Add lock_page_killable
[linux-2.6] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/backing-dev.h>
32 #include <linux/security.h>
33 #include <linux/syscalls.h>
34 #include <linux/cpuset.h>
35 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
36 #include "internal.h"
37
38 /*
39  * FIXME: remove all knowledge of the buffer layer from the core VM
40  */
41 #include <linux/buffer_head.h> /* for generic_osync_inode */
42
43 #include <asm/mman.h>
44
45 static ssize_t
46 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
47         loff_t offset, unsigned long nr_segs);
48
49 /*
50  * Shared mappings implemented 30.11.1994. It's not fully working yet,
51  * though.
52  *
53  * Shared mappings now work. 15.8.1995  Bruno.
54  *
55  * finished 'unifying' the page and buffer cache and SMP-threaded the
56  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
57  *
58  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
59  */
60
61 /*
62  * Lock ordering:
63  *
64  *  ->i_mmap_lock               (vmtruncate)
65  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
66  *      ->swap_lock             (exclusive_swap_page, others)
67  *        ->mapping->tree_lock
68  *          ->zone.lock
69  *
70  *  ->i_mutex
71  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
72  *
73  *  ->mmap_sem
74  *    ->i_mmap_lock
75  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
76  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
77  *
78  *  ->mmap_sem
79  *    ->lock_page               (access_process_vm)
80  *
81  *  ->i_mutex                   (generic_file_buffered_write)
82  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
83  *
84  *  ->i_mutex
85  *    ->i_alloc_sem             (various)
86  *
87  *  ->inode_lock
88  *    ->sb_lock                 (fs/fs-writeback.c)
89  *    ->mapping->tree_lock      (__sync_single_inode)
90  *
91  *  ->i_mmap_lock
92  *    ->anon_vma.lock           (vma_adjust)
93  *
94  *  ->anon_vma.lock
95  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
96  *
97  *  ->page_table_lock or pte_lock
98  *    ->swap_lock               (try_to_unmap_one)
99  *    ->private_lock            (try_to_unmap_one)
100  *    ->tree_lock               (try_to_unmap_one)
101  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
102  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
103  *    ->private_lock            (page_remove_rmap->set_page_dirty)
104  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
106  *    ->inode_lock              (zap_pte_range->set_page_dirty)
107  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
108  *
109  *  ->task->proc_lock
110  *    ->dcache_lock             (proc_pid_lookup)
111  */
112
113 /*
114  * Remove a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         BUG_ON(page_mapped(page));
127 }
128
129 void remove_from_page_cache(struct page *page)
130 {
131         struct address_space *mapping = page->mapping;
132
133         BUG_ON(!PageLocked(page));
134
135         write_lock_irq(&mapping->tree_lock);
136         __remove_from_page_cache(page);
137         write_unlock_irq(&mapping->tree_lock);
138 }
139
140 static int sync_page(void *word)
141 {
142         struct address_space *mapping;
143         struct page *page;
144
145         page = container_of((unsigned long *)word, struct page, flags);
146
147         /*
148          * page_mapping() is being called without PG_locked held.
149          * Some knowledge of the state and use of the page is used to
150          * reduce the requirements down to a memory barrier.
151          * The danger here is of a stale page_mapping() return value
152          * indicating a struct address_space different from the one it's
153          * associated with when it is associated with one.
154          * After smp_mb(), it's either the correct page_mapping() for
155          * the page, or an old page_mapping() and the page's own
156          * page_mapping() has gone NULL.
157          * The ->sync_page() address_space operation must tolerate
158          * page_mapping() going NULL. By an amazing coincidence,
159          * this comes about because none of the users of the page
160          * in the ->sync_page() methods make essential use of the
161          * page_mapping(), merely passing the page down to the backing
162          * device's unplug functions when it's non-NULL, which in turn
163          * ignore it for all cases but swap, where only page_private(page) is
164          * of interest. When page_mapping() does go NULL, the entire
165          * call stack gracefully ignores the page and returns.
166          * -- wli
167          */
168         smp_mb();
169         mapping = page_mapping(page);
170         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
171                 mapping->a_ops->sync_page(page);
172         io_schedule();
173         return 0;
174 }
175
176 static int sync_page_killable(void *word)
177 {
178         sync_page(word);
179         return fatal_signal_pending(current) ? -EINTR : 0;
180 }
181
182 /**
183  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
184  * @mapping:    address space structure to write
185  * @start:      offset in bytes where the range starts
186  * @end:        offset in bytes where the range ends (inclusive)
187  * @sync_mode:  enable synchronous operation
188  *
189  * Start writeback against all of a mapping's dirty pages that lie
190  * within the byte offsets <start, end> inclusive.
191  *
192  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
193  * opposed to a regular memory cleansing writeback.  The difference between
194  * these two operations is that if a dirty page/buffer is encountered, it must
195  * be waited upon, and not just skipped over.
196  */
197 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
198                                 loff_t end, int sync_mode)
199 {
200         int ret;
201         struct writeback_control wbc = {
202                 .sync_mode = sync_mode,
203                 .nr_to_write = mapping->nrpages * 2,
204                 .range_start = start,
205                 .range_end = end,
206         };
207
208         if (!mapping_cap_writeback_dirty(mapping))
209                 return 0;
210
211         ret = do_writepages(mapping, &wbc);
212         return ret;
213 }
214
215 static inline int __filemap_fdatawrite(struct address_space *mapping,
216         int sync_mode)
217 {
218         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
219 }
220
221 int filemap_fdatawrite(struct address_space *mapping)
222 {
223         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
224 }
225 EXPORT_SYMBOL(filemap_fdatawrite);
226
227 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
228                                 loff_t end)
229 {
230         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
231 }
232
233 /**
234  * filemap_flush - mostly a non-blocking flush
235  * @mapping:    target address_space
236  *
237  * This is a mostly non-blocking flush.  Not suitable for data-integrity
238  * purposes - I/O may not be started against all dirty pages.
239  */
240 int filemap_flush(struct address_space *mapping)
241 {
242         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
243 }
244 EXPORT_SYMBOL(filemap_flush);
245
246 /**
247  * wait_on_page_writeback_range - wait for writeback to complete
248  * @mapping:    target address_space
249  * @start:      beginning page index
250  * @end:        ending page index
251  *
252  * Wait for writeback to complete against pages indexed by start->end
253  * inclusive
254  */
255 int wait_on_page_writeback_range(struct address_space *mapping,
256                                 pgoff_t start, pgoff_t end)
257 {
258         struct pagevec pvec;
259         int nr_pages;
260         int ret = 0;
261         pgoff_t index;
262
263         if (end < start)
264                 return 0;
265
266         pagevec_init(&pvec, 0);
267         index = start;
268         while ((index <= end) &&
269                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
270                         PAGECACHE_TAG_WRITEBACK,
271                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
272                 unsigned i;
273
274                 for (i = 0; i < nr_pages; i++) {
275                         struct page *page = pvec.pages[i];
276
277                         /* until radix tree lookup accepts end_index */
278                         if (page->index > end)
279                                 continue;
280
281                         wait_on_page_writeback(page);
282                         if (PageError(page))
283                                 ret = -EIO;
284                 }
285                 pagevec_release(&pvec);
286                 cond_resched();
287         }
288
289         /* Check for outstanding write errors */
290         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
291                 ret = -ENOSPC;
292         if (test_and_clear_bit(AS_EIO, &mapping->flags))
293                 ret = -EIO;
294
295         return ret;
296 }
297
298 /**
299  * sync_page_range - write and wait on all pages in the passed range
300  * @inode:      target inode
301  * @mapping:    target address_space
302  * @pos:        beginning offset in pages to write
303  * @count:      number of bytes to write
304  *
305  * Write and wait upon all the pages in the passed range.  This is a "data
306  * integrity" operation.  It waits upon in-flight writeout before starting and
307  * waiting upon new writeout.  If there was an IO error, return it.
308  *
309  * We need to re-take i_mutex during the generic_osync_inode list walk because
310  * it is otherwise livelockable.
311  */
312 int sync_page_range(struct inode *inode, struct address_space *mapping,
313                         loff_t pos, loff_t count)
314 {
315         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
316         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
317         int ret;
318
319         if (!mapping_cap_writeback_dirty(mapping) || !count)
320                 return 0;
321         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
322         if (ret == 0) {
323                 mutex_lock(&inode->i_mutex);
324                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
325                 mutex_unlock(&inode->i_mutex);
326         }
327         if (ret == 0)
328                 ret = wait_on_page_writeback_range(mapping, start, end);
329         return ret;
330 }
331 EXPORT_SYMBOL(sync_page_range);
332
333 /**
334  * sync_page_range_nolock
335  * @inode:      target inode
336  * @mapping:    target address_space
337  * @pos:        beginning offset in pages to write
338  * @count:      number of bytes to write
339  *
340  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
341  * as it forces O_SYNC writers to different parts of the same file
342  * to be serialised right until io completion.
343  */
344 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
345                            loff_t pos, loff_t count)
346 {
347         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
348         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
349         int ret;
350
351         if (!mapping_cap_writeback_dirty(mapping) || !count)
352                 return 0;
353         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
354         if (ret == 0)
355                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
356         if (ret == 0)
357                 ret = wait_on_page_writeback_range(mapping, start, end);
358         return ret;
359 }
360 EXPORT_SYMBOL(sync_page_range_nolock);
361
362 /**
363  * filemap_fdatawait - wait for all under-writeback pages to complete
364  * @mapping: address space structure to wait for
365  *
366  * Walk the list of under-writeback pages of the given address space
367  * and wait for all of them.
368  */
369 int filemap_fdatawait(struct address_space *mapping)
370 {
371         loff_t i_size = i_size_read(mapping->host);
372
373         if (i_size == 0)
374                 return 0;
375
376         return wait_on_page_writeback_range(mapping, 0,
377                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
378 }
379 EXPORT_SYMBOL(filemap_fdatawait);
380
381 int filemap_write_and_wait(struct address_space *mapping)
382 {
383         int err = 0;
384
385         if (mapping->nrpages) {
386                 err = filemap_fdatawrite(mapping);
387                 /*
388                  * Even if the above returned error, the pages may be
389                  * written partially (e.g. -ENOSPC), so we wait for it.
390                  * But the -EIO is special case, it may indicate the worst
391                  * thing (e.g. bug) happened, so we avoid waiting for it.
392                  */
393                 if (err != -EIO) {
394                         int err2 = filemap_fdatawait(mapping);
395                         if (!err)
396                                 err = err2;
397                 }
398         }
399         return err;
400 }
401 EXPORT_SYMBOL(filemap_write_and_wait);
402
403 /**
404  * filemap_write_and_wait_range - write out & wait on a file range
405  * @mapping:    the address_space for the pages
406  * @lstart:     offset in bytes where the range starts
407  * @lend:       offset in bytes where the range ends (inclusive)
408  *
409  * Write out and wait upon file offsets lstart->lend, inclusive.
410  *
411  * Note that `lend' is inclusive (describes the last byte to be written) so
412  * that this function can be used to write to the very end-of-file (end = -1).
413  */
414 int filemap_write_and_wait_range(struct address_space *mapping,
415                                  loff_t lstart, loff_t lend)
416 {
417         int err = 0;
418
419         if (mapping->nrpages) {
420                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
421                                                  WB_SYNC_ALL);
422                 /* See comment of filemap_write_and_wait() */
423                 if (err != -EIO) {
424                         int err2 = wait_on_page_writeback_range(mapping,
425                                                 lstart >> PAGE_CACHE_SHIFT,
426                                                 lend >> PAGE_CACHE_SHIFT);
427                         if (!err)
428                                 err = err2;
429                 }
430         }
431         return err;
432 }
433
434 /**
435  * add_to_page_cache - add newly allocated pagecache pages
436  * @page:       page to add
437  * @mapping:    the page's address_space
438  * @offset:     page index
439  * @gfp_mask:   page allocation mode
440  *
441  * This function is used to add newly allocated pagecache pages;
442  * the page is new, so we can just run SetPageLocked() against it.
443  * The other page state flags were set by rmqueue().
444  *
445  * This function does not add the page to the LRU.  The caller must do that.
446  */
447 int add_to_page_cache(struct page *page, struct address_space *mapping,
448                 pgoff_t offset, gfp_t gfp_mask)
449 {
450         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
451
452         if (error == 0) {
453                 write_lock_irq(&mapping->tree_lock);
454                 error = radix_tree_insert(&mapping->page_tree, offset, page);
455                 if (!error) {
456                         page_cache_get(page);
457                         SetPageLocked(page);
458                         page->mapping = mapping;
459                         page->index = offset;
460                         mapping->nrpages++;
461                         __inc_zone_page_state(page, NR_FILE_PAGES);
462                 }
463                 write_unlock_irq(&mapping->tree_lock);
464                 radix_tree_preload_end();
465         }
466         return error;
467 }
468 EXPORT_SYMBOL(add_to_page_cache);
469
470 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
471                                 pgoff_t offset, gfp_t gfp_mask)
472 {
473         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
474         if (ret == 0)
475                 lru_cache_add(page);
476         return ret;
477 }
478
479 #ifdef CONFIG_NUMA
480 struct page *__page_cache_alloc(gfp_t gfp)
481 {
482         if (cpuset_do_page_mem_spread()) {
483                 int n = cpuset_mem_spread_node();
484                 return alloc_pages_node(n, gfp, 0);
485         }
486         return alloc_pages(gfp, 0);
487 }
488 EXPORT_SYMBOL(__page_cache_alloc);
489 #endif
490
491 static int __sleep_on_page_lock(void *word)
492 {
493         io_schedule();
494         return 0;
495 }
496
497 /*
498  * In order to wait for pages to become available there must be
499  * waitqueues associated with pages. By using a hash table of
500  * waitqueues where the bucket discipline is to maintain all
501  * waiters on the same queue and wake all when any of the pages
502  * become available, and for the woken contexts to check to be
503  * sure the appropriate page became available, this saves space
504  * at a cost of "thundering herd" phenomena during rare hash
505  * collisions.
506  */
507 static wait_queue_head_t *page_waitqueue(struct page *page)
508 {
509         const struct zone *zone = page_zone(page);
510
511         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
512 }
513
514 static inline void wake_up_page(struct page *page, int bit)
515 {
516         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
517 }
518
519 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
520 {
521         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
522
523         if (test_bit(bit_nr, &page->flags))
524                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
525                                                         TASK_UNINTERRUPTIBLE);
526 }
527 EXPORT_SYMBOL(wait_on_page_bit);
528
529 /**
530  * unlock_page - unlock a locked page
531  * @page: the page
532  *
533  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
534  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
535  * mechananism between PageLocked pages and PageWriteback pages is shared.
536  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
537  *
538  * The first mb is necessary to safely close the critical section opened by the
539  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
540  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
541  * parallel wait_on_page_locked()).
542  */
543 void fastcall unlock_page(struct page *page)
544 {
545         smp_mb__before_clear_bit();
546         if (!TestClearPageLocked(page))
547                 BUG();
548         smp_mb__after_clear_bit(); 
549         wake_up_page(page, PG_locked);
550 }
551 EXPORT_SYMBOL(unlock_page);
552
553 /**
554  * end_page_writeback - end writeback against a page
555  * @page: the page
556  */
557 void end_page_writeback(struct page *page)
558 {
559         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
560                 if (!test_clear_page_writeback(page))
561                         BUG();
562         }
563         smp_mb__after_clear_bit();
564         wake_up_page(page, PG_writeback);
565 }
566 EXPORT_SYMBOL(end_page_writeback);
567
568 /**
569  * __lock_page - get a lock on the page, assuming we need to sleep to get it
570  * @page: the page to lock
571  *
572  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
573  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
574  * chances are that on the second loop, the block layer's plug list is empty,
575  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
576  */
577 void fastcall __lock_page(struct page *page)
578 {
579         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
580
581         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
582                                                         TASK_UNINTERRUPTIBLE);
583 }
584 EXPORT_SYMBOL(__lock_page);
585
586 int fastcall __lock_page_killable(struct page *page)
587 {
588         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
589
590         return __wait_on_bit_lock(page_waitqueue(page), &wait,
591                                         sync_page_killable, TASK_KILLABLE);
592 }
593
594 /*
595  * Variant of lock_page that does not require the caller to hold a reference
596  * on the page's mapping.
597  */
598 void fastcall __lock_page_nosync(struct page *page)
599 {
600         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
601         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
602                                                         TASK_UNINTERRUPTIBLE);
603 }
604
605 /**
606  * find_get_page - find and get a page reference
607  * @mapping: the address_space to search
608  * @offset: the page index
609  *
610  * Is there a pagecache struct page at the given (mapping, offset) tuple?
611  * If yes, increment its refcount and return it; if no, return NULL.
612  */
613 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
614 {
615         struct page *page;
616
617         read_lock_irq(&mapping->tree_lock);
618         page = radix_tree_lookup(&mapping->page_tree, offset);
619         if (page)
620                 page_cache_get(page);
621         read_unlock_irq(&mapping->tree_lock);
622         return page;
623 }
624 EXPORT_SYMBOL(find_get_page);
625
626 /**
627  * find_lock_page - locate, pin and lock a pagecache page
628  * @mapping: the address_space to search
629  * @offset: the page index
630  *
631  * Locates the desired pagecache page, locks it, increments its reference
632  * count and returns its address.
633  *
634  * Returns zero if the page was not present. find_lock_page() may sleep.
635  */
636 struct page *find_lock_page(struct address_space *mapping,
637                                 pgoff_t offset)
638 {
639         struct page *page;
640
641 repeat:
642         read_lock_irq(&mapping->tree_lock);
643         page = radix_tree_lookup(&mapping->page_tree, offset);
644         if (page) {
645                 page_cache_get(page);
646                 if (TestSetPageLocked(page)) {
647                         read_unlock_irq(&mapping->tree_lock);
648                         __lock_page(page);
649
650                         /* Has the page been truncated while we slept? */
651                         if (unlikely(page->mapping != mapping)) {
652                                 unlock_page(page);
653                                 page_cache_release(page);
654                                 goto repeat;
655                         }
656                         VM_BUG_ON(page->index != offset);
657                         goto out;
658                 }
659         }
660         read_unlock_irq(&mapping->tree_lock);
661 out:
662         return page;
663 }
664 EXPORT_SYMBOL(find_lock_page);
665
666 /**
667  * find_or_create_page - locate or add a pagecache page
668  * @mapping: the page's address_space
669  * @index: the page's index into the mapping
670  * @gfp_mask: page allocation mode
671  *
672  * Locates a page in the pagecache.  If the page is not present, a new page
673  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
674  * LRU list.  The returned page is locked and has its reference count
675  * incremented.
676  *
677  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
678  * allocation!
679  *
680  * find_or_create_page() returns the desired page's address, or zero on
681  * memory exhaustion.
682  */
683 struct page *find_or_create_page(struct address_space *mapping,
684                 pgoff_t index, gfp_t gfp_mask)
685 {
686         struct page *page;
687         int err;
688 repeat:
689         page = find_lock_page(mapping, index);
690         if (!page) {
691                 page = __page_cache_alloc(gfp_mask);
692                 if (!page)
693                         return NULL;
694                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
695                 if (unlikely(err)) {
696                         page_cache_release(page);
697                         page = NULL;
698                         if (err == -EEXIST)
699                                 goto repeat;
700                 }
701         }
702         return page;
703 }
704 EXPORT_SYMBOL(find_or_create_page);
705
706 /**
707  * find_get_pages - gang pagecache lookup
708  * @mapping:    The address_space to search
709  * @start:      The starting page index
710  * @nr_pages:   The maximum number of pages
711  * @pages:      Where the resulting pages are placed
712  *
713  * find_get_pages() will search for and return a group of up to
714  * @nr_pages pages in the mapping.  The pages are placed at @pages.
715  * find_get_pages() takes a reference against the returned pages.
716  *
717  * The search returns a group of mapping-contiguous pages with ascending
718  * indexes.  There may be holes in the indices due to not-present pages.
719  *
720  * find_get_pages() returns the number of pages which were found.
721  */
722 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
723                             unsigned int nr_pages, struct page **pages)
724 {
725         unsigned int i;
726         unsigned int ret;
727
728         read_lock_irq(&mapping->tree_lock);
729         ret = radix_tree_gang_lookup(&mapping->page_tree,
730                                 (void **)pages, start, nr_pages);
731         for (i = 0; i < ret; i++)
732                 page_cache_get(pages[i]);
733         read_unlock_irq(&mapping->tree_lock);
734         return ret;
735 }
736
737 /**
738  * find_get_pages_contig - gang contiguous pagecache lookup
739  * @mapping:    The address_space to search
740  * @index:      The starting page index
741  * @nr_pages:   The maximum number of pages
742  * @pages:      Where the resulting pages are placed
743  *
744  * find_get_pages_contig() works exactly like find_get_pages(), except
745  * that the returned number of pages are guaranteed to be contiguous.
746  *
747  * find_get_pages_contig() returns the number of pages which were found.
748  */
749 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
750                                unsigned int nr_pages, struct page **pages)
751 {
752         unsigned int i;
753         unsigned int ret;
754
755         read_lock_irq(&mapping->tree_lock);
756         ret = radix_tree_gang_lookup(&mapping->page_tree,
757                                 (void **)pages, index, nr_pages);
758         for (i = 0; i < ret; i++) {
759                 if (pages[i]->mapping == NULL || pages[i]->index != index)
760                         break;
761
762                 page_cache_get(pages[i]);
763                 index++;
764         }
765         read_unlock_irq(&mapping->tree_lock);
766         return i;
767 }
768 EXPORT_SYMBOL(find_get_pages_contig);
769
770 /**
771  * find_get_pages_tag - find and return pages that match @tag
772  * @mapping:    the address_space to search
773  * @index:      the starting page index
774  * @tag:        the tag index
775  * @nr_pages:   the maximum number of pages
776  * @pages:      where the resulting pages are placed
777  *
778  * Like find_get_pages, except we only return pages which are tagged with
779  * @tag.   We update @index to index the next page for the traversal.
780  */
781 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
782                         int tag, unsigned int nr_pages, struct page **pages)
783 {
784         unsigned int i;
785         unsigned int ret;
786
787         read_lock_irq(&mapping->tree_lock);
788         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
789                                 (void **)pages, *index, nr_pages, tag);
790         for (i = 0; i < ret; i++)
791                 page_cache_get(pages[i]);
792         if (ret)
793                 *index = pages[ret - 1]->index + 1;
794         read_unlock_irq(&mapping->tree_lock);
795         return ret;
796 }
797 EXPORT_SYMBOL(find_get_pages_tag);
798
799 /**
800  * grab_cache_page_nowait - returns locked page at given index in given cache
801  * @mapping: target address_space
802  * @index: the page index
803  *
804  * Same as grab_cache_page(), but do not wait if the page is unavailable.
805  * This is intended for speculative data generators, where the data can
806  * be regenerated if the page couldn't be grabbed.  This routine should
807  * be safe to call while holding the lock for another page.
808  *
809  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
810  * and deadlock against the caller's locked page.
811  */
812 struct page *
813 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
814 {
815         struct page *page = find_get_page(mapping, index);
816
817         if (page) {
818                 if (!TestSetPageLocked(page))
819                         return page;
820                 page_cache_release(page);
821                 return NULL;
822         }
823         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
824         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
825                 page_cache_release(page);
826                 page = NULL;
827         }
828         return page;
829 }
830 EXPORT_SYMBOL(grab_cache_page_nowait);
831
832 /*
833  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
834  * a _large_ part of the i/o request. Imagine the worst scenario:
835  *
836  *      ---R__________________________________________B__________
837  *         ^ reading here                             ^ bad block(assume 4k)
838  *
839  * read(R) => miss => readahead(R...B) => media error => frustrating retries
840  * => failing the whole request => read(R) => read(R+1) =>
841  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
842  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
843  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
844  *
845  * It is going insane. Fix it by quickly scaling down the readahead size.
846  */
847 static void shrink_readahead_size_eio(struct file *filp,
848                                         struct file_ra_state *ra)
849 {
850         if (!ra->ra_pages)
851                 return;
852
853         ra->ra_pages /= 4;
854 }
855
856 /**
857  * do_generic_mapping_read - generic file read routine
858  * @mapping:    address_space to be read
859  * @ra:         file's readahead state
860  * @filp:       the file to read
861  * @ppos:       current file position
862  * @desc:       read_descriptor
863  * @actor:      read method
864  *
865  * This is a generic file read routine, and uses the
866  * mapping->a_ops->readpage() function for the actual low-level stuff.
867  *
868  * This is really ugly. But the goto's actually try to clarify some
869  * of the logic when it comes to error handling etc.
870  *
871  * Note the struct file* is only passed for the use of readpage.
872  * It may be NULL.
873  */
874 void do_generic_mapping_read(struct address_space *mapping,
875                              struct file_ra_state *ra,
876                              struct file *filp,
877                              loff_t *ppos,
878                              read_descriptor_t *desc,
879                              read_actor_t actor)
880 {
881         struct inode *inode = mapping->host;
882         pgoff_t index;
883         pgoff_t last_index;
884         pgoff_t prev_index;
885         unsigned long offset;      /* offset into pagecache page */
886         unsigned int prev_offset;
887         int error;
888
889         index = *ppos >> PAGE_CACHE_SHIFT;
890         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
891         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
892         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
893         offset = *ppos & ~PAGE_CACHE_MASK;
894
895         for (;;) {
896                 struct page *page;
897                 pgoff_t end_index;
898                 loff_t isize;
899                 unsigned long nr, ret;
900
901                 cond_resched();
902 find_page:
903                 page = find_get_page(mapping, index);
904                 if (!page) {
905                         page_cache_sync_readahead(mapping,
906                                         ra, filp,
907                                         index, last_index - index);
908                         page = find_get_page(mapping, index);
909                         if (unlikely(page == NULL))
910                                 goto no_cached_page;
911                 }
912                 if (PageReadahead(page)) {
913                         page_cache_async_readahead(mapping,
914                                         ra, filp, page,
915                                         index, last_index - index);
916                 }
917                 if (!PageUptodate(page))
918                         goto page_not_up_to_date;
919 page_ok:
920                 /*
921                  * i_size must be checked after we know the page is Uptodate.
922                  *
923                  * Checking i_size after the check allows us to calculate
924                  * the correct value for "nr", which means the zero-filled
925                  * part of the page is not copied back to userspace (unless
926                  * another truncate extends the file - this is desired though).
927                  */
928
929                 isize = i_size_read(inode);
930                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
931                 if (unlikely(!isize || index > end_index)) {
932                         page_cache_release(page);
933                         goto out;
934                 }
935
936                 /* nr is the maximum number of bytes to copy from this page */
937                 nr = PAGE_CACHE_SIZE;
938                 if (index == end_index) {
939                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
940                         if (nr <= offset) {
941                                 page_cache_release(page);
942                                 goto out;
943                         }
944                 }
945                 nr = nr - offset;
946
947                 /* If users can be writing to this page using arbitrary
948                  * virtual addresses, take care about potential aliasing
949                  * before reading the page on the kernel side.
950                  */
951                 if (mapping_writably_mapped(mapping))
952                         flush_dcache_page(page);
953
954                 /*
955                  * When a sequential read accesses a page several times,
956                  * only mark it as accessed the first time.
957                  */
958                 if (prev_index != index || offset != prev_offset)
959                         mark_page_accessed(page);
960                 prev_index = index;
961
962                 /*
963                  * Ok, we have the page, and it's up-to-date, so
964                  * now we can copy it to user space...
965                  *
966                  * The actor routine returns how many bytes were actually used..
967                  * NOTE! This may not be the same as how much of a user buffer
968                  * we filled up (we may be padding etc), so we can only update
969                  * "pos" here (the actor routine has to update the user buffer
970                  * pointers and the remaining count).
971                  */
972                 ret = actor(desc, page, offset, nr);
973                 offset += ret;
974                 index += offset >> PAGE_CACHE_SHIFT;
975                 offset &= ~PAGE_CACHE_MASK;
976                 prev_offset = offset;
977
978                 page_cache_release(page);
979                 if (ret == nr && desc->count)
980                         continue;
981                 goto out;
982
983 page_not_up_to_date:
984                 /* Get exclusive access to the page ... */
985                 lock_page(page);
986
987                 /* Did it get truncated before we got the lock? */
988                 if (!page->mapping) {
989                         unlock_page(page);
990                         page_cache_release(page);
991                         continue;
992                 }
993
994                 /* Did somebody else fill it already? */
995                 if (PageUptodate(page)) {
996                         unlock_page(page);
997                         goto page_ok;
998                 }
999
1000 readpage:
1001                 /* Start the actual read. The read will unlock the page. */
1002                 error = mapping->a_ops->readpage(filp, page);
1003
1004                 if (unlikely(error)) {
1005                         if (error == AOP_TRUNCATED_PAGE) {
1006                                 page_cache_release(page);
1007                                 goto find_page;
1008                         }
1009                         goto readpage_error;
1010                 }
1011
1012                 if (!PageUptodate(page)) {
1013                         lock_page(page);
1014                         if (!PageUptodate(page)) {
1015                                 if (page->mapping == NULL) {
1016                                         /*
1017                                          * invalidate_inode_pages got it
1018                                          */
1019                                         unlock_page(page);
1020                                         page_cache_release(page);
1021                                         goto find_page;
1022                                 }
1023                                 unlock_page(page);
1024                                 error = -EIO;
1025                                 shrink_readahead_size_eio(filp, ra);
1026                                 goto readpage_error;
1027                         }
1028                         unlock_page(page);
1029                 }
1030
1031                 goto page_ok;
1032
1033 readpage_error:
1034                 /* UHHUH! A synchronous read error occurred. Report it */
1035                 desc->error = error;
1036                 page_cache_release(page);
1037                 goto out;
1038
1039 no_cached_page:
1040                 /*
1041                  * Ok, it wasn't cached, so we need to create a new
1042                  * page..
1043                  */
1044                 page = page_cache_alloc_cold(mapping);
1045                 if (!page) {
1046                         desc->error = -ENOMEM;
1047                         goto out;
1048                 }
1049                 error = add_to_page_cache_lru(page, mapping,
1050                                                 index, GFP_KERNEL);
1051                 if (error) {
1052                         page_cache_release(page);
1053                         if (error == -EEXIST)
1054                                 goto find_page;
1055                         desc->error = error;
1056                         goto out;
1057                 }
1058                 goto readpage;
1059         }
1060
1061 out:
1062         ra->prev_pos = prev_index;
1063         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1064         ra->prev_pos |= prev_offset;
1065
1066         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1067         if (filp)
1068                 file_accessed(filp);
1069 }
1070 EXPORT_SYMBOL(do_generic_mapping_read);
1071
1072 int file_read_actor(read_descriptor_t *desc, struct page *page,
1073                         unsigned long offset, unsigned long size)
1074 {
1075         char *kaddr;
1076         unsigned long left, count = desc->count;
1077
1078         if (size > count)
1079                 size = count;
1080
1081         /*
1082          * Faults on the destination of a read are common, so do it before
1083          * taking the kmap.
1084          */
1085         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1086                 kaddr = kmap_atomic(page, KM_USER0);
1087                 left = __copy_to_user_inatomic(desc->arg.buf,
1088                                                 kaddr + offset, size);
1089                 kunmap_atomic(kaddr, KM_USER0);
1090                 if (left == 0)
1091                         goto success;
1092         }
1093
1094         /* Do it the slow way */
1095         kaddr = kmap(page);
1096         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1097         kunmap(page);
1098
1099         if (left) {
1100                 size -= left;
1101                 desc->error = -EFAULT;
1102         }
1103 success:
1104         desc->count = count - size;
1105         desc->written += size;
1106         desc->arg.buf += size;
1107         return size;
1108 }
1109
1110 /*
1111  * Performs necessary checks before doing a write
1112  * @iov:        io vector request
1113  * @nr_segs:    number of segments in the iovec
1114  * @count:      number of bytes to write
1115  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1116  *
1117  * Adjust number of segments and amount of bytes to write (nr_segs should be
1118  * properly initialized first). Returns appropriate error code that caller
1119  * should return or zero in case that write should be allowed.
1120  */
1121 int generic_segment_checks(const struct iovec *iov,
1122                         unsigned long *nr_segs, size_t *count, int access_flags)
1123 {
1124         unsigned long   seg;
1125         size_t cnt = 0;
1126         for (seg = 0; seg < *nr_segs; seg++) {
1127                 const struct iovec *iv = &iov[seg];
1128
1129                 /*
1130                  * If any segment has a negative length, or the cumulative
1131                  * length ever wraps negative then return -EINVAL.
1132                  */
1133                 cnt += iv->iov_len;
1134                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1135                         return -EINVAL;
1136                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1137                         continue;
1138                 if (seg == 0)
1139                         return -EFAULT;
1140                 *nr_segs = seg;
1141                 cnt -= iv->iov_len;     /* This segment is no good */
1142                 break;
1143         }
1144         *count = cnt;
1145         return 0;
1146 }
1147 EXPORT_SYMBOL(generic_segment_checks);
1148
1149 /**
1150  * generic_file_aio_read - generic filesystem read routine
1151  * @iocb:       kernel I/O control block
1152  * @iov:        io vector request
1153  * @nr_segs:    number of segments in the iovec
1154  * @pos:        current file position
1155  *
1156  * This is the "read()" routine for all filesystems
1157  * that can use the page cache directly.
1158  */
1159 ssize_t
1160 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1161                 unsigned long nr_segs, loff_t pos)
1162 {
1163         struct file *filp = iocb->ki_filp;
1164         ssize_t retval;
1165         unsigned long seg;
1166         size_t count;
1167         loff_t *ppos = &iocb->ki_pos;
1168
1169         count = 0;
1170         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1171         if (retval)
1172                 return retval;
1173
1174         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1175         if (filp->f_flags & O_DIRECT) {
1176                 loff_t size;
1177                 struct address_space *mapping;
1178                 struct inode *inode;
1179
1180                 mapping = filp->f_mapping;
1181                 inode = mapping->host;
1182                 retval = 0;
1183                 if (!count)
1184                         goto out; /* skip atime */
1185                 size = i_size_read(inode);
1186                 if (pos < size) {
1187                         retval = generic_file_direct_IO(READ, iocb,
1188                                                 iov, pos, nr_segs);
1189                         if (retval > 0)
1190                                 *ppos = pos + retval;
1191                 }
1192                 if (likely(retval != 0)) {
1193                         file_accessed(filp);
1194                         goto out;
1195                 }
1196         }
1197
1198         retval = 0;
1199         if (count) {
1200                 for (seg = 0; seg < nr_segs; seg++) {
1201                         read_descriptor_t desc;
1202
1203                         desc.written = 0;
1204                         desc.arg.buf = iov[seg].iov_base;
1205                         desc.count = iov[seg].iov_len;
1206                         if (desc.count == 0)
1207                                 continue;
1208                         desc.error = 0;
1209                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1210                         retval += desc.written;
1211                         if (desc.error) {
1212                                 retval = retval ?: desc.error;
1213                                 break;
1214                         }
1215                         if (desc.count > 0)
1216                                 break;
1217                 }
1218         }
1219 out:
1220         return retval;
1221 }
1222 EXPORT_SYMBOL(generic_file_aio_read);
1223
1224 static ssize_t
1225 do_readahead(struct address_space *mapping, struct file *filp,
1226              pgoff_t index, unsigned long nr)
1227 {
1228         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1229                 return -EINVAL;
1230
1231         force_page_cache_readahead(mapping, filp, index,
1232                                         max_sane_readahead(nr));
1233         return 0;
1234 }
1235
1236 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1237 {
1238         ssize_t ret;
1239         struct file *file;
1240
1241         ret = -EBADF;
1242         file = fget(fd);
1243         if (file) {
1244                 if (file->f_mode & FMODE_READ) {
1245                         struct address_space *mapping = file->f_mapping;
1246                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1247                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1248                         unsigned long len = end - start + 1;
1249                         ret = do_readahead(mapping, file, start, len);
1250                 }
1251                 fput(file);
1252         }
1253         return ret;
1254 }
1255
1256 #ifdef CONFIG_MMU
1257 /**
1258  * page_cache_read - adds requested page to the page cache if not already there
1259  * @file:       file to read
1260  * @offset:     page index
1261  *
1262  * This adds the requested page to the page cache if it isn't already there,
1263  * and schedules an I/O to read in its contents from disk.
1264  */
1265 static int fastcall page_cache_read(struct file * file, pgoff_t offset)
1266 {
1267         struct address_space *mapping = file->f_mapping;
1268         struct page *page; 
1269         int ret;
1270
1271         do {
1272                 page = page_cache_alloc_cold(mapping);
1273                 if (!page)
1274                         return -ENOMEM;
1275
1276                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1277                 if (ret == 0)
1278                         ret = mapping->a_ops->readpage(file, page);
1279                 else if (ret == -EEXIST)
1280                         ret = 0; /* losing race to add is OK */
1281
1282                 page_cache_release(page);
1283
1284         } while (ret == AOP_TRUNCATED_PAGE);
1285                 
1286         return ret;
1287 }
1288
1289 #define MMAP_LOTSAMISS  (100)
1290
1291 /**
1292  * filemap_fault - read in file data for page fault handling
1293  * @vma:        vma in which the fault was taken
1294  * @vmf:        struct vm_fault containing details of the fault
1295  *
1296  * filemap_fault() is invoked via the vma operations vector for a
1297  * mapped memory region to read in file data during a page fault.
1298  *
1299  * The goto's are kind of ugly, but this streamlines the normal case of having
1300  * it in the page cache, and handles the special cases reasonably without
1301  * having a lot of duplicated code.
1302  */
1303 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1304 {
1305         int error;
1306         struct file *file = vma->vm_file;
1307         struct address_space *mapping = file->f_mapping;
1308         struct file_ra_state *ra = &file->f_ra;
1309         struct inode *inode = mapping->host;
1310         struct page *page;
1311         unsigned long size;
1312         int did_readaround = 0;
1313         int ret = 0;
1314
1315         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1316         if (vmf->pgoff >= size)
1317                 return VM_FAULT_SIGBUS;
1318
1319         /* If we don't want any read-ahead, don't bother */
1320         if (VM_RandomReadHint(vma))
1321                 goto no_cached_page;
1322
1323         /*
1324          * Do we have something in the page cache already?
1325          */
1326 retry_find:
1327         page = find_lock_page(mapping, vmf->pgoff);
1328         /*
1329          * For sequential accesses, we use the generic readahead logic.
1330          */
1331         if (VM_SequentialReadHint(vma)) {
1332                 if (!page) {
1333                         page_cache_sync_readahead(mapping, ra, file,
1334                                                            vmf->pgoff, 1);
1335                         page = find_lock_page(mapping, vmf->pgoff);
1336                         if (!page)
1337                                 goto no_cached_page;
1338                 }
1339                 if (PageReadahead(page)) {
1340                         page_cache_async_readahead(mapping, ra, file, page,
1341                                                            vmf->pgoff, 1);
1342                 }
1343         }
1344
1345         if (!page) {
1346                 unsigned long ra_pages;
1347
1348                 ra->mmap_miss++;
1349
1350                 /*
1351                  * Do we miss much more than hit in this file? If so,
1352                  * stop bothering with read-ahead. It will only hurt.
1353                  */
1354                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1355                         goto no_cached_page;
1356
1357                 /*
1358                  * To keep the pgmajfault counter straight, we need to
1359                  * check did_readaround, as this is an inner loop.
1360                  */
1361                 if (!did_readaround) {
1362                         ret = VM_FAULT_MAJOR;
1363                         count_vm_event(PGMAJFAULT);
1364                 }
1365                 did_readaround = 1;
1366                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1367                 if (ra_pages) {
1368                         pgoff_t start = 0;
1369
1370                         if (vmf->pgoff > ra_pages / 2)
1371                                 start = vmf->pgoff - ra_pages / 2;
1372                         do_page_cache_readahead(mapping, file, start, ra_pages);
1373                 }
1374                 page = find_lock_page(mapping, vmf->pgoff);
1375                 if (!page)
1376                         goto no_cached_page;
1377         }
1378
1379         if (!did_readaround)
1380                 ra->mmap_miss--;
1381
1382         /*
1383          * We have a locked page in the page cache, now we need to check
1384          * that it's up-to-date. If not, it is going to be due to an error.
1385          */
1386         if (unlikely(!PageUptodate(page)))
1387                 goto page_not_uptodate;
1388
1389         /* Must recheck i_size under page lock */
1390         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1391         if (unlikely(vmf->pgoff >= size)) {
1392                 unlock_page(page);
1393                 page_cache_release(page);
1394                 return VM_FAULT_SIGBUS;
1395         }
1396
1397         /*
1398          * Found the page and have a reference on it.
1399          */
1400         mark_page_accessed(page);
1401         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1402         vmf->page = page;
1403         return ret | VM_FAULT_LOCKED;
1404
1405 no_cached_page:
1406         /*
1407          * We're only likely to ever get here if MADV_RANDOM is in
1408          * effect.
1409          */
1410         error = page_cache_read(file, vmf->pgoff);
1411
1412         /*
1413          * The page we want has now been added to the page cache.
1414          * In the unlikely event that someone removed it in the
1415          * meantime, we'll just come back here and read it again.
1416          */
1417         if (error >= 0)
1418                 goto retry_find;
1419
1420         /*
1421          * An error return from page_cache_read can result if the
1422          * system is low on memory, or a problem occurs while trying
1423          * to schedule I/O.
1424          */
1425         if (error == -ENOMEM)
1426                 return VM_FAULT_OOM;
1427         return VM_FAULT_SIGBUS;
1428
1429 page_not_uptodate:
1430         /* IO error path */
1431         if (!did_readaround) {
1432                 ret = VM_FAULT_MAJOR;
1433                 count_vm_event(PGMAJFAULT);
1434         }
1435
1436         /*
1437          * Umm, take care of errors if the page isn't up-to-date.
1438          * Try to re-read it _once_. We do this synchronously,
1439          * because there really aren't any performance issues here
1440          * and we need to check for errors.
1441          */
1442         ClearPageError(page);
1443         error = mapping->a_ops->readpage(file, page);
1444         page_cache_release(page);
1445
1446         if (!error || error == AOP_TRUNCATED_PAGE)
1447                 goto retry_find;
1448
1449         /* Things didn't work out. Return zero to tell the mm layer so. */
1450         shrink_readahead_size_eio(file, ra);
1451         return VM_FAULT_SIGBUS;
1452 }
1453 EXPORT_SYMBOL(filemap_fault);
1454
1455 struct vm_operations_struct generic_file_vm_ops = {
1456         .fault          = filemap_fault,
1457 };
1458
1459 /* This is used for a general mmap of a disk file */
1460
1461 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1462 {
1463         struct address_space *mapping = file->f_mapping;
1464
1465         if (!mapping->a_ops->readpage)
1466                 return -ENOEXEC;
1467         file_accessed(file);
1468         vma->vm_ops = &generic_file_vm_ops;
1469         vma->vm_flags |= VM_CAN_NONLINEAR;
1470         return 0;
1471 }
1472
1473 /*
1474  * This is for filesystems which do not implement ->writepage.
1475  */
1476 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1477 {
1478         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1479                 return -EINVAL;
1480         return generic_file_mmap(file, vma);
1481 }
1482 #else
1483 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1484 {
1485         return -ENOSYS;
1486 }
1487 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1488 {
1489         return -ENOSYS;
1490 }
1491 #endif /* CONFIG_MMU */
1492
1493 EXPORT_SYMBOL(generic_file_mmap);
1494 EXPORT_SYMBOL(generic_file_readonly_mmap);
1495
1496 static struct page *__read_cache_page(struct address_space *mapping,
1497                                 pgoff_t index,
1498                                 int (*filler)(void *,struct page*),
1499                                 void *data)
1500 {
1501         struct page *page;
1502         int err;
1503 repeat:
1504         page = find_get_page(mapping, index);
1505         if (!page) {
1506                 page = page_cache_alloc_cold(mapping);
1507                 if (!page)
1508                         return ERR_PTR(-ENOMEM);
1509                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1510                 if (unlikely(err)) {
1511                         page_cache_release(page);
1512                         if (err == -EEXIST)
1513                                 goto repeat;
1514                         /* Presumably ENOMEM for radix tree node */
1515                         return ERR_PTR(err);
1516                 }
1517                 err = filler(data, page);
1518                 if (err < 0) {
1519                         page_cache_release(page);
1520                         page = ERR_PTR(err);
1521                 }
1522         }
1523         return page;
1524 }
1525
1526 /*
1527  * Same as read_cache_page, but don't wait for page to become unlocked
1528  * after submitting it to the filler.
1529  */
1530 struct page *read_cache_page_async(struct address_space *mapping,
1531                                 pgoff_t index,
1532                                 int (*filler)(void *,struct page*),
1533                                 void *data)
1534 {
1535         struct page *page;
1536         int err;
1537
1538 retry:
1539         page = __read_cache_page(mapping, index, filler, data);
1540         if (IS_ERR(page))
1541                 return page;
1542         if (PageUptodate(page))
1543                 goto out;
1544
1545         lock_page(page);
1546         if (!page->mapping) {
1547                 unlock_page(page);
1548                 page_cache_release(page);
1549                 goto retry;
1550         }
1551         if (PageUptodate(page)) {
1552                 unlock_page(page);
1553                 goto out;
1554         }
1555         err = filler(data, page);
1556         if (err < 0) {
1557                 page_cache_release(page);
1558                 return ERR_PTR(err);
1559         }
1560 out:
1561         mark_page_accessed(page);
1562         return page;
1563 }
1564 EXPORT_SYMBOL(read_cache_page_async);
1565
1566 /**
1567  * read_cache_page - read into page cache, fill it if needed
1568  * @mapping:    the page's address_space
1569  * @index:      the page index
1570  * @filler:     function to perform the read
1571  * @data:       destination for read data
1572  *
1573  * Read into the page cache. If a page already exists, and PageUptodate() is
1574  * not set, try to fill the page then wait for it to become unlocked.
1575  *
1576  * If the page does not get brought uptodate, return -EIO.
1577  */
1578 struct page *read_cache_page(struct address_space *mapping,
1579                                 pgoff_t index,
1580                                 int (*filler)(void *,struct page*),
1581                                 void *data)
1582 {
1583         struct page *page;
1584
1585         page = read_cache_page_async(mapping, index, filler, data);
1586         if (IS_ERR(page))
1587                 goto out;
1588         wait_on_page_locked(page);
1589         if (!PageUptodate(page)) {
1590                 page_cache_release(page);
1591                 page = ERR_PTR(-EIO);
1592         }
1593  out:
1594         return page;
1595 }
1596 EXPORT_SYMBOL(read_cache_page);
1597
1598 /*
1599  * The logic we want is
1600  *
1601  *      if suid or (sgid and xgrp)
1602  *              remove privs
1603  */
1604 int should_remove_suid(struct dentry *dentry)
1605 {
1606         mode_t mode = dentry->d_inode->i_mode;
1607         int kill = 0;
1608
1609         /* suid always must be killed */
1610         if (unlikely(mode & S_ISUID))
1611                 kill = ATTR_KILL_SUID;
1612
1613         /*
1614          * sgid without any exec bits is just a mandatory locking mark; leave
1615          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1616          */
1617         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1618                 kill |= ATTR_KILL_SGID;
1619
1620         if (unlikely(kill && !capable(CAP_FSETID)))
1621                 return kill;
1622
1623         return 0;
1624 }
1625 EXPORT_SYMBOL(should_remove_suid);
1626
1627 int __remove_suid(struct dentry *dentry, int kill)
1628 {
1629         struct iattr newattrs;
1630
1631         newattrs.ia_valid = ATTR_FORCE | kill;
1632         return notify_change(dentry, &newattrs);
1633 }
1634
1635 int remove_suid(struct dentry *dentry)
1636 {
1637         int killsuid = should_remove_suid(dentry);
1638         int killpriv = security_inode_need_killpriv(dentry);
1639         int error = 0;
1640
1641         if (killpriv < 0)
1642                 return killpriv;
1643         if (killpriv)
1644                 error = security_inode_killpriv(dentry);
1645         if (!error && killsuid)
1646                 error = __remove_suid(dentry, killsuid);
1647
1648         return error;
1649 }
1650 EXPORT_SYMBOL(remove_suid);
1651
1652 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1653                         const struct iovec *iov, size_t base, size_t bytes)
1654 {
1655         size_t copied = 0, left = 0;
1656
1657         while (bytes) {
1658                 char __user *buf = iov->iov_base + base;
1659                 int copy = min(bytes, iov->iov_len - base);
1660
1661                 base = 0;
1662                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1663                 copied += copy;
1664                 bytes -= copy;
1665                 vaddr += copy;
1666                 iov++;
1667
1668                 if (unlikely(left))
1669                         break;
1670         }
1671         return copied - left;
1672 }
1673
1674 /*
1675  * Copy as much as we can into the page and return the number of bytes which
1676  * were sucessfully copied.  If a fault is encountered then return the number of
1677  * bytes which were copied.
1678  */
1679 size_t iov_iter_copy_from_user_atomic(struct page *page,
1680                 struct iov_iter *i, unsigned long offset, size_t bytes)
1681 {
1682         char *kaddr;
1683         size_t copied;
1684
1685         BUG_ON(!in_atomic());
1686         kaddr = kmap_atomic(page, KM_USER0);
1687         if (likely(i->nr_segs == 1)) {
1688                 int left;
1689                 char __user *buf = i->iov->iov_base + i->iov_offset;
1690                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1691                                                         buf, bytes);
1692                 copied = bytes - left;
1693         } else {
1694                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1695                                                 i->iov, i->iov_offset, bytes);
1696         }
1697         kunmap_atomic(kaddr, KM_USER0);
1698
1699         return copied;
1700 }
1701 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1702
1703 /*
1704  * This has the same sideeffects and return value as
1705  * iov_iter_copy_from_user_atomic().
1706  * The difference is that it attempts to resolve faults.
1707  * Page must not be locked.
1708  */
1709 size_t iov_iter_copy_from_user(struct page *page,
1710                 struct iov_iter *i, unsigned long offset, size_t bytes)
1711 {
1712         char *kaddr;
1713         size_t copied;
1714
1715         kaddr = kmap(page);
1716         if (likely(i->nr_segs == 1)) {
1717                 int left;
1718                 char __user *buf = i->iov->iov_base + i->iov_offset;
1719                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1720                 copied = bytes - left;
1721         } else {
1722                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1723                                                 i->iov, i->iov_offset, bytes);
1724         }
1725         kunmap(page);
1726         return copied;
1727 }
1728 EXPORT_SYMBOL(iov_iter_copy_from_user);
1729
1730 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1731 {
1732         if (likely(i->nr_segs == 1)) {
1733                 i->iov_offset += bytes;
1734         } else {
1735                 const struct iovec *iov = i->iov;
1736                 size_t base = i->iov_offset;
1737
1738                 while (bytes) {
1739                         int copy = min(bytes, iov->iov_len - base);
1740
1741                         bytes -= copy;
1742                         base += copy;
1743                         if (iov->iov_len == base) {
1744                                 iov++;
1745                                 base = 0;
1746                         }
1747                 }
1748                 i->iov = iov;
1749                 i->iov_offset = base;
1750         }
1751 }
1752
1753 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1754 {
1755         BUG_ON(i->count < bytes);
1756
1757         __iov_iter_advance_iov(i, bytes);
1758         i->count -= bytes;
1759 }
1760 EXPORT_SYMBOL(iov_iter_advance);
1761
1762 /*
1763  * Fault in the first iovec of the given iov_iter, to a maximum length
1764  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1765  * accessed (ie. because it is an invalid address).
1766  *
1767  * writev-intensive code may want this to prefault several iovecs -- that
1768  * would be possible (callers must not rely on the fact that _only_ the
1769  * first iovec will be faulted with the current implementation).
1770  */
1771 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1772 {
1773         char __user *buf = i->iov->iov_base + i->iov_offset;
1774         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1775         return fault_in_pages_readable(buf, bytes);
1776 }
1777 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1778
1779 /*
1780  * Return the count of just the current iov_iter segment.
1781  */
1782 size_t iov_iter_single_seg_count(struct iov_iter *i)
1783 {
1784         const struct iovec *iov = i->iov;
1785         if (i->nr_segs == 1)
1786                 return i->count;
1787         else
1788                 return min(i->count, iov->iov_len - i->iov_offset);
1789 }
1790 EXPORT_SYMBOL(iov_iter_single_seg_count);
1791
1792 /*
1793  * Performs necessary checks before doing a write
1794  *
1795  * Can adjust writing position or amount of bytes to write.
1796  * Returns appropriate error code that caller should return or
1797  * zero in case that write should be allowed.
1798  */
1799 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1800 {
1801         struct inode *inode = file->f_mapping->host;
1802         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1803
1804         if (unlikely(*pos < 0))
1805                 return -EINVAL;
1806
1807         if (!isblk) {
1808                 /* FIXME: this is for backwards compatibility with 2.4 */
1809                 if (file->f_flags & O_APPEND)
1810                         *pos = i_size_read(inode);
1811
1812                 if (limit != RLIM_INFINITY) {
1813                         if (*pos >= limit) {
1814                                 send_sig(SIGXFSZ, current, 0);
1815                                 return -EFBIG;
1816                         }
1817                         if (*count > limit - (typeof(limit))*pos) {
1818                                 *count = limit - (typeof(limit))*pos;
1819                         }
1820                 }
1821         }
1822
1823         /*
1824          * LFS rule
1825          */
1826         if (unlikely(*pos + *count > MAX_NON_LFS &&
1827                                 !(file->f_flags & O_LARGEFILE))) {
1828                 if (*pos >= MAX_NON_LFS) {
1829                         return -EFBIG;
1830                 }
1831                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1832                         *count = MAX_NON_LFS - (unsigned long)*pos;
1833                 }
1834         }
1835
1836         /*
1837          * Are we about to exceed the fs block limit ?
1838          *
1839          * If we have written data it becomes a short write.  If we have
1840          * exceeded without writing data we send a signal and return EFBIG.
1841          * Linus frestrict idea will clean these up nicely..
1842          */
1843         if (likely(!isblk)) {
1844                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1845                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1846                                 return -EFBIG;
1847                         }
1848                         /* zero-length writes at ->s_maxbytes are OK */
1849                 }
1850
1851                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1852                         *count = inode->i_sb->s_maxbytes - *pos;
1853         } else {
1854 #ifdef CONFIG_BLOCK
1855                 loff_t isize;
1856                 if (bdev_read_only(I_BDEV(inode)))
1857                         return -EPERM;
1858                 isize = i_size_read(inode);
1859                 if (*pos >= isize) {
1860                         if (*count || *pos > isize)
1861                                 return -ENOSPC;
1862                 }
1863
1864                 if (*pos + *count > isize)
1865                         *count = isize - *pos;
1866 #else
1867                 return -EPERM;
1868 #endif
1869         }
1870         return 0;
1871 }
1872 EXPORT_SYMBOL(generic_write_checks);
1873
1874 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1875                                 loff_t pos, unsigned len, unsigned flags,
1876                                 struct page **pagep, void **fsdata)
1877 {
1878         const struct address_space_operations *aops = mapping->a_ops;
1879
1880         if (aops->write_begin) {
1881                 return aops->write_begin(file, mapping, pos, len, flags,
1882                                                         pagep, fsdata);
1883         } else {
1884                 int ret;
1885                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1886                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1887                 struct inode *inode = mapping->host;
1888                 struct page *page;
1889 again:
1890                 page = __grab_cache_page(mapping, index);
1891                 *pagep = page;
1892                 if (!page)
1893                         return -ENOMEM;
1894
1895                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1896                         /*
1897                          * There is no way to resolve a short write situation
1898                          * for a !Uptodate page (except by double copying in
1899                          * the caller done by generic_perform_write_2copy).
1900                          *
1901                          * Instead, we have to bring it uptodate here.
1902                          */
1903                         ret = aops->readpage(file, page);
1904                         page_cache_release(page);
1905                         if (ret) {
1906                                 if (ret == AOP_TRUNCATED_PAGE)
1907                                         goto again;
1908                                 return ret;
1909                         }
1910                         goto again;
1911                 }
1912
1913                 ret = aops->prepare_write(file, page, offset, offset+len);
1914                 if (ret) {
1915                         unlock_page(page);
1916                         page_cache_release(page);
1917                         if (pos + len > inode->i_size)
1918                                 vmtruncate(inode, inode->i_size);
1919                 }
1920                 return ret;
1921         }
1922 }
1923 EXPORT_SYMBOL(pagecache_write_begin);
1924
1925 int pagecache_write_end(struct file *file, struct address_space *mapping,
1926                                 loff_t pos, unsigned len, unsigned copied,
1927                                 struct page *page, void *fsdata)
1928 {
1929         const struct address_space_operations *aops = mapping->a_ops;
1930         int ret;
1931
1932         if (aops->write_end) {
1933                 mark_page_accessed(page);
1934                 ret = aops->write_end(file, mapping, pos, len, copied,
1935                                                         page, fsdata);
1936         } else {
1937                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1938                 struct inode *inode = mapping->host;
1939
1940                 flush_dcache_page(page);
1941                 ret = aops->commit_write(file, page, offset, offset+len);
1942                 unlock_page(page);
1943                 mark_page_accessed(page);
1944                 page_cache_release(page);
1945
1946                 if (ret < 0) {
1947                         if (pos + len > inode->i_size)
1948                                 vmtruncate(inode, inode->i_size);
1949                 } else if (ret > 0)
1950                         ret = min_t(size_t, copied, ret);
1951                 else
1952                         ret = copied;
1953         }
1954
1955         return ret;
1956 }
1957 EXPORT_SYMBOL(pagecache_write_end);
1958
1959 ssize_t
1960 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1961                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1962                 size_t count, size_t ocount)
1963 {
1964         struct file     *file = iocb->ki_filp;
1965         struct address_space *mapping = file->f_mapping;
1966         struct inode    *inode = mapping->host;
1967         ssize_t         written;
1968
1969         if (count != ocount)
1970                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1971
1972         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1973         if (written > 0) {
1974                 loff_t end = pos + written;
1975                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1976                         i_size_write(inode,  end);
1977                         mark_inode_dirty(inode);
1978                 }
1979                 *ppos = end;
1980         }
1981
1982         /*
1983          * Sync the fs metadata but not the minor inode changes and
1984          * of course not the data as we did direct DMA for the IO.
1985          * i_mutex is held, which protects generic_osync_inode() from
1986          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
1987          */
1988         if ((written >= 0 || written == -EIOCBQUEUED) &&
1989             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1990                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1991                 if (err < 0)
1992                         written = err;
1993         }
1994         return written;
1995 }
1996 EXPORT_SYMBOL(generic_file_direct_write);
1997
1998 /*
1999  * Find or create a page at the given pagecache position. Return the locked
2000  * page. This function is specifically for buffered writes.
2001  */
2002 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2003 {
2004         int status;
2005         struct page *page;
2006 repeat:
2007         page = find_lock_page(mapping, index);
2008         if (likely(page))
2009                 return page;
2010
2011         page = page_cache_alloc(mapping);
2012         if (!page)
2013                 return NULL;
2014         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2015         if (unlikely(status)) {
2016                 page_cache_release(page);
2017                 if (status == -EEXIST)
2018                         goto repeat;
2019                 return NULL;
2020         }
2021         return page;
2022 }
2023 EXPORT_SYMBOL(__grab_cache_page);
2024
2025 static ssize_t generic_perform_write_2copy(struct file *file,
2026                                 struct iov_iter *i, loff_t pos)
2027 {
2028         struct address_space *mapping = file->f_mapping;
2029         const struct address_space_operations *a_ops = mapping->a_ops;
2030         struct inode *inode = mapping->host;
2031         long status = 0;
2032         ssize_t written = 0;
2033
2034         do {
2035                 struct page *src_page;
2036                 struct page *page;
2037                 pgoff_t index;          /* Pagecache index for current page */
2038                 unsigned long offset;   /* Offset into pagecache page */
2039                 unsigned long bytes;    /* Bytes to write to page */
2040                 size_t copied;          /* Bytes copied from user */
2041
2042                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2043                 index = pos >> PAGE_CACHE_SHIFT;
2044                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2045                                                 iov_iter_count(i));
2046
2047                 /*
2048                  * a non-NULL src_page indicates that we're doing the
2049                  * copy via get_user_pages and kmap.
2050                  */
2051                 src_page = NULL;
2052
2053                 /*
2054                  * Bring in the user page that we will copy from _first_.
2055                  * Otherwise there's a nasty deadlock on copying from the
2056                  * same page as we're writing to, without it being marked
2057                  * up-to-date.
2058                  *
2059                  * Not only is this an optimisation, but it is also required
2060                  * to check that the address is actually valid, when atomic
2061                  * usercopies are used, below.
2062                  */
2063                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2064                         status = -EFAULT;
2065                         break;
2066                 }
2067
2068                 page = __grab_cache_page(mapping, index);
2069                 if (!page) {
2070                         status = -ENOMEM;
2071                         break;
2072                 }
2073
2074                 /*
2075                  * non-uptodate pages cannot cope with short copies, and we
2076                  * cannot take a pagefault with the destination page locked.
2077                  * So pin the source page to copy it.
2078                  */
2079                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2080                         unlock_page(page);
2081
2082                         src_page = alloc_page(GFP_KERNEL);
2083                         if (!src_page) {
2084                                 page_cache_release(page);
2085                                 status = -ENOMEM;
2086                                 break;
2087                         }
2088
2089                         /*
2090                          * Cannot get_user_pages with a page locked for the
2091                          * same reason as we can't take a page fault with a
2092                          * page locked (as explained below).
2093                          */
2094                         copied = iov_iter_copy_from_user(src_page, i,
2095                                                                 offset, bytes);
2096                         if (unlikely(copied == 0)) {
2097                                 status = -EFAULT;
2098                                 page_cache_release(page);
2099                                 page_cache_release(src_page);
2100                                 break;
2101                         }
2102                         bytes = copied;
2103
2104                         lock_page(page);
2105                         /*
2106                          * Can't handle the page going uptodate here, because
2107                          * that means we would use non-atomic usercopies, which
2108                          * zero out the tail of the page, which can cause
2109                          * zeroes to become transiently visible. We could just
2110                          * use a non-zeroing copy, but the APIs aren't too
2111                          * consistent.
2112                          */
2113                         if (unlikely(!page->mapping || PageUptodate(page))) {
2114                                 unlock_page(page);
2115                                 page_cache_release(page);
2116                                 page_cache_release(src_page);
2117                                 continue;
2118                         }
2119                 }
2120
2121                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2122                 if (unlikely(status))
2123                         goto fs_write_aop_error;
2124
2125                 if (!src_page) {
2126                         /*
2127                          * Must not enter the pagefault handler here, because
2128                          * we hold the page lock, so we might recursively
2129                          * deadlock on the same lock, or get an ABBA deadlock
2130                          * against a different lock, or against the mmap_sem
2131                          * (which nests outside the page lock).  So increment
2132                          * preempt count, and use _atomic usercopies.
2133                          *
2134                          * The page is uptodate so we are OK to encounter a
2135                          * short copy: if unmodified parts of the page are
2136                          * marked dirty and written out to disk, it doesn't
2137                          * really matter.
2138                          */
2139                         pagefault_disable();
2140                         copied = iov_iter_copy_from_user_atomic(page, i,
2141                                                                 offset, bytes);
2142                         pagefault_enable();
2143                 } else {
2144                         void *src, *dst;
2145                         src = kmap_atomic(src_page, KM_USER0);
2146                         dst = kmap_atomic(page, KM_USER1);
2147                         memcpy(dst + offset, src + offset, bytes);
2148                         kunmap_atomic(dst, KM_USER1);
2149                         kunmap_atomic(src, KM_USER0);
2150                         copied = bytes;
2151                 }
2152                 flush_dcache_page(page);
2153
2154                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2155                 if (unlikely(status < 0))
2156                         goto fs_write_aop_error;
2157                 if (unlikely(status > 0)) /* filesystem did partial write */
2158                         copied = min_t(size_t, copied, status);
2159
2160                 unlock_page(page);
2161                 mark_page_accessed(page);
2162                 page_cache_release(page);
2163                 if (src_page)
2164                         page_cache_release(src_page);
2165
2166                 iov_iter_advance(i, copied);
2167                 pos += copied;
2168                 written += copied;
2169
2170                 balance_dirty_pages_ratelimited(mapping);
2171                 cond_resched();
2172                 continue;
2173
2174 fs_write_aop_error:
2175                 unlock_page(page);
2176                 page_cache_release(page);
2177                 if (src_page)
2178                         page_cache_release(src_page);
2179
2180                 /*
2181                  * prepare_write() may have instantiated a few blocks
2182                  * outside i_size.  Trim these off again. Don't need
2183                  * i_size_read because we hold i_mutex.
2184                  */
2185                 if (pos + bytes > inode->i_size)
2186                         vmtruncate(inode, inode->i_size);
2187                 break;
2188         } while (iov_iter_count(i));
2189
2190         return written ? written : status;
2191 }
2192
2193 static ssize_t generic_perform_write(struct file *file,
2194                                 struct iov_iter *i, loff_t pos)
2195 {
2196         struct address_space *mapping = file->f_mapping;
2197         const struct address_space_operations *a_ops = mapping->a_ops;
2198         long status = 0;
2199         ssize_t written = 0;
2200         unsigned int flags = 0;
2201
2202         /*
2203          * Copies from kernel address space cannot fail (NFSD is a big user).
2204          */
2205         if (segment_eq(get_fs(), KERNEL_DS))
2206                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2207
2208         do {
2209                 struct page *page;
2210                 pgoff_t index;          /* Pagecache index for current page */
2211                 unsigned long offset;   /* Offset into pagecache page */
2212                 unsigned long bytes;    /* Bytes to write to page */
2213                 size_t copied;          /* Bytes copied from user */
2214                 void *fsdata;
2215
2216                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2217                 index = pos >> PAGE_CACHE_SHIFT;
2218                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2219                                                 iov_iter_count(i));
2220
2221 again:
2222
2223                 /*
2224                  * Bring in the user page that we will copy from _first_.
2225                  * Otherwise there's a nasty deadlock on copying from the
2226                  * same page as we're writing to, without it being marked
2227                  * up-to-date.
2228                  *
2229                  * Not only is this an optimisation, but it is also required
2230                  * to check that the address is actually valid, when atomic
2231                  * usercopies are used, below.
2232                  */
2233                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2234                         status = -EFAULT;
2235                         break;
2236                 }
2237
2238                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2239                                                 &page, &fsdata);
2240                 if (unlikely(status))
2241                         break;
2242
2243                 pagefault_disable();
2244                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2245                 pagefault_enable();
2246                 flush_dcache_page(page);
2247
2248                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2249                                                 page, fsdata);
2250                 if (unlikely(status < 0))
2251                         break;
2252                 copied = status;
2253
2254                 cond_resched();
2255
2256                 if (unlikely(copied == 0)) {
2257                         /*
2258                          * If we were unable to copy any data at all, we must
2259                          * fall back to a single segment length write.
2260                          *
2261                          * If we didn't fallback here, we could livelock
2262                          * because not all segments in the iov can be copied at
2263                          * once without a pagefault.
2264                          */
2265                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2266                                                 iov_iter_single_seg_count(i));
2267                         goto again;
2268                 }
2269                 iov_iter_advance(i, copied);
2270                 pos += copied;
2271                 written += copied;
2272
2273                 balance_dirty_pages_ratelimited(mapping);
2274
2275         } while (iov_iter_count(i));
2276
2277         return written ? written : status;
2278 }
2279
2280 ssize_t
2281 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2282                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2283                 size_t count, ssize_t written)
2284 {
2285         struct file *file = iocb->ki_filp;
2286         struct address_space *mapping = file->f_mapping;
2287         const struct address_space_operations *a_ops = mapping->a_ops;
2288         struct inode *inode = mapping->host;
2289         ssize_t status;
2290         struct iov_iter i;
2291
2292         iov_iter_init(&i, iov, nr_segs, count, written);
2293         if (a_ops->write_begin)
2294                 status = generic_perform_write(file, &i, pos);
2295         else
2296                 status = generic_perform_write_2copy(file, &i, pos);
2297
2298         if (likely(status >= 0)) {
2299                 written += status;
2300                 *ppos = pos + status;
2301
2302                 /*
2303                  * For now, when the user asks for O_SYNC, we'll actually give
2304                  * O_DSYNC
2305                  */
2306                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2307                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2308                                 status = generic_osync_inode(inode, mapping,
2309                                                 OSYNC_METADATA|OSYNC_DATA);
2310                 }
2311         }
2312         
2313         /*
2314          * If we get here for O_DIRECT writes then we must have fallen through
2315          * to buffered writes (block instantiation inside i_size).  So we sync
2316          * the file data here, to try to honour O_DIRECT expectations.
2317          */
2318         if (unlikely(file->f_flags & O_DIRECT) && written)
2319                 status = filemap_write_and_wait(mapping);
2320
2321         return written ? written : status;
2322 }
2323 EXPORT_SYMBOL(generic_file_buffered_write);
2324
2325 static ssize_t
2326 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2327                                 unsigned long nr_segs, loff_t *ppos)
2328 {
2329         struct file *file = iocb->ki_filp;
2330         struct address_space * mapping = file->f_mapping;
2331         size_t ocount;          /* original count */
2332         size_t count;           /* after file limit checks */
2333         struct inode    *inode = mapping->host;
2334         loff_t          pos;
2335         ssize_t         written;
2336         ssize_t         err;
2337
2338         ocount = 0;
2339         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2340         if (err)
2341                 return err;
2342
2343         count = ocount;
2344         pos = *ppos;
2345
2346         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2347
2348         /* We can write back this queue in page reclaim */
2349         current->backing_dev_info = mapping->backing_dev_info;
2350         written = 0;
2351
2352         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2353         if (err)
2354                 goto out;
2355
2356         if (count == 0)
2357                 goto out;
2358
2359         err = remove_suid(file->f_path.dentry);
2360         if (err)
2361                 goto out;
2362
2363         file_update_time(file);
2364
2365         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2366         if (unlikely(file->f_flags & O_DIRECT)) {
2367                 loff_t endbyte;
2368                 ssize_t written_buffered;
2369
2370                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2371                                                         ppos, count, ocount);
2372                 if (written < 0 || written == count)
2373                         goto out;
2374                 /*
2375                  * direct-io write to a hole: fall through to buffered I/O
2376                  * for completing the rest of the request.
2377                  */
2378                 pos += written;
2379                 count -= written;
2380                 written_buffered = generic_file_buffered_write(iocb, iov,
2381                                                 nr_segs, pos, ppos, count,
2382                                                 written);
2383                 /*
2384                  * If generic_file_buffered_write() retuned a synchronous error
2385                  * then we want to return the number of bytes which were
2386                  * direct-written, or the error code if that was zero.  Note
2387                  * that this differs from normal direct-io semantics, which
2388                  * will return -EFOO even if some bytes were written.
2389                  */
2390                 if (written_buffered < 0) {
2391                         err = written_buffered;
2392                         goto out;
2393                 }
2394
2395                 /*
2396                  * We need to ensure that the page cache pages are written to
2397                  * disk and invalidated to preserve the expected O_DIRECT
2398                  * semantics.
2399                  */
2400                 endbyte = pos + written_buffered - written - 1;
2401                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2402                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2403                                             SYNC_FILE_RANGE_WRITE|
2404                                             SYNC_FILE_RANGE_WAIT_AFTER);
2405                 if (err == 0) {
2406                         written = written_buffered;
2407                         invalidate_mapping_pages(mapping,
2408                                                  pos >> PAGE_CACHE_SHIFT,
2409                                                  endbyte >> PAGE_CACHE_SHIFT);
2410                 } else {
2411                         /*
2412                          * We don't know how much we wrote, so just return
2413                          * the number of bytes which were direct-written
2414                          */
2415                 }
2416         } else {
2417                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2418                                 pos, ppos, count, written);
2419         }
2420 out:
2421         current->backing_dev_info = NULL;
2422         return written ? written : err;
2423 }
2424
2425 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2426                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2427 {
2428         struct file *file = iocb->ki_filp;
2429         struct address_space *mapping = file->f_mapping;
2430         struct inode *inode = mapping->host;
2431         ssize_t ret;
2432
2433         BUG_ON(iocb->ki_pos != pos);
2434
2435         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2436                         &iocb->ki_pos);
2437
2438         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2439                 ssize_t err;
2440
2441                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2442                 if (err < 0)
2443                         ret = err;
2444         }
2445         return ret;
2446 }
2447 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2448
2449 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2450                 unsigned long nr_segs, loff_t pos)
2451 {
2452         struct file *file = iocb->ki_filp;
2453         struct address_space *mapping = file->f_mapping;
2454         struct inode *inode = mapping->host;
2455         ssize_t ret;
2456
2457         BUG_ON(iocb->ki_pos != pos);
2458
2459         mutex_lock(&inode->i_mutex);
2460         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2461                         &iocb->ki_pos);
2462         mutex_unlock(&inode->i_mutex);
2463
2464         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2465                 ssize_t err;
2466
2467                 err = sync_page_range(inode, mapping, pos, ret);
2468                 if (err < 0)
2469                         ret = err;
2470         }
2471         return ret;
2472 }
2473 EXPORT_SYMBOL(generic_file_aio_write);
2474
2475 /*
2476  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2477  * went wrong during pagecache shootdown.
2478  */
2479 static ssize_t
2480 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2481         loff_t offset, unsigned long nr_segs)
2482 {
2483         struct file *file = iocb->ki_filp;
2484         struct address_space *mapping = file->f_mapping;
2485         ssize_t retval;
2486         size_t write_len;
2487         pgoff_t end = 0; /* silence gcc */
2488
2489         /*
2490          * If it's a write, unmap all mmappings of the file up-front.  This
2491          * will cause any pte dirty bits to be propagated into the pageframes
2492          * for the subsequent filemap_write_and_wait().
2493          */
2494         if (rw == WRITE) {
2495                 write_len = iov_length(iov, nr_segs);
2496                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2497                 if (mapping_mapped(mapping))
2498                         unmap_mapping_range(mapping, offset, write_len, 0);
2499         }
2500
2501         retval = filemap_write_and_wait(mapping);
2502         if (retval)
2503                 goto out;
2504
2505         /*
2506          * After a write we want buffered reads to be sure to go to disk to get
2507          * the new data.  We invalidate clean cached page from the region we're
2508          * about to write.  We do this *before* the write so that we can return
2509          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2510          */
2511         if (rw == WRITE && mapping->nrpages) {
2512                 retval = invalidate_inode_pages2_range(mapping,
2513                                         offset >> PAGE_CACHE_SHIFT, end);
2514                 if (retval)
2515                         goto out;
2516         }
2517
2518         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2519
2520         /*
2521          * Finally, try again to invalidate clean pages which might have been
2522          * cached by non-direct readahead, or faulted in by get_user_pages()
2523          * if the source of the write was an mmap'ed region of the file
2524          * we're writing.  Either one is a pretty crazy thing to do,
2525          * so we don't support it 100%.  If this invalidation
2526          * fails, tough, the write still worked...
2527          */
2528         if (rw == WRITE && mapping->nrpages) {
2529                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2530         }
2531 out:
2532         return retval;
2533 }
2534
2535 /**
2536  * try_to_release_page() - release old fs-specific metadata on a page
2537  *
2538  * @page: the page which the kernel is trying to free
2539  * @gfp_mask: memory allocation flags (and I/O mode)
2540  *
2541  * The address_space is to try to release any data against the page
2542  * (presumably at page->private).  If the release was successful, return `1'.
2543  * Otherwise return zero.
2544  *
2545  * The @gfp_mask argument specifies whether I/O may be performed to release
2546  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2547  *
2548  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2549  */
2550 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2551 {
2552         struct address_space * const mapping = page->mapping;
2553
2554         BUG_ON(!PageLocked(page));
2555         if (PageWriteback(page))
2556                 return 0;
2557
2558         if (mapping && mapping->a_ops->releasepage)
2559                 return mapping->a_ops->releasepage(page, gfp_mask);
2560         return try_to_free_buffers(page);
2561 }
2562
2563 EXPORT_SYMBOL(try_to_release_page);