mm: pagecache gfp flags fix
[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                 /*
745                  * We want a regular kernel memory (not highmem or DMA etc)
746                  * allocation for the radix tree nodes, but we need to honour
747                  * the context-specific requirements the caller has asked for.
748                  * GFP_RECLAIM_MASK collects those requirements.
749                  */
750                 err = add_to_page_cache_lru(page, mapping, index,
751                         (gfp_mask & GFP_RECLAIM_MASK));
752                 if (unlikely(err)) {
753                         page_cache_release(page);
754                         page = NULL;
755                         if (err == -EEXIST)
756                                 goto repeat;
757                 }
758         }
759         return page;
760 }
761 EXPORT_SYMBOL(find_or_create_page);
762
763 /**
764  * find_get_pages - gang pagecache lookup
765  * @mapping:    The address_space to search
766  * @start:      The starting page index
767  * @nr_pages:   The maximum number of pages
768  * @pages:      Where the resulting pages are placed
769  *
770  * find_get_pages() will search for and return a group of up to
771  * @nr_pages pages in the mapping.  The pages are placed at @pages.
772  * find_get_pages() takes a reference against the returned pages.
773  *
774  * The search returns a group of mapping-contiguous pages with ascending
775  * indexes.  There may be holes in the indices due to not-present pages.
776  *
777  * find_get_pages() returns the number of pages which were found.
778  */
779 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
780                             unsigned int nr_pages, struct page **pages)
781 {
782         unsigned int i;
783         unsigned int ret;
784         unsigned int nr_found;
785
786         rcu_read_lock();
787 restart:
788         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
789                                 (void ***)pages, start, nr_pages);
790         ret = 0;
791         for (i = 0; i < nr_found; i++) {
792                 struct page *page;
793 repeat:
794                 page = radix_tree_deref_slot((void **)pages[i]);
795                 if (unlikely(!page))
796                         continue;
797                 /*
798                  * this can only trigger if nr_found == 1, making livelock
799                  * a non issue.
800                  */
801                 if (unlikely(page == RADIX_TREE_RETRY))
802                         goto restart;
803
804                 if (!page_cache_get_speculative(page))
805                         goto repeat;
806
807                 /* Has the page moved? */
808                 if (unlikely(page != *((void **)pages[i]))) {
809                         page_cache_release(page);
810                         goto repeat;
811                 }
812
813                 pages[ret] = page;
814                 ret++;
815         }
816         rcu_read_unlock();
817         return ret;
818 }
819
820 /**
821  * find_get_pages_contig - gang contiguous pagecache lookup
822  * @mapping:    The address_space to search
823  * @index:      The starting page index
824  * @nr_pages:   The maximum number of pages
825  * @pages:      Where the resulting pages are placed
826  *
827  * find_get_pages_contig() works exactly like find_get_pages(), except
828  * that the returned number of pages are guaranteed to be contiguous.
829  *
830  * find_get_pages_contig() returns the number of pages which were found.
831  */
832 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
833                                unsigned int nr_pages, struct page **pages)
834 {
835         unsigned int i;
836         unsigned int ret;
837         unsigned int nr_found;
838
839         rcu_read_lock();
840 restart:
841         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
842                                 (void ***)pages, index, nr_pages);
843         ret = 0;
844         for (i = 0; i < nr_found; i++) {
845                 struct page *page;
846 repeat:
847                 page = radix_tree_deref_slot((void **)pages[i]);
848                 if (unlikely(!page))
849                         continue;
850                 /*
851                  * this can only trigger if nr_found == 1, making livelock
852                  * a non issue.
853                  */
854                 if (unlikely(page == RADIX_TREE_RETRY))
855                         goto restart;
856
857                 if (page->mapping == NULL || page->index != index)
858                         break;
859
860                 if (!page_cache_get_speculative(page))
861                         goto repeat;
862
863                 /* Has the page moved? */
864                 if (unlikely(page != *((void **)pages[i]))) {
865                         page_cache_release(page);
866                         goto repeat;
867                 }
868
869                 pages[ret] = page;
870                 ret++;
871                 index++;
872         }
873         rcu_read_unlock();
874         return ret;
875 }
876 EXPORT_SYMBOL(find_get_pages_contig);
877
878 /**
879  * find_get_pages_tag - find and return pages that match @tag
880  * @mapping:    the address_space to search
881  * @index:      the starting page index
882  * @tag:        the tag index
883  * @nr_pages:   the maximum number of pages
884  * @pages:      where the resulting pages are placed
885  *
886  * Like find_get_pages, except we only return pages which are tagged with
887  * @tag.   We update @index to index the next page for the traversal.
888  */
889 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
890                         int tag, unsigned int nr_pages, struct page **pages)
891 {
892         unsigned int i;
893         unsigned int ret;
894         unsigned int nr_found;
895
896         rcu_read_lock();
897 restart:
898         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
899                                 (void ***)pages, *index, nr_pages, tag);
900         ret = 0;
901         for (i = 0; i < nr_found; i++) {
902                 struct page *page;
903 repeat:
904                 page = radix_tree_deref_slot((void **)pages[i]);
905                 if (unlikely(!page))
906                         continue;
907                 /*
908                  * this can only trigger if nr_found == 1, making livelock
909                  * a non issue.
910                  */
911                 if (unlikely(page == RADIX_TREE_RETRY))
912                         goto restart;
913
914                 if (!page_cache_get_speculative(page))
915                         goto repeat;
916
917                 /* Has the page moved? */
918                 if (unlikely(page != *((void **)pages[i]))) {
919                         page_cache_release(page);
920                         goto repeat;
921                 }
922
923                 pages[ret] = page;
924                 ret++;
925         }
926         rcu_read_unlock();
927
928         if (ret)
929                 *index = pages[ret - 1]->index + 1;
930
931         return ret;
932 }
933 EXPORT_SYMBOL(find_get_pages_tag);
934
935 /**
936  * grab_cache_page_nowait - returns locked page at given index in given cache
937  * @mapping: target address_space
938  * @index: the page index
939  *
940  * Same as grab_cache_page(), but do not wait if the page is unavailable.
941  * This is intended for speculative data generators, where the data can
942  * be regenerated if the page couldn't be grabbed.  This routine should
943  * be safe to call while holding the lock for another page.
944  *
945  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
946  * and deadlock against the caller's locked page.
947  */
948 struct page *
949 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
950 {
951         struct page *page = find_get_page(mapping, index);
952
953         if (page) {
954                 if (trylock_page(page))
955                         return page;
956                 page_cache_release(page);
957                 return NULL;
958         }
959         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
960         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
961                 page_cache_release(page);
962                 page = NULL;
963         }
964         return page;
965 }
966 EXPORT_SYMBOL(grab_cache_page_nowait);
967
968 /*
969  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
970  * a _large_ part of the i/o request. Imagine the worst scenario:
971  *
972  *      ---R__________________________________________B__________
973  *         ^ reading here                             ^ bad block(assume 4k)
974  *
975  * read(R) => miss => readahead(R...B) => media error => frustrating retries
976  * => failing the whole request => read(R) => read(R+1) =>
977  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
978  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
979  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
980  *
981  * It is going insane. Fix it by quickly scaling down the readahead size.
982  */
983 static void shrink_readahead_size_eio(struct file *filp,
984                                         struct file_ra_state *ra)
985 {
986         if (!ra->ra_pages)
987                 return;
988
989         ra->ra_pages /= 4;
990 }
991
992 /**
993  * do_generic_file_read - generic file read routine
994  * @filp:       the file to read
995  * @ppos:       current file position
996  * @desc:       read_descriptor
997  * @actor:      read method
998  *
999  * This is a generic file read routine, and uses the
1000  * mapping->a_ops->readpage() function for the actual low-level stuff.
1001  *
1002  * This is really ugly. But the goto's actually try to clarify some
1003  * of the logic when it comes to error handling etc.
1004  */
1005 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1006                 read_descriptor_t *desc, read_actor_t actor)
1007 {
1008         struct address_space *mapping = filp->f_mapping;
1009         struct inode *inode = mapping->host;
1010         struct file_ra_state *ra = &filp->f_ra;
1011         pgoff_t index;
1012         pgoff_t last_index;
1013         pgoff_t prev_index;
1014         unsigned long offset;      /* offset into pagecache page */
1015         unsigned int prev_offset;
1016         int error;
1017
1018         index = *ppos >> PAGE_CACHE_SHIFT;
1019         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1020         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1021         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1022         offset = *ppos & ~PAGE_CACHE_MASK;
1023
1024         for (;;) {
1025                 struct page *page;
1026                 pgoff_t end_index;
1027                 loff_t isize;
1028                 unsigned long nr, ret;
1029
1030                 cond_resched();
1031 find_page:
1032                 page = find_get_page(mapping, index);
1033                 if (!page) {
1034                         page_cache_sync_readahead(mapping,
1035                                         ra, filp,
1036                                         index, last_index - index);
1037                         page = find_get_page(mapping, index);
1038                         if (unlikely(page == NULL))
1039                                 goto no_cached_page;
1040                 }
1041                 if (PageReadahead(page)) {
1042                         page_cache_async_readahead(mapping,
1043                                         ra, filp, page,
1044                                         index, last_index - index);
1045                 }
1046                 if (!PageUptodate(page)) {
1047                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1048                                         !mapping->a_ops->is_partially_uptodate)
1049                                 goto page_not_up_to_date;
1050                         if (!trylock_page(page))
1051                                 goto page_not_up_to_date;
1052                         if (!mapping->a_ops->is_partially_uptodate(page,
1053                                                                 desc, offset))
1054                                 goto page_not_up_to_date_locked;
1055                         unlock_page(page);
1056                 }
1057 page_ok:
1058                 /*
1059                  * i_size must be checked after we know the page is Uptodate.
1060                  *
1061                  * Checking i_size after the check allows us to calculate
1062                  * the correct value for "nr", which means the zero-filled
1063                  * part of the page is not copied back to userspace (unless
1064                  * another truncate extends the file - this is desired though).
1065                  */
1066
1067                 isize = i_size_read(inode);
1068                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1069                 if (unlikely(!isize || index > end_index)) {
1070                         page_cache_release(page);
1071                         goto out;
1072                 }
1073
1074                 /* nr is the maximum number of bytes to copy from this page */
1075                 nr = PAGE_CACHE_SIZE;
1076                 if (index == end_index) {
1077                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1078                         if (nr <= offset) {
1079                                 page_cache_release(page);
1080                                 goto out;
1081                         }
1082                 }
1083                 nr = nr - offset;
1084
1085                 /* If users can be writing to this page using arbitrary
1086                  * virtual addresses, take care about potential aliasing
1087                  * before reading the page on the kernel side.
1088                  */
1089                 if (mapping_writably_mapped(mapping))
1090                         flush_dcache_page(page);
1091
1092                 /*
1093                  * When a sequential read accesses a page several times,
1094                  * only mark it as accessed the first time.
1095                  */
1096                 if (prev_index != index || offset != prev_offset)
1097                         mark_page_accessed(page);
1098                 prev_index = index;
1099
1100                 /*
1101                  * Ok, we have the page, and it's up-to-date, so
1102                  * now we can copy it to user space...
1103                  *
1104                  * The actor routine returns how many bytes were actually used..
1105                  * NOTE! This may not be the same as how much of a user buffer
1106                  * we filled up (we may be padding etc), so we can only update
1107                  * "pos" here (the actor routine has to update the user buffer
1108                  * pointers and the remaining count).
1109                  */
1110                 ret = actor(desc, page, offset, nr);
1111                 offset += ret;
1112                 index += offset >> PAGE_CACHE_SHIFT;
1113                 offset &= ~PAGE_CACHE_MASK;
1114                 prev_offset = offset;
1115
1116                 page_cache_release(page);
1117                 if (ret == nr && desc->count)
1118                         continue;
1119                 goto out;
1120
1121 page_not_up_to_date:
1122                 /* Get exclusive access to the page ... */
1123                 error = lock_page_killable(page);
1124                 if (unlikely(error))
1125                         goto readpage_error;
1126
1127 page_not_up_to_date_locked:
1128                 /* Did it get truncated before we got the lock? */
1129                 if (!page->mapping) {
1130                         unlock_page(page);
1131                         page_cache_release(page);
1132                         continue;
1133                 }
1134
1135                 /* Did somebody else fill it already? */
1136                 if (PageUptodate(page)) {
1137                         unlock_page(page);
1138                         goto page_ok;
1139                 }
1140
1141 readpage:
1142                 /* Start the actual read. The read will unlock the page. */
1143                 error = mapping->a_ops->readpage(filp, page);
1144
1145                 if (unlikely(error)) {
1146                         if (error == AOP_TRUNCATED_PAGE) {
1147                                 page_cache_release(page);
1148                                 goto find_page;
1149                         }
1150                         goto readpage_error;
1151                 }
1152
1153                 if (!PageUptodate(page)) {
1154                         error = lock_page_killable(page);
1155                         if (unlikely(error))
1156                                 goto readpage_error;
1157                         if (!PageUptodate(page)) {
1158                                 if (page->mapping == NULL) {
1159                                         /*
1160                                          * invalidate_inode_pages got it
1161                                          */
1162                                         unlock_page(page);
1163                                         page_cache_release(page);
1164                                         goto find_page;
1165                                 }
1166                                 unlock_page(page);
1167                                 shrink_readahead_size_eio(filp, ra);
1168                                 error = -EIO;
1169                                 goto readpage_error;
1170                         }
1171                         unlock_page(page);
1172                 }
1173
1174                 goto page_ok;
1175
1176 readpage_error:
1177                 /* UHHUH! A synchronous read error occurred. Report it */
1178                 desc->error = error;
1179                 page_cache_release(page);
1180                 goto out;
1181
1182 no_cached_page:
1183                 /*
1184                  * Ok, it wasn't cached, so we need to create a new
1185                  * page..
1186                  */
1187                 page = page_cache_alloc_cold(mapping);
1188                 if (!page) {
1189                         desc->error = -ENOMEM;
1190                         goto out;
1191                 }
1192                 error = add_to_page_cache_lru(page, mapping,
1193                                                 index, GFP_KERNEL);
1194                 if (error) {
1195                         page_cache_release(page);
1196                         if (error == -EEXIST)
1197                                 goto find_page;
1198                         desc->error = error;
1199                         goto out;
1200                 }
1201                 goto readpage;
1202         }
1203
1204 out:
1205         ra->prev_pos = prev_index;
1206         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1207         ra->prev_pos |= prev_offset;
1208
1209         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1210         file_accessed(filp);
1211 }
1212
1213 int file_read_actor(read_descriptor_t *desc, struct page *page,
1214                         unsigned long offset, unsigned long size)
1215 {
1216         char *kaddr;
1217         unsigned long left, count = desc->count;
1218
1219         if (size > count)
1220                 size = count;
1221
1222         /*
1223          * Faults on the destination of a read are common, so do it before
1224          * taking the kmap.
1225          */
1226         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1227                 kaddr = kmap_atomic(page, KM_USER0);
1228                 left = __copy_to_user_inatomic(desc->arg.buf,
1229                                                 kaddr + offset, size);
1230                 kunmap_atomic(kaddr, KM_USER0);
1231                 if (left == 0)
1232                         goto success;
1233         }
1234
1235         /* Do it the slow way */
1236         kaddr = kmap(page);
1237         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1238         kunmap(page);
1239
1240         if (left) {
1241                 size -= left;
1242                 desc->error = -EFAULT;
1243         }
1244 success:
1245         desc->count = count - size;
1246         desc->written += size;
1247         desc->arg.buf += size;
1248         return size;
1249 }
1250
1251 /*
1252  * Performs necessary checks before doing a write
1253  * @iov:        io vector request
1254  * @nr_segs:    number of segments in the iovec
1255  * @count:      number of bytes to write
1256  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1257  *
1258  * Adjust number of segments and amount of bytes to write (nr_segs should be
1259  * properly initialized first). Returns appropriate error code that caller
1260  * should return or zero in case that write should be allowed.
1261  */
1262 int generic_segment_checks(const struct iovec *iov,
1263                         unsigned long *nr_segs, size_t *count, int access_flags)
1264 {
1265         unsigned long   seg;
1266         size_t cnt = 0;
1267         for (seg = 0; seg < *nr_segs; seg++) {
1268                 const struct iovec *iv = &iov[seg];
1269
1270                 /*
1271                  * If any segment has a negative length, or the cumulative
1272                  * length ever wraps negative then return -EINVAL.
1273                  */
1274                 cnt += iv->iov_len;
1275                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1276                         return -EINVAL;
1277                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1278                         continue;
1279                 if (seg == 0)
1280                         return -EFAULT;
1281                 *nr_segs = seg;
1282                 cnt -= iv->iov_len;     /* This segment is no good */
1283                 break;
1284         }
1285         *count = cnt;
1286         return 0;
1287 }
1288 EXPORT_SYMBOL(generic_segment_checks);
1289
1290 /**
1291  * generic_file_aio_read - generic filesystem read routine
1292  * @iocb:       kernel I/O control block
1293  * @iov:        io vector request
1294  * @nr_segs:    number of segments in the iovec
1295  * @pos:        current file position
1296  *
1297  * This is the "read()" routine for all filesystems
1298  * that can use the page cache directly.
1299  */
1300 ssize_t
1301 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1302                 unsigned long nr_segs, loff_t pos)
1303 {
1304         struct file *filp = iocb->ki_filp;
1305         ssize_t retval;
1306         unsigned long seg;
1307         size_t count;
1308         loff_t *ppos = &iocb->ki_pos;
1309
1310         count = 0;
1311         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1312         if (retval)
1313                 return retval;
1314
1315         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1316         if (filp->f_flags & O_DIRECT) {
1317                 loff_t size;
1318                 struct address_space *mapping;
1319                 struct inode *inode;
1320
1321                 mapping = filp->f_mapping;
1322                 inode = mapping->host;
1323                 if (!count)
1324                         goto out; /* skip atime */
1325                 size = i_size_read(inode);
1326                 if (pos < size) {
1327                         retval = filemap_write_and_wait_range(mapping, pos,
1328                                         pos + iov_length(iov, nr_segs) - 1);
1329                         if (!retval) {
1330                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1331                                                         iov, pos, nr_segs);
1332                         }
1333                         if (retval > 0)
1334                                 *ppos = pos + retval;
1335                         if (retval) {
1336                                 file_accessed(filp);
1337                                 goto out;
1338                         }
1339                 }
1340         }
1341
1342         for (seg = 0; seg < nr_segs; seg++) {
1343                 read_descriptor_t desc;
1344
1345                 desc.written = 0;
1346                 desc.arg.buf = iov[seg].iov_base;
1347                 desc.count = iov[seg].iov_len;
1348                 if (desc.count == 0)
1349                         continue;
1350                 desc.error = 0;
1351                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1352                 retval += desc.written;
1353                 if (desc.error) {
1354                         retval = retval ?: desc.error;
1355                         break;
1356                 }
1357                 if (desc.count > 0)
1358                         break;
1359         }
1360 out:
1361         return retval;
1362 }
1363 EXPORT_SYMBOL(generic_file_aio_read);
1364
1365 static ssize_t
1366 do_readahead(struct address_space *mapping, struct file *filp,
1367              pgoff_t index, unsigned long nr)
1368 {
1369         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1370                 return -EINVAL;
1371
1372         force_page_cache_readahead(mapping, filp, index,
1373                                         max_sane_readahead(nr));
1374         return 0;
1375 }
1376
1377 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1378 {
1379         ssize_t ret;
1380         struct file *file;
1381
1382         ret = -EBADF;
1383         file = fget(fd);
1384         if (file) {
1385                 if (file->f_mode & FMODE_READ) {
1386                         struct address_space *mapping = file->f_mapping;
1387                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1388                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1389                         unsigned long len = end - start + 1;
1390                         ret = do_readahead(mapping, file, start, len);
1391                 }
1392                 fput(file);
1393         }
1394         return ret;
1395 }
1396
1397 #ifdef CONFIG_MMU
1398 /**
1399  * page_cache_read - adds requested page to the page cache if not already there
1400  * @file:       file to read
1401  * @offset:     page index
1402  *
1403  * This adds the requested page to the page cache if it isn't already there,
1404  * and schedules an I/O to read in its contents from disk.
1405  */
1406 static int page_cache_read(struct file *file, pgoff_t offset)
1407 {
1408         struct address_space *mapping = file->f_mapping;
1409         struct page *page; 
1410         int ret;
1411
1412         do {
1413                 page = page_cache_alloc_cold(mapping);
1414                 if (!page)
1415                         return -ENOMEM;
1416
1417                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1418                 if (ret == 0)
1419                         ret = mapping->a_ops->readpage(file, page);
1420                 else if (ret == -EEXIST)
1421                         ret = 0; /* losing race to add is OK */
1422
1423                 page_cache_release(page);
1424
1425         } while (ret == AOP_TRUNCATED_PAGE);
1426                 
1427         return ret;
1428 }
1429
1430 #define MMAP_LOTSAMISS  (100)
1431
1432 /**
1433  * filemap_fault - read in file data for page fault handling
1434  * @vma:        vma in which the fault was taken
1435  * @vmf:        struct vm_fault containing details of the fault
1436  *
1437  * filemap_fault() is invoked via the vma operations vector for a
1438  * mapped memory region to read in file data during a page fault.
1439  *
1440  * The goto's are kind of ugly, but this streamlines the normal case of having
1441  * it in the page cache, and handles the special cases reasonably without
1442  * having a lot of duplicated code.
1443  */
1444 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1445 {
1446         int error;
1447         struct file *file = vma->vm_file;
1448         struct address_space *mapping = file->f_mapping;
1449         struct file_ra_state *ra = &file->f_ra;
1450         struct inode *inode = mapping->host;
1451         struct page *page;
1452         pgoff_t size;
1453         int did_readaround = 0;
1454         int ret = 0;
1455
1456         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1457         if (vmf->pgoff >= size)
1458                 return VM_FAULT_SIGBUS;
1459
1460         /* If we don't want any read-ahead, don't bother */
1461         if (VM_RandomReadHint(vma))
1462                 goto no_cached_page;
1463
1464         /*
1465          * Do we have something in the page cache already?
1466          */
1467 retry_find:
1468         page = find_lock_page(mapping, vmf->pgoff);
1469         /*
1470          * For sequential accesses, we use the generic readahead logic.
1471          */
1472         if (VM_SequentialReadHint(vma)) {
1473                 if (!page) {
1474                         page_cache_sync_readahead(mapping, ra, file,
1475                                                            vmf->pgoff, 1);
1476                         page = find_lock_page(mapping, vmf->pgoff);
1477                         if (!page)
1478                                 goto no_cached_page;
1479                 }
1480                 if (PageReadahead(page)) {
1481                         page_cache_async_readahead(mapping, ra, file, page,
1482                                                            vmf->pgoff, 1);
1483                 }
1484         }
1485
1486         if (!page) {
1487                 unsigned long ra_pages;
1488
1489                 ra->mmap_miss++;
1490
1491                 /*
1492                  * Do we miss much more than hit in this file? If so,
1493                  * stop bothering with read-ahead. It will only hurt.
1494                  */
1495                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1496                         goto no_cached_page;
1497
1498                 /*
1499                  * To keep the pgmajfault counter straight, we need to
1500                  * check did_readaround, as this is an inner loop.
1501                  */
1502                 if (!did_readaround) {
1503                         ret = VM_FAULT_MAJOR;
1504                         count_vm_event(PGMAJFAULT);
1505                 }
1506                 did_readaround = 1;
1507                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1508                 if (ra_pages) {
1509                         pgoff_t start = 0;
1510
1511                         if (vmf->pgoff > ra_pages / 2)
1512                                 start = vmf->pgoff - ra_pages / 2;
1513                         do_page_cache_readahead(mapping, file, start, ra_pages);
1514                 }
1515                 page = find_lock_page(mapping, vmf->pgoff);
1516                 if (!page)
1517                         goto no_cached_page;
1518         }
1519
1520         if (!did_readaround)
1521                 ra->mmap_miss--;
1522
1523         /*
1524          * We have a locked page in the page cache, now we need to check
1525          * that it's up-to-date. If not, it is going to be due to an error.
1526          */
1527         if (unlikely(!PageUptodate(page)))
1528                 goto page_not_uptodate;
1529
1530         /* Must recheck i_size under page lock */
1531         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1532         if (unlikely(vmf->pgoff >= size)) {
1533                 unlock_page(page);
1534                 page_cache_release(page);
1535                 return VM_FAULT_SIGBUS;
1536         }
1537
1538         /*
1539          * Found the page and have a reference on it.
1540          */
1541         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1542         vmf->page = page;
1543         return ret | VM_FAULT_LOCKED;
1544
1545 no_cached_page:
1546         /*
1547          * We're only likely to ever get here if MADV_RANDOM is in
1548          * effect.
1549          */
1550         error = page_cache_read(file, vmf->pgoff);
1551
1552         /*
1553          * The page we want has now been added to the page cache.
1554          * In the unlikely event that someone removed it in the
1555          * meantime, we'll just come back here and read it again.
1556          */
1557         if (error >= 0)
1558                 goto retry_find;
1559
1560         /*
1561          * An error return from page_cache_read can result if the
1562          * system is low on memory, or a problem occurs while trying
1563          * to schedule I/O.
1564          */
1565         if (error == -ENOMEM)
1566                 return VM_FAULT_OOM;
1567         return VM_FAULT_SIGBUS;
1568
1569 page_not_uptodate:
1570         /* IO error path */
1571         if (!did_readaround) {
1572                 ret = VM_FAULT_MAJOR;
1573                 count_vm_event(PGMAJFAULT);
1574         }
1575
1576         /*
1577          * Umm, take care of errors if the page isn't up-to-date.
1578          * Try to re-read it _once_. We do this synchronously,
1579          * because there really aren't any performance issues here
1580          * and we need to check for errors.
1581          */
1582         ClearPageError(page);
1583         error = mapping->a_ops->readpage(file, page);
1584         if (!error) {
1585                 wait_on_page_locked(page);
1586                 if (!PageUptodate(page))
1587                         error = -EIO;
1588         }
1589         page_cache_release(page);
1590
1591         if (!error || error == AOP_TRUNCATED_PAGE)
1592                 goto retry_find;
1593
1594         /* Things didn't work out. Return zero to tell the mm layer so. */
1595         shrink_readahead_size_eio(file, ra);
1596         return VM_FAULT_SIGBUS;
1597 }
1598 EXPORT_SYMBOL(filemap_fault);
1599
1600 struct vm_operations_struct generic_file_vm_ops = {
1601         .fault          = filemap_fault,
1602 };
1603
1604 /* This is used for a general mmap of a disk file */
1605
1606 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1607 {
1608         struct address_space *mapping = file->f_mapping;
1609
1610         if (!mapping->a_ops->readpage)
1611                 return -ENOEXEC;
1612         file_accessed(file);
1613         vma->vm_ops = &generic_file_vm_ops;
1614         vma->vm_flags |= VM_CAN_NONLINEAR;
1615         return 0;
1616 }
1617
1618 /*
1619  * This is for filesystems which do not implement ->writepage.
1620  */
1621 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1622 {
1623         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1624                 return -EINVAL;
1625         return generic_file_mmap(file, vma);
1626 }
1627 #else
1628 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1629 {
1630         return -ENOSYS;
1631 }
1632 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1633 {
1634         return -ENOSYS;
1635 }
1636 #endif /* CONFIG_MMU */
1637
1638 EXPORT_SYMBOL(generic_file_mmap);
1639 EXPORT_SYMBOL(generic_file_readonly_mmap);
1640
1641 static struct page *__read_cache_page(struct address_space *mapping,
1642                                 pgoff_t index,
1643                                 int (*filler)(void *,struct page*),
1644                                 void *data)
1645 {
1646         struct page *page;
1647         int err;
1648 repeat:
1649         page = find_get_page(mapping, index);
1650         if (!page) {
1651                 page = page_cache_alloc_cold(mapping);
1652                 if (!page)
1653                         return ERR_PTR(-ENOMEM);
1654                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1655                 if (unlikely(err)) {
1656                         page_cache_release(page);
1657                         if (err == -EEXIST)
1658                                 goto repeat;
1659                         /* Presumably ENOMEM for radix tree node */
1660                         return ERR_PTR(err);
1661                 }
1662                 err = filler(data, page);
1663                 if (err < 0) {
1664                         page_cache_release(page);
1665                         page = ERR_PTR(err);
1666                 }
1667         }
1668         return page;
1669 }
1670
1671 /**
1672  * read_cache_page_async - read into page cache, fill it if needed
1673  * @mapping:    the page's address_space
1674  * @index:      the page index
1675  * @filler:     function to perform the read
1676  * @data:       destination for read data
1677  *
1678  * Same as read_cache_page, but don't wait for page to become unlocked
1679  * after submitting it to the filler.
1680  *
1681  * Read into the page cache. If a page already exists, and PageUptodate() is
1682  * not set, try to fill the page but don't wait for it to become unlocked.
1683  *
1684  * If the page does not get brought uptodate, return -EIO.
1685  */
1686 struct page *read_cache_page_async(struct address_space *mapping,
1687                                 pgoff_t index,
1688                                 int (*filler)(void *,struct page*),
1689                                 void *data)
1690 {
1691         struct page *page;
1692         int err;
1693
1694 retry:
1695         page = __read_cache_page(mapping, index, filler, data);
1696         if (IS_ERR(page))
1697                 return page;
1698         if (PageUptodate(page))
1699                 goto out;
1700
1701         lock_page(page);
1702         if (!page->mapping) {
1703                 unlock_page(page);
1704                 page_cache_release(page);
1705                 goto retry;
1706         }
1707         if (PageUptodate(page)) {
1708                 unlock_page(page);
1709                 goto out;
1710         }
1711         err = filler(data, page);
1712         if (err < 0) {
1713                 page_cache_release(page);
1714                 return ERR_PTR(err);
1715         }
1716 out:
1717         mark_page_accessed(page);
1718         return page;
1719 }
1720 EXPORT_SYMBOL(read_cache_page_async);
1721
1722 /**
1723  * read_cache_page - read into page cache, fill it if needed
1724  * @mapping:    the page's address_space
1725  * @index:      the page index
1726  * @filler:     function to perform the read
1727  * @data:       destination for read data
1728  *
1729  * Read into the page cache. If a page already exists, and PageUptodate() is
1730  * not set, try to fill the page then wait for it to become unlocked.
1731  *
1732  * If the page does not get brought uptodate, return -EIO.
1733  */
1734 struct page *read_cache_page(struct address_space *mapping,
1735                                 pgoff_t index,
1736                                 int (*filler)(void *,struct page*),
1737                                 void *data)
1738 {
1739         struct page *page;
1740
1741         page = read_cache_page_async(mapping, index, filler, data);
1742         if (IS_ERR(page))
1743                 goto out;
1744         wait_on_page_locked(page);
1745         if (!PageUptodate(page)) {
1746                 page_cache_release(page);
1747                 page = ERR_PTR(-EIO);
1748         }
1749  out:
1750         return page;
1751 }
1752 EXPORT_SYMBOL(read_cache_page);
1753
1754 /*
1755  * The logic we want is
1756  *
1757  *      if suid or (sgid and xgrp)
1758  *              remove privs
1759  */
1760 int should_remove_suid(struct dentry *dentry)
1761 {
1762         mode_t mode = dentry->d_inode->i_mode;
1763         int kill = 0;
1764
1765         /* suid always must be killed */
1766         if (unlikely(mode & S_ISUID))
1767                 kill = ATTR_KILL_SUID;
1768
1769         /*
1770          * sgid without any exec bits is just a mandatory locking mark; leave
1771          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1772          */
1773         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1774                 kill |= ATTR_KILL_SGID;
1775
1776         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1777                 return kill;
1778
1779         return 0;
1780 }
1781 EXPORT_SYMBOL(should_remove_suid);
1782
1783 static int __remove_suid(struct dentry *dentry, int kill)
1784 {
1785         struct iattr newattrs;
1786
1787         newattrs.ia_valid = ATTR_FORCE | kill;
1788         return notify_change(dentry, &newattrs);
1789 }
1790
1791 int file_remove_suid(struct file *file)
1792 {
1793         struct dentry *dentry = file->f_path.dentry;
1794         int killsuid = should_remove_suid(dentry);
1795         int killpriv = security_inode_need_killpriv(dentry);
1796         int error = 0;
1797
1798         if (killpriv < 0)
1799                 return killpriv;
1800         if (killpriv)
1801                 error = security_inode_killpriv(dentry);
1802         if (!error && killsuid)
1803                 error = __remove_suid(dentry, killsuid);
1804
1805         return error;
1806 }
1807 EXPORT_SYMBOL(file_remove_suid);
1808
1809 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1810                         const struct iovec *iov, size_t base, size_t bytes)
1811 {
1812         size_t copied = 0, left = 0;
1813
1814         while (bytes) {
1815                 char __user *buf = iov->iov_base + base;
1816                 int copy = min(bytes, iov->iov_len - base);
1817
1818                 base = 0;
1819                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1820                 copied += copy;
1821                 bytes -= copy;
1822                 vaddr += copy;
1823                 iov++;
1824
1825                 if (unlikely(left))
1826                         break;
1827         }
1828         return copied - left;
1829 }
1830
1831 /*
1832  * Copy as much as we can into the page and return the number of bytes which
1833  * were sucessfully copied.  If a fault is encountered then return the number of
1834  * bytes which were copied.
1835  */
1836 size_t iov_iter_copy_from_user_atomic(struct page *page,
1837                 struct iov_iter *i, unsigned long offset, size_t bytes)
1838 {
1839         char *kaddr;
1840         size_t copied;
1841
1842         BUG_ON(!in_atomic());
1843         kaddr = kmap_atomic(page, KM_USER0);
1844         if (likely(i->nr_segs == 1)) {
1845                 int left;
1846                 char __user *buf = i->iov->iov_base + i->iov_offset;
1847                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1848                                                         buf, bytes);
1849                 copied = bytes - left;
1850         } else {
1851                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1852                                                 i->iov, i->iov_offset, bytes);
1853         }
1854         kunmap_atomic(kaddr, KM_USER0);
1855
1856         return copied;
1857 }
1858 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1859
1860 /*
1861  * This has the same sideeffects and return value as
1862  * iov_iter_copy_from_user_atomic().
1863  * The difference is that it attempts to resolve faults.
1864  * Page must not be locked.
1865  */
1866 size_t iov_iter_copy_from_user(struct page *page,
1867                 struct iov_iter *i, unsigned long offset, size_t bytes)
1868 {
1869         char *kaddr;
1870         size_t copied;
1871
1872         kaddr = kmap(page);
1873         if (likely(i->nr_segs == 1)) {
1874                 int left;
1875                 char __user *buf = i->iov->iov_base + i->iov_offset;
1876                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1877                 copied = bytes - left;
1878         } else {
1879                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1880                                                 i->iov, i->iov_offset, bytes);
1881         }
1882         kunmap(page);
1883         return copied;
1884 }
1885 EXPORT_SYMBOL(iov_iter_copy_from_user);
1886
1887 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1888 {
1889         BUG_ON(i->count < bytes);
1890
1891         if (likely(i->nr_segs == 1)) {
1892                 i->iov_offset += bytes;
1893                 i->count -= bytes;
1894         } else {
1895                 const struct iovec *iov = i->iov;
1896                 size_t base = i->iov_offset;
1897
1898                 /*
1899                  * The !iov->iov_len check ensures we skip over unlikely
1900                  * zero-length segments (without overruning the iovec).
1901                  */
1902                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1903                         int copy;
1904
1905                         copy = min(bytes, iov->iov_len - base);
1906                         BUG_ON(!i->count || i->count < copy);
1907                         i->count -= copy;
1908                         bytes -= copy;
1909                         base += copy;
1910                         if (iov->iov_len == base) {
1911                                 iov++;
1912                                 base = 0;
1913                         }
1914                 }
1915                 i->iov = iov;
1916                 i->iov_offset = base;
1917         }
1918 }
1919 EXPORT_SYMBOL(iov_iter_advance);
1920
1921 /*
1922  * Fault in the first iovec of the given iov_iter, to a maximum length
1923  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1924  * accessed (ie. because it is an invalid address).
1925  *
1926  * writev-intensive code may want this to prefault several iovecs -- that
1927  * would be possible (callers must not rely on the fact that _only_ the
1928  * first iovec will be faulted with the current implementation).
1929  */
1930 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1931 {
1932         char __user *buf = i->iov->iov_base + i->iov_offset;
1933         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1934         return fault_in_pages_readable(buf, bytes);
1935 }
1936 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1937
1938 /*
1939  * Return the count of just the current iov_iter segment.
1940  */
1941 size_t iov_iter_single_seg_count(struct iov_iter *i)
1942 {
1943         const struct iovec *iov = i->iov;
1944         if (i->nr_segs == 1)
1945                 return i->count;
1946         else
1947                 return min(i->count, iov->iov_len - i->iov_offset);
1948 }
1949 EXPORT_SYMBOL(iov_iter_single_seg_count);
1950
1951 /*
1952  * Performs necessary checks before doing a write
1953  *
1954  * Can adjust writing position or amount of bytes to write.
1955  * Returns appropriate error code that caller should return or
1956  * zero in case that write should be allowed.
1957  */
1958 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1959 {
1960         struct inode *inode = file->f_mapping->host;
1961         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1962
1963         if (unlikely(*pos < 0))
1964                 return -EINVAL;
1965
1966         if (!isblk) {
1967                 /* FIXME: this is for backwards compatibility with 2.4 */
1968                 if (file->f_flags & O_APPEND)
1969                         *pos = i_size_read(inode);
1970
1971                 if (limit != RLIM_INFINITY) {
1972                         if (*pos >= limit) {
1973                                 send_sig(SIGXFSZ, current, 0);
1974                                 return -EFBIG;
1975                         }
1976                         if (*count > limit - (typeof(limit))*pos) {
1977                                 *count = limit - (typeof(limit))*pos;
1978                         }
1979                 }
1980         }
1981
1982         /*
1983          * LFS rule
1984          */
1985         if (unlikely(*pos + *count > MAX_NON_LFS &&
1986                                 !(file->f_flags & O_LARGEFILE))) {
1987                 if (*pos >= MAX_NON_LFS) {
1988                         return -EFBIG;
1989                 }
1990                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1991                         *count = MAX_NON_LFS - (unsigned long)*pos;
1992                 }
1993         }
1994
1995         /*
1996          * Are we about to exceed the fs block limit ?
1997          *
1998          * If we have written data it becomes a short write.  If we have
1999          * exceeded without writing data we send a signal and return EFBIG.
2000          * Linus frestrict idea will clean these up nicely..
2001          */
2002         if (likely(!isblk)) {
2003                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2004                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2005                                 return -EFBIG;
2006                         }
2007                         /* zero-length writes at ->s_maxbytes are OK */
2008                 }
2009
2010                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2011                         *count = inode->i_sb->s_maxbytes - *pos;
2012         } else {
2013 #ifdef CONFIG_BLOCK
2014                 loff_t isize;
2015                 if (bdev_read_only(I_BDEV(inode)))
2016                         return -EPERM;
2017                 isize = i_size_read(inode);
2018                 if (*pos >= isize) {
2019                         if (*count || *pos > isize)
2020                                 return -ENOSPC;
2021                 }
2022
2023                 if (*pos + *count > isize)
2024                         *count = isize - *pos;
2025 #else
2026                 return -EPERM;
2027 #endif
2028         }
2029         return 0;
2030 }
2031 EXPORT_SYMBOL(generic_write_checks);
2032
2033 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2034                                 loff_t pos, unsigned len, unsigned flags,
2035                                 struct page **pagep, void **fsdata)
2036 {
2037         const struct address_space_operations *aops = mapping->a_ops;
2038
2039         return aops->write_begin(file, mapping, pos, len, flags,
2040                                                         pagep, fsdata);
2041 }
2042 EXPORT_SYMBOL(pagecache_write_begin);
2043
2044 int pagecache_write_end(struct file *file, struct address_space *mapping,
2045                                 loff_t pos, unsigned len, unsigned copied,
2046                                 struct page *page, void *fsdata)
2047 {
2048         const struct address_space_operations *aops = mapping->a_ops;
2049
2050         mark_page_accessed(page);
2051         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2052 }
2053 EXPORT_SYMBOL(pagecache_write_end);
2054
2055 ssize_t
2056 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2057                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2058                 size_t count, size_t ocount)
2059 {
2060         struct file     *file = iocb->ki_filp;
2061         struct address_space *mapping = file->f_mapping;
2062         struct inode    *inode = mapping->host;
2063         ssize_t         written;
2064         size_t          write_len;
2065         pgoff_t         end;
2066
2067         if (count != ocount)
2068                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2069
2070         write_len = iov_length(iov, *nr_segs);
2071         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2072
2073         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
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_range(mapping,
2294                                         pos, pos + written - 1);
2295
2296         return written ? written : status;
2297 }
2298 EXPORT_SYMBOL(generic_file_buffered_write);
2299
2300 static ssize_t
2301 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2302                                 unsigned long nr_segs, loff_t *ppos)
2303 {
2304         struct file *file = iocb->ki_filp;
2305         struct address_space * mapping = file->f_mapping;
2306         size_t ocount;          /* original count */
2307         size_t count;           /* after file limit checks */
2308         struct inode    *inode = mapping->host;
2309         loff_t          pos;
2310         ssize_t         written;
2311         ssize_t         err;
2312
2313         ocount = 0;
2314         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2315         if (err)
2316                 return err;
2317
2318         count = ocount;
2319         pos = *ppos;
2320
2321         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2322
2323         /* We can write back this queue in page reclaim */
2324         current->backing_dev_info = mapping->backing_dev_info;
2325         written = 0;
2326
2327         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2328         if (err)
2329                 goto out;
2330
2331         if (count == 0)
2332                 goto out;
2333
2334         err = file_remove_suid(file);
2335         if (err)
2336                 goto out;
2337
2338         file_update_time(file);
2339
2340         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2341         if (unlikely(file->f_flags & O_DIRECT)) {
2342                 loff_t endbyte;
2343                 ssize_t written_buffered;
2344
2345                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2346                                                         ppos, count, ocount);
2347                 if (written < 0 || written == count)
2348                         goto out;
2349                 /*
2350                  * direct-io write to a hole: fall through to buffered I/O
2351                  * for completing the rest of the request.
2352                  */
2353                 pos += written;
2354                 count -= written;
2355                 written_buffered = generic_file_buffered_write(iocb, iov,
2356                                                 nr_segs, pos, ppos, count,
2357                                                 written);
2358                 /*
2359                  * If generic_file_buffered_write() retuned a synchronous error
2360                  * then we want to return the number of bytes which were
2361                  * direct-written, or the error code if that was zero.  Note
2362                  * that this differs from normal direct-io semantics, which
2363                  * will return -EFOO even if some bytes were written.
2364                  */
2365                 if (written_buffered < 0) {
2366                         err = written_buffered;
2367                         goto out;
2368                 }
2369
2370                 /*
2371                  * We need to ensure that the page cache pages are written to
2372                  * disk and invalidated to preserve the expected O_DIRECT
2373                  * semantics.
2374                  */
2375                 endbyte = pos + written_buffered - written - 1;
2376                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2377                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2378                                             SYNC_FILE_RANGE_WRITE|
2379                                             SYNC_FILE_RANGE_WAIT_AFTER);
2380                 if (err == 0) {
2381                         written = written_buffered;
2382                         invalidate_mapping_pages(mapping,
2383                                                  pos >> PAGE_CACHE_SHIFT,
2384                                                  endbyte >> PAGE_CACHE_SHIFT);
2385                 } else {
2386                         /*
2387                          * We don't know how much we wrote, so just return
2388                          * the number of bytes which were direct-written
2389                          */
2390                 }
2391         } else {
2392                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2393                                 pos, ppos, count, written);
2394         }
2395 out:
2396         current->backing_dev_info = NULL;
2397         return written ? written : err;
2398 }
2399
2400 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2401                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2402 {
2403         struct file *file = iocb->ki_filp;
2404         struct address_space *mapping = file->f_mapping;
2405         struct inode *inode = mapping->host;
2406         ssize_t ret;
2407
2408         BUG_ON(iocb->ki_pos != pos);
2409
2410         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2411                         &iocb->ki_pos);
2412
2413         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2414                 ssize_t err;
2415
2416                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2417                 if (err < 0)
2418                         ret = err;
2419         }
2420         return ret;
2421 }
2422 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2423
2424 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2425                 unsigned long nr_segs, loff_t pos)
2426 {
2427         struct file *file = iocb->ki_filp;
2428         struct address_space *mapping = file->f_mapping;
2429         struct inode *inode = mapping->host;
2430         ssize_t ret;
2431
2432         BUG_ON(iocb->ki_pos != pos);
2433
2434         mutex_lock(&inode->i_mutex);
2435         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2436                         &iocb->ki_pos);
2437         mutex_unlock(&inode->i_mutex);
2438
2439         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2440                 ssize_t err;
2441
2442                 err = sync_page_range(inode, mapping, pos, ret);
2443                 if (err < 0)
2444                         ret = err;
2445         }
2446         return ret;
2447 }
2448 EXPORT_SYMBOL(generic_file_aio_write);
2449
2450 /**
2451  * try_to_release_page() - release old fs-specific metadata on a page
2452  *
2453  * @page: the page which the kernel is trying to free
2454  * @gfp_mask: memory allocation flags (and I/O mode)
2455  *
2456  * The address_space is to try to release any data against the page
2457  * (presumably at page->private).  If the release was successful, return `1'.
2458  * Otherwise return zero.
2459  *
2460  * The @gfp_mask argument specifies whether I/O may be performed to release
2461  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2462  *
2463  */
2464 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2465 {
2466         struct address_space * const mapping = page->mapping;
2467
2468         BUG_ON(!PageLocked(page));
2469         if (PageWriteback(page))
2470                 return 0;
2471
2472         if (mapping && mapping->a_ops->releasepage)
2473                 return mapping->a_ops->releasepage(page, gfp_mask);
2474         return try_to_free_buffers(page);
2475 }
2476
2477 EXPORT_SYMBOL(try_to_release_page);