9c5e6235cc743e37787c124e9f26a9e287ac8c06
[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/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for generic_osync_inode */
43
44 #include <asm/mman.h>
45
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 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         BUG_ON(page_mapped(page));
124         mem_cgroup_uncharge_cache_page(page);
125
126         /*
127          * Some filesystems seem to re-dirty the page even after
128          * the VM has canceled the dirty bit (eg ext3 journaling).
129          *
130          * Fix it up by doing a final dirty accounting check after
131          * having removed the page entirely.
132          */
133         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
134                 dec_zone_page_state(page, NR_FILE_DIRTY);
135                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
136         }
137 }
138
139 void remove_from_page_cache(struct page *page)
140 {
141         struct address_space *mapping = page->mapping;
142
143         BUG_ON(!PageLocked(page));
144
145         spin_lock_irq(&mapping->tree_lock);
146         __remove_from_page_cache(page);
147         spin_unlock_irq(&mapping->tree_lock);
148 }
149
150 static int sync_page(void *word)
151 {
152         struct address_space *mapping;
153         struct page *page;
154
155         page = container_of((unsigned long *)word, struct page, flags);
156
157         /*
158          * page_mapping() is being called without PG_locked held.
159          * Some knowledge of the state and use of the page is used to
160          * reduce the requirements down to a memory barrier.
161          * The danger here is of a stale page_mapping() return value
162          * indicating a struct address_space different from the one it's
163          * associated with when it is associated with one.
164          * After smp_mb(), it's either the correct page_mapping() for
165          * the page, or an old page_mapping() and the page's own
166          * page_mapping() has gone NULL.
167          * The ->sync_page() address_space operation must tolerate
168          * page_mapping() going NULL. By an amazing coincidence,
169          * this comes about because none of the users of the page
170          * in the ->sync_page() methods make essential use of the
171          * page_mapping(), merely passing the page down to the backing
172          * device's unplug functions when it's non-NULL, which in turn
173          * ignore it for all cases but swap, where only page_private(page) is
174          * of interest. When page_mapping() does go NULL, the entire
175          * call stack gracefully ignores the page and returns.
176          * -- wli
177          */
178         smp_mb();
179         mapping = page_mapping(page);
180         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
181                 mapping->a_ops->sync_page(page);
182         io_schedule();
183         return 0;
184 }
185
186 static int sync_page_killable(void *word)
187 {
188         sync_page(word);
189         return fatal_signal_pending(current) ? -EINTR : 0;
190 }
191
192 /**
193  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
194  * @mapping:    address space structure to write
195  * @start:      offset in bytes where the range starts
196  * @end:        offset in bytes where the range ends (inclusive)
197  * @sync_mode:  enable synchronous operation
198  *
199  * Start writeback against all of a mapping's dirty pages that lie
200  * within the byte offsets <start, end> inclusive.
201  *
202  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
203  * opposed to a regular memory cleansing writeback.  The difference between
204  * these two operations is that if a dirty page/buffer is encountered, it must
205  * be waited upon, and not just skipped over.
206  */
207 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
208                                 loff_t end, int sync_mode)
209 {
210         int ret;
211         struct writeback_control wbc = {
212                 .sync_mode = sync_mode,
213                 .nr_to_write = LONG_MAX,
214                 .range_start = start,
215                 .range_end = end,
216         };
217
218         if (!mapping_cap_writeback_dirty(mapping))
219                 return 0;
220
221         ret = do_writepages(mapping, &wbc);
222         return ret;
223 }
224
225 static inline int __filemap_fdatawrite(struct address_space *mapping,
226         int sync_mode)
227 {
228         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
229 }
230
231 int filemap_fdatawrite(struct address_space *mapping)
232 {
233         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
234 }
235 EXPORT_SYMBOL(filemap_fdatawrite);
236
237 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
238                                 loff_t end)
239 {
240         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
241 }
242 EXPORT_SYMBOL(filemap_fdatawrite_range);
243
244 /**
245  * filemap_flush - mostly a non-blocking flush
246  * @mapping:    target address_space
247  *
248  * This is a mostly non-blocking flush.  Not suitable for data-integrity
249  * purposes - I/O may not be started against all dirty pages.
250  */
251 int filemap_flush(struct address_space *mapping)
252 {
253         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
254 }
255 EXPORT_SYMBOL(filemap_flush);
256
257 /**
258  * wait_on_page_writeback_range - wait for writeback to complete
259  * @mapping:    target address_space
260  * @start:      beginning page index
261  * @end:        ending page index
262  *
263  * Wait for writeback to complete against pages indexed by start->end
264  * inclusive
265  */
266 int wait_on_page_writeback_range(struct address_space *mapping,
267                                 pgoff_t start, pgoff_t end)
268 {
269         struct pagevec pvec;
270         int nr_pages;
271         int ret = 0;
272         pgoff_t index;
273
274         if (end < start)
275                 return 0;
276
277         pagevec_init(&pvec, 0);
278         index = start;
279         while ((index <= end) &&
280                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
281                         PAGECACHE_TAG_WRITEBACK,
282                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
283                 unsigned i;
284
285                 for (i = 0; i < nr_pages; i++) {
286                         struct page *page = pvec.pages[i];
287
288                         /* until radix tree lookup accepts end_index */
289                         if (page->index > end)
290                                 continue;
291
292                         wait_on_page_writeback(page);
293                         if (PageError(page))
294                                 ret = -EIO;
295                 }
296                 pagevec_release(&pvec);
297                 cond_resched();
298         }
299
300         /* Check for outstanding write errors */
301         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
302                 ret = -ENOSPC;
303         if (test_and_clear_bit(AS_EIO, &mapping->flags))
304                 ret = -EIO;
305
306         return ret;
307 }
308
309 /**
310  * sync_page_range - write and wait on all pages in the passed range
311  * @inode:      target inode
312  * @mapping:    target address_space
313  * @pos:        beginning offset in pages to write
314  * @count:      number of bytes to write
315  *
316  * Write and wait upon all the pages in the passed range.  This is a "data
317  * integrity" operation.  It waits upon in-flight writeout before starting and
318  * waiting upon new writeout.  If there was an IO error, return it.
319  *
320  * We need to re-take i_mutex during the generic_osync_inode list walk because
321  * it is otherwise livelockable.
322  */
323 int sync_page_range(struct inode *inode, struct address_space *mapping,
324                         loff_t pos, loff_t count)
325 {
326         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
327         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
328         int ret;
329
330         if (!mapping_cap_writeback_dirty(mapping) || !count)
331                 return 0;
332         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
333         if (ret == 0) {
334                 mutex_lock(&inode->i_mutex);
335                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
336                 mutex_unlock(&inode->i_mutex);
337         }
338         if (ret == 0)
339                 ret = wait_on_page_writeback_range(mapping, start, end);
340         return ret;
341 }
342 EXPORT_SYMBOL(sync_page_range);
343
344 /**
345  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
346  * @inode:      target inode
347  * @mapping:    target address_space
348  * @pos:        beginning offset in pages to write
349  * @count:      number of bytes to write
350  *
351  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
352  * as it forces O_SYNC writers to different parts of the same file
353  * to be serialised right until io completion.
354  */
355 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
356                            loff_t pos, loff_t count)
357 {
358         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
359         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
360         int ret;
361
362         if (!mapping_cap_writeback_dirty(mapping) || !count)
363                 return 0;
364         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
365         if (ret == 0)
366                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
367         if (ret == 0)
368                 ret = wait_on_page_writeback_range(mapping, start, end);
369         return ret;
370 }
371 EXPORT_SYMBOL(sync_page_range_nolock);
372
373 /**
374  * filemap_fdatawait - wait for all under-writeback pages to complete
375  * @mapping: address space structure to wait for
376  *
377  * Walk the list of under-writeback pages of the given address space
378  * and wait for all of them.
379  */
380 int filemap_fdatawait(struct address_space *mapping)
381 {
382         loff_t i_size = i_size_read(mapping->host);
383
384         if (i_size == 0)
385                 return 0;
386
387         return wait_on_page_writeback_range(mapping, 0,
388                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
389 }
390 EXPORT_SYMBOL(filemap_fdatawait);
391
392 int filemap_write_and_wait(struct address_space *mapping)
393 {
394         int err = 0;
395
396         if (mapping->nrpages) {
397                 err = filemap_fdatawrite(mapping);
398                 /*
399                  * Even if the above returned error, the pages may be
400                  * written partially (e.g. -ENOSPC), so we wait for it.
401                  * But the -EIO is special case, it may indicate the worst
402                  * thing (e.g. bug) happened, so we avoid waiting for it.
403                  */
404                 if (err != -EIO) {
405                         int err2 = filemap_fdatawait(mapping);
406                         if (!err)
407                                 err = err2;
408                 }
409         }
410         return err;
411 }
412 EXPORT_SYMBOL(filemap_write_and_wait);
413
414 /**
415  * filemap_write_and_wait_range - write out & wait on a file range
416  * @mapping:    the address_space for the pages
417  * @lstart:     offset in bytes where the range starts
418  * @lend:       offset in bytes where the range ends (inclusive)
419  *
420  * Write out and wait upon file offsets lstart->lend, inclusive.
421  *
422  * Note that `lend' is inclusive (describes the last byte to be written) so
423  * that this function can be used to write to the very end-of-file (end = -1).
424  */
425 int filemap_write_and_wait_range(struct address_space *mapping,
426                                  loff_t lstart, loff_t lend)
427 {
428         int err = 0;
429
430         if (mapping->nrpages) {
431                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
432                                                  WB_SYNC_ALL);
433                 /* See comment of filemap_write_and_wait() */
434                 if (err != -EIO) {
435                         int err2 = wait_on_page_writeback_range(mapping,
436                                                 lstart >> PAGE_CACHE_SHIFT,
437                                                 lend >> PAGE_CACHE_SHIFT);
438                         if (!err)
439                                 err = err2;
440                 }
441         }
442         return err;
443 }
444
445 /**
446  * add_to_page_cache_locked - add a locked page to the pagecache
447  * @page:       page to add
448  * @mapping:    the page's address_space
449  * @offset:     page index
450  * @gfp_mask:   page allocation mode
451  *
452  * This function is used to add a page to the pagecache. It must be locked.
453  * This function does not add the page to the LRU.  The caller must do that.
454  */
455 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
456                 pgoff_t offset, gfp_t gfp_mask)
457 {
458         int error;
459
460         VM_BUG_ON(!PageLocked(page));
461
462         error = mem_cgroup_cache_charge(page, current->mm,
463                                         gfp_mask & ~__GFP_HIGHMEM);
464         if (error)
465                 goto out;
466
467         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
468         if (error == 0) {
469                 page_cache_get(page);
470                 page->mapping = mapping;
471                 page->index = offset;
472
473                 spin_lock_irq(&mapping->tree_lock);
474                 error = radix_tree_insert(&mapping->page_tree, offset, page);
475                 if (likely(!error)) {
476                         mapping->nrpages++;
477                         __inc_zone_page_state(page, NR_FILE_PAGES);
478                 } else {
479                         page->mapping = NULL;
480                         mem_cgroup_uncharge_cache_page(page);
481                         page_cache_release(page);
482                 }
483
484                 spin_unlock_irq(&mapping->tree_lock);
485                 radix_tree_preload_end();
486         } else
487                 mem_cgroup_uncharge_cache_page(page);
488 out:
489         return error;
490 }
491 EXPORT_SYMBOL(add_to_page_cache_locked);
492
493 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
494                                 pgoff_t offset, gfp_t gfp_mask)
495 {
496         int ret;
497
498         /*
499          * Splice_read and readahead add shmem/tmpfs pages into the page cache
500          * before shmem_readpage has a chance to mark them as SwapBacked: they
501          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
502          * (called in add_to_page_cache) needs to know where they're going too.
503          */
504         if (mapping_cap_swap_backed(mapping))
505                 SetPageSwapBacked(page);
506
507         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
508         if (ret == 0) {
509                 if (page_is_file_cache(page))
510                         lru_cache_add_file(page);
511                 else
512                         lru_cache_add_active_anon(page);
513         }
514         return ret;
515 }
516
517 #ifdef CONFIG_NUMA
518 struct page *__page_cache_alloc(gfp_t gfp)
519 {
520         if (cpuset_do_page_mem_spread()) {
521                 int n = cpuset_mem_spread_node();
522                 return alloc_pages_node(n, gfp, 0);
523         }
524         return alloc_pages(gfp, 0);
525 }
526 EXPORT_SYMBOL(__page_cache_alloc);
527 #endif
528
529 static int __sleep_on_page_lock(void *word)
530 {
531         io_schedule();
532         return 0;
533 }
534
535 /*
536  * In order to wait for pages to become available there must be
537  * waitqueues associated with pages. By using a hash table of
538  * waitqueues where the bucket discipline is to maintain all
539  * waiters on the same queue and wake all when any of the pages
540  * become available, and for the woken contexts to check to be
541  * sure the appropriate page became available, this saves space
542  * at a cost of "thundering herd" phenomena during rare hash
543  * collisions.
544  */
545 static wait_queue_head_t *page_waitqueue(struct page *page)
546 {
547         const struct zone *zone = page_zone(page);
548
549         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
550 }
551
552 static inline void wake_up_page(struct page *page, int bit)
553 {
554         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
555 }
556
557 void wait_on_page_bit(struct page *page, int bit_nr)
558 {
559         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
560
561         if (test_bit(bit_nr, &page->flags))
562                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
563                                                         TASK_UNINTERRUPTIBLE);
564 }
565 EXPORT_SYMBOL(wait_on_page_bit);
566
567 /**
568  * unlock_page - unlock a locked page
569  * @page: the page
570  *
571  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
572  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
573  * mechananism between PageLocked pages and PageWriteback pages is shared.
574  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
575  *
576  * The mb is necessary to enforce ordering between the clear_bit and the read
577  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
578  */
579 void unlock_page(struct page *page)
580 {
581         VM_BUG_ON(!PageLocked(page));
582         clear_bit_unlock(PG_locked, &page->flags);
583         smp_mb__after_clear_bit();
584         wake_up_page(page, PG_locked);
585 }
586 EXPORT_SYMBOL(unlock_page);
587
588 /**
589  * end_page_writeback - end writeback against a page
590  * @page: the page
591  */
592 void end_page_writeback(struct page *page)
593 {
594         if (TestClearPageReclaim(page))
595                 rotate_reclaimable_page(page);
596
597         if (!test_clear_page_writeback(page))
598                 BUG();
599
600         smp_mb__after_clear_bit();
601         wake_up_page(page, PG_writeback);
602 }
603 EXPORT_SYMBOL(end_page_writeback);
604
605 /**
606  * __lock_page - get a lock on the page, assuming we need to sleep to get it
607  * @page: the page to lock
608  *
609  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
610  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
611  * chances are that on the second loop, the block layer's plug list is empty,
612  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
613  */
614 void __lock_page(struct page *page)
615 {
616         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
617
618         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
619                                                         TASK_UNINTERRUPTIBLE);
620 }
621 EXPORT_SYMBOL(__lock_page);
622
623 int __lock_page_killable(struct page *page)
624 {
625         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
626
627         return __wait_on_bit_lock(page_waitqueue(page), &wait,
628                                         sync_page_killable, TASK_KILLABLE);
629 }
630
631 /**
632  * __lock_page_nosync - get a lock on the page, without calling sync_page()
633  * @page: the page to lock
634  *
635  * Variant of lock_page that does not require the caller to hold a reference
636  * on the page's mapping.
637  */
638 void __lock_page_nosync(struct page *page)
639 {
640         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
641         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
642                                                         TASK_UNINTERRUPTIBLE);
643 }
644
645 /**
646  * find_get_page - find and get a page reference
647  * @mapping: the address_space to search
648  * @offset: the page index
649  *
650  * Is there a pagecache struct page at the given (mapping, offset) tuple?
651  * If yes, increment its refcount and return it; if no, return NULL.
652  */
653 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
654 {
655         void **pagep;
656         struct page *page;
657
658         rcu_read_lock();
659 repeat:
660         page = NULL;
661         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
662         if (pagep) {
663                 page = radix_tree_deref_slot(pagep);
664                 if (unlikely(!page || page == RADIX_TREE_RETRY))
665                         goto repeat;
666
667                 if (!page_cache_get_speculative(page))
668                         goto repeat;
669
670                 /*
671                  * Has the page moved?
672                  * This is part of the lockless pagecache protocol. See
673                  * include/linux/pagemap.h for details.
674                  */
675                 if (unlikely(page != *pagep)) {
676                         page_cache_release(page);
677                         goto repeat;
678                 }
679         }
680         rcu_read_unlock();
681
682         return page;
683 }
684 EXPORT_SYMBOL(find_get_page);
685
686 /**
687  * find_lock_page - locate, pin and lock a pagecache page
688  * @mapping: the address_space to search
689  * @offset: the page index
690  *
691  * Locates the desired pagecache page, locks it, increments its reference
692  * count and returns its address.
693  *
694  * Returns zero if the page was not present. find_lock_page() may sleep.
695  */
696 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
697 {
698         struct page *page;
699
700 repeat:
701         page = find_get_page(mapping, offset);
702         if (page) {
703                 lock_page(page);
704                 /* Has the page been truncated? */
705                 if (unlikely(page->mapping != mapping)) {
706                         unlock_page(page);
707                         page_cache_release(page);
708                         goto repeat;
709                 }
710                 VM_BUG_ON(page->index != offset);
711         }
712         return page;
713 }
714 EXPORT_SYMBOL(find_lock_page);
715
716 /**
717  * find_or_create_page - locate or add a pagecache page
718  * @mapping: the page's address_space
719  * @index: the page's index into the mapping
720  * @gfp_mask: page allocation mode
721  *
722  * Locates a page in the pagecache.  If the page is not present, a new page
723  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
724  * LRU list.  The returned page is locked and has its reference count
725  * incremented.
726  *
727  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
728  * allocation!
729  *
730  * find_or_create_page() returns the desired page's address, or zero on
731  * memory exhaustion.
732  */
733 struct page *find_or_create_page(struct address_space *mapping,
734                 pgoff_t index, gfp_t gfp_mask)
735 {
736         struct page *page;
737         int err;
738 repeat:
739         page = find_lock_page(mapping, index);
740         if (!page) {
741                 page = __page_cache_alloc(gfp_mask);
742                 if (!page)
743                         return NULL;
744                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
745                 if (unlikely(err)) {
746                         page_cache_release(page);
747                         page = NULL;
748                         if (err == -EEXIST)
749                                 goto repeat;
750                 }
751         }
752         return page;
753 }
754 EXPORT_SYMBOL(find_or_create_page);
755
756 /**
757  * find_get_pages - gang pagecache lookup
758  * @mapping:    The address_space to search
759  * @start:      The starting page index
760  * @nr_pages:   The maximum number of pages
761  * @pages:      Where the resulting pages are placed
762  *
763  * find_get_pages() will search for and return a group of up to
764  * @nr_pages pages in the mapping.  The pages are placed at @pages.
765  * find_get_pages() takes a reference against the returned pages.
766  *
767  * The search returns a group of mapping-contiguous pages with ascending
768  * indexes.  There may be holes in the indices due to not-present pages.
769  *
770  * find_get_pages() returns the number of pages which were found.
771  */
772 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
773                             unsigned int nr_pages, struct page **pages)
774 {
775         unsigned int i;
776         unsigned int ret;
777         unsigned int nr_found;
778
779         rcu_read_lock();
780 restart:
781         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
782                                 (void ***)pages, start, nr_pages);
783         ret = 0;
784         for (i = 0; i < nr_found; i++) {
785                 struct page *page;
786 repeat:
787                 page = radix_tree_deref_slot((void **)pages[i]);
788                 if (unlikely(!page))
789                         continue;
790                 /*
791                  * this can only trigger if nr_found == 1, making livelock
792                  * a non issue.
793                  */
794                 if (unlikely(page == RADIX_TREE_RETRY))
795                         goto restart;
796
797                 if (!page_cache_get_speculative(page))
798                         goto repeat;
799
800                 /* Has the page moved? */
801                 if (unlikely(page != *((void **)pages[i]))) {
802                         page_cache_release(page);
803                         goto repeat;
804                 }
805
806                 pages[ret] = page;
807                 ret++;
808         }
809         rcu_read_unlock();
810         return ret;
811 }
812
813 /**
814  * find_get_pages_contig - gang contiguous pagecache lookup
815  * @mapping:    The address_space to search
816  * @index:      The starting page index
817  * @nr_pages:   The maximum number of pages
818  * @pages:      Where the resulting pages are placed
819  *
820  * find_get_pages_contig() works exactly like find_get_pages(), except
821  * that the returned number of pages are guaranteed to be contiguous.
822  *
823  * find_get_pages_contig() returns the number of pages which were found.
824  */
825 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
826                                unsigned int nr_pages, struct page **pages)
827 {
828         unsigned int i;
829         unsigned int ret;
830         unsigned int nr_found;
831
832         rcu_read_lock();
833 restart:
834         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
835                                 (void ***)pages, index, nr_pages);
836         ret = 0;
837         for (i = 0; i < nr_found; i++) {
838                 struct page *page;
839 repeat:
840                 page = radix_tree_deref_slot((void **)pages[i]);
841                 if (unlikely(!page))
842                         continue;
843                 /*
844                  * this can only trigger if nr_found == 1, making livelock
845                  * a non issue.
846                  */
847                 if (unlikely(page == RADIX_TREE_RETRY))
848                         goto restart;
849
850                 if (page->mapping == NULL || page->index != index)
851                         break;
852
853                 if (!page_cache_get_speculative(page))
854                         goto repeat;
855
856                 /* Has the page moved? */
857                 if (unlikely(page != *((void **)pages[i]))) {
858                         page_cache_release(page);
859                         goto repeat;
860                 }
861
862                 pages[ret] = page;
863                 ret++;
864                 index++;
865         }
866         rcu_read_unlock();
867         return ret;
868 }
869 EXPORT_SYMBOL(find_get_pages_contig);
870
871 /**
872  * find_get_pages_tag - find and return pages that match @tag
873  * @mapping:    the address_space to search
874  * @index:      the starting page index
875  * @tag:        the tag index
876  * @nr_pages:   the maximum number of pages
877  * @pages:      where the resulting pages are placed
878  *
879  * Like find_get_pages, except we only return pages which are tagged with
880  * @tag.   We update @index to index the next page for the traversal.
881  */
882 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
883                         int tag, unsigned int nr_pages, struct page **pages)
884 {
885         unsigned int i;
886         unsigned int ret;
887         unsigned int nr_found;
888
889         rcu_read_lock();
890 restart:
891         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
892                                 (void ***)pages, *index, nr_pages, tag);
893         ret = 0;
894         for (i = 0; i < nr_found; i++) {
895                 struct page *page;
896 repeat:
897                 page = radix_tree_deref_slot((void **)pages[i]);
898                 if (unlikely(!page))
899                         continue;
900                 /*
901                  * this can only trigger if nr_found == 1, making livelock
902                  * a non issue.
903                  */
904                 if (unlikely(page == RADIX_TREE_RETRY))
905                         goto restart;
906
907                 if (!page_cache_get_speculative(page))
908                         goto repeat;
909
910                 /* Has the page moved? */
911                 if (unlikely(page != *((void **)pages[i]))) {
912                         page_cache_release(page);
913                         goto repeat;
914                 }
915
916                 pages[ret] = page;
917                 ret++;
918         }
919         rcu_read_unlock();
920
921         if (ret)
922                 *index = pages[ret - 1]->index + 1;
923
924         return ret;
925 }
926 EXPORT_SYMBOL(find_get_pages_tag);
927
928 /**
929  * grab_cache_page_nowait - returns locked page at given index in given cache
930  * @mapping: target address_space
931  * @index: the page index
932  *
933  * Same as grab_cache_page(), but do not wait if the page is unavailable.
934  * This is intended for speculative data generators, where the data can
935  * be regenerated if the page couldn't be grabbed.  This routine should
936  * be safe to call while holding the lock for another page.
937  *
938  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
939  * and deadlock against the caller's locked page.
940  */
941 struct page *
942 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
943 {
944         struct page *page = find_get_page(mapping, index);
945
946         if (page) {
947                 if (trylock_page(page))
948                         return page;
949                 page_cache_release(page);
950                 return NULL;
951         }
952         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
953         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
954                 page_cache_release(page);
955                 page = NULL;
956         }
957         return page;
958 }
959 EXPORT_SYMBOL(grab_cache_page_nowait);
960
961 /*
962  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
963  * a _large_ part of the i/o request. Imagine the worst scenario:
964  *
965  *      ---R__________________________________________B__________
966  *         ^ reading here                             ^ bad block(assume 4k)
967  *
968  * read(R) => miss => readahead(R...B) => media error => frustrating retries
969  * => failing the whole request => read(R) => read(R+1) =>
970  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
971  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
972  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
973  *
974  * It is going insane. Fix it by quickly scaling down the readahead size.
975  */
976 static void shrink_readahead_size_eio(struct file *filp,
977                                         struct file_ra_state *ra)
978 {
979         if (!ra->ra_pages)
980                 return;
981
982         ra->ra_pages /= 4;
983 }
984
985 /**
986  * do_generic_file_read - generic file read routine
987  * @filp:       the file to read
988  * @ppos:       current file position
989  * @desc:       read_descriptor
990  * @actor:      read method
991  *
992  * This is a generic file read routine, and uses the
993  * mapping->a_ops->readpage() function for the actual low-level stuff.
994  *
995  * This is really ugly. But the goto's actually try to clarify some
996  * of the logic when it comes to error handling etc.
997  */
998 static void do_generic_file_read(struct file *filp, loff_t *ppos,
999                 read_descriptor_t *desc, read_actor_t actor)
1000 {
1001         struct address_space *mapping = filp->f_mapping;
1002         struct inode *inode = mapping->host;
1003         struct file_ra_state *ra = &filp->f_ra;
1004         pgoff_t index;
1005         pgoff_t last_index;
1006         pgoff_t prev_index;
1007         unsigned long offset;      /* offset into pagecache page */
1008         unsigned int prev_offset;
1009         int error;
1010
1011         index = *ppos >> PAGE_CACHE_SHIFT;
1012         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1013         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1014         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1015         offset = *ppos & ~PAGE_CACHE_MASK;
1016
1017         for (;;) {
1018                 struct page *page;
1019                 pgoff_t end_index;
1020                 loff_t isize;
1021                 unsigned long nr, ret;
1022
1023                 cond_resched();
1024 find_page:
1025                 page = find_get_page(mapping, index);
1026                 if (!page) {
1027                         page_cache_sync_readahead(mapping,
1028                                         ra, filp,
1029                                         index, last_index - index);
1030                         page = find_get_page(mapping, index);
1031                         if (unlikely(page == NULL))
1032                                 goto no_cached_page;
1033                 }
1034                 if (PageReadahead(page)) {
1035                         page_cache_async_readahead(mapping,
1036                                         ra, filp, page,
1037                                         index, last_index - index);
1038                 }
1039                 if (!PageUptodate(page)) {
1040                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1041                                         !mapping->a_ops->is_partially_uptodate)
1042                                 goto page_not_up_to_date;
1043                         if (!trylock_page(page))
1044                                 goto page_not_up_to_date;
1045                         if (!mapping->a_ops->is_partially_uptodate(page,
1046                                                                 desc, offset))
1047                                 goto page_not_up_to_date_locked;
1048                         unlock_page(page);
1049                 }
1050 page_ok:
1051                 /*
1052                  * i_size must be checked after we know the page is Uptodate.
1053                  *
1054                  * Checking i_size after the check allows us to calculate
1055                  * the correct value for "nr", which means the zero-filled
1056                  * part of the page is not copied back to userspace (unless
1057                  * another truncate extends the file - this is desired though).
1058                  */
1059
1060                 isize = i_size_read(inode);
1061                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1062                 if (unlikely(!isize || index > end_index)) {
1063                         page_cache_release(page);
1064                         goto out;
1065                 }
1066
1067                 /* nr is the maximum number of bytes to copy from this page */
1068                 nr = PAGE_CACHE_SIZE;
1069                 if (index == end_index) {
1070                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1071                         if (nr <= offset) {
1072                                 page_cache_release(page);
1073                                 goto out;
1074                         }
1075                 }
1076                 nr = nr - offset;
1077
1078                 /* If users can be writing to this page using arbitrary
1079                  * virtual addresses, take care about potential aliasing
1080                  * before reading the page on the kernel side.
1081                  */
1082                 if (mapping_writably_mapped(mapping))
1083                         flush_dcache_page(page);
1084
1085                 /*
1086                  * When a sequential read accesses a page several times,
1087                  * only mark it as accessed the first time.
1088                  */
1089                 if (prev_index != index || offset != prev_offset)
1090                         mark_page_accessed(page);
1091                 prev_index = index;
1092
1093                 /*
1094                  * Ok, we have the page, and it's up-to-date, so
1095                  * now we can copy it to user space...
1096                  *
1097                  * The actor routine returns how many bytes were actually used..
1098                  * NOTE! This may not be the same as how much of a user buffer
1099                  * we filled up (we may be padding etc), so we can only update
1100                  * "pos" here (the actor routine has to update the user buffer
1101                  * pointers and the remaining count).
1102                  */
1103                 ret = actor(desc, page, offset, nr);
1104                 offset += ret;
1105                 index += offset >> PAGE_CACHE_SHIFT;
1106                 offset &= ~PAGE_CACHE_MASK;
1107                 prev_offset = offset;
1108
1109                 page_cache_release(page);
1110                 if (ret == nr && desc->count)
1111                         continue;
1112                 goto out;
1113
1114 page_not_up_to_date:
1115                 /* Get exclusive access to the page ... */
1116                 error = lock_page_killable(page);
1117                 if (unlikely(error))
1118                         goto readpage_error;
1119
1120 page_not_up_to_date_locked:
1121                 /* Did it get truncated before we got the lock? */
1122                 if (!page->mapping) {
1123                         unlock_page(page);
1124                         page_cache_release(page);
1125                         continue;
1126                 }
1127
1128                 /* Did somebody else fill it already? */
1129                 if (PageUptodate(page)) {
1130                         unlock_page(page);
1131                         goto page_ok;
1132                 }
1133
1134 readpage:
1135                 /* Start the actual read. The read will unlock the page. */
1136                 error = mapping->a_ops->readpage(filp, page);
1137
1138                 if (unlikely(error)) {
1139                         if (error == AOP_TRUNCATED_PAGE) {
1140                                 page_cache_release(page);
1141                                 goto find_page;
1142                         }
1143                         goto readpage_error;
1144                 }
1145
1146                 if (!PageUptodate(page)) {
1147                         error = lock_page_killable(page);
1148                         if (unlikely(error))
1149                                 goto readpage_error;
1150                         if (!PageUptodate(page)) {
1151                                 if (page->mapping == NULL) {
1152                                         /*
1153                                          * invalidate_inode_pages got it
1154                                          */
1155                                         unlock_page(page);
1156                                         page_cache_release(page);
1157                                         goto find_page;
1158                                 }
1159                                 unlock_page(page);
1160                                 shrink_readahead_size_eio(filp, ra);
1161                                 error = -EIO;
1162                                 goto readpage_error;
1163                         }
1164                         unlock_page(page);
1165                 }
1166
1167                 goto page_ok;
1168
1169 readpage_error:
1170                 /* UHHUH! A synchronous read error occurred. Report it */
1171                 desc->error = error;
1172                 page_cache_release(page);
1173                 goto out;
1174
1175 no_cached_page:
1176                 /*
1177                  * Ok, it wasn't cached, so we need to create a new
1178                  * page..
1179                  */
1180                 page = page_cache_alloc_cold(mapping);
1181                 if (!page) {
1182                         desc->error = -ENOMEM;
1183                         goto out;
1184                 }
1185                 error = add_to_page_cache_lru(page, mapping,
1186                                                 index, GFP_KERNEL);
1187                 if (error) {
1188                         page_cache_release(page);
1189                         if (error == -EEXIST)
1190                                 goto find_page;
1191                         desc->error = error;
1192                         goto out;
1193                 }
1194                 goto readpage;
1195         }
1196
1197 out:
1198         ra->prev_pos = prev_index;
1199         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1200         ra->prev_pos |= prev_offset;
1201
1202         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1203         file_accessed(filp);
1204 }
1205
1206 int file_read_actor(read_descriptor_t *desc, struct page *page,
1207                         unsigned long offset, unsigned long size)
1208 {
1209         char *kaddr;
1210         unsigned long left, count = desc->count;
1211
1212         if (size > count)
1213                 size = count;
1214
1215         /*
1216          * Faults on the destination of a read are common, so do it before
1217          * taking the kmap.
1218          */
1219         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1220                 kaddr = kmap_atomic(page, KM_USER0);
1221                 left = __copy_to_user_inatomic(desc->arg.buf,
1222                                                 kaddr + offset, size);
1223                 kunmap_atomic(kaddr, KM_USER0);
1224                 if (left == 0)
1225                         goto success;
1226         }
1227
1228         /* Do it the slow way */
1229         kaddr = kmap(page);
1230         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1231         kunmap(page);
1232
1233         if (left) {
1234                 size -= left;
1235                 desc->error = -EFAULT;
1236         }
1237 success:
1238         desc->count = count - size;
1239         desc->written += size;
1240         desc->arg.buf += size;
1241         return size;
1242 }
1243
1244 /*
1245  * Performs necessary checks before doing a write
1246  * @iov:        io vector request
1247  * @nr_segs:    number of segments in the iovec
1248  * @count:      number of bytes to write
1249  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1250  *
1251  * Adjust number of segments and amount of bytes to write (nr_segs should be
1252  * properly initialized first). Returns appropriate error code that caller
1253  * should return or zero in case that write should be allowed.
1254  */
1255 int generic_segment_checks(const struct iovec *iov,
1256                         unsigned long *nr_segs, size_t *count, int access_flags)
1257 {
1258         unsigned long   seg;
1259         size_t cnt = 0;
1260         for (seg = 0; seg < *nr_segs; seg++) {
1261                 const struct iovec *iv = &iov[seg];
1262
1263                 /*
1264                  * If any segment has a negative length, or the cumulative
1265                  * length ever wraps negative then return -EINVAL.
1266                  */
1267                 cnt += iv->iov_len;
1268                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1269                         return -EINVAL;
1270                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1271                         continue;
1272                 if (seg == 0)
1273                         return -EFAULT;
1274                 *nr_segs = seg;
1275                 cnt -= iv->iov_len;     /* This segment is no good */
1276                 break;
1277         }
1278         *count = cnt;
1279         return 0;
1280 }
1281 EXPORT_SYMBOL(generic_segment_checks);
1282
1283 /**
1284  * generic_file_aio_read - generic filesystem read routine
1285  * @iocb:       kernel I/O control block
1286  * @iov:        io vector request
1287  * @nr_segs:    number of segments in the iovec
1288  * @pos:        current file position
1289  *
1290  * This is the "read()" routine for all filesystems
1291  * that can use the page cache directly.
1292  */
1293 ssize_t
1294 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1295                 unsigned long nr_segs, loff_t pos)
1296 {
1297         struct file *filp = iocb->ki_filp;
1298         ssize_t retval;
1299         unsigned long seg;
1300         size_t count;
1301         loff_t *ppos = &iocb->ki_pos;
1302
1303         count = 0;
1304         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1305         if (retval)
1306                 return retval;
1307
1308         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1309         if (filp->f_flags & O_DIRECT) {
1310                 loff_t size;
1311                 struct address_space *mapping;
1312                 struct inode *inode;
1313
1314                 mapping = filp->f_mapping;
1315                 inode = mapping->host;
1316                 if (!count)
1317                         goto out; /* skip atime */
1318                 size = i_size_read(inode);
1319                 if (pos < size) {
1320                         retval = filemap_write_and_wait(mapping);
1321                         if (!retval) {
1322                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1323                                                         iov, pos, nr_segs);
1324                         }
1325                         if (retval > 0)
1326                                 *ppos = pos + retval;
1327                         if (retval) {
1328                                 file_accessed(filp);
1329                                 goto out;
1330                         }
1331                 }
1332         }
1333
1334         for (seg = 0; seg < nr_segs; seg++) {
1335                 read_descriptor_t desc;
1336
1337                 desc.written = 0;
1338                 desc.arg.buf = iov[seg].iov_base;
1339                 desc.count = iov[seg].iov_len;
1340                 if (desc.count == 0)
1341                         continue;
1342                 desc.error = 0;
1343                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1344                 retval += desc.written;
1345                 if (desc.error) {
1346                         retval = retval ?: desc.error;
1347                         break;
1348                 }
1349                 if (desc.count > 0)
1350                         break;
1351         }
1352 out:
1353         return retval;
1354 }
1355 EXPORT_SYMBOL(generic_file_aio_read);
1356
1357 static ssize_t
1358 do_readahead(struct address_space *mapping, struct file *filp,
1359              pgoff_t index, unsigned long nr)
1360 {
1361         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1362                 return -EINVAL;
1363
1364         force_page_cache_readahead(mapping, filp, index,
1365                                         max_sane_readahead(nr));
1366         return 0;
1367 }
1368
1369 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1370 {
1371         ssize_t ret;
1372         struct file *file;
1373
1374         ret = -EBADF;
1375         file = fget(fd);
1376         if (file) {
1377                 if (file->f_mode & FMODE_READ) {
1378                         struct address_space *mapping = file->f_mapping;
1379                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1380                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1381                         unsigned long len = end - start + 1;
1382                         ret = do_readahead(mapping, file, start, len);
1383                 }
1384                 fput(file);
1385         }
1386         return ret;
1387 }
1388
1389 #ifdef CONFIG_MMU
1390 /**
1391  * page_cache_read - adds requested page to the page cache if not already there
1392  * @file:       file to read
1393  * @offset:     page index
1394  *
1395  * This adds the requested page to the page cache if it isn't already there,
1396  * and schedules an I/O to read in its contents from disk.
1397  */
1398 static int page_cache_read(struct file *file, pgoff_t offset)
1399 {
1400         struct address_space *mapping = file->f_mapping;
1401         struct page *page; 
1402         int ret;
1403
1404         do {
1405                 page = page_cache_alloc_cold(mapping);
1406                 if (!page)
1407                         return -ENOMEM;
1408
1409                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1410                 if (ret == 0)
1411                         ret = mapping->a_ops->readpage(file, page);
1412                 else if (ret == -EEXIST)
1413                         ret = 0; /* losing race to add is OK */
1414
1415                 page_cache_release(page);
1416
1417         } while (ret == AOP_TRUNCATED_PAGE);
1418                 
1419         return ret;
1420 }
1421
1422 #define MMAP_LOTSAMISS  (100)
1423
1424 /**
1425  * filemap_fault - read in file data for page fault handling
1426  * @vma:        vma in which the fault was taken
1427  * @vmf:        struct vm_fault containing details of the fault
1428  *
1429  * filemap_fault() is invoked via the vma operations vector for a
1430  * mapped memory region to read in file data during a page fault.
1431  *
1432  * The goto's are kind of ugly, but this streamlines the normal case of having
1433  * it in the page cache, and handles the special cases reasonably without
1434  * having a lot of duplicated code.
1435  */
1436 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1437 {
1438         int error;
1439         struct file *file = vma->vm_file;
1440         struct address_space *mapping = file->f_mapping;
1441         struct file_ra_state *ra = &file->f_ra;
1442         struct inode *inode = mapping->host;
1443         struct page *page;
1444         pgoff_t size;
1445         int did_readaround = 0;
1446         int ret = 0;
1447
1448         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1449         if (vmf->pgoff >= size)
1450                 return VM_FAULT_SIGBUS;
1451
1452         /* If we don't want any read-ahead, don't bother */
1453         if (VM_RandomReadHint(vma))
1454                 goto no_cached_page;
1455
1456         /*
1457          * Do we have something in the page cache already?
1458          */
1459 retry_find:
1460         page = find_lock_page(mapping, vmf->pgoff);
1461         /*
1462          * For sequential accesses, we use the generic readahead logic.
1463          */
1464         if (VM_SequentialReadHint(vma)) {
1465                 if (!page) {
1466                         page_cache_sync_readahead(mapping, ra, file,
1467                                                            vmf->pgoff, 1);
1468                         page = find_lock_page(mapping, vmf->pgoff);
1469                         if (!page)
1470                                 goto no_cached_page;
1471                 }
1472                 if (PageReadahead(page)) {
1473                         page_cache_async_readahead(mapping, ra, file, page,
1474                                                            vmf->pgoff, 1);
1475                 }
1476         }
1477
1478         if (!page) {
1479                 unsigned long ra_pages;
1480
1481                 ra->mmap_miss++;
1482
1483                 /*
1484                  * Do we miss much more than hit in this file? If so,
1485                  * stop bothering with read-ahead. It will only hurt.
1486                  */
1487                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1488                         goto no_cached_page;
1489
1490                 /*
1491                  * To keep the pgmajfault counter straight, we need to
1492                  * check did_readaround, as this is an inner loop.
1493                  */
1494                 if (!did_readaround) {
1495                         ret = VM_FAULT_MAJOR;
1496                         count_vm_event(PGMAJFAULT);
1497                 }
1498                 did_readaround = 1;
1499                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1500                 if (ra_pages) {
1501                         pgoff_t start = 0;
1502
1503                         if (vmf->pgoff > ra_pages / 2)
1504                                 start = vmf->pgoff - ra_pages / 2;
1505                         do_page_cache_readahead(mapping, file, start, ra_pages);
1506                 }
1507                 page = find_lock_page(mapping, vmf->pgoff);
1508                 if (!page)
1509                         goto no_cached_page;
1510         }
1511
1512         if (!did_readaround)
1513                 ra->mmap_miss--;
1514
1515         /*
1516          * We have a locked page in the page cache, now we need to check
1517          * that it's up-to-date. If not, it is going to be due to an error.
1518          */
1519         if (unlikely(!PageUptodate(page)))
1520                 goto page_not_uptodate;
1521
1522         /* Must recheck i_size under page lock */
1523         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1524         if (unlikely(vmf->pgoff >= size)) {
1525                 unlock_page(page);
1526                 page_cache_release(page);
1527                 return VM_FAULT_SIGBUS;
1528         }
1529
1530         /*
1531          * Found the page and have a reference on it.
1532          */
1533         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1534         vmf->page = page;
1535         return ret | VM_FAULT_LOCKED;
1536
1537 no_cached_page:
1538         /*
1539          * We're only likely to ever get here if MADV_RANDOM is in
1540          * effect.
1541          */
1542         error = page_cache_read(file, vmf->pgoff);
1543
1544         /*
1545          * The page we want has now been added to the page cache.
1546          * In the unlikely event that someone removed it in the
1547          * meantime, we'll just come back here and read it again.
1548          */
1549         if (error >= 0)
1550                 goto retry_find;
1551
1552         /*
1553          * An error return from page_cache_read can result if the
1554          * system is low on memory, or a problem occurs while trying
1555          * to schedule I/O.
1556          */
1557         if (error == -ENOMEM)
1558                 return VM_FAULT_OOM;
1559         return VM_FAULT_SIGBUS;
1560
1561 page_not_uptodate:
1562         /* IO error path */
1563         if (!did_readaround) {
1564                 ret = VM_FAULT_MAJOR;
1565                 count_vm_event(PGMAJFAULT);
1566         }
1567
1568         /*
1569          * Umm, take care of errors if the page isn't up-to-date.
1570          * Try to re-read it _once_. We do this synchronously,
1571          * because there really aren't any performance issues here
1572          * and we need to check for errors.
1573          */
1574         ClearPageError(page);
1575         error = mapping->a_ops->readpage(file, page);
1576         if (!error) {
1577                 wait_on_page_locked(page);
1578                 if (!PageUptodate(page))
1579                         error = -EIO;
1580         }
1581         page_cache_release(page);
1582
1583         if (!error || error == AOP_TRUNCATED_PAGE)
1584                 goto retry_find;
1585
1586         /* Things didn't work out. Return zero to tell the mm layer so. */
1587         shrink_readahead_size_eio(file, ra);
1588         return VM_FAULT_SIGBUS;
1589 }
1590 EXPORT_SYMBOL(filemap_fault);
1591
1592 struct vm_operations_struct generic_file_vm_ops = {
1593         .fault          = filemap_fault,
1594 };
1595
1596 /* This is used for a general mmap of a disk file */
1597
1598 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1599 {
1600         struct address_space *mapping = file->f_mapping;
1601
1602         if (!mapping->a_ops->readpage)
1603                 return -ENOEXEC;
1604         file_accessed(file);
1605         vma->vm_ops = &generic_file_vm_ops;
1606         vma->vm_flags |= VM_CAN_NONLINEAR;
1607         return 0;
1608 }
1609
1610 /*
1611  * This is for filesystems which do not implement ->writepage.
1612  */
1613 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1614 {
1615         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1616                 return -EINVAL;
1617         return generic_file_mmap(file, vma);
1618 }
1619 #else
1620 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1621 {
1622         return -ENOSYS;
1623 }
1624 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1625 {
1626         return -ENOSYS;
1627 }
1628 #endif /* CONFIG_MMU */
1629
1630 EXPORT_SYMBOL(generic_file_mmap);
1631 EXPORT_SYMBOL(generic_file_readonly_mmap);
1632
1633 static struct page *__read_cache_page(struct address_space *mapping,
1634                                 pgoff_t index,
1635                                 int (*filler)(void *,struct page*),
1636                                 void *data)
1637 {
1638         struct page *page;
1639         int err;
1640 repeat:
1641         page = find_get_page(mapping, index);
1642         if (!page) {
1643                 page = page_cache_alloc_cold(mapping);
1644                 if (!page)
1645                         return ERR_PTR(-ENOMEM);
1646                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1647                 if (unlikely(err)) {
1648                         page_cache_release(page);
1649                         if (err == -EEXIST)
1650                                 goto repeat;
1651                         /* Presumably ENOMEM for radix tree node */
1652                         return ERR_PTR(err);
1653                 }
1654                 err = filler(data, page);
1655                 if (err < 0) {
1656                         page_cache_release(page);
1657                         page = ERR_PTR(err);
1658                 }
1659         }
1660         return page;
1661 }
1662
1663 /**
1664  * read_cache_page_async - read into page cache, fill it if needed
1665  * @mapping:    the page's address_space
1666  * @index:      the page index
1667  * @filler:     function to perform the read
1668  * @data:       destination for read data
1669  *
1670  * Same as read_cache_page, but don't wait for page to become unlocked
1671  * after submitting it to the filler.
1672  *
1673  * Read into the page cache. If a page already exists, and PageUptodate() is
1674  * not set, try to fill the page but don't wait for it to become unlocked.
1675  *
1676  * If the page does not get brought uptodate, return -EIO.
1677  */
1678 struct page *read_cache_page_async(struct address_space *mapping,
1679                                 pgoff_t index,
1680                                 int (*filler)(void *,struct page*),
1681                                 void *data)
1682 {
1683         struct page *page;
1684         int err;
1685
1686 retry:
1687         page = __read_cache_page(mapping, index, filler, data);
1688         if (IS_ERR(page))
1689                 return page;
1690         if (PageUptodate(page))
1691                 goto out;
1692
1693         lock_page(page);
1694         if (!page->mapping) {
1695                 unlock_page(page);
1696                 page_cache_release(page);
1697                 goto retry;
1698         }
1699         if (PageUptodate(page)) {
1700                 unlock_page(page);
1701                 goto out;
1702         }
1703         err = filler(data, page);
1704         if (err < 0) {
1705                 page_cache_release(page);
1706                 return ERR_PTR(err);
1707         }
1708 out:
1709         mark_page_accessed(page);
1710         return page;
1711 }
1712 EXPORT_SYMBOL(read_cache_page_async);
1713
1714 /**
1715  * read_cache_page - read into page cache, fill it if needed
1716  * @mapping:    the page's address_space
1717  * @index:      the page index
1718  * @filler:     function to perform the read
1719  * @data:       destination for read data
1720  *
1721  * Read into the page cache. If a page already exists, and PageUptodate() is
1722  * not set, try to fill the page then wait for it to become unlocked.
1723  *
1724  * If the page does not get brought uptodate, return -EIO.
1725  */
1726 struct page *read_cache_page(struct address_space *mapping,
1727                                 pgoff_t index,
1728                                 int (*filler)(void *,struct page*),
1729                                 void *data)
1730 {
1731         struct page *page;
1732
1733         page = read_cache_page_async(mapping, index, filler, data);
1734         if (IS_ERR(page))
1735                 goto out;
1736         wait_on_page_locked(page);
1737         if (!PageUptodate(page)) {
1738                 page_cache_release(page);
1739                 page = ERR_PTR(-EIO);
1740         }
1741  out:
1742         return page;
1743 }
1744 EXPORT_SYMBOL(read_cache_page);
1745
1746 /*
1747  * The logic we want is
1748  *
1749  *      if suid or (sgid and xgrp)
1750  *              remove privs
1751  */
1752 int should_remove_suid(struct dentry *dentry)
1753 {
1754         mode_t mode = dentry->d_inode->i_mode;
1755         int kill = 0;
1756
1757         /* suid always must be killed */
1758         if (unlikely(mode & S_ISUID))
1759                 kill = ATTR_KILL_SUID;
1760
1761         /*
1762          * sgid without any exec bits is just a mandatory locking mark; leave
1763          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1764          */
1765         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1766                 kill |= ATTR_KILL_SGID;
1767
1768         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1769                 return kill;
1770
1771         return 0;
1772 }
1773 EXPORT_SYMBOL(should_remove_suid);
1774
1775 static int __remove_suid(struct dentry *dentry, int kill)
1776 {
1777         struct iattr newattrs;
1778
1779         newattrs.ia_valid = ATTR_FORCE | kill;
1780         return notify_change(dentry, &newattrs);
1781 }
1782
1783 int file_remove_suid(struct file *file)
1784 {
1785         struct dentry *dentry = file->f_path.dentry;
1786         int killsuid = should_remove_suid(dentry);
1787         int killpriv = security_inode_need_killpriv(dentry);
1788         int error = 0;
1789
1790         if (killpriv < 0)
1791                 return killpriv;
1792         if (killpriv)
1793                 error = security_inode_killpriv(dentry);
1794         if (!error && killsuid)
1795                 error = __remove_suid(dentry, killsuid);
1796
1797         return error;
1798 }
1799 EXPORT_SYMBOL(file_remove_suid);
1800
1801 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1802                         const struct iovec *iov, size_t base, size_t bytes)
1803 {
1804         size_t copied = 0, left = 0;
1805
1806         while (bytes) {
1807                 char __user *buf = iov->iov_base + base;
1808                 int copy = min(bytes, iov->iov_len - base);
1809
1810                 base = 0;
1811                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1812                 copied += copy;
1813                 bytes -= copy;
1814                 vaddr += copy;
1815                 iov++;
1816
1817                 if (unlikely(left))
1818                         break;
1819         }
1820         return copied - left;
1821 }
1822
1823 /*
1824  * Copy as much as we can into the page and return the number of bytes which
1825  * were sucessfully copied.  If a fault is encountered then return the number of
1826  * bytes which were copied.
1827  */
1828 size_t iov_iter_copy_from_user_atomic(struct page *page,
1829                 struct iov_iter *i, unsigned long offset, size_t bytes)
1830 {
1831         char *kaddr;
1832         size_t copied;
1833
1834         BUG_ON(!in_atomic());
1835         kaddr = kmap_atomic(page, KM_USER0);
1836         if (likely(i->nr_segs == 1)) {
1837                 int left;
1838                 char __user *buf = i->iov->iov_base + i->iov_offset;
1839                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1840                                                         buf, bytes);
1841                 copied = bytes - left;
1842         } else {
1843                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1844                                                 i->iov, i->iov_offset, bytes);
1845         }
1846         kunmap_atomic(kaddr, KM_USER0);
1847
1848         return copied;
1849 }
1850 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1851
1852 /*
1853  * This has the same sideeffects and return value as
1854  * iov_iter_copy_from_user_atomic().
1855  * The difference is that it attempts to resolve faults.
1856  * Page must not be locked.
1857  */
1858 size_t iov_iter_copy_from_user(struct page *page,
1859                 struct iov_iter *i, unsigned long offset, size_t bytes)
1860 {
1861         char *kaddr;
1862         size_t copied;
1863
1864         kaddr = kmap(page);
1865         if (likely(i->nr_segs == 1)) {
1866                 int left;
1867                 char __user *buf = i->iov->iov_base + i->iov_offset;
1868                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1869                 copied = bytes - left;
1870         } else {
1871                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1872                                                 i->iov, i->iov_offset, bytes);
1873         }
1874         kunmap(page);
1875         return copied;
1876 }
1877 EXPORT_SYMBOL(iov_iter_copy_from_user);
1878
1879 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1880 {
1881         BUG_ON(i->count < bytes);
1882
1883         if (likely(i->nr_segs == 1)) {
1884                 i->iov_offset += bytes;
1885                 i->count -= bytes;
1886         } else {
1887                 const struct iovec *iov = i->iov;
1888                 size_t base = i->iov_offset;
1889
1890                 /*
1891                  * The !iov->iov_len check ensures we skip over unlikely
1892                  * zero-length segments (without overruning the iovec).
1893                  */
1894                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1895                         int copy;
1896
1897                         copy = min(bytes, iov->iov_len - base);
1898                         BUG_ON(!i->count || i->count < copy);
1899                         i->count -= copy;
1900                         bytes -= copy;
1901                         base += copy;
1902                         if (iov->iov_len == base) {
1903                                 iov++;
1904                                 base = 0;
1905                         }
1906                 }
1907                 i->iov = iov;
1908                 i->iov_offset = base;
1909         }
1910 }
1911 EXPORT_SYMBOL(iov_iter_advance);
1912
1913 /*
1914  * Fault in the first iovec of the given iov_iter, to a maximum length
1915  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1916  * accessed (ie. because it is an invalid address).
1917  *
1918  * writev-intensive code may want this to prefault several iovecs -- that
1919  * would be possible (callers must not rely on the fact that _only_ the
1920  * first iovec will be faulted with the current implementation).
1921  */
1922 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1923 {
1924         char __user *buf = i->iov->iov_base + i->iov_offset;
1925         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1926         return fault_in_pages_readable(buf, bytes);
1927 }
1928 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1929
1930 /*
1931  * Return the count of just the current iov_iter segment.
1932  */
1933 size_t iov_iter_single_seg_count(struct iov_iter *i)
1934 {
1935         const struct iovec *iov = i->iov;
1936         if (i->nr_segs == 1)
1937                 return i->count;
1938         else
1939                 return min(i->count, iov->iov_len - i->iov_offset);
1940 }
1941 EXPORT_SYMBOL(iov_iter_single_seg_count);
1942
1943 /*
1944  * Performs necessary checks before doing a write
1945  *
1946  * Can adjust writing position or amount of bytes to write.
1947  * Returns appropriate error code that caller should return or
1948  * zero in case that write should be allowed.
1949  */
1950 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1951 {
1952         struct inode *inode = file->f_mapping->host;
1953         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1954
1955         if (unlikely(*pos < 0))
1956                 return -EINVAL;
1957
1958         if (!isblk) {
1959                 /* FIXME: this is for backwards compatibility with 2.4 */
1960                 if (file->f_flags & O_APPEND)
1961                         *pos = i_size_read(inode);
1962
1963                 if (limit != RLIM_INFINITY) {
1964                         if (*pos >= limit) {
1965                                 send_sig(SIGXFSZ, current, 0);
1966                                 return -EFBIG;
1967                         }
1968                         if (*count > limit - (typeof(limit))*pos) {
1969                                 *count = limit - (typeof(limit))*pos;
1970                         }
1971                 }
1972         }
1973
1974         /*
1975          * LFS rule
1976          */
1977         if (unlikely(*pos + *count > MAX_NON_LFS &&
1978                                 !(file->f_flags & O_LARGEFILE))) {
1979                 if (*pos >= MAX_NON_LFS) {
1980                         return -EFBIG;
1981                 }
1982                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1983                         *count = MAX_NON_LFS - (unsigned long)*pos;
1984                 }
1985         }
1986
1987         /*
1988          * Are we about to exceed the fs block limit ?
1989          *
1990          * If we have written data it becomes a short write.  If we have
1991          * exceeded without writing data we send a signal and return EFBIG.
1992          * Linus frestrict idea will clean these up nicely..
1993          */
1994         if (likely(!isblk)) {
1995                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1996                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1997                                 return -EFBIG;
1998                         }
1999                         /* zero-length writes at ->s_maxbytes are OK */
2000                 }
2001
2002                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2003                         *count = inode->i_sb->s_maxbytes - *pos;
2004         } else {
2005 #ifdef CONFIG_BLOCK
2006                 loff_t isize;
2007                 if (bdev_read_only(I_BDEV(inode)))
2008                         return -EPERM;
2009                 isize = i_size_read(inode);
2010                 if (*pos >= isize) {
2011                         if (*count || *pos > isize)
2012                                 return -ENOSPC;
2013                 }
2014
2015                 if (*pos + *count > isize)
2016                         *count = isize - *pos;
2017 #else
2018                 return -EPERM;
2019 #endif
2020         }
2021         return 0;
2022 }
2023 EXPORT_SYMBOL(generic_write_checks);
2024
2025 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2026                                 loff_t pos, unsigned len, unsigned flags,
2027                                 struct page **pagep, void **fsdata)
2028 {
2029         const struct address_space_operations *aops = mapping->a_ops;
2030
2031         return aops->write_begin(file, mapping, pos, len, flags,
2032                                                         pagep, fsdata);
2033 }
2034 EXPORT_SYMBOL(pagecache_write_begin);
2035
2036 int pagecache_write_end(struct file *file, struct address_space *mapping,
2037                                 loff_t pos, unsigned len, unsigned copied,
2038                                 struct page *page, void *fsdata)
2039 {
2040         const struct address_space_operations *aops = mapping->a_ops;
2041
2042         mark_page_accessed(page);
2043         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2044 }
2045 EXPORT_SYMBOL(pagecache_write_end);
2046
2047 ssize_t
2048 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2049                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2050                 size_t count, size_t ocount)
2051 {
2052         struct file     *file = iocb->ki_filp;
2053         struct address_space *mapping = file->f_mapping;
2054         struct inode    *inode = mapping->host;
2055         ssize_t         written;
2056         size_t          write_len;
2057         pgoff_t         end;
2058
2059         if (count != ocount)
2060                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2061
2062         /*
2063          * Unmap all mmappings of the file up-front.
2064          *
2065          * This will cause any pte dirty bits to be propagated into the
2066          * pageframes for the subsequent filemap_write_and_wait().
2067          */
2068         write_len = iov_length(iov, *nr_segs);
2069         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2070         if (mapping_mapped(mapping))
2071                 unmap_mapping_range(mapping, pos, write_len, 0);
2072
2073         written = filemap_write_and_wait(mapping);
2074         if (written)
2075                 goto out;
2076
2077         /*
2078          * After a write we want buffered reads to be sure to go to disk to get
2079          * the new data.  We invalidate clean cached page from the region we're
2080          * about to write.  We do this *before* the write so that we can return
2081          * without clobbering -EIOCBQUEUED from ->direct_IO().
2082          */
2083         if (mapping->nrpages) {
2084                 written = invalidate_inode_pages2_range(mapping,
2085                                         pos >> PAGE_CACHE_SHIFT, end);
2086                 /*
2087                  * If a page can not be invalidated, return 0 to fall back
2088                  * to buffered write.
2089                  */
2090                 if (written) {
2091                         if (written == -EBUSY)
2092                                 return 0;
2093                         goto out;
2094                 }
2095         }
2096
2097         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2098
2099         /*
2100          * Finally, try again to invalidate clean pages which might have been
2101          * cached by non-direct readahead, or faulted in by get_user_pages()
2102          * if the source of the write was an mmap'ed region of the file
2103          * we're writing.  Either one is a pretty crazy thing to do,
2104          * so we don't support it 100%.  If this invalidation
2105          * fails, tough, the write still worked...
2106          */
2107         if (mapping->nrpages) {
2108                 invalidate_inode_pages2_range(mapping,
2109                                               pos >> PAGE_CACHE_SHIFT, end);
2110         }
2111
2112         if (written > 0) {
2113                 loff_t end = pos + written;
2114                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2115                         i_size_write(inode,  end);
2116                         mark_inode_dirty(inode);
2117                 }
2118                 *ppos = end;
2119         }
2120
2121         /*
2122          * Sync the fs metadata but not the minor inode changes and
2123          * of course not the data as we did direct DMA for the IO.
2124          * i_mutex is held, which protects generic_osync_inode() from
2125          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2126          */
2127 out:
2128         if ((written >= 0 || written == -EIOCBQUEUED) &&
2129             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2130                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2131                 if (err < 0)
2132                         written = err;
2133         }
2134         return written;
2135 }
2136 EXPORT_SYMBOL(generic_file_direct_write);
2137
2138 /*
2139  * Find or create a page at the given pagecache position. Return the locked
2140  * page. This function is specifically for buffered writes.
2141  */
2142 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2143                                         pgoff_t index, unsigned flags)
2144 {
2145         int status;
2146         struct page *page;
2147         gfp_t gfp_notmask = 0;
2148         if (flags & AOP_FLAG_NOFS)
2149                 gfp_notmask = __GFP_FS;
2150 repeat:
2151         page = find_lock_page(mapping, index);
2152         if (likely(page))
2153                 return page;
2154
2155         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2156         if (!page)
2157                 return NULL;
2158         status = add_to_page_cache_lru(page, mapping, index,
2159                                                 GFP_KERNEL & ~gfp_notmask);
2160         if (unlikely(status)) {
2161                 page_cache_release(page);
2162                 if (status == -EEXIST)
2163                         goto repeat;
2164                 return NULL;
2165         }
2166         return page;
2167 }
2168 EXPORT_SYMBOL(grab_cache_page_write_begin);
2169
2170 static ssize_t generic_perform_write(struct file *file,
2171                                 struct iov_iter *i, loff_t pos)
2172 {
2173         struct address_space *mapping = file->f_mapping;
2174         const struct address_space_operations *a_ops = mapping->a_ops;
2175         long status = 0;
2176         ssize_t written = 0;
2177         unsigned int flags = 0;
2178
2179         /*
2180          * Copies from kernel address space cannot fail (NFSD is a big user).
2181          */
2182         if (segment_eq(get_fs(), KERNEL_DS))
2183                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2184
2185         do {
2186                 struct page *page;
2187                 pgoff_t index;          /* Pagecache index for current page */
2188                 unsigned long offset;   /* Offset into pagecache page */
2189                 unsigned long bytes;    /* Bytes to write to page */
2190                 size_t copied;          /* Bytes copied from user */
2191                 void *fsdata;
2192
2193                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2194                 index = pos >> PAGE_CACHE_SHIFT;
2195                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2196                                                 iov_iter_count(i));
2197
2198 again:
2199
2200                 /*
2201                  * Bring in the user page that we will copy from _first_.
2202                  * Otherwise there's a nasty deadlock on copying from the
2203                  * same page as we're writing to, without it being marked
2204                  * up-to-date.
2205                  *
2206                  * Not only is this an optimisation, but it is also required
2207                  * to check that the address is actually valid, when atomic
2208                  * usercopies are used, below.
2209                  */
2210                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2211                         status = -EFAULT;
2212                         break;
2213                 }
2214
2215                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2216                                                 &page, &fsdata);
2217                 if (unlikely(status))
2218                         break;
2219
2220                 pagefault_disable();
2221                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2222                 pagefault_enable();
2223                 flush_dcache_page(page);
2224
2225                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2226                                                 page, fsdata);
2227                 if (unlikely(status < 0))
2228                         break;
2229                 copied = status;
2230
2231                 cond_resched();
2232
2233                 iov_iter_advance(i, copied);
2234                 if (unlikely(copied == 0)) {
2235                         /*
2236                          * If we were unable to copy any data at all, we must
2237                          * fall back to a single segment length write.
2238                          *
2239                          * If we didn't fallback here, we could livelock
2240                          * because not all segments in the iov can be copied at
2241                          * once without a pagefault.
2242                          */
2243                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2244                                                 iov_iter_single_seg_count(i));
2245                         goto again;
2246                 }
2247                 pos += copied;
2248                 written += copied;
2249
2250                 balance_dirty_pages_ratelimited(mapping);
2251
2252         } while (iov_iter_count(i));
2253
2254         return written ? written : status;
2255 }
2256
2257 ssize_t
2258 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2259                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2260                 size_t count, ssize_t written)
2261 {
2262         struct file *file = iocb->ki_filp;
2263         struct address_space *mapping = file->f_mapping;
2264         const struct address_space_operations *a_ops = mapping->a_ops;
2265         struct inode *inode = mapping->host;
2266         ssize_t status;
2267         struct iov_iter i;
2268
2269         iov_iter_init(&i, iov, nr_segs, count, written);
2270         status = generic_perform_write(file, &i, pos);
2271
2272         if (likely(status >= 0)) {
2273                 written += status;
2274                 *ppos = pos + status;
2275
2276                 /*
2277                  * For now, when the user asks for O_SYNC, we'll actually give
2278                  * O_DSYNC
2279                  */
2280                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2281                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2282                                 status = generic_osync_inode(inode, mapping,
2283                                                 OSYNC_METADATA|OSYNC_DATA);
2284                 }
2285         }
2286         
2287         /*
2288          * If we get here for O_DIRECT writes then we must have fallen through
2289          * to buffered writes (block instantiation inside i_size).  So we sync
2290          * the file data here, to try to honour O_DIRECT expectations.
2291          */
2292         if (unlikely(file->f_flags & O_DIRECT) && written)
2293                 status = filemap_write_and_wait(mapping);
2294
2295         return written ? written : status;
2296 }
2297 EXPORT_SYMBOL(generic_file_buffered_write);
2298
2299 static ssize_t
2300 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2301                                 unsigned long nr_segs, loff_t *ppos)
2302 {
2303         struct file *file = iocb->ki_filp;
2304         struct address_space * mapping = file->f_mapping;
2305         size_t ocount;          /* original count */
2306         size_t count;           /* after file limit checks */
2307         struct inode    *inode = mapping->host;
2308         loff_t          pos;
2309         ssize_t         written;
2310         ssize_t         err;
2311
2312         ocount = 0;
2313         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2314         if (err)
2315                 return err;
2316
2317         count = ocount;
2318         pos = *ppos;
2319
2320         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2321
2322         /* We can write back this queue in page reclaim */
2323         current->backing_dev_info = mapping->backing_dev_info;
2324         written = 0;
2325
2326         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2327         if (err)
2328                 goto out;
2329
2330         if (count == 0)
2331                 goto out;
2332
2333         err = file_remove_suid(file);
2334         if (err)
2335                 goto out;
2336
2337         file_update_time(file);
2338
2339         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2340         if (unlikely(file->f_flags & O_DIRECT)) {
2341                 loff_t endbyte;
2342                 ssize_t written_buffered;
2343
2344                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2345                                                         ppos, count, ocount);
2346                 if (written < 0 || written == count)
2347                         goto out;
2348                 /*
2349                  * direct-io write to a hole: fall through to buffered I/O
2350                  * for completing the rest of the request.
2351                  */
2352                 pos += written;
2353                 count -= written;
2354                 written_buffered = generic_file_buffered_write(iocb, iov,
2355                                                 nr_segs, pos, ppos, count,
2356                                                 written);
2357                 /*
2358                  * If generic_file_buffered_write() retuned a synchronous error
2359                  * then we want to return the number of bytes which were
2360                  * direct-written, or the error code if that was zero.  Note
2361                  * that this differs from normal direct-io semantics, which
2362                  * will return -EFOO even if some bytes were written.
2363                  */
2364                 if (written_buffered < 0) {
2365                         err = written_buffered;
2366                         goto out;
2367                 }
2368
2369                 /*
2370                  * We need to ensure that the page cache pages are written to
2371                  * disk and invalidated to preserve the expected O_DIRECT
2372                  * semantics.
2373                  */
2374                 endbyte = pos + written_buffered - written - 1;
2375                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2376                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2377                                             SYNC_FILE_RANGE_WRITE|
2378                                             SYNC_FILE_RANGE_WAIT_AFTER);
2379                 if (err == 0) {
2380                         written = written_buffered;
2381                         invalidate_mapping_pages(mapping,
2382                                                  pos >> PAGE_CACHE_SHIFT,
2383                                                  endbyte >> PAGE_CACHE_SHIFT);
2384                 } else {
2385                         /*
2386                          * We don't know how much we wrote, so just return
2387                          * the number of bytes which were direct-written
2388                          */
2389                 }
2390         } else {
2391                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2392                                 pos, ppos, count, written);
2393         }
2394 out:
2395         current->backing_dev_info = NULL;
2396         return written ? written : err;
2397 }
2398
2399 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2400                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2401 {
2402         struct file *file = iocb->ki_filp;
2403         struct address_space *mapping = file->f_mapping;
2404         struct inode *inode = mapping->host;
2405         ssize_t ret;
2406
2407         BUG_ON(iocb->ki_pos != pos);
2408
2409         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2410                         &iocb->ki_pos);
2411
2412         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2413                 ssize_t err;
2414
2415                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2416                 if (err < 0)
2417                         ret = err;
2418         }
2419         return ret;
2420 }
2421 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2422
2423 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2424                 unsigned long nr_segs, loff_t pos)
2425 {
2426         struct file *file = iocb->ki_filp;
2427         struct address_space *mapping = file->f_mapping;
2428         struct inode *inode = mapping->host;
2429         ssize_t ret;
2430
2431         BUG_ON(iocb->ki_pos != pos);
2432
2433         mutex_lock(&inode->i_mutex);
2434         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2435                         &iocb->ki_pos);
2436         mutex_unlock(&inode->i_mutex);
2437
2438         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2439                 ssize_t err;
2440
2441                 err = sync_page_range(inode, mapping, pos, ret);
2442                 if (err < 0)
2443                         ret = err;
2444         }
2445         return ret;
2446 }
2447 EXPORT_SYMBOL(generic_file_aio_write);
2448
2449 /**
2450  * try_to_release_page() - release old fs-specific metadata on a page
2451  *
2452  * @page: the page which the kernel is trying to free
2453  * @gfp_mask: memory allocation flags (and I/O mode)
2454  *
2455  * The address_space is to try to release any data against the page
2456  * (presumably at page->private).  If the release was successful, return `1'.
2457  * Otherwise return zero.
2458  *
2459  * The @gfp_mask argument specifies whether I/O may be performed to release
2460  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2461  *
2462  */
2463 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2464 {
2465         struct address_space * const mapping = page->mapping;
2466
2467         BUG_ON(!PageLocked(page));
2468         if (PageWriteback(page))
2469                 return 0;
2470
2471         if (mapping && mapping->a_ops->releasepage)
2472                 return mapping->a_ops->releasepage(page, gfp_mask);
2473         return try_to_free_buffers(page);
2474 }
2475
2476 EXPORT_SYMBOL(try_to_release_page);