Merge branch 'for-2.6.24' of git://git.secretlab.ca/git/linux-2.6-mpc52xx into merge
[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/pagevec.h>
29 #include <linux/blkdev.h>
30 #include <linux/security.h>
31 #include <linux/syscalls.h>
32 #include <linux/cpuset.h>
33 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34 #include "internal.h"
35
36 /*
37  * FIXME: remove all knowledge of the buffer layer from the core VM
38  */
39 #include <linux/buffer_head.h> /* for generic_osync_inode */
40
41 #include <asm/mman.h>
42
43 static ssize_t
44 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
45         loff_t offset, unsigned long nr_segs);
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_lock               (vmtruncate)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_lock
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  ->i_mutex
82  *    ->i_alloc_sem             (various)
83  *
84  *  ->inode_lock
85  *    ->sb_lock                 (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_lock
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
103  *    ->inode_lock              (zap_pte_range->set_page_dirty)
104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
105  *
106  *  ->task->proc_lock
107  *    ->dcache_lock             (proc_pid_lookup)
108  */
109
110 /*
111  * Remove a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
114  */
115 void __remove_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         BUG_ON(page_mapped(page));
124 }
125
126 void remove_from_page_cache(struct page *page)
127 {
128         struct address_space *mapping = page->mapping;
129
130         BUG_ON(!PageLocked(page));
131
132         write_lock_irq(&mapping->tree_lock);
133         __remove_from_page_cache(page);
134         write_unlock_irq(&mapping->tree_lock);
135 }
136
137 static int sync_page(void *word)
138 {
139         struct address_space *mapping;
140         struct page *page;
141
142         page = container_of((unsigned long *)word, struct page, flags);
143
144         /*
145          * page_mapping() is being called without PG_locked held.
146          * Some knowledge of the state and use of the page is used to
147          * reduce the requirements down to a memory barrier.
148          * The danger here is of a stale page_mapping() return value
149          * indicating a struct address_space different from the one it's
150          * associated with when it is associated with one.
151          * After smp_mb(), it's either the correct page_mapping() for
152          * the page, or an old page_mapping() and the page's own
153          * page_mapping() has gone NULL.
154          * The ->sync_page() address_space operation must tolerate
155          * page_mapping() going NULL. By an amazing coincidence,
156          * this comes about because none of the users of the page
157          * in the ->sync_page() methods make essential use of the
158          * page_mapping(), merely passing the page down to the backing
159          * device's unplug functions when it's non-NULL, which in turn
160          * ignore it for all cases but swap, where only page_private(page) is
161          * of interest. When page_mapping() does go NULL, the entire
162          * call stack gracefully ignores the page and returns.
163          * -- wli
164          */
165         smp_mb();
166         mapping = page_mapping(page);
167         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
168                 mapping->a_ops->sync_page(page);
169         io_schedule();
170         return 0;
171 }
172
173 /**
174  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
175  * @mapping:    address space structure to write
176  * @start:      offset in bytes where the range starts
177  * @end:        offset in bytes where the range ends (inclusive)
178  * @sync_mode:  enable synchronous operation
179  *
180  * Start writeback against all of a mapping's dirty pages that lie
181  * within the byte offsets <start, end> inclusive.
182  *
183  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
184  * opposed to a regular memory cleansing writeback.  The difference between
185  * these two operations is that if a dirty page/buffer is encountered, it must
186  * be waited upon, and not just skipped over.
187  */
188 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
189                                 loff_t end, int sync_mode)
190 {
191         int ret;
192         struct writeback_control wbc = {
193                 .sync_mode = sync_mode,
194                 .nr_to_write = mapping->nrpages * 2,
195                 .range_start = start,
196                 .range_end = end,
197         };
198
199         if (!mapping_cap_writeback_dirty(mapping))
200                 return 0;
201
202         ret = do_writepages(mapping, &wbc);
203         return ret;
204 }
205
206 static inline int __filemap_fdatawrite(struct address_space *mapping,
207         int sync_mode)
208 {
209         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
210 }
211
212 int filemap_fdatawrite(struct address_space *mapping)
213 {
214         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
215 }
216 EXPORT_SYMBOL(filemap_fdatawrite);
217
218 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
219                                 loff_t end)
220 {
221         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
222 }
223
224 /**
225  * filemap_flush - mostly a non-blocking flush
226  * @mapping:    target address_space
227  *
228  * This is a mostly non-blocking flush.  Not suitable for data-integrity
229  * purposes - I/O may not be started against all dirty pages.
230  */
231 int filemap_flush(struct address_space *mapping)
232 {
233         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
234 }
235 EXPORT_SYMBOL(filemap_flush);
236
237 /**
238  * wait_on_page_writeback_range - wait for writeback to complete
239  * @mapping:    target address_space
240  * @start:      beginning page index
241  * @end:        ending page index
242  *
243  * Wait for writeback to complete against pages indexed by start->end
244  * inclusive
245  */
246 int wait_on_page_writeback_range(struct address_space *mapping,
247                                 pgoff_t start, pgoff_t end)
248 {
249         struct pagevec pvec;
250         int nr_pages;
251         int ret = 0;
252         pgoff_t index;
253
254         if (end < start)
255                 return 0;
256
257         pagevec_init(&pvec, 0);
258         index = start;
259         while ((index <= end) &&
260                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
261                         PAGECACHE_TAG_WRITEBACK,
262                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
263                 unsigned i;
264
265                 for (i = 0; i < nr_pages; i++) {
266                         struct page *page = pvec.pages[i];
267
268                         /* until radix tree lookup accepts end_index */
269                         if (page->index > end)
270                                 continue;
271
272                         wait_on_page_writeback(page);
273                         if (PageError(page))
274                                 ret = -EIO;
275                 }
276                 pagevec_release(&pvec);
277                 cond_resched();
278         }
279
280         /* Check for outstanding write errors */
281         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
282                 ret = -ENOSPC;
283         if (test_and_clear_bit(AS_EIO, &mapping->flags))
284                 ret = -EIO;
285
286         return ret;
287 }
288
289 /**
290  * sync_page_range - write and wait on all pages in the passed range
291  * @inode:      target inode
292  * @mapping:    target address_space
293  * @pos:        beginning offset in pages to write
294  * @count:      number of bytes to write
295  *
296  * Write and wait upon all the pages in the passed range.  This is a "data
297  * integrity" operation.  It waits upon in-flight writeout before starting and
298  * waiting upon new writeout.  If there was an IO error, return it.
299  *
300  * We need to re-take i_mutex during the generic_osync_inode list walk because
301  * it is otherwise livelockable.
302  */
303 int sync_page_range(struct inode *inode, struct address_space *mapping,
304                         loff_t pos, loff_t count)
305 {
306         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
307         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
308         int ret;
309
310         if (!mapping_cap_writeback_dirty(mapping) || !count)
311                 return 0;
312         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
313         if (ret == 0) {
314                 mutex_lock(&inode->i_mutex);
315                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
316                 mutex_unlock(&inode->i_mutex);
317         }
318         if (ret == 0)
319                 ret = wait_on_page_writeback_range(mapping, start, end);
320         return ret;
321 }
322 EXPORT_SYMBOL(sync_page_range);
323
324 /**
325  * sync_page_range_nolock
326  * @inode:      target inode
327  * @mapping:    target address_space
328  * @pos:        beginning offset in pages to write
329  * @count:      number of bytes to write
330  *
331  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
332  * as it forces O_SYNC writers to different parts of the same file
333  * to be serialised right until io completion.
334  */
335 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
336                            loff_t pos, loff_t count)
337 {
338         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
339         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
340         int ret;
341
342         if (!mapping_cap_writeback_dirty(mapping) || !count)
343                 return 0;
344         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
345         if (ret == 0)
346                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
347         if (ret == 0)
348                 ret = wait_on_page_writeback_range(mapping, start, end);
349         return ret;
350 }
351 EXPORT_SYMBOL(sync_page_range_nolock);
352
353 /**
354  * filemap_fdatawait - wait for all under-writeback pages to complete
355  * @mapping: address space structure to wait for
356  *
357  * Walk the list of under-writeback pages of the given address space
358  * and wait for all of them.
359  */
360 int filemap_fdatawait(struct address_space *mapping)
361 {
362         loff_t i_size = i_size_read(mapping->host);
363
364         if (i_size == 0)
365                 return 0;
366
367         return wait_on_page_writeback_range(mapping, 0,
368                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
369 }
370 EXPORT_SYMBOL(filemap_fdatawait);
371
372 int filemap_write_and_wait(struct address_space *mapping)
373 {
374         int err = 0;
375
376         if (mapping->nrpages) {
377                 err = filemap_fdatawrite(mapping);
378                 /*
379                  * Even if the above returned error, the pages may be
380                  * written partially (e.g. -ENOSPC), so we wait for it.
381                  * But the -EIO is special case, it may indicate the worst
382                  * thing (e.g. bug) happened, so we avoid waiting for it.
383                  */
384                 if (err != -EIO) {
385                         int err2 = filemap_fdatawait(mapping);
386                         if (!err)
387                                 err = err2;
388                 }
389         }
390         return err;
391 }
392 EXPORT_SYMBOL(filemap_write_and_wait);
393
394 /**
395  * filemap_write_and_wait_range - write out & wait on a file range
396  * @mapping:    the address_space for the pages
397  * @lstart:     offset in bytes where the range starts
398  * @lend:       offset in bytes where the range ends (inclusive)
399  *
400  * Write out and wait upon file offsets lstart->lend, inclusive.
401  *
402  * Note that `lend' is inclusive (describes the last byte to be written) so
403  * that this function can be used to write to the very end-of-file (end = -1).
404  */
405 int filemap_write_and_wait_range(struct address_space *mapping,
406                                  loff_t lstart, loff_t lend)
407 {
408         int err = 0;
409
410         if (mapping->nrpages) {
411                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
412                                                  WB_SYNC_ALL);
413                 /* See comment of filemap_write_and_wait() */
414                 if (err != -EIO) {
415                         int err2 = wait_on_page_writeback_range(mapping,
416                                                 lstart >> PAGE_CACHE_SHIFT,
417                                                 lend >> PAGE_CACHE_SHIFT);
418                         if (!err)
419                                 err = err2;
420                 }
421         }
422         return err;
423 }
424
425 /**
426  * add_to_page_cache - add newly allocated pagecache pages
427  * @page:       page to add
428  * @mapping:    the page's address_space
429  * @offset:     page index
430  * @gfp_mask:   page allocation mode
431  *
432  * This function is used to add newly allocated pagecache pages;
433  * the page is new, so we can just run SetPageLocked() against it.
434  * The other page state flags were set by rmqueue().
435  *
436  * This function does not add the page to the LRU.  The caller must do that.
437  */
438 int add_to_page_cache(struct page *page, struct address_space *mapping,
439                 pgoff_t offset, gfp_t gfp_mask)
440 {
441         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
442
443         if (error == 0) {
444                 write_lock_irq(&mapping->tree_lock);
445                 error = radix_tree_insert(&mapping->page_tree, offset, page);
446                 if (!error) {
447                         page_cache_get(page);
448                         SetPageLocked(page);
449                         page->mapping = mapping;
450                         page->index = offset;
451                         mapping->nrpages++;
452                         __inc_zone_page_state(page, NR_FILE_PAGES);
453                 }
454                 write_unlock_irq(&mapping->tree_lock);
455                 radix_tree_preload_end();
456         }
457         return error;
458 }
459 EXPORT_SYMBOL(add_to_page_cache);
460
461 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
462                                 pgoff_t offset, gfp_t gfp_mask)
463 {
464         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
465         if (ret == 0)
466                 lru_cache_add(page);
467         return ret;
468 }
469
470 #ifdef CONFIG_NUMA
471 struct page *__page_cache_alloc(gfp_t gfp)
472 {
473         if (cpuset_do_page_mem_spread()) {
474                 int n = cpuset_mem_spread_node();
475                 return alloc_pages_node(n, gfp, 0);
476         }
477         return alloc_pages(gfp, 0);
478 }
479 EXPORT_SYMBOL(__page_cache_alloc);
480 #endif
481
482 static int __sleep_on_page_lock(void *word)
483 {
484         io_schedule();
485         return 0;
486 }
487
488 /*
489  * In order to wait for pages to become available there must be
490  * waitqueues associated with pages. By using a hash table of
491  * waitqueues where the bucket discipline is to maintain all
492  * waiters on the same queue and wake all when any of the pages
493  * become available, and for the woken contexts to check to be
494  * sure the appropriate page became available, this saves space
495  * at a cost of "thundering herd" phenomena during rare hash
496  * collisions.
497  */
498 static wait_queue_head_t *page_waitqueue(struct page *page)
499 {
500         const struct zone *zone = page_zone(page);
501
502         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
503 }
504
505 static inline void wake_up_page(struct page *page, int bit)
506 {
507         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
508 }
509
510 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
511 {
512         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
513
514         if (test_bit(bit_nr, &page->flags))
515                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
516                                                         TASK_UNINTERRUPTIBLE);
517 }
518 EXPORT_SYMBOL(wait_on_page_bit);
519
520 /**
521  * unlock_page - unlock a locked page
522  * @page: the page
523  *
524  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
525  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
526  * mechananism between PageLocked pages and PageWriteback pages is shared.
527  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
528  *
529  * The first mb is necessary to safely close the critical section opened by the
530  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
531  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
532  * parallel wait_on_page_locked()).
533  */
534 void fastcall unlock_page(struct page *page)
535 {
536         smp_mb__before_clear_bit();
537         if (!TestClearPageLocked(page))
538                 BUG();
539         smp_mb__after_clear_bit(); 
540         wake_up_page(page, PG_locked);
541 }
542 EXPORT_SYMBOL(unlock_page);
543
544 /**
545  * end_page_writeback - end writeback against a page
546  * @page: the page
547  */
548 void end_page_writeback(struct page *page)
549 {
550         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
551                 if (!test_clear_page_writeback(page))
552                         BUG();
553         }
554         smp_mb__after_clear_bit();
555         wake_up_page(page, PG_writeback);
556 }
557 EXPORT_SYMBOL(end_page_writeback);
558
559 /**
560  * __lock_page - get a lock on the page, assuming we need to sleep to get it
561  * @page: the page to lock
562  *
563  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
564  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
565  * chances are that on the second loop, the block layer's plug list is empty,
566  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
567  */
568 void fastcall __lock_page(struct page *page)
569 {
570         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
571
572         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
573                                                         TASK_UNINTERRUPTIBLE);
574 }
575 EXPORT_SYMBOL(__lock_page);
576
577 /*
578  * Variant of lock_page that does not require the caller to hold a reference
579  * on the page's mapping.
580  */
581 void fastcall __lock_page_nosync(struct page *page)
582 {
583         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
584         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
585                                                         TASK_UNINTERRUPTIBLE);
586 }
587
588 /**
589  * find_get_page - find and get a page reference
590  * @mapping: the address_space to search
591  * @offset: the page index
592  *
593  * Is there a pagecache struct page at the given (mapping, offset) tuple?
594  * If yes, increment its refcount and return it; if no, return NULL.
595  */
596 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
597 {
598         struct page *page;
599
600         read_lock_irq(&mapping->tree_lock);
601         page = radix_tree_lookup(&mapping->page_tree, offset);
602         if (page)
603                 page_cache_get(page);
604         read_unlock_irq(&mapping->tree_lock);
605         return page;
606 }
607 EXPORT_SYMBOL(find_get_page);
608
609 /**
610  * find_lock_page - locate, pin and lock a pagecache page
611  * @mapping: the address_space to search
612  * @offset: the page index
613  *
614  * Locates the desired pagecache page, locks it, increments its reference
615  * count and returns its address.
616  *
617  * Returns zero if the page was not present. find_lock_page() may sleep.
618  */
619 struct page *find_lock_page(struct address_space *mapping,
620                                 pgoff_t offset)
621 {
622         struct page *page;
623
624 repeat:
625         read_lock_irq(&mapping->tree_lock);
626         page = radix_tree_lookup(&mapping->page_tree, offset);
627         if (page) {
628                 page_cache_get(page);
629                 if (TestSetPageLocked(page)) {
630                         read_unlock_irq(&mapping->tree_lock);
631                         __lock_page(page);
632
633                         /* Has the page been truncated while we slept? */
634                         if (unlikely(page->mapping != mapping)) {
635                                 unlock_page(page);
636                                 page_cache_release(page);
637                                 goto repeat;
638                         }
639                         VM_BUG_ON(page->index != offset);
640                         goto out;
641                 }
642         }
643         read_unlock_irq(&mapping->tree_lock);
644 out:
645         return page;
646 }
647 EXPORT_SYMBOL(find_lock_page);
648
649 /**
650  * find_or_create_page - locate or add a pagecache page
651  * @mapping: the page's address_space
652  * @index: the page's index into the mapping
653  * @gfp_mask: page allocation mode
654  *
655  * Locates a page in the pagecache.  If the page is not present, a new page
656  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
657  * LRU list.  The returned page is locked and has its reference count
658  * incremented.
659  *
660  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
661  * allocation!
662  *
663  * find_or_create_page() returns the desired page's address, or zero on
664  * memory exhaustion.
665  */
666 struct page *find_or_create_page(struct address_space *mapping,
667                 pgoff_t index, gfp_t gfp_mask)
668 {
669         struct page *page;
670         int err;
671 repeat:
672         page = find_lock_page(mapping, index);
673         if (!page) {
674                 page = __page_cache_alloc(gfp_mask);
675                 if (!page)
676                         return NULL;
677                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
678                 if (unlikely(err)) {
679                         page_cache_release(page);
680                         page = NULL;
681                         if (err == -EEXIST)
682                                 goto repeat;
683                 }
684         }
685         return page;
686 }
687 EXPORT_SYMBOL(find_or_create_page);
688
689 /**
690  * find_get_pages - gang pagecache lookup
691  * @mapping:    The address_space to search
692  * @start:      The starting page index
693  * @nr_pages:   The maximum number of pages
694  * @pages:      Where the resulting pages are placed
695  *
696  * find_get_pages() will search for and return a group of up to
697  * @nr_pages pages in the mapping.  The pages are placed at @pages.
698  * find_get_pages() takes a reference against the returned pages.
699  *
700  * The search returns a group of mapping-contiguous pages with ascending
701  * indexes.  There may be holes in the indices due to not-present pages.
702  *
703  * find_get_pages() returns the number of pages which were found.
704  */
705 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
706                             unsigned int nr_pages, struct page **pages)
707 {
708         unsigned int i;
709         unsigned int ret;
710
711         read_lock_irq(&mapping->tree_lock);
712         ret = radix_tree_gang_lookup(&mapping->page_tree,
713                                 (void **)pages, start, nr_pages);
714         for (i = 0; i < ret; i++)
715                 page_cache_get(pages[i]);
716         read_unlock_irq(&mapping->tree_lock);
717         return ret;
718 }
719
720 /**
721  * find_get_pages_contig - gang contiguous pagecache lookup
722  * @mapping:    The address_space to search
723  * @index:      The starting page index
724  * @nr_pages:   The maximum number of pages
725  * @pages:      Where the resulting pages are placed
726  *
727  * find_get_pages_contig() works exactly like find_get_pages(), except
728  * that the returned number of pages are guaranteed to be contiguous.
729  *
730  * find_get_pages_contig() returns the number of pages which were found.
731  */
732 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
733                                unsigned int nr_pages, struct page **pages)
734 {
735         unsigned int i;
736         unsigned int ret;
737
738         read_lock_irq(&mapping->tree_lock);
739         ret = radix_tree_gang_lookup(&mapping->page_tree,
740                                 (void **)pages, index, nr_pages);
741         for (i = 0; i < ret; i++) {
742                 if (pages[i]->mapping == NULL || pages[i]->index != index)
743                         break;
744
745                 page_cache_get(pages[i]);
746                 index++;
747         }
748         read_unlock_irq(&mapping->tree_lock);
749         return i;
750 }
751 EXPORT_SYMBOL(find_get_pages_contig);
752
753 /**
754  * find_get_pages_tag - find and return pages that match @tag
755  * @mapping:    the address_space to search
756  * @index:      the starting page index
757  * @tag:        the tag index
758  * @nr_pages:   the maximum number of pages
759  * @pages:      where the resulting pages are placed
760  *
761  * Like find_get_pages, except we only return pages which are tagged with
762  * @tag.   We update @index to index the next page for the traversal.
763  */
764 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
765                         int tag, unsigned int nr_pages, struct page **pages)
766 {
767         unsigned int i;
768         unsigned int ret;
769
770         read_lock_irq(&mapping->tree_lock);
771         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
772                                 (void **)pages, *index, nr_pages, tag);
773         for (i = 0; i < ret; i++)
774                 page_cache_get(pages[i]);
775         if (ret)
776                 *index = pages[ret - 1]->index + 1;
777         read_unlock_irq(&mapping->tree_lock);
778         return ret;
779 }
780 EXPORT_SYMBOL(find_get_pages_tag);
781
782 /**
783  * grab_cache_page_nowait - returns locked page at given index in given cache
784  * @mapping: target address_space
785  * @index: the page index
786  *
787  * Same as grab_cache_page(), but do not wait if the page is unavailable.
788  * This is intended for speculative data generators, where the data can
789  * be regenerated if the page couldn't be grabbed.  This routine should
790  * be safe to call while holding the lock for another page.
791  *
792  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
793  * and deadlock against the caller's locked page.
794  */
795 struct page *
796 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
797 {
798         struct page *page = find_get_page(mapping, index);
799
800         if (page) {
801                 if (!TestSetPageLocked(page))
802                         return page;
803                 page_cache_release(page);
804                 return NULL;
805         }
806         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
807         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
808                 page_cache_release(page);
809                 page = NULL;
810         }
811         return page;
812 }
813 EXPORT_SYMBOL(grab_cache_page_nowait);
814
815 /*
816  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
817  * a _large_ part of the i/o request. Imagine the worst scenario:
818  *
819  *      ---R__________________________________________B__________
820  *         ^ reading here                             ^ bad block(assume 4k)
821  *
822  * read(R) => miss => readahead(R...B) => media error => frustrating retries
823  * => failing the whole request => read(R) => read(R+1) =>
824  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
825  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
826  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
827  *
828  * It is going insane. Fix it by quickly scaling down the readahead size.
829  */
830 static void shrink_readahead_size_eio(struct file *filp,
831                                         struct file_ra_state *ra)
832 {
833         if (!ra->ra_pages)
834                 return;
835
836         ra->ra_pages /= 4;
837 }
838
839 /**
840  * do_generic_mapping_read - generic file read routine
841  * @mapping:    address_space to be read
842  * @_ra:        file's readahead state
843  * @filp:       the file to read
844  * @ppos:       current file position
845  * @desc:       read_descriptor
846  * @actor:      read method
847  *
848  * This is a generic file read routine, and uses the
849  * mapping->a_ops->readpage() function for the actual low-level stuff.
850  *
851  * This is really ugly. But the goto's actually try to clarify some
852  * of the logic when it comes to error handling etc.
853  *
854  * Note the struct file* is only passed for the use of readpage.
855  * It may be NULL.
856  */
857 void do_generic_mapping_read(struct address_space *mapping,
858                              struct file_ra_state *ra,
859                              struct file *filp,
860                              loff_t *ppos,
861                              read_descriptor_t *desc,
862                              read_actor_t actor)
863 {
864         struct inode *inode = mapping->host;
865         pgoff_t index;
866         pgoff_t last_index;
867         pgoff_t prev_index;
868         unsigned long offset;      /* offset into pagecache page */
869         unsigned int prev_offset;
870         int error;
871
872         index = *ppos >> PAGE_CACHE_SHIFT;
873         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
874         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
875         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
876         offset = *ppos & ~PAGE_CACHE_MASK;
877
878         for (;;) {
879                 struct page *page;
880                 pgoff_t end_index;
881                 loff_t isize;
882                 unsigned long nr, ret;
883
884                 cond_resched();
885 find_page:
886                 page = find_get_page(mapping, index);
887                 if (!page) {
888                         page_cache_sync_readahead(mapping,
889                                         ra, filp,
890                                         index, last_index - index);
891                         page = find_get_page(mapping, index);
892                         if (unlikely(page == NULL))
893                                 goto no_cached_page;
894                 }
895                 if (PageReadahead(page)) {
896                         page_cache_async_readahead(mapping,
897                                         ra, filp, page,
898                                         index, last_index - index);
899                 }
900                 if (!PageUptodate(page))
901                         goto page_not_up_to_date;
902 page_ok:
903                 /*
904                  * i_size must be checked after we know the page is Uptodate.
905                  *
906                  * Checking i_size after the check allows us to calculate
907                  * the correct value for "nr", which means the zero-filled
908                  * part of the page is not copied back to userspace (unless
909                  * another truncate extends the file - this is desired though).
910                  */
911
912                 isize = i_size_read(inode);
913                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
914                 if (unlikely(!isize || index > end_index)) {
915                         page_cache_release(page);
916                         goto out;
917                 }
918
919                 /* nr is the maximum number of bytes to copy from this page */
920                 nr = PAGE_CACHE_SIZE;
921                 if (index == end_index) {
922                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
923                         if (nr <= offset) {
924                                 page_cache_release(page);
925                                 goto out;
926                         }
927                 }
928                 nr = nr - offset;
929
930                 /* If users can be writing to this page using arbitrary
931                  * virtual addresses, take care about potential aliasing
932                  * before reading the page on the kernel side.
933                  */
934                 if (mapping_writably_mapped(mapping))
935                         flush_dcache_page(page);
936
937                 /*
938                  * When a sequential read accesses a page several times,
939                  * only mark it as accessed the first time.
940                  */
941                 if (prev_index != index || offset != prev_offset)
942                         mark_page_accessed(page);
943                 prev_index = index;
944
945                 /*
946                  * Ok, we have the page, and it's up-to-date, so
947                  * now we can copy it to user space...
948                  *
949                  * The actor routine returns how many bytes were actually used..
950                  * NOTE! This may not be the same as how much of a user buffer
951                  * we filled up (we may be padding etc), so we can only update
952                  * "pos" here (the actor routine has to update the user buffer
953                  * pointers and the remaining count).
954                  */
955                 ret = actor(desc, page, offset, nr);
956                 offset += ret;
957                 index += offset >> PAGE_CACHE_SHIFT;
958                 offset &= ~PAGE_CACHE_MASK;
959                 prev_offset = offset;
960
961                 page_cache_release(page);
962                 if (ret == nr && desc->count)
963                         continue;
964                 goto out;
965
966 page_not_up_to_date:
967                 /* Get exclusive access to the page ... */
968                 lock_page(page);
969
970                 /* Did it get truncated before we got the lock? */
971                 if (!page->mapping) {
972                         unlock_page(page);
973                         page_cache_release(page);
974                         continue;
975                 }
976
977                 /* Did somebody else fill it already? */
978                 if (PageUptodate(page)) {
979                         unlock_page(page);
980                         goto page_ok;
981                 }
982
983 readpage:
984                 /* Start the actual read. The read will unlock the page. */
985                 error = mapping->a_ops->readpage(filp, page);
986
987                 if (unlikely(error)) {
988                         if (error == AOP_TRUNCATED_PAGE) {
989                                 page_cache_release(page);
990                                 goto find_page;
991                         }
992                         goto readpage_error;
993                 }
994
995                 if (!PageUptodate(page)) {
996                         lock_page(page);
997                         if (!PageUptodate(page)) {
998                                 if (page->mapping == NULL) {
999                                         /*
1000                                          * invalidate_inode_pages got it
1001                                          */
1002                                         unlock_page(page);
1003                                         page_cache_release(page);
1004                                         goto find_page;
1005                                 }
1006                                 unlock_page(page);
1007                                 error = -EIO;
1008                                 shrink_readahead_size_eio(filp, ra);
1009                                 goto readpage_error;
1010                         }
1011                         unlock_page(page);
1012                 }
1013
1014                 goto page_ok;
1015
1016 readpage_error:
1017                 /* UHHUH! A synchronous read error occurred. Report it */
1018                 desc->error = error;
1019                 page_cache_release(page);
1020                 goto out;
1021
1022 no_cached_page:
1023                 /*
1024                  * Ok, it wasn't cached, so we need to create a new
1025                  * page..
1026                  */
1027                 page = page_cache_alloc_cold(mapping);
1028                 if (!page) {
1029                         desc->error = -ENOMEM;
1030                         goto out;
1031                 }
1032                 error = add_to_page_cache_lru(page, mapping,
1033                                                 index, GFP_KERNEL);
1034                 if (error) {
1035                         page_cache_release(page);
1036                         if (error == -EEXIST)
1037                                 goto find_page;
1038                         desc->error = error;
1039                         goto out;
1040                 }
1041                 goto readpage;
1042         }
1043
1044 out:
1045         ra->prev_pos = prev_index;
1046         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1047         ra->prev_pos |= prev_offset;
1048
1049         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1050         if (filp)
1051                 file_accessed(filp);
1052 }
1053 EXPORT_SYMBOL(do_generic_mapping_read);
1054
1055 int file_read_actor(read_descriptor_t *desc, struct page *page,
1056                         unsigned long offset, unsigned long size)
1057 {
1058         char *kaddr;
1059         unsigned long left, count = desc->count;
1060
1061         if (size > count)
1062                 size = count;
1063
1064         /*
1065          * Faults on the destination of a read are common, so do it before
1066          * taking the kmap.
1067          */
1068         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1069                 kaddr = kmap_atomic(page, KM_USER0);
1070                 left = __copy_to_user_inatomic(desc->arg.buf,
1071                                                 kaddr + offset, size);
1072                 kunmap_atomic(kaddr, KM_USER0);
1073                 if (left == 0)
1074                         goto success;
1075         }
1076
1077         /* Do it the slow way */
1078         kaddr = kmap(page);
1079         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1080         kunmap(page);
1081
1082         if (left) {
1083                 size -= left;
1084                 desc->error = -EFAULT;
1085         }
1086 success:
1087         desc->count = count - size;
1088         desc->written += size;
1089         desc->arg.buf += size;
1090         return size;
1091 }
1092
1093 /*
1094  * Performs necessary checks before doing a write
1095  * @iov:        io vector request
1096  * @nr_segs:    number of segments in the iovec
1097  * @count:      number of bytes to write
1098  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1099  *
1100  * Adjust number of segments and amount of bytes to write (nr_segs should be
1101  * properly initialized first). Returns appropriate error code that caller
1102  * should return or zero in case that write should be allowed.
1103  */
1104 int generic_segment_checks(const struct iovec *iov,
1105                         unsigned long *nr_segs, size_t *count, int access_flags)
1106 {
1107         unsigned long   seg;
1108         size_t cnt = 0;
1109         for (seg = 0; seg < *nr_segs; seg++) {
1110                 const struct iovec *iv = &iov[seg];
1111
1112                 /*
1113                  * If any segment has a negative length, or the cumulative
1114                  * length ever wraps negative then return -EINVAL.
1115                  */
1116                 cnt += iv->iov_len;
1117                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1118                         return -EINVAL;
1119                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1120                         continue;
1121                 if (seg == 0)
1122                         return -EFAULT;
1123                 *nr_segs = seg;
1124                 cnt -= iv->iov_len;     /* This segment is no good */
1125                 break;
1126         }
1127         *count = cnt;
1128         return 0;
1129 }
1130 EXPORT_SYMBOL(generic_segment_checks);
1131
1132 /**
1133  * generic_file_aio_read - generic filesystem read routine
1134  * @iocb:       kernel I/O control block
1135  * @iov:        io vector request
1136  * @nr_segs:    number of segments in the iovec
1137  * @pos:        current file position
1138  *
1139  * This is the "read()" routine for all filesystems
1140  * that can use the page cache directly.
1141  */
1142 ssize_t
1143 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1144                 unsigned long nr_segs, loff_t pos)
1145 {
1146         struct file *filp = iocb->ki_filp;
1147         ssize_t retval;
1148         unsigned long seg;
1149         size_t count;
1150         loff_t *ppos = &iocb->ki_pos;
1151
1152         count = 0;
1153         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1154         if (retval)
1155                 return retval;
1156
1157         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1158         if (filp->f_flags & O_DIRECT) {
1159                 loff_t size;
1160                 struct address_space *mapping;
1161                 struct inode *inode;
1162
1163                 mapping = filp->f_mapping;
1164                 inode = mapping->host;
1165                 retval = 0;
1166                 if (!count)
1167                         goto out; /* skip atime */
1168                 size = i_size_read(inode);
1169                 if (pos < size) {
1170                         retval = generic_file_direct_IO(READ, iocb,
1171                                                 iov, pos, nr_segs);
1172                         if (retval > 0)
1173                                 *ppos = pos + retval;
1174                 }
1175                 if (likely(retval != 0)) {
1176                         file_accessed(filp);
1177                         goto out;
1178                 }
1179         }
1180
1181         retval = 0;
1182         if (count) {
1183                 for (seg = 0; seg < nr_segs; seg++) {
1184                         read_descriptor_t desc;
1185
1186                         desc.written = 0;
1187                         desc.arg.buf = iov[seg].iov_base;
1188                         desc.count = iov[seg].iov_len;
1189                         if (desc.count == 0)
1190                                 continue;
1191                         desc.error = 0;
1192                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1193                         retval += desc.written;
1194                         if (desc.error) {
1195                                 retval = retval ?: desc.error;
1196                                 break;
1197                         }
1198                         if (desc.count > 0)
1199                                 break;
1200                 }
1201         }
1202 out:
1203         return retval;
1204 }
1205 EXPORT_SYMBOL(generic_file_aio_read);
1206
1207 static ssize_t
1208 do_readahead(struct address_space *mapping, struct file *filp,
1209              pgoff_t index, unsigned long nr)
1210 {
1211         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1212                 return -EINVAL;
1213
1214         force_page_cache_readahead(mapping, filp, index,
1215                                         max_sane_readahead(nr));
1216         return 0;
1217 }
1218
1219 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1220 {
1221         ssize_t ret;
1222         struct file *file;
1223
1224         ret = -EBADF;
1225         file = fget(fd);
1226         if (file) {
1227                 if (file->f_mode & FMODE_READ) {
1228                         struct address_space *mapping = file->f_mapping;
1229                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1230                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1231                         unsigned long len = end - start + 1;
1232                         ret = do_readahead(mapping, file, start, len);
1233                 }
1234                 fput(file);
1235         }
1236         return ret;
1237 }
1238
1239 #ifdef CONFIG_MMU
1240 /**
1241  * page_cache_read - adds requested page to the page cache if not already there
1242  * @file:       file to read
1243  * @offset:     page index
1244  *
1245  * This adds the requested page to the page cache if it isn't already there,
1246  * and schedules an I/O to read in its contents from disk.
1247  */
1248 static int fastcall page_cache_read(struct file * file, pgoff_t offset)
1249 {
1250         struct address_space *mapping = file->f_mapping;
1251         struct page *page; 
1252         int ret;
1253
1254         do {
1255                 page = page_cache_alloc_cold(mapping);
1256                 if (!page)
1257                         return -ENOMEM;
1258
1259                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1260                 if (ret == 0)
1261                         ret = mapping->a_ops->readpage(file, page);
1262                 else if (ret == -EEXIST)
1263                         ret = 0; /* losing race to add is OK */
1264
1265                 page_cache_release(page);
1266
1267         } while (ret == AOP_TRUNCATED_PAGE);
1268                 
1269         return ret;
1270 }
1271
1272 #define MMAP_LOTSAMISS  (100)
1273
1274 /**
1275  * filemap_fault - read in file data for page fault handling
1276  * @vma:        vma in which the fault was taken
1277  * @vmf:        struct vm_fault containing details of the fault
1278  *
1279  * filemap_fault() is invoked via the vma operations vector for a
1280  * mapped memory region to read in file data during a page fault.
1281  *
1282  * The goto's are kind of ugly, but this streamlines the normal case of having
1283  * it in the page cache, and handles the special cases reasonably without
1284  * having a lot of duplicated code.
1285  */
1286 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1287 {
1288         int error;
1289         struct file *file = vma->vm_file;
1290         struct address_space *mapping = file->f_mapping;
1291         struct file_ra_state *ra = &file->f_ra;
1292         struct inode *inode = mapping->host;
1293         struct page *page;
1294         unsigned long size;
1295         int did_readaround = 0;
1296         int ret = 0;
1297
1298         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1299         if (vmf->pgoff >= size)
1300                 goto outside_data_content;
1301
1302         /* If we don't want any read-ahead, don't bother */
1303         if (VM_RandomReadHint(vma))
1304                 goto no_cached_page;
1305
1306         /*
1307          * Do we have something in the page cache already?
1308          */
1309 retry_find:
1310         page = find_lock_page(mapping, vmf->pgoff);
1311         /*
1312          * For sequential accesses, we use the generic readahead logic.
1313          */
1314         if (VM_SequentialReadHint(vma)) {
1315                 if (!page) {
1316                         page_cache_sync_readahead(mapping, ra, file,
1317                                                            vmf->pgoff, 1);
1318                         page = find_lock_page(mapping, vmf->pgoff);
1319                         if (!page)
1320                                 goto no_cached_page;
1321                 }
1322                 if (PageReadahead(page)) {
1323                         page_cache_async_readahead(mapping, ra, file, page,
1324                                                            vmf->pgoff, 1);
1325                 }
1326         }
1327
1328         if (!page) {
1329                 unsigned long ra_pages;
1330
1331                 ra->mmap_miss++;
1332
1333                 /*
1334                  * Do we miss much more than hit in this file? If so,
1335                  * stop bothering with read-ahead. It will only hurt.
1336                  */
1337                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1338                         goto no_cached_page;
1339
1340                 /*
1341                  * To keep the pgmajfault counter straight, we need to
1342                  * check did_readaround, as this is an inner loop.
1343                  */
1344                 if (!did_readaround) {
1345                         ret = VM_FAULT_MAJOR;
1346                         count_vm_event(PGMAJFAULT);
1347                 }
1348                 did_readaround = 1;
1349                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1350                 if (ra_pages) {
1351                         pgoff_t start = 0;
1352
1353                         if (vmf->pgoff > ra_pages / 2)
1354                                 start = vmf->pgoff - ra_pages / 2;
1355                         do_page_cache_readahead(mapping, file, start, ra_pages);
1356                 }
1357                 page = find_lock_page(mapping, vmf->pgoff);
1358                 if (!page)
1359                         goto no_cached_page;
1360         }
1361
1362         if (!did_readaround)
1363                 ra->mmap_miss--;
1364
1365         /*
1366          * We have a locked page in the page cache, now we need to check
1367          * that it's up-to-date. If not, it is going to be due to an error.
1368          */
1369         if (unlikely(!PageUptodate(page)))
1370                 goto page_not_uptodate;
1371
1372         /* Must recheck i_size under page lock */
1373         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1374         if (unlikely(vmf->pgoff >= size)) {
1375                 unlock_page(page);
1376                 page_cache_release(page);
1377                 goto outside_data_content;
1378         }
1379
1380         /*
1381          * Found the page and have a reference on it.
1382          */
1383         mark_page_accessed(page);
1384         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1385         vmf->page = page;
1386         return ret | VM_FAULT_LOCKED;
1387
1388 outside_data_content:
1389         /*
1390          * An external ptracer can access pages that normally aren't
1391          * accessible..
1392          */
1393         if (vma->vm_mm == current->mm)
1394                 return VM_FAULT_SIGBUS;
1395
1396         /* Fall through to the non-read-ahead case */
1397 no_cached_page:
1398         /*
1399          * We're only likely to ever get here if MADV_RANDOM is in
1400          * effect.
1401          */
1402         error = page_cache_read(file, vmf->pgoff);
1403
1404         /*
1405          * The page we want has now been added to the page cache.
1406          * In the unlikely event that someone removed it in the
1407          * meantime, we'll just come back here and read it again.
1408          */
1409         if (error >= 0)
1410                 goto retry_find;
1411
1412         /*
1413          * An error return from page_cache_read can result if the
1414          * system is low on memory, or a problem occurs while trying
1415          * to schedule I/O.
1416          */
1417         if (error == -ENOMEM)
1418                 return VM_FAULT_OOM;
1419         return VM_FAULT_SIGBUS;
1420
1421 page_not_uptodate:
1422         /* IO error path */
1423         if (!did_readaround) {
1424                 ret = VM_FAULT_MAJOR;
1425                 count_vm_event(PGMAJFAULT);
1426         }
1427
1428         /*
1429          * Umm, take care of errors if the page isn't up-to-date.
1430          * Try to re-read it _once_. We do this synchronously,
1431          * because there really aren't any performance issues here
1432          * and we need to check for errors.
1433          */
1434         ClearPageError(page);
1435         error = mapping->a_ops->readpage(file, page);
1436         page_cache_release(page);
1437
1438         if (!error || error == AOP_TRUNCATED_PAGE)
1439                 goto retry_find;
1440
1441         /* Things didn't work out. Return zero to tell the mm layer so. */
1442         shrink_readahead_size_eio(file, ra);
1443         return VM_FAULT_SIGBUS;
1444 }
1445 EXPORT_SYMBOL(filemap_fault);
1446
1447 struct vm_operations_struct generic_file_vm_ops = {
1448         .fault          = filemap_fault,
1449 };
1450
1451 /* This is used for a general mmap of a disk file */
1452
1453 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1454 {
1455         struct address_space *mapping = file->f_mapping;
1456
1457         if (!mapping->a_ops->readpage)
1458                 return -ENOEXEC;
1459         file_accessed(file);
1460         vma->vm_ops = &generic_file_vm_ops;
1461         vma->vm_flags |= VM_CAN_NONLINEAR;
1462         return 0;
1463 }
1464
1465 /*
1466  * This is for filesystems which do not implement ->writepage.
1467  */
1468 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1469 {
1470         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1471                 return -EINVAL;
1472         return generic_file_mmap(file, vma);
1473 }
1474 #else
1475 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1476 {
1477         return -ENOSYS;
1478 }
1479 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1480 {
1481         return -ENOSYS;
1482 }
1483 #endif /* CONFIG_MMU */
1484
1485 EXPORT_SYMBOL(generic_file_mmap);
1486 EXPORT_SYMBOL(generic_file_readonly_mmap);
1487
1488 static struct page *__read_cache_page(struct address_space *mapping,
1489                                 pgoff_t index,
1490                                 int (*filler)(void *,struct page*),
1491                                 void *data)
1492 {
1493         struct page *page;
1494         int err;
1495 repeat:
1496         page = find_get_page(mapping, index);
1497         if (!page) {
1498                 page = page_cache_alloc_cold(mapping);
1499                 if (!page)
1500                         return ERR_PTR(-ENOMEM);
1501                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1502                 if (unlikely(err)) {
1503                         page_cache_release(page);
1504                         if (err == -EEXIST)
1505                                 goto repeat;
1506                         /* Presumably ENOMEM for radix tree node */
1507                         return ERR_PTR(err);
1508                 }
1509                 err = filler(data, page);
1510                 if (err < 0) {
1511                         page_cache_release(page);
1512                         page = ERR_PTR(err);
1513                 }
1514         }
1515         return page;
1516 }
1517
1518 /*
1519  * Same as read_cache_page, but don't wait for page to become unlocked
1520  * after submitting it to the filler.
1521  */
1522 struct page *read_cache_page_async(struct address_space *mapping,
1523                                 pgoff_t index,
1524                                 int (*filler)(void *,struct page*),
1525                                 void *data)
1526 {
1527         struct page *page;
1528         int err;
1529
1530 retry:
1531         page = __read_cache_page(mapping, index, filler, data);
1532         if (IS_ERR(page))
1533                 return page;
1534         if (PageUptodate(page))
1535                 goto out;
1536
1537         lock_page(page);
1538         if (!page->mapping) {
1539                 unlock_page(page);
1540                 page_cache_release(page);
1541                 goto retry;
1542         }
1543         if (PageUptodate(page)) {
1544                 unlock_page(page);
1545                 goto out;
1546         }
1547         err = filler(data, page);
1548         if (err < 0) {
1549                 page_cache_release(page);
1550                 return ERR_PTR(err);
1551         }
1552 out:
1553         mark_page_accessed(page);
1554         return page;
1555 }
1556 EXPORT_SYMBOL(read_cache_page_async);
1557
1558 /**
1559  * read_cache_page - read into page cache, fill it if needed
1560  * @mapping:    the page's address_space
1561  * @index:      the page index
1562  * @filler:     function to perform the read
1563  * @data:       destination for read data
1564  *
1565  * Read into the page cache. If a page already exists, and PageUptodate() is
1566  * not set, try to fill the page then wait for it to become unlocked.
1567  *
1568  * If the page does not get brought uptodate, return -EIO.
1569  */
1570 struct page *read_cache_page(struct address_space *mapping,
1571                                 pgoff_t index,
1572                                 int (*filler)(void *,struct page*),
1573                                 void *data)
1574 {
1575         struct page *page;
1576
1577         page = read_cache_page_async(mapping, index, filler, data);
1578         if (IS_ERR(page))
1579                 goto out;
1580         wait_on_page_locked(page);
1581         if (!PageUptodate(page)) {
1582                 page_cache_release(page);
1583                 page = ERR_PTR(-EIO);
1584         }
1585  out:
1586         return page;
1587 }
1588 EXPORT_SYMBOL(read_cache_page);
1589
1590 /*
1591  * The logic we want is
1592  *
1593  *      if suid or (sgid and xgrp)
1594  *              remove privs
1595  */
1596 int should_remove_suid(struct dentry *dentry)
1597 {
1598         mode_t mode = dentry->d_inode->i_mode;
1599         int kill = 0;
1600
1601         /* suid always must be killed */
1602         if (unlikely(mode & S_ISUID))
1603                 kill = ATTR_KILL_SUID;
1604
1605         /*
1606          * sgid without any exec bits is just a mandatory locking mark; leave
1607          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1608          */
1609         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1610                 kill |= ATTR_KILL_SGID;
1611
1612         if (unlikely(kill && !capable(CAP_FSETID)))
1613                 return kill;
1614
1615         return 0;
1616 }
1617 EXPORT_SYMBOL(should_remove_suid);
1618
1619 int __remove_suid(struct dentry *dentry, int kill)
1620 {
1621         struct iattr newattrs;
1622
1623         newattrs.ia_valid = ATTR_FORCE | kill;
1624         return notify_change(dentry, &newattrs);
1625 }
1626
1627 int remove_suid(struct dentry *dentry)
1628 {
1629         int kill = should_remove_suid(dentry);
1630
1631         if (unlikely(kill))
1632                 return __remove_suid(dentry, kill);
1633
1634         return 0;
1635 }
1636 EXPORT_SYMBOL(remove_suid);
1637
1638 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1639                         const struct iovec *iov, size_t base, size_t bytes)
1640 {
1641         size_t copied = 0, left = 0;
1642
1643         while (bytes) {
1644                 char __user *buf = iov->iov_base + base;
1645                 int copy = min(bytes, iov->iov_len - base);
1646
1647                 base = 0;
1648                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1649                 copied += copy;
1650                 bytes -= copy;
1651                 vaddr += copy;
1652                 iov++;
1653
1654                 if (unlikely(left))
1655                         break;
1656         }
1657         return copied - left;
1658 }
1659
1660 /*
1661  * Copy as much as we can into the page and return the number of bytes which
1662  * were sucessfully copied.  If a fault is encountered then return the number of
1663  * bytes which were copied.
1664  */
1665 size_t iov_iter_copy_from_user_atomic(struct page *page,
1666                 struct iov_iter *i, unsigned long offset, size_t bytes)
1667 {
1668         char *kaddr;
1669         size_t copied;
1670
1671         BUG_ON(!in_atomic());
1672         kaddr = kmap_atomic(page, KM_USER0);
1673         if (likely(i->nr_segs == 1)) {
1674                 int left;
1675                 char __user *buf = i->iov->iov_base + i->iov_offset;
1676                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1677                                                         buf, bytes);
1678                 copied = bytes - left;
1679         } else {
1680                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1681                                                 i->iov, i->iov_offset, bytes);
1682         }
1683         kunmap_atomic(kaddr, KM_USER0);
1684
1685         return copied;
1686 }
1687 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1688
1689 /*
1690  * This has the same sideeffects and return value as
1691  * iov_iter_copy_from_user_atomic().
1692  * The difference is that it attempts to resolve faults.
1693  * Page must not be locked.
1694  */
1695 size_t iov_iter_copy_from_user(struct page *page,
1696                 struct iov_iter *i, unsigned long offset, size_t bytes)
1697 {
1698         char *kaddr;
1699         size_t copied;
1700
1701         kaddr = kmap(page);
1702         if (likely(i->nr_segs == 1)) {
1703                 int left;
1704                 char __user *buf = i->iov->iov_base + i->iov_offset;
1705                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1706                 copied = bytes - left;
1707         } else {
1708                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1709                                                 i->iov, i->iov_offset, bytes);
1710         }
1711         kunmap(page);
1712         return copied;
1713 }
1714 EXPORT_SYMBOL(iov_iter_copy_from_user);
1715
1716 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1717 {
1718         if (likely(i->nr_segs == 1)) {
1719                 i->iov_offset += bytes;
1720         } else {
1721                 const struct iovec *iov = i->iov;
1722                 size_t base = i->iov_offset;
1723
1724                 while (bytes) {
1725                         int copy = min(bytes, iov->iov_len - base);
1726
1727                         bytes -= copy;
1728                         base += copy;
1729                         if (iov->iov_len == base) {
1730                                 iov++;
1731                                 base = 0;
1732                         }
1733                 }
1734                 i->iov = iov;
1735                 i->iov_offset = base;
1736         }
1737 }
1738
1739 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1740 {
1741         BUG_ON(i->count < bytes);
1742
1743         __iov_iter_advance_iov(i, bytes);
1744         i->count -= bytes;
1745 }
1746 EXPORT_SYMBOL(iov_iter_advance);
1747
1748 /*
1749  * Fault in the first iovec of the given iov_iter, to a maximum length
1750  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1751  * accessed (ie. because it is an invalid address).
1752  *
1753  * writev-intensive code may want this to prefault several iovecs -- that
1754  * would be possible (callers must not rely on the fact that _only_ the
1755  * first iovec will be faulted with the current implementation).
1756  */
1757 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1758 {
1759         char __user *buf = i->iov->iov_base + i->iov_offset;
1760         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1761         return fault_in_pages_readable(buf, bytes);
1762 }
1763 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1764
1765 /*
1766  * Return the count of just the current iov_iter segment.
1767  */
1768 size_t iov_iter_single_seg_count(struct iov_iter *i)
1769 {
1770         const struct iovec *iov = i->iov;
1771         if (i->nr_segs == 1)
1772                 return i->count;
1773         else
1774                 return min(i->count, iov->iov_len - i->iov_offset);
1775 }
1776 EXPORT_SYMBOL(iov_iter_single_seg_count);
1777
1778 /*
1779  * Performs necessary checks before doing a write
1780  *
1781  * Can adjust writing position or amount of bytes to write.
1782  * Returns appropriate error code that caller should return or
1783  * zero in case that write should be allowed.
1784  */
1785 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1786 {
1787         struct inode *inode = file->f_mapping->host;
1788         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1789
1790         if (unlikely(*pos < 0))
1791                 return -EINVAL;
1792
1793         if (!isblk) {
1794                 /* FIXME: this is for backwards compatibility with 2.4 */
1795                 if (file->f_flags & O_APPEND)
1796                         *pos = i_size_read(inode);
1797
1798                 if (limit != RLIM_INFINITY) {
1799                         if (*pos >= limit) {
1800                                 send_sig(SIGXFSZ, current, 0);
1801                                 return -EFBIG;
1802                         }
1803                         if (*count > limit - (typeof(limit))*pos) {
1804                                 *count = limit - (typeof(limit))*pos;
1805                         }
1806                 }
1807         }
1808
1809         /*
1810          * LFS rule
1811          */
1812         if (unlikely(*pos + *count > MAX_NON_LFS &&
1813                                 !(file->f_flags & O_LARGEFILE))) {
1814                 if (*pos >= MAX_NON_LFS) {
1815                         return -EFBIG;
1816                 }
1817                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1818                         *count = MAX_NON_LFS - (unsigned long)*pos;
1819                 }
1820         }
1821
1822         /*
1823          * Are we about to exceed the fs block limit ?
1824          *
1825          * If we have written data it becomes a short write.  If we have
1826          * exceeded without writing data we send a signal and return EFBIG.
1827          * Linus frestrict idea will clean these up nicely..
1828          */
1829         if (likely(!isblk)) {
1830                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1831                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1832                                 return -EFBIG;
1833                         }
1834                         /* zero-length writes at ->s_maxbytes are OK */
1835                 }
1836
1837                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1838                         *count = inode->i_sb->s_maxbytes - *pos;
1839         } else {
1840 #ifdef CONFIG_BLOCK
1841                 loff_t isize;
1842                 if (bdev_read_only(I_BDEV(inode)))
1843                         return -EPERM;
1844                 isize = i_size_read(inode);
1845                 if (*pos >= isize) {
1846                         if (*count || *pos > isize)
1847                                 return -ENOSPC;
1848                 }
1849
1850                 if (*pos + *count > isize)
1851                         *count = isize - *pos;
1852 #else
1853                 return -EPERM;
1854 #endif
1855         }
1856         return 0;
1857 }
1858 EXPORT_SYMBOL(generic_write_checks);
1859
1860 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1861                                 loff_t pos, unsigned len, unsigned flags,
1862                                 struct page **pagep, void **fsdata)
1863 {
1864         const struct address_space_operations *aops = mapping->a_ops;
1865
1866         if (aops->write_begin) {
1867                 return aops->write_begin(file, mapping, pos, len, flags,
1868                                                         pagep, fsdata);
1869         } else {
1870                 int ret;
1871                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1872                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1873                 struct inode *inode = mapping->host;
1874                 struct page *page;
1875 again:
1876                 page = __grab_cache_page(mapping, index);
1877                 *pagep = page;
1878                 if (!page)
1879                         return -ENOMEM;
1880
1881                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1882                         /*
1883                          * There is no way to resolve a short write situation
1884                          * for a !Uptodate page (except by double copying in
1885                          * the caller done by generic_perform_write_2copy).
1886                          *
1887                          * Instead, we have to bring it uptodate here.
1888                          */
1889                         ret = aops->readpage(file, page);
1890                         page_cache_release(page);
1891                         if (ret) {
1892                                 if (ret == AOP_TRUNCATED_PAGE)
1893                                         goto again;
1894                                 return ret;
1895                         }
1896                         goto again;
1897                 }
1898
1899                 ret = aops->prepare_write(file, page, offset, offset+len);
1900                 if (ret) {
1901                         unlock_page(page);
1902                         page_cache_release(page);
1903                         if (pos + len > inode->i_size)
1904                                 vmtruncate(inode, inode->i_size);
1905                 }
1906                 return ret;
1907         }
1908 }
1909 EXPORT_SYMBOL(pagecache_write_begin);
1910
1911 int pagecache_write_end(struct file *file, struct address_space *mapping,
1912                                 loff_t pos, unsigned len, unsigned copied,
1913                                 struct page *page, void *fsdata)
1914 {
1915         const struct address_space_operations *aops = mapping->a_ops;
1916         int ret;
1917
1918         if (aops->write_end) {
1919                 mark_page_accessed(page);
1920                 ret = aops->write_end(file, mapping, pos, len, copied,
1921                                                         page, fsdata);
1922         } else {
1923                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1924                 struct inode *inode = mapping->host;
1925
1926                 flush_dcache_page(page);
1927                 ret = aops->commit_write(file, page, offset, offset+len);
1928                 unlock_page(page);
1929                 mark_page_accessed(page);
1930                 page_cache_release(page);
1931
1932                 if (ret < 0) {
1933                         if (pos + len > inode->i_size)
1934                                 vmtruncate(inode, inode->i_size);
1935                 } else if (ret > 0)
1936                         ret = min_t(size_t, copied, ret);
1937                 else
1938                         ret = copied;
1939         }
1940
1941         return ret;
1942 }
1943 EXPORT_SYMBOL(pagecache_write_end);
1944
1945 ssize_t
1946 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1947                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1948                 size_t count, size_t ocount)
1949 {
1950         struct file     *file = iocb->ki_filp;
1951         struct address_space *mapping = file->f_mapping;
1952         struct inode    *inode = mapping->host;
1953         ssize_t         written;
1954
1955         if (count != ocount)
1956                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1957
1958         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1959         if (written > 0) {
1960                 loff_t end = pos + written;
1961                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1962                         i_size_write(inode,  end);
1963                         mark_inode_dirty(inode);
1964                 }
1965                 *ppos = end;
1966         }
1967
1968         /*
1969          * Sync the fs metadata but not the minor inode changes and
1970          * of course not the data as we did direct DMA for the IO.
1971          * i_mutex is held, which protects generic_osync_inode() from
1972          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
1973          */
1974         if ((written >= 0 || written == -EIOCBQUEUED) &&
1975             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1976                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1977                 if (err < 0)
1978                         written = err;
1979         }
1980         return written;
1981 }
1982 EXPORT_SYMBOL(generic_file_direct_write);
1983
1984 /*
1985  * Find or create a page at the given pagecache position. Return the locked
1986  * page. This function is specifically for buffered writes.
1987  */
1988 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
1989 {
1990         int status;
1991         struct page *page;
1992 repeat:
1993         page = find_lock_page(mapping, index);
1994         if (likely(page))
1995                 return page;
1996
1997         page = page_cache_alloc(mapping);
1998         if (!page)
1999                 return NULL;
2000         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2001         if (unlikely(status)) {
2002                 page_cache_release(page);
2003                 if (status == -EEXIST)
2004                         goto repeat;
2005                 return NULL;
2006         }
2007         return page;
2008 }
2009 EXPORT_SYMBOL(__grab_cache_page);
2010
2011 static ssize_t generic_perform_write_2copy(struct file *file,
2012                                 struct iov_iter *i, loff_t pos)
2013 {
2014         struct address_space *mapping = file->f_mapping;
2015         const struct address_space_operations *a_ops = mapping->a_ops;
2016         struct inode *inode = mapping->host;
2017         long status = 0;
2018         ssize_t written = 0;
2019
2020         do {
2021                 struct page *src_page;
2022                 struct page *page;
2023                 pgoff_t index;          /* Pagecache index for current page */
2024                 unsigned long offset;   /* Offset into pagecache page */
2025                 unsigned long bytes;    /* Bytes to write to page */
2026                 size_t copied;          /* Bytes copied from user */
2027
2028                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2029                 index = pos >> PAGE_CACHE_SHIFT;
2030                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2031                                                 iov_iter_count(i));
2032
2033                 /*
2034                  * a non-NULL src_page indicates that we're doing the
2035                  * copy via get_user_pages and kmap.
2036                  */
2037                 src_page = NULL;
2038
2039                 /*
2040                  * Bring in the user page that we will copy from _first_.
2041                  * Otherwise there's a nasty deadlock on copying from the
2042                  * same page as we're writing to, without it being marked
2043                  * up-to-date.
2044                  *
2045                  * Not only is this an optimisation, but it is also required
2046                  * to check that the address is actually valid, when atomic
2047                  * usercopies are used, below.
2048                  */
2049                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2050                         status = -EFAULT;
2051                         break;
2052                 }
2053
2054                 page = __grab_cache_page(mapping, index);
2055                 if (!page) {
2056                         status = -ENOMEM;
2057                         break;
2058                 }
2059
2060                 /*
2061                  * non-uptodate pages cannot cope with short copies, and we
2062                  * cannot take a pagefault with the destination page locked.
2063                  * So pin the source page to copy it.
2064                  */
2065                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2066                         unlock_page(page);
2067
2068                         src_page = alloc_page(GFP_KERNEL);
2069                         if (!src_page) {
2070                                 page_cache_release(page);
2071                                 status = -ENOMEM;
2072                                 break;
2073                         }
2074
2075                         /*
2076                          * Cannot get_user_pages with a page locked for the
2077                          * same reason as we can't take a page fault with a
2078                          * page locked (as explained below).
2079                          */
2080                         copied = iov_iter_copy_from_user(src_page, i,
2081                                                                 offset, bytes);
2082                         if (unlikely(copied == 0)) {
2083                                 status = -EFAULT;
2084                                 page_cache_release(page);
2085                                 page_cache_release(src_page);
2086                                 break;
2087                         }
2088                         bytes = copied;
2089
2090                         lock_page(page);
2091                         /*
2092                          * Can't handle the page going uptodate here, because
2093                          * that means we would use non-atomic usercopies, which
2094                          * zero out the tail of the page, which can cause
2095                          * zeroes to become transiently visible. We could just
2096                          * use a non-zeroing copy, but the APIs aren't too
2097                          * consistent.
2098                          */
2099                         if (unlikely(!page->mapping || PageUptodate(page))) {
2100                                 unlock_page(page);
2101                                 page_cache_release(page);
2102                                 page_cache_release(src_page);
2103                                 continue;
2104                         }
2105                 }
2106
2107                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2108                 if (unlikely(status))
2109                         goto fs_write_aop_error;
2110
2111                 if (!src_page) {
2112                         /*
2113                          * Must not enter the pagefault handler here, because
2114                          * we hold the page lock, so we might recursively
2115                          * deadlock on the same lock, or get an ABBA deadlock
2116                          * against a different lock, or against the mmap_sem
2117                          * (which nests outside the page lock).  So increment
2118                          * preempt count, and use _atomic usercopies.
2119                          *
2120                          * The page is uptodate so we are OK to encounter a
2121                          * short copy: if unmodified parts of the page are
2122                          * marked dirty and written out to disk, it doesn't
2123                          * really matter.
2124                          */
2125                         pagefault_disable();
2126                         copied = iov_iter_copy_from_user_atomic(page, i,
2127                                                                 offset, bytes);
2128                         pagefault_enable();
2129                 } else {
2130                         void *src, *dst;
2131                         src = kmap_atomic(src_page, KM_USER0);
2132                         dst = kmap_atomic(page, KM_USER1);
2133                         memcpy(dst + offset, src + offset, bytes);
2134                         kunmap_atomic(dst, KM_USER1);
2135                         kunmap_atomic(src, KM_USER0);
2136                         copied = bytes;
2137                 }
2138                 flush_dcache_page(page);
2139
2140                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2141                 if (unlikely(status < 0))
2142                         goto fs_write_aop_error;
2143                 if (unlikely(status > 0)) /* filesystem did partial write */
2144                         copied = min_t(size_t, copied, status);
2145
2146                 unlock_page(page);
2147                 mark_page_accessed(page);
2148                 page_cache_release(page);
2149                 if (src_page)
2150                         page_cache_release(src_page);
2151
2152                 iov_iter_advance(i, copied);
2153                 pos += copied;
2154                 written += copied;
2155
2156                 balance_dirty_pages_ratelimited(mapping);
2157                 cond_resched();
2158                 continue;
2159
2160 fs_write_aop_error:
2161                 unlock_page(page);
2162                 page_cache_release(page);
2163                 if (src_page)
2164                         page_cache_release(src_page);
2165
2166                 /*
2167                  * prepare_write() may have instantiated a few blocks
2168                  * outside i_size.  Trim these off again. Don't need
2169                  * i_size_read because we hold i_mutex.
2170                  */
2171                 if (pos + bytes > inode->i_size)
2172                         vmtruncate(inode, inode->i_size);
2173                 break;
2174         } while (iov_iter_count(i));
2175
2176         return written ? written : status;
2177 }
2178
2179 static ssize_t generic_perform_write(struct file *file,
2180                                 struct iov_iter *i, loff_t pos)
2181 {
2182         struct address_space *mapping = file->f_mapping;
2183         const struct address_space_operations *a_ops = mapping->a_ops;
2184         long status = 0;
2185         ssize_t written = 0;
2186         unsigned int flags = 0;
2187
2188         /*
2189          * Copies from kernel address space cannot fail (NFSD is a big user).
2190          */
2191         if (segment_eq(get_fs(), KERNEL_DS))
2192                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2193
2194         do {
2195                 struct page *page;
2196                 pgoff_t index;          /* Pagecache index for current page */
2197                 unsigned long offset;   /* Offset into pagecache page */
2198                 unsigned long bytes;    /* Bytes to write to page */
2199                 size_t copied;          /* Bytes copied from user */
2200                 void *fsdata;
2201
2202                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2203                 index = pos >> PAGE_CACHE_SHIFT;
2204                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2205                                                 iov_iter_count(i));
2206
2207 again:
2208
2209                 /*
2210                  * Bring in the user page that we will copy from _first_.
2211                  * Otherwise there's a nasty deadlock on copying from the
2212                  * same page as we're writing to, without it being marked
2213                  * up-to-date.
2214                  *
2215                  * Not only is this an optimisation, but it is also required
2216                  * to check that the address is actually valid, when atomic
2217                  * usercopies are used, below.
2218                  */
2219                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2220                         status = -EFAULT;
2221                         break;
2222                 }
2223
2224                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2225                                                 &page, &fsdata);
2226                 if (unlikely(status))
2227                         break;
2228
2229                 pagefault_disable();
2230                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2231                 pagefault_enable();
2232                 flush_dcache_page(page);
2233
2234                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2235                                                 page, fsdata);
2236                 if (unlikely(status < 0))
2237                         break;
2238                 copied = status;
2239
2240                 cond_resched();
2241
2242                 if (unlikely(copied == 0)) {
2243                         /*
2244                          * If we were unable to copy any data at all, we must
2245                          * fall back to a single segment length write.
2246                          *
2247                          * If we didn't fallback here, we could livelock
2248                          * because not all segments in the iov can be copied at
2249                          * once without a pagefault.
2250                          */
2251                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2252                                                 iov_iter_single_seg_count(i));
2253                         goto again;
2254                 }
2255                 iov_iter_advance(i, copied);
2256                 pos += copied;
2257                 written += copied;
2258
2259                 balance_dirty_pages_ratelimited(mapping);
2260
2261         } while (iov_iter_count(i));
2262
2263         return written ? written : status;
2264 }
2265
2266 ssize_t
2267 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2268                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2269                 size_t count, ssize_t written)
2270 {
2271         struct file *file = iocb->ki_filp;
2272         struct address_space *mapping = file->f_mapping;
2273         const struct address_space_operations *a_ops = mapping->a_ops;
2274         struct inode *inode = mapping->host;
2275         ssize_t status;
2276         struct iov_iter i;
2277
2278         iov_iter_init(&i, iov, nr_segs, count, written);
2279         if (a_ops->write_begin)
2280                 status = generic_perform_write(file, &i, pos);
2281         else
2282                 status = generic_perform_write_2copy(file, &i, pos);
2283
2284         if (likely(status >= 0)) {
2285                 written += status;
2286                 *ppos = pos + status;
2287
2288                 /*
2289                  * For now, when the user asks for O_SYNC, we'll actually give
2290                  * O_DSYNC
2291                  */
2292                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2293                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2294                                 status = generic_osync_inode(inode, mapping,
2295                                                 OSYNC_METADATA|OSYNC_DATA);
2296                 }
2297         }
2298         
2299         /*
2300          * If we get here for O_DIRECT writes then we must have fallen through
2301          * to buffered writes (block instantiation inside i_size).  So we sync
2302          * the file data here, to try to honour O_DIRECT expectations.
2303          */
2304         if (unlikely(file->f_flags & O_DIRECT) && written)
2305                 status = filemap_write_and_wait(mapping);
2306
2307         return written ? written : status;
2308 }
2309 EXPORT_SYMBOL(generic_file_buffered_write);
2310
2311 static ssize_t
2312 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2313                                 unsigned long nr_segs, loff_t *ppos)
2314 {
2315         struct file *file = iocb->ki_filp;
2316         struct address_space * mapping = file->f_mapping;
2317         size_t ocount;          /* original count */
2318         size_t count;           /* after file limit checks */
2319         struct inode    *inode = mapping->host;
2320         loff_t          pos;
2321         ssize_t         written;
2322         ssize_t         err;
2323
2324         ocount = 0;
2325         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2326         if (err)
2327                 return err;
2328
2329         count = ocount;
2330         pos = *ppos;
2331
2332         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2333
2334         /* We can write back this queue in page reclaim */
2335         current->backing_dev_info = mapping->backing_dev_info;
2336         written = 0;
2337
2338         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2339         if (err)
2340                 goto out;
2341
2342         if (count == 0)
2343                 goto out;
2344
2345         err = remove_suid(file->f_path.dentry);
2346         if (err)
2347                 goto out;
2348
2349         file_update_time(file);
2350
2351         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2352         if (unlikely(file->f_flags & O_DIRECT)) {
2353                 loff_t endbyte;
2354                 ssize_t written_buffered;
2355
2356                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2357                                                         ppos, count, ocount);
2358                 if (written < 0 || written == count)
2359                         goto out;
2360                 /*
2361                  * direct-io write to a hole: fall through to buffered I/O
2362                  * for completing the rest of the request.
2363                  */
2364                 pos += written;
2365                 count -= written;
2366                 written_buffered = generic_file_buffered_write(iocb, iov,
2367                                                 nr_segs, pos, ppos, count,
2368                                                 written);
2369                 /*
2370                  * If generic_file_buffered_write() retuned a synchronous error
2371                  * then we want to return the number of bytes which were
2372                  * direct-written, or the error code if that was zero.  Note
2373                  * that this differs from normal direct-io semantics, which
2374                  * will return -EFOO even if some bytes were written.
2375                  */
2376                 if (written_buffered < 0) {
2377                         err = written_buffered;
2378                         goto out;
2379                 }
2380
2381                 /*
2382                  * We need to ensure that the page cache pages are written to
2383                  * disk and invalidated to preserve the expected O_DIRECT
2384                  * semantics.
2385                  */
2386                 endbyte = pos + written_buffered - written - 1;
2387                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2388                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2389                                             SYNC_FILE_RANGE_WRITE|
2390                                             SYNC_FILE_RANGE_WAIT_AFTER);
2391                 if (err == 0) {
2392                         written = written_buffered;
2393                         invalidate_mapping_pages(mapping,
2394                                                  pos >> PAGE_CACHE_SHIFT,
2395                                                  endbyte >> PAGE_CACHE_SHIFT);
2396                 } else {
2397                         /*
2398                          * We don't know how much we wrote, so just return
2399                          * the number of bytes which were direct-written
2400                          */
2401                 }
2402         } else {
2403                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2404                                 pos, ppos, count, written);
2405         }
2406 out:
2407         current->backing_dev_info = NULL;
2408         return written ? written : err;
2409 }
2410
2411 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2412                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2413 {
2414         struct file *file = iocb->ki_filp;
2415         struct address_space *mapping = file->f_mapping;
2416         struct inode *inode = mapping->host;
2417         ssize_t ret;
2418
2419         BUG_ON(iocb->ki_pos != pos);
2420
2421         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2422                         &iocb->ki_pos);
2423
2424         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2425                 ssize_t err;
2426
2427                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2428                 if (err < 0)
2429                         ret = err;
2430         }
2431         return ret;
2432 }
2433 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2434
2435 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2436                 unsigned long nr_segs, loff_t pos)
2437 {
2438         struct file *file = iocb->ki_filp;
2439         struct address_space *mapping = file->f_mapping;
2440         struct inode *inode = mapping->host;
2441         ssize_t ret;
2442
2443         BUG_ON(iocb->ki_pos != pos);
2444
2445         mutex_lock(&inode->i_mutex);
2446         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2447                         &iocb->ki_pos);
2448         mutex_unlock(&inode->i_mutex);
2449
2450         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2451                 ssize_t err;
2452
2453                 err = sync_page_range(inode, mapping, pos, ret);
2454                 if (err < 0)
2455                         ret = err;
2456         }
2457         return ret;
2458 }
2459 EXPORT_SYMBOL(generic_file_aio_write);
2460
2461 /*
2462  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2463  * went wrong during pagecache shootdown.
2464  */
2465 static ssize_t
2466 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2467         loff_t offset, unsigned long nr_segs)
2468 {
2469         struct file *file = iocb->ki_filp;
2470         struct address_space *mapping = file->f_mapping;
2471         ssize_t retval;
2472         size_t write_len;
2473         pgoff_t end = 0; /* silence gcc */
2474
2475         /*
2476          * If it's a write, unmap all mmappings of the file up-front.  This
2477          * will cause any pte dirty bits to be propagated into the pageframes
2478          * for the subsequent filemap_write_and_wait().
2479          */
2480         if (rw == WRITE) {
2481                 write_len = iov_length(iov, nr_segs);
2482                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2483                 if (mapping_mapped(mapping))
2484                         unmap_mapping_range(mapping, offset, write_len, 0);
2485         }
2486
2487         retval = filemap_write_and_wait(mapping);
2488         if (retval)
2489                 goto out;
2490
2491         /*
2492          * After a write we want buffered reads to be sure to go to disk to get
2493          * the new data.  We invalidate clean cached page from the region we're
2494          * about to write.  We do this *before* the write so that we can return
2495          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2496          */
2497         if (rw == WRITE && mapping->nrpages) {
2498                 retval = invalidate_inode_pages2_range(mapping,
2499                                         offset >> PAGE_CACHE_SHIFT, end);
2500                 if (retval)
2501                         goto out;
2502         }
2503
2504         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2505         if (retval)
2506                 goto out;
2507
2508         /*
2509          * Finally, try again to invalidate clean pages which might have been
2510          * faulted in by get_user_pages() if the source of the write was an
2511          * mmap()ed region of the file we're writing.  That's a pretty crazy
2512          * thing to do, so we don't support it 100%.  If this invalidation
2513          * fails and we have -EIOCBQUEUED we ignore the failure.
2514          */
2515         if (rw == WRITE && mapping->nrpages) {
2516                 int err = invalidate_inode_pages2_range(mapping,
2517                                               offset >> PAGE_CACHE_SHIFT, end);
2518                 if (err && retval >= 0)
2519                         retval = err;
2520         }
2521 out:
2522         return retval;
2523 }
2524
2525 /**
2526  * try_to_release_page() - release old fs-specific metadata on a page
2527  *
2528  * @page: the page which the kernel is trying to free
2529  * @gfp_mask: memory allocation flags (and I/O mode)
2530  *
2531  * The address_space is to try to release any data against the page
2532  * (presumably at page->private).  If the release was successful, return `1'.
2533  * Otherwise return zero.
2534  *
2535  * The @gfp_mask argument specifies whether I/O may be performed to release
2536  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2537  *
2538  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2539  */
2540 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2541 {
2542         struct address_space * const mapping = page->mapping;
2543
2544         BUG_ON(!PageLocked(page));
2545         if (PageWriteback(page))
2546                 return 0;
2547
2548         if (mapping && mapping->a_ops->releasepage)
2549                 return mapping->a_ops->releasepage(page, gfp_mask);
2550         return try_to_free_buffers(page);
2551 }
2552
2553 EXPORT_SYMBOL(try_to_release_page);