81fb9bff0d4f9022db02f9d3194ebc5ccd0c350e
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/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  *
69  *  ->i_mutex
70  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
71  *
72  *  ->mmap_sem
73  *    ->i_mmap_lock
74  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
75  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
76  *
77  *  ->mmap_sem
78  *    ->lock_page               (access_process_vm)
79  *
80  *  ->i_mutex                   (generic_file_buffered_write)
81  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
82  *
83  *  ->i_mutex
84  *    ->i_alloc_sem             (various)
85  *
86  *  ->inode_lock
87  *    ->sb_lock                 (fs/fs-writeback.c)
88  *    ->mapping->tree_lock      (__sync_single_inode)
89  *
90  *  ->i_mmap_lock
91  *    ->anon_vma.lock           (vma_adjust)
92  *
93  *  ->anon_vma.lock
94  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
95  *
96  *  ->page_table_lock or pte_lock
97  *    ->swap_lock               (try_to_unmap_one)
98  *    ->private_lock            (try_to_unmap_one)
99  *    ->tree_lock               (try_to_unmap_one)
100  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
101  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
102  *    ->private_lock            (page_remove_rmap->set_page_dirty)
103  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
104  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (zap_pte_range->set_page_dirty)
106  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
107  *
108  *  ->task->proc_lock
109  *    ->dcache_lock             (proc_pid_lookup)
110  */
111
112 /*
113  * Remove a page from the page cache and free it. Caller has to make
114  * sure the page is locked and that nobody else uses it - or that usage
115  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
116  */
117 void __remove_from_page_cache(struct page *page)
118 {
119         struct address_space *mapping = page->mapping;
120
121         radix_tree_delete(&mapping->page_tree, page->index);
122         page->mapping = NULL;
123         mapping->nrpages--;
124         __dec_zone_page_state(page, NR_FILE_PAGES);
125         BUG_ON(page_mapped(page));
126
127         /*
128          * Some filesystems seem to re-dirty the page even after
129          * the VM has canceled the dirty bit (eg ext3 journaling).
130          *
131          * Fix it up by doing a final dirty accounting check after
132          * having removed the page entirely.
133          */
134         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
135                 dec_zone_page_state(page, NR_FILE_DIRTY);
136                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
137         }
138 }
139
140 void remove_from_page_cache(struct page *page)
141 {
142         struct address_space *mapping = page->mapping;
143
144         BUG_ON(!PageLocked(page));
145
146         write_lock_irq(&mapping->tree_lock);
147         __remove_from_page_cache(page);
148         write_unlock_irq(&mapping->tree_lock);
149 }
150
151 static int sync_page(void *word)
152 {
153         struct address_space *mapping;
154         struct page *page;
155
156         page = container_of((unsigned long *)word, struct page, flags);
157
158         /*
159          * page_mapping() is being called without PG_locked held.
160          * Some knowledge of the state and use of the page is used to
161          * reduce the requirements down to a memory barrier.
162          * The danger here is of a stale page_mapping() return value
163          * indicating a struct address_space different from the one it's
164          * associated with when it is associated with one.
165          * After smp_mb(), it's either the correct page_mapping() for
166          * the page, or an old page_mapping() and the page's own
167          * page_mapping() has gone NULL.
168          * The ->sync_page() address_space operation must tolerate
169          * page_mapping() going NULL. By an amazing coincidence,
170          * this comes about because none of the users of the page
171          * in the ->sync_page() methods make essential use of the
172          * page_mapping(), merely passing the page down to the backing
173          * device's unplug functions when it's non-NULL, which in turn
174          * ignore it for all cases but swap, where only page_private(page) is
175          * of interest. When page_mapping() does go NULL, the entire
176          * call stack gracefully ignores the page and returns.
177          * -- wli
178          */
179         smp_mb();
180         mapping = page_mapping(page);
181         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
182                 mapping->a_ops->sync_page(page);
183         io_schedule();
184         return 0;
185 }
186
187 static int sync_page_killable(void *word)
188 {
189         sync_page(word);
190         return fatal_signal_pending(current) ? -EINTR : 0;
191 }
192
193 /**
194  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
195  * @mapping:    address space structure to write
196  * @start:      offset in bytes where the range starts
197  * @end:        offset in bytes where the range ends (inclusive)
198  * @sync_mode:  enable synchronous operation
199  *
200  * Start writeback against all of a mapping's dirty pages that lie
201  * within the byte offsets <start, end> inclusive.
202  *
203  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
204  * opposed to a regular memory cleansing writeback.  The difference between
205  * these two operations is that if a dirty page/buffer is encountered, it must
206  * be waited upon, and not just skipped over.
207  */
208 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
209                                 loff_t end, int sync_mode)
210 {
211         int ret;
212         struct writeback_control wbc = {
213                 .sync_mode = sync_mode,
214                 .nr_to_write = mapping->nrpages * 2,
215                 .range_start = start,
216                 .range_end = end,
217         };
218
219         if (!mapping_cap_writeback_dirty(mapping))
220                 return 0;
221
222         ret = do_writepages(mapping, &wbc);
223         return ret;
224 }
225
226 static inline int __filemap_fdatawrite(struct address_space *mapping,
227         int sync_mode)
228 {
229         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
230 }
231
232 int filemap_fdatawrite(struct address_space *mapping)
233 {
234         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
235 }
236 EXPORT_SYMBOL(filemap_fdatawrite);
237
238 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
239                                 loff_t end)
240 {
241         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
242 }
243
244 /**
245  * filemap_flush - mostly a non-blocking flush
246  * @mapping:    target address_space
247  *
248  * This is a mostly non-blocking flush.  Not suitable for data-integrity
249  * purposes - I/O may not be started against all dirty pages.
250  */
251 int filemap_flush(struct address_space *mapping)
252 {
253         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
254 }
255 EXPORT_SYMBOL(filemap_flush);
256
257 /**
258  * wait_on_page_writeback_range - wait for writeback to complete
259  * @mapping:    target address_space
260  * @start:      beginning page index
261  * @end:        ending page index
262  *
263  * Wait for writeback to complete against pages indexed by start->end
264  * inclusive
265  */
266 int wait_on_page_writeback_range(struct address_space *mapping,
267                                 pgoff_t start, pgoff_t end)
268 {
269         struct pagevec pvec;
270         int nr_pages;
271         int ret = 0;
272         pgoff_t index;
273
274         if (end < start)
275                 return 0;
276
277         pagevec_init(&pvec, 0);
278         index = start;
279         while ((index <= end) &&
280                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
281                         PAGECACHE_TAG_WRITEBACK,
282                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
283                 unsigned i;
284
285                 for (i = 0; i < nr_pages; i++) {
286                         struct page *page = pvec.pages[i];
287
288                         /* until radix tree lookup accepts end_index */
289                         if (page->index > end)
290                                 continue;
291
292                         wait_on_page_writeback(page);
293                         if (PageError(page))
294                                 ret = -EIO;
295                 }
296                 pagevec_release(&pvec);
297                 cond_resched();
298         }
299
300         /* Check for outstanding write errors */
301         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
302                 ret = -ENOSPC;
303         if (test_and_clear_bit(AS_EIO, &mapping->flags))
304                 ret = -EIO;
305
306         return ret;
307 }
308
309 /**
310  * sync_page_range - write and wait on all pages in the passed range
311  * @inode:      target inode
312  * @mapping:    target address_space
313  * @pos:        beginning offset in pages to write
314  * @count:      number of bytes to write
315  *
316  * Write and wait upon all the pages in the passed range.  This is a "data
317  * integrity" operation.  It waits upon in-flight writeout before starting and
318  * waiting upon new writeout.  If there was an IO error, return it.
319  *
320  * We need to re-take i_mutex during the generic_osync_inode list walk because
321  * it is otherwise livelockable.
322  */
323 int sync_page_range(struct inode *inode, struct address_space *mapping,
324                         loff_t pos, loff_t count)
325 {
326         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
327         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
328         int ret;
329
330         if (!mapping_cap_writeback_dirty(mapping) || !count)
331                 return 0;
332         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
333         if (ret == 0) {
334                 mutex_lock(&inode->i_mutex);
335                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
336                 mutex_unlock(&inode->i_mutex);
337         }
338         if (ret == 0)
339                 ret = wait_on_page_writeback_range(mapping, start, end);
340         return ret;
341 }
342 EXPORT_SYMBOL(sync_page_range);
343
344 /**
345  * sync_page_range_nolock
346  * @inode:      target inode
347  * @mapping:    target address_space
348  * @pos:        beginning offset in pages to write
349  * @count:      number of bytes to write
350  *
351  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
352  * as it forces O_SYNC writers to different parts of the same file
353  * to be serialised right until io completion.
354  */
355 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
356                            loff_t pos, loff_t count)
357 {
358         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
359         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
360         int ret;
361
362         if (!mapping_cap_writeback_dirty(mapping) || !count)
363                 return 0;
364         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
365         if (ret == 0)
366                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
367         if (ret == 0)
368                 ret = wait_on_page_writeback_range(mapping, start, end);
369         return ret;
370 }
371 EXPORT_SYMBOL(sync_page_range_nolock);
372
373 /**
374  * filemap_fdatawait - wait for all under-writeback pages to complete
375  * @mapping: address space structure to wait for
376  *
377  * Walk the list of under-writeback pages of the given address space
378  * and wait for all of them.
379  */
380 int filemap_fdatawait(struct address_space *mapping)
381 {
382         loff_t i_size = i_size_read(mapping->host);
383
384         if (i_size == 0)
385                 return 0;
386
387         return wait_on_page_writeback_range(mapping, 0,
388                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
389 }
390 EXPORT_SYMBOL(filemap_fdatawait);
391
392 int filemap_write_and_wait(struct address_space *mapping)
393 {
394         int err = 0;
395
396         if (mapping->nrpages) {
397                 err = filemap_fdatawrite(mapping);
398                 /*
399                  * Even if the above returned error, the pages may be
400                  * written partially (e.g. -ENOSPC), so we wait for it.
401                  * But the -EIO is special case, it may indicate the worst
402                  * thing (e.g. bug) happened, so we avoid waiting for it.
403                  */
404                 if (err != -EIO) {
405                         int err2 = filemap_fdatawait(mapping);
406                         if (!err)
407                                 err = err2;
408                 }
409         }
410         return err;
411 }
412 EXPORT_SYMBOL(filemap_write_and_wait);
413
414 /**
415  * filemap_write_and_wait_range - write out & wait on a file range
416  * @mapping:    the address_space for the pages
417  * @lstart:     offset in bytes where the range starts
418  * @lend:       offset in bytes where the range ends (inclusive)
419  *
420  * Write out and wait upon file offsets lstart->lend, inclusive.
421  *
422  * Note that `lend' is inclusive (describes the last byte to be written) so
423  * that this function can be used to write to the very end-of-file (end = -1).
424  */
425 int filemap_write_and_wait_range(struct address_space *mapping,
426                                  loff_t lstart, loff_t lend)
427 {
428         int err = 0;
429
430         if (mapping->nrpages) {
431                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
432                                                  WB_SYNC_ALL);
433                 /* See comment of filemap_write_and_wait() */
434                 if (err != -EIO) {
435                         int err2 = wait_on_page_writeback_range(mapping,
436                                                 lstart >> PAGE_CACHE_SHIFT,
437                                                 lend >> PAGE_CACHE_SHIFT);
438                         if (!err)
439                                 err = err2;
440                 }
441         }
442         return err;
443 }
444
445 /**
446  * add_to_page_cache - add newly allocated pagecache pages
447  * @page:       page to add
448  * @mapping:    the page's address_space
449  * @offset:     page index
450  * @gfp_mask:   page allocation mode
451  *
452  * This function is used to add newly allocated pagecache pages;
453  * the page is new, so we can just run SetPageLocked() against it.
454  * The other page state flags were set by rmqueue().
455  *
456  * This function does not add the page to the LRU.  The caller must do that.
457  */
458 int add_to_page_cache(struct page *page, struct address_space *mapping,
459                 pgoff_t offset, gfp_t gfp_mask)
460 {
461         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
462
463         if (error == 0) {
464                 write_lock_irq(&mapping->tree_lock);
465                 error = radix_tree_insert(&mapping->page_tree, offset, page);
466                 if (!error) {
467                         page_cache_get(page);
468                         SetPageLocked(page);
469                         page->mapping = mapping;
470                         page->index = offset;
471                         mapping->nrpages++;
472                         __inc_zone_page_state(page, NR_FILE_PAGES);
473                 }
474                 write_unlock_irq(&mapping->tree_lock);
475                 radix_tree_preload_end();
476         }
477         return error;
478 }
479 EXPORT_SYMBOL(add_to_page_cache);
480
481 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
482                                 pgoff_t offset, gfp_t gfp_mask)
483 {
484         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
485         if (ret == 0)
486                 lru_cache_add(page);
487         return ret;
488 }
489
490 #ifdef CONFIG_NUMA
491 struct page *__page_cache_alloc(gfp_t gfp)
492 {
493         if (cpuset_do_page_mem_spread()) {
494                 int n = cpuset_mem_spread_node();
495                 return alloc_pages_node(n, gfp, 0);
496         }
497         return alloc_pages(gfp, 0);
498 }
499 EXPORT_SYMBOL(__page_cache_alloc);
500 #endif
501
502 static int __sleep_on_page_lock(void *word)
503 {
504         io_schedule();
505         return 0;
506 }
507
508 /*
509  * In order to wait for pages to become available there must be
510  * waitqueues associated with pages. By using a hash table of
511  * waitqueues where the bucket discipline is to maintain all
512  * waiters on the same queue and wake all when any of the pages
513  * become available, and for the woken contexts to check to be
514  * sure the appropriate page became available, this saves space
515  * at a cost of "thundering herd" phenomena during rare hash
516  * collisions.
517  */
518 static wait_queue_head_t *page_waitqueue(struct page *page)
519 {
520         const struct zone *zone = page_zone(page);
521
522         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
523 }
524
525 static inline void wake_up_page(struct page *page, int bit)
526 {
527         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
528 }
529
530 void wait_on_page_bit(struct page *page, int bit_nr)
531 {
532         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
533
534         if (test_bit(bit_nr, &page->flags))
535                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
536                                                         TASK_UNINTERRUPTIBLE);
537 }
538 EXPORT_SYMBOL(wait_on_page_bit);
539
540 /**
541  * unlock_page - unlock a locked page
542  * @page: the page
543  *
544  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
545  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
546  * mechananism between PageLocked pages and PageWriteback pages is shared.
547  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
548  *
549  * The first mb is necessary to safely close the critical section opened by the
550  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
551  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
552  * parallel wait_on_page_locked()).
553  */
554 void unlock_page(struct page *page)
555 {
556         smp_mb__before_clear_bit();
557         if (!TestClearPageLocked(page))
558                 BUG();
559         smp_mb__after_clear_bit(); 
560         wake_up_page(page, PG_locked);
561 }
562 EXPORT_SYMBOL(unlock_page);
563
564 /**
565  * end_page_writeback - end writeback against a page
566  * @page: the page
567  */
568 void end_page_writeback(struct page *page)
569 {
570         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
571                 if (!test_clear_page_writeback(page))
572                         BUG();
573         }
574         smp_mb__after_clear_bit();
575         wake_up_page(page, PG_writeback);
576 }
577 EXPORT_SYMBOL(end_page_writeback);
578
579 /**
580  * __lock_page - get a lock on the page, assuming we need to sleep to get it
581  * @page: the page to lock
582  *
583  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
584  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
585  * chances are that on the second loop, the block layer's plug list is empty,
586  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
587  */
588 void __lock_page(struct page *page)
589 {
590         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
591
592         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
593                                                         TASK_UNINTERRUPTIBLE);
594 }
595 EXPORT_SYMBOL(__lock_page);
596
597 int fastcall __lock_page_killable(struct page *page)
598 {
599         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
600
601         return __wait_on_bit_lock(page_waitqueue(page), &wait,
602                                         sync_page_killable, TASK_KILLABLE);
603 }
604
605 /*
606  * Variant of lock_page that does not require the caller to hold a reference
607  * on the page's mapping.
608  */
609 void __lock_page_nosync(struct page *page)
610 {
611         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
612         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
613                                                         TASK_UNINTERRUPTIBLE);
614 }
615
616 /**
617  * find_get_page - find and get a page reference
618  * @mapping: the address_space to search
619  * @offset: the page index
620  *
621  * Is there a pagecache struct page at the given (mapping, offset) tuple?
622  * If yes, increment its refcount and return it; if no, return NULL.
623  */
624 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
625 {
626         struct page *page;
627
628         read_lock_irq(&mapping->tree_lock);
629         page = radix_tree_lookup(&mapping->page_tree, offset);
630         if (page)
631                 page_cache_get(page);
632         read_unlock_irq(&mapping->tree_lock);
633         return page;
634 }
635 EXPORT_SYMBOL(find_get_page);
636
637 /**
638  * find_lock_page - locate, pin and lock a pagecache page
639  * @mapping: the address_space to search
640  * @offset: the page index
641  *
642  * Locates the desired pagecache page, locks it, increments its reference
643  * count and returns its address.
644  *
645  * Returns zero if the page was not present. find_lock_page() may sleep.
646  */
647 struct page *find_lock_page(struct address_space *mapping,
648                                 pgoff_t offset)
649 {
650         struct page *page;
651
652 repeat:
653         read_lock_irq(&mapping->tree_lock);
654         page = radix_tree_lookup(&mapping->page_tree, offset);
655         if (page) {
656                 page_cache_get(page);
657                 if (TestSetPageLocked(page)) {
658                         read_unlock_irq(&mapping->tree_lock);
659                         __lock_page(page);
660
661                         /* Has the page been truncated while we slept? */
662                         if (unlikely(page->mapping != mapping)) {
663                                 unlock_page(page);
664                                 page_cache_release(page);
665                                 goto repeat;
666                         }
667                         VM_BUG_ON(page->index != offset);
668                         goto out;
669                 }
670         }
671         read_unlock_irq(&mapping->tree_lock);
672 out:
673         return page;
674 }
675 EXPORT_SYMBOL(find_lock_page);
676
677 /**
678  * find_or_create_page - locate or add a pagecache page
679  * @mapping: the page's address_space
680  * @index: the page's index into the mapping
681  * @gfp_mask: page allocation mode
682  *
683  * Locates a page in the pagecache.  If the page is not present, a new page
684  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
685  * LRU list.  The returned page is locked and has its reference count
686  * incremented.
687  *
688  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
689  * allocation!
690  *
691  * find_or_create_page() returns the desired page's address, or zero on
692  * memory exhaustion.
693  */
694 struct page *find_or_create_page(struct address_space *mapping,
695                 pgoff_t index, gfp_t gfp_mask)
696 {
697         struct page *page;
698         int err;
699 repeat:
700         page = find_lock_page(mapping, index);
701         if (!page) {
702                 page = __page_cache_alloc(gfp_mask);
703                 if (!page)
704                         return NULL;
705                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
706                 if (unlikely(err)) {
707                         page_cache_release(page);
708                         page = NULL;
709                         if (err == -EEXIST)
710                                 goto repeat;
711                 }
712         }
713         return page;
714 }
715 EXPORT_SYMBOL(find_or_create_page);
716
717 /**
718  * find_get_pages - gang pagecache lookup
719  * @mapping:    The address_space to search
720  * @start:      The starting page index
721  * @nr_pages:   The maximum number of pages
722  * @pages:      Where the resulting pages are placed
723  *
724  * find_get_pages() will search for and return a group of up to
725  * @nr_pages pages in the mapping.  The pages are placed at @pages.
726  * find_get_pages() takes a reference against the returned pages.
727  *
728  * The search returns a group of mapping-contiguous pages with ascending
729  * indexes.  There may be holes in the indices due to not-present pages.
730  *
731  * find_get_pages() returns the number of pages which were found.
732  */
733 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
734                             unsigned int nr_pages, struct page **pages)
735 {
736         unsigned int i;
737         unsigned int ret;
738
739         read_lock_irq(&mapping->tree_lock);
740         ret = radix_tree_gang_lookup(&mapping->page_tree,
741                                 (void **)pages, start, nr_pages);
742         for (i = 0; i < ret; i++)
743                 page_cache_get(pages[i]);
744         read_unlock_irq(&mapping->tree_lock);
745         return ret;
746 }
747
748 /**
749  * find_get_pages_contig - gang contiguous pagecache lookup
750  * @mapping:    The address_space to search
751  * @index:      The starting page index
752  * @nr_pages:   The maximum number of pages
753  * @pages:      Where the resulting pages are placed
754  *
755  * find_get_pages_contig() works exactly like find_get_pages(), except
756  * that the returned number of pages are guaranteed to be contiguous.
757  *
758  * find_get_pages_contig() returns the number of pages which were found.
759  */
760 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
761                                unsigned int nr_pages, struct page **pages)
762 {
763         unsigned int i;
764         unsigned int ret;
765
766         read_lock_irq(&mapping->tree_lock);
767         ret = radix_tree_gang_lookup(&mapping->page_tree,
768                                 (void **)pages, index, nr_pages);
769         for (i = 0; i < ret; i++) {
770                 if (pages[i]->mapping == NULL || pages[i]->index != index)
771                         break;
772
773                 page_cache_get(pages[i]);
774                 index++;
775         }
776         read_unlock_irq(&mapping->tree_lock);
777         return i;
778 }
779 EXPORT_SYMBOL(find_get_pages_contig);
780
781 /**
782  * find_get_pages_tag - find and return pages that match @tag
783  * @mapping:    the address_space to search
784  * @index:      the starting page index
785  * @tag:        the tag index
786  * @nr_pages:   the maximum number of pages
787  * @pages:      where the resulting pages are placed
788  *
789  * Like find_get_pages, except we only return pages which are tagged with
790  * @tag.   We update @index to index the next page for the traversal.
791  */
792 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
793                         int tag, unsigned int nr_pages, struct page **pages)
794 {
795         unsigned int i;
796         unsigned int ret;
797
798         read_lock_irq(&mapping->tree_lock);
799         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
800                                 (void **)pages, *index, nr_pages, tag);
801         for (i = 0; i < ret; i++)
802                 page_cache_get(pages[i]);
803         if (ret)
804                 *index = pages[ret - 1]->index + 1;
805         read_unlock_irq(&mapping->tree_lock);
806         return ret;
807 }
808 EXPORT_SYMBOL(find_get_pages_tag);
809
810 /**
811  * grab_cache_page_nowait - returns locked page at given index in given cache
812  * @mapping: target address_space
813  * @index: the page index
814  *
815  * Same as grab_cache_page(), but do not wait if the page is unavailable.
816  * This is intended for speculative data generators, where the data can
817  * be regenerated if the page couldn't be grabbed.  This routine should
818  * be safe to call while holding the lock for another page.
819  *
820  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
821  * and deadlock against the caller's locked page.
822  */
823 struct page *
824 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
825 {
826         struct page *page = find_get_page(mapping, index);
827
828         if (page) {
829                 if (!TestSetPageLocked(page))
830                         return page;
831                 page_cache_release(page);
832                 return NULL;
833         }
834         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
835         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
836                 page_cache_release(page);
837                 page = NULL;
838         }
839         return page;
840 }
841 EXPORT_SYMBOL(grab_cache_page_nowait);
842
843 /*
844  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
845  * a _large_ part of the i/o request. Imagine the worst scenario:
846  *
847  *      ---R__________________________________________B__________
848  *         ^ reading here                             ^ bad block(assume 4k)
849  *
850  * read(R) => miss => readahead(R...B) => media error => frustrating retries
851  * => failing the whole request => read(R) => read(R+1) =>
852  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
853  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
854  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
855  *
856  * It is going insane. Fix it by quickly scaling down the readahead size.
857  */
858 static void shrink_readahead_size_eio(struct file *filp,
859                                         struct file_ra_state *ra)
860 {
861         if (!ra->ra_pages)
862                 return;
863
864         ra->ra_pages /= 4;
865 }
866
867 /**
868  * do_generic_mapping_read - generic file read routine
869  * @mapping:    address_space to be read
870  * @ra:         file's readahead state
871  * @filp:       the file to read
872  * @ppos:       current file position
873  * @desc:       read_descriptor
874  * @actor:      read method
875  *
876  * This is a generic file read routine, and uses the
877  * mapping->a_ops->readpage() function for the actual low-level stuff.
878  *
879  * This is really ugly. But the goto's actually try to clarify some
880  * of the logic when it comes to error handling etc.
881  *
882  * Note the struct file* is only passed for the use of readpage.
883  * It may be NULL.
884  */
885 void do_generic_mapping_read(struct address_space *mapping,
886                              struct file_ra_state *ra,
887                              struct file *filp,
888                              loff_t *ppos,
889                              read_descriptor_t *desc,
890                              read_actor_t actor)
891 {
892         struct inode *inode = mapping->host;
893         pgoff_t index;
894         pgoff_t last_index;
895         pgoff_t prev_index;
896         unsigned long offset;      /* offset into pagecache page */
897         unsigned int prev_offset;
898         int error;
899
900         index = *ppos >> PAGE_CACHE_SHIFT;
901         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
902         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
903         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
904         offset = *ppos & ~PAGE_CACHE_MASK;
905
906         for (;;) {
907                 struct page *page;
908                 pgoff_t end_index;
909                 loff_t isize;
910                 unsigned long nr, ret;
911
912                 cond_resched();
913 find_page:
914                 page = find_get_page(mapping, index);
915                 if (!page) {
916                         page_cache_sync_readahead(mapping,
917                                         ra, filp,
918                                         index, last_index - index);
919                         page = find_get_page(mapping, index);
920                         if (unlikely(page == NULL))
921                                 goto no_cached_page;
922                 }
923                 if (PageReadahead(page)) {
924                         page_cache_async_readahead(mapping,
925                                         ra, filp, page,
926                                         index, last_index - index);
927                 }
928                 if (!PageUptodate(page))
929                         goto page_not_up_to_date;
930 page_ok:
931                 /*
932                  * i_size must be checked after we know the page is Uptodate.
933                  *
934                  * Checking i_size after the check allows us to calculate
935                  * the correct value for "nr", which means the zero-filled
936                  * part of the page is not copied back to userspace (unless
937                  * another truncate extends the file - this is desired though).
938                  */
939
940                 isize = i_size_read(inode);
941                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
942                 if (unlikely(!isize || index > end_index)) {
943                         page_cache_release(page);
944                         goto out;
945                 }
946
947                 /* nr is the maximum number of bytes to copy from this page */
948                 nr = PAGE_CACHE_SIZE;
949                 if (index == end_index) {
950                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
951                         if (nr <= offset) {
952                                 page_cache_release(page);
953                                 goto out;
954                         }
955                 }
956                 nr = nr - offset;
957
958                 /* If users can be writing to this page using arbitrary
959                  * virtual addresses, take care about potential aliasing
960                  * before reading the page on the kernel side.
961                  */
962                 if (mapping_writably_mapped(mapping))
963                         flush_dcache_page(page);
964
965                 /*
966                  * When a sequential read accesses a page several times,
967                  * only mark it as accessed the first time.
968                  */
969                 if (prev_index != index || offset != prev_offset)
970                         mark_page_accessed(page);
971                 prev_index = index;
972
973                 /*
974                  * Ok, we have the page, and it's up-to-date, so
975                  * now we can copy it to user space...
976                  *
977                  * The actor routine returns how many bytes were actually used..
978                  * NOTE! This may not be the same as how much of a user buffer
979                  * we filled up (we may be padding etc), so we can only update
980                  * "pos" here (the actor routine has to update the user buffer
981                  * pointers and the remaining count).
982                  */
983                 ret = actor(desc, page, offset, nr);
984                 offset += ret;
985                 index += offset >> PAGE_CACHE_SHIFT;
986                 offset &= ~PAGE_CACHE_MASK;
987                 prev_offset = offset;
988
989                 page_cache_release(page);
990                 if (ret == nr && desc->count)
991                         continue;
992                 goto out;
993
994 page_not_up_to_date:
995                 /* Get exclusive access to the page ... */
996                 if (lock_page_killable(page))
997                         goto readpage_eio;
998
999                 /* Did it get truncated before we got the lock? */
1000                 if (!page->mapping) {
1001                         unlock_page(page);
1002                         page_cache_release(page);
1003                         continue;
1004                 }
1005
1006                 /* Did somebody else fill it already? */
1007                 if (PageUptodate(page)) {
1008                         unlock_page(page);
1009                         goto page_ok;
1010                 }
1011
1012 readpage:
1013                 /* Start the actual read. The read will unlock the page. */
1014                 error = mapping->a_ops->readpage(filp, page);
1015
1016                 if (unlikely(error)) {
1017                         if (error == AOP_TRUNCATED_PAGE) {
1018                                 page_cache_release(page);
1019                                 goto find_page;
1020                         }
1021                         goto readpage_error;
1022                 }
1023
1024                 if (!PageUptodate(page)) {
1025                         if (lock_page_killable(page))
1026                                 goto readpage_eio;
1027                         if (!PageUptodate(page)) {
1028                                 if (page->mapping == NULL) {
1029                                         /*
1030                                          * invalidate_inode_pages got it
1031                                          */
1032                                         unlock_page(page);
1033                                         page_cache_release(page);
1034                                         goto find_page;
1035                                 }
1036                                 unlock_page(page);
1037                                 shrink_readahead_size_eio(filp, ra);
1038                                 goto readpage_eio;
1039                         }
1040                         unlock_page(page);
1041                 }
1042
1043                 goto page_ok;
1044
1045 readpage_eio:
1046                 error = -EIO;
1047 readpage_error:
1048                 /* UHHUH! A synchronous read error occurred. Report it */
1049                 desc->error = error;
1050                 page_cache_release(page);
1051                 goto out;
1052
1053 no_cached_page:
1054                 /*
1055                  * Ok, it wasn't cached, so we need to create a new
1056                  * page..
1057                  */
1058                 page = page_cache_alloc_cold(mapping);
1059                 if (!page) {
1060                         desc->error = -ENOMEM;
1061                         goto out;
1062                 }
1063                 error = add_to_page_cache_lru(page, mapping,
1064                                                 index, GFP_KERNEL);
1065                 if (error) {
1066                         page_cache_release(page);
1067                         if (error == -EEXIST)
1068                                 goto find_page;
1069                         desc->error = error;
1070                         goto out;
1071                 }
1072                 goto readpage;
1073         }
1074
1075 out:
1076         ra->prev_pos = prev_index;
1077         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1078         ra->prev_pos |= prev_offset;
1079
1080         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1081         if (filp)
1082                 file_accessed(filp);
1083 }
1084 EXPORT_SYMBOL(do_generic_mapping_read);
1085
1086 int file_read_actor(read_descriptor_t *desc, struct page *page,
1087                         unsigned long offset, unsigned long size)
1088 {
1089         char *kaddr;
1090         unsigned long left, count = desc->count;
1091
1092         if (size > count)
1093                 size = count;
1094
1095         /*
1096          * Faults on the destination of a read are common, so do it before
1097          * taking the kmap.
1098          */
1099         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1100                 kaddr = kmap_atomic(page, KM_USER0);
1101                 left = __copy_to_user_inatomic(desc->arg.buf,
1102                                                 kaddr + offset, size);
1103                 kunmap_atomic(kaddr, KM_USER0);
1104                 if (left == 0)
1105                         goto success;
1106         }
1107
1108         /* Do it the slow way */
1109         kaddr = kmap(page);
1110         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1111         kunmap(page);
1112
1113         if (left) {
1114                 size -= left;
1115                 desc->error = -EFAULT;
1116         }
1117 success:
1118         desc->count = count - size;
1119         desc->written += size;
1120         desc->arg.buf += size;
1121         return size;
1122 }
1123
1124 /*
1125  * Performs necessary checks before doing a write
1126  * @iov:        io vector request
1127  * @nr_segs:    number of segments in the iovec
1128  * @count:      number of bytes to write
1129  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1130  *
1131  * Adjust number of segments and amount of bytes to write (nr_segs should be
1132  * properly initialized first). Returns appropriate error code that caller
1133  * should return or zero in case that write should be allowed.
1134  */
1135 int generic_segment_checks(const struct iovec *iov,
1136                         unsigned long *nr_segs, size_t *count, int access_flags)
1137 {
1138         unsigned long   seg;
1139         size_t cnt = 0;
1140         for (seg = 0; seg < *nr_segs; seg++) {
1141                 const struct iovec *iv = &iov[seg];
1142
1143                 /*
1144                  * If any segment has a negative length, or the cumulative
1145                  * length ever wraps negative then return -EINVAL.
1146                  */
1147                 cnt += iv->iov_len;
1148                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1149                         return -EINVAL;
1150                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1151                         continue;
1152                 if (seg == 0)
1153                         return -EFAULT;
1154                 *nr_segs = seg;
1155                 cnt -= iv->iov_len;     /* This segment is no good */
1156                 break;
1157         }
1158         *count = cnt;
1159         return 0;
1160 }
1161 EXPORT_SYMBOL(generic_segment_checks);
1162
1163 /**
1164  * generic_file_aio_read - generic filesystem read routine
1165  * @iocb:       kernel I/O control block
1166  * @iov:        io vector request
1167  * @nr_segs:    number of segments in the iovec
1168  * @pos:        current file position
1169  *
1170  * This is the "read()" routine for all filesystems
1171  * that can use the page cache directly.
1172  */
1173 ssize_t
1174 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1175                 unsigned long nr_segs, loff_t pos)
1176 {
1177         struct file *filp = iocb->ki_filp;
1178         ssize_t retval;
1179         unsigned long seg;
1180         size_t count;
1181         loff_t *ppos = &iocb->ki_pos;
1182
1183         count = 0;
1184         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1185         if (retval)
1186                 return retval;
1187
1188         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1189         if (filp->f_flags & O_DIRECT) {
1190                 loff_t size;
1191                 struct address_space *mapping;
1192                 struct inode *inode;
1193
1194                 mapping = filp->f_mapping;
1195                 inode = mapping->host;
1196                 retval = 0;
1197                 if (!count)
1198                         goto out; /* skip atime */
1199                 size = i_size_read(inode);
1200                 if (pos < size) {
1201                         retval = generic_file_direct_IO(READ, iocb,
1202                                                 iov, pos, nr_segs);
1203                         if (retval > 0)
1204                                 *ppos = pos + retval;
1205                 }
1206                 if (likely(retval != 0)) {
1207                         file_accessed(filp);
1208                         goto out;
1209                 }
1210         }
1211
1212         retval = 0;
1213         if (count) {
1214                 for (seg = 0; seg < nr_segs; seg++) {
1215                         read_descriptor_t desc;
1216
1217                         desc.written = 0;
1218                         desc.arg.buf = iov[seg].iov_base;
1219                         desc.count = iov[seg].iov_len;
1220                         if (desc.count == 0)
1221                                 continue;
1222                         desc.error = 0;
1223                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1224                         retval += desc.written;
1225                         if (desc.error) {
1226                                 retval = retval ?: desc.error;
1227                                 break;
1228                         }
1229                         if (desc.count > 0)
1230                                 break;
1231                 }
1232         }
1233 out:
1234         return retval;
1235 }
1236 EXPORT_SYMBOL(generic_file_aio_read);
1237
1238 static ssize_t
1239 do_readahead(struct address_space *mapping, struct file *filp,
1240              pgoff_t index, unsigned long nr)
1241 {
1242         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1243                 return -EINVAL;
1244
1245         force_page_cache_readahead(mapping, filp, index,
1246                                         max_sane_readahead(nr));
1247         return 0;
1248 }
1249
1250 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1251 {
1252         ssize_t ret;
1253         struct file *file;
1254
1255         ret = -EBADF;
1256         file = fget(fd);
1257         if (file) {
1258                 if (file->f_mode & FMODE_READ) {
1259                         struct address_space *mapping = file->f_mapping;
1260                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1261                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1262                         unsigned long len = end - start + 1;
1263                         ret = do_readahead(mapping, file, start, len);
1264                 }
1265                 fput(file);
1266         }
1267         return ret;
1268 }
1269
1270 #ifdef CONFIG_MMU
1271 /**
1272  * page_cache_read - adds requested page to the page cache if not already there
1273  * @file:       file to read
1274  * @offset:     page index
1275  *
1276  * This adds the requested page to the page cache if it isn't already there,
1277  * and schedules an I/O to read in its contents from disk.
1278  */
1279 static int page_cache_read(struct file *file, pgoff_t offset)
1280 {
1281         struct address_space *mapping = file->f_mapping;
1282         struct page *page; 
1283         int ret;
1284
1285         do {
1286                 page = page_cache_alloc_cold(mapping);
1287                 if (!page)
1288                         return -ENOMEM;
1289
1290                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1291                 if (ret == 0)
1292                         ret = mapping->a_ops->readpage(file, page);
1293                 else if (ret == -EEXIST)
1294                         ret = 0; /* losing race to add is OK */
1295
1296                 page_cache_release(page);
1297
1298         } while (ret == AOP_TRUNCATED_PAGE);
1299                 
1300         return ret;
1301 }
1302
1303 #define MMAP_LOTSAMISS  (100)
1304
1305 /**
1306  * filemap_fault - read in file data for page fault handling
1307  * @vma:        vma in which the fault was taken
1308  * @vmf:        struct vm_fault containing details of the fault
1309  *
1310  * filemap_fault() is invoked via the vma operations vector for a
1311  * mapped memory region to read in file data during a page fault.
1312  *
1313  * The goto's are kind of ugly, but this streamlines the normal case of having
1314  * it in the page cache, and handles the special cases reasonably without
1315  * having a lot of duplicated code.
1316  */
1317 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1318 {
1319         int error;
1320         struct file *file = vma->vm_file;
1321         struct address_space *mapping = file->f_mapping;
1322         struct file_ra_state *ra = &file->f_ra;
1323         struct inode *inode = mapping->host;
1324         struct page *page;
1325         unsigned long size;
1326         int did_readaround = 0;
1327         int ret = 0;
1328
1329         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1330         if (vmf->pgoff >= size)
1331                 return VM_FAULT_SIGBUS;
1332
1333         /* If we don't want any read-ahead, don't bother */
1334         if (VM_RandomReadHint(vma))
1335                 goto no_cached_page;
1336
1337         /*
1338          * Do we have something in the page cache already?
1339          */
1340 retry_find:
1341         page = find_lock_page(mapping, vmf->pgoff);
1342         /*
1343          * For sequential accesses, we use the generic readahead logic.
1344          */
1345         if (VM_SequentialReadHint(vma)) {
1346                 if (!page) {
1347                         page_cache_sync_readahead(mapping, ra, file,
1348                                                            vmf->pgoff, 1);
1349                         page = find_lock_page(mapping, vmf->pgoff);
1350                         if (!page)
1351                                 goto no_cached_page;
1352                 }
1353                 if (PageReadahead(page)) {
1354                         page_cache_async_readahead(mapping, ra, file, page,
1355                                                            vmf->pgoff, 1);
1356                 }
1357         }
1358
1359         if (!page) {
1360                 unsigned long ra_pages;
1361
1362                 ra->mmap_miss++;
1363
1364                 /*
1365                  * Do we miss much more than hit in this file? If so,
1366                  * stop bothering with read-ahead. It will only hurt.
1367                  */
1368                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1369                         goto no_cached_page;
1370
1371                 /*
1372                  * To keep the pgmajfault counter straight, we need to
1373                  * check did_readaround, as this is an inner loop.
1374                  */
1375                 if (!did_readaround) {
1376                         ret = VM_FAULT_MAJOR;
1377                         count_vm_event(PGMAJFAULT);
1378                 }
1379                 did_readaround = 1;
1380                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1381                 if (ra_pages) {
1382                         pgoff_t start = 0;
1383
1384                         if (vmf->pgoff > ra_pages / 2)
1385                                 start = vmf->pgoff - ra_pages / 2;
1386                         do_page_cache_readahead(mapping, file, start, ra_pages);
1387                 }
1388                 page = find_lock_page(mapping, vmf->pgoff);
1389                 if (!page)
1390                         goto no_cached_page;
1391         }
1392
1393         if (!did_readaround)
1394                 ra->mmap_miss--;
1395
1396         /*
1397          * We have a locked page in the page cache, now we need to check
1398          * that it's up-to-date. If not, it is going to be due to an error.
1399          */
1400         if (unlikely(!PageUptodate(page)))
1401                 goto page_not_uptodate;
1402
1403         /* Must recheck i_size under page lock */
1404         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1405         if (unlikely(vmf->pgoff >= size)) {
1406                 unlock_page(page);
1407                 page_cache_release(page);
1408                 return VM_FAULT_SIGBUS;
1409         }
1410
1411         /*
1412          * Found the page and have a reference on it.
1413          */
1414         mark_page_accessed(page);
1415         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1416         vmf->page = page;
1417         return ret | VM_FAULT_LOCKED;
1418
1419 no_cached_page:
1420         /*
1421          * We're only likely to ever get here if MADV_RANDOM is in
1422          * effect.
1423          */
1424         error = page_cache_read(file, vmf->pgoff);
1425
1426         /*
1427          * The page we want has now been added to the page cache.
1428          * In the unlikely event that someone removed it in the
1429          * meantime, we'll just come back here and read it again.
1430          */
1431         if (error >= 0)
1432                 goto retry_find;
1433
1434         /*
1435          * An error return from page_cache_read can result if the
1436          * system is low on memory, or a problem occurs while trying
1437          * to schedule I/O.
1438          */
1439         if (error == -ENOMEM)
1440                 return VM_FAULT_OOM;
1441         return VM_FAULT_SIGBUS;
1442
1443 page_not_uptodate:
1444         /* IO error path */
1445         if (!did_readaround) {
1446                 ret = VM_FAULT_MAJOR;
1447                 count_vm_event(PGMAJFAULT);
1448         }
1449
1450         /*
1451          * Umm, take care of errors if the page isn't up-to-date.
1452          * Try to re-read it _once_. We do this synchronously,
1453          * because there really aren't any performance issues here
1454          * and we need to check for errors.
1455          */
1456         ClearPageError(page);
1457         error = mapping->a_ops->readpage(file, page);
1458         page_cache_release(page);
1459
1460         if (!error || error == AOP_TRUNCATED_PAGE)
1461                 goto retry_find;
1462
1463         /* Things didn't work out. Return zero to tell the mm layer so. */
1464         shrink_readahead_size_eio(file, ra);
1465         return VM_FAULT_SIGBUS;
1466 }
1467 EXPORT_SYMBOL(filemap_fault);
1468
1469 struct vm_operations_struct generic_file_vm_ops = {
1470         .fault          = filemap_fault,
1471 };
1472
1473 /* This is used for a general mmap of a disk file */
1474
1475 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1476 {
1477         struct address_space *mapping = file->f_mapping;
1478
1479         if (!mapping->a_ops->readpage)
1480                 return -ENOEXEC;
1481         file_accessed(file);
1482         vma->vm_ops = &generic_file_vm_ops;
1483         vma->vm_flags |= VM_CAN_NONLINEAR;
1484         return 0;
1485 }
1486
1487 /*
1488  * This is for filesystems which do not implement ->writepage.
1489  */
1490 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1491 {
1492         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1493                 return -EINVAL;
1494         return generic_file_mmap(file, vma);
1495 }
1496 #else
1497 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1498 {
1499         return -ENOSYS;
1500 }
1501 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1502 {
1503         return -ENOSYS;
1504 }
1505 #endif /* CONFIG_MMU */
1506
1507 EXPORT_SYMBOL(generic_file_mmap);
1508 EXPORT_SYMBOL(generic_file_readonly_mmap);
1509
1510 static struct page *__read_cache_page(struct address_space *mapping,
1511                                 pgoff_t index,
1512                                 int (*filler)(void *,struct page*),
1513                                 void *data)
1514 {
1515         struct page *page;
1516         int err;
1517 repeat:
1518         page = find_get_page(mapping, index);
1519         if (!page) {
1520                 page = page_cache_alloc_cold(mapping);
1521                 if (!page)
1522                         return ERR_PTR(-ENOMEM);
1523                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1524                 if (unlikely(err)) {
1525                         page_cache_release(page);
1526                         if (err == -EEXIST)
1527                                 goto repeat;
1528                         /* Presumably ENOMEM for radix tree node */
1529                         return ERR_PTR(err);
1530                 }
1531                 err = filler(data, page);
1532                 if (err < 0) {
1533                         page_cache_release(page);
1534                         page = ERR_PTR(err);
1535                 }
1536         }
1537         return page;
1538 }
1539
1540 /*
1541  * Same as read_cache_page, but don't wait for page to become unlocked
1542  * after submitting it to the filler.
1543  */
1544 struct page *read_cache_page_async(struct address_space *mapping,
1545                                 pgoff_t index,
1546                                 int (*filler)(void *,struct page*),
1547                                 void *data)
1548 {
1549         struct page *page;
1550         int err;
1551
1552 retry:
1553         page = __read_cache_page(mapping, index, filler, data);
1554         if (IS_ERR(page))
1555                 return page;
1556         if (PageUptodate(page))
1557                 goto out;
1558
1559         lock_page(page);
1560         if (!page->mapping) {
1561                 unlock_page(page);
1562                 page_cache_release(page);
1563                 goto retry;
1564         }
1565         if (PageUptodate(page)) {
1566                 unlock_page(page);
1567                 goto out;
1568         }
1569         err = filler(data, page);
1570         if (err < 0) {
1571                 page_cache_release(page);
1572                 return ERR_PTR(err);
1573         }
1574 out:
1575         mark_page_accessed(page);
1576         return page;
1577 }
1578 EXPORT_SYMBOL(read_cache_page_async);
1579
1580 /**
1581  * read_cache_page - read into page cache, fill it if needed
1582  * @mapping:    the page's address_space
1583  * @index:      the page index
1584  * @filler:     function to perform the read
1585  * @data:       destination for read data
1586  *
1587  * Read into the page cache. If a page already exists, and PageUptodate() is
1588  * not set, try to fill the page then wait for it to become unlocked.
1589  *
1590  * If the page does not get brought uptodate, return -EIO.
1591  */
1592 struct page *read_cache_page(struct address_space *mapping,
1593                                 pgoff_t index,
1594                                 int (*filler)(void *,struct page*),
1595                                 void *data)
1596 {
1597         struct page *page;
1598
1599         page = read_cache_page_async(mapping, index, filler, data);
1600         if (IS_ERR(page))
1601                 goto out;
1602         wait_on_page_locked(page);
1603         if (!PageUptodate(page)) {
1604                 page_cache_release(page);
1605                 page = ERR_PTR(-EIO);
1606         }
1607  out:
1608         return page;
1609 }
1610 EXPORT_SYMBOL(read_cache_page);
1611
1612 /*
1613  * The logic we want is
1614  *
1615  *      if suid or (sgid and xgrp)
1616  *              remove privs
1617  */
1618 int should_remove_suid(struct dentry *dentry)
1619 {
1620         mode_t mode = dentry->d_inode->i_mode;
1621         int kill = 0;
1622
1623         /* suid always must be killed */
1624         if (unlikely(mode & S_ISUID))
1625                 kill = ATTR_KILL_SUID;
1626
1627         /*
1628          * sgid without any exec bits is just a mandatory locking mark; leave
1629          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1630          */
1631         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1632                 kill |= ATTR_KILL_SGID;
1633
1634         if (unlikely(kill && !capable(CAP_FSETID)))
1635                 return kill;
1636
1637         return 0;
1638 }
1639 EXPORT_SYMBOL(should_remove_suid);
1640
1641 int __remove_suid(struct dentry *dentry, int kill)
1642 {
1643         struct iattr newattrs;
1644
1645         newattrs.ia_valid = ATTR_FORCE | kill;
1646         return notify_change(dentry, &newattrs);
1647 }
1648
1649 int remove_suid(struct dentry *dentry)
1650 {
1651         int killsuid = should_remove_suid(dentry);
1652         int killpriv = security_inode_need_killpriv(dentry);
1653         int error = 0;
1654
1655         if (killpriv < 0)
1656                 return killpriv;
1657         if (killpriv)
1658                 error = security_inode_killpriv(dentry);
1659         if (!error && killsuid)
1660                 error = __remove_suid(dentry, killsuid);
1661
1662         return error;
1663 }
1664 EXPORT_SYMBOL(remove_suid);
1665
1666 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1667                         const struct iovec *iov, size_t base, size_t bytes)
1668 {
1669         size_t copied = 0, left = 0;
1670
1671         while (bytes) {
1672                 char __user *buf = iov->iov_base + base;
1673                 int copy = min(bytes, iov->iov_len - base);
1674
1675                 base = 0;
1676                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1677                 copied += copy;
1678                 bytes -= copy;
1679                 vaddr += copy;
1680                 iov++;
1681
1682                 if (unlikely(left))
1683                         break;
1684         }
1685         return copied - left;
1686 }
1687
1688 /*
1689  * Copy as much as we can into the page and return the number of bytes which
1690  * were sucessfully copied.  If a fault is encountered then return the number of
1691  * bytes which were copied.
1692  */
1693 size_t iov_iter_copy_from_user_atomic(struct page *page,
1694                 struct iov_iter *i, unsigned long offset, size_t bytes)
1695 {
1696         char *kaddr;
1697         size_t copied;
1698
1699         BUG_ON(!in_atomic());
1700         kaddr = kmap_atomic(page, KM_USER0);
1701         if (likely(i->nr_segs == 1)) {
1702                 int left;
1703                 char __user *buf = i->iov->iov_base + i->iov_offset;
1704                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1705                                                         buf, bytes);
1706                 copied = bytes - left;
1707         } else {
1708                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1709                                                 i->iov, i->iov_offset, bytes);
1710         }
1711         kunmap_atomic(kaddr, KM_USER0);
1712
1713         return copied;
1714 }
1715 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1716
1717 /*
1718  * This has the same sideeffects and return value as
1719  * iov_iter_copy_from_user_atomic().
1720  * The difference is that it attempts to resolve faults.
1721  * Page must not be locked.
1722  */
1723 size_t iov_iter_copy_from_user(struct page *page,
1724                 struct iov_iter *i, unsigned long offset, size_t bytes)
1725 {
1726         char *kaddr;
1727         size_t copied;
1728
1729         kaddr = kmap(page);
1730         if (likely(i->nr_segs == 1)) {
1731                 int left;
1732                 char __user *buf = i->iov->iov_base + i->iov_offset;
1733                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1734                 copied = bytes - left;
1735         } else {
1736                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1737                                                 i->iov, i->iov_offset, bytes);
1738         }
1739         kunmap(page);
1740         return copied;
1741 }
1742 EXPORT_SYMBOL(iov_iter_copy_from_user);
1743
1744 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1745 {
1746         if (likely(i->nr_segs == 1)) {
1747                 i->iov_offset += bytes;
1748         } else {
1749                 const struct iovec *iov = i->iov;
1750                 size_t base = i->iov_offset;
1751
1752                 /*
1753                  * The !iov->iov_len check ensures we skip over unlikely
1754                  * zero-length segments.
1755                  */
1756                 while (bytes || !iov->iov_len) {
1757                         int copy = min(bytes, iov->iov_len - base);
1758
1759                         bytes -= copy;
1760                         base += copy;
1761                         if (iov->iov_len == base) {
1762                                 iov++;
1763                                 base = 0;
1764                         }
1765                 }
1766                 i->iov = iov;
1767                 i->iov_offset = base;
1768         }
1769 }
1770
1771 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1772 {
1773         BUG_ON(i->count < bytes);
1774
1775         __iov_iter_advance_iov(i, bytes);
1776         i->count -= bytes;
1777 }
1778 EXPORT_SYMBOL(iov_iter_advance);
1779
1780 /*
1781  * Fault in the first iovec of the given iov_iter, to a maximum length
1782  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1783  * accessed (ie. because it is an invalid address).
1784  *
1785  * writev-intensive code may want this to prefault several iovecs -- that
1786  * would be possible (callers must not rely on the fact that _only_ the
1787  * first iovec will be faulted with the current implementation).
1788  */
1789 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1790 {
1791         char __user *buf = i->iov->iov_base + i->iov_offset;
1792         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1793         return fault_in_pages_readable(buf, bytes);
1794 }
1795 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1796
1797 /*
1798  * Return the count of just the current iov_iter segment.
1799  */
1800 size_t iov_iter_single_seg_count(struct iov_iter *i)
1801 {
1802         const struct iovec *iov = i->iov;
1803         if (i->nr_segs == 1)
1804                 return i->count;
1805         else
1806                 return min(i->count, iov->iov_len - i->iov_offset);
1807 }
1808 EXPORT_SYMBOL(iov_iter_single_seg_count);
1809
1810 /*
1811  * Performs necessary checks before doing a write
1812  *
1813  * Can adjust writing position or amount of bytes to write.
1814  * Returns appropriate error code that caller should return or
1815  * zero in case that write should be allowed.
1816  */
1817 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1818 {
1819         struct inode *inode = file->f_mapping->host;
1820         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1821
1822         if (unlikely(*pos < 0))
1823                 return -EINVAL;
1824
1825         if (!isblk) {
1826                 /* FIXME: this is for backwards compatibility with 2.4 */
1827                 if (file->f_flags & O_APPEND)
1828                         *pos = i_size_read(inode);
1829
1830                 if (limit != RLIM_INFINITY) {
1831                         if (*pos >= limit) {
1832                                 send_sig(SIGXFSZ, current, 0);
1833                                 return -EFBIG;
1834                         }
1835                         if (*count > limit - (typeof(limit))*pos) {
1836                                 *count = limit - (typeof(limit))*pos;
1837                         }
1838                 }
1839         }
1840
1841         /*
1842          * LFS rule
1843          */
1844         if (unlikely(*pos + *count > MAX_NON_LFS &&
1845                                 !(file->f_flags & O_LARGEFILE))) {
1846                 if (*pos >= MAX_NON_LFS) {
1847                         return -EFBIG;
1848                 }
1849                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1850                         *count = MAX_NON_LFS - (unsigned long)*pos;
1851                 }
1852         }
1853
1854         /*
1855          * Are we about to exceed the fs block limit ?
1856          *
1857          * If we have written data it becomes a short write.  If we have
1858          * exceeded without writing data we send a signal and return EFBIG.
1859          * Linus frestrict idea will clean these up nicely..
1860          */
1861         if (likely(!isblk)) {
1862                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1863                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1864                                 return -EFBIG;
1865                         }
1866                         /* zero-length writes at ->s_maxbytes are OK */
1867                 }
1868
1869                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1870                         *count = inode->i_sb->s_maxbytes - *pos;
1871         } else {
1872 #ifdef CONFIG_BLOCK
1873                 loff_t isize;
1874                 if (bdev_read_only(I_BDEV(inode)))
1875                         return -EPERM;
1876                 isize = i_size_read(inode);
1877                 if (*pos >= isize) {
1878                         if (*count || *pos > isize)
1879                                 return -ENOSPC;
1880                 }
1881
1882                 if (*pos + *count > isize)
1883                         *count = isize - *pos;
1884 #else
1885                 return -EPERM;
1886 #endif
1887         }
1888         return 0;
1889 }
1890 EXPORT_SYMBOL(generic_write_checks);
1891
1892 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1893                                 loff_t pos, unsigned len, unsigned flags,
1894                                 struct page **pagep, void **fsdata)
1895 {
1896         const struct address_space_operations *aops = mapping->a_ops;
1897
1898         if (aops->write_begin) {
1899                 return aops->write_begin(file, mapping, pos, len, flags,
1900                                                         pagep, fsdata);
1901         } else {
1902                 int ret;
1903                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1904                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1905                 struct inode *inode = mapping->host;
1906                 struct page *page;
1907 again:
1908                 page = __grab_cache_page(mapping, index);
1909                 *pagep = page;
1910                 if (!page)
1911                         return -ENOMEM;
1912
1913                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1914                         /*
1915                          * There is no way to resolve a short write situation
1916                          * for a !Uptodate page (except by double copying in
1917                          * the caller done by generic_perform_write_2copy).
1918                          *
1919                          * Instead, we have to bring it uptodate here.
1920                          */
1921                         ret = aops->readpage(file, page);
1922                         page_cache_release(page);
1923                         if (ret) {
1924                                 if (ret == AOP_TRUNCATED_PAGE)
1925                                         goto again;
1926                                 return ret;
1927                         }
1928                         goto again;
1929                 }
1930
1931                 ret = aops->prepare_write(file, page, offset, offset+len);
1932                 if (ret) {
1933                         unlock_page(page);
1934                         page_cache_release(page);
1935                         if (pos + len > inode->i_size)
1936                                 vmtruncate(inode, inode->i_size);
1937                 }
1938                 return ret;
1939         }
1940 }
1941 EXPORT_SYMBOL(pagecache_write_begin);
1942
1943 int pagecache_write_end(struct file *file, struct address_space *mapping,
1944                                 loff_t pos, unsigned len, unsigned copied,
1945                                 struct page *page, void *fsdata)
1946 {
1947         const struct address_space_operations *aops = mapping->a_ops;
1948         int ret;
1949
1950         if (aops->write_end) {
1951                 mark_page_accessed(page);
1952                 ret = aops->write_end(file, mapping, pos, len, copied,
1953                                                         page, fsdata);
1954         } else {
1955                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1956                 struct inode *inode = mapping->host;
1957
1958                 flush_dcache_page(page);
1959                 ret = aops->commit_write(file, page, offset, offset+len);
1960                 unlock_page(page);
1961                 mark_page_accessed(page);
1962                 page_cache_release(page);
1963
1964                 if (ret < 0) {
1965                         if (pos + len > inode->i_size)
1966                                 vmtruncate(inode, inode->i_size);
1967                 } else if (ret > 0)
1968                         ret = min_t(size_t, copied, ret);
1969                 else
1970                         ret = copied;
1971         }
1972
1973         return ret;
1974 }
1975 EXPORT_SYMBOL(pagecache_write_end);
1976
1977 ssize_t
1978 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1979                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1980                 size_t count, size_t ocount)
1981 {
1982         struct file     *file = iocb->ki_filp;
1983         struct address_space *mapping = file->f_mapping;
1984         struct inode    *inode = mapping->host;
1985         ssize_t         written;
1986
1987         if (count != ocount)
1988                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1989
1990         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1991         if (written > 0) {
1992                 loff_t end = pos + written;
1993                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1994                         i_size_write(inode,  end);
1995                         mark_inode_dirty(inode);
1996                 }
1997                 *ppos = end;
1998         }
1999
2000         /*
2001          * Sync the fs metadata but not the minor inode changes and
2002          * of course not the data as we did direct DMA for the IO.
2003          * i_mutex is held, which protects generic_osync_inode() from
2004          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2005          */
2006         if ((written >= 0 || written == -EIOCBQUEUED) &&
2007             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2008                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2009                 if (err < 0)
2010                         written = err;
2011         }
2012         return written;
2013 }
2014 EXPORT_SYMBOL(generic_file_direct_write);
2015
2016 /*
2017  * Find or create a page at the given pagecache position. Return the locked
2018  * page. This function is specifically for buffered writes.
2019  */
2020 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2021 {
2022         int status;
2023         struct page *page;
2024 repeat:
2025         page = find_lock_page(mapping, index);
2026         if (likely(page))
2027                 return page;
2028
2029         page = page_cache_alloc(mapping);
2030         if (!page)
2031                 return NULL;
2032         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2033         if (unlikely(status)) {
2034                 page_cache_release(page);
2035                 if (status == -EEXIST)
2036                         goto repeat;
2037                 return NULL;
2038         }
2039         return page;
2040 }
2041 EXPORT_SYMBOL(__grab_cache_page);
2042
2043 static ssize_t generic_perform_write_2copy(struct file *file,
2044                                 struct iov_iter *i, loff_t pos)
2045 {
2046         struct address_space *mapping = file->f_mapping;
2047         const struct address_space_operations *a_ops = mapping->a_ops;
2048         struct inode *inode = mapping->host;
2049         long status = 0;
2050         ssize_t written = 0;
2051
2052         do {
2053                 struct page *src_page;
2054                 struct page *page;
2055                 pgoff_t index;          /* Pagecache index for current page */
2056                 unsigned long offset;   /* Offset into pagecache page */
2057                 unsigned long bytes;    /* Bytes to write to page */
2058                 size_t copied;          /* Bytes copied from user */
2059
2060                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2061                 index = pos >> PAGE_CACHE_SHIFT;
2062                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2063                                                 iov_iter_count(i));
2064
2065                 /*
2066                  * a non-NULL src_page indicates that we're doing the
2067                  * copy via get_user_pages and kmap.
2068                  */
2069                 src_page = NULL;
2070
2071                 /*
2072                  * Bring in the user page that we will copy from _first_.
2073                  * Otherwise there's a nasty deadlock on copying from the
2074                  * same page as we're writing to, without it being marked
2075                  * up-to-date.
2076                  *
2077                  * Not only is this an optimisation, but it is also required
2078                  * to check that the address is actually valid, when atomic
2079                  * usercopies are used, below.
2080                  */
2081                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2082                         status = -EFAULT;
2083                         break;
2084                 }
2085
2086                 page = __grab_cache_page(mapping, index);
2087                 if (!page) {
2088                         status = -ENOMEM;
2089                         break;
2090                 }
2091
2092                 /*
2093                  * non-uptodate pages cannot cope with short copies, and we
2094                  * cannot take a pagefault with the destination page locked.
2095                  * So pin the source page to copy it.
2096                  */
2097                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2098                         unlock_page(page);
2099
2100                         src_page = alloc_page(GFP_KERNEL);
2101                         if (!src_page) {
2102                                 page_cache_release(page);
2103                                 status = -ENOMEM;
2104                                 break;
2105                         }
2106
2107                         /*
2108                          * Cannot get_user_pages with a page locked for the
2109                          * same reason as we can't take a page fault with a
2110                          * page locked (as explained below).
2111                          */
2112                         copied = iov_iter_copy_from_user(src_page, i,
2113                                                                 offset, bytes);
2114                         if (unlikely(copied == 0)) {
2115                                 status = -EFAULT;
2116                                 page_cache_release(page);
2117                                 page_cache_release(src_page);
2118                                 break;
2119                         }
2120                         bytes = copied;
2121
2122                         lock_page(page);
2123                         /*
2124                          * Can't handle the page going uptodate here, because
2125                          * that means we would use non-atomic usercopies, which
2126                          * zero out the tail of the page, which can cause
2127                          * zeroes to become transiently visible. We could just
2128                          * use a non-zeroing copy, but the APIs aren't too
2129                          * consistent.
2130                          */
2131                         if (unlikely(!page->mapping || PageUptodate(page))) {
2132                                 unlock_page(page);
2133                                 page_cache_release(page);
2134                                 page_cache_release(src_page);
2135                                 continue;
2136                         }
2137                 }
2138
2139                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2140                 if (unlikely(status))
2141                         goto fs_write_aop_error;
2142
2143                 if (!src_page) {
2144                         /*
2145                          * Must not enter the pagefault handler here, because
2146                          * we hold the page lock, so we might recursively
2147                          * deadlock on the same lock, or get an ABBA deadlock
2148                          * against a different lock, or against the mmap_sem
2149                          * (which nests outside the page lock).  So increment
2150                          * preempt count, and use _atomic usercopies.
2151                          *
2152                          * The page is uptodate so we are OK to encounter a
2153                          * short copy: if unmodified parts of the page are
2154                          * marked dirty and written out to disk, it doesn't
2155                          * really matter.
2156                          */
2157                         pagefault_disable();
2158                         copied = iov_iter_copy_from_user_atomic(page, i,
2159                                                                 offset, bytes);
2160                         pagefault_enable();
2161                 } else {
2162                         void *src, *dst;
2163                         src = kmap_atomic(src_page, KM_USER0);
2164                         dst = kmap_atomic(page, KM_USER1);
2165                         memcpy(dst + offset, src + offset, bytes);
2166                         kunmap_atomic(dst, KM_USER1);
2167                         kunmap_atomic(src, KM_USER0);
2168                         copied = bytes;
2169                 }
2170                 flush_dcache_page(page);
2171
2172                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2173                 if (unlikely(status < 0))
2174                         goto fs_write_aop_error;
2175                 if (unlikely(status > 0)) /* filesystem did partial write */
2176                         copied = min_t(size_t, copied, status);
2177
2178                 unlock_page(page);
2179                 mark_page_accessed(page);
2180                 page_cache_release(page);
2181                 if (src_page)
2182                         page_cache_release(src_page);
2183
2184                 iov_iter_advance(i, copied);
2185                 pos += copied;
2186                 written += copied;
2187
2188                 balance_dirty_pages_ratelimited(mapping);
2189                 cond_resched();
2190                 continue;
2191
2192 fs_write_aop_error:
2193                 unlock_page(page);
2194                 page_cache_release(page);
2195                 if (src_page)
2196                         page_cache_release(src_page);
2197
2198                 /*
2199                  * prepare_write() may have instantiated a few blocks
2200                  * outside i_size.  Trim these off again. Don't need
2201                  * i_size_read because we hold i_mutex.
2202                  */
2203                 if (pos + bytes > inode->i_size)
2204                         vmtruncate(inode, inode->i_size);
2205                 break;
2206         } while (iov_iter_count(i));
2207
2208         return written ? written : status;
2209 }
2210
2211 static ssize_t generic_perform_write(struct file *file,
2212                                 struct iov_iter *i, loff_t pos)
2213 {
2214         struct address_space *mapping = file->f_mapping;
2215         const struct address_space_operations *a_ops = mapping->a_ops;
2216         long status = 0;
2217         ssize_t written = 0;
2218         unsigned int flags = 0;
2219
2220         /*
2221          * Copies from kernel address space cannot fail (NFSD is a big user).
2222          */
2223         if (segment_eq(get_fs(), KERNEL_DS))
2224                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2225
2226         do {
2227                 struct page *page;
2228                 pgoff_t index;          /* Pagecache index for current page */
2229                 unsigned long offset;   /* Offset into pagecache page */
2230                 unsigned long bytes;    /* Bytes to write to page */
2231                 size_t copied;          /* Bytes copied from user */
2232                 void *fsdata;
2233
2234                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2235                 index = pos >> PAGE_CACHE_SHIFT;
2236                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2237                                                 iov_iter_count(i));
2238
2239 again:
2240
2241                 /*
2242                  * Bring in the user page that we will copy from _first_.
2243                  * Otherwise there's a nasty deadlock on copying from the
2244                  * same page as we're writing to, without it being marked
2245                  * up-to-date.
2246                  *
2247                  * Not only is this an optimisation, but it is also required
2248                  * to check that the address is actually valid, when atomic
2249                  * usercopies are used, below.
2250                  */
2251                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2252                         status = -EFAULT;
2253                         break;
2254                 }
2255
2256                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2257                                                 &page, &fsdata);
2258                 if (unlikely(status))
2259                         break;
2260
2261                 pagefault_disable();
2262                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2263                 pagefault_enable();
2264                 flush_dcache_page(page);
2265
2266                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2267                                                 page, fsdata);
2268                 if (unlikely(status < 0))
2269                         break;
2270                 copied = status;
2271
2272                 cond_resched();
2273
2274                 iov_iter_advance(i, copied);
2275                 if (unlikely(copied == 0)) {
2276                         /*
2277                          * If we were unable to copy any data at all, we must
2278                          * fall back to a single segment length write.
2279                          *
2280                          * If we didn't fallback here, we could livelock
2281                          * because not all segments in the iov can be copied at
2282                          * once without a pagefault.
2283                          */
2284                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2285                                                 iov_iter_single_seg_count(i));
2286                         goto again;
2287                 }
2288                 pos += copied;
2289                 written += copied;
2290
2291                 balance_dirty_pages_ratelimited(mapping);
2292
2293         } while (iov_iter_count(i));
2294
2295         return written ? written : status;
2296 }
2297
2298 ssize_t
2299 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2300                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2301                 size_t count, ssize_t written)
2302 {
2303         struct file *file = iocb->ki_filp;
2304         struct address_space *mapping = file->f_mapping;
2305         const struct address_space_operations *a_ops = mapping->a_ops;
2306         struct inode *inode = mapping->host;
2307         ssize_t status;
2308         struct iov_iter i;
2309
2310         iov_iter_init(&i, iov, nr_segs, count, written);
2311         if (a_ops->write_begin)
2312                 status = generic_perform_write(file, &i, pos);
2313         else
2314                 status = generic_perform_write_2copy(file, &i, pos);
2315
2316         if (likely(status >= 0)) {
2317                 written += status;
2318                 *ppos = pos + status;
2319
2320                 /*
2321                  * For now, when the user asks for O_SYNC, we'll actually give
2322                  * O_DSYNC
2323                  */
2324                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2325                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2326                                 status = generic_osync_inode(inode, mapping,
2327                                                 OSYNC_METADATA|OSYNC_DATA);
2328                 }
2329         }
2330         
2331         /*
2332          * If we get here for O_DIRECT writes then we must have fallen through
2333          * to buffered writes (block instantiation inside i_size).  So we sync
2334          * the file data here, to try to honour O_DIRECT expectations.
2335          */
2336         if (unlikely(file->f_flags & O_DIRECT) && written)
2337                 status = filemap_write_and_wait(mapping);
2338
2339         return written ? written : status;
2340 }
2341 EXPORT_SYMBOL(generic_file_buffered_write);
2342
2343 static ssize_t
2344 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2345                                 unsigned long nr_segs, loff_t *ppos)
2346 {
2347         struct file *file = iocb->ki_filp;
2348         struct address_space * mapping = file->f_mapping;
2349         size_t ocount;          /* original count */
2350         size_t count;           /* after file limit checks */
2351         struct inode    *inode = mapping->host;
2352         loff_t          pos;
2353         ssize_t         written;
2354         ssize_t         err;
2355
2356         ocount = 0;
2357         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2358         if (err)
2359                 return err;
2360
2361         count = ocount;
2362         pos = *ppos;
2363
2364         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2365
2366         /* We can write back this queue in page reclaim */
2367         current->backing_dev_info = mapping->backing_dev_info;
2368         written = 0;
2369
2370         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2371         if (err)
2372                 goto out;
2373
2374         if (count == 0)
2375                 goto out;
2376
2377         err = remove_suid(file->f_path.dentry);
2378         if (err)
2379                 goto out;
2380
2381         file_update_time(file);
2382
2383         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2384         if (unlikely(file->f_flags & O_DIRECT)) {
2385                 loff_t endbyte;
2386                 ssize_t written_buffered;
2387
2388                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2389                                                         ppos, count, ocount);
2390                 if (written < 0 || written == count)
2391                         goto out;
2392                 /*
2393                  * direct-io write to a hole: fall through to buffered I/O
2394                  * for completing the rest of the request.
2395                  */
2396                 pos += written;
2397                 count -= written;
2398                 written_buffered = generic_file_buffered_write(iocb, iov,
2399                                                 nr_segs, pos, ppos, count,
2400                                                 written);
2401                 /*
2402                  * If generic_file_buffered_write() retuned a synchronous error
2403                  * then we want to return the number of bytes which were
2404                  * direct-written, or the error code if that was zero.  Note
2405                  * that this differs from normal direct-io semantics, which
2406                  * will return -EFOO even if some bytes were written.
2407                  */
2408                 if (written_buffered < 0) {
2409                         err = written_buffered;
2410                         goto out;
2411                 }
2412
2413                 /*
2414                  * We need to ensure that the page cache pages are written to
2415                  * disk and invalidated to preserve the expected O_DIRECT
2416                  * semantics.
2417                  */
2418                 endbyte = pos + written_buffered - written - 1;
2419                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2420                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2421                                             SYNC_FILE_RANGE_WRITE|
2422                                             SYNC_FILE_RANGE_WAIT_AFTER);
2423                 if (err == 0) {
2424                         written = written_buffered;
2425                         invalidate_mapping_pages(mapping,
2426                                                  pos >> PAGE_CACHE_SHIFT,
2427                                                  endbyte >> PAGE_CACHE_SHIFT);
2428                 } else {
2429                         /*
2430                          * We don't know how much we wrote, so just return
2431                          * the number of bytes which were direct-written
2432                          */
2433                 }
2434         } else {
2435                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2436                                 pos, ppos, count, written);
2437         }
2438 out:
2439         current->backing_dev_info = NULL;
2440         return written ? written : err;
2441 }
2442
2443 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2444                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2445 {
2446         struct file *file = iocb->ki_filp;
2447         struct address_space *mapping = file->f_mapping;
2448         struct inode *inode = mapping->host;
2449         ssize_t ret;
2450
2451         BUG_ON(iocb->ki_pos != pos);
2452
2453         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2454                         &iocb->ki_pos);
2455
2456         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2457                 ssize_t err;
2458
2459                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2460                 if (err < 0)
2461                         ret = err;
2462         }
2463         return ret;
2464 }
2465 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2466
2467 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2468                 unsigned long nr_segs, loff_t pos)
2469 {
2470         struct file *file = iocb->ki_filp;
2471         struct address_space *mapping = file->f_mapping;
2472         struct inode *inode = mapping->host;
2473         ssize_t ret;
2474
2475         BUG_ON(iocb->ki_pos != pos);
2476
2477         mutex_lock(&inode->i_mutex);
2478         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2479                         &iocb->ki_pos);
2480         mutex_unlock(&inode->i_mutex);
2481
2482         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2483                 ssize_t err;
2484
2485                 err = sync_page_range(inode, mapping, pos, ret);
2486                 if (err < 0)
2487                         ret = err;
2488         }
2489         return ret;
2490 }
2491 EXPORT_SYMBOL(generic_file_aio_write);
2492
2493 /*
2494  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2495  * went wrong during pagecache shootdown.
2496  */
2497 static ssize_t
2498 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2499         loff_t offset, unsigned long nr_segs)
2500 {
2501         struct file *file = iocb->ki_filp;
2502         struct address_space *mapping = file->f_mapping;
2503         ssize_t retval;
2504         size_t write_len;
2505         pgoff_t end = 0; /* silence gcc */
2506
2507         /*
2508          * If it's a write, unmap all mmappings of the file up-front.  This
2509          * will cause any pte dirty bits to be propagated into the pageframes
2510          * for the subsequent filemap_write_and_wait().
2511          */
2512         if (rw == WRITE) {
2513                 write_len = iov_length(iov, nr_segs);
2514                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2515                 if (mapping_mapped(mapping))
2516                         unmap_mapping_range(mapping, offset, write_len, 0);
2517         }
2518
2519         retval = filemap_write_and_wait(mapping);
2520         if (retval)
2521                 goto out;
2522
2523         /*
2524          * After a write we want buffered reads to be sure to go to disk to get
2525          * the new data.  We invalidate clean cached page from the region we're
2526          * about to write.  We do this *before* the write so that we can return
2527          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2528          */
2529         if (rw == WRITE && mapping->nrpages) {
2530                 retval = invalidate_inode_pages2_range(mapping,
2531                                         offset >> PAGE_CACHE_SHIFT, end);
2532                 if (retval)
2533                         goto out;
2534         }
2535
2536         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2537
2538         /*
2539          * Finally, try again to invalidate clean pages which might have been
2540          * cached by non-direct readahead, or faulted in by get_user_pages()
2541          * if the source of the write was an mmap'ed region of the file
2542          * we're writing.  Either one is a pretty crazy thing to do,
2543          * so we don't support it 100%.  If this invalidation
2544          * fails, tough, the write still worked...
2545          */
2546         if (rw == WRITE && mapping->nrpages) {
2547                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2548         }
2549 out:
2550         return retval;
2551 }
2552
2553 /**
2554  * try_to_release_page() - release old fs-specific metadata on a page
2555  *
2556  * @page: the page which the kernel is trying to free
2557  * @gfp_mask: memory allocation flags (and I/O mode)
2558  *
2559  * The address_space is to try to release any data against the page
2560  * (presumably at page->private).  If the release was successful, return `1'.
2561  * Otherwise return zero.
2562  *
2563  * The @gfp_mask argument specifies whether I/O may be performed to release
2564  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2565  *
2566  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2567  */
2568 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2569 {
2570         struct address_space * const mapping = page->mapping;
2571
2572         BUG_ON(!PageLocked(page));
2573         if (PageWriteback(page))
2574                 return 0;
2575
2576         if (mapping && mapping->a_ops->releasepage)
2577                 return mapping->a_ops->releasepage(page, gfp_mask);
2578         return try_to_free_buffers(page);
2579 }
2580
2581 EXPORT_SYMBOL(try_to_release_page);