[PATCH] mm: more commenting on lock ordering
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/security.h>
31 #include <linux/syscalls.h>
32 #include <linux/cpuset.h>
33 #include "filemap.h"
34 #include "internal.h"
35
36 /*
37  * FIXME: remove all knowledge of the buffer layer from the core VM
38  */
39 #include <linux/buffer_head.h> /* for generic_osync_inode */
40
41 #include <asm/mman.h>
42
43 static ssize_t
44 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
45         loff_t offset, unsigned long nr_segs);
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_lock               (vmtruncate)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_lock
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  ->i_mutex
82  *    ->i_alloc_sem             (various)
83  *
84  *  ->inode_lock
85  *    ->sb_lock                 (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_lock
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
103  *    ->inode_lock              (zap_pte_range->set_page_dirty)
104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
105  *
106  *  ->task->proc_lock
107  *    ->dcache_lock             (proc_pid_lookup)
108  */
109
110 /*
111  * Remove a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
114  */
115 void __remove_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123 }
124
125 void remove_from_page_cache(struct page *page)
126 {
127         struct address_space *mapping = page->mapping;
128
129         BUG_ON(!PageLocked(page));
130
131         write_lock_irq(&mapping->tree_lock);
132         __remove_from_page_cache(page);
133         write_unlock_irq(&mapping->tree_lock);
134 }
135
136 static int sync_page(void *word)
137 {
138         struct address_space *mapping;
139         struct page *page;
140
141         page = container_of((unsigned long *)word, struct page, flags);
142
143         /*
144          * page_mapping() is being called without PG_locked held.
145          * Some knowledge of the state and use of the page is used to
146          * reduce the requirements down to a memory barrier.
147          * The danger here is of a stale page_mapping() return value
148          * indicating a struct address_space different from the one it's
149          * associated with when it is associated with one.
150          * After smp_mb(), it's either the correct page_mapping() for
151          * the page, or an old page_mapping() and the page's own
152          * page_mapping() has gone NULL.
153          * The ->sync_page() address_space operation must tolerate
154          * page_mapping() going NULL. By an amazing coincidence,
155          * this comes about because none of the users of the page
156          * in the ->sync_page() methods make essential use of the
157          * page_mapping(), merely passing the page down to the backing
158          * device's unplug functions when it's non-NULL, which in turn
159          * ignore it for all cases but swap, where only page_private(page) is
160          * of interest. When page_mapping() does go NULL, the entire
161          * call stack gracefully ignores the page and returns.
162          * -- wli
163          */
164         smp_mb();
165         mapping = page_mapping(page);
166         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
167                 mapping->a_ops->sync_page(page);
168         io_schedule();
169         return 0;
170 }
171
172 /**
173  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
174  * @mapping:    address space structure to write
175  * @start:      offset in bytes where the range starts
176  * @end:        offset in bytes where the range ends (inclusive)
177  * @sync_mode:  enable synchronous operation
178  *
179  * Start writeback against all of a mapping's dirty pages that lie
180  * within the byte offsets <start, end> inclusive.
181  *
182  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
183  * opposed to a regular memory cleansing writeback.  The difference between
184  * these two operations is that if a dirty page/buffer is encountered, it must
185  * be waited upon, and not just skipped over.
186  */
187 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
188                                 loff_t end, int sync_mode)
189 {
190         int ret;
191         struct writeback_control wbc = {
192                 .sync_mode = sync_mode,
193                 .nr_to_write = mapping->nrpages * 2,
194                 .range_start = start,
195                 .range_end = end,
196         };
197
198         if (!mapping_cap_writeback_dirty(mapping))
199                 return 0;
200
201         ret = do_writepages(mapping, &wbc);
202         return ret;
203 }
204
205 static inline int __filemap_fdatawrite(struct address_space *mapping,
206         int sync_mode)
207 {
208         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
209 }
210
211 int filemap_fdatawrite(struct address_space *mapping)
212 {
213         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
214 }
215 EXPORT_SYMBOL(filemap_fdatawrite);
216
217 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
218                                 loff_t end)
219 {
220         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
221 }
222
223 /**
224  * filemap_flush - mostly a non-blocking flush
225  * @mapping:    target address_space
226  *
227  * This is a mostly non-blocking flush.  Not suitable for data-integrity
228  * purposes - I/O may not be started against all dirty pages.
229  */
230 int filemap_flush(struct address_space *mapping)
231 {
232         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
233 }
234 EXPORT_SYMBOL(filemap_flush);
235
236 /**
237  * wait_on_page_writeback_range - wait for writeback to complete
238  * @mapping:    target address_space
239  * @start:      beginning page index
240  * @end:        ending page index
241  *
242  * Wait for writeback to complete against pages indexed by start->end
243  * inclusive
244  */
245 int wait_on_page_writeback_range(struct address_space *mapping,
246                                 pgoff_t start, pgoff_t end)
247 {
248         struct pagevec pvec;
249         int nr_pages;
250         int ret = 0;
251         pgoff_t index;
252
253         if (end < start)
254                 return 0;
255
256         pagevec_init(&pvec, 0);
257         index = start;
258         while ((index <= end) &&
259                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
260                         PAGECACHE_TAG_WRITEBACK,
261                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
262                 unsigned i;
263
264                 for (i = 0; i < nr_pages; i++) {
265                         struct page *page = pvec.pages[i];
266
267                         /* until radix tree lookup accepts end_index */
268                         if (page->index > end)
269                                 continue;
270
271                         wait_on_page_writeback(page);
272                         if (PageError(page))
273                                 ret = -EIO;
274                 }
275                 pagevec_release(&pvec);
276                 cond_resched();
277         }
278
279         /* Check for outstanding write errors */
280         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
281                 ret = -ENOSPC;
282         if (test_and_clear_bit(AS_EIO, &mapping->flags))
283                 ret = -EIO;
284
285         return ret;
286 }
287
288 /**
289  * sync_page_range - write and wait on all pages in the passed range
290  * @inode:      target inode
291  * @mapping:    target address_space
292  * @pos:        beginning offset in pages to write
293  * @count:      number of bytes to write
294  *
295  * Write and wait upon all the pages in the passed range.  This is a "data
296  * integrity" operation.  It waits upon in-flight writeout before starting and
297  * waiting upon new writeout.  If there was an IO error, return it.
298  *
299  * We need to re-take i_mutex during the generic_osync_inode list walk because
300  * it is otherwise livelockable.
301  */
302 int sync_page_range(struct inode *inode, struct address_space *mapping,
303                         loff_t pos, loff_t count)
304 {
305         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
306         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
307         int ret;
308
309         if (!mapping_cap_writeback_dirty(mapping) || !count)
310                 return 0;
311         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
312         if (ret == 0) {
313                 mutex_lock(&inode->i_mutex);
314                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
315                 mutex_unlock(&inode->i_mutex);
316         }
317         if (ret == 0)
318                 ret = wait_on_page_writeback_range(mapping, start, end);
319         return ret;
320 }
321 EXPORT_SYMBOL(sync_page_range);
322
323 /**
324  * sync_page_range_nolock
325  * @inode:      target inode
326  * @mapping:    target address_space
327  * @pos:        beginning offset in pages to write
328  * @count:      number of bytes to write
329  *
330  * Note: Holding i_mutex across sync_page_range_nolock is not a good idea
331  * as it forces O_SYNC writers to different parts of the same file
332  * to be serialised right until io completion.
333  */
334 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
335                            loff_t pos, loff_t count)
336 {
337         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
338         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
339         int ret;
340
341         if (!mapping_cap_writeback_dirty(mapping) || !count)
342                 return 0;
343         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
344         if (ret == 0)
345                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
346         if (ret == 0)
347                 ret = wait_on_page_writeback_range(mapping, start, end);
348         return ret;
349 }
350 EXPORT_SYMBOL(sync_page_range_nolock);
351
352 /**
353  * filemap_fdatawait - wait for all under-writeback pages to complete
354  * @mapping: address space structure to wait for
355  *
356  * Walk the list of under-writeback pages of the given address space
357  * and wait for all of them.
358  */
359 int filemap_fdatawait(struct address_space *mapping)
360 {
361         loff_t i_size = i_size_read(mapping->host);
362
363         if (i_size == 0)
364                 return 0;
365
366         return wait_on_page_writeback_range(mapping, 0,
367                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
368 }
369 EXPORT_SYMBOL(filemap_fdatawait);
370
371 int filemap_write_and_wait(struct address_space *mapping)
372 {
373         int err = 0;
374
375         if (mapping->nrpages) {
376                 err = filemap_fdatawrite(mapping);
377                 /*
378                  * Even if the above returned error, the pages may be
379                  * written partially (e.g. -ENOSPC), so we wait for it.
380                  * But the -EIO is special case, it may indicate the worst
381                  * thing (e.g. bug) happened, so we avoid waiting for it.
382                  */
383                 if (err != -EIO) {
384                         int err2 = filemap_fdatawait(mapping);
385                         if (!err)
386                                 err = err2;
387                 }
388         }
389         return err;
390 }
391 EXPORT_SYMBOL(filemap_write_and_wait);
392
393 /**
394  * filemap_write_and_wait_range - write out & wait on a file range
395  * @mapping:    the address_space for the pages
396  * @lstart:     offset in bytes where the range starts
397  * @lend:       offset in bytes where the range ends (inclusive)
398  *
399  * Write out and wait upon file offsets lstart->lend, inclusive.
400  *
401  * Note that `lend' is inclusive (describes the last byte to be written) so
402  * that this function can be used to write to the very end-of-file (end = -1).
403  */
404 int filemap_write_and_wait_range(struct address_space *mapping,
405                                  loff_t lstart, loff_t lend)
406 {
407         int err = 0;
408
409         if (mapping->nrpages) {
410                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
411                                                  WB_SYNC_ALL);
412                 /* See comment of filemap_write_and_wait() */
413                 if (err != -EIO) {
414                         int err2 = wait_on_page_writeback_range(mapping,
415                                                 lstart >> PAGE_CACHE_SHIFT,
416                                                 lend >> PAGE_CACHE_SHIFT);
417                         if (!err)
418                                 err = err2;
419                 }
420         }
421         return err;
422 }
423
424 /**
425  * add_to_page_cache - add newly allocated pagecache pages
426  * @page:       page to add
427  * @mapping:    the page's address_space
428  * @offset:     page index
429  * @gfp_mask:   page allocation mode
430  *
431  * This function is used to add newly allocated pagecache pages;
432  * the page is new, so we can just run SetPageLocked() against it.
433  * The other page state flags were set by rmqueue().
434  *
435  * This function does not add the page to the LRU.  The caller must do that.
436  */
437 int add_to_page_cache(struct page *page, struct address_space *mapping,
438                 pgoff_t offset, gfp_t gfp_mask)
439 {
440         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
441
442         if (error == 0) {
443                 write_lock_irq(&mapping->tree_lock);
444                 error = radix_tree_insert(&mapping->page_tree, offset, page);
445                 if (!error) {
446                         page_cache_get(page);
447                         SetPageLocked(page);
448                         page->mapping = mapping;
449                         page->index = offset;
450                         mapping->nrpages++;
451                         __inc_zone_page_state(page, NR_FILE_PAGES);
452                 }
453                 write_unlock_irq(&mapping->tree_lock);
454                 radix_tree_preload_end();
455         }
456         return error;
457 }
458 EXPORT_SYMBOL(add_to_page_cache);
459
460 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
461                                 pgoff_t offset, gfp_t gfp_mask)
462 {
463         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
464         if (ret == 0)
465                 lru_cache_add(page);
466         return ret;
467 }
468
469 #ifdef CONFIG_NUMA
470 struct page *page_cache_alloc(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  * @pos:        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 pos)
1154 {
1155         struct file *filp = iocb->ki_filp;
1156         ssize_t retval;
1157         unsigned long seg;
1158         size_t count;
1159         loff_t *ppos = &iocb->ki_pos;
1160
1161         count = 0;
1162         for (seg = 0; seg < nr_segs; seg++) {
1163                 const struct iovec *iv = &iov[seg];
1164
1165                 /*
1166                  * If any segment has a negative length, or the cumulative
1167                  * length ever wraps negative then return -EINVAL.
1168                  */
1169                 count += iv->iov_len;
1170                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1171                         return -EINVAL;
1172                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1173                         continue;
1174                 if (seg == 0)
1175                         return -EFAULT;
1176                 nr_segs = seg;
1177                 count -= iv->iov_len;   /* This segment is no good */
1178                 break;
1179         }
1180
1181         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1182         if (filp->f_flags & O_DIRECT) {
1183                 loff_t size;
1184                 struct address_space *mapping;
1185                 struct inode *inode;
1186
1187                 mapping = filp->f_mapping;
1188                 inode = mapping->host;
1189                 retval = 0;
1190                 if (!count)
1191                         goto out; /* skip atime */
1192                 size = i_size_read(inode);
1193                 if (pos < size) {
1194                         retval = generic_file_direct_IO(READ, iocb,
1195                                                 iov, pos, nr_segs);
1196                         if (retval > 0 && !is_sync_kiocb(iocb))
1197                                 retval = -EIOCBQUEUED;
1198                         if (retval > 0)
1199                                 *ppos = pos + retval;
1200                 }
1201                 if (likely(retval != 0)) {
1202                         file_accessed(filp);
1203                         goto out;
1204                 }
1205         }
1206
1207         retval = 0;
1208         if (count) {
1209                 for (seg = 0; seg < nr_segs; seg++) {
1210                         read_descriptor_t desc;
1211
1212                         desc.written = 0;
1213                         desc.arg.buf = iov[seg].iov_base;
1214                         desc.count = iov[seg].iov_len;
1215                         if (desc.count == 0)
1216                                 continue;
1217                         desc.error = 0;
1218                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1219                         retval += desc.written;
1220                         if (desc.error) {
1221                                 retval = retval ?: desc.error;
1222                                 break;
1223                         }
1224                 }
1225         }
1226 out:
1227         return retval;
1228 }
1229 EXPORT_SYMBOL(generic_file_aio_read);
1230
1231 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1232 {
1233         ssize_t written;
1234         unsigned long count = desc->count;
1235         struct file *file = desc->arg.data;
1236
1237         if (size > count)
1238                 size = count;
1239
1240         written = file->f_op->sendpage(file, page, offset,
1241                                        size, &file->f_pos, size<count);
1242         if (written < 0) {
1243                 desc->error = written;
1244                 written = 0;
1245         }
1246         desc->count = count - written;
1247         desc->written += written;
1248         return written;
1249 }
1250
1251 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1252                          size_t count, read_actor_t actor, void *target)
1253 {
1254         read_descriptor_t desc;
1255
1256         if (!count)
1257                 return 0;
1258
1259         desc.written = 0;
1260         desc.count = count;
1261         desc.arg.data = target;
1262         desc.error = 0;
1263
1264         do_generic_file_read(in_file, ppos, &desc, actor);
1265         if (desc.written)
1266                 return desc.written;
1267         return desc.error;
1268 }
1269 EXPORT_SYMBOL(generic_file_sendfile);
1270
1271 static ssize_t
1272 do_readahead(struct address_space *mapping, struct file *filp,
1273              unsigned long index, unsigned long nr)
1274 {
1275         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1276                 return -EINVAL;
1277
1278         force_page_cache_readahead(mapping, filp, index,
1279                                         max_sane_readahead(nr));
1280         return 0;
1281 }
1282
1283 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1284 {
1285         ssize_t ret;
1286         struct file *file;
1287
1288         ret = -EBADF;
1289         file = fget(fd);
1290         if (file) {
1291                 if (file->f_mode & FMODE_READ) {
1292                         struct address_space *mapping = file->f_mapping;
1293                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1294                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1295                         unsigned long len = end - start + 1;
1296                         ret = do_readahead(mapping, file, start, len);
1297                 }
1298                 fput(file);
1299         }
1300         return ret;
1301 }
1302
1303 #ifdef CONFIG_MMU
1304 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1305 /**
1306  * page_cache_read - adds requested page to the page cache if not already there
1307  * @file:       file to read
1308  * @offset:     page index
1309  *
1310  * This adds the requested page to the page cache if it isn't already there,
1311  * and schedules an I/O to read in its contents from disk.
1312  */
1313 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1314 {
1315         struct address_space *mapping = file->f_mapping;
1316         struct page *page; 
1317         int ret;
1318
1319         do {
1320                 page = page_cache_alloc_cold(mapping);
1321                 if (!page)
1322                         return -ENOMEM;
1323
1324                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1325                 if (ret == 0)
1326                         ret = mapping->a_ops->readpage(file, page);
1327                 else if (ret == -EEXIST)
1328                         ret = 0; /* losing race to add is OK */
1329
1330                 page_cache_release(page);
1331
1332         } while (ret == AOP_TRUNCATED_PAGE);
1333                 
1334         return ret;
1335 }
1336
1337 #define MMAP_LOTSAMISS  (100)
1338
1339 /**
1340  * filemap_nopage - read in file data for page fault handling
1341  * @area:       the applicable vm_area
1342  * @address:    target address to read in
1343  * @type:       returned with VM_FAULT_{MINOR,MAJOR} if not %NULL
1344  *
1345  * filemap_nopage() is invoked via the vma operations vector for a
1346  * mapped memory region to read in file data during a page fault.
1347  *
1348  * The goto's are kind of ugly, but this streamlines the normal case of having
1349  * it in the page cache, and handles the special cases reasonably without
1350  * having a lot of duplicated code.
1351  */
1352 struct page *filemap_nopage(struct vm_area_struct *area,
1353                                 unsigned long address, int *type)
1354 {
1355         int error;
1356         struct file *file = area->vm_file;
1357         struct address_space *mapping = file->f_mapping;
1358         struct file_ra_state *ra = &file->f_ra;
1359         struct inode *inode = mapping->host;
1360         struct page *page;
1361         unsigned long size, pgoff;
1362         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1363
1364         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1365
1366 retry_all:
1367         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1368         if (pgoff >= size)
1369                 goto outside_data_content;
1370
1371         /* If we don't want any read-ahead, don't bother */
1372         if (VM_RandomReadHint(area))
1373                 goto no_cached_page;
1374
1375         /*
1376          * The readahead code wants to be told about each and every page
1377          * so it can build and shrink its windows appropriately
1378          *
1379          * For sequential accesses, we use the generic readahead logic.
1380          */
1381         if (VM_SequentialReadHint(area))
1382                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1383
1384         /*
1385          * Do we have something in the page cache already?
1386          */
1387 retry_find:
1388         page = find_get_page(mapping, pgoff);
1389         if (!page) {
1390                 unsigned long ra_pages;
1391
1392                 if (VM_SequentialReadHint(area)) {
1393                         handle_ra_miss(mapping, ra, pgoff);
1394                         goto no_cached_page;
1395                 }
1396                 ra->mmap_miss++;
1397
1398                 /*
1399                  * Do we miss much more than hit in this file? If so,
1400                  * stop bothering with read-ahead. It will only hurt.
1401                  */
1402                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1403                         goto no_cached_page;
1404
1405                 /*
1406                  * To keep the pgmajfault counter straight, we need to
1407                  * check did_readaround, as this is an inner loop.
1408                  */
1409                 if (!did_readaround) {
1410                         majmin = VM_FAULT_MAJOR;
1411                         count_vm_event(PGMAJFAULT);
1412                 }
1413                 did_readaround = 1;
1414                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1415                 if (ra_pages) {
1416                         pgoff_t start = 0;
1417
1418                         if (pgoff > ra_pages / 2)
1419                                 start = pgoff - ra_pages / 2;
1420                         do_page_cache_readahead(mapping, file, start, ra_pages);
1421                 }
1422                 page = find_get_page(mapping, pgoff);
1423                 if (!page)
1424                         goto no_cached_page;
1425         }
1426
1427         if (!did_readaround)
1428                 ra->mmap_hit++;
1429
1430         /*
1431          * Ok, found a page in the page cache, now we need to check
1432          * that it's up-to-date.
1433          */
1434         if (!PageUptodate(page))
1435                 goto page_not_uptodate;
1436
1437 success:
1438         /*
1439          * Found the page and have a reference on it.
1440          */
1441         mark_page_accessed(page);
1442         if (type)
1443                 *type = majmin;
1444         return page;
1445
1446 outside_data_content:
1447         /*
1448          * An external ptracer can access pages that normally aren't
1449          * accessible..
1450          */
1451         if (area->vm_mm == current->mm)
1452                 return NOPAGE_SIGBUS;
1453         /* Fall through to the non-read-ahead case */
1454 no_cached_page:
1455         /*
1456          * We're only likely to ever get here if MADV_RANDOM is in
1457          * effect.
1458          */
1459         error = page_cache_read(file, pgoff);
1460         grab_swap_token();
1461
1462         /*
1463          * The page we want has now been added to the page cache.
1464          * In the unlikely event that someone removed it in the
1465          * meantime, we'll just come back here and read it again.
1466          */
1467         if (error >= 0)
1468                 goto retry_find;
1469
1470         /*
1471          * An error return from page_cache_read can result if the
1472          * system is low on memory, or a problem occurs while trying
1473          * to schedule I/O.
1474          */
1475         if (error == -ENOMEM)
1476                 return NOPAGE_OOM;
1477         return NOPAGE_SIGBUS;
1478
1479 page_not_uptodate:
1480         if (!did_readaround) {
1481                 majmin = VM_FAULT_MAJOR;
1482                 count_vm_event(PGMAJFAULT);
1483         }
1484         lock_page(page);
1485
1486         /* Did it get unhashed while we waited for it? */
1487         if (!page->mapping) {
1488                 unlock_page(page);
1489                 page_cache_release(page);
1490                 goto retry_all;
1491         }
1492
1493         /* Did somebody else get it up-to-date? */
1494         if (PageUptodate(page)) {
1495                 unlock_page(page);
1496                 goto success;
1497         }
1498
1499         error = mapping->a_ops->readpage(file, page);
1500         if (!error) {
1501                 wait_on_page_locked(page);
1502                 if (PageUptodate(page))
1503                         goto success;
1504         } else if (error == AOP_TRUNCATED_PAGE) {
1505                 page_cache_release(page);
1506                 goto retry_find;
1507         }
1508
1509         /*
1510          * Umm, take care of errors if the page isn't up-to-date.
1511          * Try to re-read it _once_. We do this synchronously,
1512          * because there really aren't any performance issues here
1513          * and we need to check for errors.
1514          */
1515         lock_page(page);
1516
1517         /* Somebody truncated the page on us? */
1518         if (!page->mapping) {
1519                 unlock_page(page);
1520                 page_cache_release(page);
1521                 goto retry_all;
1522         }
1523
1524         /* Somebody else successfully read it in? */
1525         if (PageUptodate(page)) {
1526                 unlock_page(page);
1527                 goto success;
1528         }
1529         ClearPageError(page);
1530         error = mapping->a_ops->readpage(file, page);
1531         if (!error) {
1532                 wait_on_page_locked(page);
1533                 if (PageUptodate(page))
1534                         goto success;
1535         } else if (error == AOP_TRUNCATED_PAGE) {
1536                 page_cache_release(page);
1537                 goto retry_find;
1538         }
1539
1540         /*
1541          * Things didn't work out. Return zero to tell the
1542          * mm layer so, possibly freeing the page cache page first.
1543          */
1544         shrink_readahead_size_eio(file, ra);
1545         page_cache_release(page);
1546         return NOPAGE_SIGBUS;
1547 }
1548 EXPORT_SYMBOL(filemap_nopage);
1549
1550 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1551                                         int nonblock)
1552 {
1553         struct address_space *mapping = file->f_mapping;
1554         struct page *page;
1555         int error;
1556
1557         /*
1558          * Do we have something in the page cache already?
1559          */
1560 retry_find:
1561         page = find_get_page(mapping, pgoff);
1562         if (!page) {
1563                 if (nonblock)
1564                         return NULL;
1565                 goto no_cached_page;
1566         }
1567
1568         /*
1569          * Ok, found a page in the page cache, now we need to check
1570          * that it's up-to-date.
1571          */
1572         if (!PageUptodate(page)) {
1573                 if (nonblock) {
1574                         page_cache_release(page);
1575                         return NULL;
1576                 }
1577                 goto page_not_uptodate;
1578         }
1579
1580 success:
1581         /*
1582          * Found the page and have a reference on it.
1583          */
1584         mark_page_accessed(page);
1585         return page;
1586
1587 no_cached_page:
1588         error = page_cache_read(file, pgoff);
1589
1590         /*
1591          * The page we want has now been added to the page cache.
1592          * In the unlikely event that someone removed it in the
1593          * meantime, we'll just come back here and read it again.
1594          */
1595         if (error >= 0)
1596                 goto retry_find;
1597
1598         /*
1599          * An error return from page_cache_read can result if the
1600          * system is low on memory, or a problem occurs while trying
1601          * to schedule I/O.
1602          */
1603         return NULL;
1604
1605 page_not_uptodate:
1606         lock_page(page);
1607
1608         /* Did it get truncated while we waited for it? */
1609         if (!page->mapping) {
1610                 unlock_page(page);
1611                 goto err;
1612         }
1613
1614         /* Did somebody else get it up-to-date? */
1615         if (PageUptodate(page)) {
1616                 unlock_page(page);
1617                 goto success;
1618         }
1619
1620         error = mapping->a_ops->readpage(file, page);
1621         if (!error) {
1622                 wait_on_page_locked(page);
1623                 if (PageUptodate(page))
1624                         goto success;
1625         } else if (error == AOP_TRUNCATED_PAGE) {
1626                 page_cache_release(page);
1627                 goto retry_find;
1628         }
1629
1630         /*
1631          * Umm, take care of errors if the page isn't up-to-date.
1632          * Try to re-read it _once_. We do this synchronously,
1633          * because there really aren't any performance issues here
1634          * and we need to check for errors.
1635          */
1636         lock_page(page);
1637
1638         /* Somebody truncated the page on us? */
1639         if (!page->mapping) {
1640                 unlock_page(page);
1641                 goto err;
1642         }
1643         /* Somebody else successfully read it in? */
1644         if (PageUptodate(page)) {
1645                 unlock_page(page);
1646                 goto success;
1647         }
1648
1649         ClearPageError(page);
1650         error = mapping->a_ops->readpage(file, page);
1651         if (!error) {
1652                 wait_on_page_locked(page);
1653                 if (PageUptodate(page))
1654                         goto success;
1655         } else if (error == AOP_TRUNCATED_PAGE) {
1656                 page_cache_release(page);
1657                 goto retry_find;
1658         }
1659
1660         /*
1661          * Things didn't work out. Return zero to tell the
1662          * mm layer so, possibly freeing the page cache page first.
1663          */
1664 err:
1665         page_cache_release(page);
1666
1667         return NULL;
1668 }
1669
1670 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1671                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1672                 int nonblock)
1673 {
1674         struct file *file = vma->vm_file;
1675         struct address_space *mapping = file->f_mapping;
1676         struct inode *inode = mapping->host;
1677         unsigned long size;
1678         struct mm_struct *mm = vma->vm_mm;
1679         struct page *page;
1680         int err;
1681
1682         if (!nonblock)
1683                 force_page_cache_readahead(mapping, vma->vm_file,
1684                                         pgoff, len >> PAGE_CACHE_SHIFT);
1685
1686 repeat:
1687         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1688         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1689                 return -EINVAL;
1690
1691         page = filemap_getpage(file, pgoff, nonblock);
1692
1693         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1694          * done in shmem_populate calling shmem_getpage */
1695         if (!page && !nonblock)
1696                 return -ENOMEM;
1697
1698         if (page) {
1699                 err = install_page(mm, vma, addr, page, prot);
1700                 if (err) {
1701                         page_cache_release(page);
1702                         return err;
1703                 }
1704         } else if (vma->vm_flags & VM_NONLINEAR) {
1705                 /* No page was found just because we can't read it in now (being
1706                  * here implies nonblock != 0), but the page may exist, so set
1707                  * the PTE to fault it in later. */
1708                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1709                 if (err)
1710                         return err;
1711         }
1712
1713         len -= PAGE_SIZE;
1714         addr += PAGE_SIZE;
1715         pgoff++;
1716         if (len)
1717                 goto repeat;
1718
1719         return 0;
1720 }
1721 EXPORT_SYMBOL(filemap_populate);
1722
1723 struct vm_operations_struct generic_file_vm_ops = {
1724         .nopage         = filemap_nopage,
1725         .populate       = filemap_populate,
1726 };
1727
1728 /* This is used for a general mmap of a disk file */
1729
1730 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1731 {
1732         struct address_space *mapping = file->f_mapping;
1733
1734         if (!mapping->a_ops->readpage)
1735                 return -ENOEXEC;
1736         file_accessed(file);
1737         vma->vm_ops = &generic_file_vm_ops;
1738         return 0;
1739 }
1740
1741 /*
1742  * This is for filesystems which do not implement ->writepage.
1743  */
1744 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1745 {
1746         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1747                 return -EINVAL;
1748         return generic_file_mmap(file, vma);
1749 }
1750 #else
1751 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1752 {
1753         return -ENOSYS;
1754 }
1755 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1756 {
1757         return -ENOSYS;
1758 }
1759 #endif /* CONFIG_MMU */
1760
1761 EXPORT_SYMBOL(generic_file_mmap);
1762 EXPORT_SYMBOL(generic_file_readonly_mmap);
1763
1764 static inline struct page *__read_cache_page(struct address_space *mapping,
1765                                 unsigned long index,
1766                                 int (*filler)(void *,struct page*),
1767                                 void *data)
1768 {
1769         struct page *page, *cached_page = NULL;
1770         int err;
1771 repeat:
1772         page = find_get_page(mapping, index);
1773         if (!page) {
1774                 if (!cached_page) {
1775                         cached_page = page_cache_alloc_cold(mapping);
1776                         if (!cached_page)
1777                                 return ERR_PTR(-ENOMEM);
1778                 }
1779                 err = add_to_page_cache_lru(cached_page, mapping,
1780                                         index, GFP_KERNEL);
1781                 if (err == -EEXIST)
1782                         goto repeat;
1783                 if (err < 0) {
1784                         /* Presumably ENOMEM for radix tree node */
1785                         page_cache_release(cached_page);
1786                         return ERR_PTR(err);
1787                 }
1788                 page = cached_page;
1789                 cached_page = NULL;
1790                 err = filler(data, page);
1791                 if (err < 0) {
1792                         page_cache_release(page);
1793                         page = ERR_PTR(err);
1794                 }
1795         }
1796         if (cached_page)
1797                 page_cache_release(cached_page);
1798         return page;
1799 }
1800
1801 /**
1802  * read_cache_page - read into page cache, fill it if needed
1803  * @mapping:    the page's address_space
1804  * @index:      the page index
1805  * @filler:     function to perform the read
1806  * @data:       destination for read data
1807  *
1808  * Read into the page cache. If a page already exists,
1809  * and PageUptodate() is not set, try to fill the page.
1810  */
1811 struct page *read_cache_page(struct address_space *mapping,
1812                                 unsigned long index,
1813                                 int (*filler)(void *,struct page*),
1814                                 void *data)
1815 {
1816         struct page *page;
1817         int err;
1818
1819 retry:
1820         page = __read_cache_page(mapping, index, filler, data);
1821         if (IS_ERR(page))
1822                 goto out;
1823         mark_page_accessed(page);
1824         if (PageUptodate(page))
1825                 goto out;
1826
1827         lock_page(page);
1828         if (!page->mapping) {
1829                 unlock_page(page);
1830                 page_cache_release(page);
1831                 goto retry;
1832         }
1833         if (PageUptodate(page)) {
1834                 unlock_page(page);
1835                 goto out;
1836         }
1837         err = filler(data, page);
1838         if (err < 0) {
1839                 page_cache_release(page);
1840                 page = ERR_PTR(err);
1841         }
1842  out:
1843         return page;
1844 }
1845 EXPORT_SYMBOL(read_cache_page);
1846
1847 /*
1848  * If the page was newly created, increment its refcount and add it to the
1849  * caller's lru-buffering pagevec.  This function is specifically for
1850  * generic_file_write().
1851  */
1852 static inline struct page *
1853 __grab_cache_page(struct address_space *mapping, unsigned long index,
1854                         struct page **cached_page, struct pagevec *lru_pvec)
1855 {
1856         int err;
1857         struct page *page;
1858 repeat:
1859         page = find_lock_page(mapping, index);
1860         if (!page) {
1861                 if (!*cached_page) {
1862                         *cached_page = page_cache_alloc(mapping);
1863                         if (!*cached_page)
1864                                 return NULL;
1865                 }
1866                 err = add_to_page_cache(*cached_page, mapping,
1867                                         index, GFP_KERNEL);
1868                 if (err == -EEXIST)
1869                         goto repeat;
1870                 if (err == 0) {
1871                         page = *cached_page;
1872                         page_cache_get(page);
1873                         if (!pagevec_add(lru_pvec, page))
1874                                 __pagevec_lru_add(lru_pvec);
1875                         *cached_page = NULL;
1876                 }
1877         }
1878         return page;
1879 }
1880
1881 /*
1882  * The logic we want is
1883  *
1884  *      if suid or (sgid and xgrp)
1885  *              remove privs
1886  */
1887 int remove_suid(struct dentry *dentry)
1888 {
1889         mode_t mode = dentry->d_inode->i_mode;
1890         int kill = 0;
1891         int result = 0;
1892
1893         /* suid always must be killed */
1894         if (unlikely(mode & S_ISUID))
1895                 kill = ATTR_KILL_SUID;
1896
1897         /*
1898          * sgid without any exec bits is just a mandatory locking mark; leave
1899          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1900          */
1901         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1902                 kill |= ATTR_KILL_SGID;
1903
1904         if (unlikely(kill && !capable(CAP_FSETID))) {
1905                 struct iattr newattrs;
1906
1907                 newattrs.ia_valid = ATTR_FORCE | kill;
1908                 result = notify_change(dentry, &newattrs);
1909         }
1910         return result;
1911 }
1912 EXPORT_SYMBOL(remove_suid);
1913
1914 size_t
1915 __filemap_copy_from_user_iovec_inatomic(char *vaddr,
1916                         const struct iovec *iov, size_t base, size_t bytes)
1917 {
1918         size_t copied = 0, left = 0;
1919
1920         while (bytes) {
1921                 char __user *buf = iov->iov_base + base;
1922                 int copy = min(bytes, iov->iov_len - base);
1923
1924                 base = 0;
1925                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1926                 copied += copy;
1927                 bytes -= copy;
1928                 vaddr += copy;
1929                 iov++;
1930
1931                 if (unlikely(left))
1932                         break;
1933         }
1934         return copied - left;
1935 }
1936
1937 /*
1938  * Performs necessary checks before doing a write
1939  *
1940  * Can adjust writing position or amount of bytes to write.
1941  * Returns appropriate error code that caller should return or
1942  * zero in case that write should be allowed.
1943  */
1944 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1945 {
1946         struct inode *inode = file->f_mapping->host;
1947         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1948
1949         if (unlikely(*pos < 0))
1950                 return -EINVAL;
1951
1952         if (!isblk) {
1953                 /* FIXME: this is for backwards compatibility with 2.4 */
1954                 if (file->f_flags & O_APPEND)
1955                         *pos = i_size_read(inode);
1956
1957                 if (limit != RLIM_INFINITY) {
1958                         if (*pos >= limit) {
1959                                 send_sig(SIGXFSZ, current, 0);
1960                                 return -EFBIG;
1961                         }
1962                         if (*count > limit - (typeof(limit))*pos) {
1963                                 *count = limit - (typeof(limit))*pos;
1964                         }
1965                 }
1966         }
1967
1968         /*
1969          * LFS rule
1970          */
1971         if (unlikely(*pos + *count > MAX_NON_LFS &&
1972                                 !(file->f_flags & O_LARGEFILE))) {
1973                 if (*pos >= MAX_NON_LFS) {
1974                         send_sig(SIGXFSZ, current, 0);
1975                         return -EFBIG;
1976                 }
1977                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1978                         *count = MAX_NON_LFS - (unsigned long)*pos;
1979                 }
1980         }
1981
1982         /*
1983          * Are we about to exceed the fs block limit ?
1984          *
1985          * If we have written data it becomes a short write.  If we have
1986          * exceeded without writing data we send a signal and return EFBIG.
1987          * Linus frestrict idea will clean these up nicely..
1988          */
1989         if (likely(!isblk)) {
1990                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1991                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1992                                 send_sig(SIGXFSZ, current, 0);
1993                                 return -EFBIG;
1994                         }
1995                         /* zero-length writes at ->s_maxbytes are OK */
1996                 }
1997
1998                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1999                         *count = inode->i_sb->s_maxbytes - *pos;
2000         } else {
2001 #ifdef CONFIG_BLOCK
2002                 loff_t isize;
2003                 if (bdev_read_only(I_BDEV(inode)))
2004                         return -EPERM;
2005                 isize = i_size_read(inode);
2006                 if (*pos >= isize) {
2007                         if (*count || *pos > isize)
2008                                 return -ENOSPC;
2009                 }
2010
2011                 if (*pos + *count > isize)
2012                         *count = isize - *pos;
2013 #else
2014                 return -EPERM;
2015 #endif
2016         }
2017         return 0;
2018 }
2019 EXPORT_SYMBOL(generic_write_checks);
2020
2021 ssize_t
2022 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2023                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2024                 size_t count, size_t ocount)
2025 {
2026         struct file     *file = iocb->ki_filp;
2027         struct address_space *mapping = file->f_mapping;
2028         struct inode    *inode = mapping->host;
2029         ssize_t         written;
2030
2031         if (count != ocount)
2032                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2033
2034         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2035         if (written > 0) {
2036                 loff_t end = pos + written;
2037                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2038                         i_size_write(inode,  end);
2039                         mark_inode_dirty(inode);
2040                 }
2041                 *ppos = end;
2042         }
2043
2044         /*
2045          * Sync the fs metadata but not the minor inode changes and
2046          * of course not the data as we did direct DMA for the IO.
2047          * i_mutex is held, which protects generic_osync_inode() from
2048          * livelocking.
2049          */
2050         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2051                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2052                 if (err < 0)
2053                         written = err;
2054         }
2055         if (written == count && !is_sync_kiocb(iocb))
2056                 written = -EIOCBQUEUED;
2057         return written;
2058 }
2059 EXPORT_SYMBOL(generic_file_direct_write);
2060
2061 ssize_t
2062 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2063                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2064                 size_t count, ssize_t written)
2065 {
2066         struct file *file = iocb->ki_filp;
2067         struct address_space * mapping = file->f_mapping;
2068         const struct address_space_operations *a_ops = mapping->a_ops;
2069         struct inode    *inode = mapping->host;
2070         long            status = 0;
2071         struct page     *page;
2072         struct page     *cached_page = NULL;
2073         size_t          bytes;
2074         struct pagevec  lru_pvec;
2075         const struct iovec *cur_iov = iov; /* current iovec */
2076         size_t          iov_base = 0;      /* offset in the current iovec */
2077         char __user     *buf;
2078
2079         pagevec_init(&lru_pvec, 0);
2080
2081         /*
2082          * handle partial DIO write.  Adjust cur_iov if needed.
2083          */
2084         if (likely(nr_segs == 1))
2085                 buf = iov->iov_base + written;
2086         else {
2087                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
2088                 buf = cur_iov->iov_base + iov_base;
2089         }
2090
2091         do {
2092                 unsigned long index;
2093                 unsigned long offset;
2094                 size_t copied;
2095
2096                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
2097                 index = pos >> PAGE_CACHE_SHIFT;
2098                 bytes = PAGE_CACHE_SIZE - offset;
2099
2100                 /* Limit the size of the copy to the caller's write size */
2101                 bytes = min(bytes, count);
2102
2103                 /*
2104                  * Limit the size of the copy to that of the current segment,
2105                  * because fault_in_pages_readable() doesn't know how to walk
2106                  * segments.
2107                  */
2108                 bytes = min(bytes, cur_iov->iov_len - iov_base);
2109
2110                 /*
2111                  * Bring in the user page that we will copy from _first_.
2112                  * Otherwise there's a nasty deadlock on copying from the
2113                  * same page as we're writing to, without it being marked
2114                  * up-to-date.
2115                  */
2116                 fault_in_pages_readable(buf, bytes);
2117
2118                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
2119                 if (!page) {
2120                         status = -ENOMEM;
2121                         break;
2122                 }
2123
2124                 if (unlikely(bytes == 0)) {
2125                         status = 0;
2126                         copied = 0;
2127                         goto zero_length_segment;
2128                 }
2129
2130                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2131                 if (unlikely(status)) {
2132                         loff_t isize = i_size_read(inode);
2133
2134                         if (status != AOP_TRUNCATED_PAGE)
2135                                 unlock_page(page);
2136                         page_cache_release(page);
2137                         if (status == AOP_TRUNCATED_PAGE)
2138                                 continue;
2139                         /*
2140                          * prepare_write() may have instantiated a few blocks
2141                          * outside i_size.  Trim these off again.
2142                          */
2143                         if (pos + bytes > isize)
2144                                 vmtruncate(inode, isize);
2145                         break;
2146                 }
2147                 if (likely(nr_segs == 1))
2148                         copied = filemap_copy_from_user(page, offset,
2149                                                         buf, bytes);
2150                 else
2151                         copied = filemap_copy_from_user_iovec(page, offset,
2152                                                 cur_iov, iov_base, bytes);
2153                 flush_dcache_page(page);
2154                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2155                 if (status == AOP_TRUNCATED_PAGE) {
2156                         page_cache_release(page);
2157                         continue;
2158                 }
2159 zero_length_segment:
2160                 if (likely(copied >= 0)) {
2161                         if (!status)
2162                                 status = copied;
2163
2164                         if (status >= 0) {
2165                                 written += status;
2166                                 count -= status;
2167                                 pos += status;
2168                                 buf += status;
2169                                 if (unlikely(nr_segs > 1)) {
2170                                         filemap_set_next_iovec(&cur_iov,
2171                                                         &iov_base, status);
2172                                         if (count)
2173                                                 buf = cur_iov->iov_base +
2174                                                         iov_base;
2175                                 } else {
2176                                         iov_base += status;
2177                                 }
2178                         }
2179                 }
2180                 if (unlikely(copied != bytes))
2181                         if (status >= 0)
2182                                 status = -EFAULT;
2183                 unlock_page(page);
2184                 mark_page_accessed(page);
2185                 page_cache_release(page);
2186                 if (status < 0)
2187                         break;
2188                 balance_dirty_pages_ratelimited(mapping);
2189                 cond_resched();
2190         } while (count);
2191         *ppos = pos;
2192
2193         if (cached_page)
2194                 page_cache_release(cached_page);
2195
2196         /*
2197          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2198          */
2199         if (likely(status >= 0)) {
2200                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2201                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2202                                 status = generic_osync_inode(inode, mapping,
2203                                                 OSYNC_METADATA|OSYNC_DATA);
2204                 }
2205         }
2206         
2207         /*
2208          * If we get here for O_DIRECT writes then we must have fallen through
2209          * to buffered writes (block instantiation inside i_size).  So we sync
2210          * the file data here, to try to honour O_DIRECT expectations.
2211          */
2212         if (unlikely(file->f_flags & O_DIRECT) && written)
2213                 status = filemap_write_and_wait(mapping);
2214
2215         pagevec_lru_add(&lru_pvec);
2216         return written ? written : status;
2217 }
2218 EXPORT_SYMBOL(generic_file_buffered_write);
2219
2220 static ssize_t
2221 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2222                                 unsigned long nr_segs, loff_t *ppos)
2223 {
2224         struct file *file = iocb->ki_filp;
2225         struct address_space * mapping = file->f_mapping;
2226         size_t ocount;          /* original count */
2227         size_t count;           /* after file limit checks */
2228         struct inode    *inode = mapping->host;
2229         unsigned long   seg;
2230         loff_t          pos;
2231         ssize_t         written;
2232         ssize_t         err;
2233
2234         ocount = 0;
2235         for (seg = 0; seg < nr_segs; seg++) {
2236                 const struct iovec *iv = &iov[seg];
2237
2238                 /*
2239                  * If any segment has a negative length, or the cumulative
2240                  * length ever wraps negative then return -EINVAL.
2241                  */
2242                 ocount += iv->iov_len;
2243                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2244                         return -EINVAL;
2245                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2246                         continue;
2247                 if (seg == 0)
2248                         return -EFAULT;
2249                 nr_segs = seg;
2250                 ocount -= iv->iov_len;  /* This segment is no good */
2251                 break;
2252         }
2253
2254         count = ocount;
2255         pos = *ppos;
2256
2257         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2258
2259         /* We can write back this queue in page reclaim */
2260         current->backing_dev_info = mapping->backing_dev_info;
2261         written = 0;
2262
2263         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2264         if (err)
2265                 goto out;
2266
2267         if (count == 0)
2268                 goto out;
2269
2270         err = remove_suid(file->f_dentry);
2271         if (err)
2272                 goto out;
2273
2274         file_update_time(file);
2275
2276         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2277         if (unlikely(file->f_flags & O_DIRECT)) {
2278                 loff_t endbyte;
2279                 ssize_t written_buffered;
2280
2281                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2282                                                         ppos, count, ocount);
2283                 if (written < 0 || written == count)
2284                         goto out;
2285                 /*
2286                  * direct-io write to a hole: fall through to buffered I/O
2287                  * for completing the rest of the request.
2288                  */
2289                 pos += written;
2290                 count -= written;
2291                 written_buffered = generic_file_buffered_write(iocb, iov,
2292                                                 nr_segs, pos, ppos, count,
2293                                                 written);
2294                 /*
2295                  * If generic_file_buffered_write() retuned a synchronous error
2296                  * then we want to return the number of bytes which were
2297                  * direct-written, or the error code if that was zero.  Note
2298                  * that this differs from normal direct-io semantics, which
2299                  * will return -EFOO even if some bytes were written.
2300                  */
2301                 if (written_buffered < 0) {
2302                         err = written_buffered;
2303                         goto out;
2304                 }
2305
2306                 /*
2307                  * We need to ensure that the page cache pages are written to
2308                  * disk and invalidated to preserve the expected O_DIRECT
2309                  * semantics.
2310                  */
2311                 endbyte = pos + written_buffered - written - 1;
2312                 err = do_sync_file_range(file, pos, endbyte,
2313                                          SYNC_FILE_RANGE_WAIT_BEFORE|
2314                                          SYNC_FILE_RANGE_WRITE|
2315                                          SYNC_FILE_RANGE_WAIT_AFTER);
2316                 if (err == 0) {
2317                         written = written_buffered;
2318                         invalidate_mapping_pages(mapping,
2319                                                  pos >> PAGE_CACHE_SHIFT,
2320                                                  endbyte >> PAGE_CACHE_SHIFT);
2321                 } else {
2322                         /*
2323                          * We don't know how much we wrote, so just return
2324                          * the number of bytes which were direct-written
2325                          */
2326                 }
2327         } else {
2328                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2329                                 pos, ppos, count, written);
2330         }
2331 out:
2332         current->backing_dev_info = NULL;
2333         return written ? written : err;
2334 }
2335
2336 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2337                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2338 {
2339         struct file *file = iocb->ki_filp;
2340         struct address_space *mapping = file->f_mapping;
2341         struct inode *inode = mapping->host;
2342         ssize_t ret;
2343
2344         BUG_ON(iocb->ki_pos != pos);
2345
2346         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2347                         &iocb->ki_pos);
2348
2349         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2350                 ssize_t err;
2351
2352                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2353                 if (err < 0)
2354                         ret = err;
2355         }
2356         return ret;
2357 }
2358 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2359
2360 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2361                 unsigned long nr_segs, loff_t pos)
2362 {
2363         struct file *file = iocb->ki_filp;
2364         struct address_space *mapping = file->f_mapping;
2365         struct inode *inode = mapping->host;
2366         ssize_t ret;
2367
2368         BUG_ON(iocb->ki_pos != pos);
2369
2370         mutex_lock(&inode->i_mutex);
2371         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2372                         &iocb->ki_pos);
2373         mutex_unlock(&inode->i_mutex);
2374
2375         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2376                 ssize_t err;
2377
2378                 err = sync_page_range(inode, mapping, pos, ret);
2379                 if (err < 0)
2380                         ret = err;
2381         }
2382         return ret;
2383 }
2384 EXPORT_SYMBOL(generic_file_aio_write);
2385
2386 /*
2387  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2388  * went wrong during pagecache shootdown.
2389  */
2390 static ssize_t
2391 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2392         loff_t offset, unsigned long nr_segs)
2393 {
2394         struct file *file = iocb->ki_filp;
2395         struct address_space *mapping = file->f_mapping;
2396         ssize_t retval;
2397         size_t write_len = 0;
2398
2399         /*
2400          * If it's a write, unmap all mmappings of the file up-front.  This
2401          * will cause any pte dirty bits to be propagated into the pageframes
2402          * for the subsequent filemap_write_and_wait().
2403          */
2404         if (rw == WRITE) {
2405                 write_len = iov_length(iov, nr_segs);
2406                 if (mapping_mapped(mapping))
2407                         unmap_mapping_range(mapping, offset, write_len, 0);
2408         }
2409
2410         retval = filemap_write_and_wait(mapping);
2411         if (retval == 0) {
2412                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2413                                                 offset, nr_segs);
2414                 if (rw == WRITE && mapping->nrpages) {
2415                         pgoff_t end = (offset + write_len - 1)
2416                                                 >> PAGE_CACHE_SHIFT;
2417                         int err = invalidate_inode_pages2_range(mapping,
2418                                         offset >> PAGE_CACHE_SHIFT, end);
2419                         if (err)
2420                                 retval = err;
2421                 }
2422         }
2423         return retval;
2424 }
2425
2426 /**
2427  * try_to_release_page() - release old fs-specific metadata on a page
2428  *
2429  * @page: the page which the kernel is trying to free
2430  * @gfp_mask: memory allocation flags (and I/O mode)
2431  *
2432  * The address_space is to try to release any data against the page
2433  * (presumably at page->private).  If the release was successful, return `1'.
2434  * Otherwise return zero.
2435  *
2436  * The @gfp_mask argument specifies whether I/O may be performed to release
2437  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2438  *
2439  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2440  */
2441 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2442 {
2443         struct address_space * const mapping = page->mapping;
2444
2445         BUG_ON(!PageLocked(page));
2446         if (PageWriteback(page))
2447                 return 0;
2448
2449         if (mapping && mapping->a_ops->releasepage)
2450                 return mapping->a_ops->releasepage(page, gfp_mask);
2451         return try_to_free_buffers(page);
2452 }
2453
2454 EXPORT_SYMBOL(try_to_release_page);