generic_file_aio_read() cleanups
[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
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (vmtruncate)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  */
108
109 /*
110  * Remove a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
113  */
114 void __remove_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         mem_cgroup_uncharge_page(page);
119         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         BUG_ON(page_mapped(page));
124
125         /*
126          * Some filesystems seem to re-dirty the page even after
127          * the VM has canceled the dirty bit (eg ext3 journaling).
128          *
129          * Fix it up by doing a final dirty accounting check after
130          * having removed the page entirely.
131          */
132         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
133                 dec_zone_page_state(page, NR_FILE_DIRTY);
134                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
135         }
136 }
137
138 void remove_from_page_cache(struct page *page)
139 {
140         struct address_space *mapping = page->mapping;
141
142         BUG_ON(!PageLocked(page));
143
144         write_lock_irq(&mapping->tree_lock);
145         __remove_from_page_cache(page);
146         write_unlock_irq(&mapping->tree_lock);
147 }
148
149 static int sync_page(void *word)
150 {
151         struct address_space *mapping;
152         struct page *page;
153
154         page = container_of((unsigned long *)word, struct page, flags);
155
156         /*
157          * page_mapping() is being called without PG_locked held.
158          * Some knowledge of the state and use of the page is used to
159          * reduce the requirements down to a memory barrier.
160          * The danger here is of a stale page_mapping() return value
161          * indicating a struct address_space different from the one it's
162          * associated with when it is associated with one.
163          * After smp_mb(), it's either the correct page_mapping() for
164          * the page, or an old page_mapping() and the page's own
165          * page_mapping() has gone NULL.
166          * The ->sync_page() address_space operation must tolerate
167          * page_mapping() going NULL. By an amazing coincidence,
168          * this comes about because none of the users of the page
169          * in the ->sync_page() methods make essential use of the
170          * page_mapping(), merely passing the page down to the backing
171          * device's unplug functions when it's non-NULL, which in turn
172          * ignore it for all cases but swap, where only page_private(page) is
173          * of interest. When page_mapping() does go NULL, the entire
174          * call stack gracefully ignores the page and returns.
175          * -- wli
176          */
177         smp_mb();
178         mapping = page_mapping(page);
179         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
180                 mapping->a_ops->sync_page(page);
181         io_schedule();
182         return 0;
183 }
184
185 static int sync_page_killable(void *word)
186 {
187         sync_page(word);
188         return fatal_signal_pending(current) ? -EINTR : 0;
189 }
190
191 /**
192  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
193  * @mapping:    address space structure to write
194  * @start:      offset in bytes where the range starts
195  * @end:        offset in bytes where the range ends (inclusive)
196  * @sync_mode:  enable synchronous operation
197  *
198  * Start writeback against all of a mapping's dirty pages that lie
199  * within the byte offsets <start, end> inclusive.
200  *
201  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
202  * opposed to a regular memory cleansing writeback.  The difference between
203  * these two operations is that if a dirty page/buffer is encountered, it must
204  * be waited upon, and not just skipped over.
205  */
206 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
207                                 loff_t end, int sync_mode)
208 {
209         int ret;
210         struct writeback_control wbc = {
211                 .sync_mode = sync_mode,
212                 .nr_to_write = mapping->nrpages * 2,
213                 .range_start = start,
214                 .range_end = end,
215         };
216
217         if (!mapping_cap_writeback_dirty(mapping))
218                 return 0;
219
220         ret = do_writepages(mapping, &wbc);
221         return ret;
222 }
223
224 static inline int __filemap_fdatawrite(struct address_space *mapping,
225         int sync_mode)
226 {
227         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
228 }
229
230 int filemap_fdatawrite(struct address_space *mapping)
231 {
232         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
233 }
234 EXPORT_SYMBOL(filemap_fdatawrite);
235
236 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
237                                 loff_t end)
238 {
239         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite_range);
242
243 /**
244  * filemap_flush - mostly a non-blocking flush
245  * @mapping:    target address_space
246  *
247  * This is a mostly non-blocking flush.  Not suitable for data-integrity
248  * purposes - I/O may not be started against all dirty pages.
249  */
250 int filemap_flush(struct address_space *mapping)
251 {
252         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
253 }
254 EXPORT_SYMBOL(filemap_flush);
255
256 /**
257  * wait_on_page_writeback_range - wait for writeback to complete
258  * @mapping:    target address_space
259  * @start:      beginning page index
260  * @end:        ending page index
261  *
262  * Wait for writeback to complete against pages indexed by start->end
263  * inclusive
264  */
265 int wait_on_page_writeback_range(struct address_space *mapping,
266                                 pgoff_t start, pgoff_t end)
267 {
268         struct pagevec pvec;
269         int nr_pages;
270         int ret = 0;
271         pgoff_t index;
272
273         if (end < start)
274                 return 0;
275
276         pagevec_init(&pvec, 0);
277         index = start;
278         while ((index <= end) &&
279                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
280                         PAGECACHE_TAG_WRITEBACK,
281                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
282                 unsigned i;
283
284                 for (i = 0; i < nr_pages; i++) {
285                         struct page *page = pvec.pages[i];
286
287                         /* until radix tree lookup accepts end_index */
288                         if (page->index > end)
289                                 continue;
290
291                         wait_on_page_writeback(page);
292                         if (PageError(page))
293                                 ret = -EIO;
294                 }
295                 pagevec_release(&pvec);
296                 cond_resched();
297         }
298
299         /* Check for outstanding write errors */
300         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
301                 ret = -ENOSPC;
302         if (test_and_clear_bit(AS_EIO, &mapping->flags))
303                 ret = -EIO;
304
305         return ret;
306 }
307
308 /**
309  * sync_page_range - write and wait on all pages in the passed range
310  * @inode:      target inode
311  * @mapping:    target address_space
312  * @pos:        beginning offset in pages to write
313  * @count:      number of bytes to write
314  *
315  * Write and wait upon all the pages in the passed range.  This is a "data
316  * integrity" operation.  It waits upon in-flight writeout before starting and
317  * waiting upon new writeout.  If there was an IO error, return it.
318  *
319  * We need to re-take i_mutex during the generic_osync_inode list walk because
320  * it is otherwise livelockable.
321  */
322 int sync_page_range(struct inode *inode, struct address_space *mapping,
323                         loff_t pos, loff_t count)
324 {
325         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
326         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
327         int ret;
328
329         if (!mapping_cap_writeback_dirty(mapping) || !count)
330                 return 0;
331         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
332         if (ret == 0) {
333                 mutex_lock(&inode->i_mutex);
334                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
335                 mutex_unlock(&inode->i_mutex);
336         }
337         if (ret == 0)
338                 ret = wait_on_page_writeback_range(mapping, start, end);
339         return ret;
340 }
341 EXPORT_SYMBOL(sync_page_range);
342
343 /**
344  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
345  * @inode:      target inode
346  * @mapping:    target address_space
347  * @pos:        beginning offset in pages to write
348  * @count:      number of bytes to write
349  *
350  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
351  * as it forces O_SYNC writers to different parts of the same file
352  * to be serialised right until io completion.
353  */
354 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
355                            loff_t pos, loff_t count)
356 {
357         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
358         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
359         int ret;
360
361         if (!mapping_cap_writeback_dirty(mapping) || !count)
362                 return 0;
363         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
364         if (ret == 0)
365                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
366         if (ret == 0)
367                 ret = wait_on_page_writeback_range(mapping, start, end);
368         return ret;
369 }
370 EXPORT_SYMBOL(sync_page_range_nolock);
371
372 /**
373  * filemap_fdatawait - wait for all under-writeback pages to complete
374  * @mapping: address space structure to wait for
375  *
376  * Walk the list of under-writeback pages of the given address space
377  * and wait for all of them.
378  */
379 int filemap_fdatawait(struct address_space *mapping)
380 {
381         loff_t i_size = i_size_read(mapping->host);
382
383         if (i_size == 0)
384                 return 0;
385
386         return wait_on_page_writeback_range(mapping, 0,
387                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
388 }
389 EXPORT_SYMBOL(filemap_fdatawait);
390
391 int filemap_write_and_wait(struct address_space *mapping)
392 {
393         int err = 0;
394
395         if (mapping->nrpages) {
396                 err = filemap_fdatawrite(mapping);
397                 /*
398                  * Even if the above returned error, the pages may be
399                  * written partially (e.g. -ENOSPC), so we wait for it.
400                  * But the -EIO is special case, it may indicate the worst
401                  * thing (e.g. bug) happened, so we avoid waiting for it.
402                  */
403                 if (err != -EIO) {
404                         int err2 = filemap_fdatawait(mapping);
405                         if (!err)
406                                 err = err2;
407                 }
408         }
409         return err;
410 }
411 EXPORT_SYMBOL(filemap_write_and_wait);
412
413 /**
414  * filemap_write_and_wait_range - write out & wait on a file range
415  * @mapping:    the address_space for the pages
416  * @lstart:     offset in bytes where the range starts
417  * @lend:       offset in bytes where the range ends (inclusive)
418  *
419  * Write out and wait upon file offsets lstart->lend, inclusive.
420  *
421  * Note that `lend' is inclusive (describes the last byte to be written) so
422  * that this function can be used to write to the very end-of-file (end = -1).
423  */
424 int filemap_write_and_wait_range(struct address_space *mapping,
425                                  loff_t lstart, loff_t lend)
426 {
427         int err = 0;
428
429         if (mapping->nrpages) {
430                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
431                                                  WB_SYNC_ALL);
432                 /* See comment of filemap_write_and_wait() */
433                 if (err != -EIO) {
434                         int err2 = wait_on_page_writeback_range(mapping,
435                                                 lstart >> PAGE_CACHE_SHIFT,
436                                                 lend >> PAGE_CACHE_SHIFT);
437                         if (!err)
438                                 err = err2;
439                 }
440         }
441         return err;
442 }
443
444 /**
445  * add_to_page_cache - add newly allocated pagecache pages
446  * @page:       page to add
447  * @mapping:    the page's address_space
448  * @offset:     page index
449  * @gfp_mask:   page allocation mode
450  *
451  * This function is used to add newly allocated pagecache pages;
452  * the page is new, so we can just run SetPageLocked() against it.
453  * The other page state flags were set by rmqueue().
454  *
455  * This function does not add the page to the LRU.  The caller must do that.
456  */
457 int add_to_page_cache(struct page *page, struct address_space *mapping,
458                 pgoff_t offset, gfp_t gfp_mask)
459 {
460         int error = mem_cgroup_cache_charge(page, current->mm,
461                                         gfp_mask & ~__GFP_HIGHMEM);
462         if (error)
463                 goto out;
464
465         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
466         if (error == 0) {
467                 write_lock_irq(&mapping->tree_lock);
468                 error = radix_tree_insert(&mapping->page_tree, offset, page);
469                 if (!error) {
470                         page_cache_get(page);
471                         SetPageLocked(page);
472                         page->mapping = mapping;
473                         page->index = offset;
474                         mapping->nrpages++;
475                         __inc_zone_page_state(page, NR_FILE_PAGES);
476                 } else
477                         mem_cgroup_uncharge_page(page);
478
479                 write_unlock_irq(&mapping->tree_lock);
480                 radix_tree_preload_end();
481         } else
482                 mem_cgroup_uncharge_page(page);
483 out:
484         return error;
485 }
486 EXPORT_SYMBOL(add_to_page_cache);
487
488 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
489                                 pgoff_t offset, gfp_t gfp_mask)
490 {
491         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
492         if (ret == 0)
493                 lru_cache_add(page);
494         return ret;
495 }
496
497 #ifdef CONFIG_NUMA
498 struct page *__page_cache_alloc(gfp_t gfp)
499 {
500         if (cpuset_do_page_mem_spread()) {
501                 int n = cpuset_mem_spread_node();
502                 return alloc_pages_node(n, gfp, 0);
503         }
504         return alloc_pages(gfp, 0);
505 }
506 EXPORT_SYMBOL(__page_cache_alloc);
507 #endif
508
509 static int __sleep_on_page_lock(void *word)
510 {
511         io_schedule();
512         return 0;
513 }
514
515 /*
516  * In order to wait for pages to become available there must be
517  * waitqueues associated with pages. By using a hash table of
518  * waitqueues where the bucket discipline is to maintain all
519  * waiters on the same queue and wake all when any of the pages
520  * become available, and for the woken contexts to check to be
521  * sure the appropriate page became available, this saves space
522  * at a cost of "thundering herd" phenomena during rare hash
523  * collisions.
524  */
525 static wait_queue_head_t *page_waitqueue(struct page *page)
526 {
527         const struct zone *zone = page_zone(page);
528
529         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
530 }
531
532 static inline void wake_up_page(struct page *page, int bit)
533 {
534         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
535 }
536
537 void wait_on_page_bit(struct page *page, int bit_nr)
538 {
539         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
540
541         if (test_bit(bit_nr, &page->flags))
542                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
543                                                         TASK_UNINTERRUPTIBLE);
544 }
545 EXPORT_SYMBOL(wait_on_page_bit);
546
547 /**
548  * unlock_page - unlock a locked page
549  * @page: the page
550  *
551  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
552  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
553  * mechananism between PageLocked pages and PageWriteback pages is shared.
554  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
555  *
556  * The first mb is necessary to safely close the critical section opened by the
557  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
558  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
559  * parallel wait_on_page_locked()).
560  */
561 void unlock_page(struct page *page)
562 {
563         smp_mb__before_clear_bit();
564         if (!TestClearPageLocked(page))
565                 BUG();
566         smp_mb__after_clear_bit(); 
567         wake_up_page(page, PG_locked);
568 }
569 EXPORT_SYMBOL(unlock_page);
570
571 /**
572  * end_page_writeback - end writeback against a page
573  * @page: the page
574  */
575 void end_page_writeback(struct page *page)
576 {
577         if (TestClearPageReclaim(page))
578                 rotate_reclaimable_page(page);
579
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  * __lock_page_nosync - get a lock on the page, without calling sync_page()
616  * @page: the page to lock
617  *
618  * Variant of lock_page that does not require the caller to hold a reference
619  * on the page's mapping.
620  */
621 void __lock_page_nosync(struct page *page)
622 {
623         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
624         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
625                                                         TASK_UNINTERRUPTIBLE);
626 }
627
628 /**
629  * find_get_page - find and get a page reference
630  * @mapping: the address_space to search
631  * @offset: the page index
632  *
633  * Is there a pagecache struct page at the given (mapping, offset) tuple?
634  * If yes, increment its refcount and return it; if no, return NULL.
635  */
636 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
637 {
638         struct page *page;
639
640         read_lock_irq(&mapping->tree_lock);
641         page = radix_tree_lookup(&mapping->page_tree, offset);
642         if (page)
643                 page_cache_get(page);
644         read_unlock_irq(&mapping->tree_lock);
645         return page;
646 }
647 EXPORT_SYMBOL(find_get_page);
648
649 /**
650  * find_lock_page - locate, pin and lock a pagecache page
651  * @mapping: the address_space to search
652  * @offset: the page index
653  *
654  * Locates the desired pagecache page, locks it, increments its reference
655  * count and returns its address.
656  *
657  * Returns zero if the page was not present. find_lock_page() may sleep.
658  */
659 struct page *find_lock_page(struct address_space *mapping,
660                                 pgoff_t offset)
661 {
662         struct page *page;
663
664 repeat:
665         read_lock_irq(&mapping->tree_lock);
666         page = radix_tree_lookup(&mapping->page_tree, offset);
667         if (page) {
668                 page_cache_get(page);
669                 if (TestSetPageLocked(page)) {
670                         read_unlock_irq(&mapping->tree_lock);
671                         __lock_page(page);
672
673                         /* Has the page been truncated while we slept? */
674                         if (unlikely(page->mapping != mapping)) {
675                                 unlock_page(page);
676                                 page_cache_release(page);
677                                 goto repeat;
678                         }
679                         VM_BUG_ON(page->index != offset);
680                         goto out;
681                 }
682         }
683         read_unlock_irq(&mapping->tree_lock);
684 out:
685         return page;
686 }
687 EXPORT_SYMBOL(find_lock_page);
688
689 /**
690  * find_or_create_page - locate or add a pagecache page
691  * @mapping: the page's address_space
692  * @index: the page's index into the mapping
693  * @gfp_mask: page allocation mode
694  *
695  * Locates a page in the pagecache.  If the page is not present, a new page
696  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
697  * LRU list.  The returned page is locked and has its reference count
698  * incremented.
699  *
700  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
701  * allocation!
702  *
703  * find_or_create_page() returns the desired page's address, or zero on
704  * memory exhaustion.
705  */
706 struct page *find_or_create_page(struct address_space *mapping,
707                 pgoff_t index, gfp_t gfp_mask)
708 {
709         struct page *page;
710         int err;
711 repeat:
712         page = find_lock_page(mapping, index);
713         if (!page) {
714                 page = __page_cache_alloc(gfp_mask);
715                 if (!page)
716                         return NULL;
717                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
718                 if (unlikely(err)) {
719                         page_cache_release(page);
720                         page = NULL;
721                         if (err == -EEXIST)
722                                 goto repeat;
723                 }
724         }
725         return page;
726 }
727 EXPORT_SYMBOL(find_or_create_page);
728
729 /**
730  * find_get_pages - gang pagecache lookup
731  * @mapping:    The address_space to search
732  * @start:      The starting page index
733  * @nr_pages:   The maximum number of pages
734  * @pages:      Where the resulting pages are placed
735  *
736  * find_get_pages() will search for and return a group of up to
737  * @nr_pages pages in the mapping.  The pages are placed at @pages.
738  * find_get_pages() takes a reference against the returned pages.
739  *
740  * The search returns a group of mapping-contiguous pages with ascending
741  * indexes.  There may be holes in the indices due to not-present pages.
742  *
743  * find_get_pages() returns the number of pages which were found.
744  */
745 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
746                             unsigned int nr_pages, struct page **pages)
747 {
748         unsigned int i;
749         unsigned int ret;
750
751         read_lock_irq(&mapping->tree_lock);
752         ret = radix_tree_gang_lookup(&mapping->page_tree,
753                                 (void **)pages, start, nr_pages);
754         for (i = 0; i < ret; i++)
755                 page_cache_get(pages[i]);
756         read_unlock_irq(&mapping->tree_lock);
757         return ret;
758 }
759
760 /**
761  * find_get_pages_contig - gang contiguous pagecache lookup
762  * @mapping:    The address_space to search
763  * @index:      The starting page index
764  * @nr_pages:   The maximum number of pages
765  * @pages:      Where the resulting pages are placed
766  *
767  * find_get_pages_contig() works exactly like find_get_pages(), except
768  * that the returned number of pages are guaranteed to be contiguous.
769  *
770  * find_get_pages_contig() returns the number of pages which were found.
771  */
772 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
773                                unsigned int nr_pages, struct page **pages)
774 {
775         unsigned int i;
776         unsigned int ret;
777
778         read_lock_irq(&mapping->tree_lock);
779         ret = radix_tree_gang_lookup(&mapping->page_tree,
780                                 (void **)pages, index, nr_pages);
781         for (i = 0; i < ret; i++) {
782                 if (pages[i]->mapping == NULL || pages[i]->index != index)
783                         break;
784
785                 page_cache_get(pages[i]);
786                 index++;
787         }
788         read_unlock_irq(&mapping->tree_lock);
789         return i;
790 }
791 EXPORT_SYMBOL(find_get_pages_contig);
792
793 /**
794  * find_get_pages_tag - find and return pages that match @tag
795  * @mapping:    the address_space to search
796  * @index:      the starting page index
797  * @tag:        the tag index
798  * @nr_pages:   the maximum number of pages
799  * @pages:      where the resulting pages are placed
800  *
801  * Like find_get_pages, except we only return pages which are tagged with
802  * @tag.   We update @index to index the next page for the traversal.
803  */
804 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
805                         int tag, unsigned int nr_pages, struct page **pages)
806 {
807         unsigned int i;
808         unsigned int ret;
809
810         read_lock_irq(&mapping->tree_lock);
811         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
812                                 (void **)pages, *index, nr_pages, tag);
813         for (i = 0; i < ret; i++)
814                 page_cache_get(pages[i]);
815         if (ret)
816                 *index = pages[ret - 1]->index + 1;
817         read_unlock_irq(&mapping->tree_lock);
818         return ret;
819 }
820 EXPORT_SYMBOL(find_get_pages_tag);
821
822 /**
823  * grab_cache_page_nowait - returns locked page at given index in given cache
824  * @mapping: target address_space
825  * @index: the page index
826  *
827  * Same as grab_cache_page(), but do not wait if the page is unavailable.
828  * This is intended for speculative data generators, where the data can
829  * be regenerated if the page couldn't be grabbed.  This routine should
830  * be safe to call while holding the lock for another page.
831  *
832  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
833  * and deadlock against the caller's locked page.
834  */
835 struct page *
836 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
837 {
838         struct page *page = find_get_page(mapping, index);
839
840         if (page) {
841                 if (!TestSetPageLocked(page))
842                         return page;
843                 page_cache_release(page);
844                 return NULL;
845         }
846         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
847         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
848                 page_cache_release(page);
849                 page = NULL;
850         }
851         return page;
852 }
853 EXPORT_SYMBOL(grab_cache_page_nowait);
854
855 /*
856  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
857  * a _large_ part of the i/o request. Imagine the worst scenario:
858  *
859  *      ---R__________________________________________B__________
860  *         ^ reading here                             ^ bad block(assume 4k)
861  *
862  * read(R) => miss => readahead(R...B) => media error => frustrating retries
863  * => failing the whole request => read(R) => read(R+1) =>
864  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
865  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
866  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
867  *
868  * It is going insane. Fix it by quickly scaling down the readahead size.
869  */
870 static void shrink_readahead_size_eio(struct file *filp,
871                                         struct file_ra_state *ra)
872 {
873         if (!ra->ra_pages)
874                 return;
875
876         ra->ra_pages /= 4;
877 }
878
879 /**
880  * do_generic_file_read - generic file read routine
881  * @filp:       the file to read
882  * @ppos:       current file position
883  * @desc:       read_descriptor
884  * @actor:      read method
885  *
886  * This is a generic file read routine, and uses the
887  * mapping->a_ops->readpage() function for the actual low-level stuff.
888  *
889  * This is really ugly. But the goto's actually try to clarify some
890  * of the logic when it comes to error handling etc.
891  */
892 static void do_generic_file_read(struct file *filp, loff_t *ppos,
893                 read_descriptor_t *desc, read_actor_t actor)
894 {
895         struct address_space *mapping = filp->f_mapping;
896         struct inode *inode = mapping->host;
897         struct file_ra_state *ra = &filp->f_ra;
898         pgoff_t index;
899         pgoff_t last_index;
900         pgoff_t prev_index;
901         unsigned long offset;      /* offset into pagecache page */
902         unsigned int prev_offset;
903         int error;
904
905         index = *ppos >> PAGE_CACHE_SHIFT;
906         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
907         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
908         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
909         offset = *ppos & ~PAGE_CACHE_MASK;
910
911         for (;;) {
912                 struct page *page;
913                 pgoff_t end_index;
914                 loff_t isize;
915                 unsigned long nr, ret;
916
917                 cond_resched();
918 find_page:
919                 page = find_get_page(mapping, index);
920                 if (!page) {
921                         page_cache_sync_readahead(mapping,
922                                         ra, filp,
923                                         index, last_index - index);
924                         page = find_get_page(mapping, index);
925                         if (unlikely(page == NULL))
926                                 goto no_cached_page;
927                 }
928                 if (PageReadahead(page)) {
929                         page_cache_async_readahead(mapping,
930                                         ra, filp, page,
931                                         index, last_index - index);
932                 }
933                 if (!PageUptodate(page))
934                         goto page_not_up_to_date;
935 page_ok:
936                 /*
937                  * i_size must be checked after we know the page is Uptodate.
938                  *
939                  * Checking i_size after the check allows us to calculate
940                  * the correct value for "nr", which means the zero-filled
941                  * part of the page is not copied back to userspace (unless
942                  * another truncate extends the file - this is desired though).
943                  */
944
945                 isize = i_size_read(inode);
946                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
947                 if (unlikely(!isize || index > end_index)) {
948                         page_cache_release(page);
949                         goto out;
950                 }
951
952                 /* nr is the maximum number of bytes to copy from this page */
953                 nr = PAGE_CACHE_SIZE;
954                 if (index == end_index) {
955                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
956                         if (nr <= offset) {
957                                 page_cache_release(page);
958                                 goto out;
959                         }
960                 }
961                 nr = nr - offset;
962
963                 /* If users can be writing to this page using arbitrary
964                  * virtual addresses, take care about potential aliasing
965                  * before reading the page on the kernel side.
966                  */
967                 if (mapping_writably_mapped(mapping))
968                         flush_dcache_page(page);
969
970                 /*
971                  * When a sequential read accesses a page several times,
972                  * only mark it as accessed the first time.
973                  */
974                 if (prev_index != index || offset != prev_offset)
975                         mark_page_accessed(page);
976                 prev_index = index;
977
978                 /*
979                  * Ok, we have the page, and it's up-to-date, so
980                  * now we can copy it to user space...
981                  *
982                  * The actor routine returns how many bytes were actually used..
983                  * NOTE! This may not be the same as how much of a user buffer
984                  * we filled up (we may be padding etc), so we can only update
985                  * "pos" here (the actor routine has to update the user buffer
986                  * pointers and the remaining count).
987                  */
988                 ret = actor(desc, page, offset, nr);
989                 offset += ret;
990                 index += offset >> PAGE_CACHE_SHIFT;
991                 offset &= ~PAGE_CACHE_MASK;
992                 prev_offset = offset;
993
994                 page_cache_release(page);
995                 if (ret == nr && desc->count)
996                         continue;
997                 goto out;
998
999 page_not_up_to_date:
1000                 /* Get exclusive access to the page ... */
1001                 if (lock_page_killable(page))
1002                         goto readpage_eio;
1003
1004                 /* Did it get truncated before we got the lock? */
1005                 if (!page->mapping) {
1006                         unlock_page(page);
1007                         page_cache_release(page);
1008                         continue;
1009                 }
1010
1011                 /* Did somebody else fill it already? */
1012                 if (PageUptodate(page)) {
1013                         unlock_page(page);
1014                         goto page_ok;
1015                 }
1016
1017 readpage:
1018                 /* Start the actual read. The read will unlock the page. */
1019                 error = mapping->a_ops->readpage(filp, page);
1020
1021                 if (unlikely(error)) {
1022                         if (error == AOP_TRUNCATED_PAGE) {
1023                                 page_cache_release(page);
1024                                 goto find_page;
1025                         }
1026                         goto readpage_error;
1027                 }
1028
1029                 if (!PageUptodate(page)) {
1030                         if (lock_page_killable(page))
1031                                 goto readpage_eio;
1032                         if (!PageUptodate(page)) {
1033                                 if (page->mapping == NULL) {
1034                                         /*
1035                                          * invalidate_inode_pages got it
1036                                          */
1037                                         unlock_page(page);
1038                                         page_cache_release(page);
1039                                         goto find_page;
1040                                 }
1041                                 unlock_page(page);
1042                                 shrink_readahead_size_eio(filp, ra);
1043                                 goto readpage_eio;
1044                         }
1045                         unlock_page(page);
1046                 }
1047
1048                 goto page_ok;
1049
1050 readpage_eio:
1051                 error = -EIO;
1052 readpage_error:
1053                 /* UHHUH! A synchronous read error occurred. Report it */
1054                 desc->error = error;
1055                 page_cache_release(page);
1056                 goto out;
1057
1058 no_cached_page:
1059                 /*
1060                  * Ok, it wasn't cached, so we need to create a new
1061                  * page..
1062                  */
1063                 page = page_cache_alloc_cold(mapping);
1064                 if (!page) {
1065                         desc->error = -ENOMEM;
1066                         goto out;
1067                 }
1068                 error = add_to_page_cache_lru(page, mapping,
1069                                                 index, GFP_KERNEL);
1070                 if (error) {
1071                         page_cache_release(page);
1072                         if (error == -EEXIST)
1073                                 goto find_page;
1074                         desc->error = error;
1075                         goto out;
1076                 }
1077                 goto readpage;
1078         }
1079
1080 out:
1081         ra->prev_pos = prev_index;
1082         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1083         ra->prev_pos |= prev_offset;
1084
1085         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1086         if (filp)
1087                 file_accessed(filp);
1088 }
1089
1090 int file_read_actor(read_descriptor_t *desc, struct page *page,
1091                         unsigned long offset, unsigned long size)
1092 {
1093         char *kaddr;
1094         unsigned long left, count = desc->count;
1095
1096         if (size > count)
1097                 size = count;
1098
1099         /*
1100          * Faults on the destination of a read are common, so do it before
1101          * taking the kmap.
1102          */
1103         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1104                 kaddr = kmap_atomic(page, KM_USER0);
1105                 left = __copy_to_user_inatomic(desc->arg.buf,
1106                                                 kaddr + offset, size);
1107                 kunmap_atomic(kaddr, KM_USER0);
1108                 if (left == 0)
1109                         goto success;
1110         }
1111
1112         /* Do it the slow way */
1113         kaddr = kmap(page);
1114         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1115         kunmap(page);
1116
1117         if (left) {
1118                 size -= left;
1119                 desc->error = -EFAULT;
1120         }
1121 success:
1122         desc->count = count - size;
1123         desc->written += size;
1124         desc->arg.buf += size;
1125         return size;
1126 }
1127
1128 /*
1129  * Performs necessary checks before doing a write
1130  * @iov:        io vector request
1131  * @nr_segs:    number of segments in the iovec
1132  * @count:      number of bytes to write
1133  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1134  *
1135  * Adjust number of segments and amount of bytes to write (nr_segs should be
1136  * properly initialized first). Returns appropriate error code that caller
1137  * should return or zero in case that write should be allowed.
1138  */
1139 int generic_segment_checks(const struct iovec *iov,
1140                         unsigned long *nr_segs, size_t *count, int access_flags)
1141 {
1142         unsigned long   seg;
1143         size_t cnt = 0;
1144         for (seg = 0; seg < *nr_segs; seg++) {
1145                 const struct iovec *iv = &iov[seg];
1146
1147                 /*
1148                  * If any segment has a negative length, or the cumulative
1149                  * length ever wraps negative then return -EINVAL.
1150                  */
1151                 cnt += iv->iov_len;
1152                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1153                         return -EINVAL;
1154                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1155                         continue;
1156                 if (seg == 0)
1157                         return -EFAULT;
1158                 *nr_segs = seg;
1159                 cnt -= iv->iov_len;     /* This segment is no good */
1160                 break;
1161         }
1162         *count = cnt;
1163         return 0;
1164 }
1165 EXPORT_SYMBOL(generic_segment_checks);
1166
1167 /**
1168  * generic_file_aio_read - generic filesystem read routine
1169  * @iocb:       kernel I/O control block
1170  * @iov:        io vector request
1171  * @nr_segs:    number of segments in the iovec
1172  * @pos:        current file position
1173  *
1174  * This is the "read()" routine for all filesystems
1175  * that can use the page cache directly.
1176  */
1177 ssize_t
1178 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1179                 unsigned long nr_segs, loff_t pos)
1180 {
1181         struct file *filp = iocb->ki_filp;
1182         ssize_t retval;
1183         unsigned long seg;
1184         size_t count;
1185         loff_t *ppos = &iocb->ki_pos;
1186
1187         count = 0;
1188         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1189         if (retval)
1190                 return retval;
1191
1192         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1193         if (filp->f_flags & O_DIRECT) {
1194                 loff_t size;
1195                 struct address_space *mapping;
1196                 struct inode *inode;
1197
1198                 mapping = filp->f_mapping;
1199                 inode = mapping->host;
1200                 if (!count)
1201                         goto out; /* skip atime */
1202                 size = i_size_read(inode);
1203                 if (pos < size) {
1204                         retval = filemap_write_and_wait(mapping);
1205                         if (!retval) {
1206                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1207                                                         iov, pos, nr_segs);
1208                         }
1209                         if (retval > 0)
1210                                 *ppos = pos + retval;
1211                         if (retval) {
1212                                 file_accessed(filp);
1213                                 goto out;
1214                         }
1215                 }
1216         }
1217
1218         for (seg = 0; seg < nr_segs; seg++) {
1219                 read_descriptor_t desc;
1220
1221                 desc.written = 0;
1222                 desc.arg.buf = iov[seg].iov_base;
1223                 desc.count = iov[seg].iov_len;
1224                 if (desc.count == 0)
1225                         continue;
1226                 desc.error = 0;
1227                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1228                 retval += desc.written;
1229                 if (desc.error) {
1230                         retval = retval ?: desc.error;
1231                         break;
1232                 }
1233                 if (desc.count > 0)
1234                         break;
1235         }
1236 out:
1237         return retval;
1238 }
1239 EXPORT_SYMBOL(generic_file_aio_read);
1240
1241 static ssize_t
1242 do_readahead(struct address_space *mapping, struct file *filp,
1243              pgoff_t index, unsigned long nr)
1244 {
1245         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1246                 return -EINVAL;
1247
1248         force_page_cache_readahead(mapping, filp, index,
1249                                         max_sane_readahead(nr));
1250         return 0;
1251 }
1252
1253 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1254 {
1255         ssize_t ret;
1256         struct file *file;
1257
1258         ret = -EBADF;
1259         file = fget(fd);
1260         if (file) {
1261                 if (file->f_mode & FMODE_READ) {
1262                         struct address_space *mapping = file->f_mapping;
1263                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1264                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1265                         unsigned long len = end - start + 1;
1266                         ret = do_readahead(mapping, file, start, len);
1267                 }
1268                 fput(file);
1269         }
1270         return ret;
1271 }
1272
1273 #ifdef CONFIG_MMU
1274 /**
1275  * page_cache_read - adds requested page to the page cache if not already there
1276  * @file:       file to read
1277  * @offset:     page index
1278  *
1279  * This adds the requested page to the page cache if it isn't already there,
1280  * and schedules an I/O to read in its contents from disk.
1281  */
1282 static int page_cache_read(struct file *file, pgoff_t offset)
1283 {
1284         struct address_space *mapping = file->f_mapping;
1285         struct page *page; 
1286         int ret;
1287
1288         do {
1289                 page = page_cache_alloc_cold(mapping);
1290                 if (!page)
1291                         return -ENOMEM;
1292
1293                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1294                 if (ret == 0)
1295                         ret = mapping->a_ops->readpage(file, page);
1296                 else if (ret == -EEXIST)
1297                         ret = 0; /* losing race to add is OK */
1298
1299                 page_cache_release(page);
1300
1301         } while (ret == AOP_TRUNCATED_PAGE);
1302                 
1303         return ret;
1304 }
1305
1306 #define MMAP_LOTSAMISS  (100)
1307
1308 /**
1309  * filemap_fault - read in file data for page fault handling
1310  * @vma:        vma in which the fault was taken
1311  * @vmf:        struct vm_fault containing details of the fault
1312  *
1313  * filemap_fault() is invoked via the vma operations vector for a
1314  * mapped memory region to read in file data during a page fault.
1315  *
1316  * The goto's are kind of ugly, but this streamlines the normal case of having
1317  * it in the page cache, and handles the special cases reasonably without
1318  * having a lot of duplicated code.
1319  */
1320 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1321 {
1322         int error;
1323         struct file *file = vma->vm_file;
1324         struct address_space *mapping = file->f_mapping;
1325         struct file_ra_state *ra = &file->f_ra;
1326         struct inode *inode = mapping->host;
1327         struct page *page;
1328         pgoff_t size;
1329         int did_readaround = 0;
1330         int ret = 0;
1331
1332         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1333         if (vmf->pgoff >= size)
1334                 return VM_FAULT_SIGBUS;
1335
1336         /* If we don't want any read-ahead, don't bother */
1337         if (VM_RandomReadHint(vma))
1338                 goto no_cached_page;
1339
1340         /*
1341          * Do we have something in the page cache already?
1342          */
1343 retry_find:
1344         page = find_lock_page(mapping, vmf->pgoff);
1345         /*
1346          * For sequential accesses, we use the generic readahead logic.
1347          */
1348         if (VM_SequentialReadHint(vma)) {
1349                 if (!page) {
1350                         page_cache_sync_readahead(mapping, ra, file,
1351                                                            vmf->pgoff, 1);
1352                         page = find_lock_page(mapping, vmf->pgoff);
1353                         if (!page)
1354                                 goto no_cached_page;
1355                 }
1356                 if (PageReadahead(page)) {
1357                         page_cache_async_readahead(mapping, ra, file, page,
1358                                                            vmf->pgoff, 1);
1359                 }
1360         }
1361
1362         if (!page) {
1363                 unsigned long ra_pages;
1364
1365                 ra->mmap_miss++;
1366
1367                 /*
1368                  * Do we miss much more than hit in this file? If so,
1369                  * stop bothering with read-ahead. It will only hurt.
1370                  */
1371                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1372                         goto no_cached_page;
1373
1374                 /*
1375                  * To keep the pgmajfault counter straight, we need to
1376                  * check did_readaround, as this is an inner loop.
1377                  */
1378                 if (!did_readaround) {
1379                         ret = VM_FAULT_MAJOR;
1380                         count_vm_event(PGMAJFAULT);
1381                 }
1382                 did_readaround = 1;
1383                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1384                 if (ra_pages) {
1385                         pgoff_t start = 0;
1386
1387                         if (vmf->pgoff > ra_pages / 2)
1388                                 start = vmf->pgoff - ra_pages / 2;
1389                         do_page_cache_readahead(mapping, file, start, ra_pages);
1390                 }
1391                 page = find_lock_page(mapping, vmf->pgoff);
1392                 if (!page)
1393                         goto no_cached_page;
1394         }
1395
1396         if (!did_readaround)
1397                 ra->mmap_miss--;
1398
1399         /*
1400          * We have a locked page in the page cache, now we need to check
1401          * that it's up-to-date. If not, it is going to be due to an error.
1402          */
1403         if (unlikely(!PageUptodate(page)))
1404                 goto page_not_uptodate;
1405
1406         /* Must recheck i_size under page lock */
1407         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1408         if (unlikely(vmf->pgoff >= size)) {
1409                 unlock_page(page);
1410                 page_cache_release(page);
1411                 return VM_FAULT_SIGBUS;
1412         }
1413
1414         /*
1415          * Found the page and have a reference on it.
1416          */
1417         mark_page_accessed(page);
1418         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1419         vmf->page = page;
1420         return ret | VM_FAULT_LOCKED;
1421
1422 no_cached_page:
1423         /*
1424          * We're only likely to ever get here if MADV_RANDOM is in
1425          * effect.
1426          */
1427         error = page_cache_read(file, vmf->pgoff);
1428
1429         /*
1430          * The page we want has now been added to the page cache.
1431          * In the unlikely event that someone removed it in the
1432          * meantime, we'll just come back here and read it again.
1433          */
1434         if (error >= 0)
1435                 goto retry_find;
1436
1437         /*
1438          * An error return from page_cache_read can result if the
1439          * system is low on memory, or a problem occurs while trying
1440          * to schedule I/O.
1441          */
1442         if (error == -ENOMEM)
1443                 return VM_FAULT_OOM;
1444         return VM_FAULT_SIGBUS;
1445
1446 page_not_uptodate:
1447         /* IO error path */
1448         if (!did_readaround) {
1449                 ret = VM_FAULT_MAJOR;
1450                 count_vm_event(PGMAJFAULT);
1451         }
1452
1453         /*
1454          * Umm, take care of errors if the page isn't up-to-date.
1455          * Try to re-read it _once_. We do this synchronously,
1456          * because there really aren't any performance issues here
1457          * and we need to check for errors.
1458          */
1459         ClearPageError(page);
1460         error = mapping->a_ops->readpage(file, page);
1461         if (!error) {
1462                 wait_on_page_locked(page);
1463                 if (!PageUptodate(page))
1464                         error = -EIO;
1465         }
1466         page_cache_release(page);
1467
1468         if (!error || error == AOP_TRUNCATED_PAGE)
1469                 goto retry_find;
1470
1471         /* Things didn't work out. Return zero to tell the mm layer so. */
1472         shrink_readahead_size_eio(file, ra);
1473         return VM_FAULT_SIGBUS;
1474 }
1475 EXPORT_SYMBOL(filemap_fault);
1476
1477 struct vm_operations_struct generic_file_vm_ops = {
1478         .fault          = filemap_fault,
1479 };
1480
1481 /* This is used for a general mmap of a disk file */
1482
1483 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1484 {
1485         struct address_space *mapping = file->f_mapping;
1486
1487         if (!mapping->a_ops->readpage)
1488                 return -ENOEXEC;
1489         file_accessed(file);
1490         vma->vm_ops = &generic_file_vm_ops;
1491         vma->vm_flags |= VM_CAN_NONLINEAR;
1492         return 0;
1493 }
1494
1495 /*
1496  * This is for filesystems which do not implement ->writepage.
1497  */
1498 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1499 {
1500         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1501                 return -EINVAL;
1502         return generic_file_mmap(file, vma);
1503 }
1504 #else
1505 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1506 {
1507         return -ENOSYS;
1508 }
1509 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1510 {
1511         return -ENOSYS;
1512 }
1513 #endif /* CONFIG_MMU */
1514
1515 EXPORT_SYMBOL(generic_file_mmap);
1516 EXPORT_SYMBOL(generic_file_readonly_mmap);
1517
1518 static struct page *__read_cache_page(struct address_space *mapping,
1519                                 pgoff_t index,
1520                                 int (*filler)(void *,struct page*),
1521                                 void *data)
1522 {
1523         struct page *page;
1524         int err;
1525 repeat:
1526         page = find_get_page(mapping, index);
1527         if (!page) {
1528                 page = page_cache_alloc_cold(mapping);
1529                 if (!page)
1530                         return ERR_PTR(-ENOMEM);
1531                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1532                 if (unlikely(err)) {
1533                         page_cache_release(page);
1534                         if (err == -EEXIST)
1535                                 goto repeat;
1536                         /* Presumably ENOMEM for radix tree node */
1537                         return ERR_PTR(err);
1538                 }
1539                 err = filler(data, page);
1540                 if (err < 0) {
1541                         page_cache_release(page);
1542                         page = ERR_PTR(err);
1543                 }
1544         }
1545         return page;
1546 }
1547
1548 /**
1549  * read_cache_page_async - read into page cache, fill it if needed
1550  * @mapping:    the page's address_space
1551  * @index:      the page index
1552  * @filler:     function to perform the read
1553  * @data:       destination for read data
1554  *
1555  * Same as read_cache_page, but don't wait for page to become unlocked
1556  * after submitting it to the filler.
1557  *
1558  * Read into the page cache. If a page already exists, and PageUptodate() is
1559  * not set, try to fill the page but don't wait for it to become unlocked.
1560  *
1561  * If the page does not get brought uptodate, return -EIO.
1562  */
1563 struct page *read_cache_page_async(struct address_space *mapping,
1564                                 pgoff_t index,
1565                                 int (*filler)(void *,struct page*),
1566                                 void *data)
1567 {
1568         struct page *page;
1569         int err;
1570
1571 retry:
1572         page = __read_cache_page(mapping, index, filler, data);
1573         if (IS_ERR(page))
1574                 return page;
1575         if (PageUptodate(page))
1576                 goto out;
1577
1578         lock_page(page);
1579         if (!page->mapping) {
1580                 unlock_page(page);
1581                 page_cache_release(page);
1582                 goto retry;
1583         }
1584         if (PageUptodate(page)) {
1585                 unlock_page(page);
1586                 goto out;
1587         }
1588         err = filler(data, page);
1589         if (err < 0) {
1590                 page_cache_release(page);
1591                 return ERR_PTR(err);
1592         }
1593 out:
1594         mark_page_accessed(page);
1595         return page;
1596 }
1597 EXPORT_SYMBOL(read_cache_page_async);
1598
1599 /**
1600  * read_cache_page - read into page cache, fill it if needed
1601  * @mapping:    the page's address_space
1602  * @index:      the page index
1603  * @filler:     function to perform the read
1604  * @data:       destination for read data
1605  *
1606  * Read into the page cache. If a page already exists, and PageUptodate() is
1607  * not set, try to fill the page then wait for it to become unlocked.
1608  *
1609  * If the page does not get brought uptodate, return -EIO.
1610  */
1611 struct page *read_cache_page(struct address_space *mapping,
1612                                 pgoff_t index,
1613                                 int (*filler)(void *,struct page*),
1614                                 void *data)
1615 {
1616         struct page *page;
1617
1618         page = read_cache_page_async(mapping, index, filler, data);
1619         if (IS_ERR(page))
1620                 goto out;
1621         wait_on_page_locked(page);
1622         if (!PageUptodate(page)) {
1623                 page_cache_release(page);
1624                 page = ERR_PTR(-EIO);
1625         }
1626  out:
1627         return page;
1628 }
1629 EXPORT_SYMBOL(read_cache_page);
1630
1631 /*
1632  * The logic we want is
1633  *
1634  *      if suid or (sgid and xgrp)
1635  *              remove privs
1636  */
1637 int should_remove_suid(struct dentry *dentry)
1638 {
1639         mode_t mode = dentry->d_inode->i_mode;
1640         int kill = 0;
1641
1642         /* suid always must be killed */
1643         if (unlikely(mode & S_ISUID))
1644                 kill = ATTR_KILL_SUID;
1645
1646         /*
1647          * sgid without any exec bits is just a mandatory locking mark; leave
1648          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1649          */
1650         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1651                 kill |= ATTR_KILL_SGID;
1652
1653         if (unlikely(kill && !capable(CAP_FSETID)))
1654                 return kill;
1655
1656         return 0;
1657 }
1658 EXPORT_SYMBOL(should_remove_suid);
1659
1660 static int __remove_suid(struct dentry *dentry, int kill)
1661 {
1662         struct iattr newattrs;
1663
1664         newattrs.ia_valid = ATTR_FORCE | kill;
1665         return notify_change(dentry, &newattrs);
1666 }
1667
1668 int remove_suid(struct dentry *dentry)
1669 {
1670         int killsuid = should_remove_suid(dentry);
1671         int killpriv = security_inode_need_killpriv(dentry);
1672         int error = 0;
1673
1674         if (killpriv < 0)
1675                 return killpriv;
1676         if (killpriv)
1677                 error = security_inode_killpriv(dentry);
1678         if (!error && killsuid)
1679                 error = __remove_suid(dentry, killsuid);
1680
1681         return error;
1682 }
1683 EXPORT_SYMBOL(remove_suid);
1684
1685 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1686                         const struct iovec *iov, size_t base, size_t bytes)
1687 {
1688         size_t copied = 0, left = 0;
1689
1690         while (bytes) {
1691                 char __user *buf = iov->iov_base + base;
1692                 int copy = min(bytes, iov->iov_len - base);
1693
1694                 base = 0;
1695                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1696                 copied += copy;
1697                 bytes -= copy;
1698                 vaddr += copy;
1699                 iov++;
1700
1701                 if (unlikely(left))
1702                         break;
1703         }
1704         return copied - left;
1705 }
1706
1707 /*
1708  * Copy as much as we can into the page and return the number of bytes which
1709  * were sucessfully copied.  If a fault is encountered then return the number of
1710  * bytes which were copied.
1711  */
1712 size_t iov_iter_copy_from_user_atomic(struct page *page,
1713                 struct iov_iter *i, unsigned long offset, size_t bytes)
1714 {
1715         char *kaddr;
1716         size_t copied;
1717
1718         BUG_ON(!in_atomic());
1719         kaddr = kmap_atomic(page, KM_USER0);
1720         if (likely(i->nr_segs == 1)) {
1721                 int left;
1722                 char __user *buf = i->iov->iov_base + i->iov_offset;
1723                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1724                                                         buf, bytes);
1725                 copied = bytes - left;
1726         } else {
1727                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1728                                                 i->iov, i->iov_offset, bytes);
1729         }
1730         kunmap_atomic(kaddr, KM_USER0);
1731
1732         return copied;
1733 }
1734 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1735
1736 /*
1737  * This has the same sideeffects and return value as
1738  * iov_iter_copy_from_user_atomic().
1739  * The difference is that it attempts to resolve faults.
1740  * Page must not be locked.
1741  */
1742 size_t iov_iter_copy_from_user(struct page *page,
1743                 struct iov_iter *i, unsigned long offset, size_t bytes)
1744 {
1745         char *kaddr;
1746         size_t copied;
1747
1748         kaddr = kmap(page);
1749         if (likely(i->nr_segs == 1)) {
1750                 int left;
1751                 char __user *buf = i->iov->iov_base + i->iov_offset;
1752                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1753                 copied = bytes - left;
1754         } else {
1755                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1756                                                 i->iov, i->iov_offset, bytes);
1757         }
1758         kunmap(page);
1759         return copied;
1760 }
1761 EXPORT_SYMBOL(iov_iter_copy_from_user);
1762
1763 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1764 {
1765         BUG_ON(i->count < bytes);
1766
1767         if (likely(i->nr_segs == 1)) {
1768                 i->iov_offset += bytes;
1769                 i->count -= bytes;
1770         } else {
1771                 const struct iovec *iov = i->iov;
1772                 size_t base = i->iov_offset;
1773
1774                 /*
1775                  * The !iov->iov_len check ensures we skip over unlikely
1776                  * zero-length segments (without overruning the iovec).
1777                  */
1778                 while (bytes || unlikely(!iov->iov_len && i->count)) {
1779                         int copy;
1780
1781                         copy = min(bytes, iov->iov_len - base);
1782                         BUG_ON(!i->count || i->count < copy);
1783                         i->count -= copy;
1784                         bytes -= copy;
1785                         base += copy;
1786                         if (iov->iov_len == base) {
1787                                 iov++;
1788                                 base = 0;
1789                         }
1790                 }
1791                 i->iov = iov;
1792                 i->iov_offset = base;
1793         }
1794 }
1795 EXPORT_SYMBOL(iov_iter_advance);
1796
1797 /*
1798  * Fault in the first iovec of the given iov_iter, to a maximum length
1799  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1800  * accessed (ie. because it is an invalid address).
1801  *
1802  * writev-intensive code may want this to prefault several iovecs -- that
1803  * would be possible (callers must not rely on the fact that _only_ the
1804  * first iovec will be faulted with the current implementation).
1805  */
1806 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1807 {
1808         char __user *buf = i->iov->iov_base + i->iov_offset;
1809         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1810         return fault_in_pages_readable(buf, bytes);
1811 }
1812 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1813
1814 /*
1815  * Return the count of just the current iov_iter segment.
1816  */
1817 size_t iov_iter_single_seg_count(struct iov_iter *i)
1818 {
1819         const struct iovec *iov = i->iov;
1820         if (i->nr_segs == 1)
1821                 return i->count;
1822         else
1823                 return min(i->count, iov->iov_len - i->iov_offset);
1824 }
1825 EXPORT_SYMBOL(iov_iter_single_seg_count);
1826
1827 /*
1828  * Performs necessary checks before doing a write
1829  *
1830  * Can adjust writing position or amount of bytes to write.
1831  * Returns appropriate error code that caller should return or
1832  * zero in case that write should be allowed.
1833  */
1834 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1835 {
1836         struct inode *inode = file->f_mapping->host;
1837         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1838
1839         if (unlikely(*pos < 0))
1840                 return -EINVAL;
1841
1842         if (!isblk) {
1843                 /* FIXME: this is for backwards compatibility with 2.4 */
1844                 if (file->f_flags & O_APPEND)
1845                         *pos = i_size_read(inode);
1846
1847                 if (limit != RLIM_INFINITY) {
1848                         if (*pos >= limit) {
1849                                 send_sig(SIGXFSZ, current, 0);
1850                                 return -EFBIG;
1851                         }
1852                         if (*count > limit - (typeof(limit))*pos) {
1853                                 *count = limit - (typeof(limit))*pos;
1854                         }
1855                 }
1856         }
1857
1858         /*
1859          * LFS rule
1860          */
1861         if (unlikely(*pos + *count > MAX_NON_LFS &&
1862                                 !(file->f_flags & O_LARGEFILE))) {
1863                 if (*pos >= MAX_NON_LFS) {
1864                         return -EFBIG;
1865                 }
1866                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1867                         *count = MAX_NON_LFS - (unsigned long)*pos;
1868                 }
1869         }
1870
1871         /*
1872          * Are we about to exceed the fs block limit ?
1873          *
1874          * If we have written data it becomes a short write.  If we have
1875          * exceeded without writing data we send a signal and return EFBIG.
1876          * Linus frestrict idea will clean these up nicely..
1877          */
1878         if (likely(!isblk)) {
1879                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1880                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1881                                 return -EFBIG;
1882                         }
1883                         /* zero-length writes at ->s_maxbytes are OK */
1884                 }
1885
1886                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1887                         *count = inode->i_sb->s_maxbytes - *pos;
1888         } else {
1889 #ifdef CONFIG_BLOCK
1890                 loff_t isize;
1891                 if (bdev_read_only(I_BDEV(inode)))
1892                         return -EPERM;
1893                 isize = i_size_read(inode);
1894                 if (*pos >= isize) {
1895                         if (*count || *pos > isize)
1896                                 return -ENOSPC;
1897                 }
1898
1899                 if (*pos + *count > isize)
1900                         *count = isize - *pos;
1901 #else
1902                 return -EPERM;
1903 #endif
1904         }
1905         return 0;
1906 }
1907 EXPORT_SYMBOL(generic_write_checks);
1908
1909 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1910                                 loff_t pos, unsigned len, unsigned flags,
1911                                 struct page **pagep, void **fsdata)
1912 {
1913         const struct address_space_operations *aops = mapping->a_ops;
1914
1915         if (aops->write_begin) {
1916                 return aops->write_begin(file, mapping, pos, len, flags,
1917                                                         pagep, fsdata);
1918         } else {
1919                 int ret;
1920                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1921                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1922                 struct inode *inode = mapping->host;
1923                 struct page *page;
1924 again:
1925                 page = __grab_cache_page(mapping, index);
1926                 *pagep = page;
1927                 if (!page)
1928                         return -ENOMEM;
1929
1930                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1931                         /*
1932                          * There is no way to resolve a short write situation
1933                          * for a !Uptodate page (except by double copying in
1934                          * the caller done by generic_perform_write_2copy).
1935                          *
1936                          * Instead, we have to bring it uptodate here.
1937                          */
1938                         ret = aops->readpage(file, page);
1939                         page_cache_release(page);
1940                         if (ret) {
1941                                 if (ret == AOP_TRUNCATED_PAGE)
1942                                         goto again;
1943                                 return ret;
1944                         }
1945                         goto again;
1946                 }
1947
1948                 ret = aops->prepare_write(file, page, offset, offset+len);
1949                 if (ret) {
1950                         unlock_page(page);
1951                         page_cache_release(page);
1952                         if (pos + len > inode->i_size)
1953                                 vmtruncate(inode, inode->i_size);
1954                 }
1955                 return ret;
1956         }
1957 }
1958 EXPORT_SYMBOL(pagecache_write_begin);
1959
1960 int pagecache_write_end(struct file *file, struct address_space *mapping,
1961                                 loff_t pos, unsigned len, unsigned copied,
1962                                 struct page *page, void *fsdata)
1963 {
1964         const struct address_space_operations *aops = mapping->a_ops;
1965         int ret;
1966
1967         if (aops->write_end) {
1968                 mark_page_accessed(page);
1969                 ret = aops->write_end(file, mapping, pos, len, copied,
1970                                                         page, fsdata);
1971         } else {
1972                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1973                 struct inode *inode = mapping->host;
1974
1975                 flush_dcache_page(page);
1976                 ret = aops->commit_write(file, page, offset, offset+len);
1977                 unlock_page(page);
1978                 mark_page_accessed(page);
1979                 page_cache_release(page);
1980
1981                 if (ret < 0) {
1982                         if (pos + len > inode->i_size)
1983                                 vmtruncate(inode, inode->i_size);
1984                 } else if (ret > 0)
1985                         ret = min_t(size_t, copied, ret);
1986                 else
1987                         ret = copied;
1988         }
1989
1990         return ret;
1991 }
1992 EXPORT_SYMBOL(pagecache_write_end);
1993
1994 ssize_t
1995 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1996                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1997                 size_t count, size_t ocount)
1998 {
1999         struct file     *file = iocb->ki_filp;
2000         struct address_space *mapping = file->f_mapping;
2001         struct inode    *inode = mapping->host;
2002         ssize_t         written;
2003         size_t          write_len;
2004         pgoff_t         end;
2005
2006         if (count != ocount)
2007                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2008
2009         /*
2010          * Unmap all mmappings of the file up-front.
2011          *
2012          * This will cause any pte dirty bits to be propagated into the
2013          * pageframes for the subsequent filemap_write_and_wait().
2014          */
2015         write_len = iov_length(iov, *nr_segs);
2016         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2017         if (mapping_mapped(mapping))
2018                 unmap_mapping_range(mapping, pos, write_len, 0);
2019
2020         written = filemap_write_and_wait(mapping);
2021         if (written)
2022                 goto out;
2023
2024         /*
2025          * After a write we want buffered reads to be sure to go to disk to get
2026          * the new data.  We invalidate clean cached page from the region we're
2027          * about to write.  We do this *before* the write so that we can return
2028          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2029          */
2030         if (mapping->nrpages) {
2031                 written = invalidate_inode_pages2_range(mapping,
2032                                         pos >> PAGE_CACHE_SHIFT, end);
2033                 if (written)
2034                         goto out;
2035         }
2036
2037         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2038
2039         /*
2040          * Finally, try again to invalidate clean pages which might have been
2041          * cached by non-direct readahead, or faulted in by get_user_pages()
2042          * if the source of the write was an mmap'ed region of the file
2043          * we're writing.  Either one is a pretty crazy thing to do,
2044          * so we don't support it 100%.  If this invalidation
2045          * fails, tough, the write still worked...
2046          */
2047         if (mapping->nrpages) {
2048                 invalidate_inode_pages2_range(mapping,
2049                                               pos >> PAGE_CACHE_SHIFT, end);
2050         }
2051
2052         if (written > 0) {
2053                 loff_t end = pos + written;
2054                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2055                         i_size_write(inode,  end);
2056                         mark_inode_dirty(inode);
2057                 }
2058                 *ppos = end;
2059         }
2060
2061         /*
2062          * Sync the fs metadata but not the minor inode changes and
2063          * of course not the data as we did direct DMA for the IO.
2064          * i_mutex is held, which protects generic_osync_inode() from
2065          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2066          */
2067 out:
2068         if ((written >= 0 || written == -EIOCBQUEUED) &&
2069             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2070                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2071                 if (err < 0)
2072                         written = err;
2073         }
2074         return written;
2075 }
2076 EXPORT_SYMBOL(generic_file_direct_write);
2077
2078 /*
2079  * Find or create a page at the given pagecache position. Return the locked
2080  * page. This function is specifically for buffered writes.
2081  */
2082 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2083 {
2084         int status;
2085         struct page *page;
2086 repeat:
2087         page = find_lock_page(mapping, index);
2088         if (likely(page))
2089                 return page;
2090
2091         page = page_cache_alloc(mapping);
2092         if (!page)
2093                 return NULL;
2094         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2095         if (unlikely(status)) {
2096                 page_cache_release(page);
2097                 if (status == -EEXIST)
2098                         goto repeat;
2099                 return NULL;
2100         }
2101         return page;
2102 }
2103 EXPORT_SYMBOL(__grab_cache_page);
2104
2105 static ssize_t generic_perform_write_2copy(struct file *file,
2106                                 struct iov_iter *i, loff_t pos)
2107 {
2108         struct address_space *mapping = file->f_mapping;
2109         const struct address_space_operations *a_ops = mapping->a_ops;
2110         struct inode *inode = mapping->host;
2111         long status = 0;
2112         ssize_t written = 0;
2113
2114         do {
2115                 struct page *src_page;
2116                 struct page *page;
2117                 pgoff_t index;          /* Pagecache index for current page */
2118                 unsigned long offset;   /* Offset into pagecache page */
2119                 unsigned long bytes;    /* Bytes to write to page */
2120                 size_t copied;          /* Bytes copied from user */
2121
2122                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2123                 index = pos >> PAGE_CACHE_SHIFT;
2124                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2125                                                 iov_iter_count(i));
2126
2127                 /*
2128                  * a non-NULL src_page indicates that we're doing the
2129                  * copy via get_user_pages and kmap.
2130                  */
2131                 src_page = NULL;
2132
2133                 /*
2134                  * Bring in the user page that we will copy from _first_.
2135                  * Otherwise there's a nasty deadlock on copying from the
2136                  * same page as we're writing to, without it being marked
2137                  * up-to-date.
2138                  *
2139                  * Not only is this an optimisation, but it is also required
2140                  * to check that the address is actually valid, when atomic
2141                  * usercopies are used, below.
2142                  */
2143                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2144                         status = -EFAULT;
2145                         break;
2146                 }
2147
2148                 page = __grab_cache_page(mapping, index);
2149                 if (!page) {
2150                         status = -ENOMEM;
2151                         break;
2152                 }
2153
2154                 /*
2155                  * non-uptodate pages cannot cope with short copies, and we
2156                  * cannot take a pagefault with the destination page locked.
2157                  * So pin the source page to copy it.
2158                  */
2159                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2160                         unlock_page(page);
2161
2162                         src_page = alloc_page(GFP_KERNEL);
2163                         if (!src_page) {
2164                                 page_cache_release(page);
2165                                 status = -ENOMEM;
2166                                 break;
2167                         }
2168
2169                         /*
2170                          * Cannot get_user_pages with a page locked for the
2171                          * same reason as we can't take a page fault with a
2172                          * page locked (as explained below).
2173                          */
2174                         copied = iov_iter_copy_from_user(src_page, i,
2175                                                                 offset, bytes);
2176                         if (unlikely(copied == 0)) {
2177                                 status = -EFAULT;
2178                                 page_cache_release(page);
2179                                 page_cache_release(src_page);
2180                                 break;
2181                         }
2182                         bytes = copied;
2183
2184                         lock_page(page);
2185                         /*
2186                          * Can't handle the page going uptodate here, because
2187                          * that means we would use non-atomic usercopies, which
2188                          * zero out the tail of the page, which can cause
2189                          * zeroes to become transiently visible. We could just
2190                          * use a non-zeroing copy, but the APIs aren't too
2191                          * consistent.
2192                          */
2193                         if (unlikely(!page->mapping || PageUptodate(page))) {
2194                                 unlock_page(page);
2195                                 page_cache_release(page);
2196                                 page_cache_release(src_page);
2197                                 continue;
2198                         }
2199                 }
2200
2201                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2202                 if (unlikely(status))
2203                         goto fs_write_aop_error;
2204
2205                 if (!src_page) {
2206                         /*
2207                          * Must not enter the pagefault handler here, because
2208                          * we hold the page lock, so we might recursively
2209                          * deadlock on the same lock, or get an ABBA deadlock
2210                          * against a different lock, or against the mmap_sem
2211                          * (which nests outside the page lock).  So increment
2212                          * preempt count, and use _atomic usercopies.
2213                          *
2214                          * The page is uptodate so we are OK to encounter a
2215                          * short copy: if unmodified parts of the page are
2216                          * marked dirty and written out to disk, it doesn't
2217                          * really matter.
2218                          */
2219                         pagefault_disable();
2220                         copied = iov_iter_copy_from_user_atomic(page, i,
2221                                                                 offset, bytes);
2222                         pagefault_enable();
2223                 } else {
2224                         void *src, *dst;
2225                         src = kmap_atomic(src_page, KM_USER0);
2226                         dst = kmap_atomic(page, KM_USER1);
2227                         memcpy(dst + offset, src + offset, bytes);
2228                         kunmap_atomic(dst, KM_USER1);
2229                         kunmap_atomic(src, KM_USER0);
2230                         copied = bytes;
2231                 }
2232                 flush_dcache_page(page);
2233
2234                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2235                 if (unlikely(status < 0))
2236                         goto fs_write_aop_error;
2237                 if (unlikely(status > 0)) /* filesystem did partial write */
2238                         copied = min_t(size_t, copied, status);
2239
2240                 unlock_page(page);
2241                 mark_page_accessed(page);
2242                 page_cache_release(page);
2243                 if (src_page)
2244                         page_cache_release(src_page);
2245
2246                 iov_iter_advance(i, copied);
2247                 pos += copied;
2248                 written += copied;
2249
2250                 balance_dirty_pages_ratelimited(mapping);
2251                 cond_resched();
2252                 continue;
2253
2254 fs_write_aop_error:
2255                 unlock_page(page);
2256                 page_cache_release(page);
2257                 if (src_page)
2258                         page_cache_release(src_page);
2259
2260                 /*
2261                  * prepare_write() may have instantiated a few blocks
2262                  * outside i_size.  Trim these off again. Don't need
2263                  * i_size_read because we hold i_mutex.
2264                  */
2265                 if (pos + bytes > inode->i_size)
2266                         vmtruncate(inode, inode->i_size);
2267                 break;
2268         } while (iov_iter_count(i));
2269
2270         return written ? written : status;
2271 }
2272
2273 static ssize_t generic_perform_write(struct file *file,
2274                                 struct iov_iter *i, loff_t pos)
2275 {
2276         struct address_space *mapping = file->f_mapping;
2277         const struct address_space_operations *a_ops = mapping->a_ops;
2278         long status = 0;
2279         ssize_t written = 0;
2280         unsigned int flags = 0;
2281
2282         /*
2283          * Copies from kernel address space cannot fail (NFSD is a big user).
2284          */
2285         if (segment_eq(get_fs(), KERNEL_DS))
2286                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2287
2288         do {
2289                 struct page *page;
2290                 pgoff_t index;          /* Pagecache index for current page */
2291                 unsigned long offset;   /* Offset into pagecache page */
2292                 unsigned long bytes;    /* Bytes to write to page */
2293                 size_t copied;          /* Bytes copied from user */
2294                 void *fsdata;
2295
2296                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2297                 index = pos >> PAGE_CACHE_SHIFT;
2298                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2299                                                 iov_iter_count(i));
2300
2301 again:
2302
2303                 /*
2304                  * Bring in the user page that we will copy from _first_.
2305                  * Otherwise there's a nasty deadlock on copying from the
2306                  * same page as we're writing to, without it being marked
2307                  * up-to-date.
2308                  *
2309                  * Not only is this an optimisation, but it is also required
2310                  * to check that the address is actually valid, when atomic
2311                  * usercopies are used, below.
2312                  */
2313                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2314                         status = -EFAULT;
2315                         break;
2316                 }
2317
2318                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2319                                                 &page, &fsdata);
2320                 if (unlikely(status))
2321                         break;
2322
2323                 pagefault_disable();
2324                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2325                 pagefault_enable();
2326                 flush_dcache_page(page);
2327
2328                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2329                                                 page, fsdata);
2330                 if (unlikely(status < 0))
2331                         break;
2332                 copied = status;
2333
2334                 cond_resched();
2335
2336                 iov_iter_advance(i, copied);
2337                 if (unlikely(copied == 0)) {
2338                         /*
2339                          * If we were unable to copy any data at all, we must
2340                          * fall back to a single segment length write.
2341                          *
2342                          * If we didn't fallback here, we could livelock
2343                          * because not all segments in the iov can be copied at
2344                          * once without a pagefault.
2345                          */
2346                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2347                                                 iov_iter_single_seg_count(i));
2348                         goto again;
2349                 }
2350                 pos += copied;
2351                 written += copied;
2352
2353                 balance_dirty_pages_ratelimited(mapping);
2354
2355         } while (iov_iter_count(i));
2356
2357         return written ? written : status;
2358 }
2359
2360 ssize_t
2361 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2362                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2363                 size_t count, ssize_t written)
2364 {
2365         struct file *file = iocb->ki_filp;
2366         struct address_space *mapping = file->f_mapping;
2367         const struct address_space_operations *a_ops = mapping->a_ops;
2368         struct inode *inode = mapping->host;
2369         ssize_t status;
2370         struct iov_iter i;
2371
2372         iov_iter_init(&i, iov, nr_segs, count, written);
2373         if (a_ops->write_begin)
2374                 status = generic_perform_write(file, &i, pos);
2375         else
2376                 status = generic_perform_write_2copy(file, &i, pos);
2377
2378         if (likely(status >= 0)) {
2379                 written += status;
2380                 *ppos = pos + status;
2381
2382                 /*
2383                  * For now, when the user asks for O_SYNC, we'll actually give
2384                  * O_DSYNC
2385                  */
2386                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2387                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2388                                 status = generic_osync_inode(inode, mapping,
2389                                                 OSYNC_METADATA|OSYNC_DATA);
2390                 }
2391         }
2392         
2393         /*
2394          * If we get here for O_DIRECT writes then we must have fallen through
2395          * to buffered writes (block instantiation inside i_size).  So we sync
2396          * the file data here, to try to honour O_DIRECT expectations.
2397          */
2398         if (unlikely(file->f_flags & O_DIRECT) && written)
2399                 status = filemap_write_and_wait(mapping);
2400
2401         return written ? written : status;
2402 }
2403 EXPORT_SYMBOL(generic_file_buffered_write);
2404
2405 static ssize_t
2406 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2407                                 unsigned long nr_segs, loff_t *ppos)
2408 {
2409         struct file *file = iocb->ki_filp;
2410         struct address_space * mapping = file->f_mapping;
2411         size_t ocount;          /* original count */
2412         size_t count;           /* after file limit checks */
2413         struct inode    *inode = mapping->host;
2414         loff_t          pos;
2415         ssize_t         written;
2416         ssize_t         err;
2417
2418         ocount = 0;
2419         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2420         if (err)
2421                 return err;
2422
2423         count = ocount;
2424         pos = *ppos;
2425
2426         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2427
2428         /* We can write back this queue in page reclaim */
2429         current->backing_dev_info = mapping->backing_dev_info;
2430         written = 0;
2431
2432         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2433         if (err)
2434                 goto out;
2435
2436         if (count == 0)
2437                 goto out;
2438
2439         err = remove_suid(file->f_path.dentry);
2440         if (err)
2441                 goto out;
2442
2443         file_update_time(file);
2444
2445         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2446         if (unlikely(file->f_flags & O_DIRECT)) {
2447                 loff_t endbyte;
2448                 ssize_t written_buffered;
2449
2450                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2451                                                         ppos, count, ocount);
2452                 if (written < 0 || written == count)
2453                         goto out;
2454                 /*
2455                  * direct-io write to a hole: fall through to buffered I/O
2456                  * for completing the rest of the request.
2457                  */
2458                 pos += written;
2459                 count -= written;
2460                 written_buffered = generic_file_buffered_write(iocb, iov,
2461                                                 nr_segs, pos, ppos, count,
2462                                                 written);
2463                 /*
2464                  * If generic_file_buffered_write() retuned a synchronous error
2465                  * then we want to return the number of bytes which were
2466                  * direct-written, or the error code if that was zero.  Note
2467                  * that this differs from normal direct-io semantics, which
2468                  * will return -EFOO even if some bytes were written.
2469                  */
2470                 if (written_buffered < 0) {
2471                         err = written_buffered;
2472                         goto out;
2473                 }
2474
2475                 /*
2476                  * We need to ensure that the page cache pages are written to
2477                  * disk and invalidated to preserve the expected O_DIRECT
2478                  * semantics.
2479                  */
2480                 endbyte = pos + written_buffered - written - 1;
2481                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2482                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2483                                             SYNC_FILE_RANGE_WRITE|
2484                                             SYNC_FILE_RANGE_WAIT_AFTER);
2485                 if (err == 0) {
2486                         written = written_buffered;
2487                         invalidate_mapping_pages(mapping,
2488                                                  pos >> PAGE_CACHE_SHIFT,
2489                                                  endbyte >> PAGE_CACHE_SHIFT);
2490                 } else {
2491                         /*
2492                          * We don't know how much we wrote, so just return
2493                          * the number of bytes which were direct-written
2494                          */
2495                 }
2496         } else {
2497                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2498                                 pos, ppos, count, written);
2499         }
2500 out:
2501         current->backing_dev_info = NULL;
2502         return written ? written : err;
2503 }
2504
2505 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2506                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2507 {
2508         struct file *file = iocb->ki_filp;
2509         struct address_space *mapping = file->f_mapping;
2510         struct inode *inode = mapping->host;
2511         ssize_t ret;
2512
2513         BUG_ON(iocb->ki_pos != pos);
2514
2515         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2516                         &iocb->ki_pos);
2517
2518         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2519                 ssize_t err;
2520
2521                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2522                 if (err < 0)
2523                         ret = err;
2524         }
2525         return ret;
2526 }
2527 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2528
2529 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2530                 unsigned long nr_segs, loff_t pos)
2531 {
2532         struct file *file = iocb->ki_filp;
2533         struct address_space *mapping = file->f_mapping;
2534         struct inode *inode = mapping->host;
2535         ssize_t ret;
2536
2537         BUG_ON(iocb->ki_pos != pos);
2538
2539         mutex_lock(&inode->i_mutex);
2540         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2541                         &iocb->ki_pos);
2542         mutex_unlock(&inode->i_mutex);
2543
2544         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2545                 ssize_t err;
2546
2547                 err = sync_page_range(inode, mapping, pos, ret);
2548                 if (err < 0)
2549                         ret = err;
2550         }
2551         return ret;
2552 }
2553 EXPORT_SYMBOL(generic_file_aio_write);
2554
2555 /**
2556  * try_to_release_page() - release old fs-specific metadata on a page
2557  *
2558  * @page: the page which the kernel is trying to free
2559  * @gfp_mask: memory allocation flags (and I/O mode)
2560  *
2561  * The address_space is to try to release any data against the page
2562  * (presumably at page->private).  If the release was successful, return `1'.
2563  * Otherwise return zero.
2564  *
2565  * The @gfp_mask argument specifies whether I/O may be performed to release
2566  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2567  *
2568  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2569  */
2570 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2571 {
2572         struct address_space * const mapping = page->mapping;
2573
2574         BUG_ON(!PageLocked(page));
2575         if (PageWriteback(page))
2576                 return 0;
2577
2578         if (mapping && mapping->a_ops->releasepage)
2579                 return mapping->a_ops->releasepage(page, gfp_mask);
2580         return try_to_free_buffers(page);
2581 }
2582
2583 EXPORT_SYMBOL(try_to_release_page);