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