MIPS: Remove duplicated #include
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
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 try_to_free_buffers */
43
44 #include <asm/mman.h>
45
46 /*
47  * Shared mappings implemented 30.11.1994. It's not fully working yet,
48  * though.
49  *
50  * Shared mappings now work. 15.8.1995  Bruno.
51  *
52  * finished 'unifying' the page and buffer cache and SMP-threaded the
53  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
54  *
55  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
56  */
57
58 /*
59  * Lock ordering:
60  *
61  *  ->i_mmap_lock               (truncate_pagecache)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  *
108  *  (code doesn't rely on that order, so you could switch it around)
109  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
110  *    ->i_mmap_lock
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 the mapping's tree_lock.
117  */
118 void __remove_from_page_cache(struct page *page)
119 {
120         struct address_space *mapping = page->mapping;
121
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         if (PageSwapBacked(page))
127                 __dec_zone_page_state(page, NR_SHMEM);
128         BUG_ON(page_mapped(page));
129
130         /*
131          * Some filesystems seem to re-dirty the page even after
132          * the VM has canceled the dirty bit (eg ext3 journaling).
133          *
134          * Fix it up by doing a final dirty accounting check after
135          * having removed the page entirely.
136          */
137         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
138                 dec_zone_page_state(page, NR_FILE_DIRTY);
139                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
140         }
141 }
142
143 void remove_from_page_cache(struct page *page)
144 {
145         struct address_space *mapping = page->mapping;
146
147         BUG_ON(!PageLocked(page));
148
149         spin_lock_irq(&mapping->tree_lock);
150         __remove_from_page_cache(page);
151         spin_unlock_irq(&mapping->tree_lock);
152         mem_cgroup_uncharge_cache_page(page);
153 }
154
155 static int sync_page(void *word)
156 {
157         struct address_space *mapping;
158         struct page *page;
159
160         page = container_of((unsigned long *)word, struct page, flags);
161
162         /*
163          * page_mapping() is being called without PG_locked held.
164          * Some knowledge of the state and use of the page is used to
165          * reduce the requirements down to a memory barrier.
166          * The danger here is of a stale page_mapping() return value
167          * indicating a struct address_space different from the one it's
168          * associated with when it is associated with one.
169          * After smp_mb(), it's either the correct page_mapping() for
170          * the page, or an old page_mapping() and the page's own
171          * page_mapping() has gone NULL.
172          * The ->sync_page() address_space operation must tolerate
173          * page_mapping() going NULL. By an amazing coincidence,
174          * this comes about because none of the users of the page
175          * in the ->sync_page() methods make essential use of the
176          * page_mapping(), merely passing the page down to the backing
177          * device's unplug functions when it's non-NULL, which in turn
178          * ignore it for all cases but swap, where only page_private(page) is
179          * of interest. When page_mapping() does go NULL, the entire
180          * call stack gracefully ignores the page and returns.
181          * -- wli
182          */
183         smp_mb();
184         mapping = page_mapping(page);
185         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
186                 mapping->a_ops->sync_page(page);
187         io_schedule();
188         return 0;
189 }
190
191 static int sync_page_killable(void *word)
192 {
193         sync_page(word);
194         return fatal_signal_pending(current) ? -EINTR : 0;
195 }
196
197 /**
198  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
199  * @mapping:    address space structure to write
200  * @start:      offset in bytes where the range starts
201  * @end:        offset in bytes where the range ends (inclusive)
202  * @sync_mode:  enable synchronous operation
203  *
204  * Start writeback against all of a mapping's dirty pages that lie
205  * within the byte offsets <start, end> inclusive.
206  *
207  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
208  * opposed to a regular memory cleansing writeback.  The difference between
209  * these two operations is that if a dirty page/buffer is encountered, it must
210  * be waited upon, and not just skipped over.
211  */
212 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
213                                 loff_t end, int sync_mode)
214 {
215         int ret;
216         struct writeback_control wbc = {
217                 .sync_mode = sync_mode,
218                 .nr_to_write = LONG_MAX,
219                 .range_start = start,
220                 .range_end = end,
221         };
222
223         if (!mapping_cap_writeback_dirty(mapping))
224                 return 0;
225
226         ret = do_writepages(mapping, &wbc);
227         return ret;
228 }
229
230 static inline int __filemap_fdatawrite(struct address_space *mapping,
231         int sync_mode)
232 {
233         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
234 }
235
236 int filemap_fdatawrite(struct address_space *mapping)
237 {
238         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
239 }
240 EXPORT_SYMBOL(filemap_fdatawrite);
241
242 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
243                                 loff_t end)
244 {
245         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
246 }
247 EXPORT_SYMBOL(filemap_fdatawrite_range);
248
249 /**
250  * filemap_flush - mostly a non-blocking flush
251  * @mapping:    target address_space
252  *
253  * This is a mostly non-blocking flush.  Not suitable for data-integrity
254  * purposes - I/O may not be started against all dirty pages.
255  */
256 int filemap_flush(struct address_space *mapping)
257 {
258         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
259 }
260 EXPORT_SYMBOL(filemap_flush);
261
262 /**
263  * wait_on_page_writeback_range - wait for writeback to complete
264  * @mapping:    target address_space
265  * @start:      beginning page index
266  * @end:        ending page index
267  *
268  * Wait for writeback to complete against pages indexed by start->end
269  * inclusive
270  */
271 int wait_on_page_writeback_range(struct address_space *mapping,
272                                 pgoff_t start, pgoff_t end)
273 {
274         struct pagevec pvec;
275         int nr_pages;
276         int ret = 0;
277         pgoff_t index;
278
279         if (end < start)
280                 return 0;
281
282         pagevec_init(&pvec, 0);
283         index = start;
284         while ((index <= end) &&
285                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
286                         PAGECACHE_TAG_WRITEBACK,
287                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
288                 unsigned i;
289
290                 for (i = 0; i < nr_pages; i++) {
291                         struct page *page = pvec.pages[i];
292
293                         /* until radix tree lookup accepts end_index */
294                         if (page->index > end)
295                                 continue;
296
297                         wait_on_page_writeback(page);
298                         if (PageError(page))
299                                 ret = -EIO;
300                 }
301                 pagevec_release(&pvec);
302                 cond_resched();
303         }
304
305         /* Check for outstanding write errors */
306         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
307                 ret = -ENOSPC;
308         if (test_and_clear_bit(AS_EIO, &mapping->flags))
309                 ret = -EIO;
310
311         return ret;
312 }
313
314 /**
315  * filemap_fdatawait_range - wait for all under-writeback pages to complete in a given range
316  * @mapping: address space structure to wait for
317  * @start:      offset in bytes where the range starts
318  * @end:        offset in bytes where the range ends (inclusive)
319  *
320  * Walk the list of under-writeback pages of the given address space
321  * in the given range and wait for all of them.
322  *
323  * This is just a simple wrapper so that callers don't have to convert offsets
324  * to page indexes themselves
325  */
326 int filemap_fdatawait_range(struct address_space *mapping, loff_t start,
327                             loff_t end)
328 {
329         return wait_on_page_writeback_range(mapping, start >> PAGE_CACHE_SHIFT,
330                                             end >> PAGE_CACHE_SHIFT);
331 }
332 EXPORT_SYMBOL(filemap_fdatawait_range);
333
334 /**
335  * filemap_fdatawait - wait for all under-writeback pages to complete
336  * @mapping: address space structure to wait for
337  *
338  * Walk the list of under-writeback pages of the given address space
339  * and wait for all of them.
340  */
341 int filemap_fdatawait(struct address_space *mapping)
342 {
343         loff_t i_size = i_size_read(mapping->host);
344
345         if (i_size == 0)
346                 return 0;
347
348         return wait_on_page_writeback_range(mapping, 0,
349                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
350 }
351 EXPORT_SYMBOL(filemap_fdatawait);
352
353 int filemap_write_and_wait(struct address_space *mapping)
354 {
355         int err = 0;
356
357         if (mapping->nrpages) {
358                 err = filemap_fdatawrite(mapping);
359                 /*
360                  * Even if the above returned error, the pages may be
361                  * written partially (e.g. -ENOSPC), so we wait for it.
362                  * But the -EIO is special case, it may indicate the worst
363                  * thing (e.g. bug) happened, so we avoid waiting for it.
364                  */
365                 if (err != -EIO) {
366                         int err2 = filemap_fdatawait(mapping);
367                         if (!err)
368                                 err = err2;
369                 }
370         }
371         return err;
372 }
373 EXPORT_SYMBOL(filemap_write_and_wait);
374
375 /**
376  * filemap_write_and_wait_range - write out & wait on a file range
377  * @mapping:    the address_space for the pages
378  * @lstart:     offset in bytes where the range starts
379  * @lend:       offset in bytes where the range ends (inclusive)
380  *
381  * Write out and wait upon file offsets lstart->lend, inclusive.
382  *
383  * Note that `lend' is inclusive (describes the last byte to be written) so
384  * that this function can be used to write to the very end-of-file (end = -1).
385  */
386 int filemap_write_and_wait_range(struct address_space *mapping,
387                                  loff_t lstart, loff_t lend)
388 {
389         int err = 0;
390
391         if (mapping->nrpages) {
392                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
393                                                  WB_SYNC_ALL);
394                 /* See comment of filemap_write_and_wait() */
395                 if (err != -EIO) {
396                         int err2 = wait_on_page_writeback_range(mapping,
397                                                 lstart >> PAGE_CACHE_SHIFT,
398                                                 lend >> PAGE_CACHE_SHIFT);
399                         if (!err)
400                                 err = err2;
401                 }
402         }
403         return err;
404 }
405 EXPORT_SYMBOL(filemap_write_and_wait_range);
406
407 /**
408  * add_to_page_cache_locked - add a locked page to the pagecache
409  * @page:       page to add
410  * @mapping:    the page's address_space
411  * @offset:     page index
412  * @gfp_mask:   page allocation mode
413  *
414  * This function is used to add a page to the pagecache. It must be locked.
415  * This function does not add the page to the LRU.  The caller must do that.
416  */
417 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
418                 pgoff_t offset, gfp_t gfp_mask)
419 {
420         int error;
421
422         VM_BUG_ON(!PageLocked(page));
423
424         error = mem_cgroup_cache_charge(page, current->mm,
425                                         gfp_mask & GFP_RECLAIM_MASK);
426         if (error)
427                 goto out;
428
429         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
430         if (error == 0) {
431                 page_cache_get(page);
432                 page->mapping = mapping;
433                 page->index = offset;
434
435                 spin_lock_irq(&mapping->tree_lock);
436                 error = radix_tree_insert(&mapping->page_tree, offset, page);
437                 if (likely(!error)) {
438                         mapping->nrpages++;
439                         __inc_zone_page_state(page, NR_FILE_PAGES);
440                         if (PageSwapBacked(page))
441                                 __inc_zone_page_state(page, NR_SHMEM);
442                         spin_unlock_irq(&mapping->tree_lock);
443                 } else {
444                         page->mapping = NULL;
445                         spin_unlock_irq(&mapping->tree_lock);
446                         mem_cgroup_uncharge_cache_page(page);
447                         page_cache_release(page);
448                 }
449                 radix_tree_preload_end();
450         } else
451                 mem_cgroup_uncharge_cache_page(page);
452 out:
453         return error;
454 }
455 EXPORT_SYMBOL(add_to_page_cache_locked);
456
457 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
458                                 pgoff_t offset, gfp_t gfp_mask)
459 {
460         int ret;
461
462         /*
463          * Splice_read and readahead add shmem/tmpfs pages into the page cache
464          * before shmem_readpage has a chance to mark them as SwapBacked: they
465          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
466          * (called in add_to_page_cache) needs to know where they're going too.
467          */
468         if (mapping_cap_swap_backed(mapping))
469                 SetPageSwapBacked(page);
470
471         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
472         if (ret == 0) {
473                 if (page_is_file_cache(page))
474                         lru_cache_add_file(page);
475                 else
476                         lru_cache_add_active_anon(page);
477         }
478         return ret;
479 }
480 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
481
482 #ifdef CONFIG_NUMA
483 struct page *__page_cache_alloc(gfp_t gfp)
484 {
485         if (cpuset_do_page_mem_spread()) {
486                 int n = cpuset_mem_spread_node();
487                 return alloc_pages_exact_node(n, gfp, 0);
488         }
489         return alloc_pages(gfp, 0);
490 }
491 EXPORT_SYMBOL(__page_cache_alloc);
492 #endif
493
494 static int __sleep_on_page_lock(void *word)
495 {
496         io_schedule();
497         return 0;
498 }
499
500 /*
501  * In order to wait for pages to become available there must be
502  * waitqueues associated with pages. By using a hash table of
503  * waitqueues where the bucket discipline is to maintain all
504  * waiters on the same queue and wake all when any of the pages
505  * become available, and for the woken contexts to check to be
506  * sure the appropriate page became available, this saves space
507  * at a cost of "thundering herd" phenomena during rare hash
508  * collisions.
509  */
510 static wait_queue_head_t *page_waitqueue(struct page *page)
511 {
512         const struct zone *zone = page_zone(page);
513
514         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
515 }
516
517 static inline void wake_up_page(struct page *page, int bit)
518 {
519         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
520 }
521
522 void wait_on_page_bit(struct page *page, int bit_nr)
523 {
524         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
525
526         if (test_bit(bit_nr, &page->flags))
527                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
528                                                         TASK_UNINTERRUPTIBLE);
529 }
530 EXPORT_SYMBOL(wait_on_page_bit);
531
532 /**
533  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
534  * @page: Page defining the wait queue of interest
535  * @waiter: Waiter to add to the queue
536  *
537  * Add an arbitrary @waiter to the wait queue for the nominated @page.
538  */
539 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
540 {
541         wait_queue_head_t *q = page_waitqueue(page);
542         unsigned long flags;
543
544         spin_lock_irqsave(&q->lock, flags);
545         __add_wait_queue(q, waiter);
546         spin_unlock_irqrestore(&q->lock, flags);
547 }
548 EXPORT_SYMBOL_GPL(add_page_wait_queue);
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 mb is necessary to enforce ordering between the clear_bit and the read
560  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
561  */
562 void unlock_page(struct page *page)
563 {
564         VM_BUG_ON(!PageLocked(page));
565         clear_bit_unlock(PG_locked, &page->flags);
566         smp_mb__after_clear_bit();
567         wake_up_page(page, PG_locked);
568 }
569 EXPORT_SYMBOL(unlock_page);
570
571 /**
572  * end_page_writeback - end writeback against a page
573  * @page: the page
574  */
575 void end_page_writeback(struct page *page)
576 {
577         if (TestClearPageReclaim(page))
578                 rotate_reclaimable_page(page);
579
580         if (!test_clear_page_writeback(page))
581                 BUG();
582
583         smp_mb__after_clear_bit();
584         wake_up_page(page, PG_writeback);
585 }
586 EXPORT_SYMBOL(end_page_writeback);
587
588 /**
589  * __lock_page - get a lock on the page, assuming we need to sleep to get it
590  * @page: the page to lock
591  *
592  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
593  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
594  * chances are that on the second loop, the block layer's plug list is empty,
595  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
596  */
597 void __lock_page(struct page *page)
598 {
599         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
600
601         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
602                                                         TASK_UNINTERRUPTIBLE);
603 }
604 EXPORT_SYMBOL(__lock_page);
605
606 int __lock_page_killable(struct page *page)
607 {
608         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
609
610         return __wait_on_bit_lock(page_waitqueue(page), &wait,
611                                         sync_page_killable, TASK_KILLABLE);
612 }
613 EXPORT_SYMBOL_GPL(__lock_page_killable);
614
615 /**
616  * __lock_page_nosync - get a lock on the page, without calling sync_page()
617  * @page: the page to lock
618  *
619  * Variant of lock_page that does not require the caller to hold a reference
620  * on the page's mapping.
621  */
622 void __lock_page_nosync(struct page *page)
623 {
624         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
625         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
626                                                         TASK_UNINTERRUPTIBLE);
627 }
628
629 /**
630  * find_get_page - find and get a page reference
631  * @mapping: the address_space to search
632  * @offset: the page index
633  *
634  * Is there a pagecache struct page at the given (mapping, offset) tuple?
635  * If yes, increment its refcount and return it; if no, return NULL.
636  */
637 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
638 {
639         void **pagep;
640         struct page *page;
641
642         rcu_read_lock();
643 repeat:
644         page = NULL;
645         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
646         if (pagep) {
647                 page = radix_tree_deref_slot(pagep);
648                 if (unlikely(!page || page == RADIX_TREE_RETRY))
649                         goto repeat;
650
651                 if (!page_cache_get_speculative(page))
652                         goto repeat;
653
654                 /*
655                  * Has the page moved?
656                  * This is part of the lockless pagecache protocol. See
657                  * include/linux/pagemap.h for details.
658                  */
659                 if (unlikely(page != *pagep)) {
660                         page_cache_release(page);
661                         goto repeat;
662                 }
663         }
664         rcu_read_unlock();
665
666         return page;
667 }
668 EXPORT_SYMBOL(find_get_page);
669
670 /**
671  * find_lock_page - locate, pin and lock a pagecache page
672  * @mapping: the address_space to search
673  * @offset: the page index
674  *
675  * Locates the desired pagecache page, locks it, increments its reference
676  * count and returns its address.
677  *
678  * Returns zero if the page was not present. find_lock_page() may sleep.
679  */
680 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
681 {
682         struct page *page;
683
684 repeat:
685         page = find_get_page(mapping, offset);
686         if (page) {
687                 lock_page(page);
688                 /* Has the page been truncated? */
689                 if (unlikely(page->mapping != mapping)) {
690                         unlock_page(page);
691                         page_cache_release(page);
692                         goto repeat;
693                 }
694                 VM_BUG_ON(page->index != offset);
695         }
696         return page;
697 }
698 EXPORT_SYMBOL(find_lock_page);
699
700 /**
701  * find_or_create_page - locate or add a pagecache page
702  * @mapping: the page's address_space
703  * @index: the page's index into the mapping
704  * @gfp_mask: page allocation mode
705  *
706  * Locates a page in the pagecache.  If the page is not present, a new page
707  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
708  * LRU list.  The returned page is locked and has its reference count
709  * incremented.
710  *
711  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
712  * allocation!
713  *
714  * find_or_create_page() returns the desired page's address, or zero on
715  * memory exhaustion.
716  */
717 struct page *find_or_create_page(struct address_space *mapping,
718                 pgoff_t index, gfp_t gfp_mask)
719 {
720         struct page *page;
721         int err;
722 repeat:
723         page = find_lock_page(mapping, index);
724         if (!page) {
725                 page = __page_cache_alloc(gfp_mask);
726                 if (!page)
727                         return NULL;
728                 /*
729                  * We want a regular kernel memory (not highmem or DMA etc)
730                  * allocation for the radix tree nodes, but we need to honour
731                  * the context-specific requirements the caller has asked for.
732                  * GFP_RECLAIM_MASK collects those requirements.
733                  */
734                 err = add_to_page_cache_lru(page, mapping, index,
735                         (gfp_mask & GFP_RECLAIM_MASK));
736                 if (unlikely(err)) {
737                         page_cache_release(page);
738                         page = NULL;
739                         if (err == -EEXIST)
740                                 goto repeat;
741                 }
742         }
743         return page;
744 }
745 EXPORT_SYMBOL(find_or_create_page);
746
747 /**
748  * find_get_pages - gang pagecache lookup
749  * @mapping:    The address_space to search
750  * @start:      The starting page index
751  * @nr_pages:   The maximum number of pages
752  * @pages:      Where the resulting pages are placed
753  *
754  * find_get_pages() will search for and return a group of up to
755  * @nr_pages pages in the mapping.  The pages are placed at @pages.
756  * find_get_pages() takes a reference against the returned pages.
757  *
758  * The search returns a group of mapping-contiguous pages with ascending
759  * indexes.  There may be holes in the indices due to not-present pages.
760  *
761  * find_get_pages() returns the number of pages which were found.
762  */
763 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
764                             unsigned int nr_pages, struct page **pages)
765 {
766         unsigned int i;
767         unsigned int ret;
768         unsigned int nr_found;
769
770         rcu_read_lock();
771 restart:
772         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
773                                 (void ***)pages, start, nr_pages);
774         ret = 0;
775         for (i = 0; i < nr_found; i++) {
776                 struct page *page;
777 repeat:
778                 page = radix_tree_deref_slot((void **)pages[i]);
779                 if (unlikely(!page))
780                         continue;
781                 /*
782                  * this can only trigger if nr_found == 1, making livelock
783                  * a non issue.
784                  */
785                 if (unlikely(page == RADIX_TREE_RETRY))
786                         goto restart;
787
788                 if (!page_cache_get_speculative(page))
789                         goto repeat;
790
791                 /* Has the page moved? */
792                 if (unlikely(page != *((void **)pages[i]))) {
793                         page_cache_release(page);
794                         goto repeat;
795                 }
796
797                 pages[ret] = page;
798                 ret++;
799         }
800         rcu_read_unlock();
801         return ret;
802 }
803
804 /**
805  * find_get_pages_contig - gang contiguous pagecache lookup
806  * @mapping:    The address_space to search
807  * @index:      The starting page index
808  * @nr_pages:   The maximum number of pages
809  * @pages:      Where the resulting pages are placed
810  *
811  * find_get_pages_contig() works exactly like find_get_pages(), except
812  * that the returned number of pages are guaranteed to be contiguous.
813  *
814  * find_get_pages_contig() returns the number of pages which were found.
815  */
816 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
817                                unsigned int nr_pages, struct page **pages)
818 {
819         unsigned int i;
820         unsigned int ret;
821         unsigned int nr_found;
822
823         rcu_read_lock();
824 restart:
825         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
826                                 (void ***)pages, index, nr_pages);
827         ret = 0;
828         for (i = 0; i < nr_found; i++) {
829                 struct page *page;
830 repeat:
831                 page = radix_tree_deref_slot((void **)pages[i]);
832                 if (unlikely(!page))
833                         continue;
834                 /*
835                  * this can only trigger if nr_found == 1, making livelock
836                  * a non issue.
837                  */
838                 if (unlikely(page == RADIX_TREE_RETRY))
839                         goto restart;
840
841                 if (page->mapping == NULL || page->index != index)
842                         break;
843
844                 if (!page_cache_get_speculative(page))
845                         goto repeat;
846
847                 /* Has the page moved? */
848                 if (unlikely(page != *((void **)pages[i]))) {
849                         page_cache_release(page);
850                         goto repeat;
851                 }
852
853                 pages[ret] = page;
854                 ret++;
855                 index++;
856         }
857         rcu_read_unlock();
858         return ret;
859 }
860 EXPORT_SYMBOL(find_get_pages_contig);
861
862 /**
863  * find_get_pages_tag - find and return pages that match @tag
864  * @mapping:    the address_space to search
865  * @index:      the starting page index
866  * @tag:        the tag index
867  * @nr_pages:   the maximum number of pages
868  * @pages:      where the resulting pages are placed
869  *
870  * Like find_get_pages, except we only return pages which are tagged with
871  * @tag.   We update @index to index the next page for the traversal.
872  */
873 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
874                         int tag, unsigned int nr_pages, struct page **pages)
875 {
876         unsigned int i;
877         unsigned int ret;
878         unsigned int nr_found;
879
880         rcu_read_lock();
881 restart:
882         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
883                                 (void ***)pages, *index, nr_pages, tag);
884         ret = 0;
885         for (i = 0; i < nr_found; i++) {
886                 struct page *page;
887 repeat:
888                 page = radix_tree_deref_slot((void **)pages[i]);
889                 if (unlikely(!page))
890                         continue;
891                 /*
892                  * this can only trigger if nr_found == 1, making livelock
893                  * a non issue.
894                  */
895                 if (unlikely(page == RADIX_TREE_RETRY))
896                         goto restart;
897
898                 if (!page_cache_get_speculative(page))
899                         goto repeat;
900
901                 /* Has the page moved? */
902                 if (unlikely(page != *((void **)pages[i]))) {
903                         page_cache_release(page);
904                         goto repeat;
905                 }
906
907                 pages[ret] = page;
908                 ret++;
909         }
910         rcu_read_unlock();
911
912         if (ret)
913                 *index = pages[ret - 1]->index + 1;
914
915         return ret;
916 }
917 EXPORT_SYMBOL(find_get_pages_tag);
918
919 /**
920  * grab_cache_page_nowait - returns locked page at given index in given cache
921  * @mapping: target address_space
922  * @index: the page index
923  *
924  * Same as grab_cache_page(), but do not wait if the page is unavailable.
925  * This is intended for speculative data generators, where the data can
926  * be regenerated if the page couldn't be grabbed.  This routine should
927  * be safe to call while holding the lock for another page.
928  *
929  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
930  * and deadlock against the caller's locked page.
931  */
932 struct page *
933 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
934 {
935         struct page *page = find_get_page(mapping, index);
936
937         if (page) {
938                 if (trylock_page(page))
939                         return page;
940                 page_cache_release(page);
941                 return NULL;
942         }
943         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
944         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
945                 page_cache_release(page);
946                 page = NULL;
947         }
948         return page;
949 }
950 EXPORT_SYMBOL(grab_cache_page_nowait);
951
952 /*
953  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
954  * a _large_ part of the i/o request. Imagine the worst scenario:
955  *
956  *      ---R__________________________________________B__________
957  *         ^ reading here                             ^ bad block(assume 4k)
958  *
959  * read(R) => miss => readahead(R...B) => media error => frustrating retries
960  * => failing the whole request => read(R) => read(R+1) =>
961  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
962  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
963  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
964  *
965  * It is going insane. Fix it by quickly scaling down the readahead size.
966  */
967 static void shrink_readahead_size_eio(struct file *filp,
968                                         struct file_ra_state *ra)
969 {
970         ra->ra_pages /= 4;
971 }
972
973 /**
974  * do_generic_file_read - generic file read routine
975  * @filp:       the file to read
976  * @ppos:       current file position
977  * @desc:       read_descriptor
978  * @actor:      read method
979  *
980  * This is a generic file read routine, and uses the
981  * mapping->a_ops->readpage() function for the actual low-level stuff.
982  *
983  * This is really ugly. But the goto's actually try to clarify some
984  * of the logic when it comes to error handling etc.
985  */
986 static void do_generic_file_read(struct file *filp, loff_t *ppos,
987                 read_descriptor_t *desc, read_actor_t actor)
988 {
989         struct address_space *mapping = filp->f_mapping;
990         struct inode *inode = mapping->host;
991         struct file_ra_state *ra = &filp->f_ra;
992         pgoff_t index;
993         pgoff_t last_index;
994         pgoff_t prev_index;
995         unsigned long offset;      /* offset into pagecache page */
996         unsigned int prev_offset;
997         int error;
998
999         index = *ppos >> PAGE_CACHE_SHIFT;
1000         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1001         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1002         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1003         offset = *ppos & ~PAGE_CACHE_MASK;
1004
1005         for (;;) {
1006                 struct page *page;
1007                 pgoff_t end_index;
1008                 loff_t isize;
1009                 unsigned long nr, ret;
1010
1011                 cond_resched();
1012 find_page:
1013                 page = find_get_page(mapping, index);
1014                 if (!page) {
1015                         page_cache_sync_readahead(mapping,
1016                                         ra, filp,
1017                                         index, last_index - index);
1018                         page = find_get_page(mapping, index);
1019                         if (unlikely(page == NULL))
1020                                 goto no_cached_page;
1021                 }
1022                 if (PageReadahead(page)) {
1023                         page_cache_async_readahead(mapping,
1024                                         ra, filp, page,
1025                                         index, last_index - index);
1026                 }
1027                 if (!PageUptodate(page)) {
1028                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1029                                         !mapping->a_ops->is_partially_uptodate)
1030                                 goto page_not_up_to_date;
1031                         if (!trylock_page(page))
1032                                 goto page_not_up_to_date;
1033                         if (!mapping->a_ops->is_partially_uptodate(page,
1034                                                                 desc, offset))
1035                                 goto page_not_up_to_date_locked;
1036                         unlock_page(page);
1037                 }
1038 page_ok:
1039                 /*
1040                  * i_size must be checked after we know the page is Uptodate.
1041                  *
1042                  * Checking i_size after the check allows us to calculate
1043                  * the correct value for "nr", which means the zero-filled
1044                  * part of the page is not copied back to userspace (unless
1045                  * another truncate extends the file - this is desired though).
1046                  */
1047
1048                 isize = i_size_read(inode);
1049                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1050                 if (unlikely(!isize || index > end_index)) {
1051                         page_cache_release(page);
1052                         goto out;
1053                 }
1054
1055                 /* nr is the maximum number of bytes to copy from this page */
1056                 nr = PAGE_CACHE_SIZE;
1057                 if (index == end_index) {
1058                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1059                         if (nr <= offset) {
1060                                 page_cache_release(page);
1061                                 goto out;
1062                         }
1063                 }
1064                 nr = nr - offset;
1065
1066                 /* If users can be writing to this page using arbitrary
1067                  * virtual addresses, take care about potential aliasing
1068                  * before reading the page on the kernel side.
1069                  */
1070                 if (mapping_writably_mapped(mapping))
1071                         flush_dcache_page(page);
1072
1073                 /*
1074                  * When a sequential read accesses a page several times,
1075                  * only mark it as accessed the first time.
1076                  */
1077                 if (prev_index != index || offset != prev_offset)
1078                         mark_page_accessed(page);
1079                 prev_index = index;
1080
1081                 /*
1082                  * Ok, we have the page, and it's up-to-date, so
1083                  * now we can copy it to user space...
1084                  *
1085                  * The actor routine returns how many bytes were actually used..
1086                  * NOTE! This may not be the same as how much of a user buffer
1087                  * we filled up (we may be padding etc), so we can only update
1088                  * "pos" here (the actor routine has to update the user buffer
1089                  * pointers and the remaining count).
1090                  */
1091                 ret = actor(desc, page, offset, nr);
1092                 offset += ret;
1093                 index += offset >> PAGE_CACHE_SHIFT;
1094                 offset &= ~PAGE_CACHE_MASK;
1095                 prev_offset = offset;
1096
1097                 page_cache_release(page);
1098                 if (ret == nr && desc->count)
1099                         continue;
1100                 goto out;
1101
1102 page_not_up_to_date:
1103                 /* Get exclusive access to the page ... */
1104                 error = lock_page_killable(page);
1105                 if (unlikely(error))
1106                         goto readpage_error;
1107
1108 page_not_up_to_date_locked:
1109                 /* Did it get truncated before we got the lock? */
1110                 if (!page->mapping) {
1111                         unlock_page(page);
1112                         page_cache_release(page);
1113                         continue;
1114                 }
1115
1116                 /* Did somebody else fill it already? */
1117                 if (PageUptodate(page)) {
1118                         unlock_page(page);
1119                         goto page_ok;
1120                 }
1121
1122 readpage:
1123                 /* Start the actual read. The read will unlock the page. */
1124                 error = mapping->a_ops->readpage(filp, page);
1125
1126                 if (unlikely(error)) {
1127                         if (error == AOP_TRUNCATED_PAGE) {
1128                                 page_cache_release(page);
1129                                 goto find_page;
1130                         }
1131                         goto readpage_error;
1132                 }
1133
1134                 if (!PageUptodate(page)) {
1135                         error = lock_page_killable(page);
1136                         if (unlikely(error))
1137                                 goto readpage_error;
1138                         if (!PageUptodate(page)) {
1139                                 if (page->mapping == NULL) {
1140                                         /*
1141                                          * invalidate_inode_pages got it
1142                                          */
1143                                         unlock_page(page);
1144                                         page_cache_release(page);
1145                                         goto find_page;
1146                                 }
1147                                 unlock_page(page);
1148                                 shrink_readahead_size_eio(filp, ra);
1149                                 error = -EIO;
1150                                 goto readpage_error;
1151                         }
1152                         unlock_page(page);
1153                 }
1154
1155                 goto page_ok;
1156
1157 readpage_error:
1158                 /* UHHUH! A synchronous read error occurred. Report it */
1159                 desc->error = error;
1160                 page_cache_release(page);
1161                 goto out;
1162
1163 no_cached_page:
1164                 /*
1165                  * Ok, it wasn't cached, so we need to create a new
1166                  * page..
1167                  */
1168                 page = page_cache_alloc_cold(mapping);
1169                 if (!page) {
1170                         desc->error = -ENOMEM;
1171                         goto out;
1172                 }
1173                 error = add_to_page_cache_lru(page, mapping,
1174                                                 index, GFP_KERNEL);
1175                 if (error) {
1176                         page_cache_release(page);
1177                         if (error == -EEXIST)
1178                                 goto find_page;
1179                         desc->error = error;
1180                         goto out;
1181                 }
1182                 goto readpage;
1183         }
1184
1185 out:
1186         ra->prev_pos = prev_index;
1187         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1188         ra->prev_pos |= prev_offset;
1189
1190         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1191         file_accessed(filp);
1192 }
1193
1194 int file_read_actor(read_descriptor_t *desc, struct page *page,
1195                         unsigned long offset, unsigned long size)
1196 {
1197         char *kaddr;
1198         unsigned long left, count = desc->count;
1199
1200         if (size > count)
1201                 size = count;
1202
1203         /*
1204          * Faults on the destination of a read are common, so do it before
1205          * taking the kmap.
1206          */
1207         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1208                 kaddr = kmap_atomic(page, KM_USER0);
1209                 left = __copy_to_user_inatomic(desc->arg.buf,
1210                                                 kaddr + offset, size);
1211                 kunmap_atomic(kaddr, KM_USER0);
1212                 if (left == 0)
1213                         goto success;
1214         }
1215
1216         /* Do it the slow way */
1217         kaddr = kmap(page);
1218         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1219         kunmap(page);
1220
1221         if (left) {
1222                 size -= left;
1223                 desc->error = -EFAULT;
1224         }
1225 success:
1226         desc->count = count - size;
1227         desc->written += size;
1228         desc->arg.buf += size;
1229         return size;
1230 }
1231
1232 /*
1233  * Performs necessary checks before doing a write
1234  * @iov:        io vector request
1235  * @nr_segs:    number of segments in the iovec
1236  * @count:      number of bytes to write
1237  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1238  *
1239  * Adjust number of segments and amount of bytes to write (nr_segs should be
1240  * properly initialized first). Returns appropriate error code that caller
1241  * should return or zero in case that write should be allowed.
1242  */
1243 int generic_segment_checks(const struct iovec *iov,
1244                         unsigned long *nr_segs, size_t *count, int access_flags)
1245 {
1246         unsigned long   seg;
1247         size_t cnt = 0;
1248         for (seg = 0; seg < *nr_segs; seg++) {
1249                 const struct iovec *iv = &iov[seg];
1250
1251                 /*
1252                  * If any segment has a negative length, or the cumulative
1253                  * length ever wraps negative then return -EINVAL.
1254                  */
1255                 cnt += iv->iov_len;
1256                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1257                         return -EINVAL;
1258                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1259                         continue;
1260                 if (seg == 0)
1261                         return -EFAULT;
1262                 *nr_segs = seg;
1263                 cnt -= iv->iov_len;     /* This segment is no good */
1264                 break;
1265         }
1266         *count = cnt;
1267         return 0;
1268 }
1269 EXPORT_SYMBOL(generic_segment_checks);
1270
1271 /**
1272  * generic_file_aio_read - generic filesystem read routine
1273  * @iocb:       kernel I/O control block
1274  * @iov:        io vector request
1275  * @nr_segs:    number of segments in the iovec
1276  * @pos:        current file position
1277  *
1278  * This is the "read()" routine for all filesystems
1279  * that can use the page cache directly.
1280  */
1281 ssize_t
1282 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1283                 unsigned long nr_segs, loff_t pos)
1284 {
1285         struct file *filp = iocb->ki_filp;
1286         ssize_t retval;
1287         unsigned long seg;
1288         size_t count;
1289         loff_t *ppos = &iocb->ki_pos;
1290
1291         count = 0;
1292         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1293         if (retval)
1294                 return retval;
1295
1296         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1297         if (filp->f_flags & O_DIRECT) {
1298                 loff_t size;
1299                 struct address_space *mapping;
1300                 struct inode *inode;
1301
1302                 mapping = filp->f_mapping;
1303                 inode = mapping->host;
1304                 if (!count)
1305                         goto out; /* skip atime */
1306                 size = i_size_read(inode);
1307                 if (pos < size) {
1308                         retval = filemap_write_and_wait_range(mapping, pos,
1309                                         pos + iov_length(iov, nr_segs) - 1);
1310                         if (!retval) {
1311                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1312                                                         iov, pos, nr_segs);
1313                         }
1314                         if (retval > 0)
1315                                 *ppos = pos + retval;
1316                         if (retval) {
1317                                 file_accessed(filp);
1318                                 goto out;
1319                         }
1320                 }
1321         }
1322
1323         for (seg = 0; seg < nr_segs; seg++) {
1324                 read_descriptor_t desc;
1325
1326                 desc.written = 0;
1327                 desc.arg.buf = iov[seg].iov_base;
1328                 desc.count = iov[seg].iov_len;
1329                 if (desc.count == 0)
1330                         continue;
1331                 desc.error = 0;
1332                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1333                 retval += desc.written;
1334                 if (desc.error) {
1335                         retval = retval ?: desc.error;
1336                         break;
1337                 }
1338                 if (desc.count > 0)
1339                         break;
1340         }
1341 out:
1342         return retval;
1343 }
1344 EXPORT_SYMBOL(generic_file_aio_read);
1345
1346 static ssize_t
1347 do_readahead(struct address_space *mapping, struct file *filp,
1348              pgoff_t index, unsigned long nr)
1349 {
1350         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1351                 return -EINVAL;
1352
1353         force_page_cache_readahead(mapping, filp, index, nr);
1354         return 0;
1355 }
1356
1357 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1358 {
1359         ssize_t ret;
1360         struct file *file;
1361
1362         ret = -EBADF;
1363         file = fget(fd);
1364         if (file) {
1365                 if (file->f_mode & FMODE_READ) {
1366                         struct address_space *mapping = file->f_mapping;
1367                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1368                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1369                         unsigned long len = end - start + 1;
1370                         ret = do_readahead(mapping, file, start, len);
1371                 }
1372                 fput(file);
1373         }
1374         return ret;
1375 }
1376 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1377 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1378 {
1379         return SYSC_readahead((int) fd, offset, (size_t) count);
1380 }
1381 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1382 #endif
1383
1384 #ifdef CONFIG_MMU
1385 /**
1386  * page_cache_read - adds requested page to the page cache if not already there
1387  * @file:       file to read
1388  * @offset:     page index
1389  *
1390  * This adds the requested page to the page cache if it isn't already there,
1391  * and schedules an I/O to read in its contents from disk.
1392  */
1393 static int page_cache_read(struct file *file, pgoff_t offset)
1394 {
1395         struct address_space *mapping = file->f_mapping;
1396         struct page *page; 
1397         int ret;
1398
1399         do {
1400                 page = page_cache_alloc_cold(mapping);
1401                 if (!page)
1402                         return -ENOMEM;
1403
1404                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1405                 if (ret == 0)
1406                         ret = mapping->a_ops->readpage(file, page);
1407                 else if (ret == -EEXIST)
1408                         ret = 0; /* losing race to add is OK */
1409
1410                 page_cache_release(page);
1411
1412         } while (ret == AOP_TRUNCATED_PAGE);
1413                 
1414         return ret;
1415 }
1416
1417 #define MMAP_LOTSAMISS  (100)
1418
1419 /*
1420  * Synchronous readahead happens when we don't even find
1421  * a page in the page cache at all.
1422  */
1423 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1424                                    struct file_ra_state *ra,
1425                                    struct file *file,
1426                                    pgoff_t offset)
1427 {
1428         unsigned long ra_pages;
1429         struct address_space *mapping = file->f_mapping;
1430
1431         /* If we don't want any read-ahead, don't bother */
1432         if (VM_RandomReadHint(vma))
1433                 return;
1434
1435         if (VM_SequentialReadHint(vma) ||
1436                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1437                 page_cache_sync_readahead(mapping, ra, file, offset,
1438                                           ra->ra_pages);
1439                 return;
1440         }
1441
1442         if (ra->mmap_miss < INT_MAX)
1443                 ra->mmap_miss++;
1444
1445         /*
1446          * Do we miss much more than hit in this file? If so,
1447          * stop bothering with read-ahead. It will only hurt.
1448          */
1449         if (ra->mmap_miss > MMAP_LOTSAMISS)
1450                 return;
1451
1452         /*
1453          * mmap read-around
1454          */
1455         ra_pages = max_sane_readahead(ra->ra_pages);
1456         if (ra_pages) {
1457                 ra->start = max_t(long, 0, offset - ra_pages/2);
1458                 ra->size = ra_pages;
1459                 ra->async_size = 0;
1460                 ra_submit(ra, mapping, file);
1461         }
1462 }
1463
1464 /*
1465  * Asynchronous readahead happens when we find the page and PG_readahead,
1466  * so we want to possibly extend the readahead further..
1467  */
1468 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1469                                     struct file_ra_state *ra,
1470                                     struct file *file,
1471                                     struct page *page,
1472                                     pgoff_t offset)
1473 {
1474         struct address_space *mapping = file->f_mapping;
1475
1476         /* If we don't want any read-ahead, don't bother */
1477         if (VM_RandomReadHint(vma))
1478                 return;
1479         if (ra->mmap_miss > 0)
1480                 ra->mmap_miss--;
1481         if (PageReadahead(page))
1482                 page_cache_async_readahead(mapping, ra, file,
1483                                            page, offset, ra->ra_pages);
1484 }
1485
1486 /**
1487  * filemap_fault - read in file data for page fault handling
1488  * @vma:        vma in which the fault was taken
1489  * @vmf:        struct vm_fault containing details of the fault
1490  *
1491  * filemap_fault() is invoked via the vma operations vector for a
1492  * mapped memory region to read in file data during a page fault.
1493  *
1494  * The goto's are kind of ugly, but this streamlines the normal case of having
1495  * it in the page cache, and handles the special cases reasonably without
1496  * having a lot of duplicated code.
1497  */
1498 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1499 {
1500         int error;
1501         struct file *file = vma->vm_file;
1502         struct address_space *mapping = file->f_mapping;
1503         struct file_ra_state *ra = &file->f_ra;
1504         struct inode *inode = mapping->host;
1505         pgoff_t offset = vmf->pgoff;
1506         struct page *page;
1507         pgoff_t size;
1508         int ret = 0;
1509
1510         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1511         if (offset >= size)
1512                 return VM_FAULT_SIGBUS;
1513
1514         /*
1515          * Do we have something in the page cache already?
1516          */
1517         page = find_get_page(mapping, offset);
1518         if (likely(page)) {
1519                 /*
1520                  * We found the page, so try async readahead before
1521                  * waiting for the lock.
1522                  */
1523                 do_async_mmap_readahead(vma, ra, file, page, offset);
1524                 lock_page(page);
1525
1526                 /* Did it get truncated? */
1527                 if (unlikely(page->mapping != mapping)) {
1528                         unlock_page(page);
1529                         put_page(page);
1530                         goto no_cached_page;
1531                 }
1532         } else {
1533                 /* No page in the page cache at all */
1534                 do_sync_mmap_readahead(vma, ra, file, offset);
1535                 count_vm_event(PGMAJFAULT);
1536                 ret = VM_FAULT_MAJOR;
1537 retry_find:
1538                 page = find_lock_page(mapping, offset);
1539                 if (!page)
1540                         goto no_cached_page;
1541         }
1542
1543         /*
1544          * We have a locked page in the page cache, now we need to check
1545          * that it's up-to-date. If not, it is going to be due to an error.
1546          */
1547         if (unlikely(!PageUptodate(page)))
1548                 goto page_not_uptodate;
1549
1550         /*
1551          * Found the page and have a reference on it.
1552          * We must recheck i_size under page lock.
1553          */
1554         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1555         if (unlikely(offset >= size)) {
1556                 unlock_page(page);
1557                 page_cache_release(page);
1558                 return VM_FAULT_SIGBUS;
1559         }
1560
1561         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1562         vmf->page = page;
1563         return ret | VM_FAULT_LOCKED;
1564
1565 no_cached_page:
1566         /*
1567          * We're only likely to ever get here if MADV_RANDOM is in
1568          * effect.
1569          */
1570         error = page_cache_read(file, offset);
1571
1572         /*
1573          * The page we want has now been added to the page cache.
1574          * In the unlikely event that someone removed it in the
1575          * meantime, we'll just come back here and read it again.
1576          */
1577         if (error >= 0)
1578                 goto retry_find;
1579
1580         /*
1581          * An error return from page_cache_read can result if the
1582          * system is low on memory, or a problem occurs while trying
1583          * to schedule I/O.
1584          */
1585         if (error == -ENOMEM)
1586                 return VM_FAULT_OOM;
1587         return VM_FAULT_SIGBUS;
1588
1589 page_not_uptodate:
1590         /*
1591          * Umm, take care of errors if the page isn't up-to-date.
1592          * Try to re-read it _once_. We do this synchronously,
1593          * because there really aren't any performance issues here
1594          * and we need to check for errors.
1595          */
1596         ClearPageError(page);
1597         error = mapping->a_ops->readpage(file, page);
1598         if (!error) {
1599                 wait_on_page_locked(page);
1600                 if (!PageUptodate(page))
1601                         error = -EIO;
1602         }
1603         page_cache_release(page);
1604
1605         if (!error || error == AOP_TRUNCATED_PAGE)
1606                 goto retry_find;
1607
1608         /* Things didn't work out. Return zero to tell the mm layer so. */
1609         shrink_readahead_size_eio(file, ra);
1610         return VM_FAULT_SIGBUS;
1611 }
1612 EXPORT_SYMBOL(filemap_fault);
1613
1614 const struct vm_operations_struct generic_file_vm_ops = {
1615         .fault          = filemap_fault,
1616 };
1617
1618 /* This is used for a general mmap of a disk file */
1619
1620 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1621 {
1622         struct address_space *mapping = file->f_mapping;
1623
1624         if (!mapping->a_ops->readpage)
1625                 return -ENOEXEC;
1626         file_accessed(file);
1627         vma->vm_ops = &generic_file_vm_ops;
1628         vma->vm_flags |= VM_CAN_NONLINEAR;
1629         return 0;
1630 }
1631
1632 /*
1633  * This is for filesystems which do not implement ->writepage.
1634  */
1635 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1636 {
1637         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1638                 return -EINVAL;
1639         return generic_file_mmap(file, vma);
1640 }
1641 #else
1642 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1643 {
1644         return -ENOSYS;
1645 }
1646 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1647 {
1648         return -ENOSYS;
1649 }
1650 #endif /* CONFIG_MMU */
1651
1652 EXPORT_SYMBOL(generic_file_mmap);
1653 EXPORT_SYMBOL(generic_file_readonly_mmap);
1654
1655 static struct page *__read_cache_page(struct address_space *mapping,
1656                                 pgoff_t index,
1657                                 int (*filler)(void *,struct page*),
1658                                 void *data)
1659 {
1660         struct page *page;
1661         int err;
1662 repeat:
1663         page = find_get_page(mapping, index);
1664         if (!page) {
1665                 page = page_cache_alloc_cold(mapping);
1666                 if (!page)
1667                         return ERR_PTR(-ENOMEM);
1668                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1669                 if (unlikely(err)) {
1670                         page_cache_release(page);
1671                         if (err == -EEXIST)
1672                                 goto repeat;
1673                         /* Presumably ENOMEM for radix tree node */
1674                         return ERR_PTR(err);
1675                 }
1676                 err = filler(data, page);
1677                 if (err < 0) {
1678                         page_cache_release(page);
1679                         page = ERR_PTR(err);
1680                 }
1681         }
1682         return page;
1683 }
1684
1685 /**
1686  * read_cache_page_async - read into page cache, fill it if needed
1687  * @mapping:    the page's address_space
1688  * @index:      the page index
1689  * @filler:     function to perform the read
1690  * @data:       destination for read data
1691  *
1692  * Same as read_cache_page, but don't wait for page to become unlocked
1693  * after submitting it to the filler.
1694  *
1695  * Read into the page cache. If a page already exists, and PageUptodate() is
1696  * not set, try to fill the page but don't wait for it to become unlocked.
1697  *
1698  * If the page does not get brought uptodate, return -EIO.
1699  */
1700 struct page *read_cache_page_async(struct address_space *mapping,
1701                                 pgoff_t index,
1702                                 int (*filler)(void *,struct page*),
1703                                 void *data)
1704 {
1705         struct page *page;
1706         int err;
1707
1708 retry:
1709         page = __read_cache_page(mapping, index, filler, data);
1710         if (IS_ERR(page))
1711                 return page;
1712         if (PageUptodate(page))
1713                 goto out;
1714
1715         lock_page(page);
1716         if (!page->mapping) {
1717                 unlock_page(page);
1718                 page_cache_release(page);
1719                 goto retry;
1720         }
1721         if (PageUptodate(page)) {
1722                 unlock_page(page);
1723                 goto out;
1724         }
1725         err = filler(data, page);
1726         if (err < 0) {
1727                 page_cache_release(page);
1728                 return ERR_PTR(err);
1729         }
1730 out:
1731         mark_page_accessed(page);
1732         return page;
1733 }
1734 EXPORT_SYMBOL(read_cache_page_async);
1735
1736 /**
1737  * read_cache_page - read into page cache, fill it if needed
1738  * @mapping:    the page's address_space
1739  * @index:      the page index
1740  * @filler:     function to perform the read
1741  * @data:       destination for read data
1742  *
1743  * Read into the page cache. If a page already exists, and PageUptodate() is
1744  * not set, try to fill the page then wait for it to become unlocked.
1745  *
1746  * If the page does not get brought uptodate, return -EIO.
1747  */
1748 struct page *read_cache_page(struct address_space *mapping,
1749                                 pgoff_t index,
1750                                 int (*filler)(void *,struct page*),
1751                                 void *data)
1752 {
1753         struct page *page;
1754
1755         page = read_cache_page_async(mapping, index, filler, data);
1756         if (IS_ERR(page))
1757                 goto out;
1758         wait_on_page_locked(page);
1759         if (!PageUptodate(page)) {
1760                 page_cache_release(page);
1761                 page = ERR_PTR(-EIO);
1762         }
1763  out:
1764         return page;
1765 }
1766 EXPORT_SYMBOL(read_cache_page);
1767
1768 /*
1769  * The logic we want is
1770  *
1771  *      if suid or (sgid and xgrp)
1772  *              remove privs
1773  */
1774 int should_remove_suid(struct dentry *dentry)
1775 {
1776         mode_t mode = dentry->d_inode->i_mode;
1777         int kill = 0;
1778
1779         /* suid always must be killed */
1780         if (unlikely(mode & S_ISUID))
1781                 kill = ATTR_KILL_SUID;
1782
1783         /*
1784          * sgid without any exec bits is just a mandatory locking mark; leave
1785          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1786          */
1787         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1788                 kill |= ATTR_KILL_SGID;
1789
1790         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1791                 return kill;
1792
1793         return 0;
1794 }
1795 EXPORT_SYMBOL(should_remove_suid);
1796
1797 static int __remove_suid(struct dentry *dentry, int kill)
1798 {
1799         struct iattr newattrs;
1800
1801         newattrs.ia_valid = ATTR_FORCE | kill;
1802         return notify_change(dentry, &newattrs);
1803 }
1804
1805 int file_remove_suid(struct file *file)
1806 {
1807         struct dentry *dentry = file->f_path.dentry;
1808         int killsuid = should_remove_suid(dentry);
1809         int killpriv = security_inode_need_killpriv(dentry);
1810         int error = 0;
1811
1812         if (killpriv < 0)
1813                 return killpriv;
1814         if (killpriv)
1815                 error = security_inode_killpriv(dentry);
1816         if (!error && killsuid)
1817                 error = __remove_suid(dentry, killsuid);
1818
1819         return error;
1820 }
1821 EXPORT_SYMBOL(file_remove_suid);
1822
1823 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1824                         const struct iovec *iov, size_t base, size_t bytes)
1825 {
1826         size_t copied = 0, left = 0;
1827
1828         while (bytes) {
1829                 char __user *buf = iov->iov_base + base;
1830                 int copy = min(bytes, iov->iov_len - base);
1831
1832                 base = 0;
1833                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1834                 copied += copy;
1835                 bytes -= copy;
1836                 vaddr += copy;
1837                 iov++;
1838
1839                 if (unlikely(left))
1840                         break;
1841         }
1842         return copied - left;
1843 }
1844
1845 /*
1846  * Copy as much as we can into the page and return the number of bytes which
1847  * were sucessfully copied.  If a fault is encountered then return the number of
1848  * bytes which were copied.
1849  */
1850 size_t iov_iter_copy_from_user_atomic(struct page *page,
1851                 struct iov_iter *i, unsigned long offset, size_t bytes)
1852 {
1853         char *kaddr;
1854         size_t copied;
1855
1856         BUG_ON(!in_atomic());
1857         kaddr = kmap_atomic(page, KM_USER0);
1858         if (likely(i->nr_segs == 1)) {
1859                 int left;
1860                 char __user *buf = i->iov->iov_base + i->iov_offset;
1861                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1862                 copied = bytes - left;
1863         } else {
1864                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1865                                                 i->iov, i->iov_offset, bytes);
1866         }
1867         kunmap_atomic(kaddr, KM_USER0);
1868
1869         return copied;
1870 }
1871 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1872
1873 /*
1874  * This has the same sideeffects and return value as
1875  * iov_iter_copy_from_user_atomic().
1876  * The difference is that it attempts to resolve faults.
1877  * Page must not be locked.
1878  */
1879 size_t iov_iter_copy_from_user(struct page *page,
1880                 struct iov_iter *i, unsigned long offset, size_t bytes)
1881 {
1882         char *kaddr;
1883         size_t copied;
1884
1885         kaddr = kmap(page);
1886         if (likely(i->nr_segs == 1)) {
1887                 int left;
1888                 char __user *buf = i->iov->iov_base + i->iov_offset;
1889                 left = __copy_from_user(kaddr + offset, buf, bytes);
1890                 copied = bytes - left;
1891         } else {
1892                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1893                                                 i->iov, i->iov_offset, bytes);
1894         }
1895         kunmap(page);
1896         return copied;
1897 }
1898 EXPORT_SYMBOL(iov_iter_copy_from_user);
1899
1900 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1901 {
1902         BUG_ON(i->count < bytes);
1903
1904         if (likely(i->nr_segs == 1)) {
1905                 i->iov_offset += bytes;
1906                 i->count -= bytes;
1907         } else {
1908                 const struct iovec *iov = i->iov;
1909                 size_t base = i->iov_offset;
1910
1911                 /*
1912                  * The !iov->iov_len check ensures we skip over unlikely
1913                  * zero-length segments (without overruning the iovec).
1914                  */
1915                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1916                         int copy;
1917
1918                         copy = min(bytes, iov->iov_len - base);
1919                         BUG_ON(!i->count || i->count < copy);
1920                         i->count -= copy;
1921                         bytes -= copy;
1922                         base += copy;
1923                         if (iov->iov_len == base) {
1924                                 iov++;
1925                                 base = 0;
1926                         }
1927                 }
1928                 i->iov = iov;
1929                 i->iov_offset = base;
1930         }
1931 }
1932 EXPORT_SYMBOL(iov_iter_advance);
1933
1934 /*
1935  * Fault in the first iovec of the given iov_iter, to a maximum length
1936  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1937  * accessed (ie. because it is an invalid address).
1938  *
1939  * writev-intensive code may want this to prefault several iovecs -- that
1940  * would be possible (callers must not rely on the fact that _only_ the
1941  * first iovec will be faulted with the current implementation).
1942  */
1943 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1944 {
1945         char __user *buf = i->iov->iov_base + i->iov_offset;
1946         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1947         return fault_in_pages_readable(buf, bytes);
1948 }
1949 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1950
1951 /*
1952  * Return the count of just the current iov_iter segment.
1953  */
1954 size_t iov_iter_single_seg_count(struct iov_iter *i)
1955 {
1956         const struct iovec *iov = i->iov;
1957         if (i->nr_segs == 1)
1958                 return i->count;
1959         else
1960                 return min(i->count, iov->iov_len - i->iov_offset);
1961 }
1962 EXPORT_SYMBOL(iov_iter_single_seg_count);
1963
1964 /*
1965  * Performs necessary checks before doing a write
1966  *
1967  * Can adjust writing position or amount of bytes to write.
1968  * Returns appropriate error code that caller should return or
1969  * zero in case that write should be allowed.
1970  */
1971 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1972 {
1973         struct inode *inode = file->f_mapping->host;
1974         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1975
1976         if (unlikely(*pos < 0))
1977                 return -EINVAL;
1978
1979         if (!isblk) {
1980                 /* FIXME: this is for backwards compatibility with 2.4 */
1981                 if (file->f_flags & O_APPEND)
1982                         *pos = i_size_read(inode);
1983
1984                 if (limit != RLIM_INFINITY) {
1985                         if (*pos >= limit) {
1986                                 send_sig(SIGXFSZ, current, 0);
1987                                 return -EFBIG;
1988                         }
1989                         if (*count > limit - (typeof(limit))*pos) {
1990                                 *count = limit - (typeof(limit))*pos;
1991                         }
1992                 }
1993         }
1994
1995         /*
1996          * LFS rule
1997          */
1998         if (unlikely(*pos + *count > MAX_NON_LFS &&
1999                                 !(file->f_flags & O_LARGEFILE))) {
2000                 if (*pos >= MAX_NON_LFS) {
2001                         return -EFBIG;
2002                 }
2003                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2004                         *count = MAX_NON_LFS - (unsigned long)*pos;
2005                 }
2006         }
2007
2008         /*
2009          * Are we about to exceed the fs block limit ?
2010          *
2011          * If we have written data it becomes a short write.  If we have
2012          * exceeded without writing data we send a signal and return EFBIG.
2013          * Linus frestrict idea will clean these up nicely..
2014          */
2015         if (likely(!isblk)) {
2016                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2017                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2018                                 return -EFBIG;
2019                         }
2020                         /* zero-length writes at ->s_maxbytes are OK */
2021                 }
2022
2023                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2024                         *count = inode->i_sb->s_maxbytes - *pos;
2025         } else {
2026 #ifdef CONFIG_BLOCK
2027                 loff_t isize;
2028                 if (bdev_read_only(I_BDEV(inode)))
2029                         return -EPERM;
2030                 isize = i_size_read(inode);
2031                 if (*pos >= isize) {
2032                         if (*count || *pos > isize)
2033                                 return -ENOSPC;
2034                 }
2035
2036                 if (*pos + *count > isize)
2037                         *count = isize - *pos;
2038 #else
2039                 return -EPERM;
2040 #endif
2041         }
2042         return 0;
2043 }
2044 EXPORT_SYMBOL(generic_write_checks);
2045
2046 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2047                                 loff_t pos, unsigned len, unsigned flags,
2048                                 struct page **pagep, void **fsdata)
2049 {
2050         const struct address_space_operations *aops = mapping->a_ops;
2051
2052         return aops->write_begin(file, mapping, pos, len, flags,
2053                                                         pagep, fsdata);
2054 }
2055 EXPORT_SYMBOL(pagecache_write_begin);
2056
2057 int pagecache_write_end(struct file *file, struct address_space *mapping,
2058                                 loff_t pos, unsigned len, unsigned copied,
2059                                 struct page *page, void *fsdata)
2060 {
2061         const struct address_space_operations *aops = mapping->a_ops;
2062
2063         mark_page_accessed(page);
2064         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2065 }
2066 EXPORT_SYMBOL(pagecache_write_end);
2067
2068 ssize_t
2069 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2070                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2071                 size_t count, size_t ocount)
2072 {
2073         struct file     *file = iocb->ki_filp;
2074         struct address_space *mapping = file->f_mapping;
2075         struct inode    *inode = mapping->host;
2076         ssize_t         written;
2077         size_t          write_len;
2078         pgoff_t         end;
2079
2080         if (count != ocount)
2081                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2082
2083         write_len = iov_length(iov, *nr_segs);
2084         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2085
2086         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2087         if (written)
2088                 goto out;
2089
2090         /*
2091          * After a write we want buffered reads to be sure to go to disk to get
2092          * the new data.  We invalidate clean cached page from the region we're
2093          * about to write.  We do this *before* the write so that we can return
2094          * without clobbering -EIOCBQUEUED from ->direct_IO().
2095          */
2096         if (mapping->nrpages) {
2097                 written = invalidate_inode_pages2_range(mapping,
2098                                         pos >> PAGE_CACHE_SHIFT, end);
2099                 /*
2100                  * If a page can not be invalidated, return 0 to fall back
2101                  * to buffered write.
2102                  */
2103                 if (written) {
2104                         if (written == -EBUSY)
2105                                 return 0;
2106                         goto out;
2107                 }
2108         }
2109
2110         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2111
2112         /*
2113          * Finally, try again to invalidate clean pages which might have been
2114          * cached by non-direct readahead, or faulted in by get_user_pages()
2115          * if the source of the write was an mmap'ed region of the file
2116          * we're writing.  Either one is a pretty crazy thing to do,
2117          * so we don't support it 100%.  If this invalidation
2118          * fails, tough, the write still worked...
2119          */
2120         if (mapping->nrpages) {
2121                 invalidate_inode_pages2_range(mapping,
2122                                               pos >> PAGE_CACHE_SHIFT, end);
2123         }
2124
2125         if (written > 0) {
2126                 loff_t end = pos + written;
2127                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2128                         i_size_write(inode,  end);
2129                         mark_inode_dirty(inode);
2130                 }
2131                 *ppos = end;
2132         }
2133 out:
2134         return written;
2135 }
2136 EXPORT_SYMBOL(generic_file_direct_write);
2137
2138 /*
2139  * Find or create a page at the given pagecache position. Return the locked
2140  * page. This function is specifically for buffered writes.
2141  */
2142 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2143                                         pgoff_t index, unsigned flags)
2144 {
2145         int status;
2146         struct page *page;
2147         gfp_t gfp_notmask = 0;
2148         if (flags & AOP_FLAG_NOFS)
2149                 gfp_notmask = __GFP_FS;
2150 repeat:
2151         page = find_lock_page(mapping, index);
2152         if (likely(page))
2153                 return page;
2154
2155         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2156         if (!page)
2157                 return NULL;
2158         status = add_to_page_cache_lru(page, mapping, index,
2159                                                 GFP_KERNEL & ~gfp_notmask);
2160         if (unlikely(status)) {
2161                 page_cache_release(page);
2162                 if (status == -EEXIST)
2163                         goto repeat;
2164                 return NULL;
2165         }
2166         return page;
2167 }
2168 EXPORT_SYMBOL(grab_cache_page_write_begin);
2169
2170 static ssize_t generic_perform_write(struct file *file,
2171                                 struct iov_iter *i, loff_t pos)
2172 {
2173         struct address_space *mapping = file->f_mapping;
2174         const struct address_space_operations *a_ops = mapping->a_ops;
2175         long status = 0;
2176         ssize_t written = 0;
2177         unsigned int flags = 0;
2178
2179         /*
2180          * Copies from kernel address space cannot fail (NFSD is a big user).
2181          */
2182         if (segment_eq(get_fs(), KERNEL_DS))
2183                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2184
2185         do {
2186                 struct page *page;
2187                 pgoff_t index;          /* Pagecache index for current page */
2188                 unsigned long offset;   /* Offset into pagecache page */
2189                 unsigned long bytes;    /* Bytes to write to page */
2190                 size_t copied;          /* Bytes copied from user */
2191                 void *fsdata;
2192
2193                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2194                 index = pos >> PAGE_CACHE_SHIFT;
2195                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2196                                                 iov_iter_count(i));
2197
2198 again:
2199
2200                 /*
2201                  * Bring in the user page that we will copy from _first_.
2202                  * Otherwise there's a nasty deadlock on copying from the
2203                  * same page as we're writing to, without it being marked
2204                  * up-to-date.
2205                  *
2206                  * Not only is this an optimisation, but it is also required
2207                  * to check that the address is actually valid, when atomic
2208                  * usercopies are used, below.
2209                  */
2210                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2211                         status = -EFAULT;
2212                         break;
2213                 }
2214
2215                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2216                                                 &page, &fsdata);
2217                 if (unlikely(status))
2218                         break;
2219
2220                 pagefault_disable();
2221                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2222                 pagefault_enable();
2223                 flush_dcache_page(page);
2224
2225                 mark_page_accessed(page);
2226                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2227                                                 page, fsdata);
2228                 if (unlikely(status < 0))
2229                         break;
2230                 copied = status;
2231
2232                 cond_resched();
2233
2234                 iov_iter_advance(i, copied);
2235                 if (unlikely(copied == 0)) {
2236                         /*
2237                          * If we were unable to copy any data at all, we must
2238                          * fall back to a single segment length write.
2239                          *
2240                          * If we didn't fallback here, we could livelock
2241                          * because not all segments in the iov can be copied at
2242                          * once without a pagefault.
2243                          */
2244                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2245                                                 iov_iter_single_seg_count(i));
2246                         goto again;
2247                 }
2248                 pos += copied;
2249                 written += copied;
2250
2251                 balance_dirty_pages_ratelimited(mapping);
2252
2253         } while (iov_iter_count(i));
2254
2255         return written ? written : status;
2256 }
2257
2258 ssize_t
2259 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2260                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2261                 size_t count, ssize_t written)
2262 {
2263         struct file *file = iocb->ki_filp;
2264         struct address_space *mapping = file->f_mapping;
2265         ssize_t status;
2266         struct iov_iter i;
2267
2268         iov_iter_init(&i, iov, nr_segs, count, written);
2269         status = generic_perform_write(file, &i, pos);
2270
2271         if (likely(status >= 0)) {
2272                 written += status;
2273                 *ppos = pos + status;
2274         }
2275         
2276         /*
2277          * If we get here for O_DIRECT writes then we must have fallen through
2278          * to buffered writes (block instantiation inside i_size).  So we sync
2279          * the file data here, to try to honour O_DIRECT expectations.
2280          */
2281         if (unlikely(file->f_flags & O_DIRECT) && written)
2282                 status = filemap_write_and_wait_range(mapping,
2283                                         pos, pos + written - 1);
2284
2285         return written ? written : status;
2286 }
2287 EXPORT_SYMBOL(generic_file_buffered_write);
2288
2289 /**
2290  * __generic_file_aio_write - write data to a file
2291  * @iocb:       IO state structure (file, offset, etc.)
2292  * @iov:        vector with data to write
2293  * @nr_segs:    number of segments in the vector
2294  * @ppos:       position where to write
2295  *
2296  * This function does all the work needed for actually writing data to a
2297  * file. It does all basic checks, removes SUID from the file, updates
2298  * modification times and calls proper subroutines depending on whether we
2299  * do direct IO or a standard buffered write.
2300  *
2301  * It expects i_mutex to be grabbed unless we work on a block device or similar
2302  * object which does not need locking at all.
2303  *
2304  * This function does *not* take care of syncing data in case of O_SYNC write.
2305  * A caller has to handle it. This is mainly due to the fact that we want to
2306  * avoid syncing under i_mutex.
2307  */
2308 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2309                                  unsigned long nr_segs, loff_t *ppos)
2310 {
2311         struct file *file = iocb->ki_filp;
2312         struct address_space * mapping = file->f_mapping;
2313         size_t ocount;          /* original count */
2314         size_t count;           /* after file limit checks */
2315         struct inode    *inode = mapping->host;
2316         loff_t          pos;
2317         ssize_t         written;
2318         ssize_t         err;
2319
2320         ocount = 0;
2321         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2322         if (err)
2323                 return err;
2324
2325         count = ocount;
2326         pos = *ppos;
2327
2328         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2329
2330         /* We can write back this queue in page reclaim */
2331         current->backing_dev_info = mapping->backing_dev_info;
2332         written = 0;
2333
2334         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2335         if (err)
2336                 goto out;
2337
2338         if (count == 0)
2339                 goto out;
2340
2341         err = file_remove_suid(file);
2342         if (err)
2343                 goto out;
2344
2345         file_update_time(file);
2346
2347         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2348         if (unlikely(file->f_flags & O_DIRECT)) {
2349                 loff_t endbyte;
2350                 ssize_t written_buffered;
2351
2352                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2353                                                         ppos, count, ocount);
2354                 if (written < 0 || written == count)
2355                         goto out;
2356                 /*
2357                  * direct-io write to a hole: fall through to buffered I/O
2358                  * for completing the rest of the request.
2359                  */
2360                 pos += written;
2361                 count -= written;
2362                 written_buffered = generic_file_buffered_write(iocb, iov,
2363                                                 nr_segs, pos, ppos, count,
2364                                                 written);
2365                 /*
2366                  * If generic_file_buffered_write() retuned a synchronous error
2367                  * then we want to return the number of bytes which were
2368                  * direct-written, or the error code if that was zero.  Note
2369                  * that this differs from normal direct-io semantics, which
2370                  * will return -EFOO even if some bytes were written.
2371                  */
2372                 if (written_buffered < 0) {
2373                         err = written_buffered;
2374                         goto out;
2375                 }
2376
2377                 /*
2378                  * We need to ensure that the page cache pages are written to
2379                  * disk and invalidated to preserve the expected O_DIRECT
2380                  * semantics.
2381                  */
2382                 endbyte = pos + written_buffered - written - 1;
2383                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2384                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2385                                             SYNC_FILE_RANGE_WRITE|
2386                                             SYNC_FILE_RANGE_WAIT_AFTER);
2387                 if (err == 0) {
2388                         written = written_buffered;
2389                         invalidate_mapping_pages(mapping,
2390                                                  pos >> PAGE_CACHE_SHIFT,
2391                                                  endbyte >> PAGE_CACHE_SHIFT);
2392                 } else {
2393                         /*
2394                          * We don't know how much we wrote, so just return
2395                          * the number of bytes which were direct-written
2396                          */
2397                 }
2398         } else {
2399                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2400                                 pos, ppos, count, written);
2401         }
2402 out:
2403         current->backing_dev_info = NULL;
2404         return written ? written : err;
2405 }
2406 EXPORT_SYMBOL(__generic_file_aio_write);
2407
2408 /**
2409  * generic_file_aio_write - write data to a file
2410  * @iocb:       IO state structure
2411  * @iov:        vector with data to write
2412  * @nr_segs:    number of segments in the vector
2413  * @pos:        position in file where to write
2414  *
2415  * This is a wrapper around __generic_file_aio_write() to be used by most
2416  * filesystems. It takes care of syncing the file in case of O_SYNC file
2417  * and acquires i_mutex as needed.
2418  */
2419 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2420                 unsigned long nr_segs, loff_t pos)
2421 {
2422         struct file *file = iocb->ki_filp;
2423         struct inode *inode = file->f_mapping->host;
2424         ssize_t ret;
2425
2426         BUG_ON(iocb->ki_pos != pos);
2427
2428         mutex_lock(&inode->i_mutex);
2429         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2430         mutex_unlock(&inode->i_mutex);
2431
2432         if (ret > 0 || ret == -EIOCBQUEUED) {
2433                 ssize_t err;
2434
2435                 err = generic_write_sync(file, pos, ret);
2436                 if (err < 0 && ret > 0)
2437                         ret = err;
2438         }
2439         return ret;
2440 }
2441 EXPORT_SYMBOL(generic_file_aio_write);
2442
2443 /**
2444  * try_to_release_page() - release old fs-specific metadata on a page
2445  *
2446  * @page: the page which the kernel is trying to free
2447  * @gfp_mask: memory allocation flags (and I/O mode)
2448  *
2449  * The address_space is to try to release any data against the page
2450  * (presumably at page->private).  If the release was successful, return `1'.
2451  * Otherwise return zero.
2452  *
2453  * This may also be called if PG_fscache is set on a page, indicating that the
2454  * page is known to the local caching routines.
2455  *
2456  * The @gfp_mask argument specifies whether I/O may be performed to release
2457  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2458  *
2459  */
2460 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2461 {
2462         struct address_space * const mapping = page->mapping;
2463
2464         BUG_ON(!PageLocked(page));
2465         if (PageWriteback(page))
2466                 return 0;
2467
2468         if (mapping && mapping->a_ops->releasepage)
2469                 return mapping->a_ops->releasepage(page, gfp_mask);
2470         return try_to_free_buffers(page);
2471 }
2472
2473 EXPORT_SYMBOL(try_to_release_page);