Use pgoff_t instead of unsigned long
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/backing-dev.h>
32 #include <linux/security.h>
33 #include <linux/syscalls.h>
34 #include <linux/cpuset.h>
35 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
36 #include <linux/memcontrol.h>
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for generic_osync_inode */
43
44 #include <asm/mman.h>
45
46 static ssize_t
47 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
48         loff_t offset, unsigned long nr_segs);
49
50 /*
51  * Shared mappings implemented 30.11.1994. It's not fully working yet,
52  * though.
53  *
54  * Shared mappings now work. 15.8.1995  Bruno.
55  *
56  * finished 'unifying' the page and buffer cache and SMP-threaded the
57  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
58  *
59  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
60  */
61
62 /*
63  * Lock ordering:
64  *
65  *  ->i_mmap_lock               (vmtruncate)
66  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
67  *      ->swap_lock             (exclusive_swap_page, others)
68  *        ->mapping->tree_lock
69  *
70  *  ->i_mutex
71  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
72  *
73  *  ->mmap_sem
74  *    ->i_mmap_lock
75  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
76  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
77  *
78  *  ->mmap_sem
79  *    ->lock_page               (access_process_vm)
80  *
81  *  ->i_mutex                   (generic_file_buffered_write)
82  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
83  *
84  *  ->i_mutex
85  *    ->i_alloc_sem             (various)
86  *
87  *  ->inode_lock
88  *    ->sb_lock                 (fs/fs-writeback.c)
89  *    ->mapping->tree_lock      (__sync_single_inode)
90  *
91  *  ->i_mmap_lock
92  *    ->anon_vma.lock           (vma_adjust)
93  *
94  *  ->anon_vma.lock
95  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
96  *
97  *  ->page_table_lock or pte_lock
98  *    ->swap_lock               (try_to_unmap_one)
99  *    ->private_lock            (try_to_unmap_one)
100  *    ->tree_lock               (try_to_unmap_one)
101  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
102  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
103  *    ->private_lock            (page_remove_rmap->set_page_dirty)
104  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
106  *    ->inode_lock              (zap_pte_range->set_page_dirty)
107  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
108  *
109  *  ->task->proc_lock
110  *    ->dcache_lock             (proc_pid_lookup)
111  */
112
113 /*
114  * Remove a page from the page cache and free it. Caller has to make
115  * sure the page is locked and that nobody else uses it - or that usage
116  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         mem_cgroup_uncharge_page(page);
123         radix_tree_delete(&mapping->page_tree, page->index);
124         page->mapping = NULL;
125         mapping->nrpages--;
126         __dec_zone_page_state(page, NR_FILE_PAGES);
127         BUG_ON(page_mapped(page));
128
129         /*
130          * Some filesystems seem to re-dirty the page even after
131          * the VM has canceled the dirty bit (eg ext3 journaling).
132          *
133          * Fix it up by doing a final dirty accounting check after
134          * having removed the page entirely.
135          */
136         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
137                 dec_zone_page_state(page, NR_FILE_DIRTY);
138                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
139         }
140 }
141
142 void remove_from_page_cache(struct page *page)
143 {
144         struct address_space *mapping = page->mapping;
145
146         BUG_ON(!PageLocked(page));
147
148         write_lock_irq(&mapping->tree_lock);
149         __remove_from_page_cache(page);
150         write_unlock_irq(&mapping->tree_lock);
151 }
152
153 static int sync_page(void *word)
154 {
155         struct address_space *mapping;
156         struct page *page;
157
158         page = container_of((unsigned long *)word, struct page, flags);
159
160         /*
161          * page_mapping() is being called without PG_locked held.
162          * Some knowledge of the state and use of the page is used to
163          * reduce the requirements down to a memory barrier.
164          * The danger here is of a stale page_mapping() return value
165          * indicating a struct address_space different from the one it's
166          * associated with when it is associated with one.
167          * After smp_mb(), it's either the correct page_mapping() for
168          * the page, or an old page_mapping() and the page's own
169          * page_mapping() has gone NULL.
170          * The ->sync_page() address_space operation must tolerate
171          * page_mapping() going NULL. By an amazing coincidence,
172          * this comes about because none of the users of the page
173          * in the ->sync_page() methods make essential use of the
174          * page_mapping(), merely passing the page down to the backing
175          * device's unplug functions when it's non-NULL, which in turn
176          * ignore it for all cases but swap, where only page_private(page) is
177          * of interest. When page_mapping() does go NULL, the entire
178          * call stack gracefully ignores the page and returns.
179          * -- wli
180          */
181         smp_mb();
182         mapping = page_mapping(page);
183         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
184                 mapping->a_ops->sync_page(page);
185         io_schedule();
186         return 0;
187 }
188
189 static int sync_page_killable(void *word)
190 {
191         sync_page(word);
192         return fatal_signal_pending(current) ? -EINTR : 0;
193 }
194
195 /**
196  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
197  * @mapping:    address space structure to write
198  * @start:      offset in bytes where the range starts
199  * @end:        offset in bytes where the range ends (inclusive)
200  * @sync_mode:  enable synchronous operation
201  *
202  * Start writeback against all of a mapping's dirty pages that lie
203  * within the byte offsets <start, end> inclusive.
204  *
205  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
206  * opposed to a regular memory cleansing writeback.  The difference between
207  * these two operations is that if a dirty page/buffer is encountered, it must
208  * be waited upon, and not just skipped over.
209  */
210 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
211                                 loff_t end, int sync_mode)
212 {
213         int ret;
214         struct writeback_control wbc = {
215                 .sync_mode = sync_mode,
216                 .nr_to_write = mapping->nrpages * 2,
217                 .range_start = start,
218                 .range_end = end,
219         };
220
221         if (!mapping_cap_writeback_dirty(mapping))
222                 return 0;
223
224         ret = do_writepages(mapping, &wbc);
225         return ret;
226 }
227
228 static inline int __filemap_fdatawrite(struct address_space *mapping,
229         int sync_mode)
230 {
231         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
232 }
233
234 int filemap_fdatawrite(struct address_space *mapping)
235 {
236         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
237 }
238 EXPORT_SYMBOL(filemap_fdatawrite);
239
240 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
241                                 loff_t end)
242 {
243         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
244 }
245
246 /**
247  * filemap_flush - mostly a non-blocking flush
248  * @mapping:    target address_space
249  *
250  * This is a mostly non-blocking flush.  Not suitable for data-integrity
251  * purposes - I/O may not be started against all dirty pages.
252  */
253 int filemap_flush(struct address_space *mapping)
254 {
255         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
256 }
257 EXPORT_SYMBOL(filemap_flush);
258
259 /**
260  * wait_on_page_writeback_range - wait for writeback to complete
261  * @mapping:    target address_space
262  * @start:      beginning page index
263  * @end:        ending page index
264  *
265  * Wait for writeback to complete against pages indexed by start->end
266  * inclusive
267  */
268 int wait_on_page_writeback_range(struct address_space *mapping,
269                                 pgoff_t start, pgoff_t end)
270 {
271         struct pagevec pvec;
272         int nr_pages;
273         int ret = 0;
274         pgoff_t index;
275
276         if (end < start)
277                 return 0;
278
279         pagevec_init(&pvec, 0);
280         index = start;
281         while ((index <= end) &&
282                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
283                         PAGECACHE_TAG_WRITEBACK,
284                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
285                 unsigned i;
286
287                 for (i = 0; i < nr_pages; i++) {
288                         struct page *page = pvec.pages[i];
289
290                         /* until radix tree lookup accepts end_index */
291                         if (page->index > end)
292                                 continue;
293
294                         wait_on_page_writeback(page);
295                         if (PageError(page))
296                                 ret = -EIO;
297                 }
298                 pagevec_release(&pvec);
299                 cond_resched();
300         }
301
302         /* Check for outstanding write errors */
303         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
304                 ret = -ENOSPC;
305         if (test_and_clear_bit(AS_EIO, &mapping->flags))
306                 ret = -EIO;
307
308         return ret;
309 }
310
311 /**
312  * sync_page_range - write and wait on all pages in the passed range
313  * @inode:      target inode
314  * @mapping:    target address_space
315  * @pos:        beginning offset in pages to write
316  * @count:      number of bytes to write
317  *
318  * Write and wait upon all the pages in the passed range.  This is a "data
319  * integrity" operation.  It waits upon in-flight writeout before starting and
320  * waiting upon new writeout.  If there was an IO error, return it.
321  *
322  * We need to re-take i_mutex during the generic_osync_inode list walk because
323  * it is otherwise livelockable.
324  */
325 int sync_page_range(struct inode *inode, struct address_space *mapping,
326                         loff_t pos, loff_t count)
327 {
328         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
329         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
330         int ret;
331
332         if (!mapping_cap_writeback_dirty(mapping) || !count)
333                 return 0;
334         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
335         if (ret == 0) {
336                 mutex_lock(&inode->i_mutex);
337                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
338                 mutex_unlock(&inode->i_mutex);
339         }
340         if (ret == 0)
341                 ret = wait_on_page_writeback_range(mapping, start, end);
342         return ret;
343 }
344 EXPORT_SYMBOL(sync_page_range);
345
346 /**
347  * sync_page_range_nolock
348  * @inode:      target inode
349  * @mapping:    target address_space
350  * @pos:        beginning offset in pages to write
351  * @count:      number of bytes to write
352  *
353  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
354  * as it forces O_SYNC writers to different parts of the same file
355  * to be serialised right until io completion.
356  */
357 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
358                            loff_t pos, loff_t count)
359 {
360         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
361         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
362         int ret;
363
364         if (!mapping_cap_writeback_dirty(mapping) || !count)
365                 return 0;
366         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
367         if (ret == 0)
368                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
369         if (ret == 0)
370                 ret = wait_on_page_writeback_range(mapping, start, end);
371         return ret;
372 }
373 EXPORT_SYMBOL(sync_page_range_nolock);
374
375 /**
376  * filemap_fdatawait - wait for all under-writeback pages to complete
377  * @mapping: address space structure to wait for
378  *
379  * Walk the list of under-writeback pages of the given address space
380  * and wait for all of them.
381  */
382 int filemap_fdatawait(struct address_space *mapping)
383 {
384         loff_t i_size = i_size_read(mapping->host);
385
386         if (i_size == 0)
387                 return 0;
388
389         return wait_on_page_writeback_range(mapping, 0,
390                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
391 }
392 EXPORT_SYMBOL(filemap_fdatawait);
393
394 int filemap_write_and_wait(struct address_space *mapping)
395 {
396         int err = 0;
397
398         if (mapping->nrpages) {
399                 err = filemap_fdatawrite(mapping);
400                 /*
401                  * Even if the above returned error, the pages may be
402                  * written partially (e.g. -ENOSPC), so we wait for it.
403                  * But the -EIO is special case, it may indicate the worst
404                  * thing (e.g. bug) happened, so we avoid waiting for it.
405                  */
406                 if (err != -EIO) {
407                         int err2 = filemap_fdatawait(mapping);
408                         if (!err)
409                                 err = err2;
410                 }
411         }
412         return err;
413 }
414 EXPORT_SYMBOL(filemap_write_and_wait);
415
416 /**
417  * filemap_write_and_wait_range - write out & wait on a file range
418  * @mapping:    the address_space for the pages
419  * @lstart:     offset in bytes where the range starts
420  * @lend:       offset in bytes where the range ends (inclusive)
421  *
422  * Write out and wait upon file offsets lstart->lend, inclusive.
423  *
424  * Note that `lend' is inclusive (describes the last byte to be written) so
425  * that this function can be used to write to the very end-of-file (end = -1).
426  */
427 int filemap_write_and_wait_range(struct address_space *mapping,
428                                  loff_t lstart, loff_t lend)
429 {
430         int err = 0;
431
432         if (mapping->nrpages) {
433                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
434                                                  WB_SYNC_ALL);
435                 /* See comment of filemap_write_and_wait() */
436                 if (err != -EIO) {
437                         int err2 = wait_on_page_writeback_range(mapping,
438                                                 lstart >> PAGE_CACHE_SHIFT,
439                                                 lend >> PAGE_CACHE_SHIFT);
440                         if (!err)
441                                 err = err2;
442                 }
443         }
444         return err;
445 }
446
447 /**
448  * add_to_page_cache - add newly allocated pagecache pages
449  * @page:       page to add
450  * @mapping:    the page's address_space
451  * @offset:     page index
452  * @gfp_mask:   page allocation mode
453  *
454  * This function is used to add newly allocated pagecache pages;
455  * the page is new, so we can just run SetPageLocked() against it.
456  * The other page state flags were set by rmqueue().
457  *
458  * This function does not add the page to the LRU.  The caller must do that.
459  */
460 int add_to_page_cache(struct page *page, struct address_space *mapping,
461                 pgoff_t offset, gfp_t gfp_mask)
462 {
463         int error = mem_cgroup_cache_charge(page, current->mm,
464                                         gfp_mask & ~__GFP_HIGHMEM);
465         if (error)
466                 goto out;
467
468         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
469         if (error == 0) {
470                 write_lock_irq(&mapping->tree_lock);
471                 error = radix_tree_insert(&mapping->page_tree, offset, page);
472                 if (!error) {
473                         page_cache_get(page);
474                         SetPageLocked(page);
475                         page->mapping = mapping;
476                         page->index = offset;
477                         mapping->nrpages++;
478                         __inc_zone_page_state(page, NR_FILE_PAGES);
479                 } else
480                         mem_cgroup_uncharge_page(page);
481
482                 write_unlock_irq(&mapping->tree_lock);
483                 radix_tree_preload_end();
484         } else
485                 mem_cgroup_uncharge_page(page);
486 out:
487         return error;
488 }
489 EXPORT_SYMBOL(add_to_page_cache);
490
491 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
492                                 pgoff_t offset, gfp_t gfp_mask)
493 {
494         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
495         if (ret == 0)
496                 lru_cache_add(page);
497         return ret;
498 }
499
500 #ifdef CONFIG_NUMA
501 struct page *__page_cache_alloc(gfp_t gfp)
502 {
503         if (cpuset_do_page_mem_spread()) {
504                 int n = cpuset_mem_spread_node();
505                 return alloc_pages_node(n, gfp, 0);
506         }
507         return alloc_pages(gfp, 0);
508 }
509 EXPORT_SYMBOL(__page_cache_alloc);
510 #endif
511
512 static int __sleep_on_page_lock(void *word)
513 {
514         io_schedule();
515         return 0;
516 }
517
518 /*
519  * In order to wait for pages to become available there must be
520  * waitqueues associated with pages. By using a hash table of
521  * waitqueues where the bucket discipline is to maintain all
522  * waiters on the same queue and wake all when any of the pages
523  * become available, and for the woken contexts to check to be
524  * sure the appropriate page became available, this saves space
525  * at a cost of "thundering herd" phenomena during rare hash
526  * collisions.
527  */
528 static wait_queue_head_t *page_waitqueue(struct page *page)
529 {
530         const struct zone *zone = page_zone(page);
531
532         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
533 }
534
535 static inline void wake_up_page(struct page *page, int bit)
536 {
537         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
538 }
539
540 void wait_on_page_bit(struct page *page, int bit_nr)
541 {
542         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
543
544         if (test_bit(bit_nr, &page->flags))
545                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
546                                                         TASK_UNINTERRUPTIBLE);
547 }
548 EXPORT_SYMBOL(wait_on_page_bit);
549
550 /**
551  * unlock_page - unlock a locked page
552  * @page: the page
553  *
554  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
555  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
556  * mechananism between PageLocked pages and PageWriteback pages is shared.
557  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
558  *
559  * The first mb is necessary to safely close the critical section opened by the
560  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
561  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
562  * parallel wait_on_page_locked()).
563  */
564 void unlock_page(struct page *page)
565 {
566         smp_mb__before_clear_bit();
567         if (!TestClearPageLocked(page))
568                 BUG();
569         smp_mb__after_clear_bit(); 
570         wake_up_page(page, PG_locked);
571 }
572 EXPORT_SYMBOL(unlock_page);
573
574 /**
575  * end_page_writeback - end writeback against a page
576  * @page: the page
577  */
578 void end_page_writeback(struct page *page)
579 {
580         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
581                 if (!test_clear_page_writeback(page))
582                         BUG();
583         }
584         smp_mb__after_clear_bit();
585         wake_up_page(page, PG_writeback);
586 }
587 EXPORT_SYMBOL(end_page_writeback);
588
589 /**
590  * __lock_page - get a lock on the page, assuming we need to sleep to get it
591  * @page: the page to lock
592  *
593  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
594  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
595  * chances are that on the second loop, the block layer's plug list is empty,
596  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
597  */
598 void __lock_page(struct page *page)
599 {
600         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
601
602         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
603                                                         TASK_UNINTERRUPTIBLE);
604 }
605 EXPORT_SYMBOL(__lock_page);
606
607 int fastcall __lock_page_killable(struct page *page)
608 {
609         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
610
611         return __wait_on_bit_lock(page_waitqueue(page), &wait,
612                                         sync_page_killable, TASK_KILLABLE);
613 }
614
615 /*
616  * Variant of lock_page that does not require the caller to hold a reference
617  * on the page's mapping.
618  */
619 void __lock_page_nosync(struct page *page)
620 {
621         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
622         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
623                                                         TASK_UNINTERRUPTIBLE);
624 }
625
626 /**
627  * find_get_page - find and get a page reference
628  * @mapping: the address_space to search
629  * @offset: the page index
630  *
631  * Is there a pagecache struct page at the given (mapping, offset) tuple?
632  * If yes, increment its refcount and return it; if no, return NULL.
633  */
634 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
635 {
636         struct page *page;
637
638         read_lock_irq(&mapping->tree_lock);
639         page = radix_tree_lookup(&mapping->page_tree, offset);
640         if (page)
641                 page_cache_get(page);
642         read_unlock_irq(&mapping->tree_lock);
643         return page;
644 }
645 EXPORT_SYMBOL(find_get_page);
646
647 /**
648  * find_lock_page - locate, pin and lock a pagecache page
649  * @mapping: the address_space to search
650  * @offset: the page index
651  *
652  * Locates the desired pagecache page, locks it, increments its reference
653  * count and returns its address.
654  *
655  * Returns zero if the page was not present. find_lock_page() may sleep.
656  */
657 struct page *find_lock_page(struct address_space *mapping,
658                                 pgoff_t offset)
659 {
660         struct page *page;
661
662 repeat:
663         read_lock_irq(&mapping->tree_lock);
664         page = radix_tree_lookup(&mapping->page_tree, offset);
665         if (page) {
666                 page_cache_get(page);
667                 if (TestSetPageLocked(page)) {
668                         read_unlock_irq(&mapping->tree_lock);
669                         __lock_page(page);
670
671                         /* Has the page been truncated while we slept? */
672                         if (unlikely(page->mapping != mapping)) {
673                                 unlock_page(page);
674                                 page_cache_release(page);
675                                 goto repeat;
676                         }
677                         VM_BUG_ON(page->index != offset);
678                         goto out;
679                 }
680         }
681         read_unlock_irq(&mapping->tree_lock);
682 out:
683         return page;
684 }
685 EXPORT_SYMBOL(find_lock_page);
686
687 /**
688  * find_or_create_page - locate or add a pagecache page
689  * @mapping: the page's address_space
690  * @index: the page's index into the mapping
691  * @gfp_mask: page allocation mode
692  *
693  * Locates a page in the pagecache.  If the page is not present, a new page
694  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
695  * LRU list.  The returned page is locked and has its reference count
696  * incremented.
697  *
698  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
699  * allocation!
700  *
701  * find_or_create_page() returns the desired page's address, or zero on
702  * memory exhaustion.
703  */
704 struct page *find_or_create_page(struct address_space *mapping,
705                 pgoff_t index, gfp_t gfp_mask)
706 {
707         struct page *page;
708         int err;
709 repeat:
710         page = find_lock_page(mapping, index);
711         if (!page) {
712                 page = __page_cache_alloc(gfp_mask);
713                 if (!page)
714                         return NULL;
715                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
716                 if (unlikely(err)) {
717                         page_cache_release(page);
718                         page = NULL;
719                         if (err == -EEXIST)
720                                 goto repeat;
721                 }
722         }
723         return page;
724 }
725 EXPORT_SYMBOL(find_or_create_page);
726
727 /**
728  * find_get_pages - gang pagecache lookup
729  * @mapping:    The address_space to search
730  * @start:      The starting page index
731  * @nr_pages:   The maximum number of pages
732  * @pages:      Where the resulting pages are placed
733  *
734  * find_get_pages() will search for and return a group of up to
735  * @nr_pages pages in the mapping.  The pages are placed at @pages.
736  * find_get_pages() takes a reference against the returned pages.
737  *
738  * The search returns a group of mapping-contiguous pages with ascending
739  * indexes.  There may be holes in the indices due to not-present pages.
740  *
741  * find_get_pages() returns the number of pages which were found.
742  */
743 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
744                             unsigned int nr_pages, struct page **pages)
745 {
746         unsigned int i;
747         unsigned int ret;
748
749         read_lock_irq(&mapping->tree_lock);
750         ret = radix_tree_gang_lookup(&mapping->page_tree,
751                                 (void **)pages, start, nr_pages);
752         for (i = 0; i < ret; i++)
753                 page_cache_get(pages[i]);
754         read_unlock_irq(&mapping->tree_lock);
755         return ret;
756 }
757
758 /**
759  * find_get_pages_contig - gang contiguous pagecache lookup
760  * @mapping:    The address_space to search
761  * @index:      The starting page index
762  * @nr_pages:   The maximum number of pages
763  * @pages:      Where the resulting pages are placed
764  *
765  * find_get_pages_contig() works exactly like find_get_pages(), except
766  * that the returned number of pages are guaranteed to be contiguous.
767  *
768  * find_get_pages_contig() returns the number of pages which were found.
769  */
770 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
771                                unsigned int nr_pages, struct page **pages)
772 {
773         unsigned int i;
774         unsigned int ret;
775
776         read_lock_irq(&mapping->tree_lock);
777         ret = radix_tree_gang_lookup(&mapping->page_tree,
778                                 (void **)pages, index, nr_pages);
779         for (i = 0; i < ret; i++) {
780                 if (pages[i]->mapping == NULL || pages[i]->index != index)
781                         break;
782
783                 page_cache_get(pages[i]);
784                 index++;
785         }
786         read_unlock_irq(&mapping->tree_lock);
787         return i;
788 }
789 EXPORT_SYMBOL(find_get_pages_contig);
790
791 /**
792  * find_get_pages_tag - find and return pages that match @tag
793  * @mapping:    the address_space to search
794  * @index:      the starting page index
795  * @tag:        the tag index
796  * @nr_pages:   the maximum number of pages
797  * @pages:      where the resulting pages are placed
798  *
799  * Like find_get_pages, except we only return pages which are tagged with
800  * @tag.   We update @index to index the next page for the traversal.
801  */
802 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
803                         int tag, unsigned int nr_pages, struct page **pages)
804 {
805         unsigned int i;
806         unsigned int ret;
807
808         read_lock_irq(&mapping->tree_lock);
809         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
810                                 (void **)pages, *index, nr_pages, tag);
811         for (i = 0; i < ret; i++)
812                 page_cache_get(pages[i]);
813         if (ret)
814                 *index = pages[ret - 1]->index + 1;
815         read_unlock_irq(&mapping->tree_lock);
816         return ret;
817 }
818 EXPORT_SYMBOL(find_get_pages_tag);
819
820 /**
821  * grab_cache_page_nowait - returns locked page at given index in given cache
822  * @mapping: target address_space
823  * @index: the page index
824  *
825  * Same as grab_cache_page(), but do not wait if the page is unavailable.
826  * This is intended for speculative data generators, where the data can
827  * be regenerated if the page couldn't be grabbed.  This routine should
828  * be safe to call while holding the lock for another page.
829  *
830  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
831  * and deadlock against the caller's locked page.
832  */
833 struct page *
834 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
835 {
836         struct page *page = find_get_page(mapping, index);
837
838         if (page) {
839                 if (!TestSetPageLocked(page))
840                         return page;
841                 page_cache_release(page);
842                 return NULL;
843         }
844         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
845         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
846                 page_cache_release(page);
847                 page = NULL;
848         }
849         return page;
850 }
851 EXPORT_SYMBOL(grab_cache_page_nowait);
852
853 /*
854  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
855  * a _large_ part of the i/o request. Imagine the worst scenario:
856  *
857  *      ---R__________________________________________B__________
858  *         ^ reading here                             ^ bad block(assume 4k)
859  *
860  * read(R) => miss => readahead(R...B) => media error => frustrating retries
861  * => failing the whole request => read(R) => read(R+1) =>
862  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
863  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
864  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
865  *
866  * It is going insane. Fix it by quickly scaling down the readahead size.
867  */
868 static void shrink_readahead_size_eio(struct file *filp,
869                                         struct file_ra_state *ra)
870 {
871         if (!ra->ra_pages)
872                 return;
873
874         ra->ra_pages /= 4;
875 }
876
877 /**
878  * do_generic_mapping_read - generic file read routine
879  * @mapping:    address_space to be read
880  * @ra:         file's readahead state
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  * Note the struct file* is only passed for the use of readpage.
893  * It may be NULL.
894  */
895 void do_generic_mapping_read(struct address_space *mapping,
896                              struct file_ra_state *ra,
897                              struct file *filp,
898                              loff_t *ppos,
899                              read_descriptor_t *desc,
900                              read_actor_t actor)
901 {
902         struct inode *inode = mapping->host;
903         pgoff_t index;
904         pgoff_t last_index;
905         pgoff_t prev_index;
906         unsigned long offset;      /* offset into pagecache page */
907         unsigned int prev_offset;
908         int error;
909
910         index = *ppos >> PAGE_CACHE_SHIFT;
911         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
912         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
913         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
914         offset = *ppos & ~PAGE_CACHE_MASK;
915
916         for (;;) {
917                 struct page *page;
918                 pgoff_t end_index;
919                 loff_t isize;
920                 unsigned long nr, ret;
921
922                 cond_resched();
923 find_page:
924                 page = find_get_page(mapping, index);
925                 if (!page) {
926                         page_cache_sync_readahead(mapping,
927                                         ra, filp,
928                                         index, last_index - index);
929                         page = find_get_page(mapping, index);
930                         if (unlikely(page == NULL))
931                                 goto no_cached_page;
932                 }
933                 if (PageReadahead(page)) {
934                         page_cache_async_readahead(mapping,
935                                         ra, filp, page,
936                                         index, last_index - index);
937                 }
938                 if (!PageUptodate(page))
939                         goto page_not_up_to_date;
940 page_ok:
941                 /*
942                  * i_size must be checked after we know the page is Uptodate.
943                  *
944                  * Checking i_size after the check allows us to calculate
945                  * the correct value for "nr", which means the zero-filled
946                  * part of the page is not copied back to userspace (unless
947                  * another truncate extends the file - this is desired though).
948                  */
949
950                 isize = i_size_read(inode);
951                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
952                 if (unlikely(!isize || index > end_index)) {
953                         page_cache_release(page);
954                         goto out;
955                 }
956
957                 /* nr is the maximum number of bytes to copy from this page */
958                 nr = PAGE_CACHE_SIZE;
959                 if (index == end_index) {
960                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
961                         if (nr <= offset) {
962                                 page_cache_release(page);
963                                 goto out;
964                         }
965                 }
966                 nr = nr - offset;
967
968                 /* If users can be writing to this page using arbitrary
969                  * virtual addresses, take care about potential aliasing
970                  * before reading the page on the kernel side.
971                  */
972                 if (mapping_writably_mapped(mapping))
973                         flush_dcache_page(page);
974
975                 /*
976                  * When a sequential read accesses a page several times,
977                  * only mark it as accessed the first time.
978                  */
979                 if (prev_index != index || offset != prev_offset)
980                         mark_page_accessed(page);
981                 prev_index = index;
982
983                 /*
984                  * Ok, we have the page, and it's up-to-date, so
985                  * now we can copy it to user space...
986                  *
987                  * The actor routine returns how many bytes were actually used..
988                  * NOTE! This may not be the same as how much of a user buffer
989                  * we filled up (we may be padding etc), so we can only update
990                  * "pos" here (the actor routine has to update the user buffer
991                  * pointers and the remaining count).
992                  */
993                 ret = actor(desc, page, offset, nr);
994                 offset += ret;
995                 index += offset >> PAGE_CACHE_SHIFT;
996                 offset &= ~PAGE_CACHE_MASK;
997                 prev_offset = offset;
998
999                 page_cache_release(page);
1000                 if (ret == nr && desc->count)
1001                         continue;
1002                 goto out;
1003
1004 page_not_up_to_date:
1005                 /* Get exclusive access to the page ... */
1006                 if (lock_page_killable(page))
1007                         goto readpage_eio;
1008
1009                 /* Did it get truncated before we got the lock? */
1010                 if (!page->mapping) {
1011                         unlock_page(page);
1012                         page_cache_release(page);
1013                         continue;
1014                 }
1015
1016                 /* Did somebody else fill it already? */
1017                 if (PageUptodate(page)) {
1018                         unlock_page(page);
1019                         goto page_ok;
1020                 }
1021
1022 readpage:
1023                 /* Start the actual read. The read will unlock the page. */
1024                 error = mapping->a_ops->readpage(filp, page);
1025
1026                 if (unlikely(error)) {
1027                         if (error == AOP_TRUNCATED_PAGE) {
1028                                 page_cache_release(page);
1029                                 goto find_page;
1030                         }
1031                         goto readpage_error;
1032                 }
1033
1034                 if (!PageUptodate(page)) {
1035                         if (lock_page_killable(page))
1036                                 goto readpage_eio;
1037                         if (!PageUptodate(page)) {
1038                                 if (page->mapping == NULL) {
1039                                         /*
1040                                          * invalidate_inode_pages got it
1041                                          */
1042                                         unlock_page(page);
1043                                         page_cache_release(page);
1044                                         goto find_page;
1045                                 }
1046                                 unlock_page(page);
1047                                 shrink_readahead_size_eio(filp, ra);
1048                                 goto readpage_eio;
1049                         }
1050                         unlock_page(page);
1051                 }
1052
1053                 goto page_ok;
1054
1055 readpage_eio:
1056                 error = -EIO;
1057 readpage_error:
1058                 /* UHHUH! A synchronous read error occurred. Report it */
1059                 desc->error = error;
1060                 page_cache_release(page);
1061                 goto out;
1062
1063 no_cached_page:
1064                 /*
1065                  * Ok, it wasn't cached, so we need to create a new
1066                  * page..
1067                  */
1068                 page = page_cache_alloc_cold(mapping);
1069                 if (!page) {
1070                         desc->error = -ENOMEM;
1071                         goto out;
1072                 }
1073                 error = add_to_page_cache_lru(page, mapping,
1074                                                 index, GFP_KERNEL);
1075                 if (error) {
1076                         page_cache_release(page);
1077                         if (error == -EEXIST)
1078                                 goto find_page;
1079                         desc->error = error;
1080                         goto out;
1081                 }
1082                 goto readpage;
1083         }
1084
1085 out:
1086         ra->prev_pos = prev_index;
1087         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1088         ra->prev_pos |= prev_offset;
1089
1090         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1091         if (filp)
1092                 file_accessed(filp);
1093 }
1094 EXPORT_SYMBOL(do_generic_mapping_read);
1095
1096 int file_read_actor(read_descriptor_t *desc, struct page *page,
1097                         unsigned long offset, unsigned long size)
1098 {
1099         char *kaddr;
1100         unsigned long left, count = desc->count;
1101
1102         if (size > count)
1103                 size = count;
1104
1105         /*
1106          * Faults on the destination of a read are common, so do it before
1107          * taking the kmap.
1108          */
1109         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1110                 kaddr = kmap_atomic(page, KM_USER0);
1111                 left = __copy_to_user_inatomic(desc->arg.buf,
1112                                                 kaddr + offset, size);
1113                 kunmap_atomic(kaddr, KM_USER0);
1114                 if (left == 0)
1115                         goto success;
1116         }
1117
1118         /* Do it the slow way */
1119         kaddr = kmap(page);
1120         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1121         kunmap(page);
1122
1123         if (left) {
1124                 size -= left;
1125                 desc->error = -EFAULT;
1126         }
1127 success:
1128         desc->count = count - size;
1129         desc->written += size;
1130         desc->arg.buf += size;
1131         return size;
1132 }
1133
1134 /*
1135  * Performs necessary checks before doing a write
1136  * @iov:        io vector request
1137  * @nr_segs:    number of segments in the iovec
1138  * @count:      number of bytes to write
1139  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1140  *
1141  * Adjust number of segments and amount of bytes to write (nr_segs should be
1142  * properly initialized first). Returns appropriate error code that caller
1143  * should return or zero in case that write should be allowed.
1144  */
1145 int generic_segment_checks(const struct iovec *iov,
1146                         unsigned long *nr_segs, size_t *count, int access_flags)
1147 {
1148         unsigned long   seg;
1149         size_t cnt = 0;
1150         for (seg = 0; seg < *nr_segs; seg++) {
1151                 const struct iovec *iv = &iov[seg];
1152
1153                 /*
1154                  * If any segment has a negative length, or the cumulative
1155                  * length ever wraps negative then return -EINVAL.
1156                  */
1157                 cnt += iv->iov_len;
1158                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1159                         return -EINVAL;
1160                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1161                         continue;
1162                 if (seg == 0)
1163                         return -EFAULT;
1164                 *nr_segs = seg;
1165                 cnt -= iv->iov_len;     /* This segment is no good */
1166                 break;
1167         }
1168         *count = cnt;
1169         return 0;
1170 }
1171 EXPORT_SYMBOL(generic_segment_checks);
1172
1173 /**
1174  * generic_file_aio_read - generic filesystem read routine
1175  * @iocb:       kernel I/O control block
1176  * @iov:        io vector request
1177  * @nr_segs:    number of segments in the iovec
1178  * @pos:        current file position
1179  *
1180  * This is the "read()" routine for all filesystems
1181  * that can use the page cache directly.
1182  */
1183 ssize_t
1184 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1185                 unsigned long nr_segs, loff_t pos)
1186 {
1187         struct file *filp = iocb->ki_filp;
1188         ssize_t retval;
1189         unsigned long seg;
1190         size_t count;
1191         loff_t *ppos = &iocb->ki_pos;
1192
1193         count = 0;
1194         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1195         if (retval)
1196                 return retval;
1197
1198         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1199         if (filp->f_flags & O_DIRECT) {
1200                 loff_t size;
1201                 struct address_space *mapping;
1202                 struct inode *inode;
1203
1204                 mapping = filp->f_mapping;
1205                 inode = mapping->host;
1206                 retval = 0;
1207                 if (!count)
1208                         goto out; /* skip atime */
1209                 size = i_size_read(inode);
1210                 if (pos < size) {
1211                         retval = generic_file_direct_IO(READ, iocb,
1212                                                 iov, pos, nr_segs);
1213                         if (retval > 0)
1214                                 *ppos = pos + retval;
1215                 }
1216                 if (likely(retval != 0)) {
1217                         file_accessed(filp);
1218                         goto out;
1219                 }
1220         }
1221
1222         retval = 0;
1223         if (count) {
1224                 for (seg = 0; seg < nr_segs; seg++) {
1225                         read_descriptor_t desc;
1226
1227                         desc.written = 0;
1228                         desc.arg.buf = iov[seg].iov_base;
1229                         desc.count = iov[seg].iov_len;
1230                         if (desc.count == 0)
1231                                 continue;
1232                         desc.error = 0;
1233                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1234                         retval += desc.written;
1235                         if (desc.error) {
1236                                 retval = retval ?: desc.error;
1237                                 break;
1238                         }
1239                         if (desc.count > 0)
1240                                 break;
1241                 }
1242         }
1243 out:
1244         return retval;
1245 }
1246 EXPORT_SYMBOL(generic_file_aio_read);
1247
1248 static ssize_t
1249 do_readahead(struct address_space *mapping, struct file *filp,
1250              pgoff_t index, unsigned long nr)
1251 {
1252         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1253                 return -EINVAL;
1254
1255         force_page_cache_readahead(mapping, filp, index,
1256                                         max_sane_readahead(nr));
1257         return 0;
1258 }
1259
1260 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1261 {
1262         ssize_t ret;
1263         struct file *file;
1264
1265         ret = -EBADF;
1266         file = fget(fd);
1267         if (file) {
1268                 if (file->f_mode & FMODE_READ) {
1269                         struct address_space *mapping = file->f_mapping;
1270                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1271                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1272                         unsigned long len = end - start + 1;
1273                         ret = do_readahead(mapping, file, start, len);
1274                 }
1275                 fput(file);
1276         }
1277         return ret;
1278 }
1279
1280 #ifdef CONFIG_MMU
1281 /**
1282  * page_cache_read - adds requested page to the page cache if not already there
1283  * @file:       file to read
1284  * @offset:     page index
1285  *
1286  * This adds the requested page to the page cache if it isn't already there,
1287  * and schedules an I/O to read in its contents from disk.
1288  */
1289 static int page_cache_read(struct file *file, pgoff_t offset)
1290 {
1291         struct address_space *mapping = file->f_mapping;
1292         struct page *page; 
1293         int ret;
1294
1295         do {
1296                 page = page_cache_alloc_cold(mapping);
1297                 if (!page)
1298                         return -ENOMEM;
1299
1300                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1301                 if (ret == 0)
1302                         ret = mapping->a_ops->readpage(file, page);
1303                 else if (ret == -EEXIST)
1304                         ret = 0; /* losing race to add is OK */
1305
1306                 page_cache_release(page);
1307
1308         } while (ret == AOP_TRUNCATED_PAGE);
1309                 
1310         return ret;
1311 }
1312
1313 #define MMAP_LOTSAMISS  (100)
1314
1315 /**
1316  * filemap_fault - read in file data for page fault handling
1317  * @vma:        vma in which the fault was taken
1318  * @vmf:        struct vm_fault containing details of the fault
1319  *
1320  * filemap_fault() is invoked via the vma operations vector for a
1321  * mapped memory region to read in file data during a page fault.
1322  *
1323  * The goto's are kind of ugly, but this streamlines the normal case of having
1324  * it in the page cache, and handles the special cases reasonably without
1325  * having a lot of duplicated code.
1326  */
1327 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1328 {
1329         int error;
1330         struct file *file = vma->vm_file;
1331         struct address_space *mapping = file->f_mapping;
1332         struct file_ra_state *ra = &file->f_ra;
1333         struct inode *inode = mapping->host;
1334         struct page *page;
1335         pgoff_t size;
1336         int did_readaround = 0;
1337         int ret = 0;
1338
1339         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1340         if (vmf->pgoff >= size)
1341                 return VM_FAULT_SIGBUS;
1342
1343         /* If we don't want any read-ahead, don't bother */
1344         if (VM_RandomReadHint(vma))
1345                 goto no_cached_page;
1346
1347         /*
1348          * Do we have something in the page cache already?
1349          */
1350 retry_find:
1351         page = find_lock_page(mapping, vmf->pgoff);
1352         /*
1353          * For sequential accesses, we use the generic readahead logic.
1354          */
1355         if (VM_SequentialReadHint(vma)) {
1356                 if (!page) {
1357                         page_cache_sync_readahead(mapping, ra, file,
1358                                                            vmf->pgoff, 1);
1359                         page = find_lock_page(mapping, vmf->pgoff);
1360                         if (!page)
1361                                 goto no_cached_page;
1362                 }
1363                 if (PageReadahead(page)) {
1364                         page_cache_async_readahead(mapping, ra, file, page,
1365                                                            vmf->pgoff, 1);
1366                 }
1367         }
1368
1369         if (!page) {
1370                 unsigned long ra_pages;
1371
1372                 ra->mmap_miss++;
1373
1374                 /*
1375                  * Do we miss much more than hit in this file? If so,
1376                  * stop bothering with read-ahead. It will only hurt.
1377                  */
1378                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1379                         goto no_cached_page;
1380
1381                 /*
1382                  * To keep the pgmajfault counter straight, we need to
1383                  * check did_readaround, as this is an inner loop.
1384                  */
1385                 if (!did_readaround) {
1386                         ret = VM_FAULT_MAJOR;
1387                         count_vm_event(PGMAJFAULT);
1388                 }
1389                 did_readaround = 1;
1390                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1391                 if (ra_pages) {
1392                         pgoff_t start = 0;
1393
1394                         if (vmf->pgoff > ra_pages / 2)
1395                                 start = vmf->pgoff - ra_pages / 2;
1396                         do_page_cache_readahead(mapping, file, start, ra_pages);
1397                 }
1398                 page = find_lock_page(mapping, vmf->pgoff);
1399                 if (!page)
1400                         goto no_cached_page;
1401         }
1402
1403         if (!did_readaround)
1404                 ra->mmap_miss--;
1405
1406         /*
1407          * We have a locked page in the page cache, now we need to check
1408          * that it's up-to-date. If not, it is going to be due to an error.
1409          */
1410         if (unlikely(!PageUptodate(page)))
1411                 goto page_not_uptodate;
1412
1413         /* Must recheck i_size under page lock */
1414         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1415         if (unlikely(vmf->pgoff >= size)) {
1416                 unlock_page(page);
1417                 page_cache_release(page);
1418                 return VM_FAULT_SIGBUS;
1419         }
1420
1421         /*
1422          * Found the page and have a reference on it.
1423          */
1424         mark_page_accessed(page);
1425         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1426         vmf->page = page;
1427         return ret | VM_FAULT_LOCKED;
1428
1429 no_cached_page:
1430         /*
1431          * We're only likely to ever get here if MADV_RANDOM is in
1432          * effect.
1433          */
1434         error = page_cache_read(file, vmf->pgoff);
1435
1436         /*
1437          * The page we want has now been added to the page cache.
1438          * In the unlikely event that someone removed it in the
1439          * meantime, we'll just come back here and read it again.
1440          */
1441         if (error >= 0)
1442                 goto retry_find;
1443
1444         /*
1445          * An error return from page_cache_read can result if the
1446          * system is low on memory, or a problem occurs while trying
1447          * to schedule I/O.
1448          */
1449         if (error == -ENOMEM)
1450                 return VM_FAULT_OOM;
1451         return VM_FAULT_SIGBUS;
1452
1453 page_not_uptodate:
1454         /* IO error path */
1455         if (!did_readaround) {
1456                 ret = VM_FAULT_MAJOR;
1457                 count_vm_event(PGMAJFAULT);
1458         }
1459
1460         /*
1461          * Umm, take care of errors if the page isn't up-to-date.
1462          * Try to re-read it _once_. We do this synchronously,
1463          * because there really aren't any performance issues here
1464          * and we need to check for errors.
1465          */
1466         ClearPageError(page);
1467         error = mapping->a_ops->readpage(file, page);
1468         page_cache_release(page);
1469
1470         if (!error || error == AOP_TRUNCATED_PAGE)
1471                 goto retry_find;
1472
1473         /* Things didn't work out. Return zero to tell the mm layer so. */
1474         shrink_readahead_size_eio(file, ra);
1475         return VM_FAULT_SIGBUS;
1476 }
1477 EXPORT_SYMBOL(filemap_fault);
1478
1479 struct vm_operations_struct generic_file_vm_ops = {
1480         .fault          = filemap_fault,
1481 };
1482
1483 /* This is used for a general mmap of a disk file */
1484
1485 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1486 {
1487         struct address_space *mapping = file->f_mapping;
1488
1489         if (!mapping->a_ops->readpage)
1490                 return -ENOEXEC;
1491         file_accessed(file);
1492         vma->vm_ops = &generic_file_vm_ops;
1493         vma->vm_flags |= VM_CAN_NONLINEAR;
1494         return 0;
1495 }
1496
1497 /*
1498  * This is for filesystems which do not implement ->writepage.
1499  */
1500 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1501 {
1502         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1503                 return -EINVAL;
1504         return generic_file_mmap(file, vma);
1505 }
1506 #else
1507 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1508 {
1509         return -ENOSYS;
1510 }
1511 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1512 {
1513         return -ENOSYS;
1514 }
1515 #endif /* CONFIG_MMU */
1516
1517 EXPORT_SYMBOL(generic_file_mmap);
1518 EXPORT_SYMBOL(generic_file_readonly_mmap);
1519
1520 static struct page *__read_cache_page(struct address_space *mapping,
1521                                 pgoff_t index,
1522                                 int (*filler)(void *,struct page*),
1523                                 void *data)
1524 {
1525         struct page *page;
1526         int err;
1527 repeat:
1528         page = find_get_page(mapping, index);
1529         if (!page) {
1530                 page = page_cache_alloc_cold(mapping);
1531                 if (!page)
1532                         return ERR_PTR(-ENOMEM);
1533                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1534                 if (unlikely(err)) {
1535                         page_cache_release(page);
1536                         if (err == -EEXIST)
1537                                 goto repeat;
1538                         /* Presumably ENOMEM for radix tree node */
1539                         return ERR_PTR(err);
1540                 }
1541                 err = filler(data, page);
1542                 if (err < 0) {
1543                         page_cache_release(page);
1544                         page = ERR_PTR(err);
1545                 }
1546         }
1547         return page;
1548 }
1549
1550 /*
1551  * Same as read_cache_page, but don't wait for page to become unlocked
1552  * after submitting it to the filler.
1553  */
1554 struct page *read_cache_page_async(struct address_space *mapping,
1555                                 pgoff_t index,
1556                                 int (*filler)(void *,struct page*),
1557                                 void *data)
1558 {
1559         struct page *page;
1560         int err;
1561
1562 retry:
1563         page = __read_cache_page(mapping, index, filler, data);
1564         if (IS_ERR(page))
1565                 return page;
1566         if (PageUptodate(page))
1567                 goto out;
1568
1569         lock_page(page);
1570         if (!page->mapping) {
1571                 unlock_page(page);
1572                 page_cache_release(page);
1573                 goto retry;
1574         }
1575         if (PageUptodate(page)) {
1576                 unlock_page(page);
1577                 goto out;
1578         }
1579         err = filler(data, page);
1580         if (err < 0) {
1581                 page_cache_release(page);
1582                 return ERR_PTR(err);
1583         }
1584 out:
1585         mark_page_accessed(page);
1586         return page;
1587 }
1588 EXPORT_SYMBOL(read_cache_page_async);
1589
1590 /**
1591  * read_cache_page - read into page cache, fill it if needed
1592  * @mapping:    the page's address_space
1593  * @index:      the page index
1594  * @filler:     function to perform the read
1595  * @data:       destination for read data
1596  *
1597  * Read into the page cache. If a page already exists, and PageUptodate() is
1598  * not set, try to fill the page then wait for it to become unlocked.
1599  *
1600  * If the page does not get brought uptodate, return -EIO.
1601  */
1602 struct page *read_cache_page(struct address_space *mapping,
1603                                 pgoff_t index,
1604                                 int (*filler)(void *,struct page*),
1605                                 void *data)
1606 {
1607         struct page *page;
1608
1609         page = read_cache_page_async(mapping, index, filler, data);
1610         if (IS_ERR(page))
1611                 goto out;
1612         wait_on_page_locked(page);
1613         if (!PageUptodate(page)) {
1614                 page_cache_release(page);
1615                 page = ERR_PTR(-EIO);
1616         }
1617  out:
1618         return page;
1619 }
1620 EXPORT_SYMBOL(read_cache_page);
1621
1622 /*
1623  * The logic we want is
1624  *
1625  *      if suid or (sgid and xgrp)
1626  *              remove privs
1627  */
1628 int should_remove_suid(struct dentry *dentry)
1629 {
1630         mode_t mode = dentry->d_inode->i_mode;
1631         int kill = 0;
1632
1633         /* suid always must be killed */
1634         if (unlikely(mode & S_ISUID))
1635                 kill = ATTR_KILL_SUID;
1636
1637         /*
1638          * sgid without any exec bits is just a mandatory locking mark; leave
1639          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1640          */
1641         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1642                 kill |= ATTR_KILL_SGID;
1643
1644         if (unlikely(kill && !capable(CAP_FSETID)))
1645                 return kill;
1646
1647         return 0;
1648 }
1649 EXPORT_SYMBOL(should_remove_suid);
1650
1651 int __remove_suid(struct dentry *dentry, int kill)
1652 {
1653         struct iattr newattrs;
1654
1655         newattrs.ia_valid = ATTR_FORCE | kill;
1656         return notify_change(dentry, &newattrs);
1657 }
1658
1659 int remove_suid(struct dentry *dentry)
1660 {
1661         int killsuid = should_remove_suid(dentry);
1662         int killpriv = security_inode_need_killpriv(dentry);
1663         int error = 0;
1664
1665         if (killpriv < 0)
1666                 return killpriv;
1667         if (killpriv)
1668                 error = security_inode_killpriv(dentry);
1669         if (!error && killsuid)
1670                 error = __remove_suid(dentry, killsuid);
1671
1672         return error;
1673 }
1674 EXPORT_SYMBOL(remove_suid);
1675
1676 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1677                         const struct iovec *iov, size_t base, size_t bytes)
1678 {
1679         size_t copied = 0, left = 0;
1680
1681         while (bytes) {
1682                 char __user *buf = iov->iov_base + base;
1683                 int copy = min(bytes, iov->iov_len - base);
1684
1685                 base = 0;
1686                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1687                 copied += copy;
1688                 bytes -= copy;
1689                 vaddr += copy;
1690                 iov++;
1691
1692                 if (unlikely(left))
1693                         break;
1694         }
1695         return copied - left;
1696 }
1697
1698 /*
1699  * Copy as much as we can into the page and return the number of bytes which
1700  * were sucessfully copied.  If a fault is encountered then return the number of
1701  * bytes which were copied.
1702  */
1703 size_t iov_iter_copy_from_user_atomic(struct page *page,
1704                 struct iov_iter *i, unsigned long offset, size_t bytes)
1705 {
1706         char *kaddr;
1707         size_t copied;
1708
1709         BUG_ON(!in_atomic());
1710         kaddr = kmap_atomic(page, KM_USER0);
1711         if (likely(i->nr_segs == 1)) {
1712                 int left;
1713                 char __user *buf = i->iov->iov_base + i->iov_offset;
1714                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1715                                                         buf, bytes);
1716                 copied = bytes - left;
1717         } else {
1718                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1719                                                 i->iov, i->iov_offset, bytes);
1720         }
1721         kunmap_atomic(kaddr, KM_USER0);
1722
1723         return copied;
1724 }
1725 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1726
1727 /*
1728  * This has the same sideeffects and return value as
1729  * iov_iter_copy_from_user_atomic().
1730  * The difference is that it attempts to resolve faults.
1731  * Page must not be locked.
1732  */
1733 size_t iov_iter_copy_from_user(struct page *page,
1734                 struct iov_iter *i, unsigned long offset, size_t bytes)
1735 {
1736         char *kaddr;
1737         size_t copied;
1738
1739         kaddr = kmap(page);
1740         if (likely(i->nr_segs == 1)) {
1741                 int left;
1742                 char __user *buf = i->iov->iov_base + i->iov_offset;
1743                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1744                 copied = bytes - left;
1745         } else {
1746                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1747                                                 i->iov, i->iov_offset, bytes);
1748         }
1749         kunmap(page);
1750         return copied;
1751 }
1752 EXPORT_SYMBOL(iov_iter_copy_from_user);
1753
1754 static void __iov_iter_advance_iov(struct iov_iter *i, size_t bytes)
1755 {
1756         if (likely(i->nr_segs == 1)) {
1757                 i->iov_offset += bytes;
1758         } else {
1759                 const struct iovec *iov = i->iov;
1760                 size_t base = i->iov_offset;
1761
1762                 /*
1763                  * The !iov->iov_len check ensures we skip over unlikely
1764                  * zero-length segments.
1765                  */
1766                 while (bytes || !iov->iov_len) {
1767                         int copy = min(bytes, iov->iov_len - base);
1768
1769                         bytes -= copy;
1770                         base += copy;
1771                         if (iov->iov_len == base) {
1772                                 iov++;
1773                                 base = 0;
1774                         }
1775                 }
1776                 i->iov = iov;
1777                 i->iov_offset = base;
1778         }
1779 }
1780
1781 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1782 {
1783         BUG_ON(i->count < bytes);
1784
1785         __iov_iter_advance_iov(i, bytes);
1786         i->count -= bytes;
1787 }
1788 EXPORT_SYMBOL(iov_iter_advance);
1789
1790 /*
1791  * Fault in the first iovec of the given iov_iter, to a maximum length
1792  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1793  * accessed (ie. because it is an invalid address).
1794  *
1795  * writev-intensive code may want this to prefault several iovecs -- that
1796  * would be possible (callers must not rely on the fact that _only_ the
1797  * first iovec will be faulted with the current implementation).
1798  */
1799 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1800 {
1801         char __user *buf = i->iov->iov_base + i->iov_offset;
1802         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1803         return fault_in_pages_readable(buf, bytes);
1804 }
1805 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1806
1807 /*
1808  * Return the count of just the current iov_iter segment.
1809  */
1810 size_t iov_iter_single_seg_count(struct iov_iter *i)
1811 {
1812         const struct iovec *iov = i->iov;
1813         if (i->nr_segs == 1)
1814                 return i->count;
1815         else
1816                 return min(i->count, iov->iov_len - i->iov_offset);
1817 }
1818 EXPORT_SYMBOL(iov_iter_single_seg_count);
1819
1820 /*
1821  * Performs necessary checks before doing a write
1822  *
1823  * Can adjust writing position or amount of bytes to write.
1824  * Returns appropriate error code that caller should return or
1825  * zero in case that write should be allowed.
1826  */
1827 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1828 {
1829         struct inode *inode = file->f_mapping->host;
1830         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1831
1832         if (unlikely(*pos < 0))
1833                 return -EINVAL;
1834
1835         if (!isblk) {
1836                 /* FIXME: this is for backwards compatibility with 2.4 */
1837                 if (file->f_flags & O_APPEND)
1838                         *pos = i_size_read(inode);
1839
1840                 if (limit != RLIM_INFINITY) {
1841                         if (*pos >= limit) {
1842                                 send_sig(SIGXFSZ, current, 0);
1843                                 return -EFBIG;
1844                         }
1845                         if (*count > limit - (typeof(limit))*pos) {
1846                                 *count = limit - (typeof(limit))*pos;
1847                         }
1848                 }
1849         }
1850
1851         /*
1852          * LFS rule
1853          */
1854         if (unlikely(*pos + *count > MAX_NON_LFS &&
1855                                 !(file->f_flags & O_LARGEFILE))) {
1856                 if (*pos >= MAX_NON_LFS) {
1857                         return -EFBIG;
1858                 }
1859                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1860                         *count = MAX_NON_LFS - (unsigned long)*pos;
1861                 }
1862         }
1863
1864         /*
1865          * Are we about to exceed the fs block limit ?
1866          *
1867          * If we have written data it becomes a short write.  If we have
1868          * exceeded without writing data we send a signal and return EFBIG.
1869          * Linus frestrict idea will clean these up nicely..
1870          */
1871         if (likely(!isblk)) {
1872                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1873                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1874                                 return -EFBIG;
1875                         }
1876                         /* zero-length writes at ->s_maxbytes are OK */
1877                 }
1878
1879                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1880                         *count = inode->i_sb->s_maxbytes - *pos;
1881         } else {
1882 #ifdef CONFIG_BLOCK
1883                 loff_t isize;
1884                 if (bdev_read_only(I_BDEV(inode)))
1885                         return -EPERM;
1886                 isize = i_size_read(inode);
1887                 if (*pos >= isize) {
1888                         if (*count || *pos > isize)
1889                                 return -ENOSPC;
1890                 }
1891
1892                 if (*pos + *count > isize)
1893                         *count = isize - *pos;
1894 #else
1895                 return -EPERM;
1896 #endif
1897         }
1898         return 0;
1899 }
1900 EXPORT_SYMBOL(generic_write_checks);
1901
1902 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1903                                 loff_t pos, unsigned len, unsigned flags,
1904                                 struct page **pagep, void **fsdata)
1905 {
1906         const struct address_space_operations *aops = mapping->a_ops;
1907
1908         if (aops->write_begin) {
1909                 return aops->write_begin(file, mapping, pos, len, flags,
1910                                                         pagep, fsdata);
1911         } else {
1912                 int ret;
1913                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1914                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1915                 struct inode *inode = mapping->host;
1916                 struct page *page;
1917 again:
1918                 page = __grab_cache_page(mapping, index);
1919                 *pagep = page;
1920                 if (!page)
1921                         return -ENOMEM;
1922
1923                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1924                         /*
1925                          * There is no way to resolve a short write situation
1926                          * for a !Uptodate page (except by double copying in
1927                          * the caller done by generic_perform_write_2copy).
1928                          *
1929                          * Instead, we have to bring it uptodate here.
1930                          */
1931                         ret = aops->readpage(file, page);
1932                         page_cache_release(page);
1933                         if (ret) {
1934                                 if (ret == AOP_TRUNCATED_PAGE)
1935                                         goto again;
1936                                 return ret;
1937                         }
1938                         goto again;
1939                 }
1940
1941                 ret = aops->prepare_write(file, page, offset, offset+len);
1942                 if (ret) {
1943                         unlock_page(page);
1944                         page_cache_release(page);
1945                         if (pos + len > inode->i_size)
1946                                 vmtruncate(inode, inode->i_size);
1947                 }
1948                 return ret;
1949         }
1950 }
1951 EXPORT_SYMBOL(pagecache_write_begin);
1952
1953 int pagecache_write_end(struct file *file, struct address_space *mapping,
1954                                 loff_t pos, unsigned len, unsigned copied,
1955                                 struct page *page, void *fsdata)
1956 {
1957         const struct address_space_operations *aops = mapping->a_ops;
1958         int ret;
1959
1960         if (aops->write_end) {
1961                 mark_page_accessed(page);
1962                 ret = aops->write_end(file, mapping, pos, len, copied,
1963                                                         page, fsdata);
1964         } else {
1965                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1966                 struct inode *inode = mapping->host;
1967
1968                 flush_dcache_page(page);
1969                 ret = aops->commit_write(file, page, offset, offset+len);
1970                 unlock_page(page);
1971                 mark_page_accessed(page);
1972                 page_cache_release(page);
1973
1974                 if (ret < 0) {
1975                         if (pos + len > inode->i_size)
1976                                 vmtruncate(inode, inode->i_size);
1977                 } else if (ret > 0)
1978                         ret = min_t(size_t, copied, ret);
1979                 else
1980                         ret = copied;
1981         }
1982
1983         return ret;
1984 }
1985 EXPORT_SYMBOL(pagecache_write_end);
1986
1987 ssize_t
1988 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1989                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1990                 size_t count, size_t ocount)
1991 {
1992         struct file     *file = iocb->ki_filp;
1993         struct address_space *mapping = file->f_mapping;
1994         struct inode    *inode = mapping->host;
1995         ssize_t         written;
1996
1997         if (count != ocount)
1998                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1999
2000         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2001         if (written > 0) {
2002                 loff_t end = pos + written;
2003                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2004                         i_size_write(inode,  end);
2005                         mark_inode_dirty(inode);
2006                 }
2007                 *ppos = end;
2008         }
2009
2010         /*
2011          * Sync the fs metadata but not the minor inode changes and
2012          * of course not the data as we did direct DMA for the IO.
2013          * i_mutex is held, which protects generic_osync_inode() from
2014          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2015          */
2016         if ((written >= 0 || written == -EIOCBQUEUED) &&
2017             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2018                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2019                 if (err < 0)
2020                         written = err;
2021         }
2022         return written;
2023 }
2024 EXPORT_SYMBOL(generic_file_direct_write);
2025
2026 /*
2027  * Find or create a page at the given pagecache position. Return the locked
2028  * page. This function is specifically for buffered writes.
2029  */
2030 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2031 {
2032         int status;
2033         struct page *page;
2034 repeat:
2035         page = find_lock_page(mapping, index);
2036         if (likely(page))
2037                 return page;
2038
2039         page = page_cache_alloc(mapping);
2040         if (!page)
2041                 return NULL;
2042         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2043         if (unlikely(status)) {
2044                 page_cache_release(page);
2045                 if (status == -EEXIST)
2046                         goto repeat;
2047                 return NULL;
2048         }
2049         return page;
2050 }
2051 EXPORT_SYMBOL(__grab_cache_page);
2052
2053 static ssize_t generic_perform_write_2copy(struct file *file,
2054                                 struct iov_iter *i, loff_t pos)
2055 {
2056         struct address_space *mapping = file->f_mapping;
2057         const struct address_space_operations *a_ops = mapping->a_ops;
2058         struct inode *inode = mapping->host;
2059         long status = 0;
2060         ssize_t written = 0;
2061
2062         do {
2063                 struct page *src_page;
2064                 struct page *page;
2065                 pgoff_t index;          /* Pagecache index for current page */
2066                 unsigned long offset;   /* Offset into pagecache page */
2067                 unsigned long bytes;    /* Bytes to write to page */
2068                 size_t copied;          /* Bytes copied from user */
2069
2070                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2071                 index = pos >> PAGE_CACHE_SHIFT;
2072                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2073                                                 iov_iter_count(i));
2074
2075                 /*
2076                  * a non-NULL src_page indicates that we're doing the
2077                  * copy via get_user_pages and kmap.
2078                  */
2079                 src_page = NULL;
2080
2081                 /*
2082                  * Bring in the user page that we will copy from _first_.
2083                  * Otherwise there's a nasty deadlock on copying from the
2084                  * same page as we're writing to, without it being marked
2085                  * up-to-date.
2086                  *
2087                  * Not only is this an optimisation, but it is also required
2088                  * to check that the address is actually valid, when atomic
2089                  * usercopies are used, below.
2090                  */
2091                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2092                         status = -EFAULT;
2093                         break;
2094                 }
2095
2096                 page = __grab_cache_page(mapping, index);
2097                 if (!page) {
2098                         status = -ENOMEM;
2099                         break;
2100                 }
2101
2102                 /*
2103                  * non-uptodate pages cannot cope with short copies, and we
2104                  * cannot take a pagefault with the destination page locked.
2105                  * So pin the source page to copy it.
2106                  */
2107                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2108                         unlock_page(page);
2109
2110                         src_page = alloc_page(GFP_KERNEL);
2111                         if (!src_page) {
2112                                 page_cache_release(page);
2113                                 status = -ENOMEM;
2114                                 break;
2115                         }
2116
2117                         /*
2118                          * Cannot get_user_pages with a page locked for the
2119                          * same reason as we can't take a page fault with a
2120                          * page locked (as explained below).
2121                          */
2122                         copied = iov_iter_copy_from_user(src_page, i,
2123                                                                 offset, bytes);
2124                         if (unlikely(copied == 0)) {
2125                                 status = -EFAULT;
2126                                 page_cache_release(page);
2127                                 page_cache_release(src_page);
2128                                 break;
2129                         }
2130                         bytes = copied;
2131
2132                         lock_page(page);
2133                         /*
2134                          * Can't handle the page going uptodate here, because
2135                          * that means we would use non-atomic usercopies, which
2136                          * zero out the tail of the page, which can cause
2137                          * zeroes to become transiently visible. We could just
2138                          * use a non-zeroing copy, but the APIs aren't too
2139                          * consistent.
2140                          */
2141                         if (unlikely(!page->mapping || PageUptodate(page))) {
2142                                 unlock_page(page);
2143                                 page_cache_release(page);
2144                                 page_cache_release(src_page);
2145                                 continue;
2146                         }
2147                 }
2148
2149                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2150                 if (unlikely(status))
2151                         goto fs_write_aop_error;
2152
2153                 if (!src_page) {
2154                         /*
2155                          * Must not enter the pagefault handler here, because
2156                          * we hold the page lock, so we might recursively
2157                          * deadlock on the same lock, or get an ABBA deadlock
2158                          * against a different lock, or against the mmap_sem
2159                          * (which nests outside the page lock).  So increment
2160                          * preempt count, and use _atomic usercopies.
2161                          *
2162                          * The page is uptodate so we are OK to encounter a
2163                          * short copy: if unmodified parts of the page are
2164                          * marked dirty and written out to disk, it doesn't
2165                          * really matter.
2166                          */
2167                         pagefault_disable();
2168                         copied = iov_iter_copy_from_user_atomic(page, i,
2169                                                                 offset, bytes);
2170                         pagefault_enable();
2171                 } else {
2172                         void *src, *dst;
2173                         src = kmap_atomic(src_page, KM_USER0);
2174                         dst = kmap_atomic(page, KM_USER1);
2175                         memcpy(dst + offset, src + offset, bytes);
2176                         kunmap_atomic(dst, KM_USER1);
2177                         kunmap_atomic(src, KM_USER0);
2178                         copied = bytes;
2179                 }
2180                 flush_dcache_page(page);
2181
2182                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2183                 if (unlikely(status < 0))
2184                         goto fs_write_aop_error;
2185                 if (unlikely(status > 0)) /* filesystem did partial write */
2186                         copied = min_t(size_t, copied, status);
2187
2188                 unlock_page(page);
2189                 mark_page_accessed(page);
2190                 page_cache_release(page);
2191                 if (src_page)
2192                         page_cache_release(src_page);
2193
2194                 iov_iter_advance(i, copied);
2195                 pos += copied;
2196                 written += copied;
2197
2198                 balance_dirty_pages_ratelimited(mapping);
2199                 cond_resched();
2200                 continue;
2201
2202 fs_write_aop_error:
2203                 unlock_page(page);
2204                 page_cache_release(page);
2205                 if (src_page)
2206                         page_cache_release(src_page);
2207
2208                 /*
2209                  * prepare_write() may have instantiated a few blocks
2210                  * outside i_size.  Trim these off again. Don't need
2211                  * i_size_read because we hold i_mutex.
2212                  */
2213                 if (pos + bytes > inode->i_size)
2214                         vmtruncate(inode, inode->i_size);
2215                 break;
2216         } while (iov_iter_count(i));
2217
2218         return written ? written : status;
2219 }
2220
2221 static ssize_t generic_perform_write(struct file *file,
2222                                 struct iov_iter *i, loff_t pos)
2223 {
2224         struct address_space *mapping = file->f_mapping;
2225         const struct address_space_operations *a_ops = mapping->a_ops;
2226         long status = 0;
2227         ssize_t written = 0;
2228         unsigned int flags = 0;
2229
2230         /*
2231          * Copies from kernel address space cannot fail (NFSD is a big user).
2232          */
2233         if (segment_eq(get_fs(), KERNEL_DS))
2234                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2235
2236         do {
2237                 struct page *page;
2238                 pgoff_t index;          /* Pagecache index for current page */
2239                 unsigned long offset;   /* Offset into pagecache page */
2240                 unsigned long bytes;    /* Bytes to write to page */
2241                 size_t copied;          /* Bytes copied from user */
2242                 void *fsdata;
2243
2244                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2245                 index = pos >> PAGE_CACHE_SHIFT;
2246                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2247                                                 iov_iter_count(i));
2248
2249 again:
2250
2251                 /*
2252                  * Bring in the user page that we will copy from _first_.
2253                  * Otherwise there's a nasty deadlock on copying from the
2254                  * same page as we're writing to, without it being marked
2255                  * up-to-date.
2256                  *
2257                  * Not only is this an optimisation, but it is also required
2258                  * to check that the address is actually valid, when atomic
2259                  * usercopies are used, below.
2260                  */
2261                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2262                         status = -EFAULT;
2263                         break;
2264                 }
2265
2266                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2267                                                 &page, &fsdata);
2268                 if (unlikely(status))
2269                         break;
2270
2271                 pagefault_disable();
2272                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2273                 pagefault_enable();
2274                 flush_dcache_page(page);
2275
2276                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2277                                                 page, fsdata);
2278                 if (unlikely(status < 0))
2279                         break;
2280                 copied = status;
2281
2282                 cond_resched();
2283
2284                 iov_iter_advance(i, copied);
2285                 if (unlikely(copied == 0)) {
2286                         /*
2287                          * If we were unable to copy any data at all, we must
2288                          * fall back to a single segment length write.
2289                          *
2290                          * If we didn't fallback here, we could livelock
2291                          * because not all segments in the iov can be copied at
2292                          * once without a pagefault.
2293                          */
2294                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2295                                                 iov_iter_single_seg_count(i));
2296                         goto again;
2297                 }
2298                 pos += copied;
2299                 written += copied;
2300
2301                 balance_dirty_pages_ratelimited(mapping);
2302
2303         } while (iov_iter_count(i));
2304
2305         return written ? written : status;
2306 }
2307
2308 ssize_t
2309 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2310                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2311                 size_t count, ssize_t written)
2312 {
2313         struct file *file = iocb->ki_filp;
2314         struct address_space *mapping = file->f_mapping;
2315         const struct address_space_operations *a_ops = mapping->a_ops;
2316         struct inode *inode = mapping->host;
2317         ssize_t status;
2318         struct iov_iter i;
2319
2320         iov_iter_init(&i, iov, nr_segs, count, written);
2321         if (a_ops->write_begin)
2322                 status = generic_perform_write(file, &i, pos);
2323         else
2324                 status = generic_perform_write_2copy(file, &i, pos);
2325
2326         if (likely(status >= 0)) {
2327                 written += status;
2328                 *ppos = pos + status;
2329
2330                 /*
2331                  * For now, when the user asks for O_SYNC, we'll actually give
2332                  * O_DSYNC
2333                  */
2334                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2335                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2336                                 status = generic_osync_inode(inode, mapping,
2337                                                 OSYNC_METADATA|OSYNC_DATA);
2338                 }
2339         }
2340         
2341         /*
2342          * If we get here for O_DIRECT writes then we must have fallen through
2343          * to buffered writes (block instantiation inside i_size).  So we sync
2344          * the file data here, to try to honour O_DIRECT expectations.
2345          */
2346         if (unlikely(file->f_flags & O_DIRECT) && written)
2347                 status = filemap_write_and_wait(mapping);
2348
2349         return written ? written : status;
2350 }
2351 EXPORT_SYMBOL(generic_file_buffered_write);
2352
2353 static ssize_t
2354 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2355                                 unsigned long nr_segs, loff_t *ppos)
2356 {
2357         struct file *file = iocb->ki_filp;
2358         struct address_space * mapping = file->f_mapping;
2359         size_t ocount;          /* original count */
2360         size_t count;           /* after file limit checks */
2361         struct inode    *inode = mapping->host;
2362         loff_t          pos;
2363         ssize_t         written;
2364         ssize_t         err;
2365
2366         ocount = 0;
2367         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2368         if (err)
2369                 return err;
2370
2371         count = ocount;
2372         pos = *ppos;
2373
2374         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2375
2376         /* We can write back this queue in page reclaim */
2377         current->backing_dev_info = mapping->backing_dev_info;
2378         written = 0;
2379
2380         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2381         if (err)
2382                 goto out;
2383
2384         if (count == 0)
2385                 goto out;
2386
2387         err = remove_suid(file->f_path.dentry);
2388         if (err)
2389                 goto out;
2390
2391         file_update_time(file);
2392
2393         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2394         if (unlikely(file->f_flags & O_DIRECT)) {
2395                 loff_t endbyte;
2396                 ssize_t written_buffered;
2397
2398                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2399                                                         ppos, count, ocount);
2400                 if (written < 0 || written == count)
2401                         goto out;
2402                 /*
2403                  * direct-io write to a hole: fall through to buffered I/O
2404                  * for completing the rest of the request.
2405                  */
2406                 pos += written;
2407                 count -= written;
2408                 written_buffered = generic_file_buffered_write(iocb, iov,
2409                                                 nr_segs, pos, ppos, count,
2410                                                 written);
2411                 /*
2412                  * If generic_file_buffered_write() retuned a synchronous error
2413                  * then we want to return the number of bytes which were
2414                  * direct-written, or the error code if that was zero.  Note
2415                  * that this differs from normal direct-io semantics, which
2416                  * will return -EFOO even if some bytes were written.
2417                  */
2418                 if (written_buffered < 0) {
2419                         err = written_buffered;
2420                         goto out;
2421                 }
2422
2423                 /*
2424                  * We need to ensure that the page cache pages are written to
2425                  * disk and invalidated to preserve the expected O_DIRECT
2426                  * semantics.
2427                  */
2428                 endbyte = pos + written_buffered - written - 1;
2429                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2430                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2431                                             SYNC_FILE_RANGE_WRITE|
2432                                             SYNC_FILE_RANGE_WAIT_AFTER);
2433                 if (err == 0) {
2434                         written = written_buffered;
2435                         invalidate_mapping_pages(mapping,
2436                                                  pos >> PAGE_CACHE_SHIFT,
2437                                                  endbyte >> PAGE_CACHE_SHIFT);
2438                 } else {
2439                         /*
2440                          * We don't know how much we wrote, so just return
2441                          * the number of bytes which were direct-written
2442                          */
2443                 }
2444         } else {
2445                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2446                                 pos, ppos, count, written);
2447         }
2448 out:
2449         current->backing_dev_info = NULL;
2450         return written ? written : err;
2451 }
2452
2453 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2454                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2455 {
2456         struct file *file = iocb->ki_filp;
2457         struct address_space *mapping = file->f_mapping;
2458         struct inode *inode = mapping->host;
2459         ssize_t ret;
2460
2461         BUG_ON(iocb->ki_pos != pos);
2462
2463         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2464                         &iocb->ki_pos);
2465
2466         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2467                 ssize_t err;
2468
2469                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2470                 if (err < 0)
2471                         ret = err;
2472         }
2473         return ret;
2474 }
2475 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2476
2477 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2478                 unsigned long nr_segs, loff_t pos)
2479 {
2480         struct file *file = iocb->ki_filp;
2481         struct address_space *mapping = file->f_mapping;
2482         struct inode *inode = mapping->host;
2483         ssize_t ret;
2484
2485         BUG_ON(iocb->ki_pos != pos);
2486
2487         mutex_lock(&inode->i_mutex);
2488         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2489                         &iocb->ki_pos);
2490         mutex_unlock(&inode->i_mutex);
2491
2492         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2493                 ssize_t err;
2494
2495                 err = sync_page_range(inode, mapping, pos, ret);
2496                 if (err < 0)
2497                         ret = err;
2498         }
2499         return ret;
2500 }
2501 EXPORT_SYMBOL(generic_file_aio_write);
2502
2503 /*
2504  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2505  * went wrong during pagecache shootdown.
2506  */
2507 static ssize_t
2508 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2509         loff_t offset, unsigned long nr_segs)
2510 {
2511         struct file *file = iocb->ki_filp;
2512         struct address_space *mapping = file->f_mapping;
2513         ssize_t retval;
2514         size_t write_len;
2515         pgoff_t end = 0; /* silence gcc */
2516
2517         /*
2518          * If it's a write, unmap all mmappings of the file up-front.  This
2519          * will cause any pte dirty bits to be propagated into the pageframes
2520          * for the subsequent filemap_write_and_wait().
2521          */
2522         if (rw == WRITE) {
2523                 write_len = iov_length(iov, nr_segs);
2524                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2525                 if (mapping_mapped(mapping))
2526                         unmap_mapping_range(mapping, offset, write_len, 0);
2527         }
2528
2529         retval = filemap_write_and_wait(mapping);
2530         if (retval)
2531                 goto out;
2532
2533         /*
2534          * After a write we want buffered reads to be sure to go to disk to get
2535          * the new data.  We invalidate clean cached page from the region we're
2536          * about to write.  We do this *before* the write so that we can return
2537          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2538          */
2539         if (rw == WRITE && mapping->nrpages) {
2540                 retval = invalidate_inode_pages2_range(mapping,
2541                                         offset >> PAGE_CACHE_SHIFT, end);
2542                 if (retval)
2543                         goto out;
2544         }
2545
2546         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2547
2548         /*
2549          * Finally, try again to invalidate clean pages which might have been
2550          * cached by non-direct readahead, or faulted in by get_user_pages()
2551          * if the source of the write was an mmap'ed region of the file
2552          * we're writing.  Either one is a pretty crazy thing to do,
2553          * so we don't support it 100%.  If this invalidation
2554          * fails, tough, the write still worked...
2555          */
2556         if (rw == WRITE && mapping->nrpages) {
2557                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2558         }
2559 out:
2560         return retval;
2561 }
2562
2563 /**
2564  * try_to_release_page() - release old fs-specific metadata on a page
2565  *
2566  * @page: the page which the kernel is trying to free
2567  * @gfp_mask: memory allocation flags (and I/O mode)
2568  *
2569  * The address_space is to try to release any data against the page
2570  * (presumably at page->private).  If the release was successful, return `1'.
2571  * Otherwise return zero.
2572  *
2573  * The @gfp_mask argument specifies whether I/O may be performed to release
2574  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2575  *
2576  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2577  */
2578 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2579 {
2580         struct address_space * const mapping = page->mapping;
2581
2582         BUG_ON(!PageLocked(page));
2583         if (PageWriteback(page))
2584                 return 0;
2585
2586         if (mapping && mapping->a_ops->releasepage)
2587                 return mapping->a_ops->releasepage(page, gfp_mask);
2588         return try_to_free_buffers(page);
2589 }
2590
2591 EXPORT_SYMBOL(try_to_release_page);