[MTD] [NAND] Add Blackfin BF52x support in bf5xx_nand driver
[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/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include "internal.h"
36
37 /*
38  * FIXME: remove all knowledge of the buffer layer from the core VM
39  */
40 #include <linux/buffer_head.h> /* for generic_osync_inode */
41
42 #include <asm/mman.h>
43
44 static ssize_t
45 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
46         loff_t offset, unsigned long nr_segs);
47
48 /*
49  * Shared mappings implemented 30.11.1994. It's not fully working yet,
50  * though.
51  *
52  * Shared mappings now work. 15.8.1995  Bruno.
53  *
54  * finished 'unifying' the page and buffer cache and SMP-threaded the
55  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
56  *
57  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
58  */
59
60 /*
61  * Lock ordering:
62  *
63  *  ->i_mmap_lock               (vmtruncate)
64  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
65  *      ->swap_lock             (exclusive_swap_page, others)
66  *        ->mapping->tree_lock
67  *          ->zone.lock
68  *
69  *  ->i_mutex
70  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
71  *
72  *  ->mmap_sem
73  *    ->i_mmap_lock
74  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
75  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
76  *
77  *  ->mmap_sem
78  *    ->lock_page               (access_process_vm)
79  *
80  *  ->i_mutex                   (generic_file_buffered_write)
81  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
82  *
83  *  ->i_mutex
84  *    ->i_alloc_sem             (various)
85  *
86  *  ->inode_lock
87  *    ->sb_lock                 (fs/fs-writeback.c)
88  *    ->mapping->tree_lock      (__sync_single_inode)
89  *
90  *  ->i_mmap_lock
91  *    ->anon_vma.lock           (vma_adjust)
92  *
93  *  ->anon_vma.lock
94  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
95  *
96  *  ->page_table_lock or pte_lock
97  *    ->swap_lock               (try_to_unmap_one)
98  *    ->private_lock            (try_to_unmap_one)
99  *    ->tree_lock               (try_to_unmap_one)
100  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
101  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
102  *    ->private_lock            (page_remove_rmap->set_page_dirty)
103  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
104  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (zap_pte_range->set_page_dirty)
106  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
107  *
108  *  ->task->proc_lock
109  *    ->dcache_lock             (proc_pid_lookup)
110  */
111
112 /*
113  * Remove a page from the page cache and free it. Caller has to make
114  * sure the page is locked and that nobody else uses it - or that usage
115  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
116  */
117 void __remove_from_page_cache(struct page *page)
118 {
119         struct address_space *mapping = page->mapping;
120
121         radix_tree_delete(&mapping->page_tree, page->index);
122         page->mapping = NULL;
123         mapping->nrpages--;
124         __dec_zone_page_state(page, NR_FILE_PAGES);
125         BUG_ON(page_mapped(page));
126 }
127
128 void remove_from_page_cache(struct page *page)
129 {
130         struct address_space *mapping = page->mapping;
131
132         BUG_ON(!PageLocked(page));
133
134         write_lock_irq(&mapping->tree_lock);
135         __remove_from_page_cache(page);
136         write_unlock_irq(&mapping->tree_lock);
137 }
138
139 static int sync_page(void *word)
140 {
141         struct address_space *mapping;
142         struct page *page;
143
144         page = container_of((unsigned long *)word, struct page, flags);
145
146         /*
147          * page_mapping() is being called without PG_locked held.
148          * Some knowledge of the state and use of the page is used to
149          * reduce the requirements down to a memory barrier.
150          * The danger here is of a stale page_mapping() return value
151          * indicating a struct address_space different from the one it's
152          * associated with when it is associated with one.
153          * After smp_mb(), it's either the correct page_mapping() for
154          * the page, or an old page_mapping() and the page's own
155          * page_mapping() has gone NULL.
156          * The ->sync_page() address_space operation must tolerate
157          * page_mapping() going NULL. By an amazing coincidence,
158          * this comes about because none of the users of the page
159          * in the ->sync_page() methods make essential use of the
160          * page_mapping(), merely passing the page down to the backing
161          * device's unplug functions when it's non-NULL, which in turn
162          * ignore it for all cases but swap, where only page_private(page) is
163          * of interest. When page_mapping() does go NULL, the entire
164          * call stack gracefully ignores the page and returns.
165          * -- wli
166          */
167         smp_mb();
168         mapping = page_mapping(page);
169         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
170                 mapping->a_ops->sync_page(page);
171         io_schedule();
172         return 0;
173 }
174
175 /**
176  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
177  * @mapping:    address space structure to write
178  * @start:      offset in bytes where the range starts
179  * @end:        offset in bytes where the range ends (inclusive)
180  * @sync_mode:  enable synchronous operation
181  *
182  * Start writeback against all of a mapping's dirty pages that lie
183  * within the byte offsets <start, end> inclusive.
184  *
185  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
186  * opposed to a regular memory cleansing writeback.  The difference between
187  * these two operations is that if a dirty page/buffer is encountered, it must
188  * be waited upon, and not just skipped over.
189  */
190 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
191                                 loff_t end, int sync_mode)
192 {
193         int ret;
194         struct writeback_control wbc = {
195                 .sync_mode = sync_mode,
196                 .nr_to_write = mapping->nrpages * 2,
197                 .range_start = start,
198                 .range_end = end,
199         };
200
201         if (!mapping_cap_writeback_dirty(mapping))
202                 return 0;
203
204         ret = do_writepages(mapping, &wbc);
205         return ret;
206 }
207
208 static inline int __filemap_fdatawrite(struct address_space *mapping,
209         int sync_mode)
210 {
211         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
212 }
213
214 int filemap_fdatawrite(struct address_space *mapping)
215 {
216         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
217 }
218 EXPORT_SYMBOL(filemap_fdatawrite);
219
220 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
221                                 loff_t end)
222 {
223         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
224 }
225
226 /**
227  * filemap_flush - mostly a non-blocking flush
228  * @mapping:    target address_space
229  *
230  * This is a mostly non-blocking flush.  Not suitable for data-integrity
231  * purposes - I/O may not be started against all dirty pages.
232  */
233 int filemap_flush(struct address_space *mapping)
234 {
235         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
236 }
237 EXPORT_SYMBOL(filemap_flush);
238
239 /**
240  * wait_on_page_writeback_range - wait for writeback to complete
241  * @mapping:    target address_space
242  * @start:      beginning page index
243  * @end:        ending page index
244  *
245  * Wait for writeback to complete against pages indexed by start->end
246  * inclusive
247  */
248 int wait_on_page_writeback_range(struct address_space *mapping,
249                                 pgoff_t start, pgoff_t end)
250 {
251         struct pagevec pvec;
252         int nr_pages;
253         int ret = 0;
254         pgoff_t index;
255
256         if (end < start)
257                 return 0;
258
259         pagevec_init(&pvec, 0);
260         index = start;
261         while ((index <= end) &&
262                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
263                         PAGECACHE_TAG_WRITEBACK,
264                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
265                 unsigned i;
266
267                 for (i = 0; i < nr_pages; i++) {
268                         struct page *page = pvec.pages[i];
269
270                         /* until radix tree lookup accepts end_index */
271                         if (page->index > end)
272                                 continue;
273
274                         wait_on_page_writeback(page);
275                         if (PageError(page))
276                                 ret = -EIO;
277                 }
278                 pagevec_release(&pvec);
279                 cond_resched();
280         }
281
282         /* Check for outstanding write errors */
283         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
284                 ret = -ENOSPC;
285         if (test_and_clear_bit(AS_EIO, &mapping->flags))
286                 ret = -EIO;
287
288         return ret;
289 }
290
291 /**
292  * sync_page_range - write and wait on all pages in the passed range
293  * @inode:      target inode
294  * @mapping:    target address_space
295  * @pos:        beginning offset in pages to write
296  * @count:      number of bytes to write
297  *
298  * Write and wait upon all the pages in the passed range.  This is a "data
299  * integrity" operation.  It waits upon in-flight writeout before starting and
300  * waiting upon new writeout.  If there was an IO error, return it.
301  *
302  * We need to re-take i_mutex during the generic_osync_inode list walk because
303  * it is otherwise livelockable.
304  */
305 int sync_page_range(struct inode *inode, struct address_space *mapping,
306                         loff_t pos, loff_t count)
307 {
308         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
309         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
310         int ret;
311
312         if (!mapping_cap_writeback_dirty(mapping) || !count)
313                 return 0;
314         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
315         if (ret == 0) {
316                 mutex_lock(&inode->i_mutex);
317                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
318                 mutex_unlock(&inode->i_mutex);
319         }
320         if (ret == 0)
321                 ret = wait_on_page_writeback_range(mapping, start, end);
322         return ret;
323 }
324 EXPORT_SYMBOL(sync_page_range);
325
326 /**
327  * sync_page_range_nolock
328  * @inode:      target inode
329  * @mapping:    target address_space
330  * @pos:        beginning offset in pages to write
331  * @count:      number of bytes to write
332  *
333  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
334  * as it forces O_SYNC writers to different parts of the same file
335  * to be serialised right until io completion.
336  */
337 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
338                            loff_t pos, loff_t count)
339 {
340         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
341         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
342         int ret;
343
344         if (!mapping_cap_writeback_dirty(mapping) || !count)
345                 return 0;
346         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
347         if (ret == 0)
348                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
349         if (ret == 0)
350                 ret = wait_on_page_writeback_range(mapping, start, end);
351         return ret;
352 }
353 EXPORT_SYMBOL(sync_page_range_nolock);
354
355 /**
356  * filemap_fdatawait - wait for all under-writeback pages to complete
357  * @mapping: address space structure to wait for
358  *
359  * Walk the list of under-writeback pages of the given address space
360  * and wait for all of them.
361  */
362 int filemap_fdatawait(struct address_space *mapping)
363 {
364         loff_t i_size = i_size_read(mapping->host);
365
366         if (i_size == 0)
367                 return 0;
368
369         return wait_on_page_writeback_range(mapping, 0,
370                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
371 }
372 EXPORT_SYMBOL(filemap_fdatawait);
373
374 int filemap_write_and_wait(struct address_space *mapping)
375 {
376         int err = 0;
377
378         if (mapping->nrpages) {
379                 err = filemap_fdatawrite(mapping);
380                 /*
381                  * Even if the above returned error, the pages may be
382                  * written partially (e.g. -ENOSPC), so we wait for it.
383                  * But the -EIO is special case, it may indicate the worst
384                  * thing (e.g. bug) happened, so we avoid waiting for it.
385                  */
386                 if (err != -EIO) {
387                         int err2 = filemap_fdatawait(mapping);
388                         if (!err)
389                                 err = err2;
390                 }
391         }
392         return err;
393 }
394 EXPORT_SYMBOL(filemap_write_and_wait);
395
396 /**
397  * filemap_write_and_wait_range - write out & wait on a file range
398  * @mapping:    the address_space for the pages
399  * @lstart:     offset in bytes where the range starts
400  * @lend:       offset in bytes where the range ends (inclusive)
401  *
402  * Write out and wait upon file offsets lstart->lend, inclusive.
403  *
404  * Note that `lend' is inclusive (describes the last byte to be written) so
405  * that this function can be used to write to the very end-of-file (end = -1).
406  */
407 int filemap_write_and_wait_range(struct address_space *mapping,
408                                  loff_t lstart, loff_t lend)
409 {
410         int err = 0;
411
412         if (mapping->nrpages) {
413                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
414                                                  WB_SYNC_ALL);
415                 /* See comment of filemap_write_and_wait() */
416                 if (err != -EIO) {
417                         int err2 = wait_on_page_writeback_range(mapping,
418                                                 lstart >> PAGE_CACHE_SHIFT,
419                                                 lend >> PAGE_CACHE_SHIFT);
420                         if (!err)
421                                 err = err2;
422                 }
423         }
424         return err;
425 }
426
427 /**
428  * add_to_page_cache - add newly allocated pagecache pages
429  * @page:       page to add
430  * @mapping:    the page's address_space
431  * @offset:     page index
432  * @gfp_mask:   page allocation mode
433  *
434  * This function is used to add newly allocated pagecache pages;
435  * the page is new, so we can just run SetPageLocked() against it.
436  * The other page state flags were set by rmqueue().
437  *
438  * This function does not add the page to the LRU.  The caller must do that.
439  */
440 int add_to_page_cache(struct page *page, struct address_space *mapping,
441                 pgoff_t offset, gfp_t gfp_mask)
442 {
443         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
444
445         if (error == 0) {
446                 write_lock_irq(&mapping->tree_lock);
447                 error = radix_tree_insert(&mapping->page_tree, offset, page);
448                 if (!error) {
449                         page_cache_get(page);
450                         SetPageLocked(page);
451                         page->mapping = mapping;
452                         page->index = offset;
453                         mapping->nrpages++;
454                         __inc_zone_page_state(page, NR_FILE_PAGES);
455                 }
456                 write_unlock_irq(&mapping->tree_lock);
457                 radix_tree_preload_end();
458         }
459         return error;
460 }
461 EXPORT_SYMBOL(add_to_page_cache);
462
463 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
464                                 pgoff_t offset, gfp_t gfp_mask)
465 {
466         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
467         if (ret == 0)
468                 lru_cache_add(page);
469         return ret;
470 }
471
472 #ifdef CONFIG_NUMA
473 struct page *__page_cache_alloc(gfp_t gfp)
474 {
475         if (cpuset_do_page_mem_spread()) {
476                 int n = cpuset_mem_spread_node();
477                 return alloc_pages_node(n, gfp, 0);
478         }
479         return alloc_pages(gfp, 0);
480 }
481 EXPORT_SYMBOL(__page_cache_alloc);
482 #endif
483
484 static int __sleep_on_page_lock(void *word)
485 {
486         io_schedule();
487         return 0;
488 }
489
490 /*
491  * In order to wait for pages to become available there must be
492  * waitqueues associated with pages. By using a hash table of
493  * waitqueues where the bucket discipline is to maintain all
494  * waiters on the same queue and wake all when any of the pages
495  * become available, and for the woken contexts to check to be
496  * sure the appropriate page became available, this saves space
497  * at a cost of "thundering herd" phenomena during rare hash
498  * collisions.
499  */
500 static wait_queue_head_t *page_waitqueue(struct page *page)
501 {
502         const struct zone *zone = page_zone(page);
503
504         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
505 }
506
507 static inline void wake_up_page(struct page *page, int bit)
508 {
509         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
510 }
511
512 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
513 {
514         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
515
516         if (test_bit(bit_nr, &page->flags))
517                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
518                                                         TASK_UNINTERRUPTIBLE);
519 }
520 EXPORT_SYMBOL(wait_on_page_bit);
521
522 /**
523  * unlock_page - unlock a locked page
524  * @page: the page
525  *
526  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
527  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
528  * mechananism between PageLocked pages and PageWriteback pages is shared.
529  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
530  *
531  * The first mb is necessary to safely close the critical section opened by the
532  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
533  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
534  * parallel wait_on_page_locked()).
535  */
536 void fastcall unlock_page(struct page *page)
537 {
538         smp_mb__before_clear_bit();
539         if (!TestClearPageLocked(page))
540                 BUG();
541         smp_mb__after_clear_bit(); 
542         wake_up_page(page, PG_locked);
543 }
544 EXPORT_SYMBOL(unlock_page);
545
546 /**
547  * end_page_writeback - end writeback against a page
548  * @page: the page
549  */
550 void end_page_writeback(struct page *page)
551 {
552         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
553                 if (!test_clear_page_writeback(page))
554                         BUG();
555         }
556         smp_mb__after_clear_bit();
557         wake_up_page(page, PG_writeback);
558 }
559 EXPORT_SYMBOL(end_page_writeback);
560
561 /**
562  * __lock_page - get a lock on the page, assuming we need to sleep to get it
563  * @page: the page to lock
564  *
565  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
566  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
567  * chances are that on the second loop, the block layer's plug list is empty,
568  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
569  */
570 void fastcall __lock_page(struct page *page)
571 {
572         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
573
574         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
575                                                         TASK_UNINTERRUPTIBLE);
576 }
577 EXPORT_SYMBOL(__lock_page);
578
579 /*
580  * Variant of lock_page that does not require the caller to hold a reference
581  * on the page's mapping.
582  */
583 void fastcall __lock_page_nosync(struct page *page)
584 {
585         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
586         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
587                                                         TASK_UNINTERRUPTIBLE);
588 }
589
590 /**
591  * find_get_page - find and get a page reference
592  * @mapping: the address_space to search
593  * @offset: the page index
594  *
595  * Is there a pagecache struct page at the given (mapping, offset) tuple?
596  * If yes, increment its refcount and return it; if no, return NULL.
597  */
598 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
599 {
600         struct page *page;
601
602         read_lock_irq(&mapping->tree_lock);
603         page = radix_tree_lookup(&mapping->page_tree, offset);
604         if (page)
605                 page_cache_get(page);
606         read_unlock_irq(&mapping->tree_lock);
607         return page;
608 }
609 EXPORT_SYMBOL(find_get_page);
610
611 /**
612  * find_lock_page - locate, pin and lock a pagecache page
613  * @mapping: the address_space to search
614  * @offset: the page index
615  *
616  * Locates the desired pagecache page, locks it, increments its reference
617  * count and returns its address.
618  *
619  * Returns zero if the page was not present. find_lock_page() may sleep.
620  */
621 struct page *find_lock_page(struct address_space *mapping,
622                                 pgoff_t offset)
623 {
624         struct page *page;
625
626 repeat:
627         read_lock_irq(&mapping->tree_lock);
628         page = radix_tree_lookup(&mapping->page_tree, offset);
629         if (page) {
630                 page_cache_get(page);
631                 if (TestSetPageLocked(page)) {
632                         read_unlock_irq(&mapping->tree_lock);
633                         __lock_page(page);
634
635                         /* Has the page been truncated while we slept? */
636                         if (unlikely(page->mapping != mapping)) {
637                                 unlock_page(page);
638                                 page_cache_release(page);
639                                 goto repeat;
640                         }
641                         VM_BUG_ON(page->index != offset);
642                         goto out;
643                 }
644         }
645         read_unlock_irq(&mapping->tree_lock);
646 out:
647         return page;
648 }
649 EXPORT_SYMBOL(find_lock_page);
650
651 /**
652  * find_or_create_page - locate or add a pagecache page
653  * @mapping: the page's address_space
654  * @index: the page's index into the mapping
655  * @gfp_mask: page allocation mode
656  *
657  * Locates a page in the pagecache.  If the page is not present, a new page
658  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
659  * LRU list.  The returned page is locked and has its reference count
660  * incremented.
661  *
662  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
663  * allocation!
664  *
665  * find_or_create_page() returns the desired page's address, or zero on
666  * memory exhaustion.
667  */
668 struct page *find_or_create_page(struct address_space *mapping,
669                 pgoff_t index, gfp_t gfp_mask)
670 {
671         struct page *page;
672         int err;
673 repeat:
674         page = find_lock_page(mapping, index);
675         if (!page) {
676                 page = __page_cache_alloc(gfp_mask);
677                 if (!page)
678                         return NULL;
679                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
680                 if (unlikely(err)) {
681                         page_cache_release(page);
682                         page = NULL;
683                         if (err == -EEXIST)
684                                 goto repeat;
685                 }
686         }
687         return page;
688 }
689 EXPORT_SYMBOL(find_or_create_page);
690
691 /**
692  * find_get_pages - gang pagecache lookup
693  * @mapping:    The address_space to search
694  * @start:      The starting page index
695  * @nr_pages:   The maximum number of pages
696  * @pages:      Where the resulting pages are placed
697  *
698  * find_get_pages() will search for and return a group of up to
699  * @nr_pages pages in the mapping.  The pages are placed at @pages.
700  * find_get_pages() takes a reference against the returned pages.
701  *
702  * The search returns a group of mapping-contiguous pages with ascending
703  * indexes.  There may be holes in the indices due to not-present pages.
704  *
705  * find_get_pages() returns the number of pages which were found.
706  */
707 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
708                             unsigned int nr_pages, struct page **pages)
709 {
710         unsigned int i;
711         unsigned int ret;
712
713         read_lock_irq(&mapping->tree_lock);
714         ret = radix_tree_gang_lookup(&mapping->page_tree,
715                                 (void **)pages, start, nr_pages);
716         for (i = 0; i < ret; i++)
717                 page_cache_get(pages[i]);
718         read_unlock_irq(&mapping->tree_lock);
719         return ret;
720 }
721
722 /**
723  * find_get_pages_contig - gang contiguous pagecache lookup
724  * @mapping:    The address_space to search
725  * @index:      The starting page index
726  * @nr_pages:   The maximum number of pages
727  * @pages:      Where the resulting pages are placed
728  *
729  * find_get_pages_contig() works exactly like find_get_pages(), except
730  * that the returned number of pages are guaranteed to be contiguous.
731  *
732  * find_get_pages_contig() returns the number of pages which were found.
733  */
734 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
735                                unsigned int nr_pages, struct page **pages)
736 {
737         unsigned int i;
738         unsigned int ret;
739
740         read_lock_irq(&mapping->tree_lock);
741         ret = radix_tree_gang_lookup(&mapping->page_tree,
742                                 (void **)pages, index, nr_pages);
743         for (i = 0; i < ret; i++) {
744                 if (pages[i]->mapping == NULL || pages[i]->index != index)
745                         break;
746
747                 page_cache_get(pages[i]);
748                 index++;
749         }
750         read_unlock_irq(&mapping->tree_lock);
751         return i;
752 }
753 EXPORT_SYMBOL(find_get_pages_contig);
754
755 /**
756  * find_get_pages_tag - find and return pages that match @tag
757  * @mapping:    the address_space to search
758  * @index:      the starting page index
759  * @tag:        the tag index
760  * @nr_pages:   the maximum number of pages
761  * @pages:      where the resulting pages are placed
762  *
763  * Like find_get_pages, except we only return pages which are tagged with
764  * @tag.   We update @index to index the next page for the traversal.
765  */
766 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
767                         int tag, unsigned int nr_pages, struct page **pages)
768 {
769         unsigned int i;
770         unsigned int ret;
771
772         read_lock_irq(&mapping->tree_lock);
773         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
774                                 (void **)pages, *index, nr_pages, tag);
775         for (i = 0; i < ret; i++)
776                 page_cache_get(pages[i]);
777         if (ret)
778                 *index = pages[ret - 1]->index + 1;
779         read_unlock_irq(&mapping->tree_lock);
780         return ret;
781 }
782 EXPORT_SYMBOL(find_get_pages_tag);
783
784 /**
785  * grab_cache_page_nowait - returns locked page at given index in given cache
786  * @mapping: target address_space
787  * @index: the page index
788  *
789  * Same as grab_cache_page(), but do not wait if the page is unavailable.
790  * This is intended for speculative data generators, where the data can
791  * be regenerated if the page couldn't be grabbed.  This routine should
792  * be safe to call while holding the lock for another page.
793  *
794  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
795  * and deadlock against the caller's locked page.
796  */
797 struct page *
798 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
799 {
800         struct page *page = find_get_page(mapping, index);
801
802         if (page) {
803                 if (!TestSetPageLocked(page))
804                         return page;
805                 page_cache_release(page);
806                 return NULL;
807         }
808         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
809         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
810                 page_cache_release(page);
811                 page = NULL;
812         }
813         return page;
814 }
815 EXPORT_SYMBOL(grab_cache_page_nowait);
816
817 /*
818  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
819  * a _large_ part of the i/o request. Imagine the worst scenario:
820  *
821  *      ---R__________________________________________B__________
822  *         ^ reading here                             ^ bad block(assume 4k)
823  *
824  * read(R) => miss => readahead(R...B) => media error => frustrating retries
825  * => failing the whole request => read(R) => read(R+1) =>
826  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
827  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
828  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
829  *
830  * It is going insane. Fix it by quickly scaling down the readahead size.
831  */
832 static void shrink_readahead_size_eio(struct file *filp,
833                                         struct file_ra_state *ra)
834 {
835         if (!ra->ra_pages)
836                 return;
837
838         ra->ra_pages /= 4;
839 }
840
841 /**
842  * do_generic_mapping_read - generic file read routine
843  * @mapping:    address_space to be read
844  * @ra:         file's readahead state
845  * @filp:       the file to read
846  * @ppos:       current file position
847  * @desc:       read_descriptor
848  * @actor:      read method
849  *
850  * This is a generic file read routine, and uses the
851  * mapping->a_ops->readpage() function for the actual low-level stuff.
852  *
853  * This is really ugly. But the goto's actually try to clarify some
854  * of the logic when it comes to error handling etc.
855  *
856  * Note the struct file* is only passed for the use of readpage.
857  * It may be NULL.
858  */
859 void do_generic_mapping_read(struct address_space *mapping,
860                              struct file_ra_state *ra,
861                              struct file *filp,
862                              loff_t *ppos,
863                              read_descriptor_t *desc,
864                              read_actor_t actor)
865 {
866         struct inode *inode = mapping->host;
867         pgoff_t index;
868         pgoff_t last_index;
869         pgoff_t prev_index;
870         unsigned long offset;      /* offset into pagecache page */
871         unsigned int prev_offset;
872         int error;
873
874         index = *ppos >> PAGE_CACHE_SHIFT;
875         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
876         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
877         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
878         offset = *ppos & ~PAGE_CACHE_MASK;
879
880         for (;;) {
881                 struct page *page;
882                 pgoff_t end_index;
883                 loff_t isize;
884                 unsigned long nr, ret;
885
886                 cond_resched();
887 find_page:
888                 page = find_get_page(mapping, index);
889                 if (!page) {
890                         page_cache_sync_readahead(mapping,
891                                         ra, filp,
892                                         index, last_index - index);
893                         page = find_get_page(mapping, index);
894                         if (unlikely(page == NULL))
895                                 goto no_cached_page;
896                 }
897                 if (PageReadahead(page)) {
898                         page_cache_async_readahead(mapping,
899                                         ra, filp, page,
900                                         index, last_index - index);
901                 }
902                 if (!PageUptodate(page))
903                         goto page_not_up_to_date;
904 page_ok:
905                 /*
906                  * i_size must be checked after we know the page is Uptodate.
907                  *
908                  * Checking i_size after the check allows us to calculate
909                  * the correct value for "nr", which means the zero-filled
910                  * part of the page is not copied back to userspace (unless
911                  * another truncate extends the file - this is desired though).
912                  */
913
914                 isize = i_size_read(inode);
915                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
916                 if (unlikely(!isize || index > end_index)) {
917                         page_cache_release(page);
918                         goto out;
919                 }
920
921                 /* nr is the maximum number of bytes to copy from this page */
922                 nr = PAGE_CACHE_SIZE;
923                 if (index == end_index) {
924                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
925                         if (nr <= offset) {
926                                 page_cache_release(page);
927                                 goto out;
928                         }
929                 }
930                 nr = nr - offset;
931
932                 /* If users can be writing to this page using arbitrary
933                  * virtual addresses, take care about potential aliasing
934                  * before reading the page on the kernel side.
935                  */
936                 if (mapping_writably_mapped(mapping))
937                         flush_dcache_page(page);
938
939                 /*
940                  * When a sequential read accesses a page several times,
941                  * only mark it as accessed the first time.
942                  */
943                 if (prev_index != index || offset != prev_offset)
944                         mark_page_accessed(page);
945                 prev_index = index;
946
947                 /*
948                  * Ok, we have the page, and it's up-to-date, so
949                  * now we can copy it to user space...
950                  *
951                  * The actor routine returns how many bytes were actually used..
952                  * NOTE! This may not be the same as how much of a user buffer
953                  * we filled up (we may be padding etc), so we can only update
954                  * "pos" here (the actor routine has to update the user buffer
955                  * pointers and the remaining count).
956                  */
957                 ret = actor(desc, page, offset, nr);
958                 offset += ret;
959                 index += offset >> PAGE_CACHE_SHIFT;
960                 offset &= ~PAGE_CACHE_MASK;
961                 prev_offset = offset;
962
963                 page_cache_release(page);
964                 if (ret == nr && desc->count)
965                         continue;
966                 goto out;
967
968 page_not_up_to_date:
969                 /* Get exclusive access to the page ... */
970                 lock_page(page);
971
972                 /* Did it get truncated before we got the lock? */
973                 if (!page->mapping) {
974                         unlock_page(page);
975                         page_cache_release(page);
976                         continue;
977                 }
978
979                 /* Did somebody else fill it already? */
980                 if (PageUptodate(page)) {
981                         unlock_page(page);
982                         goto page_ok;
983                 }
984
985 readpage:
986                 /* Start the actual read. The read will unlock the page. */
987                 error = mapping->a_ops->readpage(filp, page);
988
989                 if (unlikely(error)) {
990                         if (error == AOP_TRUNCATED_PAGE) {
991                                 page_cache_release(page);
992                                 goto find_page;
993                         }
994                         goto readpage_error;
995                 }
996
997                 if (!PageUptodate(page)) {
998                         lock_page(page);
999                         if (!PageUptodate(page)) {
1000                                 if (page->mapping == NULL) {
1001                                         /*
1002                                          * invalidate_inode_pages got it
1003                                          */
1004                                         unlock_page(page);
1005                                         page_cache_release(page);
1006                                         goto find_page;
1007                                 }
1008                                 unlock_page(page);
1009                                 error = -EIO;
1010                                 shrink_readahead_size_eio(filp, ra);
1011                                 goto readpage_error;
1012                         }
1013                         unlock_page(page);
1014                 }
1015
1016                 goto page_ok;
1017
1018 readpage_error:
1019                 /* UHHUH! A synchronous read error occurred. Report it */
1020                 desc->error = error;
1021                 page_cache_release(page);
1022                 goto out;
1023
1024 no_cached_page:
1025                 /*
1026                  * Ok, it wasn't cached, so we need to create a new
1027                  * page..
1028                  */
1029                 page = page_cache_alloc_cold(mapping);
1030                 if (!page) {
1031                         desc->error = -ENOMEM;
1032                         goto out;
1033                 }
1034                 error = add_to_page_cache_lru(page, mapping,
1035                                                 index, GFP_KERNEL);
1036                 if (error) {
1037                         page_cache_release(page);
1038                         if (error == -EEXIST)
1039                                 goto find_page;
1040                         desc->error = error;
1041                         goto out;
1042                 }
1043                 goto readpage;
1044         }
1045
1046 out:
1047         ra->prev_pos = prev_index;
1048         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1049         ra->prev_pos |= prev_offset;
1050
1051         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1052         if (filp)
1053                 file_accessed(filp);
1054 }
1055 EXPORT_SYMBOL(do_generic_mapping_read);
1056
1057 int file_read_actor(read_descriptor_t *desc, struct page *page,
1058                         unsigned long offset, unsigned long size)
1059 {
1060         char *kaddr;
1061         unsigned long left, count = desc->count;
1062
1063         if (size > count)
1064                 size = count;
1065
1066         /*
1067          * Faults on the destination of a read are common, so do it before
1068          * taking the kmap.
1069          */
1070         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1071                 kaddr = kmap_atomic(page, KM_USER0);
1072                 left = __copy_to_user_inatomic(desc->arg.buf,
1073                                                 kaddr + offset, size);
1074                 kunmap_atomic(kaddr, KM_USER0);
1075                 if (left == 0)
1076                         goto success;
1077         }
1078
1079         /* Do it the slow way */
1080         kaddr = kmap(page);
1081         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1082         kunmap(page);
1083
1084         if (left) {
1085                 size -= left;
1086                 desc->error = -EFAULT;
1087         }
1088 success:
1089         desc->count = count - size;
1090         desc->written += size;
1091         desc->arg.buf += size;
1092         return size;
1093 }
1094
1095 /*
1096  * Performs necessary checks before doing a write
1097  * @iov:        io vector request
1098  * @nr_segs:    number of segments in the iovec
1099  * @count:      number of bytes to write
1100  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1101  *
1102  * Adjust number of segments and amount of bytes to write (nr_segs should be
1103  * properly initialized first). Returns appropriate error code that caller
1104  * should return or zero in case that write should be allowed.
1105  */
1106 int generic_segment_checks(const struct iovec *iov,
1107                         unsigned long *nr_segs, size_t *count, int access_flags)
1108 {
1109         unsigned long   seg;
1110         size_t cnt = 0;
1111         for (seg = 0; seg < *nr_segs; seg++) {
1112                 const struct iovec *iv = &iov[seg];
1113
1114                 /*
1115                  * If any segment has a negative length, or the cumulative
1116                  * length ever wraps negative then return -EINVAL.
1117                  */
1118                 cnt += iv->iov_len;
1119                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1120                         return -EINVAL;
1121                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1122                         continue;
1123                 if (seg == 0)
1124                         return -EFAULT;
1125                 *nr_segs = seg;
1126                 cnt -= iv->iov_len;     /* This segment is no good */
1127                 break;
1128         }
1129         *count = cnt;
1130         return 0;
1131 }
1132 EXPORT_SYMBOL(generic_segment_checks);
1133
1134 /**
1135  * generic_file_aio_read - generic filesystem read routine
1136  * @iocb:       kernel I/O control block
1137  * @iov:        io vector request
1138  * @nr_segs:    number of segments in the iovec
1139  * @pos:        current file position
1140  *
1141  * This is the "read()" routine for all filesystems
1142  * that can use the page cache directly.
1143  */
1144 ssize_t
1145 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1146                 unsigned long nr_segs, loff_t pos)
1147 {
1148         struct file *filp = iocb->ki_filp;
1149         ssize_t retval;
1150         unsigned long seg;
1151         size_t count;
1152         loff_t *ppos = &iocb->ki_pos;
1153
1154         count = 0;
1155         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1156         if (retval)
1157                 return retval;
1158
1159         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1160         if (filp->f_flags & O_DIRECT) {
1161                 loff_t size;
1162                 struct address_space *mapping;
1163                 struct inode *inode;
1164
1165                 mapping = filp->f_mapping;
1166                 inode = mapping->host;
1167                 retval = 0;
1168                 if (!count)
1169                         goto out; /* skip atime */
1170                 size = i_size_read(inode);
1171                 if (pos < size) {
1172                         retval = generic_file_direct_IO(READ, iocb,
1173                                                 iov, pos, nr_segs);
1174                         if (retval > 0)
1175                                 *ppos = pos + retval;
1176                 }
1177                 if (likely(retval != 0)) {
1178                         file_accessed(filp);
1179                         goto out;
1180                 }
1181         }
1182
1183         retval = 0;
1184         if (count) {
1185                 for (seg = 0; seg < nr_segs; seg++) {
1186                         read_descriptor_t desc;
1187
1188                         desc.written = 0;
1189                         desc.arg.buf = iov[seg].iov_base;
1190                         desc.count = iov[seg].iov_len;
1191                         if (desc.count == 0)
1192                                 continue;
1193                         desc.error = 0;
1194                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1195                         retval += desc.written;
1196                         if (desc.error) {
1197                                 retval = retval ?: desc.error;
1198                                 break;
1199                         }
1200                         if (desc.count > 0)
1201                                 break;
1202                 }
1203         }
1204 out:
1205         return retval;
1206 }
1207 EXPORT_SYMBOL(generic_file_aio_read);
1208
1209 static ssize_t
1210 do_readahead(struct address_space *mapping, struct file *filp,
1211              pgoff_t index, unsigned long nr)
1212 {
1213         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1214                 return -EINVAL;
1215
1216         force_page_cache_readahead(mapping, filp, index,
1217                                         max_sane_readahead(nr));
1218         return 0;
1219 }
1220
1221 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1222 {
1223         ssize_t ret;
1224         struct file *file;
1225
1226         ret = -EBADF;
1227         file = fget(fd);
1228         if (file) {
1229                 if (file->f_mode & FMODE_READ) {
1230                         struct address_space *mapping = file->f_mapping;
1231                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1232                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1233                         unsigned long len = end - start + 1;
1234                         ret = do_readahead(mapping, file, start, len);
1235                 }
1236                 fput(file);
1237         }
1238         return ret;
1239 }
1240
1241 #ifdef CONFIG_MMU
1242 /**
1243  * page_cache_read - adds requested page to the page cache if not already there
1244  * @file:       file to read
1245  * @offset:     page index
1246  *
1247  * This adds the requested page to the page cache if it isn't already there,
1248  * and schedules an I/O to read in its contents from disk.
1249  */
1250 static int fastcall page_cache_read(struct file * file, pgoff_t offset)
1251 {
1252         struct address_space *mapping = file->f_mapping;
1253         struct page *page; 
1254         int ret;
1255
1256         do {
1257                 page = page_cache_alloc_cold(mapping);
1258                 if (!page)
1259                         return -ENOMEM;
1260
1261                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1262                 if (ret == 0)
1263                         ret = mapping->a_ops->readpage(file, page);
1264                 else if (ret == -EEXIST)
1265                         ret = 0; /* losing race to add is OK */
1266
1267                 page_cache_release(page);
1268
1269         } while (ret == AOP_TRUNCATED_PAGE);
1270                 
1271         return ret;
1272 }
1273
1274 #define MMAP_LOTSAMISS  (100)
1275
1276 /**
1277  * filemap_fault - read in file data for page fault handling
1278  * @vma:        vma in which the fault was taken
1279  * @vmf:        struct vm_fault containing details of the fault
1280  *
1281  * filemap_fault() is invoked via the vma operations vector for a
1282  * mapped memory region to read in file data during a page fault.
1283  *
1284  * The goto's are kind of ugly, but this streamlines the normal case of having
1285  * it in the page cache, and handles the special cases reasonably without
1286  * having a lot of duplicated code.
1287  */
1288 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1289 {
1290         int error;
1291         struct file *file = vma->vm_file;
1292         struct address_space *mapping = file->f_mapping;
1293         struct file_ra_state *ra = &file->f_ra;
1294         struct inode *inode = mapping->host;
1295         struct page *page;
1296         unsigned long size;
1297         int did_readaround = 0;
1298         int ret = 0;
1299
1300         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1301         if (vmf->pgoff >= size)
1302                 goto outside_data_content;
1303
1304         /* If we don't want any read-ahead, don't bother */
1305         if (VM_RandomReadHint(vma))
1306                 goto no_cached_page;
1307
1308         /*
1309          * Do we have something in the page cache already?
1310          */
1311 retry_find:
1312         page = find_lock_page(mapping, vmf->pgoff);
1313         /*
1314          * For sequential accesses, we use the generic readahead logic.
1315          */
1316         if (VM_SequentialReadHint(vma)) {
1317                 if (!page) {
1318                         page_cache_sync_readahead(mapping, ra, file,
1319                                                            vmf->pgoff, 1);
1320                         page = find_lock_page(mapping, vmf->pgoff);
1321                         if (!page)
1322                                 goto no_cached_page;
1323                 }
1324                 if (PageReadahead(page)) {
1325                         page_cache_async_readahead(mapping, ra, file, page,
1326                                                            vmf->pgoff, 1);
1327                 }
1328         }
1329
1330         if (!page) {
1331                 unsigned long ra_pages;
1332
1333                 ra->mmap_miss++;
1334
1335                 /*
1336                  * Do we miss much more than hit in this file? If so,
1337                  * stop bothering with read-ahead. It will only hurt.
1338                  */
1339                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1340                         goto no_cached_page;
1341
1342                 /*
1343                  * To keep the pgmajfault counter straight, we need to
1344                  * check did_readaround, as this is an inner loop.
1345                  */
1346                 if (!did_readaround) {
1347                         ret = VM_FAULT_MAJOR;
1348                         count_vm_event(PGMAJFAULT);
1349                 }
1350                 did_readaround = 1;
1351                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1352                 if (ra_pages) {
1353                         pgoff_t start = 0;
1354
1355                         if (vmf->pgoff > ra_pages / 2)
1356                                 start = vmf->pgoff - ra_pages / 2;
1357                         do_page_cache_readahead(mapping, file, start, ra_pages);
1358                 }
1359                 page = find_lock_page(mapping, vmf->pgoff);
1360                 if (!page)
1361                         goto no_cached_page;
1362         }
1363
1364         if (!did_readaround)
1365                 ra->mmap_miss--;
1366
1367         /*
1368          * We have a locked page in the page cache, now we need to check
1369          * that it's up-to-date. If not, it is going to be due to an error.
1370          */
1371         if (unlikely(!PageUptodate(page)))
1372                 goto page_not_uptodate;
1373
1374         /* Must recheck i_size under page lock */
1375         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1376         if (unlikely(vmf->pgoff >= size)) {
1377                 unlock_page(page);
1378                 page_cache_release(page);
1379                 goto outside_data_content;
1380         }
1381
1382         /*
1383          * Found the page and have a reference on it.
1384          */
1385         mark_page_accessed(page);
1386         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1387         vmf->page = page;
1388         return ret | VM_FAULT_LOCKED;
1389
1390 outside_data_content:
1391         /*
1392          * An external ptracer can access pages that normally aren't
1393          * accessible..
1394          */
1395         if (vma->vm_mm == current->mm)
1396                 return VM_FAULT_SIGBUS;
1397
1398         /* Fall through to the non-read-ahead case */
1399 no_cached_page:
1400         /*
1401          * We're only likely to ever get here if MADV_RANDOM is in
1402          * effect.
1403          */
1404         error = page_cache_read(file, vmf->pgoff);
1405
1406         /*
1407          * The page we want has now been added to the page cache.
1408          * In the unlikely event that someone removed it in the
1409          * meantime, we'll just come back here and read it again.
1410          */
1411         if (error >= 0)
1412                 goto retry_find;
1413
1414         /*
1415          * An error return from page_cache_read can result if the
1416          * system is low on memory, or a problem occurs while trying
1417          * to schedule I/O.
1418          */
1419         if (error == -ENOMEM)
1420                 return VM_FAULT_OOM;
1421         return VM_FAULT_SIGBUS;
1422
1423 page_not_uptodate:
1424         /* IO error path */
1425         if (!did_readaround) {
1426                 ret = VM_FAULT_MAJOR;
1427                 count_vm_event(PGMAJFAULT);
1428         }
1429
1430         /*
1431          * Umm, take care of errors if the page isn't up-to-date.
1432          * Try to re-read it _once_. We do this synchronously,
1433          * because there really aren't any performance issues here
1434          * and we need to check for errors.
1435          */
1436         ClearPageError(page);
1437         error = mapping->a_ops->readpage(file, page);
1438         page_cache_release(page);
1439
1440         if (!error || error == AOP_TRUNCATED_PAGE)
1441                 goto retry_find;
1442
1443         /* Things didn't work out. Return zero to tell the mm layer so. */
1444         shrink_readahead_size_eio(file, ra);
1445         return VM_FAULT_SIGBUS;
1446 }
1447 EXPORT_SYMBOL(filemap_fault);
1448
1449 struct vm_operations_struct generic_file_vm_ops = {
1450         .fault          = filemap_fault,
1451 };
1452
1453 /* This is used for a general mmap of a disk file */
1454
1455 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1456 {
1457         struct address_space *mapping = file->f_mapping;
1458
1459         if (!mapping->a_ops->readpage)
1460                 return -ENOEXEC;
1461         file_accessed(file);
1462         vma->vm_ops = &generic_file_vm_ops;
1463         vma->vm_flags |= VM_CAN_NONLINEAR;
1464         return 0;
1465 }
1466
1467 /*
1468  * This is for filesystems which do not implement ->writepage.
1469  */
1470 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1471 {
1472         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1473                 return -EINVAL;
1474         return generic_file_mmap(file, vma);
1475 }
1476 #else
1477 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1478 {
1479         return -ENOSYS;
1480 }
1481 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1482 {
1483         return -ENOSYS;
1484 }
1485 #endif /* CONFIG_MMU */
1486
1487 EXPORT_SYMBOL(generic_file_mmap);
1488 EXPORT_SYMBOL(generic_file_readonly_mmap);
1489
1490 static struct page *__read_cache_page(struct address_space *mapping,
1491                                 pgoff_t index,
1492                                 int (*filler)(void *,struct page*),
1493                                 void *data)
1494 {
1495         struct page *page;
1496         int err;
1497 repeat:
1498         page = find_get_page(mapping, index);
1499         if (!page) {
1500                 page = page_cache_alloc_cold(mapping);
1501                 if (!page)
1502                         return ERR_PTR(-ENOMEM);
1503                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1504                 if (unlikely(err)) {
1505                         page_cache_release(page);
1506                         if (err == -EEXIST)
1507                                 goto repeat;
1508                         /* Presumably ENOMEM for radix tree node */
1509                         return ERR_PTR(err);
1510                 }
1511                 err = filler(data, page);
1512                 if (err < 0) {
1513                         page_cache_release(page);
1514                         page = ERR_PTR(err);
1515                 }
1516         }
1517         return page;
1518 }
1519
1520 /*
1521  * Same as read_cache_page, but don't wait for page to become unlocked
1522  * after submitting it to the filler.
1523  */
1524 struct page *read_cache_page_async(struct address_space *mapping,
1525                                 pgoff_t index,
1526                                 int (*filler)(void *,struct page*),
1527                                 void *data)
1528 {
1529         struct page *page;
1530         int err;
1531
1532 retry:
1533         page = __read_cache_page(mapping, index, filler, data);
1534         if (IS_ERR(page))
1535                 return page;
1536         if (PageUptodate(page))
1537                 goto out;
1538
1539         lock_page(page);
1540         if (!page->mapping) {
1541                 unlock_page(page);
1542                 page_cache_release(page);
1543                 goto retry;
1544         }
1545         if (PageUptodate(page)) {
1546                 unlock_page(page);
1547                 goto out;
1548         }
1549         err = filler(data, page);
1550         if (err < 0) {
1551                 page_cache_release(page);
1552                 return ERR_PTR(err);
1553         }
1554 out:
1555         mark_page_accessed(page);
1556         return page;
1557 }
1558 EXPORT_SYMBOL(read_cache_page_async);
1559
1560 /**
1561  * read_cache_page - read into page cache, fill it if needed
1562  * @mapping:    the page's address_space
1563  * @index:      the page index
1564  * @filler:     function to perform the read
1565  * @data:       destination for read data
1566  *
1567  * Read into the page cache. If a page already exists, and PageUptodate() is
1568  * not set, try to fill the page then wait for it to become unlocked.
1569  *
1570  * If the page does not get brought uptodate, return -EIO.
1571  */
1572 struct page *read_cache_page(struct address_space *mapping,
1573                                 pgoff_t index,
1574                                 int (*filler)(void *,struct page*),
1575                                 void *data)
1576 {
1577         struct page *page;
1578
1579         page = read_cache_page_async(mapping, index, filler, data);
1580         if (IS_ERR(page))
1581                 goto out;
1582         wait_on_page_locked(page);
1583         if (!PageUptodate(page)) {
1584                 page_cache_release(page);
1585                 page = ERR_PTR(-EIO);
1586         }
1587  out:
1588         return page;
1589 }
1590 EXPORT_SYMBOL(read_cache_page);
1591
1592 /*
1593  * The logic we want is
1594  *
1595  *      if suid or (sgid and xgrp)
1596  *              remove privs
1597  */
1598 int should_remove_suid(struct dentry *dentry)
1599 {
1600         mode_t mode = dentry->d_inode->i_mode;
1601         int kill = 0;
1602
1603         /* suid always must be killed */
1604         if (unlikely(mode & S_ISUID))
1605                 kill = ATTR_KILL_SUID;
1606
1607         /*
1608          * sgid without any exec bits is just a mandatory locking mark; leave
1609          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1610          */
1611         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1612                 kill |= ATTR_KILL_SGID;
1613
1614         if (unlikely(kill && !capable(CAP_FSETID)))
1615                 return kill;
1616
1617         return 0;
1618 }
1619 EXPORT_SYMBOL(should_remove_suid);
1620
1621 int __remove_suid(struct dentry *dentry, int kill)
1622 {
1623         struct iattr newattrs;
1624
1625         newattrs.ia_valid = ATTR_FORCE | kill;
1626         return notify_change(dentry, &newattrs);
1627 }
1628
1629 int remove_suid(struct dentry *dentry)
1630 {
1631         int killsuid = should_remove_suid(dentry);
1632         int killpriv = security_inode_need_killpriv(dentry);
1633         int error = 0;
1634
1635         if (killpriv < 0)
1636                 return killpriv;
1637         if (killpriv)
1638                 error = security_inode_killpriv(dentry);
1639         if (!error && killsuid)
1640                 error = __remove_suid(dentry, killsuid);
1641
1642         return error;
1643 }
1644 EXPORT_SYMBOL(remove_suid);
1645
1646 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1647                         const struct iovec *iov, size_t base, size_t bytes)
1648 {
1649         size_t copied = 0, left = 0;
1650
1651         while (bytes) {
1652                 char __user *buf = iov->iov_base + base;
1653                 int copy = min(bytes, iov->iov_len - base);
1654
1655                 base = 0;
1656                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1657                 copied += copy;
1658                 bytes -= copy;
1659                 vaddr += copy;
1660                 iov++;
1661
1662                 if (unlikely(left))
1663                         break;
1664         }
1665         return copied - left;
1666 }
1667
1668 /*
1669  * Copy as much as we can into the page and return the number of bytes which
1670  * were sucessfully copied.  If a fault is encountered then return the number of
1671  * bytes which were copied.
1672  */
1673 size_t iov_iter_copy_from_user_atomic(struct page *page,
1674                 struct iov_iter *i, unsigned long offset, size_t bytes)
1675 {
1676         char *kaddr;
1677         size_t copied;
1678
1679         BUG_ON(!in_atomic());
1680         kaddr = kmap_atomic(page, KM_USER0);
1681         if (likely(i->nr_segs == 1)) {
1682                 int left;
1683                 char __user *buf = i->iov->iov_base + i->iov_offset;
1684                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1685                                                         buf, bytes);
1686                 copied = bytes - left;
1687         } else {
1688                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1689                                                 i->iov, i->iov_offset, bytes);
1690         }
1691         kunmap_atomic(kaddr, KM_USER0);
1692
1693         return copied;
1694 }
1695 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1696
1697 /*
1698  * This has the same sideeffects and return value as
1699  * iov_iter_copy_from_user_atomic().
1700  * The difference is that it attempts to resolve faults.
1701  * Page must not be locked.
1702  */
1703 size_t iov_iter_copy_from_user(struct page *page,
1704                 struct iov_iter *i, unsigned long offset, size_t bytes)
1705 {
1706         char *kaddr;
1707         size_t copied;
1708
1709         kaddr = kmap(page);
1710         if (likely(i->nr_segs == 1)) {
1711                 int left;
1712                 char __user *buf = i->iov->iov_base + i->iov_offset;
1713                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1714                 copied = bytes - left;
1715         } else {
1716                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1717                                                 i->iov, i->iov_offset, bytes);
1718         }
1719         kunmap(page);
1720         return copied;
1721 }
1722 EXPORT_SYMBOL(iov_iter_copy_from_user);
1723
1724 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1725 {
1726         if (likely(i->nr_segs == 1)) {
1727                 i->iov_offset += bytes;
1728         } else {
1729                 const struct iovec *iov = i->iov;
1730                 size_t base = i->iov_offset;
1731
1732                 while (bytes) {
1733                         int copy = min(bytes, iov->iov_len - base);
1734
1735                         bytes -= copy;
1736                         base += copy;
1737                         if (iov->iov_len == base) {
1738                                 iov++;
1739                                 base = 0;
1740                         }
1741                 }
1742                 i->iov = iov;
1743                 i->iov_offset = base;
1744         }
1745 }
1746
1747 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1748 {
1749         BUG_ON(i->count < bytes);
1750
1751         __iov_iter_advance_iov(i, bytes);
1752         i->count -= bytes;
1753 }
1754 EXPORT_SYMBOL(iov_iter_advance);
1755
1756 /*
1757  * Fault in the first iovec of the given iov_iter, to a maximum length
1758  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1759  * accessed (ie. because it is an invalid address).
1760  *
1761  * writev-intensive code may want this to prefault several iovecs -- that
1762  * would be possible (callers must not rely on the fact that _only_ the
1763  * first iovec will be faulted with the current implementation).
1764  */
1765 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1766 {
1767         char __user *buf = i->iov->iov_base + i->iov_offset;
1768         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1769         return fault_in_pages_readable(buf, bytes);
1770 }
1771 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1772
1773 /*
1774  * Return the count of just the current iov_iter segment.
1775  */
1776 size_t iov_iter_single_seg_count(struct iov_iter *i)
1777 {
1778         const struct iovec *iov = i->iov;
1779         if (i->nr_segs == 1)
1780                 return i->count;
1781         else
1782                 return min(i->count, iov->iov_len - i->iov_offset);
1783 }
1784 EXPORT_SYMBOL(iov_iter_single_seg_count);
1785
1786 /*
1787  * Performs necessary checks before doing a write
1788  *
1789  * Can adjust writing position or amount of bytes to write.
1790  * Returns appropriate error code that caller should return or
1791  * zero in case that write should be allowed.
1792  */
1793 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1794 {
1795         struct inode *inode = file->f_mapping->host;
1796         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1797
1798         if (unlikely(*pos < 0))
1799                 return -EINVAL;
1800
1801         if (!isblk) {
1802                 /* FIXME: this is for backwards compatibility with 2.4 */
1803                 if (file->f_flags & O_APPEND)
1804                         *pos = i_size_read(inode);
1805
1806                 if (limit != RLIM_INFINITY) {
1807                         if (*pos >= limit) {
1808                                 send_sig(SIGXFSZ, current, 0);
1809                                 return -EFBIG;
1810                         }
1811                         if (*count > limit - (typeof(limit))*pos) {
1812                                 *count = limit - (typeof(limit))*pos;
1813                         }
1814                 }
1815         }
1816
1817         /*
1818          * LFS rule
1819          */
1820         if (unlikely(*pos + *count > MAX_NON_LFS &&
1821                                 !(file->f_flags & O_LARGEFILE))) {
1822                 if (*pos >= MAX_NON_LFS) {
1823                         return -EFBIG;
1824                 }
1825                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1826                         *count = MAX_NON_LFS - (unsigned long)*pos;
1827                 }
1828         }
1829
1830         /*
1831          * Are we about to exceed the fs block limit ?
1832          *
1833          * If we have written data it becomes a short write.  If we have
1834          * exceeded without writing data we send a signal and return EFBIG.
1835          * Linus frestrict idea will clean these up nicely..
1836          */
1837         if (likely(!isblk)) {
1838                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1839                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1840                                 return -EFBIG;
1841                         }
1842                         /* zero-length writes at ->s_maxbytes are OK */
1843                 }
1844
1845                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1846                         *count = inode->i_sb->s_maxbytes - *pos;
1847         } else {
1848 #ifdef CONFIG_BLOCK
1849                 loff_t isize;
1850                 if (bdev_read_only(I_BDEV(inode)))
1851                         return -EPERM;
1852                 isize = i_size_read(inode);
1853                 if (*pos >= isize) {
1854                         if (*count || *pos > isize)
1855                                 return -ENOSPC;
1856                 }
1857
1858                 if (*pos + *count > isize)
1859                         *count = isize - *pos;
1860 #else
1861                 return -EPERM;
1862 #endif
1863         }
1864         return 0;
1865 }
1866 EXPORT_SYMBOL(generic_write_checks);
1867
1868 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1869                                 loff_t pos, unsigned len, unsigned flags,
1870                                 struct page **pagep, void **fsdata)
1871 {
1872         const struct address_space_operations *aops = mapping->a_ops;
1873
1874         if (aops->write_begin) {
1875                 return aops->write_begin(file, mapping, pos, len, flags,
1876                                                         pagep, fsdata);
1877         } else {
1878                 int ret;
1879                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1880                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1881                 struct inode *inode = mapping->host;
1882                 struct page *page;
1883 again:
1884                 page = __grab_cache_page(mapping, index);
1885                 *pagep = page;
1886                 if (!page)
1887                         return -ENOMEM;
1888
1889                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1890                         /*
1891                          * There is no way to resolve a short write situation
1892                          * for a !Uptodate page (except by double copying in
1893                          * the caller done by generic_perform_write_2copy).
1894                          *
1895                          * Instead, we have to bring it uptodate here.
1896                          */
1897                         ret = aops->readpage(file, page);
1898                         page_cache_release(page);
1899                         if (ret) {
1900                                 if (ret == AOP_TRUNCATED_PAGE)
1901                                         goto again;
1902                                 return ret;
1903                         }
1904                         goto again;
1905                 }
1906
1907                 ret = aops->prepare_write(file, page, offset, offset+len);
1908                 if (ret) {
1909                         unlock_page(page);
1910                         page_cache_release(page);
1911                         if (pos + len > inode->i_size)
1912                                 vmtruncate(inode, inode->i_size);
1913                 }
1914                 return ret;
1915         }
1916 }
1917 EXPORT_SYMBOL(pagecache_write_begin);
1918
1919 int pagecache_write_end(struct file *file, struct address_space *mapping,
1920                                 loff_t pos, unsigned len, unsigned copied,
1921                                 struct page *page, void *fsdata)
1922 {
1923         const struct address_space_operations *aops = mapping->a_ops;
1924         int ret;
1925
1926         if (aops->write_end) {
1927                 mark_page_accessed(page);
1928                 ret = aops->write_end(file, mapping, pos, len, copied,
1929                                                         page, fsdata);
1930         } else {
1931                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1932                 struct inode *inode = mapping->host;
1933
1934                 flush_dcache_page(page);
1935                 ret = aops->commit_write(file, page, offset, offset+len);
1936                 unlock_page(page);
1937                 mark_page_accessed(page);
1938                 page_cache_release(page);
1939
1940                 if (ret < 0) {
1941                         if (pos + len > inode->i_size)
1942                                 vmtruncate(inode, inode->i_size);
1943                 } else if (ret > 0)
1944                         ret = min_t(size_t, copied, ret);
1945                 else
1946                         ret = copied;
1947         }
1948
1949         return ret;
1950 }
1951 EXPORT_SYMBOL(pagecache_write_end);
1952
1953 ssize_t
1954 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1955                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1956                 size_t count, size_t ocount)
1957 {
1958         struct file     *file = iocb->ki_filp;
1959         struct address_space *mapping = file->f_mapping;
1960         struct inode    *inode = mapping->host;
1961         ssize_t         written;
1962
1963         if (count != ocount)
1964                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1965
1966         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1967         if (written > 0) {
1968                 loff_t end = pos + written;
1969                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1970                         i_size_write(inode,  end);
1971                         mark_inode_dirty(inode);
1972                 }
1973                 *ppos = end;
1974         }
1975
1976         /*
1977          * Sync the fs metadata but not the minor inode changes and
1978          * of course not the data as we did direct DMA for the IO.
1979          * i_mutex is held, which protects generic_osync_inode() from
1980          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
1981          */
1982         if ((written >= 0 || written == -EIOCBQUEUED) &&
1983             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1984                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1985                 if (err < 0)
1986                         written = err;
1987         }
1988         return written;
1989 }
1990 EXPORT_SYMBOL(generic_file_direct_write);
1991
1992 /*
1993  * Find or create a page at the given pagecache position. Return the locked
1994  * page. This function is specifically for buffered writes.
1995  */
1996 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
1997 {
1998         int status;
1999         struct page *page;
2000 repeat:
2001         page = find_lock_page(mapping, index);
2002         if (likely(page))
2003                 return page;
2004
2005         page = page_cache_alloc(mapping);
2006         if (!page)
2007                 return NULL;
2008         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2009         if (unlikely(status)) {
2010                 page_cache_release(page);
2011                 if (status == -EEXIST)
2012                         goto repeat;
2013                 return NULL;
2014         }
2015         return page;
2016 }
2017 EXPORT_SYMBOL(__grab_cache_page);
2018
2019 static ssize_t generic_perform_write_2copy(struct file *file,
2020                                 struct iov_iter *i, loff_t pos)
2021 {
2022         struct address_space *mapping = file->f_mapping;
2023         const struct address_space_operations *a_ops = mapping->a_ops;
2024         struct inode *inode = mapping->host;
2025         long status = 0;
2026         ssize_t written = 0;
2027
2028         do {
2029                 struct page *src_page;
2030                 struct page *page;
2031                 pgoff_t index;          /* Pagecache index for current page */
2032                 unsigned long offset;   /* Offset into pagecache page */
2033                 unsigned long bytes;    /* Bytes to write to page */
2034                 size_t copied;          /* Bytes copied from user */
2035
2036                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2037                 index = pos >> PAGE_CACHE_SHIFT;
2038                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2039                                                 iov_iter_count(i));
2040
2041                 /*
2042                  * a non-NULL src_page indicates that we're doing the
2043                  * copy via get_user_pages and kmap.
2044                  */
2045                 src_page = NULL;
2046
2047                 /*
2048                  * Bring in the user page that we will copy from _first_.
2049                  * Otherwise there's a nasty deadlock on copying from the
2050                  * same page as we're writing to, without it being marked
2051                  * up-to-date.
2052                  *
2053                  * Not only is this an optimisation, but it is also required
2054                  * to check that the address is actually valid, when atomic
2055                  * usercopies are used, below.
2056                  */
2057                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2058                         status = -EFAULT;
2059                         break;
2060                 }
2061
2062                 page = __grab_cache_page(mapping, index);
2063                 if (!page) {
2064                         status = -ENOMEM;
2065                         break;
2066                 }
2067
2068                 /*
2069                  * non-uptodate pages cannot cope with short copies, and we
2070                  * cannot take a pagefault with the destination page locked.
2071                  * So pin the source page to copy it.
2072                  */
2073                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2074                         unlock_page(page);
2075
2076                         src_page = alloc_page(GFP_KERNEL);
2077                         if (!src_page) {
2078                                 page_cache_release(page);
2079                                 status = -ENOMEM;
2080                                 break;
2081                         }
2082
2083                         /*
2084                          * Cannot get_user_pages with a page locked for the
2085                          * same reason as we can't take a page fault with a
2086                          * page locked (as explained below).
2087                          */
2088                         copied = iov_iter_copy_from_user(src_page, i,
2089                                                                 offset, bytes);
2090                         if (unlikely(copied == 0)) {
2091                                 status = -EFAULT;
2092                                 page_cache_release(page);
2093                                 page_cache_release(src_page);
2094                                 break;
2095                         }
2096                         bytes = copied;
2097
2098                         lock_page(page);
2099                         /*
2100                          * Can't handle the page going uptodate here, because
2101                          * that means we would use non-atomic usercopies, which
2102                          * zero out the tail of the page, which can cause
2103                          * zeroes to become transiently visible. We could just
2104                          * use a non-zeroing copy, but the APIs aren't too
2105                          * consistent.
2106                          */
2107                         if (unlikely(!page->mapping || PageUptodate(page))) {
2108                                 unlock_page(page);
2109                                 page_cache_release(page);
2110                                 page_cache_release(src_page);
2111                                 continue;
2112                         }
2113                 }
2114
2115                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2116                 if (unlikely(status))
2117                         goto fs_write_aop_error;
2118
2119                 if (!src_page) {
2120                         /*
2121                          * Must not enter the pagefault handler here, because
2122                          * we hold the page lock, so we might recursively
2123                          * deadlock on the same lock, or get an ABBA deadlock
2124                          * against a different lock, or against the mmap_sem
2125                          * (which nests outside the page lock).  So increment
2126                          * preempt count, and use _atomic usercopies.
2127                          *
2128                          * The page is uptodate so we are OK to encounter a
2129                          * short copy: if unmodified parts of the page are
2130                          * marked dirty and written out to disk, it doesn't
2131                          * really matter.
2132                          */
2133                         pagefault_disable();
2134                         copied = iov_iter_copy_from_user_atomic(page, i,
2135                                                                 offset, bytes);
2136                         pagefault_enable();
2137                 } else {
2138                         void *src, *dst;
2139                         src = kmap_atomic(src_page, KM_USER0);
2140                         dst = kmap_atomic(page, KM_USER1);
2141                         memcpy(dst + offset, src + offset, bytes);
2142                         kunmap_atomic(dst, KM_USER1);
2143                         kunmap_atomic(src, KM_USER0);
2144                         copied = bytes;
2145                 }
2146                 flush_dcache_page(page);
2147
2148                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2149                 if (unlikely(status < 0))
2150                         goto fs_write_aop_error;
2151                 if (unlikely(status > 0)) /* filesystem did partial write */
2152                         copied = min_t(size_t, copied, status);
2153
2154                 unlock_page(page);
2155                 mark_page_accessed(page);
2156                 page_cache_release(page);
2157                 if (src_page)
2158                         page_cache_release(src_page);
2159
2160                 iov_iter_advance(i, copied);
2161                 pos += copied;
2162                 written += copied;
2163
2164                 balance_dirty_pages_ratelimited(mapping);
2165                 cond_resched();
2166                 continue;
2167
2168 fs_write_aop_error:
2169                 unlock_page(page);
2170                 page_cache_release(page);
2171                 if (src_page)
2172                         page_cache_release(src_page);
2173
2174                 /*
2175                  * prepare_write() may have instantiated a few blocks
2176                  * outside i_size.  Trim these off again. Don't need
2177                  * i_size_read because we hold i_mutex.
2178                  */
2179                 if (pos + bytes > inode->i_size)
2180                         vmtruncate(inode, inode->i_size);
2181                 break;
2182         } while (iov_iter_count(i));
2183
2184         return written ? written : status;
2185 }
2186
2187 static ssize_t generic_perform_write(struct file *file,
2188                                 struct iov_iter *i, loff_t pos)
2189 {
2190         struct address_space *mapping = file->f_mapping;
2191         const struct address_space_operations *a_ops = mapping->a_ops;
2192         long status = 0;
2193         ssize_t written = 0;
2194         unsigned int flags = 0;
2195
2196         /*
2197          * Copies from kernel address space cannot fail (NFSD is a big user).
2198          */
2199         if (segment_eq(get_fs(), KERNEL_DS))
2200                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2201
2202         do {
2203                 struct page *page;
2204                 pgoff_t index;          /* Pagecache index for current page */
2205                 unsigned long offset;   /* Offset into pagecache page */
2206                 unsigned long bytes;    /* Bytes to write to page */
2207                 size_t copied;          /* Bytes copied from user */
2208                 void *fsdata;
2209
2210                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2211                 index = pos >> PAGE_CACHE_SHIFT;
2212                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2213                                                 iov_iter_count(i));
2214
2215 again:
2216
2217                 /*
2218                  * Bring in the user page that we will copy from _first_.
2219                  * Otherwise there's a nasty deadlock on copying from the
2220                  * same page as we're writing to, without it being marked
2221                  * up-to-date.
2222                  *
2223                  * Not only is this an optimisation, but it is also required
2224                  * to check that the address is actually valid, when atomic
2225                  * usercopies are used, below.
2226                  */
2227                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2228                         status = -EFAULT;
2229                         break;
2230                 }
2231
2232                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2233                                                 &page, &fsdata);
2234                 if (unlikely(status))
2235                         break;
2236
2237                 pagefault_disable();
2238                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2239                 pagefault_enable();
2240                 flush_dcache_page(page);
2241
2242                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2243                                                 page, fsdata);
2244                 if (unlikely(status < 0))
2245                         break;
2246                 copied = status;
2247
2248                 cond_resched();
2249
2250                 if (unlikely(copied == 0)) {
2251                         /*
2252                          * If we were unable to copy any data at all, we must
2253                          * fall back to a single segment length write.
2254                          *
2255                          * If we didn't fallback here, we could livelock
2256                          * because not all segments in the iov can be copied at
2257                          * once without a pagefault.
2258                          */
2259                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2260                                                 iov_iter_single_seg_count(i));
2261                         goto again;
2262                 }
2263                 iov_iter_advance(i, copied);
2264                 pos += copied;
2265                 written += copied;
2266
2267                 balance_dirty_pages_ratelimited(mapping);
2268
2269         } while (iov_iter_count(i));
2270
2271         return written ? written : status;
2272 }
2273
2274 ssize_t
2275 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2276                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2277                 size_t count, ssize_t written)
2278 {
2279         struct file *file = iocb->ki_filp;
2280         struct address_space *mapping = file->f_mapping;
2281         const struct address_space_operations *a_ops = mapping->a_ops;
2282         struct inode *inode = mapping->host;
2283         ssize_t status;
2284         struct iov_iter i;
2285
2286         iov_iter_init(&i, iov, nr_segs, count, written);
2287         if (a_ops->write_begin)
2288                 status = generic_perform_write(file, &i, pos);
2289         else
2290                 status = generic_perform_write_2copy(file, &i, pos);
2291
2292         if (likely(status >= 0)) {
2293                 written += status;
2294                 *ppos = pos + status;
2295
2296                 /*
2297                  * For now, when the user asks for O_SYNC, we'll actually give
2298                  * O_DSYNC
2299                  */
2300                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2301                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2302                                 status = generic_osync_inode(inode, mapping,
2303                                                 OSYNC_METADATA|OSYNC_DATA);
2304                 }
2305         }
2306         
2307         /*
2308          * If we get here for O_DIRECT writes then we must have fallen through
2309          * to buffered writes (block instantiation inside i_size).  So we sync
2310          * the file data here, to try to honour O_DIRECT expectations.
2311          */
2312         if (unlikely(file->f_flags & O_DIRECT) && written)
2313                 status = filemap_write_and_wait(mapping);
2314
2315         return written ? written : status;
2316 }
2317 EXPORT_SYMBOL(generic_file_buffered_write);
2318
2319 static ssize_t
2320 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2321                                 unsigned long nr_segs, loff_t *ppos)
2322 {
2323         struct file *file = iocb->ki_filp;
2324         struct address_space * mapping = file->f_mapping;
2325         size_t ocount;          /* original count */
2326         size_t count;           /* after file limit checks */
2327         struct inode    *inode = mapping->host;
2328         loff_t          pos;
2329         ssize_t         written;
2330         ssize_t         err;
2331
2332         ocount = 0;
2333         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2334         if (err)
2335                 return err;
2336
2337         count = ocount;
2338         pos = *ppos;
2339
2340         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2341
2342         /* We can write back this queue in page reclaim */
2343         current->backing_dev_info = mapping->backing_dev_info;
2344         written = 0;
2345
2346         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2347         if (err)
2348                 goto out;
2349
2350         if (count == 0)
2351                 goto out;
2352
2353         err = remove_suid(file->f_path.dentry);
2354         if (err)
2355                 goto out;
2356
2357         file_update_time(file);
2358
2359         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2360         if (unlikely(file->f_flags & O_DIRECT)) {
2361                 loff_t endbyte;
2362                 ssize_t written_buffered;
2363
2364                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2365                                                         ppos, count, ocount);
2366                 if (written < 0 || written == count)
2367                         goto out;
2368                 /*
2369                  * direct-io write to a hole: fall through to buffered I/O
2370                  * for completing the rest of the request.
2371                  */
2372                 pos += written;
2373                 count -= written;
2374                 written_buffered = generic_file_buffered_write(iocb, iov,
2375                                                 nr_segs, pos, ppos, count,
2376                                                 written);
2377                 /*
2378                  * If generic_file_buffered_write() retuned a synchronous error
2379                  * then we want to return the number of bytes which were
2380                  * direct-written, or the error code if that was zero.  Note
2381                  * that this differs from normal direct-io semantics, which
2382                  * will return -EFOO even if some bytes were written.
2383                  */
2384                 if (written_buffered < 0) {
2385                         err = written_buffered;
2386                         goto out;
2387                 }
2388
2389                 /*
2390                  * We need to ensure that the page cache pages are written to
2391                  * disk and invalidated to preserve the expected O_DIRECT
2392                  * semantics.
2393                  */
2394                 endbyte = pos + written_buffered - written - 1;
2395                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2396                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2397                                             SYNC_FILE_RANGE_WRITE|
2398                                             SYNC_FILE_RANGE_WAIT_AFTER);
2399                 if (err == 0) {
2400                         written = written_buffered;
2401                         invalidate_mapping_pages(mapping,
2402                                                  pos >> PAGE_CACHE_SHIFT,
2403                                                  endbyte >> PAGE_CACHE_SHIFT);
2404                 } else {
2405                         /*
2406                          * We don't know how much we wrote, so just return
2407                          * the number of bytes which were direct-written
2408                          */
2409                 }
2410         } else {
2411                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2412                                 pos, ppos, count, written);
2413         }
2414 out:
2415         current->backing_dev_info = NULL;
2416         return written ? written : err;
2417 }
2418
2419 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2420                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2421 {
2422         struct file *file = iocb->ki_filp;
2423         struct address_space *mapping = file->f_mapping;
2424         struct inode *inode = mapping->host;
2425         ssize_t ret;
2426
2427         BUG_ON(iocb->ki_pos != pos);
2428
2429         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2430                         &iocb->ki_pos);
2431
2432         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2433                 ssize_t err;
2434
2435                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2436                 if (err < 0)
2437                         ret = err;
2438         }
2439         return ret;
2440 }
2441 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2442
2443 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2444                 unsigned long nr_segs, loff_t pos)
2445 {
2446         struct file *file = iocb->ki_filp;
2447         struct address_space *mapping = file->f_mapping;
2448         struct inode *inode = mapping->host;
2449         ssize_t ret;
2450
2451         BUG_ON(iocb->ki_pos != pos);
2452
2453         mutex_lock(&inode->i_mutex);
2454         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2455                         &iocb->ki_pos);
2456         mutex_unlock(&inode->i_mutex);
2457
2458         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2459                 ssize_t err;
2460
2461                 err = sync_page_range(inode, mapping, pos, ret);
2462                 if (err < 0)
2463                         ret = err;
2464         }
2465         return ret;
2466 }
2467 EXPORT_SYMBOL(generic_file_aio_write);
2468
2469 /*
2470  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2471  * went wrong during pagecache shootdown.
2472  */
2473 static ssize_t
2474 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2475         loff_t offset, unsigned long nr_segs)
2476 {
2477         struct file *file = iocb->ki_filp;
2478         struct address_space *mapping = file->f_mapping;
2479         ssize_t retval;
2480         size_t write_len;
2481         pgoff_t end = 0; /* silence gcc */
2482
2483         /*
2484          * If it's a write, unmap all mmappings of the file up-front.  This
2485          * will cause any pte dirty bits to be propagated into the pageframes
2486          * for the subsequent filemap_write_and_wait().
2487          */
2488         if (rw == WRITE) {
2489                 write_len = iov_length(iov, nr_segs);
2490                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2491                 if (mapping_mapped(mapping))
2492                         unmap_mapping_range(mapping, offset, write_len, 0);
2493         }
2494
2495         retval = filemap_write_and_wait(mapping);
2496         if (retval)
2497                 goto out;
2498
2499         /*
2500          * After a write we want buffered reads to be sure to go to disk to get
2501          * the new data.  We invalidate clean cached page from the region we're
2502          * about to write.  We do this *before* the write so that we can return
2503          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2504          */
2505         if (rw == WRITE && mapping->nrpages) {
2506                 retval = invalidate_inode_pages2_range(mapping,
2507                                         offset >> PAGE_CACHE_SHIFT, end);
2508                 if (retval)
2509                         goto out;
2510         }
2511
2512         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2513         if (retval)
2514                 goto out;
2515
2516         /*
2517          * Finally, try again to invalidate clean pages which might have been
2518          * faulted in by get_user_pages() if the source of the write was an
2519          * mmap()ed region of the file we're writing.  That's a pretty crazy
2520          * thing to do, so we don't support it 100%.  If this invalidation
2521          * fails and we have -EIOCBQUEUED we ignore the failure.
2522          */
2523         if (rw == WRITE && mapping->nrpages) {
2524                 int err = invalidate_inode_pages2_range(mapping,
2525                                               offset >> PAGE_CACHE_SHIFT, end);
2526                 if (err && retval >= 0)
2527                         retval = err;
2528         }
2529 out:
2530         return retval;
2531 }
2532
2533 /**
2534  * try_to_release_page() - release old fs-specific metadata on a page
2535  *
2536  * @page: the page which the kernel is trying to free
2537  * @gfp_mask: memory allocation flags (and I/O mode)
2538  *
2539  * The address_space is to try to release any data against the page
2540  * (presumably at page->private).  If the release was successful, return `1'.
2541  * Otherwise return zero.
2542  *
2543  * The @gfp_mask argument specifies whether I/O may be performed to release
2544  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2545  *
2546  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2547  */
2548 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2549 {
2550         struct address_space * const mapping = page->mapping;
2551
2552         BUG_ON(!PageLocked(page));
2553         if (PageWriteback(page))
2554                 return 0;
2555
2556         if (mapping && mapping->a_ops->releasepage)
2557                 return mapping->a_ops->releasepage(page, gfp_mask);
2558         return try_to_free_buffers(page);
2559 }
2560
2561 EXPORT_SYMBOL(try_to_release_page);