Merge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[linux-2.6] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/backing-dev.h>
32 #include <linux/security.h>
33 #include <linux/syscalls.h>
34 #include <linux/cpuset.h>
35 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
36 #include <linux/memcontrol.h>
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 static ssize_t
47 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
48         loff_t offset, unsigned long nr_segs);
49
50 /*
51  * Shared mappings implemented 30.11.1994. It's not fully working yet,
52  * though.
53  *
54  * Shared mappings now work. 15.8.1995  Bruno.
55  *
56  * finished 'unifying' the page and buffer cache and SMP-threaded the
57  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
58  *
59  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
60  */
61
62 /*
63  * Lock ordering:
64  *
65  *  ->i_mmap_lock               (vmtruncate)
66  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
67  *      ->swap_lock             (exclusive_swap_page, others)
68  *        ->mapping->tree_lock
69  *
70  *  ->i_mutex
71  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
72  *
73  *  ->mmap_sem
74  *    ->i_mmap_lock
75  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
76  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
77  *
78  *  ->mmap_sem
79  *    ->lock_page               (access_process_vm)
80  *
81  *  ->i_mutex                   (generic_file_buffered_write)
82  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
83  *
84  *  ->i_mutex
85  *    ->i_alloc_sem             (various)
86  *
87  *  ->inode_lock
88  *    ->sb_lock                 (fs/fs-writeback.c)
89  *    ->mapping->tree_lock      (__sync_single_inode)
90  *
91  *  ->i_mmap_lock
92  *    ->anon_vma.lock           (vma_adjust)
93  *
94  *  ->anon_vma.lock
95  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
96  *
97  *  ->page_table_lock or pte_lock
98  *    ->swap_lock               (try_to_unmap_one)
99  *    ->private_lock            (try_to_unmap_one)
100  *    ->tree_lock               (try_to_unmap_one)
101  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
102  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
103  *    ->private_lock            (page_remove_rmap->set_page_dirty)
104  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
106  *    ->inode_lock              (zap_pte_range->set_page_dirty)
107  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
108  *
109  *  ->task->proc_lock
110  *    ->dcache_lock             (proc_pid_lookup)
111  */
112
113 /*
114  * Remove a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         mem_cgroup_uncharge_page(page);
123         radix_tree_delete(&mapping->page_tree, page->index);
124         page->mapping = NULL;
125         mapping->nrpages--;
126         __dec_zone_page_state(page, NR_FILE_PAGES);
127         BUG_ON(page_mapped(page));
128
129         /*
130          * Some filesystems seem to re-dirty the page even after
131          * the VM has canceled the dirty bit (eg ext3 journaling).
132          *
133          * Fix it up by doing a final dirty accounting check after
134          * having removed the page entirely.
135          */
136         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
137                 dec_zone_page_state(page, NR_FILE_DIRTY);
138                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
139         }
140 }
141
142 void remove_from_page_cache(struct page *page)
143 {
144         struct address_space *mapping = page->mapping;
145
146         BUG_ON(!PageLocked(page));
147
148         write_lock_irq(&mapping->tree_lock);
149         __remove_from_page_cache(page);
150         write_unlock_irq(&mapping->tree_lock);
151 }
152
153 static int sync_page(void *word)
154 {
155         struct address_space *mapping;
156         struct page *page;
157
158         page = container_of((unsigned long *)word, struct page, flags);
159
160         /*
161          * page_mapping() is being called without PG_locked held.
162          * Some knowledge of the state and use of the page is used to
163          * reduce the requirements down to a memory barrier.
164          * The danger here is of a stale page_mapping() return value
165          * indicating a struct address_space different from the one it's
166          * associated with when it is associated with one.
167          * After smp_mb(), it's either the correct page_mapping() for
168          * the page, or an old page_mapping() and the page's own
169          * page_mapping() has gone NULL.
170          * The ->sync_page() address_space operation must tolerate
171          * page_mapping() going NULL. By an amazing coincidence,
172          * this comes about because none of the users of the page
173          * in the ->sync_page() methods make essential use of the
174          * page_mapping(), merely passing the page down to the backing
175          * device's unplug functions when it's non-NULL, which in turn
176          * ignore it for all cases but swap, where only page_private(page) is
177          * of interest. When page_mapping() does go NULL, the entire
178          * call stack gracefully ignores the page and returns.
179          * -- wli
180          */
181         smp_mb();
182         mapping = page_mapping(page);
183         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
184                 mapping->a_ops->sync_page(page);
185         io_schedule();
186         return 0;
187 }
188
189 static int sync_page_killable(void *word)
190 {
191         sync_page(word);
192         return fatal_signal_pending(current) ? -EINTR : 0;
193 }
194
195 /**
196  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
197  * @mapping:    address space structure to write
198  * @start:      offset in bytes where the range starts
199  * @end:        offset in bytes where the range ends (inclusive)
200  * @sync_mode:  enable synchronous operation
201  *
202  * Start writeback against all of a mapping's dirty pages that lie
203  * within the byte offsets <start, end> inclusive.
204  *
205  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
206  * opposed to a regular memory cleansing writeback.  The difference between
207  * these two operations is that if a dirty page/buffer is encountered, it must
208  * be waited upon, and not just skipped over.
209  */
210 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
211                                 loff_t end, int sync_mode)
212 {
213         int ret;
214         struct writeback_control wbc = {
215                 .sync_mode = sync_mode,
216                 .nr_to_write = mapping->nrpages * 2,
217                 .range_start = start,
218                 .range_end = end,
219         };
220
221         if (!mapping_cap_writeback_dirty(mapping))
222                 return 0;
223
224         ret = do_writepages(mapping, &wbc);
225         return ret;
226 }
227
228 static inline int __filemap_fdatawrite(struct address_space *mapping,
229         int sync_mode)
230 {
231         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
232 }
233
234 int filemap_fdatawrite(struct address_space *mapping)
235 {
236         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
237 }
238 EXPORT_SYMBOL(filemap_fdatawrite);
239
240 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
241                                 loff_t end)
242 {
243         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
244 }
245
246 /**
247  * filemap_flush - mostly a non-blocking flush
248  * @mapping:    target address_space
249  *
250  * This is a mostly non-blocking flush.  Not suitable for data-integrity
251  * purposes - I/O may not be started against all dirty pages.
252  */
253 int filemap_flush(struct address_space *mapping)
254 {
255         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
256 }
257 EXPORT_SYMBOL(filemap_flush);
258
259 /**
260  * wait_on_page_writeback_range - wait for writeback to complete
261  * @mapping:    target address_space
262  * @start:      beginning page index
263  * @end:        ending page index
264  *
265  * Wait for writeback to complete against pages indexed by start->end
266  * inclusive
267  */
268 int wait_on_page_writeback_range(struct address_space *mapping,
269                                 pgoff_t start, pgoff_t end)
270 {
271         struct pagevec pvec;
272         int nr_pages;
273         int ret = 0;
274         pgoff_t index;
275
276         if (end < start)
277                 return 0;
278
279         pagevec_init(&pvec, 0);
280         index = start;
281         while ((index <= end) &&
282                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
283                         PAGECACHE_TAG_WRITEBACK,
284                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
285                 unsigned i;
286
287                 for (i = 0; i < nr_pages; i++) {
288                         struct page *page = pvec.pages[i];
289
290                         /* until radix tree lookup accepts end_index */
291                         if (page->index > end)
292                                 continue;
293
294                         wait_on_page_writeback(page);
295                         if (PageError(page))
296                                 ret = -EIO;
297                 }
298                 pagevec_release(&pvec);
299                 cond_resched();
300         }
301
302         /* Check for outstanding write errors */
303         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
304                 ret = -ENOSPC;
305         if (test_and_clear_bit(AS_EIO, &mapping->flags))
306                 ret = -EIO;
307
308         return ret;
309 }
310
311 /**
312  * sync_page_range - write and wait on all pages in the passed range
313  * @inode:      target inode
314  * @mapping:    target address_space
315  * @pos:        beginning offset in pages to write
316  * @count:      number of bytes to write
317  *
318  * Write and wait upon all the pages in the passed range.  This is a "data
319  * integrity" operation.  It waits upon in-flight writeout before starting and
320  * waiting upon new writeout.  If there was an IO error, return it.
321  *
322  * We need to re-take i_mutex during the generic_osync_inode list walk because
323  * it is otherwise livelockable.
324  */
325 int sync_page_range(struct inode *inode, struct address_space *mapping,
326                         loff_t pos, loff_t count)
327 {
328         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
329         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
330         int ret;
331
332         if (!mapping_cap_writeback_dirty(mapping) || !count)
333                 return 0;
334         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
335         if (ret == 0) {
336                 mutex_lock(&inode->i_mutex);
337                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
338                 mutex_unlock(&inode->i_mutex);
339         }
340         if (ret == 0)
341                 ret = wait_on_page_writeback_range(mapping, start, end);
342         return ret;
343 }
344 EXPORT_SYMBOL(sync_page_range);
345
346 /**
347  * sync_page_range_nolock
348  * @inode:      target inode
349  * @mapping:    target address_space
350  * @pos:        beginning offset in pages to write
351  * @count:      number of bytes to write
352  *
353  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
354  * as it forces O_SYNC writers to different parts of the same file
355  * to be serialised right until io completion.
356  */
357 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
358                            loff_t pos, loff_t count)
359 {
360         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
361         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
362         int ret;
363
364         if (!mapping_cap_writeback_dirty(mapping) || !count)
365                 return 0;
366         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
367         if (ret == 0)
368                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
369         if (ret == 0)
370                 ret = wait_on_page_writeback_range(mapping, start, end);
371         return ret;
372 }
373 EXPORT_SYMBOL(sync_page_range_nolock);
374
375 /**
376  * filemap_fdatawait - wait for all under-writeback pages to complete
377  * @mapping: address space structure to wait for
378  *
379  * Walk the list of under-writeback pages of the given address space
380  * and wait for all of them.
381  */
382 int filemap_fdatawait(struct address_space *mapping)
383 {
384         loff_t i_size = i_size_read(mapping->host);
385
386         if (i_size == 0)
387                 return 0;
388
389         return wait_on_page_writeback_range(mapping, 0,
390                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
391 }
392 EXPORT_SYMBOL(filemap_fdatawait);
393
394 int filemap_write_and_wait(struct address_space *mapping)
395 {
396         int err = 0;
397
398         if (mapping->nrpages) {
399                 err = filemap_fdatawrite(mapping);
400                 /*
401                  * Even if the above returned error, the pages may be
402                  * written partially (e.g. -ENOSPC), so we wait for it.
403                  * But the -EIO is special case, it may indicate the worst
404                  * thing (e.g. bug) happened, so we avoid waiting for it.
405                  */
406                 if (err != -EIO) {
407                         int err2 = filemap_fdatawait(mapping);
408                         if (!err)
409                                 err = err2;
410                 }
411         }
412         return err;
413 }
414 EXPORT_SYMBOL(filemap_write_and_wait);
415
416 /**
417  * filemap_write_and_wait_range - write out & wait on a file range
418  * @mapping:    the address_space for the pages
419  * @lstart:     offset in bytes where the range starts
420  * @lend:       offset in bytes where the range ends (inclusive)
421  *
422  * Write out and wait upon file offsets lstart->lend, inclusive.
423  *
424  * Note that `lend' is inclusive (describes the last byte to be written) so
425  * that this function can be used to write to the very end-of-file (end = -1).
426  */
427 int filemap_write_and_wait_range(struct address_space *mapping,
428                                  loff_t lstart, loff_t lend)
429 {
430         int err = 0;
431
432         if (mapping->nrpages) {
433                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
434                                                  WB_SYNC_ALL);
435                 /* See comment of filemap_write_and_wait() */
436                 if (err != -EIO) {
437                         int err2 = wait_on_page_writeback_range(mapping,
438                                                 lstart >> PAGE_CACHE_SHIFT,
439                                                 lend >> PAGE_CACHE_SHIFT);
440                         if (!err)
441                                 err = err2;
442                 }
443         }
444         return err;
445 }
446
447 /**
448  * add_to_page_cache - add newly allocated pagecache pages
449  * @page:       page to add
450  * @mapping:    the page's address_space
451  * @offset:     page index
452  * @gfp_mask:   page allocation mode
453  *
454  * This function is used to add newly allocated pagecache pages;
455  * the page is new, so we can just run SetPageLocked() against it.
456  * The other page state flags were set by rmqueue().
457  *
458  * This function does not add the page to the LRU.  The caller must do that.
459  */
460 int add_to_page_cache(struct page *page, struct address_space *mapping,
461                 pgoff_t offset, gfp_t gfp_mask)
462 {
463         int error = mem_cgroup_cache_charge(page, current->mm,
464                                         gfp_mask & ~__GFP_HIGHMEM);
465         if (error)
466                 goto out;
467
468         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
469         if (error == 0) {
470                 write_lock_irq(&mapping->tree_lock);
471                 error = radix_tree_insert(&mapping->page_tree, offset, page);
472                 if (!error) {
473                         page_cache_get(page);
474                         SetPageLocked(page);
475                         page->mapping = mapping;
476                         page->index = offset;
477                         mapping->nrpages++;
478                         __inc_zone_page_state(page, NR_FILE_PAGES);
479                 } else
480                         mem_cgroup_uncharge_page(page);
481
482                 write_unlock_irq(&mapping->tree_lock);
483                 radix_tree_preload_end();
484         } else
485                 mem_cgroup_uncharge_page(page);
486 out:
487         return error;
488 }
489 EXPORT_SYMBOL(add_to_page_cache);
490
491 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
492                                 pgoff_t offset, gfp_t gfp_mask)
493 {
494         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
495         if (ret == 0)
496                 lru_cache_add(page);
497         return ret;
498 }
499
500 #ifdef CONFIG_NUMA
501 struct page *__page_cache_alloc(gfp_t gfp)
502 {
503         if (cpuset_do_page_mem_spread()) {
504                 int n = cpuset_mem_spread_node();
505                 return alloc_pages_node(n, gfp, 0);
506         }
507         return alloc_pages(gfp, 0);
508 }
509 EXPORT_SYMBOL(__page_cache_alloc);
510 #endif
511
512 static int __sleep_on_page_lock(void *word)
513 {
514         io_schedule();
515         return 0;
516 }
517
518 /*
519  * In order to wait for pages to become available there must be
520  * waitqueues associated with pages. By using a hash table of
521  * waitqueues where the bucket discipline is to maintain all
522  * waiters on the same queue and wake all when any of the pages
523  * become available, and for the woken contexts to check to be
524  * sure the appropriate page became available, this saves space
525  * at a cost of "thundering herd" phenomena during rare hash
526  * collisions.
527  */
528 static wait_queue_head_t *page_waitqueue(struct page *page)
529 {
530         const struct zone *zone = page_zone(page);
531
532         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
533 }
534
535 static inline void wake_up_page(struct page *page, int bit)
536 {
537         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
538 }
539
540 void wait_on_page_bit(struct page *page, int bit_nr)
541 {
542         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
543
544         if (test_bit(bit_nr, &page->flags))
545                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
546                                                         TASK_UNINTERRUPTIBLE);
547 }
548 EXPORT_SYMBOL(wait_on_page_bit);
549
550 /**
551  * unlock_page - unlock a locked page
552  * @page: the page
553  *
554  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
555  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
556  * mechananism between PageLocked pages and PageWriteback pages is shared.
557  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
558  *
559  * The first mb is necessary to safely close the critical section opened by the
560  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
561  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
562  * parallel wait_on_page_locked()).
563  */
564 void unlock_page(struct page *page)
565 {
566         smp_mb__before_clear_bit();
567         if (!TestClearPageLocked(page))
568                 BUG();
569         smp_mb__after_clear_bit(); 
570         wake_up_page(page, PG_locked);
571 }
572 EXPORT_SYMBOL(unlock_page);
573
574 /**
575  * end_page_writeback - end writeback against a page
576  * @page: the page
577  */
578 void end_page_writeback(struct page *page)
579 {
580         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
581                 if (!test_clear_page_writeback(page))
582                         BUG();
583         }
584         smp_mb__after_clear_bit();
585         wake_up_page(page, PG_writeback);
586 }
587 EXPORT_SYMBOL(end_page_writeback);
588
589 /**
590  * __lock_page - get a lock on the page, assuming we need to sleep to get it
591  * @page: the page to lock
592  *
593  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
594  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
595  * chances are that on the second loop, the block layer's plug list is empty,
596  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
597  */
598 void __lock_page(struct page *page)
599 {
600         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
601
602         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
603                                                         TASK_UNINTERRUPTIBLE);
604 }
605 EXPORT_SYMBOL(__lock_page);
606
607 int __lock_page_killable(struct page *page)
608 {
609         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
610
611         return __wait_on_bit_lock(page_waitqueue(page), &wait,
612                                         sync_page_killable, TASK_KILLABLE);
613 }
614
615 /*
616  * Variant of lock_page that does not require the caller to hold a reference
617  * on the page's mapping.
618  */
619 void __lock_page_nosync(struct page *page)
620 {
621         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
622         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
623                                                         TASK_UNINTERRUPTIBLE);
624 }
625
626 /**
627  * find_get_page - find and get a page reference
628  * @mapping: the address_space to search
629  * @offset: the page index
630  *
631  * Is there a pagecache struct page at the given (mapping, offset) tuple?
632  * If yes, increment its refcount and return it; if no, return NULL.
633  */
634 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
635 {
636         struct page *page;
637
638         read_lock_irq(&mapping->tree_lock);
639         page = radix_tree_lookup(&mapping->page_tree, offset);
640         if (page)
641                 page_cache_get(page);
642         read_unlock_irq(&mapping->tree_lock);
643         return page;
644 }
645 EXPORT_SYMBOL(find_get_page);
646
647 /**
648  * find_lock_page - locate, pin and lock a pagecache page
649  * @mapping: the address_space to search
650  * @offset: the page index
651  *
652  * Locates the desired pagecache page, locks it, increments its reference
653  * count and returns its address.
654  *
655  * Returns zero if the page was not present. find_lock_page() may sleep.
656  */
657 struct page *find_lock_page(struct address_space *mapping,
658                                 pgoff_t offset)
659 {
660         struct page *page;
661
662 repeat:
663         read_lock_irq(&mapping->tree_lock);
664         page = radix_tree_lookup(&mapping->page_tree, offset);
665         if (page) {
666                 page_cache_get(page);
667                 if (TestSetPageLocked(page)) {
668                         read_unlock_irq(&mapping->tree_lock);
669                         __lock_page(page);
670
671                         /* Has the page been truncated while we slept? */
672                         if (unlikely(page->mapping != mapping)) {
673                                 unlock_page(page);
674                                 page_cache_release(page);
675                                 goto repeat;
676                         }
677                         VM_BUG_ON(page->index != offset);
678                         goto out;
679                 }
680         }
681         read_unlock_irq(&mapping->tree_lock);
682 out:
683         return page;
684 }
685 EXPORT_SYMBOL(find_lock_page);
686
687 /**
688  * find_or_create_page - locate or add a pagecache page
689  * @mapping: the page's address_space
690  * @index: the page's index into the mapping
691  * @gfp_mask: page allocation mode
692  *
693  * Locates a page in the pagecache.  If the page is not present, a new page
694  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
695  * LRU list.  The returned page is locked and has its reference count
696  * incremented.
697  *
698  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
699  * allocation!
700  *
701  * find_or_create_page() returns the desired page's address, or zero on
702  * memory exhaustion.
703  */
704 struct page *find_or_create_page(struct address_space *mapping,
705                 pgoff_t index, gfp_t gfp_mask)
706 {
707         struct page *page;
708         int err;
709 repeat:
710         page = find_lock_page(mapping, index);
711         if (!page) {
712                 page = __page_cache_alloc(gfp_mask);
713                 if (!page)
714                         return NULL;
715                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
716                 if (unlikely(err)) {
717                         page_cache_release(page);
718                         page = NULL;
719                         if (err == -EEXIST)
720                                 goto repeat;
721                 }
722         }
723         return page;
724 }
725 EXPORT_SYMBOL(find_or_create_page);
726
727 /**
728  * find_get_pages - gang pagecache lookup
729  * @mapping:    The address_space to search
730  * @start:      The starting page index
731  * @nr_pages:   The maximum number of pages
732  * @pages:      Where the resulting pages are placed
733  *
734  * find_get_pages() will search for and return a group of up to
735  * @nr_pages pages in the mapping.  The pages are placed at @pages.
736  * find_get_pages() takes a reference against the returned pages.
737  *
738  * The search returns a group of mapping-contiguous pages with ascending
739  * indexes.  There may be holes in the indices due to not-present pages.
740  *
741  * find_get_pages() returns the number of pages which were found.
742  */
743 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
744                             unsigned int nr_pages, struct page **pages)
745 {
746         unsigned int i;
747         unsigned int ret;
748
749         read_lock_irq(&mapping->tree_lock);
750         ret = radix_tree_gang_lookup(&mapping->page_tree,
751                                 (void **)pages, start, nr_pages);
752         for (i = 0; i < ret; i++)
753                 page_cache_get(pages[i]);
754         read_unlock_irq(&mapping->tree_lock);
755         return ret;
756 }
757
758 /**
759  * find_get_pages_contig - gang contiguous pagecache lookup
760  * @mapping:    The address_space to search
761  * @index:      The starting page index
762  * @nr_pages:   The maximum number of pages
763  * @pages:      Where the resulting pages are placed
764  *
765  * find_get_pages_contig() works exactly like find_get_pages(), except
766  * that the returned number of pages are guaranteed to be contiguous.
767  *
768  * find_get_pages_contig() returns the number of pages which were found.
769  */
770 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
771                                unsigned int nr_pages, struct page **pages)
772 {
773         unsigned int i;
774         unsigned int ret;
775
776         read_lock_irq(&mapping->tree_lock);
777         ret = radix_tree_gang_lookup(&mapping->page_tree,
778                                 (void **)pages, index, nr_pages);
779         for (i = 0; i < ret; i++) {
780                 if (pages[i]->mapping == NULL || pages[i]->index != index)
781                         break;
782
783                 page_cache_get(pages[i]);
784                 index++;
785         }
786         read_unlock_irq(&mapping->tree_lock);
787         return i;
788 }
789 EXPORT_SYMBOL(find_get_pages_contig);
790
791 /**
792  * find_get_pages_tag - find and return pages that match @tag
793  * @mapping:    the address_space to search
794  * @index:      the starting page index
795  * @tag:        the tag index
796  * @nr_pages:   the maximum number of pages
797  * @pages:      where the resulting pages are placed
798  *
799  * Like find_get_pages, except we only return pages which are tagged with
800  * @tag.   We update @index to index the next page for the traversal.
801  */
802 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
803                         int tag, unsigned int nr_pages, struct page **pages)
804 {
805         unsigned int i;
806         unsigned int ret;
807
808         read_lock_irq(&mapping->tree_lock);
809         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
810                                 (void **)pages, *index, nr_pages, tag);
811         for (i = 0; i < ret; i++)
812                 page_cache_get(pages[i]);
813         if (ret)
814                 *index = pages[ret - 1]->index + 1;
815         read_unlock_irq(&mapping->tree_lock);
816         return ret;
817 }
818 EXPORT_SYMBOL(find_get_pages_tag);
819
820 /**
821  * grab_cache_page_nowait - returns locked page at given index in given cache
822  * @mapping: target address_space
823  * @index: the page index
824  *
825  * Same as grab_cache_page(), but do not wait if the page is unavailable.
826  * This is intended for speculative data generators, where the data can
827  * be regenerated if the page couldn't be grabbed.  This routine should
828  * be safe to call while holding the lock for another page.
829  *
830  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
831  * and deadlock against the caller's locked page.
832  */
833 struct page *
834 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
835 {
836         struct page *page = find_get_page(mapping, index);
837
838         if (page) {
839                 if (!TestSetPageLocked(page))
840                         return page;
841                 page_cache_release(page);
842                 return NULL;
843         }
844         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
845         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
846                 page_cache_release(page);
847                 page = NULL;
848         }
849         return page;
850 }
851 EXPORT_SYMBOL(grab_cache_page_nowait);
852
853 /*
854  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
855  * a _large_ part of the i/o request. Imagine the worst scenario:
856  *
857  *      ---R__________________________________________B__________
858  *         ^ reading here                             ^ bad block(assume 4k)
859  *
860  * read(R) => miss => readahead(R...B) => media error => frustrating retries
861  * => failing the whole request => read(R) => read(R+1) =>
862  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
863  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
864  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
865  *
866  * It is going insane. Fix it by quickly scaling down the readahead size.
867  */
868 static void shrink_readahead_size_eio(struct file *filp,
869                                         struct file_ra_state *ra)
870 {
871         if (!ra->ra_pages)
872                 return;
873
874         ra->ra_pages /= 4;
875 }
876
877 /**
878  * do_generic_file_read - generic file read routine
879  * @filp:       the file to read
880  * @ppos:       current file position
881  * @desc:       read_descriptor
882  * @actor:      read method
883  *
884  * This is a generic file read routine, and uses the
885  * mapping->a_ops->readpage() function for the actual low-level stuff.
886  *
887  * This is really ugly. But the goto's actually try to clarify some
888  * of the logic when it comes to error handling etc.
889  */
890 static void do_generic_file_read(struct file *filp, loff_t *ppos,
891                 read_descriptor_t *desc, read_actor_t actor)
892 {
893         struct address_space *mapping = filp->f_mapping;
894         struct inode *inode = mapping->host;
895         struct file_ra_state *ra = &filp->f_ra;
896         pgoff_t index;
897         pgoff_t last_index;
898         pgoff_t prev_index;
899         unsigned long offset;      /* offset into pagecache page */
900         unsigned int prev_offset;
901         int error;
902
903         index = *ppos >> PAGE_CACHE_SHIFT;
904         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
905         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
906         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
907         offset = *ppos & ~PAGE_CACHE_MASK;
908
909         for (;;) {
910                 struct page *page;
911                 pgoff_t end_index;
912                 loff_t isize;
913                 unsigned long nr, ret;
914
915                 cond_resched();
916 find_page:
917                 page = find_get_page(mapping, index);
918                 if (!page) {
919                         page_cache_sync_readahead(mapping,
920                                         ra, filp,
921                                         index, last_index - index);
922                         page = find_get_page(mapping, index);
923                         if (unlikely(page == NULL))
924                                 goto no_cached_page;
925                 }
926                 if (PageReadahead(page)) {
927                         page_cache_async_readahead(mapping,
928                                         ra, filp, page,
929                                         index, last_index - index);
930                 }
931                 if (!PageUptodate(page))
932                         goto page_not_up_to_date;
933 page_ok:
934                 /*
935                  * i_size must be checked after we know the page is Uptodate.
936                  *
937                  * Checking i_size after the check allows us to calculate
938                  * the correct value for "nr", which means the zero-filled
939                  * part of the page is not copied back to userspace (unless
940                  * another truncate extends the file - this is desired though).
941                  */
942
943                 isize = i_size_read(inode);
944                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
945                 if (unlikely(!isize || index > end_index)) {
946                         page_cache_release(page);
947                         goto out;
948                 }
949
950                 /* nr is the maximum number of bytes to copy from this page */
951                 nr = PAGE_CACHE_SIZE;
952                 if (index == end_index) {
953                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
954                         if (nr <= offset) {
955                                 page_cache_release(page);
956                                 goto out;
957                         }
958                 }
959                 nr = nr - offset;
960
961                 /* If users can be writing to this page using arbitrary
962                  * virtual addresses, take care about potential aliasing
963                  * before reading the page on the kernel side.
964                  */
965                 if (mapping_writably_mapped(mapping))
966                         flush_dcache_page(page);
967
968                 /*
969                  * When a sequential read accesses a page several times,
970                  * only mark it as accessed the first time.
971                  */
972                 if (prev_index != index || offset != prev_offset)
973                         mark_page_accessed(page);
974                 prev_index = index;
975
976                 /*
977                  * Ok, we have the page, and it's up-to-date, so
978                  * now we can copy it to user space...
979                  *
980                  * The actor routine returns how many bytes were actually used..
981                  * NOTE! This may not be the same as how much of a user buffer
982                  * we filled up (we may be padding etc), so we can only update
983                  * "pos" here (the actor routine has to update the user buffer
984                  * pointers and the remaining count).
985                  */
986                 ret = actor(desc, page, offset, nr);
987                 offset += ret;
988                 index += offset >> PAGE_CACHE_SHIFT;
989                 offset &= ~PAGE_CACHE_MASK;
990                 prev_offset = offset;
991
992                 page_cache_release(page);
993                 if (ret == nr && desc->count)
994                         continue;
995                 goto out;
996
997 page_not_up_to_date:
998                 /* Get exclusive access to the page ... */
999                 if (lock_page_killable(page))
1000                         goto readpage_eio;
1001
1002                 /* Did it get truncated before we got the lock? */
1003                 if (!page->mapping) {
1004                         unlock_page(page);
1005                         page_cache_release(page);
1006                         continue;
1007                 }
1008
1009                 /* Did somebody else fill it already? */
1010                 if (PageUptodate(page)) {
1011                         unlock_page(page);
1012                         goto page_ok;
1013                 }
1014
1015 readpage:
1016                 /* Start the actual read. The read will unlock the page. */
1017                 error = mapping->a_ops->readpage(filp, page);
1018
1019                 if (unlikely(error)) {
1020                         if (error == AOP_TRUNCATED_PAGE) {
1021                                 page_cache_release(page);
1022                                 goto find_page;
1023                         }
1024                         goto readpage_error;
1025                 }
1026
1027                 if (!PageUptodate(page)) {
1028                         if (lock_page_killable(page))
1029                                 goto readpage_eio;
1030                         if (!PageUptodate(page)) {
1031                                 if (page->mapping == NULL) {
1032                                         /*
1033                                          * invalidate_inode_pages got it
1034                                          */
1035                                         unlock_page(page);
1036                                         page_cache_release(page);
1037                                         goto find_page;
1038                                 }
1039                                 unlock_page(page);
1040                                 shrink_readahead_size_eio(filp, ra);
1041                                 goto readpage_eio;
1042                         }
1043                         unlock_page(page);
1044                 }
1045
1046                 goto page_ok;
1047
1048 readpage_eio:
1049                 error = -EIO;
1050 readpage_error:
1051                 /* UHHUH! A synchronous read error occurred. Report it */
1052                 desc->error = error;
1053                 page_cache_release(page);
1054                 goto out;
1055
1056 no_cached_page:
1057                 /*
1058                  * Ok, it wasn't cached, so we need to create a new
1059                  * page..
1060                  */
1061                 page = page_cache_alloc_cold(mapping);
1062                 if (!page) {
1063                         desc->error = -ENOMEM;
1064                         goto out;
1065                 }
1066                 error = add_to_page_cache_lru(page, mapping,
1067                                                 index, GFP_KERNEL);
1068                 if (error) {
1069                         page_cache_release(page);
1070                         if (error == -EEXIST)
1071                                 goto find_page;
1072                         desc->error = error;
1073                         goto out;
1074                 }
1075                 goto readpage;
1076         }
1077
1078 out:
1079         ra->prev_pos = prev_index;
1080         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1081         ra->prev_pos |= prev_offset;
1082
1083         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1084         if (filp)
1085                 file_accessed(filp);
1086 }
1087
1088 int file_read_actor(read_descriptor_t *desc, struct page *page,
1089                         unsigned long offset, unsigned long size)
1090 {
1091         char *kaddr;
1092         unsigned long left, count = desc->count;
1093
1094         if (size > count)
1095                 size = count;
1096
1097         /*
1098          * Faults on the destination of a read are common, so do it before
1099          * taking the kmap.
1100          */
1101         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1102                 kaddr = kmap_atomic(page, KM_USER0);
1103                 left = __copy_to_user_inatomic(desc->arg.buf,
1104                                                 kaddr + offset, size);
1105                 kunmap_atomic(kaddr, KM_USER0);
1106                 if (left == 0)
1107                         goto success;
1108         }
1109
1110         /* Do it the slow way */
1111         kaddr = kmap(page);
1112         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1113         kunmap(page);
1114
1115         if (left) {
1116                 size -= left;
1117                 desc->error = -EFAULT;
1118         }
1119 success:
1120         desc->count = count - size;
1121         desc->written += size;
1122         desc->arg.buf += size;
1123         return size;
1124 }
1125
1126 /*
1127  * Performs necessary checks before doing a write
1128  * @iov:        io vector request
1129  * @nr_segs:    number of segments in the iovec
1130  * @count:      number of bytes to write
1131  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1132  *
1133  * Adjust number of segments and amount of bytes to write (nr_segs should be
1134  * properly initialized first). Returns appropriate error code that caller
1135  * should return or zero in case that write should be allowed.
1136  */
1137 int generic_segment_checks(const struct iovec *iov,
1138                         unsigned long *nr_segs, size_t *count, int access_flags)
1139 {
1140         unsigned long   seg;
1141         size_t cnt = 0;
1142         for (seg = 0; seg < *nr_segs; seg++) {
1143                 const struct iovec *iv = &iov[seg];
1144
1145                 /*
1146                  * If any segment has a negative length, or the cumulative
1147                  * length ever wraps negative then return -EINVAL.
1148                  */
1149                 cnt += iv->iov_len;
1150                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1151                         return -EINVAL;
1152                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1153                         continue;
1154                 if (seg == 0)
1155                         return -EFAULT;
1156                 *nr_segs = seg;
1157                 cnt -= iv->iov_len;     /* This segment is no good */
1158                 break;
1159         }
1160         *count = cnt;
1161         return 0;
1162 }
1163 EXPORT_SYMBOL(generic_segment_checks);
1164
1165 /**
1166  * generic_file_aio_read - generic filesystem read routine
1167  * @iocb:       kernel I/O control block
1168  * @iov:        io vector request
1169  * @nr_segs:    number of segments in the iovec
1170  * @pos:        current file position
1171  *
1172  * This is the "read()" routine for all filesystems
1173  * that can use the page cache directly.
1174  */
1175 ssize_t
1176 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1177                 unsigned long nr_segs, loff_t pos)
1178 {
1179         struct file *filp = iocb->ki_filp;
1180         ssize_t retval;
1181         unsigned long seg;
1182         size_t count;
1183         loff_t *ppos = &iocb->ki_pos;
1184
1185         count = 0;
1186         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1187         if (retval)
1188                 return retval;
1189
1190         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1191         if (filp->f_flags & O_DIRECT) {
1192                 loff_t size;
1193                 struct address_space *mapping;
1194                 struct inode *inode;
1195
1196                 mapping = filp->f_mapping;
1197                 inode = mapping->host;
1198                 retval = 0;
1199                 if (!count)
1200                         goto out; /* skip atime */
1201                 size = i_size_read(inode);
1202                 if (pos < size) {
1203                         retval = generic_file_direct_IO(READ, iocb,
1204                                                 iov, pos, nr_segs);
1205                         if (retval > 0)
1206                                 *ppos = pos + retval;
1207                 }
1208                 if (likely(retval != 0)) {
1209                         file_accessed(filp);
1210                         goto out;
1211                 }
1212         }
1213
1214         retval = 0;
1215         if (count) {
1216                 for (seg = 0; seg < nr_segs; seg++) {
1217                         read_descriptor_t desc;
1218
1219                         desc.written = 0;
1220                         desc.arg.buf = iov[seg].iov_base;
1221                         desc.count = iov[seg].iov_len;
1222                         if (desc.count == 0)
1223                                 continue;
1224                         desc.error = 0;
1225                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1226                         retval += desc.written;
1227                         if (desc.error) {
1228                                 retval = retval ?: desc.error;
1229                                 break;
1230                         }
1231                         if (desc.count > 0)
1232                                 break;
1233                 }
1234         }
1235 out:
1236         return retval;
1237 }
1238 EXPORT_SYMBOL(generic_file_aio_read);
1239
1240 static ssize_t
1241 do_readahead(struct address_space *mapping, struct file *filp,
1242              pgoff_t index, unsigned long nr)
1243 {
1244         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1245                 return -EINVAL;
1246
1247         force_page_cache_readahead(mapping, filp, index,
1248                                         max_sane_readahead(nr));
1249         return 0;
1250 }
1251
1252 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1253 {
1254         ssize_t ret;
1255         struct file *file;
1256
1257         ret = -EBADF;
1258         file = fget(fd);
1259         if (file) {
1260                 if (file->f_mode & FMODE_READ) {
1261                         struct address_space *mapping = file->f_mapping;
1262                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1263                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1264                         unsigned long len = end - start + 1;
1265                         ret = do_readahead(mapping, file, start, len);
1266                 }
1267                 fput(file);
1268         }
1269         return ret;
1270 }
1271
1272 #ifdef CONFIG_MMU
1273 /**
1274  * page_cache_read - adds requested page to the page cache if not already there
1275  * @file:       file to read
1276  * @offset:     page index
1277  *
1278  * This adds the requested page to the page cache if it isn't already there,
1279  * and schedules an I/O to read in its contents from disk.
1280  */
1281 static int page_cache_read(struct file *file, pgoff_t offset)
1282 {
1283         struct address_space *mapping = file->f_mapping;
1284         struct page *page; 
1285         int ret;
1286
1287         do {
1288                 page = page_cache_alloc_cold(mapping);
1289                 if (!page)
1290                         return -ENOMEM;
1291
1292                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1293                 if (ret == 0)
1294                         ret = mapping->a_ops->readpage(file, page);
1295                 else if (ret == -EEXIST)
1296                         ret = 0; /* losing race to add is OK */
1297
1298                 page_cache_release(page);
1299
1300         } while (ret == AOP_TRUNCATED_PAGE);
1301                 
1302         return ret;
1303 }
1304
1305 #define MMAP_LOTSAMISS  (100)
1306
1307 /**
1308  * filemap_fault - read in file data for page fault handling
1309  * @vma:        vma in which the fault was taken
1310  * @vmf:        struct vm_fault containing details of the fault
1311  *
1312  * filemap_fault() is invoked via the vma operations vector for a
1313  * mapped memory region to read in file data during a page fault.
1314  *
1315  * The goto's are kind of ugly, but this streamlines the normal case of having
1316  * it in the page cache, and handles the special cases reasonably without
1317  * having a lot of duplicated code.
1318  */
1319 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1320 {
1321         int error;
1322         struct file *file = vma->vm_file;
1323         struct address_space *mapping = file->f_mapping;
1324         struct file_ra_state *ra = &file->f_ra;
1325         struct inode *inode = mapping->host;
1326         struct page *page;
1327         pgoff_t size;
1328         int did_readaround = 0;
1329         int ret = 0;
1330
1331         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1332         if (vmf->pgoff >= size)
1333                 return VM_FAULT_SIGBUS;
1334
1335         /* If we don't want any read-ahead, don't bother */
1336         if (VM_RandomReadHint(vma))
1337                 goto no_cached_page;
1338
1339         /*
1340          * Do we have something in the page cache already?
1341          */
1342 retry_find:
1343         page = find_lock_page(mapping, vmf->pgoff);
1344         /*
1345          * For sequential accesses, we use the generic readahead logic.
1346          */
1347         if (VM_SequentialReadHint(vma)) {
1348                 if (!page) {
1349                         page_cache_sync_readahead(mapping, ra, file,
1350                                                            vmf->pgoff, 1);
1351                         page = find_lock_page(mapping, vmf->pgoff);
1352                         if (!page)
1353                                 goto no_cached_page;
1354                 }
1355                 if (PageReadahead(page)) {
1356                         page_cache_async_readahead(mapping, ra, file, page,
1357                                                            vmf->pgoff, 1);
1358                 }
1359         }
1360
1361         if (!page) {
1362                 unsigned long ra_pages;
1363
1364                 ra->mmap_miss++;
1365
1366                 /*
1367                  * Do we miss much more than hit in this file? If so,
1368                  * stop bothering with read-ahead. It will only hurt.
1369                  */
1370                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1371                         goto no_cached_page;
1372
1373                 /*
1374                  * To keep the pgmajfault counter straight, we need to
1375                  * check did_readaround, as this is an inner loop.
1376                  */
1377                 if (!did_readaround) {
1378                         ret = VM_FAULT_MAJOR;
1379                         count_vm_event(PGMAJFAULT);
1380                 }
1381                 did_readaround = 1;
1382                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1383                 if (ra_pages) {
1384                         pgoff_t start = 0;
1385
1386                         if (vmf->pgoff > ra_pages / 2)
1387                                 start = vmf->pgoff - ra_pages / 2;
1388                         do_page_cache_readahead(mapping, file, start, ra_pages);
1389                 }
1390                 page = find_lock_page(mapping, vmf->pgoff);
1391                 if (!page)
1392                         goto no_cached_page;
1393         }
1394
1395         if (!did_readaround)
1396                 ra->mmap_miss--;
1397
1398         /*
1399          * We have a locked page in the page cache, now we need to check
1400          * that it's up-to-date. If not, it is going to be due to an error.
1401          */
1402         if (unlikely(!PageUptodate(page)))
1403                 goto page_not_uptodate;
1404
1405         /* Must recheck i_size under page lock */
1406         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1407         if (unlikely(vmf->pgoff >= size)) {
1408                 unlock_page(page);
1409                 page_cache_release(page);
1410                 return VM_FAULT_SIGBUS;
1411         }
1412
1413         /*
1414          * Found the page and have a reference on it.
1415          */
1416         mark_page_accessed(page);
1417         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1418         vmf->page = page;
1419         return ret | VM_FAULT_LOCKED;
1420
1421 no_cached_page:
1422         /*
1423          * We're only likely to ever get here if MADV_RANDOM is in
1424          * effect.
1425          */
1426         error = page_cache_read(file, vmf->pgoff);
1427
1428         /*
1429          * The page we want has now been added to the page cache.
1430          * In the unlikely event that someone removed it in the
1431          * meantime, we'll just come back here and read it again.
1432          */
1433         if (error >= 0)
1434                 goto retry_find;
1435
1436         /*
1437          * An error return from page_cache_read can result if the
1438          * system is low on memory, or a problem occurs while trying
1439          * to schedule I/O.
1440          */
1441         if (error == -ENOMEM)
1442                 return VM_FAULT_OOM;
1443         return VM_FAULT_SIGBUS;
1444
1445 page_not_uptodate:
1446         /* IO error path */
1447         if (!did_readaround) {
1448                 ret = VM_FAULT_MAJOR;
1449                 count_vm_event(PGMAJFAULT);
1450         }
1451
1452         /*
1453          * Umm, take care of errors if the page isn't up-to-date.
1454          * Try to re-read it _once_. We do this synchronously,
1455          * because there really aren't any performance issues here
1456          * and we need to check for errors.
1457          */
1458         ClearPageError(page);
1459         error = mapping->a_ops->readpage(file, page);
1460         page_cache_release(page);
1461
1462         if (!error || error == AOP_TRUNCATED_PAGE)
1463                 goto retry_find;
1464
1465         /* Things didn't work out. Return zero to tell the mm layer so. */
1466         shrink_readahead_size_eio(file, ra);
1467         return VM_FAULT_SIGBUS;
1468 }
1469 EXPORT_SYMBOL(filemap_fault);
1470
1471 struct vm_operations_struct generic_file_vm_ops = {
1472         .fault          = filemap_fault,
1473 };
1474
1475 /* This is used for a general mmap of a disk file */
1476
1477 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1478 {
1479         struct address_space *mapping = file->f_mapping;
1480
1481         if (!mapping->a_ops->readpage)
1482                 return -ENOEXEC;
1483         file_accessed(file);
1484         vma->vm_ops = &generic_file_vm_ops;
1485         vma->vm_flags |= VM_CAN_NONLINEAR;
1486         return 0;
1487 }
1488
1489 /*
1490  * This is for filesystems which do not implement ->writepage.
1491  */
1492 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1493 {
1494         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1495                 return -EINVAL;
1496         return generic_file_mmap(file, vma);
1497 }
1498 #else
1499 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1500 {
1501         return -ENOSYS;
1502 }
1503 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1504 {
1505         return -ENOSYS;
1506 }
1507 #endif /* CONFIG_MMU */
1508
1509 EXPORT_SYMBOL(generic_file_mmap);
1510 EXPORT_SYMBOL(generic_file_readonly_mmap);
1511
1512 static struct page *__read_cache_page(struct address_space *mapping,
1513                                 pgoff_t index,
1514                                 int (*filler)(void *,struct page*),
1515                                 void *data)
1516 {
1517         struct page *page;
1518         int err;
1519 repeat:
1520         page = find_get_page(mapping, index);
1521         if (!page) {
1522                 page = page_cache_alloc_cold(mapping);
1523                 if (!page)
1524                         return ERR_PTR(-ENOMEM);
1525                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1526                 if (unlikely(err)) {
1527                         page_cache_release(page);
1528                         if (err == -EEXIST)
1529                                 goto repeat;
1530                         /* Presumably ENOMEM for radix tree node */
1531                         return ERR_PTR(err);
1532                 }
1533                 err = filler(data, page);
1534                 if (err < 0) {
1535                         page_cache_release(page);
1536                         page = ERR_PTR(err);
1537                 }
1538         }
1539         return page;
1540 }
1541
1542 /*
1543  * Same as read_cache_page, but don't wait for page to become unlocked
1544  * after submitting it to the filler.
1545  */
1546 struct page *read_cache_page_async(struct address_space *mapping,
1547                                 pgoff_t index,
1548                                 int (*filler)(void *,struct page*),
1549                                 void *data)
1550 {
1551         struct page *page;
1552         int err;
1553
1554 retry:
1555         page = __read_cache_page(mapping, index, filler, data);
1556         if (IS_ERR(page))
1557                 return page;
1558         if (PageUptodate(page))
1559                 goto out;
1560
1561         lock_page(page);
1562         if (!page->mapping) {
1563                 unlock_page(page);
1564                 page_cache_release(page);
1565                 goto retry;
1566         }
1567         if (PageUptodate(page)) {
1568                 unlock_page(page);
1569                 goto out;
1570         }
1571         err = filler(data, page);
1572         if (err < 0) {
1573                 page_cache_release(page);
1574                 return ERR_PTR(err);
1575         }
1576 out:
1577         mark_page_accessed(page);
1578         return page;
1579 }
1580 EXPORT_SYMBOL(read_cache_page_async);
1581
1582 /**
1583  * read_cache_page - read into page cache, fill it if needed
1584  * @mapping:    the page's address_space
1585  * @index:      the page index
1586  * @filler:     function to perform the read
1587  * @data:       destination for read data
1588  *
1589  * Read into the page cache. If a page already exists, and PageUptodate() is
1590  * not set, try to fill the page then wait for it to become unlocked.
1591  *
1592  * If the page does not get brought uptodate, return -EIO.
1593  */
1594 struct page *read_cache_page(struct address_space *mapping,
1595                                 pgoff_t index,
1596                                 int (*filler)(void *,struct page*),
1597                                 void *data)
1598 {
1599         struct page *page;
1600
1601         page = read_cache_page_async(mapping, index, filler, data);
1602         if (IS_ERR(page))
1603                 goto out;
1604         wait_on_page_locked(page);
1605         if (!PageUptodate(page)) {
1606                 page_cache_release(page);
1607                 page = ERR_PTR(-EIO);
1608         }
1609  out:
1610         return page;
1611 }
1612 EXPORT_SYMBOL(read_cache_page);
1613
1614 /*
1615  * The logic we want is
1616  *
1617  *      if suid or (sgid and xgrp)
1618  *              remove privs
1619  */
1620 int should_remove_suid(struct dentry *dentry)
1621 {
1622         mode_t mode = dentry->d_inode->i_mode;
1623         int kill = 0;
1624
1625         /* suid always must be killed */
1626         if (unlikely(mode & S_ISUID))
1627                 kill = ATTR_KILL_SUID;
1628
1629         /*
1630          * sgid without any exec bits is just a mandatory locking mark; leave
1631          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1632          */
1633         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1634                 kill |= ATTR_KILL_SGID;
1635
1636         if (unlikely(kill && !capable(CAP_FSETID)))
1637                 return kill;
1638
1639         return 0;
1640 }
1641 EXPORT_SYMBOL(should_remove_suid);
1642
1643 int __remove_suid(struct dentry *dentry, int kill)
1644 {
1645         struct iattr newattrs;
1646
1647         newattrs.ia_valid = ATTR_FORCE | kill;
1648         return notify_change(dentry, &newattrs);
1649 }
1650
1651 int remove_suid(struct dentry *dentry)
1652 {
1653         int killsuid = should_remove_suid(dentry);
1654         int killpriv = security_inode_need_killpriv(dentry);
1655         int error = 0;
1656
1657         if (killpriv < 0)
1658                 return killpriv;
1659         if (killpriv)
1660                 error = security_inode_killpriv(dentry);
1661         if (!error && killsuid)
1662                 error = __remove_suid(dentry, killsuid);
1663
1664         return error;
1665 }
1666 EXPORT_SYMBOL(remove_suid);
1667
1668 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1669                         const struct iovec *iov, size_t base, size_t bytes)
1670 {
1671         size_t copied = 0, left = 0;
1672
1673         while (bytes) {
1674                 char __user *buf = iov->iov_base + base;
1675                 int copy = min(bytes, iov->iov_len - base);
1676
1677                 base = 0;
1678                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1679                 copied += copy;
1680                 bytes -= copy;
1681                 vaddr += copy;
1682                 iov++;
1683
1684                 if (unlikely(left))
1685                         break;
1686         }
1687         return copied - left;
1688 }
1689
1690 /*
1691  * Copy as much as we can into the page and return the number of bytes which
1692  * were sucessfully copied.  If a fault is encountered then return the number of
1693  * bytes which were copied.
1694  */
1695 size_t iov_iter_copy_from_user_atomic(struct page *page,
1696                 struct iov_iter *i, unsigned long offset, size_t bytes)
1697 {
1698         char *kaddr;
1699         size_t copied;
1700
1701         BUG_ON(!in_atomic());
1702         kaddr = kmap_atomic(page, KM_USER0);
1703         if (likely(i->nr_segs == 1)) {
1704                 int left;
1705                 char __user *buf = i->iov->iov_base + i->iov_offset;
1706                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1707                                                         buf, bytes);
1708                 copied = bytes - left;
1709         } else {
1710                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1711                                                 i->iov, i->iov_offset, bytes);
1712         }
1713         kunmap_atomic(kaddr, KM_USER0);
1714
1715         return copied;
1716 }
1717 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1718
1719 /*
1720  * This has the same sideeffects and return value as
1721  * iov_iter_copy_from_user_atomic().
1722  * The difference is that it attempts to resolve faults.
1723  * Page must not be locked.
1724  */
1725 size_t iov_iter_copy_from_user(struct page *page,
1726                 struct iov_iter *i, unsigned long offset, size_t bytes)
1727 {
1728         char *kaddr;
1729         size_t copied;
1730
1731         kaddr = kmap(page);
1732         if (likely(i->nr_segs == 1)) {
1733                 int left;
1734                 char __user *buf = i->iov->iov_base + i->iov_offset;
1735                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1736                 copied = bytes - left;
1737         } else {
1738                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1739                                                 i->iov, i->iov_offset, bytes);
1740         }
1741         kunmap(page);
1742         return copied;
1743 }
1744 EXPORT_SYMBOL(iov_iter_copy_from_user);
1745
1746 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1747 {
1748         if (likely(i->nr_segs == 1)) {
1749                 i->iov_offset += bytes;
1750         } else {
1751                 const struct iovec *iov = i->iov;
1752                 size_t base = i->iov_offset;
1753
1754                 /*
1755                  * The !iov->iov_len check ensures we skip over unlikely
1756                  * zero-length segments.
1757                  */
1758                 while (bytes || !iov->iov_len) {
1759                         int copy = min(bytes, iov->iov_len - base);
1760
1761                         bytes -= copy;
1762                         base += copy;
1763                         if (iov->iov_len == base) {
1764                                 iov++;
1765                                 base = 0;
1766                         }
1767                 }
1768                 i->iov = iov;
1769                 i->iov_offset = base;
1770         }
1771 }
1772
1773 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1774 {
1775         BUG_ON(i->count < bytes);
1776
1777         __iov_iter_advance_iov(i, bytes);
1778         i->count -= bytes;
1779 }
1780 EXPORT_SYMBOL(iov_iter_advance);
1781
1782 /*
1783  * Fault in the first iovec of the given iov_iter, to a maximum length
1784  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1785  * accessed (ie. because it is an invalid address).
1786  *
1787  * writev-intensive code may want this to prefault several iovecs -- that
1788  * would be possible (callers must not rely on the fact that _only_ the
1789  * first iovec will be faulted with the current implementation).
1790  */
1791 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1792 {
1793         char __user *buf = i->iov->iov_base + i->iov_offset;
1794         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1795         return fault_in_pages_readable(buf, bytes);
1796 }
1797 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1798
1799 /*
1800  * Return the count of just the current iov_iter segment.
1801  */
1802 size_t iov_iter_single_seg_count(struct iov_iter *i)
1803 {
1804         const struct iovec *iov = i->iov;
1805         if (i->nr_segs == 1)
1806                 return i->count;
1807         else
1808                 return min(i->count, iov->iov_len - i->iov_offset);
1809 }
1810 EXPORT_SYMBOL(iov_iter_single_seg_count);
1811
1812 /*
1813  * Performs necessary checks before doing a write
1814  *
1815  * Can adjust writing position or amount of bytes to write.
1816  * Returns appropriate error code that caller should return or
1817  * zero in case that write should be allowed.
1818  */
1819 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1820 {
1821         struct inode *inode = file->f_mapping->host;
1822         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1823
1824         if (unlikely(*pos < 0))
1825                 return -EINVAL;
1826
1827         if (!isblk) {
1828                 /* FIXME: this is for backwards compatibility with 2.4 */
1829                 if (file->f_flags & O_APPEND)
1830                         *pos = i_size_read(inode);
1831
1832                 if (limit != RLIM_INFINITY) {
1833                         if (*pos >= limit) {
1834                                 send_sig(SIGXFSZ, current, 0);
1835                                 return -EFBIG;
1836                         }
1837                         if (*count > limit - (typeof(limit))*pos) {
1838                                 *count = limit - (typeof(limit))*pos;
1839                         }
1840                 }
1841         }
1842
1843         /*
1844          * LFS rule
1845          */
1846         if (unlikely(*pos + *count > MAX_NON_LFS &&
1847                                 !(file->f_flags & O_LARGEFILE))) {
1848                 if (*pos >= MAX_NON_LFS) {
1849                         return -EFBIG;
1850                 }
1851                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1852                         *count = MAX_NON_LFS - (unsigned long)*pos;
1853                 }
1854         }
1855
1856         /*
1857          * Are we about to exceed the fs block limit ?
1858          *
1859          * If we have written data it becomes a short write.  If we have
1860          * exceeded without writing data we send a signal and return EFBIG.
1861          * Linus frestrict idea will clean these up nicely..
1862          */
1863         if (likely(!isblk)) {
1864                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1865                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1866                                 return -EFBIG;
1867                         }
1868                         /* zero-length writes at ->s_maxbytes are OK */
1869                 }
1870
1871                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1872                         *count = inode->i_sb->s_maxbytes - *pos;
1873         } else {
1874 #ifdef CONFIG_BLOCK
1875                 loff_t isize;
1876                 if (bdev_read_only(I_BDEV(inode)))
1877                         return -EPERM;
1878                 isize = i_size_read(inode);
1879                 if (*pos >= isize) {
1880                         if (*count || *pos > isize)
1881                                 return -ENOSPC;
1882                 }
1883
1884                 if (*pos + *count > isize)
1885                         *count = isize - *pos;
1886 #else
1887                 return -EPERM;
1888 #endif
1889         }
1890         return 0;
1891 }
1892 EXPORT_SYMBOL(generic_write_checks);
1893
1894 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1895                                 loff_t pos, unsigned len, unsigned flags,
1896                                 struct page **pagep, void **fsdata)
1897 {
1898         const struct address_space_operations *aops = mapping->a_ops;
1899
1900         if (aops->write_begin) {
1901                 return aops->write_begin(file, mapping, pos, len, flags,
1902                                                         pagep, fsdata);
1903         } else {
1904                 int ret;
1905                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1906                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1907                 struct inode *inode = mapping->host;
1908                 struct page *page;
1909 again:
1910                 page = __grab_cache_page(mapping, index);
1911                 *pagep = page;
1912                 if (!page)
1913                         return -ENOMEM;
1914
1915                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1916                         /*
1917                          * There is no way to resolve a short write situation
1918                          * for a !Uptodate page (except by double copying in
1919                          * the caller done by generic_perform_write_2copy).
1920                          *
1921                          * Instead, we have to bring it uptodate here.
1922                          */
1923                         ret = aops->readpage(file, page);
1924                         page_cache_release(page);
1925                         if (ret) {
1926                                 if (ret == AOP_TRUNCATED_PAGE)
1927                                         goto again;
1928                                 return ret;
1929                         }
1930                         goto again;
1931                 }
1932
1933                 ret = aops->prepare_write(file, page, offset, offset+len);
1934                 if (ret) {
1935                         unlock_page(page);
1936                         page_cache_release(page);
1937                         if (pos + len > inode->i_size)
1938                                 vmtruncate(inode, inode->i_size);
1939                 }
1940                 return ret;
1941         }
1942 }
1943 EXPORT_SYMBOL(pagecache_write_begin);
1944
1945 int pagecache_write_end(struct file *file, struct address_space *mapping,
1946                                 loff_t pos, unsigned len, unsigned copied,
1947                                 struct page *page, void *fsdata)
1948 {
1949         const struct address_space_operations *aops = mapping->a_ops;
1950         int ret;
1951
1952         if (aops->write_end) {
1953                 mark_page_accessed(page);
1954                 ret = aops->write_end(file, mapping, pos, len, copied,
1955                                                         page, fsdata);
1956         } else {
1957                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1958                 struct inode *inode = mapping->host;
1959
1960                 flush_dcache_page(page);
1961                 ret = aops->commit_write(file, page, offset, offset+len);
1962                 unlock_page(page);
1963                 mark_page_accessed(page);
1964                 page_cache_release(page);
1965
1966                 if (ret < 0) {
1967                         if (pos + len > inode->i_size)
1968                                 vmtruncate(inode, inode->i_size);
1969                 } else if (ret > 0)
1970                         ret = min_t(size_t, copied, ret);
1971                 else
1972                         ret = copied;
1973         }
1974
1975         return ret;
1976 }
1977 EXPORT_SYMBOL(pagecache_write_end);
1978
1979 ssize_t
1980 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1981                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1982                 size_t count, size_t ocount)
1983 {
1984         struct file     *file = iocb->ki_filp;
1985         struct address_space *mapping = file->f_mapping;
1986         struct inode    *inode = mapping->host;
1987         ssize_t         written;
1988
1989         if (count != ocount)
1990                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1991
1992         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1993         if (written > 0) {
1994                 loff_t end = pos + written;
1995                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1996                         i_size_write(inode,  end);
1997                         mark_inode_dirty(inode);
1998                 }
1999                 *ppos = end;
2000         }
2001
2002         /*
2003          * Sync the fs metadata but not the minor inode changes and
2004          * of course not the data as we did direct DMA for the IO.
2005          * i_mutex is held, which protects generic_osync_inode() from
2006          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2007          */
2008         if ((written >= 0 || written == -EIOCBQUEUED) &&
2009             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2010                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2011                 if (err < 0)
2012                         written = err;
2013         }
2014         return written;
2015 }
2016 EXPORT_SYMBOL(generic_file_direct_write);
2017
2018 /*
2019  * Find or create a page at the given pagecache position. Return the locked
2020  * page. This function is specifically for buffered writes.
2021  */
2022 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2023 {
2024         int status;
2025         struct page *page;
2026 repeat:
2027         page = find_lock_page(mapping, index);
2028         if (likely(page))
2029                 return page;
2030
2031         page = page_cache_alloc(mapping);
2032         if (!page)
2033                 return NULL;
2034         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2035         if (unlikely(status)) {
2036                 page_cache_release(page);
2037                 if (status == -EEXIST)
2038                         goto repeat;
2039                 return NULL;
2040         }
2041         return page;
2042 }
2043 EXPORT_SYMBOL(__grab_cache_page);
2044
2045 static ssize_t generic_perform_write_2copy(struct file *file,
2046                                 struct iov_iter *i, loff_t pos)
2047 {
2048         struct address_space *mapping = file->f_mapping;
2049         const struct address_space_operations *a_ops = mapping->a_ops;
2050         struct inode *inode = mapping->host;
2051         long status = 0;
2052         ssize_t written = 0;
2053
2054         do {
2055                 struct page *src_page;
2056                 struct page *page;
2057                 pgoff_t index;          /* Pagecache index for current page */
2058                 unsigned long offset;   /* Offset into pagecache page */
2059                 unsigned long bytes;    /* Bytes to write to page */
2060                 size_t copied;          /* Bytes copied from user */
2061
2062                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2063                 index = pos >> PAGE_CACHE_SHIFT;
2064                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2065                                                 iov_iter_count(i));
2066
2067                 /*
2068                  * a non-NULL src_page indicates that we're doing the
2069                  * copy via get_user_pages and kmap.
2070                  */
2071                 src_page = NULL;
2072
2073                 /*
2074                  * Bring in the user page that we will copy from _first_.
2075                  * Otherwise there's a nasty deadlock on copying from the
2076                  * same page as we're writing to, without it being marked
2077                  * up-to-date.
2078                  *
2079                  * Not only is this an optimisation, but it is also required
2080                  * to check that the address is actually valid, when atomic
2081                  * usercopies are used, below.
2082                  */
2083                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2084                         status = -EFAULT;
2085                         break;
2086                 }
2087
2088                 page = __grab_cache_page(mapping, index);
2089                 if (!page) {
2090                         status = -ENOMEM;
2091                         break;
2092                 }
2093
2094                 /*
2095                  * non-uptodate pages cannot cope with short copies, and we
2096                  * cannot take a pagefault with the destination page locked.
2097                  * So pin the source page to copy it.
2098                  */
2099                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2100                         unlock_page(page);
2101
2102                         src_page = alloc_page(GFP_KERNEL);
2103                         if (!src_page) {
2104                                 page_cache_release(page);
2105                                 status = -ENOMEM;
2106                                 break;
2107                         }
2108
2109                         /*
2110                          * Cannot get_user_pages with a page locked for the
2111                          * same reason as we can't take a page fault with a
2112                          * page locked (as explained below).
2113                          */
2114                         copied = iov_iter_copy_from_user(src_page, i,
2115                                                                 offset, bytes);
2116                         if (unlikely(copied == 0)) {
2117                                 status = -EFAULT;
2118                                 page_cache_release(page);
2119                                 page_cache_release(src_page);
2120                                 break;
2121                         }
2122                         bytes = copied;
2123
2124                         lock_page(page);
2125                         /*
2126                          * Can't handle the page going uptodate here, because
2127                          * that means we would use non-atomic usercopies, which
2128                          * zero out the tail of the page, which can cause
2129                          * zeroes to become transiently visible. We could just
2130                          * use a non-zeroing copy, but the APIs aren't too
2131                          * consistent.
2132                          */
2133                         if (unlikely(!page->mapping || PageUptodate(page))) {
2134                                 unlock_page(page);
2135                                 page_cache_release(page);
2136                                 page_cache_release(src_page);
2137                                 continue;
2138                         }
2139                 }
2140
2141                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2142                 if (unlikely(status))
2143                         goto fs_write_aop_error;
2144
2145                 if (!src_page) {
2146                         /*
2147                          * Must not enter the pagefault handler here, because
2148                          * we hold the page lock, so we might recursively
2149                          * deadlock on the same lock, or get an ABBA deadlock
2150                          * against a different lock, or against the mmap_sem
2151                          * (which nests outside the page lock).  So increment
2152                          * preempt count, and use _atomic usercopies.
2153                          *
2154                          * The page is uptodate so we are OK to encounter a
2155                          * short copy: if unmodified parts of the page are
2156                          * marked dirty and written out to disk, it doesn't
2157                          * really matter.
2158                          */
2159                         pagefault_disable();
2160                         copied = iov_iter_copy_from_user_atomic(page, i,
2161                                                                 offset, bytes);
2162                         pagefault_enable();
2163                 } else {
2164                         void *src, *dst;
2165                         src = kmap_atomic(src_page, KM_USER0);
2166                         dst = kmap_atomic(page, KM_USER1);
2167                         memcpy(dst + offset, src + offset, bytes);
2168                         kunmap_atomic(dst, KM_USER1);
2169                         kunmap_atomic(src, KM_USER0);
2170                         copied = bytes;
2171                 }
2172                 flush_dcache_page(page);
2173
2174                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2175                 if (unlikely(status < 0))
2176                         goto fs_write_aop_error;
2177                 if (unlikely(status > 0)) /* filesystem did partial write */
2178                         copied = min_t(size_t, copied, status);
2179
2180                 unlock_page(page);
2181                 mark_page_accessed(page);
2182                 page_cache_release(page);
2183                 if (src_page)
2184                         page_cache_release(src_page);
2185
2186                 iov_iter_advance(i, copied);
2187                 pos += copied;
2188                 written += copied;
2189
2190                 balance_dirty_pages_ratelimited(mapping);
2191                 cond_resched();
2192                 continue;
2193
2194 fs_write_aop_error:
2195                 unlock_page(page);
2196                 page_cache_release(page);
2197                 if (src_page)
2198                         page_cache_release(src_page);
2199
2200                 /*
2201                  * prepare_write() may have instantiated a few blocks
2202                  * outside i_size.  Trim these off again. Don't need
2203                  * i_size_read because we hold i_mutex.
2204                  */
2205                 if (pos + bytes > inode->i_size)
2206                         vmtruncate(inode, inode->i_size);
2207                 break;
2208         } while (iov_iter_count(i));
2209
2210         return written ? written : status;
2211 }
2212
2213 static ssize_t generic_perform_write(struct file *file,
2214                                 struct iov_iter *i, loff_t pos)
2215 {
2216         struct address_space *mapping = file->f_mapping;
2217         const struct address_space_operations *a_ops = mapping->a_ops;
2218         long status = 0;
2219         ssize_t written = 0;
2220         unsigned int flags = 0;
2221
2222         /*
2223          * Copies from kernel address space cannot fail (NFSD is a big user).
2224          */
2225         if (segment_eq(get_fs(), KERNEL_DS))
2226                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2227
2228         do {
2229                 struct page *page;
2230                 pgoff_t index;          /* Pagecache index for current page */
2231                 unsigned long offset;   /* Offset into pagecache page */
2232                 unsigned long bytes;    /* Bytes to write to page */
2233                 size_t copied;          /* Bytes copied from user */
2234                 void *fsdata;
2235
2236                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2237                 index = pos >> PAGE_CACHE_SHIFT;
2238                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2239                                                 iov_iter_count(i));
2240
2241 again:
2242
2243                 /*
2244                  * Bring in the user page that we will copy from _first_.
2245                  * Otherwise there's a nasty deadlock on copying from the
2246                  * same page as we're writing to, without it being marked
2247                  * up-to-date.
2248                  *
2249                  * Not only is this an optimisation, but it is also required
2250                  * to check that the address is actually valid, when atomic
2251                  * usercopies are used, below.
2252                  */
2253                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2254                         status = -EFAULT;
2255                         break;
2256                 }
2257
2258                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2259                                                 &page, &fsdata);
2260                 if (unlikely(status))
2261                         break;
2262
2263                 pagefault_disable();
2264                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2265                 pagefault_enable();
2266                 flush_dcache_page(page);
2267
2268                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2269                                                 page, fsdata);
2270                 if (unlikely(status < 0))
2271                         break;
2272                 copied = status;
2273
2274                 cond_resched();
2275
2276                 iov_iter_advance(i, copied);
2277                 if (unlikely(copied == 0)) {
2278                         /*
2279                          * If we were unable to copy any data at all, we must
2280                          * fall back to a single segment length write.
2281                          *
2282                          * If we didn't fallback here, we could livelock
2283                          * because not all segments in the iov can be copied at
2284                          * once without a pagefault.
2285                          */
2286                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2287                                                 iov_iter_single_seg_count(i));
2288                         goto again;
2289                 }
2290                 pos += copied;
2291                 written += copied;
2292
2293                 balance_dirty_pages_ratelimited(mapping);
2294
2295         } while (iov_iter_count(i));
2296
2297         return written ? written : status;
2298 }
2299
2300 ssize_t
2301 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2302                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2303                 size_t count, ssize_t written)
2304 {
2305         struct file *file = iocb->ki_filp;
2306         struct address_space *mapping = file->f_mapping;
2307         const struct address_space_operations *a_ops = mapping->a_ops;
2308         struct inode *inode = mapping->host;
2309         ssize_t status;
2310         struct iov_iter i;
2311
2312         iov_iter_init(&i, iov, nr_segs, count, written);
2313         if (a_ops->write_begin)
2314                 status = generic_perform_write(file, &i, pos);
2315         else
2316                 status = generic_perform_write_2copy(file, &i, pos);
2317
2318         if (likely(status >= 0)) {
2319                 written += status;
2320                 *ppos = pos + status;
2321
2322                 /*
2323                  * For now, when the user asks for O_SYNC, we'll actually give
2324                  * O_DSYNC
2325                  */
2326                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2327                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2328                                 status = generic_osync_inode(inode, mapping,
2329                                                 OSYNC_METADATA|OSYNC_DATA);
2330                 }
2331         }
2332         
2333         /*
2334          * If we get here for O_DIRECT writes then we must have fallen through
2335          * to buffered writes (block instantiation inside i_size).  So we sync
2336          * the file data here, to try to honour O_DIRECT expectations.
2337          */
2338         if (unlikely(file->f_flags & O_DIRECT) && written)
2339                 status = filemap_write_and_wait(mapping);
2340
2341         return written ? written : status;
2342 }
2343 EXPORT_SYMBOL(generic_file_buffered_write);
2344
2345 static ssize_t
2346 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2347                                 unsigned long nr_segs, loff_t *ppos)
2348 {
2349         struct file *file = iocb->ki_filp;
2350         struct address_space * mapping = file->f_mapping;
2351         size_t ocount;          /* original count */
2352         size_t count;           /* after file limit checks */
2353         struct inode    *inode = mapping->host;
2354         loff_t          pos;
2355         ssize_t         written;
2356         ssize_t         err;
2357
2358         ocount = 0;
2359         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2360         if (err)
2361                 return err;
2362
2363         count = ocount;
2364         pos = *ppos;
2365
2366         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2367
2368         /* We can write back this queue in page reclaim */
2369         current->backing_dev_info = mapping->backing_dev_info;
2370         written = 0;
2371
2372         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2373         if (err)
2374                 goto out;
2375
2376         if (count == 0)
2377                 goto out;
2378
2379         err = remove_suid(file->f_path.dentry);
2380         if (err)
2381                 goto out;
2382
2383         file_update_time(file);
2384
2385         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2386         if (unlikely(file->f_flags & O_DIRECT)) {
2387                 loff_t endbyte;
2388                 ssize_t written_buffered;
2389
2390                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2391                                                         ppos, count, ocount);
2392                 if (written < 0 || written == count)
2393                         goto out;
2394                 /*
2395                  * direct-io write to a hole: fall through to buffered I/O
2396                  * for completing the rest of the request.
2397                  */
2398                 pos += written;
2399                 count -= written;
2400                 written_buffered = generic_file_buffered_write(iocb, iov,
2401                                                 nr_segs, pos, ppos, count,
2402                                                 written);
2403                 /*
2404                  * If generic_file_buffered_write() retuned a synchronous error
2405                  * then we want to return the number of bytes which were
2406                  * direct-written, or the error code if that was zero.  Note
2407                  * that this differs from normal direct-io semantics, which
2408                  * will return -EFOO even if some bytes were written.
2409                  */
2410                 if (written_buffered < 0) {
2411                         err = written_buffered;
2412                         goto out;
2413                 }
2414
2415                 /*
2416                  * We need to ensure that the page cache pages are written to
2417                  * disk and invalidated to preserve the expected O_DIRECT
2418                  * semantics.
2419                  */
2420                 endbyte = pos + written_buffered - written - 1;
2421                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2422                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2423                                             SYNC_FILE_RANGE_WRITE|
2424                                             SYNC_FILE_RANGE_WAIT_AFTER);
2425                 if (err == 0) {
2426                         written = written_buffered;
2427                         invalidate_mapping_pages(mapping,
2428                                                  pos >> PAGE_CACHE_SHIFT,
2429                                                  endbyte >> PAGE_CACHE_SHIFT);
2430                 } else {
2431                         /*
2432                          * We don't know how much we wrote, so just return
2433                          * the number of bytes which were direct-written
2434                          */
2435                 }
2436         } else {
2437                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2438                                 pos, ppos, count, written);
2439         }
2440 out:
2441         current->backing_dev_info = NULL;
2442         return written ? written : err;
2443 }
2444
2445 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2446                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2447 {
2448         struct file *file = iocb->ki_filp;
2449         struct address_space *mapping = file->f_mapping;
2450         struct inode *inode = mapping->host;
2451         ssize_t ret;
2452
2453         BUG_ON(iocb->ki_pos != pos);
2454
2455         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2456                         &iocb->ki_pos);
2457
2458         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2459                 ssize_t err;
2460
2461                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2462                 if (err < 0)
2463                         ret = err;
2464         }
2465         return ret;
2466 }
2467 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2468
2469 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2470                 unsigned long nr_segs, loff_t pos)
2471 {
2472         struct file *file = iocb->ki_filp;
2473         struct address_space *mapping = file->f_mapping;
2474         struct inode *inode = mapping->host;
2475         ssize_t ret;
2476
2477         BUG_ON(iocb->ki_pos != pos);
2478
2479         mutex_lock(&inode->i_mutex);
2480         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2481                         &iocb->ki_pos);
2482         mutex_unlock(&inode->i_mutex);
2483
2484         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2485                 ssize_t err;
2486
2487                 err = sync_page_range(inode, mapping, pos, ret);
2488                 if (err < 0)
2489                         ret = err;
2490         }
2491         return ret;
2492 }
2493 EXPORT_SYMBOL(generic_file_aio_write);
2494
2495 /*
2496  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2497  * went wrong during pagecache shootdown.
2498  */
2499 static ssize_t
2500 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2501         loff_t offset, unsigned long nr_segs)
2502 {
2503         struct file *file = iocb->ki_filp;
2504         struct address_space *mapping = file->f_mapping;
2505         ssize_t retval;
2506         size_t write_len;
2507         pgoff_t end = 0; /* silence gcc */
2508
2509         /*
2510          * If it's a write, unmap all mmappings of the file up-front.  This
2511          * will cause any pte dirty bits to be propagated into the pageframes
2512          * for the subsequent filemap_write_and_wait().
2513          */
2514         if (rw == WRITE) {
2515                 write_len = iov_length(iov, nr_segs);
2516                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2517                 if (mapping_mapped(mapping))
2518                         unmap_mapping_range(mapping, offset, write_len, 0);
2519         }
2520
2521         retval = filemap_write_and_wait(mapping);
2522         if (retval)
2523                 goto out;
2524
2525         /*
2526          * After a write we want buffered reads to be sure to go to disk to get
2527          * the new data.  We invalidate clean cached page from the region we're
2528          * about to write.  We do this *before* the write so that we can return
2529          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2530          */
2531         if (rw == WRITE && mapping->nrpages) {
2532                 retval = invalidate_inode_pages2_range(mapping,
2533                                         offset >> PAGE_CACHE_SHIFT, end);
2534                 if (retval)
2535                         goto out;
2536         }
2537
2538         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2539
2540         /*
2541          * Finally, try again to invalidate clean pages which might have been
2542          * cached by non-direct readahead, or faulted in by get_user_pages()
2543          * if the source of the write was an mmap'ed region of the file
2544          * we're writing.  Either one is a pretty crazy thing to do,
2545          * so we don't support it 100%.  If this invalidation
2546          * fails, tough, the write still worked...
2547          */
2548         if (rw == WRITE && mapping->nrpages) {
2549                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2550         }
2551 out:
2552         return retval;
2553 }
2554
2555 /**
2556  * try_to_release_page() - release old fs-specific metadata on a page
2557  *
2558  * @page: the page which the kernel is trying to free
2559  * @gfp_mask: memory allocation flags (and I/O mode)
2560  *
2561  * The address_space is to try to release any data against the page
2562  * (presumably at page->private).  If the release was successful, return `1'.
2563  * Otherwise return zero.
2564  *
2565  * The @gfp_mask argument specifies whether I/O may be performed to release
2566  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2567  *
2568  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2569  */
2570 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2571 {
2572         struct address_space * const mapping = page->mapping;
2573
2574         BUG_ON(!PageLocked(page));
2575         if (PageWriteback(page))
2576                 return 0;
2577
2578         if (mapping && mapping->a_ops->releasepage)
2579                 return mapping->a_ops->releasepage(page, gfp_mask);
2580         return try_to_free_buffers(page);
2581 }
2582
2583 EXPORT_SYMBOL(try_to_release_page);