Merge branch 'master' into gfs2
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/security.h>
31 #include <linux/syscalls.h>
32 #include <linux/cpuset.h>
33 #include "filemap.h"
34 #include "internal.h"
35
36 /*
37  * FIXME: remove all knowledge of the buffer layer from the core VM
38  */
39 #include <linux/buffer_head.h> /* for generic_osync_inode */
40
41 #include <asm/mman.h>
42
43 static ssize_t
44 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
45         loff_t offset, unsigned long nr_segs);
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_lock               (vmtruncate)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_lock
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->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 static int __sleep_on_page_lock(void *word)
492 {
493         io_schedule();
494         return 0;
495 }
496
497 /*
498  * In order to wait for pages to become available there must be
499  * waitqueues associated with pages. By using a hash table of
500  * waitqueues where the bucket discipline is to maintain all
501  * waiters on the same queue and wake all when any of the pages
502  * become available, and for the woken contexts to check to be
503  * sure the appropriate page became available, this saves space
504  * at a cost of "thundering herd" phenomena during rare hash
505  * collisions.
506  */
507 static wait_queue_head_t *page_waitqueue(struct page *page)
508 {
509         const struct zone *zone = page_zone(page);
510
511         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
512 }
513
514 static inline void wake_up_page(struct page *page, int bit)
515 {
516         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
517 }
518
519 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
520 {
521         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
522
523         if (test_bit(bit_nr, &page->flags))
524                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
525                                                         TASK_UNINTERRUPTIBLE);
526 }
527 EXPORT_SYMBOL(wait_on_page_bit);
528
529 /**
530  * unlock_page - unlock a locked page
531  * @page: the page
532  *
533  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
534  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
535  * mechananism between PageLocked pages and PageWriteback pages is shared.
536  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
537  *
538  * The first mb is necessary to safely close the critical section opened by the
539  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
540  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
541  * parallel wait_on_page_locked()).
542  */
543 void fastcall unlock_page(struct page *page)
544 {
545         smp_mb__before_clear_bit();
546         if (!TestClearPageLocked(page))
547                 BUG();
548         smp_mb__after_clear_bit(); 
549         wake_up_page(page, PG_locked);
550 }
551 EXPORT_SYMBOL(unlock_page);
552
553 /**
554  * end_page_writeback - end writeback against a page
555  * @page: the page
556  */
557 void end_page_writeback(struct page *page)
558 {
559         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
560                 if (!test_clear_page_writeback(page))
561                         BUG();
562         }
563         smp_mb__after_clear_bit();
564         wake_up_page(page, PG_writeback);
565 }
566 EXPORT_SYMBOL(end_page_writeback);
567
568 /**
569  * __lock_page - get a lock on the page, assuming we need to sleep to get it
570  * @page: the page to lock
571  *
572  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
573  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
574  * chances are that on the second loop, the block layer's plug list is empty,
575  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
576  */
577 void fastcall __lock_page(struct page *page)
578 {
579         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
580
581         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
582                                                         TASK_UNINTERRUPTIBLE);
583 }
584 EXPORT_SYMBOL(__lock_page);
585
586 /*
587  * Variant of lock_page that does not require the caller to hold a reference
588  * on the page's mapping.
589  */
590 void fastcall __lock_page_nosync(struct page *page)
591 {
592         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
593         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
594                                                         TASK_UNINTERRUPTIBLE);
595 }
596
597 /**
598  * find_get_page - find and get a page reference
599  * @mapping: the address_space to search
600  * @offset: the page index
601  *
602  * Is there a pagecache struct page at the given (mapping, offset) tuple?
603  * If yes, increment its refcount and return it; if no, return NULL.
604  */
605 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
606 {
607         struct page *page;
608
609         read_lock_irq(&mapping->tree_lock);
610         page = radix_tree_lookup(&mapping->page_tree, offset);
611         if (page)
612                 page_cache_get(page);
613         read_unlock_irq(&mapping->tree_lock);
614         return page;
615 }
616 EXPORT_SYMBOL(find_get_page);
617
618 /**
619  * find_trylock_page - find and lock a page
620  * @mapping: the address_space to search
621  * @offset: the page index
622  *
623  * Same as find_get_page(), but trylock it instead of incrementing the count.
624  */
625 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
626 {
627         struct page *page;
628
629         read_lock_irq(&mapping->tree_lock);
630         page = radix_tree_lookup(&mapping->page_tree, offset);
631         if (page && TestSetPageLocked(page))
632                 page = NULL;
633         read_unlock_irq(&mapping->tree_lock);
634         return page;
635 }
636 EXPORT_SYMBOL(find_trylock_page);
637
638 /**
639  * find_lock_page - locate, pin and lock a pagecache page
640  * @mapping: the address_space to search
641  * @offset: the page index
642  *
643  * Locates the desired pagecache page, locks it, increments its reference
644  * count and returns its address.
645  *
646  * Returns zero if the page was not present. find_lock_page() may sleep.
647  */
648 struct page *find_lock_page(struct address_space *mapping,
649                                 unsigned long offset)
650 {
651         struct page *page;
652
653         read_lock_irq(&mapping->tree_lock);
654 repeat:
655         page = radix_tree_lookup(&mapping->page_tree, offset);
656         if (page) {
657                 page_cache_get(page);
658                 if (TestSetPageLocked(page)) {
659                         read_unlock_irq(&mapping->tree_lock);
660                         __lock_page(page);
661                         read_lock_irq(&mapping->tree_lock);
662
663                         /* Has the page been truncated while we slept? */
664                         if (unlikely(page->mapping != mapping ||
665                                      page->index != offset)) {
666                                 unlock_page(page);
667                                 page_cache_release(page);
668                                 goto repeat;
669                         }
670                 }
671         }
672         read_unlock_irq(&mapping->tree_lock);
673         return page;
674 }
675 EXPORT_SYMBOL(find_lock_page);
676
677 /**
678  * find_or_create_page - locate or add a pagecache page
679  * @mapping: the page's address_space
680  * @index: the page's index into the mapping
681  * @gfp_mask: page allocation mode
682  *
683  * Locates a page in the pagecache.  If the page is not present, a new page
684  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
685  * LRU list.  The returned page is locked and has its reference count
686  * incremented.
687  *
688  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
689  * allocation!
690  *
691  * find_or_create_page() returns the desired page's address, or zero on
692  * memory exhaustion.
693  */
694 struct page *find_or_create_page(struct address_space *mapping,
695                 unsigned long index, gfp_t gfp_mask)
696 {
697         struct page *page, *cached_page = NULL;
698         int err;
699 repeat:
700         page = find_lock_page(mapping, index);
701         if (!page) {
702                 if (!cached_page) {
703                         cached_page = alloc_page(gfp_mask);
704                         if (!cached_page)
705                                 return NULL;
706                 }
707                 err = add_to_page_cache_lru(cached_page, mapping,
708                                         index, gfp_mask);
709                 if (!err) {
710                         page = cached_page;
711                         cached_page = NULL;
712                 } else if (err == -EEXIST)
713                         goto repeat;
714         }
715         if (cached_page)
716                 page_cache_release(cached_page);
717         return page;
718 }
719 EXPORT_SYMBOL(find_or_create_page);
720
721 /**
722  * find_get_pages - gang pagecache lookup
723  * @mapping:    The address_space to search
724  * @start:      The starting page index
725  * @nr_pages:   The maximum number of pages
726  * @pages:      Where the resulting pages are placed
727  *
728  * find_get_pages() will search for and return a group of up to
729  * @nr_pages pages in the mapping.  The pages are placed at @pages.
730  * find_get_pages() takes a reference against the returned pages.
731  *
732  * The search returns a group of mapping-contiguous pages with ascending
733  * indexes.  There may be holes in the indices due to not-present pages.
734  *
735  * find_get_pages() returns the number of pages which were found.
736  */
737 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
738                             unsigned int nr_pages, struct page **pages)
739 {
740         unsigned int i;
741         unsigned int ret;
742
743         read_lock_irq(&mapping->tree_lock);
744         ret = radix_tree_gang_lookup(&mapping->page_tree,
745                                 (void **)pages, start, nr_pages);
746         for (i = 0; i < ret; i++)
747                 page_cache_get(pages[i]);
748         read_unlock_irq(&mapping->tree_lock);
749         return ret;
750 }
751
752 /**
753  * find_get_pages_contig - gang contiguous pagecache lookup
754  * @mapping:    The address_space to search
755  * @index:      The starting page index
756  * @nr_pages:   The maximum number of pages
757  * @pages:      Where the resulting pages are placed
758  *
759  * find_get_pages_contig() works exactly like find_get_pages(), except
760  * that the returned number of pages are guaranteed to be contiguous.
761  *
762  * find_get_pages_contig() returns the number of pages which were found.
763  */
764 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
765                                unsigned int nr_pages, struct page **pages)
766 {
767         unsigned int i;
768         unsigned int ret;
769
770         read_lock_irq(&mapping->tree_lock);
771         ret = radix_tree_gang_lookup(&mapping->page_tree,
772                                 (void **)pages, index, nr_pages);
773         for (i = 0; i < ret; i++) {
774                 if (pages[i]->mapping == NULL || pages[i]->index != index)
775                         break;
776
777                 page_cache_get(pages[i]);
778                 index++;
779         }
780         read_unlock_irq(&mapping->tree_lock);
781         return i;
782 }
783
784 /**
785  * find_get_pages_tag - find and return pages that match @tag
786  * @mapping:    the address_space to search
787  * @index:      the starting page index
788  * @tag:        the tag index
789  * @nr_pages:   the maximum number of pages
790  * @pages:      where the resulting pages are placed
791  *
792  * Like find_get_pages, except we only return pages which are tagged with
793  * @tag.   We update @index to index the next page for the traversal.
794  */
795 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
796                         int tag, unsigned int nr_pages, struct page **pages)
797 {
798         unsigned int i;
799         unsigned int ret;
800
801         read_lock_irq(&mapping->tree_lock);
802         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
803                                 (void **)pages, *index, nr_pages, tag);
804         for (i = 0; i < ret; i++)
805                 page_cache_get(pages[i]);
806         if (ret)
807                 *index = pages[ret - 1]->index + 1;
808         read_unlock_irq(&mapping->tree_lock);
809         return ret;
810 }
811
812 /**
813  * grab_cache_page_nowait - returns locked page at given index in given cache
814  * @mapping: target address_space
815  * @index: the page index
816  *
817  * Same as grab_cache_page, but do not wait if the page is unavailable.
818  * This is intended for speculative data generators, where the data can
819  * be regenerated if the page couldn't be grabbed.  This routine should
820  * be safe to call while holding the lock for another page.
821  *
822  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
823  * and deadlock against the caller's locked page.
824  */
825 struct page *
826 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
827 {
828         struct page *page = find_get_page(mapping, index);
829         gfp_t gfp_mask;
830
831         if (page) {
832                 if (!TestSetPageLocked(page))
833                         return page;
834                 page_cache_release(page);
835                 return NULL;
836         }
837         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
838         page = alloc_pages(gfp_mask, 0);
839         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
840                 page_cache_release(page);
841                 page = NULL;
842         }
843         return page;
844 }
845 EXPORT_SYMBOL(grab_cache_page_nowait);
846
847 /*
848  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
849  * a _large_ part of the i/o request. Imagine the worst scenario:
850  *
851  *      ---R__________________________________________B__________
852  *         ^ reading here                             ^ bad block(assume 4k)
853  *
854  * read(R) => miss => readahead(R...B) => media error => frustrating retries
855  * => failing the whole request => read(R) => read(R+1) =>
856  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
857  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
858  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
859  *
860  * It is going insane. Fix it by quickly scaling down the readahead size.
861  */
862 static void shrink_readahead_size_eio(struct file *filp,
863                                         struct file_ra_state *ra)
864 {
865         if (!ra->ra_pages)
866                 return;
867
868         ra->ra_pages /= 4;
869 }
870
871 /**
872  * do_generic_mapping_read - generic file read routine
873  * @mapping:    address_space to be read
874  * @_ra:        file's readahead state
875  * @filp:       the file to read
876  * @ppos:       current file position
877  * @desc:       read_descriptor
878  * @actor:      read method
879  *
880  * This is a generic file read routine, and uses the
881  * mapping->a_ops->readpage() function for the actual low-level stuff.
882  *
883  * This is really ugly. But the goto's actually try to clarify some
884  * of the logic when it comes to error handling etc.
885  *
886  * Note the struct file* is only passed for the use of readpage.
887  * It may be NULL.
888  */
889 void do_generic_mapping_read(struct address_space *mapping,
890                              struct file_ra_state *_ra,
891                              struct file *filp,
892                              loff_t *ppos,
893                              read_descriptor_t *desc,
894                              read_actor_t actor)
895 {
896         struct inode *inode = mapping->host;
897         unsigned long index;
898         unsigned long end_index;
899         unsigned long offset;
900         unsigned long last_index;
901         unsigned long next_index;
902         unsigned long prev_index;
903         loff_t isize;
904         struct page *cached_page;
905         int error;
906         struct file_ra_state ra = *_ra;
907
908         cached_page = NULL;
909         index = *ppos >> PAGE_CACHE_SHIFT;
910         next_index = index;
911         prev_index = ra.prev_page;
912         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
913         offset = *ppos & ~PAGE_CACHE_MASK;
914
915         isize = i_size_read(inode);
916         if (!isize)
917                 goto out;
918
919         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
920         for (;;) {
921                 struct page *page;
922                 unsigned long nr, ret;
923
924                 /* nr is the maximum number of bytes to copy from this page */
925                 nr = PAGE_CACHE_SIZE;
926                 if (index >= end_index) {
927                         if (index > end_index)
928                                 goto out;
929                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
930                         if (nr <= offset) {
931                                 goto out;
932                         }
933                 }
934                 nr = nr - offset;
935
936                 cond_resched();
937                 if (index == next_index)
938                         next_index = page_cache_readahead(mapping, &ra, filp,
939                                         index, last_index - index);
940
941 find_page:
942                 page = find_get_page(mapping, index);
943                 if (unlikely(page == NULL)) {
944                         handle_ra_miss(mapping, &ra, index);
945                         goto no_cached_page;
946                 }
947                 if (!PageUptodate(page))
948                         goto page_not_up_to_date;
949 page_ok:
950
951                 /* If users can be writing to this page using arbitrary
952                  * virtual addresses, take care about potential aliasing
953                  * before reading the page on the kernel side.
954                  */
955                 if (mapping_writably_mapped(mapping))
956                         flush_dcache_page(page);
957
958                 /*
959                  * When (part of) the same page is read multiple times
960                  * in succession, only mark it as accessed the first time.
961                  */
962                 if (prev_index != index)
963                         mark_page_accessed(page);
964                 prev_index = index;
965
966                 /*
967                  * Ok, we have the page, and it's up-to-date, so
968                  * now we can copy it to user space...
969                  *
970                  * The actor routine returns how many bytes were actually used..
971                  * NOTE! This may not be the same as how much of a user buffer
972                  * we filled up (we may be padding etc), so we can only update
973                  * "pos" here (the actor routine has to update the user buffer
974                  * pointers and the remaining count).
975                  */
976                 ret = actor(desc, page, offset, nr);
977                 offset += ret;
978                 index += offset >> PAGE_CACHE_SHIFT;
979                 offset &= ~PAGE_CACHE_MASK;
980
981                 page_cache_release(page);
982                 if (ret == nr && desc->count)
983                         continue;
984                 goto out;
985
986 page_not_up_to_date:
987                 /* Get exclusive access to the page ... */
988                 lock_page(page);
989
990                 /* Did it get truncated before we got the lock? */
991                 if (!page->mapping) {
992                         unlock_page(page);
993                         page_cache_release(page);
994                         continue;
995                 }
996
997                 /* Did somebody else fill it already? */
998                 if (PageUptodate(page)) {
999                         unlock_page(page);
1000                         goto page_ok;
1001                 }
1002
1003 readpage:
1004                 /* Start the actual read. The read will unlock the page. */
1005                 error = mapping->a_ops->readpage(filp, page);
1006
1007                 if (unlikely(error)) {
1008                         if (error == AOP_TRUNCATED_PAGE) {
1009                                 page_cache_release(page);
1010                                 goto find_page;
1011                         }
1012                         goto readpage_error;
1013                 }
1014
1015                 if (!PageUptodate(page)) {
1016                         lock_page(page);
1017                         if (!PageUptodate(page)) {
1018                                 if (page->mapping == NULL) {
1019                                         /*
1020                                          * invalidate_inode_pages got it
1021                                          */
1022                                         unlock_page(page);
1023                                         page_cache_release(page);
1024                                         goto find_page;
1025                                 }
1026                                 unlock_page(page);
1027                                 error = -EIO;
1028                                 shrink_readahead_size_eio(filp, &ra);
1029                                 goto readpage_error;
1030                         }
1031                         unlock_page(page);
1032                 }
1033
1034                 /*
1035                  * i_size must be checked after we have done ->readpage.
1036                  *
1037                  * Checking i_size after the readpage allows us to calculate
1038                  * the correct value for "nr", which means the zero-filled
1039                  * part of the page is not copied back to userspace (unless
1040                  * another truncate extends the file - this is desired though).
1041                  */
1042                 isize = i_size_read(inode);
1043                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1044                 if (unlikely(!isize || index > end_index)) {
1045                         page_cache_release(page);
1046                         goto out;
1047                 }
1048
1049                 /* nr is the maximum number of bytes to copy from this page */
1050                 nr = PAGE_CACHE_SIZE;
1051                 if (index == end_index) {
1052                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1053                         if (nr <= offset) {
1054                                 page_cache_release(page);
1055                                 goto out;
1056                         }
1057                 }
1058                 nr = nr - offset;
1059                 goto page_ok;
1060
1061 readpage_error:
1062                 /* UHHUH! A synchronous read error occurred. Report it */
1063                 desc->error = error;
1064                 page_cache_release(page);
1065                 goto out;
1066
1067 no_cached_page:
1068                 /*
1069                  * Ok, it wasn't cached, so we need to create a new
1070                  * page..
1071                  */
1072                 if (!cached_page) {
1073                         cached_page = page_cache_alloc_cold(mapping);
1074                         if (!cached_page) {
1075                                 desc->error = -ENOMEM;
1076                                 goto out;
1077                         }
1078                 }
1079                 error = add_to_page_cache_lru(cached_page, mapping,
1080                                                 index, GFP_KERNEL);
1081                 if (error) {
1082                         if (error == -EEXIST)
1083                                 goto find_page;
1084                         desc->error = error;
1085                         goto out;
1086                 }
1087                 page = cached_page;
1088                 cached_page = NULL;
1089                 goto readpage;
1090         }
1091
1092 out:
1093         *_ra = ra;
1094
1095         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1096         if (cached_page)
1097                 page_cache_release(cached_page);
1098         if (filp)
1099                 file_accessed(filp);
1100 }
1101 EXPORT_SYMBOL(do_generic_mapping_read);
1102
1103 int file_read_actor(read_descriptor_t *desc, struct page *page,
1104                         unsigned long offset, unsigned long size)
1105 {
1106         char *kaddr;
1107         unsigned long left, count = desc->count;
1108
1109         if (size > count)
1110                 size = count;
1111
1112         /*
1113          * Faults on the destination of a read are common, so do it before
1114          * taking the kmap.
1115          */
1116         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1117                 kaddr = kmap_atomic(page, KM_USER0);
1118                 left = __copy_to_user_inatomic(desc->arg.buf,
1119                                                 kaddr + offset, size);
1120                 kunmap_atomic(kaddr, KM_USER0);
1121                 if (left == 0)
1122                         goto success;
1123         }
1124
1125         /* Do it the slow way */
1126         kaddr = kmap(page);
1127         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1128         kunmap(page);
1129
1130         if (left) {
1131                 size -= left;
1132                 desc->error = -EFAULT;
1133         }
1134 success:
1135         desc->count = count - size;
1136         desc->written += size;
1137         desc->arg.buf += size;
1138         return size;
1139 }
1140
1141 /**
1142  * __generic_file_aio_read - generic filesystem read routine
1143  * @iocb:       kernel I/O control block
1144  * @iov:        io vector request
1145  * @nr_segs:    number of segments in the iovec
1146  * @ppos:       current file position
1147  *
1148  * This is the "read()" routine for all filesystems
1149  * that can use the page cache directly.
1150  */
1151 ssize_t
1152 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1153                 unsigned long nr_segs, loff_t *ppos)
1154 {
1155         struct file *filp = iocb->ki_filp;
1156         ssize_t retval;
1157         unsigned long seg;
1158         size_t count;
1159
1160         count = 0;
1161         for (seg = 0; seg < nr_segs; seg++) {
1162                 const struct iovec *iv = &iov[seg];
1163
1164                 /*
1165                  * If any segment has a negative length, or the cumulative
1166                  * length ever wraps negative then return -EINVAL.
1167                  */
1168                 count += iv->iov_len;
1169                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1170                         return -EINVAL;
1171                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1172                         continue;
1173                 if (seg == 0)
1174                         return -EFAULT;
1175                 nr_segs = seg;
1176                 count -= iv->iov_len;   /* This segment is no good */
1177                 break;
1178         }
1179
1180         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1181         if (filp->f_flags & O_DIRECT) {
1182                 loff_t pos = *ppos, size;
1183                 struct address_space *mapping;
1184                 struct inode *inode;
1185
1186                 mapping = filp->f_mapping;
1187                 inode = mapping->host;
1188                 retval = 0;
1189                 if (!count)
1190                         goto out; /* skip atime */
1191                 size = i_size_read(inode);
1192                 if (pos < size) {
1193                         retval = generic_file_direct_IO(READ, iocb,
1194                                                 iov, pos, nr_segs);
1195                         if (retval > 0 && !is_sync_kiocb(iocb))
1196                                 retval = -EIOCBQUEUED;
1197                         if (retval > 0)
1198                                 *ppos = pos + retval;
1199                 }
1200                 if (likely(retval != 0)) {
1201                         file_accessed(filp);
1202                         goto out;
1203                 }
1204         }
1205
1206         retval = 0;
1207         if (count) {
1208                 for (seg = 0; seg < nr_segs; seg++) {
1209                         read_descriptor_t desc;
1210
1211                         desc.written = 0;
1212                         desc.arg.buf = iov[seg].iov_base;
1213                         desc.count = iov[seg].iov_len;
1214                         if (desc.count == 0)
1215                                 continue;
1216                         desc.error = 0;
1217                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1218                         retval += desc.written;
1219                         if (desc.error) {
1220                                 retval = retval ?: desc.error;
1221                                 break;
1222                         }
1223                 }
1224         }
1225 out:
1226         return retval;
1227 }
1228 EXPORT_SYMBOL(__generic_file_aio_read);
1229
1230 ssize_t
1231 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1232 {
1233         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1234
1235         BUG_ON(iocb->ki_pos != pos);
1236         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1237 }
1238 EXPORT_SYMBOL(generic_file_aio_read);
1239
1240 ssize_t
1241 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1242 {
1243         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1244         struct kiocb kiocb;
1245         ssize_t ret;
1246
1247         init_sync_kiocb(&kiocb, filp);
1248         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1249         if (-EIOCBQUEUED == ret)
1250                 ret = wait_on_sync_kiocb(&kiocb);
1251         return ret;
1252 }
1253 EXPORT_SYMBOL(generic_file_read);
1254
1255 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1256 {
1257         ssize_t written;
1258         unsigned long count = desc->count;
1259         struct file *file = desc->arg.data;
1260
1261         if (size > count)
1262                 size = count;
1263
1264         written = file->f_op->sendpage(file, page, offset,
1265                                        size, &file->f_pos, size<count);
1266         if (written < 0) {
1267                 desc->error = written;
1268                 written = 0;
1269         }
1270         desc->count = count - written;
1271         desc->written += written;
1272         return written;
1273 }
1274
1275 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1276                          size_t count, read_actor_t actor, void *target)
1277 {
1278         read_descriptor_t desc;
1279
1280         if (!count)
1281                 return 0;
1282
1283         desc.written = 0;
1284         desc.count = count;
1285         desc.arg.data = target;
1286         desc.error = 0;
1287
1288         do_generic_file_read(in_file, ppos, &desc, actor);
1289         if (desc.written)
1290                 return desc.written;
1291         return desc.error;
1292 }
1293 EXPORT_SYMBOL(generic_file_sendfile);
1294
1295 static ssize_t
1296 do_readahead(struct address_space *mapping, struct file *filp,
1297              unsigned long index, unsigned long nr)
1298 {
1299         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1300                 return -EINVAL;
1301
1302         force_page_cache_readahead(mapping, filp, index,
1303                                         max_sane_readahead(nr));
1304         return 0;
1305 }
1306
1307 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1308 {
1309         ssize_t ret;
1310         struct file *file;
1311
1312         ret = -EBADF;
1313         file = fget(fd);
1314         if (file) {
1315                 if (file->f_mode & FMODE_READ) {
1316                         struct address_space *mapping = file->f_mapping;
1317                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1318                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1319                         unsigned long len = end - start + 1;
1320                         ret = do_readahead(mapping, file, start, len);
1321                 }
1322                 fput(file);
1323         }
1324         return ret;
1325 }
1326
1327 #ifdef CONFIG_MMU
1328 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1329 /**
1330  * page_cache_read - adds requested page to the page cache if not already there
1331  * @file:       file to read
1332  * @offset:     page index
1333  *
1334  * This adds the requested page to the page cache if it isn't already there,
1335  * and schedules an I/O to read in its contents from disk.
1336  */
1337 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1338 {
1339         struct address_space *mapping = file->f_mapping;
1340         struct page *page; 
1341         int ret;
1342
1343         do {
1344                 page = page_cache_alloc_cold(mapping);
1345                 if (!page)
1346                         return -ENOMEM;
1347
1348                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1349                 if (ret == 0)
1350                         ret = mapping->a_ops->readpage(file, page);
1351                 else if (ret == -EEXIST)
1352                         ret = 0; /* losing race to add is OK */
1353
1354                 page_cache_release(page);
1355
1356         } while (ret == AOP_TRUNCATED_PAGE);
1357                 
1358         return ret;
1359 }
1360
1361 #define MMAP_LOTSAMISS  (100)
1362
1363 /**
1364  * filemap_nopage - read in file data for page fault handling
1365  * @area:       the applicable vm_area
1366  * @address:    target address to read in
1367  * @type:       returned with VM_FAULT_{MINOR,MAJOR} if not %NULL
1368  *
1369  * filemap_nopage() is invoked via the vma operations vector for a
1370  * mapped memory region to read in file data during a page fault.
1371  *
1372  * The goto's are kind of ugly, but this streamlines the normal case of having
1373  * it in the page cache, and handles the special cases reasonably without
1374  * having a lot of duplicated code.
1375  */
1376 struct page *filemap_nopage(struct vm_area_struct *area,
1377                                 unsigned long address, int *type)
1378 {
1379         int error;
1380         struct file *file = area->vm_file;
1381         struct address_space *mapping = file->f_mapping;
1382         struct file_ra_state *ra = &file->f_ra;
1383         struct inode *inode = mapping->host;
1384         struct page *page;
1385         unsigned long size, pgoff;
1386         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1387
1388         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1389
1390 retry_all:
1391         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1392         if (pgoff >= size)
1393                 goto outside_data_content;
1394
1395         /* If we don't want any read-ahead, don't bother */
1396         if (VM_RandomReadHint(area))
1397                 goto no_cached_page;
1398
1399         /*
1400          * The readahead code wants to be told about each and every page
1401          * so it can build and shrink its windows appropriately
1402          *
1403          * For sequential accesses, we use the generic readahead logic.
1404          */
1405         if (VM_SequentialReadHint(area))
1406                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1407
1408         /*
1409          * Do we have something in the page cache already?
1410          */
1411 retry_find:
1412         page = find_get_page(mapping, pgoff);
1413         if (!page) {
1414                 unsigned long ra_pages;
1415
1416                 if (VM_SequentialReadHint(area)) {
1417                         handle_ra_miss(mapping, ra, pgoff);
1418                         goto no_cached_page;
1419                 }
1420                 ra->mmap_miss++;
1421
1422                 /*
1423                  * Do we miss much more than hit in this file? If so,
1424                  * stop bothering with read-ahead. It will only hurt.
1425                  */
1426                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1427                         goto no_cached_page;
1428
1429                 /*
1430                  * To keep the pgmajfault counter straight, we need to
1431                  * check did_readaround, as this is an inner loop.
1432                  */
1433                 if (!did_readaround) {
1434                         majmin = VM_FAULT_MAJOR;
1435                         count_vm_event(PGMAJFAULT);
1436                 }
1437                 did_readaround = 1;
1438                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1439                 if (ra_pages) {
1440                         pgoff_t start = 0;
1441
1442                         if (pgoff > ra_pages / 2)
1443                                 start = pgoff - ra_pages / 2;
1444                         do_page_cache_readahead(mapping, file, start, ra_pages);
1445                 }
1446                 page = find_get_page(mapping, pgoff);
1447                 if (!page)
1448                         goto no_cached_page;
1449         }
1450
1451         if (!did_readaround)
1452                 ra->mmap_hit++;
1453
1454         /*
1455          * Ok, found a page in the page cache, now we need to check
1456          * that it's up-to-date.
1457          */
1458         if (!PageUptodate(page))
1459                 goto page_not_uptodate;
1460
1461 success:
1462         /*
1463          * Found the page and have a reference on it.
1464          */
1465         mark_page_accessed(page);
1466         if (type)
1467                 *type = majmin;
1468         return page;
1469
1470 outside_data_content:
1471         /*
1472          * An external ptracer can access pages that normally aren't
1473          * accessible..
1474          */
1475         if (area->vm_mm == current->mm)
1476                 return NULL;
1477         /* Fall through to the non-read-ahead case */
1478 no_cached_page:
1479         /*
1480          * We're only likely to ever get here if MADV_RANDOM is in
1481          * effect.
1482          */
1483         error = page_cache_read(file, pgoff);
1484         grab_swap_token();
1485
1486         /*
1487          * The page we want has now been added to the page cache.
1488          * In the unlikely event that someone removed it in the
1489          * meantime, we'll just come back here and read it again.
1490          */
1491         if (error >= 0)
1492                 goto retry_find;
1493
1494         /*
1495          * An error return from page_cache_read can result if the
1496          * system is low on memory, or a problem occurs while trying
1497          * to schedule I/O.
1498          */
1499         if (error == -ENOMEM)
1500                 return NOPAGE_OOM;
1501         return NULL;
1502
1503 page_not_uptodate:
1504         if (!did_readaround) {
1505                 majmin = VM_FAULT_MAJOR;
1506                 count_vm_event(PGMAJFAULT);
1507         }
1508         lock_page(page);
1509
1510         /* Did it get unhashed while we waited for it? */
1511         if (!page->mapping) {
1512                 unlock_page(page);
1513                 page_cache_release(page);
1514                 goto retry_all;
1515         }
1516
1517         /* Did somebody else get it up-to-date? */
1518         if (PageUptodate(page)) {
1519                 unlock_page(page);
1520                 goto success;
1521         }
1522
1523         error = mapping->a_ops->readpage(file, page);
1524         if (!error) {
1525                 wait_on_page_locked(page);
1526                 if (PageUptodate(page))
1527                         goto success;
1528         } else if (error == AOP_TRUNCATED_PAGE) {
1529                 page_cache_release(page);
1530                 goto retry_find;
1531         }
1532
1533         /*
1534          * Umm, take care of errors if the page isn't up-to-date.
1535          * Try to re-read it _once_. We do this synchronously,
1536          * because there really aren't any performance issues here
1537          * and we need to check for errors.
1538          */
1539         lock_page(page);
1540
1541         /* Somebody truncated the page on us? */
1542         if (!page->mapping) {
1543                 unlock_page(page);
1544                 page_cache_release(page);
1545                 goto retry_all;
1546         }
1547
1548         /* Somebody else successfully read it in? */
1549         if (PageUptodate(page)) {
1550                 unlock_page(page);
1551                 goto success;
1552         }
1553         ClearPageError(page);
1554         error = mapping->a_ops->readpage(file, page);
1555         if (!error) {
1556                 wait_on_page_locked(page);
1557                 if (PageUptodate(page))
1558                         goto success;
1559         } else if (error == AOP_TRUNCATED_PAGE) {
1560                 page_cache_release(page);
1561                 goto retry_find;
1562         }
1563
1564         /*
1565          * Things didn't work out. Return zero to tell the
1566          * mm layer so, possibly freeing the page cache page first.
1567          */
1568         shrink_readahead_size_eio(file, ra);
1569         page_cache_release(page);
1570         return NULL;
1571 }
1572 EXPORT_SYMBOL(filemap_nopage);
1573
1574 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1575                                         int nonblock)
1576 {
1577         struct address_space *mapping = file->f_mapping;
1578         struct page *page;
1579         int error;
1580
1581         /*
1582          * Do we have something in the page cache already?
1583          */
1584 retry_find:
1585         page = find_get_page(mapping, pgoff);
1586         if (!page) {
1587                 if (nonblock)
1588                         return NULL;
1589                 goto no_cached_page;
1590         }
1591
1592         /*
1593          * Ok, found a page in the page cache, now we need to check
1594          * that it's up-to-date.
1595          */
1596         if (!PageUptodate(page)) {
1597                 if (nonblock) {
1598                         page_cache_release(page);
1599                         return NULL;
1600                 }
1601                 goto page_not_uptodate;
1602         }
1603
1604 success:
1605         /*
1606          * Found the page and have a reference on it.
1607          */
1608         mark_page_accessed(page);
1609         return page;
1610
1611 no_cached_page:
1612         error = page_cache_read(file, pgoff);
1613
1614         /*
1615          * The page we want has now been added to the page cache.
1616          * In the unlikely event that someone removed it in the
1617          * meantime, we'll just come back here and read it again.
1618          */
1619         if (error >= 0)
1620                 goto retry_find;
1621
1622         /*
1623          * An error return from page_cache_read can result if the
1624          * system is low on memory, or a problem occurs while trying
1625          * to schedule I/O.
1626          */
1627         return NULL;
1628
1629 page_not_uptodate:
1630         lock_page(page);
1631
1632         /* Did it get truncated while we waited for it? */
1633         if (!page->mapping) {
1634                 unlock_page(page);
1635                 goto err;
1636         }
1637
1638         /* Did somebody else get it up-to-date? */
1639         if (PageUptodate(page)) {
1640                 unlock_page(page);
1641                 goto success;
1642         }
1643
1644         error = mapping->a_ops->readpage(file, page);
1645         if (!error) {
1646                 wait_on_page_locked(page);
1647                 if (PageUptodate(page))
1648                         goto success;
1649         } else if (error == AOP_TRUNCATED_PAGE) {
1650                 page_cache_release(page);
1651                 goto retry_find;
1652         }
1653
1654         /*
1655          * Umm, take care of errors if the page isn't up-to-date.
1656          * Try to re-read it _once_. We do this synchronously,
1657          * because there really aren't any performance issues here
1658          * and we need to check for errors.
1659          */
1660         lock_page(page);
1661
1662         /* Somebody truncated the page on us? */
1663         if (!page->mapping) {
1664                 unlock_page(page);
1665                 goto err;
1666         }
1667         /* Somebody else successfully read it in? */
1668         if (PageUptodate(page)) {
1669                 unlock_page(page);
1670                 goto success;
1671         }
1672
1673         ClearPageError(page);
1674         error = mapping->a_ops->readpage(file, page);
1675         if (!error) {
1676                 wait_on_page_locked(page);
1677                 if (PageUptodate(page))
1678                         goto success;
1679         } else if (error == AOP_TRUNCATED_PAGE) {
1680                 page_cache_release(page);
1681                 goto retry_find;
1682         }
1683
1684         /*
1685          * Things didn't work out. Return zero to tell the
1686          * mm layer so, possibly freeing the page cache page first.
1687          */
1688 err:
1689         page_cache_release(page);
1690
1691         return NULL;
1692 }
1693
1694 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1695                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1696                 int nonblock)
1697 {
1698         struct file *file = vma->vm_file;
1699         struct address_space *mapping = file->f_mapping;
1700         struct inode *inode = mapping->host;
1701         unsigned long size;
1702         struct mm_struct *mm = vma->vm_mm;
1703         struct page *page;
1704         int err;
1705
1706         if (!nonblock)
1707                 force_page_cache_readahead(mapping, vma->vm_file,
1708                                         pgoff, len >> PAGE_CACHE_SHIFT);
1709
1710 repeat:
1711         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1712         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1713                 return -EINVAL;
1714
1715         page = filemap_getpage(file, pgoff, nonblock);
1716
1717         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1718          * done in shmem_populate calling shmem_getpage */
1719         if (!page && !nonblock)
1720                 return -ENOMEM;
1721
1722         if (page) {
1723                 err = install_page(mm, vma, addr, page, prot);
1724                 if (err) {
1725                         page_cache_release(page);
1726                         return err;
1727                 }
1728         } else if (vma->vm_flags & VM_NONLINEAR) {
1729                 /* No page was found just because we can't read it in now (being
1730                  * here implies nonblock != 0), but the page may exist, so set
1731                  * the PTE to fault it in later. */
1732                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1733                 if (err)
1734                         return err;
1735         }
1736
1737         len -= PAGE_SIZE;
1738         addr += PAGE_SIZE;
1739         pgoff++;
1740         if (len)
1741                 goto repeat;
1742
1743         return 0;
1744 }
1745 EXPORT_SYMBOL(filemap_populate);
1746
1747 struct vm_operations_struct generic_file_vm_ops = {
1748         .nopage         = filemap_nopage,
1749         .populate       = filemap_populate,
1750 };
1751
1752 /* This is used for a general mmap of a disk file */
1753
1754 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1755 {
1756         struct address_space *mapping = file->f_mapping;
1757
1758         if (!mapping->a_ops->readpage)
1759                 return -ENOEXEC;
1760         file_accessed(file);
1761         vma->vm_ops = &generic_file_vm_ops;
1762         return 0;
1763 }
1764
1765 /*
1766  * This is for filesystems which do not implement ->writepage.
1767  */
1768 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1769 {
1770         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1771                 return -EINVAL;
1772         return generic_file_mmap(file, vma);
1773 }
1774 #else
1775 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1776 {
1777         return -ENOSYS;
1778 }
1779 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1780 {
1781         return -ENOSYS;
1782 }
1783 #endif /* CONFIG_MMU */
1784
1785 EXPORT_SYMBOL(generic_file_mmap);
1786 EXPORT_SYMBOL(generic_file_readonly_mmap);
1787
1788 static inline struct page *__read_cache_page(struct address_space *mapping,
1789                                 unsigned long index,
1790                                 int (*filler)(void *,struct page*),
1791                                 void *data)
1792 {
1793         struct page *page, *cached_page = NULL;
1794         int err;
1795 repeat:
1796         page = find_get_page(mapping, index);
1797         if (!page) {
1798                 if (!cached_page) {
1799                         cached_page = page_cache_alloc_cold(mapping);
1800                         if (!cached_page)
1801                                 return ERR_PTR(-ENOMEM);
1802                 }
1803                 err = add_to_page_cache_lru(cached_page, mapping,
1804                                         index, GFP_KERNEL);
1805                 if (err == -EEXIST)
1806                         goto repeat;
1807                 if (err < 0) {
1808                         /* Presumably ENOMEM for radix tree node */
1809                         page_cache_release(cached_page);
1810                         return ERR_PTR(err);
1811                 }
1812                 page = cached_page;
1813                 cached_page = NULL;
1814                 err = filler(data, page);
1815                 if (err < 0) {
1816                         page_cache_release(page);
1817                         page = ERR_PTR(err);
1818                 }
1819         }
1820         if (cached_page)
1821                 page_cache_release(cached_page);
1822         return page;
1823 }
1824
1825 /**
1826  * read_cache_page - read into page cache, fill it if needed
1827  * @mapping:    the page's address_space
1828  * @index:      the page index
1829  * @filler:     function to perform the read
1830  * @data:       destination for read data
1831  *
1832  * Read into the page cache. If a page already exists,
1833  * and PageUptodate() is not set, try to fill the page.
1834  */
1835 struct page *read_cache_page(struct address_space *mapping,
1836                                 unsigned long index,
1837                                 int (*filler)(void *,struct page*),
1838                                 void *data)
1839 {
1840         struct page *page;
1841         int err;
1842
1843 retry:
1844         page = __read_cache_page(mapping, index, filler, data);
1845         if (IS_ERR(page))
1846                 goto out;
1847         mark_page_accessed(page);
1848         if (PageUptodate(page))
1849                 goto out;
1850
1851         lock_page(page);
1852         if (!page->mapping) {
1853                 unlock_page(page);
1854                 page_cache_release(page);
1855                 goto retry;
1856         }
1857         if (PageUptodate(page)) {
1858                 unlock_page(page);
1859                 goto out;
1860         }
1861         err = filler(data, page);
1862         if (err < 0) {
1863                 page_cache_release(page);
1864                 page = ERR_PTR(err);
1865         }
1866  out:
1867         return page;
1868 }
1869 EXPORT_SYMBOL(read_cache_page);
1870
1871 /*
1872  * If the page was newly created, increment its refcount and add it to the
1873  * caller's lru-buffering pagevec.  This function is specifically for
1874  * generic_file_write().
1875  */
1876 static inline struct page *
1877 __grab_cache_page(struct address_space *mapping, unsigned long index,
1878                         struct page **cached_page, struct pagevec *lru_pvec)
1879 {
1880         int err;
1881         struct page *page;
1882 repeat:
1883         page = find_lock_page(mapping, index);
1884         if (!page) {
1885                 if (!*cached_page) {
1886                         *cached_page = page_cache_alloc(mapping);
1887                         if (!*cached_page)
1888                                 return NULL;
1889                 }
1890                 err = add_to_page_cache(*cached_page, mapping,
1891                                         index, GFP_KERNEL);
1892                 if (err == -EEXIST)
1893                         goto repeat;
1894                 if (err == 0) {
1895                         page = *cached_page;
1896                         page_cache_get(page);
1897                         if (!pagevec_add(lru_pvec, page))
1898                                 __pagevec_lru_add(lru_pvec);
1899                         *cached_page = NULL;
1900                 }
1901         }
1902         return page;
1903 }
1904
1905 /*
1906  * The logic we want is
1907  *
1908  *      if suid or (sgid and xgrp)
1909  *              remove privs
1910  */
1911 int remove_suid(struct dentry *dentry)
1912 {
1913         mode_t mode = dentry->d_inode->i_mode;
1914         int kill = 0;
1915         int result = 0;
1916
1917         /* suid always must be killed */
1918         if (unlikely(mode & S_ISUID))
1919                 kill = ATTR_KILL_SUID;
1920
1921         /*
1922          * sgid without any exec bits is just a mandatory locking mark; leave
1923          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1924          */
1925         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1926                 kill |= ATTR_KILL_SGID;
1927
1928         if (unlikely(kill && !capable(CAP_FSETID))) {
1929                 struct iattr newattrs;
1930
1931                 newattrs.ia_valid = ATTR_FORCE | kill;
1932                 result = notify_change(dentry, &newattrs);
1933         }
1934         return result;
1935 }
1936 EXPORT_SYMBOL(remove_suid);
1937
1938 size_t
1939 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1940                         const struct iovec *iov, size_t base, size_t bytes)
1941 {
1942         size_t copied = 0, left = 0;
1943
1944         while (bytes) {
1945                 char __user *buf = iov->iov_base + base;
1946                 int copy = min(bytes, iov->iov_len - base);
1947
1948                 base = 0;
1949                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1950                 copied += copy;
1951                 bytes -= copy;
1952                 vaddr += copy;
1953                 iov++;
1954
1955                 if (unlikely(left))
1956                         break;
1957         }
1958         return copied - left;
1959 }
1960
1961 /*
1962  * Performs necessary checks before doing a write
1963  *
1964  * Can adjust writing position or amount of bytes to write.
1965  * Returns appropriate error code that caller should return or
1966  * zero in case that write should be allowed.
1967  */
1968 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1969 {
1970         struct inode *inode = file->f_mapping->host;
1971         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1972
1973         if (unlikely(*pos < 0))
1974                 return -EINVAL;
1975
1976         if (!isblk) {
1977                 /* FIXME: this is for backwards compatibility with 2.4 */
1978                 if (file->f_flags & O_APPEND)
1979                         *pos = i_size_read(inode);
1980
1981                 if (limit != RLIM_INFINITY) {
1982                         if (*pos >= limit) {
1983                                 send_sig(SIGXFSZ, current, 0);
1984                                 return -EFBIG;
1985                         }
1986                         if (*count > limit - (typeof(limit))*pos) {
1987                                 *count = limit - (typeof(limit))*pos;
1988                         }
1989                 }
1990         }
1991
1992         /*
1993          * LFS rule
1994          */
1995         if (unlikely(*pos + *count > MAX_NON_LFS &&
1996                                 !(file->f_flags & O_LARGEFILE))) {
1997                 if (*pos >= MAX_NON_LFS) {
1998                         send_sig(SIGXFSZ, current, 0);
1999                         return -EFBIG;
2000                 }
2001                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2002                         *count = MAX_NON_LFS - (unsigned long)*pos;
2003                 }
2004         }
2005
2006         /*
2007          * Are we about to exceed the fs block limit ?
2008          *
2009          * If we have written data it becomes a short write.  If we have
2010          * exceeded without writing data we send a signal and return EFBIG.
2011          * Linus frestrict idea will clean these up nicely..
2012          */
2013         if (likely(!isblk)) {
2014                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2015                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2016                                 send_sig(SIGXFSZ, current, 0);
2017                                 return -EFBIG;
2018                         }
2019                         /* zero-length writes at ->s_maxbytes are OK */
2020                 }
2021
2022                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2023                         *count = inode->i_sb->s_maxbytes - *pos;
2024         } else {
2025                 loff_t isize;
2026                 if (bdev_read_only(I_BDEV(inode)))
2027                         return -EPERM;
2028                 isize = i_size_read(inode);
2029                 if (*pos >= isize) {
2030                         if (*count || *pos > isize)
2031                                 return -ENOSPC;
2032                 }
2033
2034                 if (*pos + *count > isize)
2035                         *count = isize - *pos;
2036         }
2037         return 0;
2038 }
2039 EXPORT_SYMBOL(generic_write_checks);
2040
2041 ssize_t
2042 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2043                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2044                 size_t count, size_t ocount)
2045 {
2046         struct file     *file = iocb->ki_filp;
2047         struct address_space *mapping = file->f_mapping;
2048         struct inode    *inode = mapping->host;
2049         ssize_t         written;
2050
2051         if (count != ocount)
2052                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2053
2054         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2055         if (written > 0) {
2056                 loff_t end = pos + written;
2057                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2058                         i_size_write(inode,  end);
2059                         mark_inode_dirty(inode);
2060                 }
2061                 *ppos = end;
2062         }
2063
2064         /*
2065          * Sync the fs metadata but not the minor inode changes and
2066          * of course not the data as we did direct DMA for the IO.
2067          * i_mutex is held, which protects generic_osync_inode() from
2068          * livelocking.
2069          */
2070         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2071                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2072                 if (err < 0)
2073                         written = err;
2074         }
2075         if (written == count && !is_sync_kiocb(iocb))
2076                 written = -EIOCBQUEUED;
2077         return written;
2078 }
2079 EXPORT_SYMBOL(generic_file_direct_write);
2080
2081 ssize_t
2082 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2083                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2084                 size_t count, ssize_t written)
2085 {
2086         struct file *file = iocb->ki_filp;
2087         struct address_space * mapping = file->f_mapping;
2088         const struct address_space_operations *a_ops = mapping->a_ops;
2089         struct inode    *inode = mapping->host;
2090         long            status = 0;
2091         struct page     *page;
2092         struct page     *cached_page = NULL;
2093         size_t          bytes;
2094         struct pagevec  lru_pvec;
2095         const struct iovec *cur_iov = iov; /* current iovec */
2096         size_t          iov_base = 0;      /* offset in the current iovec */
2097         char __user     *buf;
2098
2099         pagevec_init(&lru_pvec, 0);
2100
2101         /*
2102          * handle partial DIO write.  Adjust cur_iov if needed.
2103          */
2104         if (likely(nr_segs == 1))
2105                 buf = iov->iov_base + written;
2106         else {
2107                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
2108                 buf = cur_iov->iov_base + iov_base;
2109         }
2110
2111         do {
2112                 unsigned long index;
2113                 unsigned long offset;
2114                 size_t copied;
2115
2116                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
2117                 index = pos >> PAGE_CACHE_SHIFT;
2118                 bytes = PAGE_CACHE_SIZE - offset;
2119
2120                 /* Limit the size of the copy to the caller's write size */
2121                 bytes = min(bytes, count);
2122
2123                 /*
2124                  * Limit the size of the copy to that of the current segment,
2125                  * because fault_in_pages_readable() doesn't know how to walk
2126                  * segments.
2127                  */
2128                 bytes = min(bytes, cur_iov->iov_len - iov_base);
2129
2130                 /*
2131                  * Bring in the user page that we will copy from _first_.
2132                  * Otherwise there's a nasty deadlock on copying from the
2133                  * same page as we're writing to, without it being marked
2134                  * up-to-date.
2135                  */
2136                 fault_in_pages_readable(buf, bytes);
2137
2138                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
2139                 if (!page) {
2140                         status = -ENOMEM;
2141                         break;
2142                 }
2143
2144                 if (unlikely(bytes == 0)) {
2145                         status = 0;
2146                         copied = 0;
2147                         goto zero_length_segment;
2148                 }
2149
2150                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2151                 if (unlikely(status)) {
2152                         loff_t isize = i_size_read(inode);
2153
2154                         if (status != AOP_TRUNCATED_PAGE)
2155                                 unlock_page(page);
2156                         page_cache_release(page);
2157                         if (status == AOP_TRUNCATED_PAGE)
2158                                 continue;
2159                         /*
2160                          * prepare_write() may have instantiated a few blocks
2161                          * outside i_size.  Trim these off again.
2162                          */
2163                         if (pos + bytes > isize)
2164                                 vmtruncate(inode, isize);
2165                         break;
2166                 }
2167                 if (likely(nr_segs == 1))
2168                         copied = filemap_copy_from_user(page, offset,
2169                                                         buf, bytes);
2170                 else
2171                         copied = filemap_copy_from_user_iovec(page, offset,
2172                                                 cur_iov, iov_base, bytes);
2173                 flush_dcache_page(page);
2174                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2175                 if (status == AOP_TRUNCATED_PAGE) {
2176                         page_cache_release(page);
2177                         continue;
2178                 }
2179 zero_length_segment:
2180                 if (likely(copied >= 0)) {
2181                         if (!status)
2182                                 status = copied;
2183
2184                         if (status >= 0) {
2185                                 written += status;
2186                                 count -= status;
2187                                 pos += status;
2188                                 buf += status;
2189                                 if (unlikely(nr_segs > 1)) {
2190                                         filemap_set_next_iovec(&cur_iov,
2191                                                         &iov_base, status);
2192                                         if (count)
2193                                                 buf = cur_iov->iov_base +
2194                                                         iov_base;
2195                                 } else {
2196                                         iov_base += status;
2197                                 }
2198                         }
2199                 }
2200                 if (unlikely(copied != bytes))
2201                         if (status >= 0)
2202                                 status = -EFAULT;
2203                 unlock_page(page);
2204                 mark_page_accessed(page);
2205                 page_cache_release(page);
2206                 if (status < 0)
2207                         break;
2208                 balance_dirty_pages_ratelimited(mapping);
2209                 cond_resched();
2210         } while (count);
2211         *ppos = pos;
2212
2213         if (cached_page)
2214                 page_cache_release(cached_page);
2215
2216         /*
2217          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2218          */
2219         if (likely(status >= 0)) {
2220                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2221                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2222                                 status = generic_osync_inode(inode, mapping,
2223                                                 OSYNC_METADATA|OSYNC_DATA);
2224                 }
2225         }
2226         
2227         /*
2228          * If we get here for O_DIRECT writes then we must have fallen through
2229          * to buffered writes (block instantiation inside i_size).  So we sync
2230          * the file data here, to try to honour O_DIRECT expectations.
2231          */
2232         if (unlikely(file->f_flags & O_DIRECT) && written)
2233                 status = filemap_write_and_wait(mapping);
2234
2235         pagevec_lru_add(&lru_pvec);
2236         return written ? written : status;
2237 }
2238 EXPORT_SYMBOL(generic_file_buffered_write);
2239
2240 static ssize_t
2241 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2242                                 unsigned long nr_segs, loff_t *ppos)
2243 {
2244         struct file *file = iocb->ki_filp;
2245         const struct address_space * mapping = file->f_mapping;
2246         size_t ocount;          /* original count */
2247         size_t count;           /* after file limit checks */
2248         struct inode    *inode = mapping->host;
2249         unsigned long   seg;
2250         loff_t          pos;
2251         ssize_t         written;
2252         ssize_t         err;
2253
2254         ocount = 0;
2255         for (seg = 0; seg < nr_segs; seg++) {
2256                 const struct iovec *iv = &iov[seg];
2257
2258                 /*
2259                  * If any segment has a negative length, or the cumulative
2260                  * length ever wraps negative then return -EINVAL.
2261                  */
2262                 ocount += iv->iov_len;
2263                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2264                         return -EINVAL;
2265                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2266                         continue;
2267                 if (seg == 0)
2268                         return -EFAULT;
2269                 nr_segs = seg;
2270                 ocount -= iv->iov_len;  /* This segment is no good */
2271                 break;
2272         }
2273
2274         count = ocount;
2275         pos = *ppos;
2276
2277         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2278
2279         /* We can write back this queue in page reclaim */
2280         current->backing_dev_info = mapping->backing_dev_info;
2281         written = 0;
2282
2283         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2284         if (err)
2285                 goto out;
2286
2287         if (count == 0)
2288                 goto out;
2289
2290         err = remove_suid(file->f_dentry);
2291         if (err)
2292                 goto out;
2293
2294         file_update_time(file);
2295
2296         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2297         if (unlikely(file->f_flags & O_DIRECT)) {
2298                 written = generic_file_direct_write(iocb, iov,
2299                                 &nr_segs, pos, ppos, count, ocount);
2300                 if (written < 0 || written == count)
2301                         goto out;
2302                 /*
2303                  * direct-io write to a hole: fall through to buffered I/O
2304                  * for completing the rest of the request.
2305                  */
2306                 pos += written;
2307                 count -= written;
2308         }
2309
2310         written = generic_file_buffered_write(iocb, iov, nr_segs,
2311                         pos, ppos, count, written);
2312 out:
2313         current->backing_dev_info = NULL;
2314         return written ? written : err;
2315 }
2316 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2317
2318 ssize_t
2319 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2320                                 unsigned long nr_segs, loff_t *ppos)
2321 {
2322         struct file *file = iocb->ki_filp;
2323         struct address_space *mapping = file->f_mapping;
2324         struct inode *inode = mapping->host;
2325         ssize_t ret;
2326         loff_t pos = *ppos;
2327
2328         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2329
2330         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2331                 int err;
2332
2333                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2334                 if (err < 0)
2335                         ret = err;
2336         }
2337         return ret;
2338 }
2339
2340 static ssize_t
2341 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2342                                 unsigned long nr_segs, loff_t *ppos)
2343 {
2344         struct kiocb kiocb;
2345         ssize_t ret;
2346
2347         init_sync_kiocb(&kiocb, file);
2348         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2349         if (ret == -EIOCBQUEUED)
2350                 ret = wait_on_sync_kiocb(&kiocb);
2351         return ret;
2352 }
2353
2354 ssize_t
2355 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2356                                 unsigned long nr_segs, loff_t *ppos)
2357 {
2358         struct kiocb kiocb;
2359         ssize_t ret;
2360
2361         init_sync_kiocb(&kiocb, file);
2362         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2363         if (-EIOCBQUEUED == ret)
2364                 ret = wait_on_sync_kiocb(&kiocb);
2365         return ret;
2366 }
2367 EXPORT_SYMBOL(generic_file_write_nolock);
2368
2369 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2370                                size_t count, loff_t pos)
2371 {
2372         struct file *file = iocb->ki_filp;
2373         struct address_space *mapping = file->f_mapping;
2374         struct inode *inode = mapping->host;
2375         ssize_t ret;
2376         struct iovec local_iov = { .iov_base = (void __user *)buf,
2377                                         .iov_len = count };
2378
2379         BUG_ON(iocb->ki_pos != pos);
2380
2381         mutex_lock(&inode->i_mutex);
2382         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2383                                                 &iocb->ki_pos);
2384         mutex_unlock(&inode->i_mutex);
2385
2386         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2387                 ssize_t err;
2388
2389                 err = sync_page_range(inode, mapping, pos, ret);
2390                 if (err < 0)
2391                         ret = err;
2392         }
2393         return ret;
2394 }
2395 EXPORT_SYMBOL(generic_file_aio_write);
2396
2397 ssize_t generic_file_write(struct file *file, const char __user *buf,
2398                            size_t count, loff_t *ppos)
2399 {
2400         struct address_space *mapping = file->f_mapping;
2401         struct inode *inode = mapping->host;
2402         ssize_t ret;
2403         struct iovec local_iov = { .iov_base = (void __user *)buf,
2404                                         .iov_len = count };
2405
2406         mutex_lock(&inode->i_mutex);
2407         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2408         mutex_unlock(&inode->i_mutex);
2409
2410         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2411                 ssize_t err;
2412
2413                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2414                 if (err < 0)
2415                         ret = err;
2416         }
2417         return ret;
2418 }
2419 EXPORT_SYMBOL(generic_file_write);
2420
2421 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2422                         unsigned long nr_segs, loff_t *ppos)
2423 {
2424         struct kiocb kiocb;
2425         ssize_t ret;
2426
2427         init_sync_kiocb(&kiocb, filp);
2428         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2429         if (-EIOCBQUEUED == ret)
2430                 ret = wait_on_sync_kiocb(&kiocb);
2431         return ret;
2432 }
2433 EXPORT_SYMBOL(generic_file_readv);
2434
2435 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2436                         unsigned long nr_segs, loff_t *ppos)
2437 {
2438         struct address_space *mapping = file->f_mapping;
2439         struct inode *inode = mapping->host;
2440         ssize_t ret;
2441
2442         mutex_lock(&inode->i_mutex);
2443         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2444         mutex_unlock(&inode->i_mutex);
2445
2446         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2447                 int err;
2448
2449                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2450                 if (err < 0)
2451                         ret = err;
2452         }
2453         return ret;
2454 }
2455 EXPORT_SYMBOL(generic_file_writev);
2456
2457 /*
2458  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2459  * went wrong during pagecache shootdown.
2460  */
2461 static ssize_t
2462 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2463         loff_t offset, unsigned long nr_segs)
2464 {
2465         struct file *file = iocb->ki_filp;
2466         struct address_space *mapping = file->f_mapping;
2467         ssize_t retval;
2468         size_t write_len = 0;
2469
2470         /*
2471          * If it's a write, unmap all mmappings of the file up-front.  This
2472          * will cause any pte dirty bits to be propagated into the pageframes
2473          * for the subsequent filemap_write_and_wait().
2474          */
2475         if (rw == WRITE) {
2476                 write_len = iov_length(iov, nr_segs);
2477                 if (mapping_mapped(mapping))
2478                         unmap_mapping_range(mapping, offset, write_len, 0);
2479         }
2480
2481         retval = filemap_write_and_wait(mapping);
2482         if (retval == 0) {
2483                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2484                                                 offset, nr_segs);
2485                 if (rw == WRITE && mapping->nrpages) {
2486                         pgoff_t end = (offset + write_len - 1)
2487                                                 >> PAGE_CACHE_SHIFT;
2488                         int err = invalidate_inode_pages2_range(mapping,
2489                                         offset >> PAGE_CACHE_SHIFT, end);
2490                         if (err)
2491                                 retval = err;
2492                 }
2493         }
2494         return retval;
2495 }