page);
1050                 goto out;
1051
1052 no_cached_page:
1053                 /*
1054                  * Ok, it wasn't cached, so we need to create a new
1055                  * page..
1056                  */
1057                 if (!cached_page) {
1058                         cached_page = page_cache_alloc_cold(mapping);
1059                         if (!cached_page) {
1060                                 desc->error = -ENOMEM;
1061                                 goto out;
1062                         }
1063                 }
1064                 error = add_to_page_cache_lru(cached_page, mapping,
1065                                                 index, GFP_KERNEL);
1066                 if (error) {
1067                         if (error == -EEXIST)
1068                                 goto find_page;
1069                         desc->error = error;
1070                         goto out;
1071                 }
1072                 page = cached_page;
1073                 cached_page = NULL;
1074                 goto readpage;
1075         }
1076
1077 out:
1078         *_ra = ra;
1079
1080         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1081         if (cached_page)
1082                 page_cache_release(cached_page);
1083         if (filp)
1084                 file_accessed(filp);
1085 }
1086 EXPORT_SYMBOL(do_generic_mapping_read);
1087
1088 int file_read_actor(read_descriptor_t *desc, struct page *page,
1089                         unsigned long offset, unsigned long size)
1090 {
1091         char *kaddr;
1092         unsigned long left, count = desc->count;
1093
1094         if (size > count)
1095                 size = count;
1096
1097         /*
1098          * Faults on the destination of a read are common, so do it before
1099          * taking the kmap.
1100          */
1101         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1102                 kaddr = kmap_atomic(page, KM_USER0);
1103                 left = __copy_to_user_inatomic(desc->arg.buf,
1104                                                 kaddr + offset, size);
1105                 kunmap_atomic(kaddr, KM_USER0);
1106                 if (left == 0)
1107                         goto success;
1108         }
1109
1110         /* Do it the slow way */
1111         kaddr = kmap(page);
1112         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1113         kunmap(page);
1114
1115         if (left) {
1116                 size -= left;
1117                 desc->error = -EFAULT;
1118         }
1119 success:
1120         desc->count = count - size;
1121         desc->written += size;
1122         desc->arg.buf += size;
1123         return size;
1124 }
1125
1126 /**
1127  * __generic_file_aio_read - generic filesystem read routine
1128  * @iocb:       kernel I/O control block
1129  * @iov:        io vector request
1130  * @nr_segs:    number of segments in the iovec
1131  * @ppos:       current file position
1132  *
1133  * This is the "read()" routine for all filesystems
1134  * that can use the page cache directly.
1135  */
1136 ssize_t
1137 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1138                 unsigned long nr_segs, loff_t *ppos)
1139 {
1140         struct file *filp = iocb->ki_filp;
1141         ssize_t retval;
1142         unsigned long seg;
1143         size_t count;
1144
1145         count = 0;
1146         for (seg = 0; seg < nr_segs; seg++) {
1147                 const struct iovec *iv = &iov[seg];
1148
1149                 /*
1150                  * If any segment has a negative length, or the cumulative
1151                  * length ever wraps negative then return -EINVAL.
1152                  */
1153                 count += iv->iov_len;
1154                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1155                         return -EINVAL;
1156                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1157                         continue;
1158                 if (seg == 0)
1159                         return -EFAULT;
1160                 nr_segs = seg;
1161                 count -= iv->iov_len;   /* This segment is no good */
1162                 break;
1163         }
1164
1165         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1166         if (filp->f_flags & O_DIRECT) {
1167                 loff_t pos = *ppos, size;
1168                 struct address_space *mapping;
1169                 struct inode *inode;
1170
1171                 mapping = filp->f_mapping;
1172                 inode = mapping->host;
1173                 retval = 0;
1174                 if (!count)
1175                         goto out; /* skip atime */
1176                 size = i_size_read(inode);
1177                 if (pos < size) {
1178                         retval = generic_file_direct_IO(READ, iocb,
1179                                                 iov, pos, nr_segs);
1180                         if (retval > 0 && !is_sync_kiocb(iocb))
1181                                 retval = -EIOCBQUEUED;
1182                         if (retval > 0)
1183                                 *ppos = pos + retval;
1184                 }
1185                 file_accessed(filp);
1186                 goto out;
1187         }
1188
1189         retval = 0;
1190         if (count) {
1191                 for (seg = 0; seg < nr_segs; seg++) {
1192                         read_descriptor_t desc;
1193
1194                         desc.written = 0;
1195                         desc.arg.buf = iov[seg].iov_base;
1196                         desc.count = iov[seg].iov_len;
1197                         if (desc.count == 0)
1198                                 continue;
1199                         desc.error = 0;
1200                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1201                         retval += desc.written;
1202                         if (desc.error) {
1203                                 retval = retval ?: desc.error;
1204                                 break;
1205                         }
1206                 }
1207         }
1208 out:
1209         return retval;
1210 }
1211 EXPORT_SYMBOL(__generic_file_aio_read);
1212
1213 ssize_t
1214 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1215 {
1216         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1217
1218         BUG_ON(iocb->ki_pos != pos);
1219         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1220 }
1221 EXPORT_SYMBOL(generic_file_aio_read);
1222
1223 ssize_t
1224 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1225 {
1226         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1227         struct kiocb kiocb;
1228         ssize_t ret;
1229
1230         init_sync_kiocb(&kiocb, filp);
1231         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1232         if (-EIOCBQUEUED == ret)
1233                 ret = wait_on_sync_kiocb(&kiocb);
1234         return ret;
1235 }
1236 EXPORT_SYMBOL(generic_file_read);
1237
1238 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1239 {
1240         ssize_t written;
1241         unsigned long count = desc->count;
1242         struct file *file = desc->arg.data;
1243
1244         if (size > count)
1245                 size = count;
1246
1247         written = file->f_op->sendpage(file, page, offset,
1248                                        size, &file->f_pos, size<count);
1249         if (written < 0) {
1250                 desc->error = written;
1251                 written = 0;
1252         }
1253         desc->count = count - written;
1254         desc->written += written;
1255         return written;
1256 }
1257
1258 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1259                          size_t count, read_actor_t actor, void *target)
1260 {
1261         read_descriptor_t desc;
1262
1263         if (!count)
1264                 return 0;
1265
1266         desc.written = 0;
1267         desc.count = count;
1268         desc.arg.data = target;
1269         desc.error = 0;
1270
1271         do_generic_file_read(in_file, ppos, &desc, actor);
1272         if (desc.written)
1273                 return desc.written;
1274         return desc.error;
1275 }
1276 EXPORT_SYMBOL(generic_file_sendfile);
1277
1278 static ssize_t
1279 do_readahead(struct address_space *mapping, struct file *filp,
1280              unsigned long index, unsigned long nr)
1281 {
1282         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1283                 return -EINVAL;
1284
1285         force_page_cache_readahead(mapping, filp, index,
1286                                         max_sane_readahead(nr));
1287         return 0;
1288 }
1289
1290 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1291 {
1292         ssize_t ret;
1293         struct file *file;
1294
1295         ret = -EBADF;
1296         file = fget(fd);
1297         if (file) {
1298                 if (file->f_mode & FMODE_READ) {
1299                         struct address_space *mapping = file->f_mapping;
1300                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1301                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1302                         unsigned long len = end - start + 1;
1303                         ret = do_readahead(mapping, file, start, len);
1304                 }
1305                 fput(file);
1306         }
1307         return ret;
1308 }
1309
1310 #ifdef CONFIG_MMU
1311 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1312 /**
1313  * page_cache_read - adds requested page to the page cache if not already there
1314  * @file:       file to read
1315  * @offset:     page index
1316  *
1317  * This adds the requested page to the page cache if it isn't already there,
1318  * and schedules an I/O to read in its contents from disk.
1319  */
1320 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1321 {
1322         struct address_space *mapping = file->f_mapping;
1323         struct page *page; 
1324         int ret;
1325
1326         do {
1327                 page = page_cache_alloc_cold(mapping);
1328                 if (!page)
1329                         return -ENOMEM;
1330
1331                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1332                 if (ret == 0)
1333                         ret = mapping->a_ops->readpage(file, page);
1334                 else if (ret == -EEXIST)
1335                         ret = 0; /* losing race to add is OK */
1336
1337                 page_cache_release(page);
1338
1339         } while (ret == AOP_TRUNCATED_PAGE);
1340                 
1341         return ret;
1342 }
1343
1344 #define MMAP_LOTSAMISS  (100)
1345
1346 /**
1347  * filemap_nopage - read in file data for page fault handling
1348  * @area:       the applicable vm_area
1349  * @address:    target address to read in
1350  * @type:       returned with VM_FAULT_{MINOR,MAJOR} if not %NULL
1351  *
1352  * filemap_nopage() is invoked via the vma operations vector for a
1353  * mapped memory region to read in file data during a page fault.
1354  *
1355  * The goto's are kind of ugly, but this streamlines the normal case of having
1356  * it in the page cache, and handles the special cases reasonably without
1357  * having a lot of duplicated code.
1358  */
1359 struct page *filemap_nopage(struct vm_area_struct *area,
1360                                 unsigned long address, int *type)
1361 {
1362         int error;
1363         struct file *file = area->vm_file;
1364         struct address_space *mapping = file->f_mapping;
1365         struct file_ra_state *ra = &file->f_ra;
1366         struct inode *inode = mapping->host;
1367         struct page *page;
1368         unsigned long size, pgoff;
1369         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1370
1371         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1372
1373 retry_all:
1374         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1375         if (pgoff >= size)
1376                 goto outside_data_content;
1377
1378         /* If we don't want any read-ahead, don't bother */
1379         if (VM_RandomReadHint(area))
1380                 goto no_cached_page;
1381
1382         /*
1383          * The readahead code wants to be told about each and every page
1384          * so it can build and shrink its windows appropriately
1385          *
1386          * For sequential accesses, we use the generic readahead logic.
1387          */
1388         if (VM_SequentialReadHint(area))
1389                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1390
1391         /*
1392          * Do we have something in the page cache already?
1393          */
1394 retry_find:
1395         page = find_get_page(mapping, pgoff);
1396         if (!page) {
1397                 unsigned long ra_pages;
1398
1399                 if (VM_SequentialReadHint(area)) {
1400                         handle_ra_miss(mapping, ra, pgoff);
1401                         goto no_cached_page;
1402                 }
1403                 ra->mmap_miss++;
1404
1405                 /*
1406                  * Do we miss much more than hit in this file? If so,
1407                  * stop bothering with read-ahead. It will only hurt.
1408                  */
1409                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1410                         goto no_cached_page;
1411
1412                 /*
1413                  * To keep the pgmajfault counter straight, we need to
1414                  * check did_readaround, as this is an inner loop.
1415                  */
1416                 if (!did_readaround) {
1417                         majmin = VM_FAULT_MAJOR;
1418                         count_vm_event(PGMAJFAULT);
1419                 }
1420                 did_readaround = 1;
1421                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1422                 if (ra_pages) {
1423                         pgoff_t start = 0;
1424
1425                         if (pgoff > ra_pages / 2)
1426                                 start = pgoff - ra_pages / 2;
1427                         do_page_cache_readahead(mapping, file, start, ra_pages);
1428                 }
1429                 page = find_get_page(mapping, pgoff);
1430                 if (!page)
1431                         goto no_cached_page;
1432         }
1433
1434         if (!did_readaround)
1435                 ra->mmap_hit++;
1436
1437         /*
1438          * Ok, found a page in the page cache, now we need to check
1439          * that it's up-to-date.
1440          */
1441         if (!PageUptodate(page))
1442                 goto page_not_uptodate;
1443
1444 success:
1445         /*
1446          * Found the page and have a reference on it.
1447          */
1448         mark_page_accessed(page);
1449         if (type)
1450                 *type = majmin;
1451         return page;
1452
1453 outside_data_content:
1454         /*
1455          * An external ptracer can access pages that normally aren't
1456          * accessible..
1457          */
1458         if (area->vm_mm == current->mm)
1459                 return NULL;
1460         /* Fall through to the non-read-ahead case */
1461 no_cached_page:
1462         /*
1463          * We're only likely to ever get here if MADV_RANDOM is in
1464          * effect.
1465          */
1466         error = page_cache_read(file, pgoff);
1467         grab_swap_token();
1468
1469         /*
1470          * The page we want has now been added to the page cache.
1471          * In the unlikely event that someone removed it in the
1472          * meantime, we'll just come back here and read it again.
1473          */
1474         if (error >= 0)
1475                 goto retry_find;
1476
1477         /*
1478          * An error return from page_cache_read can result if the
1479          * system is low on memory, or a problem occurs while trying
1480          * to schedule I/O.
1481          */
1482         if (error == -ENOMEM)
1483                 return NOPAGE_OOM;
1484         return NULL;
1485
1486 page_not_uptodate:
1487         if (!did_readaround) {
1488                 majmin = VM_FAULT_MAJOR;
1489                 count_vm_event(PGMAJFAULT);
1490         }
1491         lock_page(page);
1492
1493         /* Did it get unhashed while we waited for it? */
1494         if (!page->mapping) {
1495                 unlock_page(page);
1496                 page_cache_release(page);
1497                 goto retry_all;
1498         }
1499
1500         /* Did somebody else get it up-to-date? */
1501         if (PageUptodate(page)) {
1502                 unlock_page(page);
1503                 goto success;
1504         }
1505
1506         error = mapping->a_ops->readpage(file, page);
1507         if (!error) {
1508                 wait_on_page_locked(page);
1509                 if (PageUptodate(page))
1510                         goto success;
1511         } else if (error == AOP_TRUNCATED_PAGE) {
1512                 page_cache_release(page);
1513                 goto retry_find;
1514         }
1515
1516         /*
1517          * Umm, take care of errors if the page isn't up-to-date.
1518          * Try to re-read it _once_. We do this synchronously,
1519          * because there really aren't any performance issues here
1520          * and we need to check for errors.
1521          */
1522         lock_page(page);
1523
1524         /* Somebody truncated the page on us? */
1525         if (!page->mapping) {
1526                 unlock_page(page);
1527                 page_cache_release(page);
1528                 goto retry_all;
1529         }
1530
1531         /* Somebody else successfully read it in? */
1532         if (PageUptodate(page)) {
1533                 unlock_page(page);
1534                 goto success;
1535         }
1536         ClearPageError(page);
1537         error = mapping->a_ops->readpage(file, page);
1538         if (!error) {
1539                 wait_on_page_locked(page);
1540                 if (PageUptodate(page))
1541                         goto success;
1542         } else if (error == AOP_TRUNCATED_PAGE) {
1543                 page_cache_release(page);
1544                 goto retry_find;
1545         }
1546
1547         /*
1548          * Things didn't work out. Return zero to tell the
1549          * mm layer so, possibly freeing the page cache page first.
1550          */
1551         shrink_readahead_size_eio(file, ra);
1552         page_cache_release(page);
1553         return NULL;
1554 }
1555 EXPORT_SYMBOL(filemap_nopage);
1556
1557 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1558                                         int nonblock)
1559 {
1560         struct address_space *mapping = file->f_mapping;
1561         struct page *page;
1562         int error;
1563
1564         /*
1565          * Do we have something in the page cache already?
1566          */
1567 retry_find:
1568         page = find_get_page(mapping, pgoff);
1569         if (!page) {
1570                 if (nonblock)
1571                         return NULL;
1572                 goto no_cached_page;
1573         }
1574
1575         /*
1576          * Ok, found a page in the page cache, now we need to check
1577          * that it's up-to-date.
1578          */
1579         if (!PageUptodate(page)) {
1580                 if (nonblock) {
1581                         page_cache_release(page);
1582                         return NULL;
1583                 }
1584                 goto page_not_uptodate;
1585         }
1586
1587 success:
1588         /*
1589          * Found the page and have a reference on it.
1590          */
1591         mark_page_accessed(page);
1592         return page;
1593
1594 no_cached_page:
1595         error = page_cache_read(file, pgoff);
1596
1597         /*
1598          * The page we want has now been added to the page cache.
1599          * In the unlikely event that someone removed it in the
1600          * meantime, we'll just come back here and read it again.
1601          */
1602         if (error >= 0)
1603                 goto retry_find;
1604
1605         /*
1606          * An error return from page_cache_read can result if the
1607          * system is low on memory, or a problem occurs while trying
1608          * to schedule I/O.
1609          */
1610         return NULL;
1611
1612 page_not_uptodate:
1613         lock_page(page);
1614
1615         /* Did it get unhashed while we waited for it? */
1616         if (!page->mapping) {
1617                 unlock_page(page);
1618                 goto err;
1619         }
1620
1621         /* Did somebody else get it up-to-date? */
1622         if (PageUptodate(page)) {
1623                 unlock_page(page);
1624                 goto success;
1625         }
1626
1627         error = mapping->a_ops->readpage(file, page);
1628         if (!error) {
1629                 wait_on_page_locked(page);
1630                 if (PageUptodate(page))
1631                         goto success;
1632         } else if (error == AOP_TRUNCATED_PAGE) {
1633                 page_cache_release(page);
1634                 goto retry_find;
1635         }
1636
1637         /*
1638          * Umm, take care of errors if the page isn't up-to-date.
1639          * Try to re-read it _once_. We do this synchronously,
1640          * because there really aren't any performance issues here
1641          * and we need to check for errors.
1642          */
1643         lock_page(page);
1644
1645         /* Somebody truncated the page on us? */
1646         if (!page->mapping) {
1647                 unlock_page(page);
1648                 goto err;
1649         }
1650         /* Somebody else successfully read it in? */
1651         if (PageUptodate(page)) {
1652                 unlock_page(page);
1653                 goto success;
1654         }
1655
1656         ClearPageError(page);
1657         error = mapping->a_ops->readpage(file, page);
1658         if (!error) {
1659                 wait_on_page_locked(page);
1660                 if (PageUptodate(page))
1661                         goto success;
1662         } else if (error == AOP_TRUNCATED_PAGE) {
1663                 page_cache_release(page);
1664                 goto retry_find;
1665         }
1666
1667         /*
1668          * Things didn't work out. Return zero to tell the
1669          * mm layer so, possibly freeing the page cache page first.
1670          */
1671 err:
1672         page_cache_release(page);
1673
1674         return NULL;
1675 }
1676
1677 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1678                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1679                 int nonblock)
1680 {
1681         struct file *file = vma->vm_file;
1682         struct address_space *mapping = file->f_mapping;
1683         struct inode *inode = mapping->host;
1684         unsigned long size;
1685         struct mm_struct *mm = vma->vm_mm;
1686         struct page *page;
1687         int err;
1688
1689         if (!nonblock)
1690                 force_page_cache_readahead(mapping, vma->vm_file,
1691                                         pgoff, len >> PAGE_CACHE_SHIFT);
1692
1693 repeat:
1694         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1695         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1696                 return -EINVAL;
1697
1698         page = filemap_getpage(file, pgoff, nonblock);
1699
1700         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1701          * done in shmem_populate calling shmem_getpage */
1702         if (!page && !nonblock)
1703                 return -ENOMEM;
1704
1705         if (page) {
1706                 err = install_page(mm, vma, addr, page, prot);
1707                 if (err) {
1708                         page_cache_release(page);
1709                         return err;
1710                 }
1711         } else if (vma->vm_flags & VM_NONLINEAR) {
1712                 /* No page was found just because we can't read it in now (being
1713                  * here implies nonblock != 0), but the page may exist, so set
1714                  * the PTE to fault it in later. */
1715                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1716                 if (err)
1717                         return err;
1718         }
1719
1720         len -= PAGE_SIZE;
1721         addr += PAGE_SIZE;
1722         pgoff++;
1723         if (len)
1724                 goto repeat;
1725
1726         return 0;
1727 }
1728 EXPORT_SYMBOL(filemap_populate);
1729
1730 struct vm_operations_struct generic_file_vm_ops = {
1731         .nopage         = filemap_nopage,
1732         .populate       = filemap_populate,
1733 };
1734
1735 /* This is used for a general mmap of a disk file */
1736
1737 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1738 {
1739         struct address_space *mapping = file->f_mapping;
1740
1741         if (!mapping->a_ops->readpage)
1742                 return -ENOEXEC;
1743         file_accessed(file);
1744         vma->vm_ops = &generic_file_vm_ops;
1745         return 0;
1746 }
1747
1748 /*
1749  * This is for filesystems which do not implement ->writepage.
1750  */
1751 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1752 {
1753         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1754                 return -EINVAL;
1755         return generic_file_mmap(file, vma);
1756 }
1757 #else
1758 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1759 {
1760         return -ENOSYS;
1761 }
1762 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1763 {
1764         return -ENOSYS;
1765 }
1766 #endif /* CONFIG_MMU */
1767
1768 EXPORT_SYMBOL(generic_file_mmap);
1769 EXPORT_SYMBOL(generic_file_readonly_mmap);
1770
1771 static inline struct page *__read_cache_page(struct address_space *mapping,
1772                                 unsigned long index,
1773                                 int (*filler)(void *,struct page*),
1774                                 void *data)
1775 {
1776         struct page *page, *cached_page = NULL;
1777         int err;
1778 repeat:
1779         page = find_get_page(mapping, index);
1780         if (!page) {
1781                 if (!cached_page) {
1782                         cached_page = page_cache_alloc_cold(mapping);
1783                         if (!cached_page)
1784                                 return ERR_PTR(-ENOMEM);
1785                 }
1786                 err = add_to_page_cache_lru(cached_page, mapping,
1787                                         index, GFP_KERNEL);
1788                 if (err == -EEXIST)
1789                         goto repeat;
1790                 if (err < 0) {
1791                         /* Presumably ENOMEM for radix tree node */
1792                         page_cache_release(cached_page);
1793                         return ERR_PTR(err);
1794                 }
1795                 page = cached_page;
1796                 cached_page = NULL;
1797                 err = filler(data, page);
1798                 if (err < 0) {
1799                         page_cache_release(page);
1800                         page = ERR_PTR(err);
1801                 }
1802         }
1803         if (cached_page)
1804                 page_cache_release(cached_page);
1805         return page;
1806 }
1807
1808 /**
1809  * read_cache_page - read into page cache, fill it if needed
1810  * @mapping:    the page's address_space
1811  * @index:      the page index
1812  * @filler:     function to perform the read
1813  * @data:       destination for read data
1814  *
1815  * Read into the page cache. If a page already exists,
1816  * and PageUptodate() is not set, try to fill the page.
1817  */
1818 struct page *read_cache_page(struct address_space *mapping,
1819                                 unsigned long index,
1820                                 int (*filler)(void *,struct page*),
1821                                 void *data)
1822 {
1823         struct page *page;
1824         int err;
1825
1826 retry:
1827         page = __read_cache_page(mapping, index, filler, data);
1828         if (IS_ERR(page))
1829                 goto out;
1830         mark_page_accessed(page);
1831         if (PageUptodate(page))
1832                 goto out;
1833
1834         lock_page(page);
1835         if (!page->mapping) {
1836                 unlock_page(page);
1837                 page_cache_release(page);
1838                 goto retry;
1839         }
1840         if (PageUptodate(page)) {
1841                 unlock_page(page);
1842                 goto out;
1843         }
1844         err = filler(data, page);
1845         if (err < 0) {
1846                 page_cache_release(page);
1847                 page = ERR_PTR(err);
1848         }
1849  out:
1850         return page;
1851 }
1852 EXPORT_SYMBOL(read_cache_page);
1853
1854 /*
1855  * If the page was newly created, increment its refcount and add it to the
1856  * caller's lru-buffering pagevec.  This function is specifically for
1857  * generic_file_write().
1858  */
1859 static inline struct page *
1860 __grab_cache_page(struct address_space *mapping, unsigned long index,
1861                         struct page **cached_page, struct pagevec *lru_pvec)
1862 {
1863         int err;
1864         struct page *page;
1865 repeat:
1866         page = find_lock_page(mapping, index);
1867         if (!page) {
1868                 if (!*cached_page) {
1869                         *cached_page = page_cache_alloc(mapping);
1870                         if (!*cached_page)
1871                                 return NULL;
1872                 }
1873                 err = add_to_page_cache(*cached_page, mapping,
1874                                         index, GFP_KERNEL);
1875                 if (err == -EEXIST)
1876                         goto repeat;
1877                 if (err == 0) {
1878                         page = *cached_page;
1879                         page_cache_get(page);
1880                         if (!pagevec_add(lru_pvec, page))
1881                                 __pagevec_lru_add(lru_pvec);
1882                         *cached_page = NULL;
1883                 }
1884         }
1885         return page;
1886 }
1887
1888 /*
1889  * The logic we want is
1890  *
1891  *      if suid or (sgid and xgrp)
1892  *              remove privs
1893  */
1894 int remove_suid(struct dentry *dentry)
1895 {
1896         mode_t mode = dentry->d_inode->i_mode;
1897         int kill = 0;
1898         int result = 0;
1899
1900         /* suid always must be killed */
1901         if (unlikely(mode & S_ISUID))
1902                 kill = ATTR_KILL_SUID;
1903
1904         /*
1905          * sgid without any exec bits is just a mandatory locking mark; leave
1906          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1907          */
1908         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1909                 kill |= ATTR_KILL_SGID;
1910
1911         if (unlikely(kill && !capable(CAP_FSETID))) {
1912                 struct iattr newattrs;
1913
1914                 newattrs.ia_valid = ATTR_FORCE | kill;
1915                 result = notify_change(dentry, &newattrs);
1916         }
1917         return result;
1918 }
1919 EXPORT_SYMBOL(remove_suid);
1920
1921 size_t
1922 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1923                         const struct iovec *iov, size_t base, size_t bytes)
1924 {
1925         size_t copied = 0, left = 0;
1926
1927         while (bytes) {
1928                 char __user *buf = iov->iov_base + base;
1929                 int copy = min(bytes, iov->iov_len - base);
1930
1931                 base = 0;
1932                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1933                 copied += copy;
1934                 bytes -= copy;
1935                 vaddr += copy;
1936                 iov++;
1937
1938                 if (unlikely(left))
1939                         break;
1940         }
1941         return copied - left;
1942 }
1943
1944 /*
1945  * Performs necessary checks before doing a write
1946  *
1947  * Can adjust writing position or amount of bytes to write.
1948  * Returns appropriate error code that caller should return or
1949  * zero in case that write should be allowed.
1950  */
1951 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1952 {
1953         struct inode *inode = file->f_mapping->host;
1954         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1955
1956         if (unlikely(*pos < 0))
1957                 return -EINVAL;
1958
1959         if (!isblk) {
1960                 /* FIXME: this is for backwards compatibility with 2.4 */
1961                 if (file->f_flags & O_APPEND)
1962                         *pos = i_size_read(inode);
1963
1964                 if (limit != RLIM_INFINITY) {
1965                         if (*pos >= limit) {
1966                                 send_sig(SIGXFSZ, current, 0);
1967                                 return -EFBIG;
1968                         }
1969                         if (*count > limit - (typeof(limit))*pos) {
1970                                 *count = limit - (typeof(limit))*pos;
1971                         }
1972                 }
1973         }
1974
1975         /*
1976          * LFS rule
1977          */
1978         if (unlikely(*pos + *count > MAX_NON_LFS &&
1979                                 !(file->f_flags & O_LARGEFILE))) {
1980                 if (*pos >= MAX_NON_LFS) {
1981                         send_sig(SIGXFSZ, current, 0);
1982                         return -EFBIG;
1983                 }
1984                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1985                         *count = MAX_NON_LFS - (unsigned long)*pos;
1986                 }
1987         }
1988
1989         /*
1990          * Are we about to exceed the fs block limit ?
1991          *
1992          * If we have written data it becomes a short write.  If we have
1993          * exceeded without writing data we send a signal and return EFBIG.
1994          * Linus frestrict idea will clean these up nicely..
1995          */
1996         if (likely(!isblk)) {
1997                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1998                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1999                                 send_sig(SIGXFSZ, current, 0);
2000                                 return -EFBIG;
2001                         }
2002                         /* zero-length writes at ->s_maxbytes are OK */
2003                 }
2004
2005                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2006                         *count = inode->i_sb->s_maxbytes - *pos;
2007         } else {
2008                 loff_t isize;
2009                 if (bdev_read_only(I_BDEV(inode)))
2010                         return -EPERM;
2011                 isize = i_size_read(inode);
2012                 if (*pos >= isize) {
2013                         if (*count || *pos > isize)
2014                                 return -ENOSPC;
2015                 }
2016
2017                 if (*pos + *count > isize)
2018                         *count = isize - *pos;
2019         }
2020         return 0;
2021 }
2022 EXPORT_SYMBOL(generic_write_checks);
2023
2024 ssize_t
2025 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2026                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2027                 size_t count, size_t ocount)
2028 {
2029         struct file     *file = iocb->ki_filp;
2030         struct address_space *mapping = file->f_mapping;
2031         struct inode    *inode = mapping->host;
2032         ssize_t         written;
2033
2034         if (count != ocount)
2035                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2036
2037         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2038         if (written > 0) {
2039                 loff_t end = pos + written;
2040                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2041                         i_size_write(inode,  end);
2042                         mark_inode_dirty(inode);
2043                 }
2044                 *ppos = end;
2045         }
2046
2047         /*
2048          * Sync the fs metadata but not the minor inode changes and
2049          * of course not the data as we did direct DMA for the IO.
2050          * i_mutex is held, which protects generic_osync_inode() from
2051          * livelocking.
2052          */
2053         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2054                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2055                 if (err < 0)
2056                         written = err;
2057         }
2058         if (written == count && !is_sync_kiocb(iocb))
2059                 written = -EIOCBQUEUED;
2060         return written;
2061 }
2062 EXPORT_SYMBOL(generic_file_direct_write);
2063
2064 ssize_t
2065 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2066                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2067                 size_t count, ssize_t written)
2068 {
2069         struct file *file = iocb->ki_filp;
2070         struct address_space * mapping = file->f_mapping;
2071         const struct address_space_operations *a_ops = mapping->a_ops;
2072         struct inode    *inode = mapping->host;
2073         long            status = 0;
2074         struct page     *page;
2075         struct page     *cached_page = NULL;
2076         size_t          bytes;
2077         struct pagevec  lru_pvec;
2078         const struct iovec *cur_iov = iov; /* current iovec */
2079         size_t          iov_base = 0;      /* offset in the current iovec */
2080         char __user     *buf;
2081
2082         pagevec_init(&lru_pvec, 0);
2083
2084         /*
2085          * handle partial DIO write.  Adjust cur_iov if needed.
2086          */
2087         if (likely(nr_segs == 1))
2088                 buf = iov->iov_base + written;
2089         else {
2090                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
2091                 buf = cur_iov->iov_base + iov_base;
2092         }
2093
2094         do {
2095                 unsigned long index;
2096                 unsigned long offset;
2097                 size_t copied;
2098
2099                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
2100                 index = pos >> PAGE_CACHE_SHIFT;
2101                 bytes = PAGE_CACHE_SIZE - offset;
2102
2103                 /* Limit the size of the copy to the caller's write size */
2104                 bytes = min(bytes, count);
2105
2106                 /*
2107                  * Limit the size of the copy to that of the current segment,
2108                  * because fault_in_pages_readable() doesn't know how to walk
2109                  * segments.
2110                  */
2111                 bytes = min(bytes, cur_iov->iov_len - iov_base);
2112
2113                 /*
2114                  * Bring in the user page that we will copy from _first_.
2115                  * Otherwise there's a nasty deadlock on copying from the
2116                  * same page as we're writing to, without it being marked
2117                  * up-to-date.
2118                  */
2119                 fault_in_pages_readable(buf, bytes);
2120
2121                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
2122                 if (!page) {
2123                         status = -ENOMEM;
2124                         break;
2125                 }
2126
2127                 if (unlikely(bytes == 0)) {
2128                         status = 0;
2129                         copied = 0;
2130                         goto zero_length_segment;
2131                 }
2132
2133                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2134                 if (unlikely(status)) {
2135                         loff_t isize = i_size_read(inode);
2136
2137                         if (status != AOP_TRUNCATED_PAGE)
2138                                 unlock_page(page);
2139                         page_cache_release(page);
2140                         if (status == AOP_TRUNCATED_PAGE)
2141                                 continue;
2142                         /*
2143                          * prepare_write() may have instantiated a few blocks
2144                          * outside i_size.  Trim these off again.
2145                          */
2146                         if (pos + bytes > isize)
2147                                 vmtruncate(inode, isize);
2148                         break;
2149                 }
2150                 if (likely(nr_segs == 1))
2151                         copied = filemap_copy_from_user(page, offset,
2152                                                         buf, bytes);
2153                 else
2154                         copied = filemap_copy_from_user_iovec(page, offset,
2155                                                 cur_iov, iov_base, bytes);
2156                 flush_dcache_page(page);
2157                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2158                 if (status == AOP_TRUNCATED_PAGE) {
2159                         page_cache_release(page);
2160                         continue;
2161                 }
2162 zero_length_segment:
2163                 if (likely(copied >= 0)) {
2164                         if (!status)
2165                                 status = copied;
2166
2167                         if (status >= 0) {
2168                                 written += status;
2169                                 count -= status;
2170                                 pos += status;
2171                                 buf += status;
2172                                 if (unlikely(nr_segs > 1)) {
2173                                         filemap_set_next_iovec(&cur_iov,
2174                                                         &iov_base, status);
2175                                         if (count)
2176                                                 buf = cur_iov->iov_base +
2177                                                         iov_base;
2178                                 } else {
2179                                         iov_base += status;
2180                                 }
2181                         }
2182                 }
2183                 if (unlikely(copied != bytes))
2184                         if (status >= 0)
2185                                 status = -EFAULT;
2186                 unlock_page(page);
2187                 mark_page_accessed(page);
2188                 page_cache_release(page);
2189                 if (status < 0)
2190                         break;
2191                 balance_dirty_pages_ratelimited(mapping);
2192                 cond_resched();
2193         } while (count);
2194         *ppos = pos;
2195
2196         if (cached_page)
2197                 page_cache_release(cached_page);
2198
2199         /*
2200          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2201          */
2202         if (likely(status >= 0)) {
2203                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2204                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2205                                 status = generic_osync_inode(inode, mapping,
2206                                                 OSYNC_METADATA|OSYNC_DATA);
2207                 }
2208         }
2209         
2210         /*
2211          * If we get here for O_DIRECT writes then we must have fallen through
2212          * to buffered writes (block instantiation inside i_size).  So we sync
2213          * the file data here, to try to honour O_DIRECT expectations.
2214          */
2215         if (unlikely(file->f_flags & O_DIRECT) && written)
2216                 status = filemap_write_and_wait(mapping);
2217
2218         pagevec_lru_add(&lru_pvec);
2219         return written ? written : status;
2220 }
2221 EXPORT_SYMBOL(generic_file_buffered_write);
2222
2223 static ssize_t
2224 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2225                                 unsigned long nr_segs, loff_t *ppos)
2226 {
2227         struct file *file = iocb->ki_filp;
2228         const struct address_space * mapping = file->f_mapping;
2229         size_t ocount;          /* original count */
2230         size_t count;           /* after file limit checks */
2231         struct inode    *inode = mapping->host;
2232         unsigned long   seg;
2233         loff_t          pos;
2234         ssize_t         written;
2235         ssize_t         err;
2236
2237         ocount = 0;
2238         for (seg = 0; seg < nr_segs; seg++) {
2239                 const struct iovec *iv = &iov[seg];
2240
2241                 /*
2242                  * If any segment has a negative length, or the cumulative
2243                  * length ever wraps negative then return -EINVAL.
2244                  */
2245                 ocount += iv->iov_len;
2246                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2247                         return -EINVAL;
2248                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2249                         continue;
2250                 if (seg == 0)
2251                         return -EFAULT;
2252                 nr_segs = seg;
2253                 ocount -= iv->iov_len;  /* This segment is no good */
2254                 break;
2255         }
2256
2257         count = ocount;
2258         pos = *ppos;
2259
2260         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2261
2262         /* We can write back this queue in page reclaim */
2263         current->backing_dev_info = mapping->backing_dev_info;
2264         written = 0;
2265
2266         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2267         if (err)
2268                 goto out;
2269
2270         if (count == 0)
2271                 goto out;
2272
2273         err = remove_suid(file->f_dentry);
2274         if (err)
2275                 goto out;
2276
2277         file_update_time(file);
2278
2279         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2280         if (unlikely(file->f_flags & O_DIRECT)) {
2281                 written = generic_file_direct_write(iocb, iov,
2282                                 &nr_segs, pos, ppos, count, ocount);
2283                 if (written < 0 || written == count)
2284                         goto out;
2285                 /*
2286                  * direct-io write to a hole: fall through to buffered I/O
2287                  * for completing the rest of the request.
2288                  */
2289                 pos += written;
2290                 count -= written;
2291         }
2292
2293         written = generic_file_buffered_write(iocb, iov, nr_segs,
2294                         pos, ppos, count, written);
2295 out:
2296         current->backing_dev_info = NULL;
2297         return written ? written : err;
2298 }
2299 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2300
2301 ssize_t
2302 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2303                                 unsigned long nr_segs, loff_t *ppos)
2304 {
2305         struct file *file = iocb->ki_filp;
2306         struct address_space *mapping = file->f_mapping;
2307         struct inode *inode = mapping->host;
2308         ssize_t ret;
2309         loff_t pos = *ppos;
2310
2311         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2312
2313         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2314                 int err;
2315
2316                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2317                 if (err < 0)
2318                         ret = err;
2319         }
2320         return ret;
2321 }
2322
2323 static ssize_t
2324 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2325                                 unsigned long nr_segs, loff_t *ppos)
2326 {
2327         struct kiocb kiocb;
2328         ssize_t ret;
2329
2330         init_sync_kiocb(&kiocb, file);
2331         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2332         if (ret == -EIOCBQUEUED)
2333                 ret = wait_on_sync_kiocb(&kiocb);
2334         return ret;
2335 }
2336
2337 ssize_t
2338 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2339                                 unsigned long nr_segs, loff_t *ppos)
2340 {
2341         struct kiocb kiocb;
2342         ssize_t ret;
2343
2344         init_sync_kiocb(&kiocb, file);
2345         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2346         if (-EIOCBQUEUED == ret)
2347                 ret = wait_on_sync_kiocb(&kiocb);
2348         return ret;
2349 }
2350 EXPORT_SYMBOL(generic_file_write_nolock);
2351
2352 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2353                                size_t count, loff_t pos)
2354 {
2355         struct file *file = iocb->ki_filp;
2356         struct address_space *mapping = file->f_mapping;
2357         struct inode *inode = mapping->host;
2358         ssize_t ret;
2359         struct iovec local_iov = { .iov_base = (void __user *)buf,
2360                                         .iov_len = count };
2361
2362         BUG_ON(iocb->ki_pos != pos);
2363
2364         mutex_lock(&inode->i_mutex);
2365         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2366                                                 &iocb->ki_pos);
2367         mutex_unlock(&inode->i_mutex);
2368
2369         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2370                 ssize_t err;
2371
2372                 err = sync_page_range(inode, mapping, pos, ret);
2373                 if (err < 0)
2374                         ret = err;
2375         }
2376         return ret;
2377 }
2378 EXPORT_SYMBOL(generic_file_aio_write);
2379
2380 ssize_t generic_file_write(struct file *file, const char __user *buf,
2381                            size_t count, loff_t *ppos)
2382 {
2383         struct address_space *mapping = file->f_mapping;
2384         struct inode *inode = mapping->host;
2385         ssize_t ret;
2386         struct iovec local_iov = { .iov_base = (void __user *)buf,
2387                                         .iov_len = count };
2388
2389         mutex_lock(&inode->i_mutex);
2390         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2391         mutex_unlock(&inode->i_mutex);
2392
2393         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2394                 ssize_t err;
2395
2396                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2397                 if (err < 0)
2398                         ret = err;
2399         }
2400         return ret;
2401 }
2402 EXPORT_SYMBOL(generic_file_write);
2403
2404 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2405                         unsigned long nr_segs, loff_t *ppos)
2406 {
2407         struct kiocb kiocb;
2408         ssize_t ret;
2409
2410         init_sync_kiocb(&kiocb, filp);
2411         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2412         if (-EIOCBQUEUED == ret)
2413                 ret = wait_on_sync_kiocb(&kiocb);
2414         return ret;
2415 }
2416 EXPORT_SYMBOL(generic_file_readv);
2417
2418 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2419                         unsigned long nr_segs, loff_t *ppos)
2420 {
2421         struct address_space *mapping = file->f_mapping;
2422         struct inode *inode = mapping->host;
2423         ssize_t ret;
2424
2425         mutex_lock(&inode->i_mutex);
2426         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2427         mutex_unlock(&inode->i_mutex);
2428
2429         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2430                 int err;
2431
2432                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2433                 if (err < 0)
2434                         ret = err;
2435         }
2436         return ret;
2437 }
2438 EXPORT_SYMBOL(generic_file_writev);
2439
2440 /*
2441  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2442  * went wrong during pagecache shootdown.
2443  */
2444 static ssize_t
2445 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2446         loff_t offset, unsigned long nr_segs)
2447 {
2448         struct file *file = iocb->ki_filp;
2449         struct address_space *mapping = file->f_mapping;
2450         ssize_t retval;
2451         size_t write_len = 0;
2452
2453         /*
2454          * If it's a write, unmap all mmappings of the file up-front.  This
2455          * will cause any pte dirty bits to be propagated into the pageframes
2456          * for the subsequent filemap_write_and_wait().
2457          */
2458         if (rw == WRITE) {
2459                 write_len = iov_length(iov, nr_segs);
2460                 if (mapping_mapped(mapping))
2461                         unmap_mapping_range(mapping, offset, write_len, 0);
2462         }
2463
2464         retval = filemap_write_and_wait(mapping);
2465         if (retval == 0) {
2466                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2467                                                 offset, nr_segs);
2468                 if (rw == WRITE && mapping->nrpages) {
2469                         pgoff_t end = (offset + write_len - 1)
2470                                                 >> PAGE_CACHE_SHIFT;
2471                         int err = invalidate_inode_pages2_range(mapping,
2472                                         offset >> PAGE_CACHE_SHIFT, end);
2473                         if (err)
2474                                 retval = err;
2475                 }
2476         }
2477         return retval;
2478 }