dd51c68e2b868adbb403b5967c4e7c6be357e722
[linux-3.10.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               (vmtruncate)
62  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
63  *      ->swap_lock             (exclusive_swap_page, others)
64  *        ->mapping->tree_lock
65  *
66  *  ->i_mutex
67  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
68  *
69  *  ->mmap_sem
70  *    ->i_mmap_lock
71  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
72  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
73  *
74  *  ->mmap_sem
75  *    ->lock_page               (access_process_vm)
76  *
77  *  ->i_mutex                   (generic_file_buffered_write)
78  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
79  *
80  *  ->i_mutex
81  *    ->i_alloc_sem             (various)
82  *
83  *  ->inode_lock
84  *    ->sb_lock                 (fs/fs-writeback.c)
85  *    ->mapping->tree_lock      (__sync_single_inode)
86  *
87  *  ->i_mmap_lock
88  *    ->anon_vma.lock           (vma_adjust)
89  *
90  *  ->anon_vma.lock
91  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
92  *
93  *  ->page_table_lock or pte_lock
94  *    ->swap_lock               (try_to_unmap_one)
95  *    ->private_lock            (try_to_unmap_one)
96  *    ->tree_lock               (try_to_unmap_one)
97  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
98  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
99  *    ->private_lock            (page_remove_rmap->set_page_dirty)
100  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
101  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (zap_pte_range->set_page_dirty)
103  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
104  *
105  *  ->task->proc_lock
106  *    ->dcache_lock             (proc_pid_lookup)
107  */
108
109 /*
110  * Remove a page from the page cache and free it. Caller has to make
111  * sure the page is locked and that nobody else uses it - or that usage
112  * is safe.  The caller must hold the mapping's tree_lock.
113  */
114 void __remove_from_page_cache(struct page *page)
115 {
116         struct address_space *mapping = page->mapping;
117
118         radix_tree_delete(&mapping->page_tree, page->index);
119         page->mapping = NULL;
120         mapping->nrpages--;
121         __dec_zone_page_state(page, NR_FILE_PAGES);
122         BUG_ON(page_mapped(page));
123
124         /*
125          * Some filesystems seem to re-dirty the page even after
126          * the VM has canceled the dirty bit (eg ext3 journaling).
127          *
128          * Fix it up by doing a final dirty accounting check after
129          * having removed the page entirely.
130          */
131         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
132                 dec_zone_page_state(page, NR_FILE_DIRTY);
133                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
134         }
135 }
136
137 void remove_from_page_cache(struct page *page)
138 {
139         struct address_space *mapping = page->mapping;
140
141         BUG_ON(!PageLocked(page));
142
143         spin_lock_irq(&mapping->tree_lock);
144         __remove_from_page_cache(page);
145         spin_unlock_irq(&mapping->tree_lock);
146         mem_cgroup_uncharge_cache_page(page);
147 }
148
149 static int sync_page(void *word)
150 {
151         struct address_space *mapping;
152         struct page *page;
153
154         page = container_of((unsigned long *)word, struct page, flags);
155
156         /*
157          * page_mapping() is being called without PG_locked held.
158          * Some knowledge of the state and use of the page is used to
159          * reduce the requirements down to a memory barrier.
160          * The danger here is of a stale page_mapping() return value
161          * indicating a struct address_space different from the one it's
162          * associated with when it is associated with one.
163          * After smp_mb(), it's either the correct page_mapping() for
164          * the page, or an old page_mapping() and the page's own
165          * page_mapping() has gone NULL.
166          * The ->sync_page() address_space operation must tolerate
167          * page_mapping() going NULL. By an amazing coincidence,
168          * this comes about because none of the users of the page
169          * in the ->sync_page() methods make essential use of the
170          * page_mapping(), merely passing the page down to the backing
171          * device's unplug functions when it's non-NULL, which in turn
172          * ignore it for all cases but swap, where only page_private(page) is
173          * of interest. When page_mapping() does go NULL, the entire
174          * call stack gracefully ignores the page and returns.
175          * -- wli
176          */
177         smp_mb();
178         mapping = page_mapping(page);
179         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
180                 mapping->a_ops->sync_page(page);
181         io_schedule();
182         return 0;
183 }
184
185 static int sync_page_killable(void *word)
186 {
187         sync_page(word);
188         return fatal_signal_pending(current) ? -EINTR : 0;
189 }
190
191 /**
192  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
193  * @mapping:    address space structure to write
194  * @start:      offset in bytes where the range starts
195  * @end:        offset in bytes where the range ends (inclusive)
196  * @sync_mode:  enable synchronous operation
197  *
198  * Start writeback against all of a mapping's dirty pages that lie
199  * within the byte offsets <start, end> inclusive.
200  *
201  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
202  * opposed to a regular memory cleansing writeback.  The difference between
203  * these two operations is that if a dirty page/buffer is encountered, it must
204  * be waited upon, and not just skipped over.
205  */
206 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
207                                 loff_t end, int sync_mode)
208 {
209         int ret;
210         struct writeback_control wbc = {
211                 .sync_mode = sync_mode,
212                 .nr_to_write = LONG_MAX,
213                 .range_start = start,
214                 .range_end = end,
215         };
216
217         if (!mapping_cap_writeback_dirty(mapping))
218                 return 0;
219
220         ret = do_writepages(mapping, &wbc);
221         return ret;
222 }
223
224 static inline int __filemap_fdatawrite(struct address_space *mapping,
225         int sync_mode)
226 {
227         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
228 }
229
230 int filemap_fdatawrite(struct address_space *mapping)
231 {
232         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
233 }
234 EXPORT_SYMBOL(filemap_fdatawrite);
235
236 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
237                                 loff_t end)
238 {
239         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
240 }
241 EXPORT_SYMBOL(filemap_fdatawrite_range);
242
243 /**
244  * filemap_flush - mostly a non-blocking flush
245  * @mapping:    target address_space
246  *
247  * This is a mostly non-blocking flush.  Not suitable for data-integrity
248  * purposes - I/O may not be started against all dirty pages.
249  */
250 int filemap_flush(struct address_space *mapping)
251 {
252         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
253 }
254 EXPORT_SYMBOL(filemap_flush);
255
256 /**
257  * wait_on_page_writeback_range - wait for writeback to complete
258  * @mapping:    target address_space
259  * @start:      beginning page index
260  * @end:        ending page index
261  *
262  * Wait for writeback to complete against pages indexed by start->end
263  * inclusive
264  */
265 int wait_on_page_writeback_range(struct address_space *mapping,
266                                 pgoff_t start, pgoff_t end)
267 {
268         struct pagevec pvec;
269         int nr_pages;
270         int ret = 0;
271         pgoff_t index;
272
273         if (end < start)
274                 return 0;
275
276         pagevec_init(&pvec, 0);
277         index = start;
278         while ((index <= end) &&
279                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
280                         PAGECACHE_TAG_WRITEBACK,
281                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
282                 unsigned i;
283
284                 for (i = 0; i < nr_pages; i++) {
285                         struct page *page = pvec.pages[i];
286
287                         /* until radix tree lookup accepts end_index */
288                         if (page->index > end)
289                                 continue;
290
291                         wait_on_page_writeback(page);
292                         if (PageError(page))
293                                 ret = -EIO;
294                 }
295                 pagevec_release(&pvec);
296                 cond_resched();
297         }
298
299         /* Check for outstanding write errors */
300         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
301                 ret = -ENOSPC;
302         if (test_and_clear_bit(AS_EIO, &mapping->flags))
303                 ret = -EIO;
304
305         return ret;
306 }
307
308 /**
309  * filemap_fdatawait_range - wait for all under-writeback pages to complete in a given range
310  * @mapping: address space structure to wait for
311  * @start:      offset in bytes where the range starts
312  * @end:        offset in bytes where the range ends (inclusive)
313  *
314  * Walk the list of under-writeback pages of the given address space
315  * in the given range and wait for all of them.
316  *
317  * This is just a simple wrapper so that callers don't have to convert offsets
318  * to page indexes themselves
319  */
320 int filemap_fdatawait_range(struct address_space *mapping, loff_t start,
321                             loff_t end)
322 {
323         return wait_on_page_writeback_range(mapping, start >> PAGE_CACHE_SHIFT,
324                                             end >> PAGE_CACHE_SHIFT);
325 }
326 EXPORT_SYMBOL(filemap_fdatawait_range);
327
328 /**
329  * filemap_fdatawait - wait for all under-writeback pages to complete
330  * @mapping: address space structure to wait for
331  *
332  * Walk the list of under-writeback pages of the given address space
333  * and wait for all of them.
334  */
335 int filemap_fdatawait(struct address_space *mapping)
336 {
337         loff_t i_size = i_size_read(mapping->host);
338
339         if (i_size == 0)
340                 return 0;
341
342         return wait_on_page_writeback_range(mapping, 0,
343                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
344 }
345 EXPORT_SYMBOL(filemap_fdatawait);
346
347 int filemap_write_and_wait(struct address_space *mapping)
348 {
349         int err = 0;
350
351         if (mapping->nrpages) {
352                 err = filemap_fdatawrite(mapping);
353                 /*
354                  * Even if the above returned error, the pages may be
355                  * written partially (e.g. -ENOSPC), so we wait for it.
356                  * But the -EIO is special case, it may indicate the worst
357                  * thing (e.g. bug) happened, so we avoid waiting for it.
358                  */
359                 if (err != -EIO) {
360                         int err2 = filemap_fdatawait(mapping);
361                         if (!err)
362                                 err = err2;
363                 }
364         }
365         return err;
366 }
367 EXPORT_SYMBOL(filemap_write_and_wait);
368
369 /**
370  * filemap_write_and_wait_range - write out & wait on a file range
371  * @mapping:    the address_space for the pages
372  * @lstart:     offset in bytes where the range starts
373  * @lend:       offset in bytes where the range ends (inclusive)
374  *
375  * Write out and wait upon file offsets lstart->lend, inclusive.
376  *
377  * Note that `lend' is inclusive (describes the last byte to be written) so
378  * that this function can be used to write to the very end-of-file (end = -1).
379  */
380 int filemap_write_and_wait_range(struct address_space *mapping,
381                                  loff_t lstart, loff_t lend)
382 {
383         int err = 0;
384
385         if (mapping->nrpages) {
386                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
387                                                  WB_SYNC_ALL);
388                 /* See comment of filemap_write_and_wait() */
389                 if (err != -EIO) {
390                         int err2 = wait_on_page_writeback_range(mapping,
391                                                 lstart >> PAGE_CACHE_SHIFT,
392                                                 lend >> PAGE_CACHE_SHIFT);
393                         if (!err)
394                                 err = err2;
395                 }
396         }
397         return err;
398 }
399 EXPORT_SYMBOL(filemap_write_and_wait_range);
400
401 /**
402  * add_to_page_cache_locked - add a locked page to the pagecache
403  * @page:       page to add
404  * @mapping:    the page's address_space
405  * @offset:     page index
406  * @gfp_mask:   page allocation mode
407  *
408  * This function is used to add a page to the pagecache. It must be locked.
409  * This function does not add the page to the LRU.  The caller must do that.
410  */
411 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
412                 pgoff_t offset, gfp_t gfp_mask)
413 {
414         int error;
415
416         VM_BUG_ON(!PageLocked(page));
417
418         error = mem_cgroup_cache_charge(page, current->mm,
419                                         gfp_mask & GFP_RECLAIM_MASK);
420         if (error)
421                 goto out;
422
423         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
424         if (error == 0) {
425                 page_cache_get(page);
426                 page->mapping = mapping;
427                 page->index = offset;
428
429                 spin_lock_irq(&mapping->tree_lock);
430                 error = radix_tree_insert(&mapping->page_tree, offset, page);
431                 if (likely(!error)) {
432                         mapping->nrpages++;
433                         __inc_zone_page_state(page, NR_FILE_PAGES);
434                         spin_unlock_irq(&mapping->tree_lock);
435                 } else {
436                         page->mapping = NULL;
437                         spin_unlock_irq(&mapping->tree_lock);
438                         mem_cgroup_uncharge_cache_page(page);
439                         page_cache_release(page);
440                 }
441                 radix_tree_preload_end();
442         } else
443                 mem_cgroup_uncharge_cache_page(page);
444 out:
445         return error;
446 }
447 EXPORT_SYMBOL(add_to_page_cache_locked);
448
449 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
450                                 pgoff_t offset, gfp_t gfp_mask)
451 {
452         int ret;
453
454         /*
455          * Splice_read and readahead add shmem/tmpfs pages into the page cache
456          * before shmem_readpage has a chance to mark them as SwapBacked: they
457          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
458          * (called in add_to_page_cache) needs to know where they're going too.
459          */
460         if (mapping_cap_swap_backed(mapping))
461                 SetPageSwapBacked(page);
462
463         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
464         if (ret == 0) {
465                 if (page_is_file_cache(page))
466                         lru_cache_add_file(page);
467                 else
468                         lru_cache_add_active_anon(page);
469         }
470         return ret;
471 }
472 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
473
474 #ifdef CONFIG_NUMA
475 struct page *__page_cache_alloc(gfp_t gfp)
476 {
477         if (cpuset_do_page_mem_spread()) {
478                 int n = cpuset_mem_spread_node();
479                 return alloc_pages_exact_node(n, gfp, 0);
480         }
481         return alloc_pages(gfp, 0);
482 }
483 EXPORT_SYMBOL(__page_cache_alloc);
484 #endif
485
486 static int __sleep_on_page_lock(void *word)
487 {
488         io_schedule();
489         return 0;
490 }
491
492 /*
493  * In order to wait for pages to become available there must be
494  * waitqueues associated with pages. By using a hash table of
495  * waitqueues where the bucket discipline is to maintain all
496  * waiters on the same queue and wake all when any of the pages
497  * become available, and for the woken contexts to check to be
498  * sure the appropriate page became available, this saves space
499  * at a cost of "thundering herd" phenomena during rare hash
500  * collisions.
501  */
502 static wait_queue_head_t *page_waitqueue(struct page *page)
503 {
504         const struct zone *zone = page_zone(page);
505
506         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
507 }
508
509 static inline void wake_up_page(struct page *page, int bit)
510 {
511         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
512 }
513
514 void wait_on_page_bit(struct page *page, int bit_nr)
515 {
516         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
517
518         if (test_bit(bit_nr, &page->flags))
519                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
520                                                         TASK_UNINTERRUPTIBLE);
521 }
522 EXPORT_SYMBOL(wait_on_page_bit);
523
524 /**
525  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
526  * @page: Page defining the wait queue of interest
527  * @waiter: Waiter to add to the queue
528  *
529  * Add an arbitrary @waiter to the wait queue for the nominated @page.
530  */
531 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
532 {
533         wait_queue_head_t *q = page_waitqueue(page);
534         unsigned long flags;
535
536         spin_lock_irqsave(&q->lock, flags);
537         __add_wait_queue(q, waiter);
538         spin_unlock_irqrestore(&q->lock, flags);
539 }
540 EXPORT_SYMBOL_GPL(add_page_wait_queue);
541
542 /**
543  * unlock_page - unlock a locked page
544  * @page: the page
545  *
546  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
547  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
548  * mechananism between PageLocked pages and PageWriteback pages is shared.
549  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
550  *
551  * The mb is necessary to enforce ordering between the clear_bit and the read
552  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
553  */
554 void unlock_page(struct page *page)
555 {
556         VM_BUG_ON(!PageLocked(page));
557         clear_bit_unlock(PG_locked, &page->flags);
558         smp_mb__after_clear_bit();
559         wake_up_page(page, PG_locked);
560 }
561 EXPORT_SYMBOL(unlock_page);
562
563 /**
564  * end_page_writeback - end writeback against a page
565  * @page: the page
566  */
567 void end_page_writeback(struct page *page)
568 {
569         if (TestClearPageReclaim(page))
570                 rotate_reclaimable_page(page);
571
572         if (!test_clear_page_writeback(page))
573                 BUG();
574
575         smp_mb__after_clear_bit();
576         wake_up_page(page, PG_writeback);
577 }
578 EXPORT_SYMBOL(end_page_writeback);
579
580 /**
581  * __lock_page - get a lock on the page, assuming we need to sleep to get it
582  * @page: the page to lock
583  *
584  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
585  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
586  * chances are that on the second loop, the block layer's plug list is empty,
587  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
588  */
589 void __lock_page(struct page *page)
590 {
591         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
592
593         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
594                                                         TASK_UNINTERRUPTIBLE);
595 }
596 EXPORT_SYMBOL(__lock_page);
597
598 int __lock_page_killable(struct page *page)
599 {
600         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
601
602         return __wait_on_bit_lock(page_waitqueue(page), &wait,
603                                         sync_page_killable, TASK_KILLABLE);
604 }
605 EXPORT_SYMBOL_GPL(__lock_page_killable);
606
607 /**
608  * __lock_page_nosync - get a lock on the page, without calling sync_page()
609  * @page: the page to lock
610  *
611  * Variant of lock_page that does not require the caller to hold a reference
612  * on the page's mapping.
613  */
614 void __lock_page_nosync(struct page *page)
615 {
616         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
617         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
618                                                         TASK_UNINTERRUPTIBLE);
619 }
620
621 /**
622  * find_get_page - find and get a page reference
623  * @mapping: the address_space to search
624  * @offset: the page index
625  *
626  * Is there a pagecache struct page at the given (mapping, offset) tuple?
627  * If yes, increment its refcount and return it; if no, return NULL.
628  */
629 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
630 {
631         void **pagep;
632         struct page *page;
633
634         rcu_read_lock();
635 repeat:
636         page = NULL;
637         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
638         if (pagep) {
639                 page = radix_tree_deref_slot(pagep);
640                 if (unlikely(!page || page == RADIX_TREE_RETRY))
641                         goto repeat;
642
643                 if (!page_cache_get_speculative(page))
644                         goto repeat;
645
646                 /*
647                  * Has the page moved?
648                  * This is part of the lockless pagecache protocol. See
649                  * include/linux/pagemap.h for details.
650                  */
651                 if (unlikely(page != *pagep)) {
652                         page_cache_release(page);
653                         goto repeat;
654                 }
655         }
656         rcu_read_unlock();
657
658         return page;
659 }
660 EXPORT_SYMBOL(find_get_page);
661
662 /**
663  * find_lock_page - locate, pin and lock a pagecache page
664  * @mapping: the address_space to search
665  * @offset: the page index
666  *
667  * Locates the desired pagecache page, locks it, increments its reference
668  * count and returns its address.
669  *
670  * Returns zero if the page was not present. find_lock_page() may sleep.
671  */
672 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
673 {
674         struct page *page;
675
676 repeat:
677         page = find_get_page(mapping, offset);
678         if (page) {
679                 lock_page(page);
680                 /* Has the page been truncated? */
681                 if (unlikely(page->mapping != mapping)) {
682                         unlock_page(page);
683                         page_cache_release(page);
684                         goto repeat;
685                 }
686                 VM_BUG_ON(page->index != offset);
687         }
688         return page;
689 }
690 EXPORT_SYMBOL(find_lock_page);
691
692 /**
693  * find_or_create_page - locate or add a pagecache page
694  * @mapping: the page's address_space
695  * @index: the page's index into the mapping
696  * @gfp_mask: page allocation mode
697  *
698  * Locates a page in the pagecache.  If the page is not present, a new page
699  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
700  * LRU list.  The returned page is locked and has its reference count
701  * incremented.
702  *
703  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
704  * allocation!
705  *
706  * find_or_create_page() returns the desired page's address, or zero on
707  * memory exhaustion.
708  */
709 struct page *find_or_create_page(struct address_space *mapping,
710                 pgoff_t index, gfp_t gfp_mask)
711 {
712         struct page *page;
713         int err;
714 repeat:
715         page = find_lock_page(mapping, index);
716         if (!page) {
717                 page = __page_cache_alloc(gfp_mask);
718                 if (!page)
719                         return NULL;
720                 /*
721                  * We want a regular kernel memory (not highmem or DMA etc)
722                  * allocation for the radix tree nodes, but we need to honour
723                  * the context-specific requirements the caller has asked for.
724                  * GFP_RECLAIM_MASK collects those requirements.
725                  */
726                 err = add_to_page_cache_lru(page, mapping, index,
727                         (gfp_mask & GFP_RECLAIM_MASK));
728                 if (unlikely(err)) {
729                         page_cache_release(page);
730                         page = NULL;
731                         if (err == -EEXIST)
732                                 goto repeat;
733                 }
734         }
735         return page;
736 }
737 EXPORT_SYMBOL(find_or_create_page);
738
739 /**
740  * find_get_pages - gang pagecache lookup
741  * @mapping:    The address_space to search
742  * @start:      The starting page index
743  * @nr_pages:   The maximum number of pages
744  * @pages:      Where the resulting pages are placed
745  *
746  * find_get_pages() will search for and return a group of up to
747  * @nr_pages pages in the mapping.  The pages are placed at @pages.
748  * find_get_pages() takes a reference against the returned pages.
749  *
750  * The search returns a group of mapping-contiguous pages with ascending
751  * indexes.  There may be holes in the indices due to not-present pages.
752  *
753  * find_get_pages() returns the number of pages which were found.
754  */
755 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
756                             unsigned int nr_pages, struct page **pages)
757 {
758         unsigned int i;
759         unsigned int ret;
760         unsigned int nr_found;
761
762         rcu_read_lock();
763 restart:
764         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
765                                 (void ***)pages, start, nr_pages);
766         ret = 0;
767         for (i = 0; i < nr_found; i++) {
768                 struct page *page;
769 repeat:
770                 page = radix_tree_deref_slot((void **)pages[i]);
771                 if (unlikely(!page))
772                         continue;
773                 /*
774                  * this can only trigger if nr_found == 1, making livelock
775                  * a non issue.
776                  */
777                 if (unlikely(page == RADIX_TREE_RETRY))
778                         goto restart;
779
780                 if (!page_cache_get_speculative(page))
781                         goto repeat;
782
783                 /* Has the page moved? */
784                 if (unlikely(page != *((void **)pages[i]))) {
785                         page_cache_release(page);
786                         goto repeat;
787                 }
788
789                 pages[ret] = page;
790                 ret++;
791         }
792         rcu_read_unlock();
793         return ret;
794 }
795
796 /**
797  * find_get_pages_contig - gang contiguous pagecache lookup
798  * @mapping:    The address_space to search
799  * @index:      The starting page index
800  * @nr_pages:   The maximum number of pages
801  * @pages:      Where the resulting pages are placed
802  *
803  * find_get_pages_contig() works exactly like find_get_pages(), except
804  * that the returned number of pages are guaranteed to be contiguous.
805  *
806  * find_get_pages_contig() returns the number of pages which were found.
807  */
808 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
809                                unsigned int nr_pages, struct page **pages)
810 {
811         unsigned int i;
812         unsigned int ret;
813         unsigned int nr_found;
814
815         rcu_read_lock();
816 restart:
817         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
818                                 (void ***)pages, index, nr_pages);
819         ret = 0;
820         for (i = 0; i < nr_found; i++) {
821                 struct page *page;
822 repeat:
823                 page = radix_tree_deref_slot((void **)pages[i]);
824                 if (unlikely(!page))
825                         continue;
826                 /*
827                  * this can only trigger if nr_found == 1, making livelock
828                  * a non issue.
829                  */
830                 if (unlikely(page == RADIX_TREE_RETRY))
831                         goto restart;
832
833                 if (page->mapping == NULL || page->index != index)
834                         break;
835
836                 if (!page_cache_get_speculative(page))
837                         goto repeat;
838
839                 /* Has the page moved? */
840                 if (unlikely(page != *((void **)pages[i]))) {
841                         page_cache_release(page);
842                         goto repeat;
843                 }
844
845                 pages[ret] = page;
846                 ret++;
847                 index++;
848         }
849         rcu_read_unlock();
850         return ret;
851 }
852 EXPORT_SYMBOL(find_get_pages_contig);
853
854 /**
855  * find_get_pages_tag - find and return pages that match @tag
856  * @mapping:    the address_space to search
857  * @index:      the starting page index
858  * @tag:        the tag index
859  * @nr_pages:   the maximum number of pages
860  * @pages:      where the resulting pages are placed
861  *
862  * Like find_get_pages, except we only return pages which are tagged with
863  * @tag.   We update @index to index the next page for the traversal.
864  */
865 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
866                         int tag, unsigned int nr_pages, struct page **pages)
867 {
868         unsigned int i;
869         unsigned int ret;
870         unsigned int nr_found;
871
872         rcu_read_lock();
873 restart:
874         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
875                                 (void ***)pages, *index, nr_pages, tag);
876         ret = 0;
877         for (i = 0; i < nr_found; i++) {
878                 struct page *page;
879 repeat:
880                 page = radix_tree_deref_slot((void **)pages[i]);
881                 if (unlikely(!page))
882                         continue;
883                 /*
884                  * this can only trigger if nr_found == 1, making livelock
885                  * a non issue.
886                  */
887                 if (unlikely(page == RADIX_TREE_RETRY))
888                         goto restart;
889
890                 if (!page_cache_get_speculative(page))
891                         goto repeat;
892
893                 /* Has the page moved? */
894                 if (unlikely(page != *((void **)pages[i]))) {
895                         page_cache_release(page);
896                         goto repeat;
897                 }
898
899                 pages[ret] = page;
900                 ret++;
901         }
902         rcu_read_unlock();
903
904         if (ret)
905                 *index = pages[ret - 1]->index + 1;
906
907         return ret;
908 }
909 EXPORT_SYMBOL(find_get_pages_tag);
910
911 /**
912  * grab_cache_page_nowait - returns locked page at given index in given cache
913  * @mapping: target address_space
914  * @index: the page index
915  *
916  * Same as grab_cache_page(), but do not wait if the page is unavailable.
917  * This is intended for speculative data generators, where the data can
918  * be regenerated if the page couldn't be grabbed.  This routine should
919  * be safe to call while holding the lock for another page.
920  *
921  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
922  * and deadlock against the caller's locked page.
923  */
924 struct page *
925 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
926 {
927         struct page *page = find_get_page(mapping, index);
928
929         if (page) {
930                 if (trylock_page(page))
931                         return page;
932                 page_cache_release(page);
933                 return NULL;
934         }
935         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
936         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
937                 page_cache_release(page);
938                 page = NULL;
939         }
940         return page;
941 }
942 EXPORT_SYMBOL(grab_cache_page_nowait);
943
944 /*
945  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
946  * a _large_ part of the i/o request. Imagine the worst scenario:
947  *
948  *      ---R__________________________________________B__________
949  *         ^ reading here                             ^ bad block(assume 4k)
950  *
951  * read(R) => miss => readahead(R...B) => media error => frustrating retries
952  * => failing the whole request => read(R) => read(R+1) =>
953  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
954  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
955  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
956  *
957  * It is going insane. Fix it by quickly scaling down the readahead size.
958  */
959 static void shrink_readahead_size_eio(struct file *filp,
960                                         struct file_ra_state *ra)
961 {
962         ra->ra_pages /= 4;
963 }
964
965 /**
966  * do_generic_file_read - generic file read routine
967  * @filp:       the file to read
968  * @ppos:       current file position
969  * @desc:       read_descriptor
970  * @actor:      read method
971  *
972  * This is a generic file read routine, and uses the
973  * mapping->a_ops->readpage() function for the actual low-level stuff.
974  *
975  * This is really ugly. But the goto's actually try to clarify some
976  * of the logic when it comes to error handling etc.
977  */
978 static void do_generic_file_read(struct file *filp, loff_t *ppos,
979                 read_descriptor_t *desc, read_actor_t actor)
980 {
981         struct address_space *mapping = filp->f_mapping;
982         struct inode *inode = mapping->host;
983         struct file_ra_state *ra = &filp->f_ra;
984         pgoff_t index;
985         pgoff_t last_index;
986         pgoff_t prev_index;
987         unsigned long offset;      /* offset into pagecache page */
988         unsigned int prev_offset;
989         int error;
990
991         index = *ppos >> PAGE_CACHE_SHIFT;
992         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
993         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
994         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
995         offset = *ppos & ~PAGE_CACHE_MASK;
996
997         for (;;) {
998                 struct page *page;
999                 pgoff_t end_index;
1000                 loff_t isize;
1001                 unsigned long nr, ret;
1002
1003                 cond_resched();
1004 find_page:
1005                 page = find_get_page(mapping, index);
1006                 if (!page) {
1007                         page_cache_sync_readahead(mapping,
1008                                         ra, filp,
1009                                         index, last_index - index);
1010                         page = find_get_page(mapping, index);
1011                         if (unlikely(page == NULL))
1012                                 goto no_cached_page;
1013                 }
1014                 if (PageReadahead(page)) {
1015                         page_cache_async_readahead(mapping,
1016                                         ra, filp, page,
1017                                         index, last_index - index);
1018                 }
1019                 if (!PageUptodate(page)) {
1020                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1021                                         !mapping->a_ops->is_partially_uptodate)
1022                                 goto page_not_up_to_date;
1023                         if (!trylock_page(page))
1024                                 goto page_not_up_to_date;
1025                         if (!mapping->a_ops->is_partially_uptodate(page,
1026                                                                 desc, offset))
1027                                 goto page_not_up_to_date_locked;
1028                         unlock_page(page);
1029                 }
1030 page_ok:
1031                 /*
1032                  * i_size must be checked after we know the page is Uptodate.
1033                  *
1034                  * Checking i_size after the check allows us to calculate
1035                  * the correct value for "nr", which means the zero-filled
1036                  * part of the page is not copied back to userspace (unless
1037                  * another truncate extends the file - this is desired though).
1038                  */
1039
1040                 isize = i_size_read(inode);
1041                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1042                 if (unlikely(!isize || index > end_index)) {
1043                         page_cache_release(page);
1044                         goto out;
1045                 }
1046
1047                 /* nr is the maximum number of bytes to copy from this page */
1048                 nr = PAGE_CACHE_SIZE;
1049                 if (index == end_index) {
1050                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1051                         if (nr <= offset) {
1052                                 page_cache_release(page);
1053                                 goto out;
1054                         }
1055                 }
1056                 nr = nr - offset;
1057
1058                 /* If users can be writing to this page using arbitrary
1059                  * virtual addresses, take care about potential aliasing
1060                  * before reading the page on the kernel side.
1061                  */
1062                 if (mapping_writably_mapped(mapping))
1063                         flush_dcache_page(page);
1064
1065                 /*
1066                  * When a sequential read accesses a page several times,
1067                  * only mark it as accessed the first time.
1068                  */
1069                 if (prev_index != index || offset != prev_offset)
1070                         mark_page_accessed(page);
1071                 prev_index = index;
1072
1073                 /*
1074                  * Ok, we have the page, and it's up-to-date, so
1075                  * now we can copy it to user space...
1076                  *
1077                  * The actor routine returns how many bytes were actually used..
1078                  * NOTE! This may not be the same as how much of a user buffer
1079                  * we filled up (we may be padding etc), so we can only update
1080                  * "pos" here (the actor routine has to update the user buffer
1081                  * pointers and the remaining count).
1082                  */
1083                 ret = actor(desc, page, offset, nr);
1084                 offset += ret;
1085                 index += offset >> PAGE_CACHE_SHIFT;
1086                 offset &= ~PAGE_CACHE_MASK;
1087                 prev_offset = offset;
1088
1089                 page_cache_release(page);
1090                 if (ret == nr && desc->count)
1091                         continue;
1092                 goto out;
1093
1094 page_not_up_to_date:
1095                 /* Get exclusive access to the page ... */
1096                 error = lock_page_killable(page);
1097                 if (unlikely(error))
1098                         goto readpage_error;
1099
1100 page_not_up_to_date_locked:
1101                 /* Did it get truncated before we got the lock? */
1102                 if (!page->mapping) {
1103                         unlock_page(page);
1104                         page_cache_release(page);
1105                         continue;
1106                 }
1107
1108                 /* Did somebody else fill it already? */
1109                 if (PageUptodate(page)) {
1110                         unlock_page(page);
1111                         goto page_ok;
1112                 }
1113
1114 readpage:
1115                 /* Start the actual read. The read will unlock the page. */
1116                 error = mapping->a_ops->readpage(filp, page);
1117
1118                 if (unlikely(error)) {
1119                         if (error == AOP_TRUNCATED_PAGE) {
1120                                 page_cache_release(page);
1121                                 goto find_page;
1122                         }
1123                         goto readpage_error;
1124                 }
1125
1126                 if (!PageUptodate(page)) {
1127                         error = lock_page_killable(page);
1128                         if (unlikely(error))
1129                                 goto readpage_error;
1130                         if (!PageUptodate(page)) {
1131                                 if (page->mapping == NULL) {
1132                                         /*
1133                                          * invalidate_inode_pages got it
1134                                          */
1135                                         unlock_page(page);
1136                                         page_cache_release(page);
1137                                         goto find_page;
1138                                 }
1139                                 unlock_page(page);
1140                                 shrink_readahead_size_eio(filp, ra);
1141                                 error = -EIO;
1142                                 goto readpage_error;
1143                         }
1144                         unlock_page(page);
1145                 }
1146
1147                 goto page_ok;
1148
1149 readpage_error:
1150                 /* UHHUH! A synchronous read error occurred. Report it */
1151                 desc->error = error;
1152                 page_cache_release(page);
1153                 goto out;
1154
1155 no_cached_page:
1156                 /*
1157                  * Ok, it wasn't cached, so we need to create a new
1158                  * page..
1159                  */
1160                 page = page_cache_alloc_cold(mapping);
1161                 if (!page) {
1162                         desc->error = -ENOMEM;
1163                         goto out;
1164                 }
1165                 error = add_to_page_cache_lru(page, mapping,
1166                                                 index, GFP_KERNEL);
1167                 if (error) {
1168                         page_cache_release(page);
1169                         if (error == -EEXIST)
1170                                 goto find_page;
1171                         desc->error = error;
1172                         goto out;
1173                 }
1174                 goto readpage;
1175         }
1176
1177 out:
1178         ra->prev_pos = prev_index;
1179         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1180         ra->prev_pos |= prev_offset;
1181
1182         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1183         file_accessed(filp);
1184 }
1185
1186 int file_read_actor(read_descriptor_t *desc, struct page *page,
1187                         unsigned long offset, unsigned long size)
1188 {
1189         char *kaddr;
1190         unsigned long left, count = desc->count;
1191
1192         if (size > count)
1193                 size = count;
1194
1195         /*
1196          * Faults on the destination of a read are common, so do it before
1197          * taking the kmap.
1198          */
1199         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1200                 kaddr = kmap_atomic(page, KM_USER0);
1201                 left = __copy_to_user_inatomic(desc->arg.buf,
1202                                                 kaddr + offset, size);
1203                 kunmap_atomic(kaddr, KM_USER0);
1204                 if (left == 0)
1205                         goto success;
1206         }
1207
1208         /* Do it the slow way */
1209         kaddr = kmap(page);
1210         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1211         kunmap(page);
1212
1213         if (left) {
1214                 size -= left;
1215                 desc->error = -EFAULT;
1216         }
1217 success:
1218         desc->count = count - size;
1219         desc->written += size;
1220         desc->arg.buf += size;
1221         return size;
1222 }
1223
1224 /*
1225  * Performs necessary checks before doing a write
1226  * @iov:        io vector request
1227  * @nr_segs:    number of segments in the iovec
1228  * @count:      number of bytes to write
1229  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1230  *
1231  * Adjust number of segments and amount of bytes to write (nr_segs should be
1232  * properly initialized first). Returns appropriate error code that caller
1233  * should return or zero in case that write should be allowed.
1234  */
1235 int generic_segment_checks(const struct iovec *iov,
1236                         unsigned long *nr_segs, size_t *count, int access_flags)
1237 {
1238         unsigned long   seg;
1239         size_t cnt = 0;
1240         for (seg = 0; seg < *nr_segs; seg++) {
1241                 const struct iovec *iv = &iov[seg];
1242
1243                 /*
1244                  * If any segment has a negative length, or the cumulative
1245                  * length ever wraps negative then return -EINVAL.
1246                  */
1247                 cnt += iv->iov_len;
1248                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1249                         return -EINVAL;
1250                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1251                         continue;
1252                 if (seg == 0)
1253                         return -EFAULT;
1254                 *nr_segs = seg;
1255                 cnt -= iv->iov_len;     /* This segment is no good */
1256                 break;
1257         }
1258         *count = cnt;
1259         return 0;
1260 }
1261 EXPORT_SYMBOL(generic_segment_checks);
1262
1263 /**
1264  * generic_file_aio_read - generic filesystem read routine
1265  * @iocb:       kernel I/O control block
1266  * @iov:        io vector request
1267  * @nr_segs:    number of segments in the iovec
1268  * @pos:        current file position
1269  *
1270  * This is the "read()" routine for all filesystems
1271  * that can use the page cache directly.
1272  */
1273 ssize_t
1274 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1275                 unsigned long nr_segs, loff_t pos)
1276 {
1277         struct file *filp = iocb->ki_filp;
1278         ssize_t retval;
1279         unsigned long seg;
1280         size_t count;
1281         loff_t *ppos = &iocb->ki_pos;
1282
1283         count = 0;
1284         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1285         if (retval)
1286                 return retval;
1287
1288         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1289         if (filp->f_flags & O_DIRECT) {
1290                 loff_t size;
1291                 struct address_space *mapping;
1292                 struct inode *inode;
1293
1294                 mapping = filp->f_mapping;
1295                 inode = mapping->host;
1296                 if (!count)
1297                         goto out; /* skip atime */
1298                 size = i_size_read(inode);
1299                 if (pos < size) {
1300                         retval = filemap_write_and_wait_range(mapping, pos,
1301                                         pos + iov_length(iov, nr_segs) - 1);
1302                         if (!retval) {
1303                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1304                                                         iov, pos, nr_segs);
1305                         }
1306                         if (retval > 0)
1307                                 *ppos = pos + retval;
1308                         if (retval) {
1309                                 file_accessed(filp);
1310                                 goto out;
1311                         }
1312                 }
1313         }
1314
1315         for (seg = 0; seg < nr_segs; seg++) {
1316                 read_descriptor_t desc;
1317
1318                 desc.written = 0;
1319                 desc.arg.buf = iov[seg].iov_base;
1320                 desc.count = iov[seg].iov_len;
1321                 if (desc.count == 0)
1322                         continue;
1323                 desc.error = 0;
1324                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1325                 retval += desc.written;
1326                 if (desc.error) {
1327                         retval = retval ?: desc.error;
1328                         break;
1329                 }
1330                 if (desc.count > 0)
1331                         break;
1332         }
1333 out:
1334         return retval;
1335 }
1336 EXPORT_SYMBOL(generic_file_aio_read);
1337
1338 static ssize_t
1339 do_readahead(struct address_space *mapping, struct file *filp,
1340              pgoff_t index, unsigned long nr)
1341 {
1342         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1343                 return -EINVAL;
1344
1345         force_page_cache_readahead(mapping, filp, index, nr);
1346         return 0;
1347 }
1348
1349 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1350 {
1351         ssize_t ret;
1352         struct file *file;
1353
1354         ret = -EBADF;
1355         file = fget(fd);
1356         if (file) {
1357                 if (file->f_mode & FMODE_READ) {
1358                         struct address_space *mapping = file->f_mapping;
1359                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1360                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1361                         unsigned long len = end - start + 1;
1362                         ret = do_readahead(mapping, file, start, len);
1363                 }
1364                 fput(file);
1365         }
1366         return ret;
1367 }
1368 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1369 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1370 {
1371         return SYSC_readahead((int) fd, offset, (size_t) count);
1372 }
1373 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1374 #endif
1375
1376 #ifdef CONFIG_MMU
1377 /**
1378  * page_cache_read - adds requested page to the page cache if not already there
1379  * @file:       file to read
1380  * @offset:     page index
1381  *
1382  * This adds the requested page to the page cache if it isn't already there,
1383  * and schedules an I/O to read in its contents from disk.
1384  */
1385 static int page_cache_read(struct file *file, pgoff_t offset)
1386 {
1387         struct address_space *mapping = file->f_mapping;
1388         struct page *page; 
1389         int ret;
1390
1391         do {
1392                 page = page_cache_alloc_cold(mapping);
1393                 if (!page)
1394                         return -ENOMEM;
1395
1396                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1397                 if (ret == 0)
1398                         ret = mapping->a_ops->readpage(file, page);
1399                 else if (ret == -EEXIST)
1400                         ret = 0; /* losing race to add is OK */
1401
1402                 page_cache_release(page);
1403
1404         } while (ret == AOP_TRUNCATED_PAGE);
1405                 
1406         return ret;
1407 }
1408
1409 #define MMAP_LOTSAMISS  (100)
1410
1411 /*
1412  * Synchronous readahead happens when we don't even find
1413  * a page in the page cache at all.
1414  */
1415 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1416                                    struct file_ra_state *ra,
1417                                    struct file *file,
1418                                    pgoff_t offset)
1419 {
1420         unsigned long ra_pages;
1421         struct address_space *mapping = file->f_mapping;
1422
1423         /* If we don't want any read-ahead, don't bother */
1424         if (VM_RandomReadHint(vma))
1425                 return;
1426
1427         if (VM_SequentialReadHint(vma) ||
1428                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1429                 page_cache_sync_readahead(mapping, ra, file, offset,
1430                                           ra->ra_pages);
1431                 return;
1432         }
1433
1434         if (ra->mmap_miss < INT_MAX)
1435                 ra->mmap_miss++;
1436
1437         /*
1438          * Do we miss much more than hit in this file? If so,
1439          * stop bothering with read-ahead. It will only hurt.
1440          */
1441         if (ra->mmap_miss > MMAP_LOTSAMISS)
1442                 return;
1443
1444         /*
1445          * mmap read-around
1446          */
1447         ra_pages = max_sane_readahead(ra->ra_pages);
1448         if (ra_pages) {
1449                 ra->start = max_t(long, 0, offset - ra_pages/2);
1450                 ra->size = ra_pages;
1451                 ra->async_size = 0;
1452                 ra_submit(ra, mapping, file);
1453         }
1454 }
1455
1456 /*
1457  * Asynchronous readahead happens when we find the page and PG_readahead,
1458  * so we want to possibly extend the readahead further..
1459  */
1460 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1461                                     struct file_ra_state *ra,
1462                                     struct file *file,
1463                                     struct page *page,
1464                                     pgoff_t offset)
1465 {
1466         struct address_space *mapping = file->f_mapping;
1467
1468         /* If we don't want any read-ahead, don't bother */
1469         if (VM_RandomReadHint(vma))
1470                 return;
1471         if (ra->mmap_miss > 0)
1472                 ra->mmap_miss--;
1473         if (PageReadahead(page))
1474                 page_cache_async_readahead(mapping, ra, file,
1475                                            page, offset, ra->ra_pages);
1476 }
1477
1478 /**
1479  * filemap_fault - read in file data for page fault handling
1480  * @vma:        vma in which the fault was taken
1481  * @vmf:        struct vm_fault containing details of the fault
1482  *
1483  * filemap_fault() is invoked via the vma operations vector for a
1484  * mapped memory region to read in file data during a page fault.
1485  *
1486  * The goto's are kind of ugly, but this streamlines the normal case of having
1487  * it in the page cache, and handles the special cases reasonably without
1488  * having a lot of duplicated code.
1489  */
1490 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1491 {
1492         int error;
1493         struct file *file = vma->vm_file;
1494         struct address_space *mapping = file->f_mapping;
1495         struct file_ra_state *ra = &file->f_ra;
1496         struct inode *inode = mapping->host;
1497         pgoff_t offset = vmf->pgoff;
1498         struct page *page;
1499         pgoff_t size;
1500         int ret = 0;
1501
1502         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1503         if (offset >= size)
1504                 return VM_FAULT_SIGBUS;
1505
1506         /*
1507          * Do we have something in the page cache already?
1508          */
1509         page = find_get_page(mapping, offset);
1510         if (likely(page)) {
1511                 /*
1512                  * We found the page, so try async readahead before
1513                  * waiting for the lock.
1514                  */
1515                 do_async_mmap_readahead(vma, ra, file, page, offset);
1516                 lock_page(page);
1517
1518                 /* Did it get truncated? */
1519                 if (unlikely(page->mapping != mapping)) {
1520                         unlock_page(page);
1521                         put_page(page);
1522                         goto no_cached_page;
1523                 }
1524         } else {
1525                 /* No page in the page cache at all */
1526                 do_sync_mmap_readahead(vma, ra, file, offset);
1527                 count_vm_event(PGMAJFAULT);
1528                 ret = VM_FAULT_MAJOR;
1529 retry_find:
1530                 page = find_lock_page(mapping, offset);
1531                 if (!page)
1532                         goto no_cached_page;
1533         }
1534
1535         /*
1536          * We have a locked page in the page cache, now we need to check
1537          * that it's up-to-date. If not, it is going to be due to an error.
1538          */
1539         if (unlikely(!PageUptodate(page)))
1540                 goto page_not_uptodate;
1541
1542         /*
1543          * Found the page and have a reference on it.
1544          * We must recheck i_size under page lock.
1545          */
1546         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1547         if (unlikely(offset >= size)) {
1548                 unlock_page(page);
1549                 page_cache_release(page);
1550                 return VM_FAULT_SIGBUS;
1551         }
1552
1553         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1554         vmf->page = page;
1555         return ret | VM_FAULT_LOCKED;
1556
1557 no_cached_page:
1558         /*
1559          * We're only likely to ever get here if MADV_RANDOM is in
1560          * effect.
1561          */
1562         error = page_cache_read(file, offset);
1563
1564         /*
1565          * The page we want has now been added to the page cache.
1566          * In the unlikely event that someone removed it in the
1567          * meantime, we'll just come back here and read it again.
1568          */
1569         if (error >= 0)
1570                 goto retry_find;
1571
1572         /*
1573          * An error return from page_cache_read can result if the
1574          * system is low on memory, or a problem occurs while trying
1575          * to schedule I/O.
1576          */
1577         if (error == -ENOMEM)
1578                 return VM_FAULT_OOM;
1579         return VM_FAULT_SIGBUS;
1580
1581 page_not_uptodate:
1582         /*
1583          * Umm, take care of errors if the page isn't up-to-date.
1584          * Try to re-read it _once_. We do this synchronously,
1585          * because there really aren't any performance issues here
1586          * and we need to check for errors.
1587          */
1588         ClearPageError(page);
1589         error = mapping->a_ops->readpage(file, page);
1590         if (!error) {
1591                 wait_on_page_locked(page);
1592                 if (!PageUptodate(page))
1593                         error = -EIO;
1594         }
1595         page_cache_release(page);
1596
1597         if (!error || error == AOP_TRUNCATED_PAGE)
1598                 goto retry_find;
1599
1600         /* Things didn't work out. Return zero to tell the mm layer so. */
1601         shrink_readahead_size_eio(file, ra);
1602         return VM_FAULT_SIGBUS;
1603 }
1604 EXPORT_SYMBOL(filemap_fault);
1605
1606 struct vm_operations_struct generic_file_vm_ops = {
1607         .fault          = filemap_fault,
1608 };
1609
1610 /* This is used for a general mmap of a disk file */
1611
1612 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1613 {
1614         struct address_space *mapping = file->f_mapping;
1615
1616         if (!mapping->a_ops->readpage)
1617                 return -ENOEXEC;
1618         file_accessed(file);
1619         vma->vm_ops = &generic_file_vm_ops;
1620         vma->vm_flags |= VM_CAN_NONLINEAR;
1621         return 0;
1622 }
1623
1624 /*
1625  * This is for filesystems which do not implement ->writepage.
1626  */
1627 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1628 {
1629         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1630                 return -EINVAL;
1631         return generic_file_mmap(file, vma);
1632 }
1633 #else
1634 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1635 {
1636         return -ENOSYS;
1637 }
1638 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1639 {
1640         return -ENOSYS;
1641 }
1642 #endif /* CONFIG_MMU */
1643
1644 EXPORT_SYMBOL(generic_file_mmap);
1645 EXPORT_SYMBOL(generic_file_readonly_mmap);
1646
1647 static struct page *__read_cache_page(struct address_space *mapping,
1648                                 pgoff_t index,
1649                                 int (*filler)(void *,struct page*),
1650                                 void *data)
1651 {
1652         struct page *page;
1653         int err;
1654 repeat:
1655         page = find_get_page(mapping, index);
1656         if (!page) {
1657                 page = page_cache_alloc_cold(mapping);
1658                 if (!page)
1659                         return ERR_PTR(-ENOMEM);
1660                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1661                 if (unlikely(err)) {
1662                         page_cache_release(page);
1663                         if (err == -EEXIST)
1664                                 goto repeat;
1665                         /* Presumably ENOMEM for radix tree node */
1666                         return ERR_PTR(err);
1667                 }
1668                 err = filler(data, page);
1669                 if (err < 0) {
1670                         page_cache_release(page);
1671                         page = ERR_PTR(err);
1672                 }
1673         }
1674         return page;
1675 }
1676
1677 /**
1678  * read_cache_page_async - read into page cache, fill it if needed
1679  * @mapping:    the page's address_space
1680  * @index:      the page index
1681  * @filler:     function to perform the read
1682  * @data:       destination for read data
1683  *
1684  * Same as read_cache_page, but don't wait for page to become unlocked
1685  * after submitting it to the filler.
1686  *
1687  * Read into the page cache. If a page already exists, and PageUptodate() is
1688  * not set, try to fill the page but don't wait for it to become unlocked.
1689  *
1690  * If the page does not get brought uptodate, return -EIO.
1691  */
1692 struct page *read_cache_page_async(struct address_space *mapping,
1693                                 pgoff_t index,
1694                                 int (*filler)(void *,struct page*),
1695                                 void *data)
1696 {
1697         struct page *page;
1698         int err;
1699
1700 retry:
1701         page = __read_cache_page(mapping, index, filler, data);
1702         if (IS_ERR(page))
1703                 return page;
1704         if (PageUptodate(page))
1705                 goto out;
1706
1707         lock_page(page);
1708         if (!page->mapping) {
1709                 unlock_page(page);
1710                 page_cache_release(page);
1711                 goto retry;
1712         }
1713         if (PageUptodate(page)) {
1714                 unlock_page(page);
1715                 goto out;
1716         }
1717         err = filler(data, page);
1718         if (err < 0) {
1719                 page_cache_release(page);
1720                 return ERR_PTR(err);
1721         }
1722 out:
1723         mark_page_accessed(page);
1724         return page;
1725 }
1726 EXPORT_SYMBOL(read_cache_page_async);
1727
1728 /**
1729  * read_cache_page - read into page cache, fill it if needed
1730  * @mapping:    the page's address_space
1731  * @index:      the page index
1732  * @filler:     function to perform the read
1733  * @data:       destination for read data
1734  *
1735  * Read into the page cache. If a page already exists, and PageUptodate() is
1736  * not set, try to fill the page then wait for it to become unlocked.
1737  *
1738  * If the page does not get brought uptodate, return -EIO.
1739  */
1740 struct page *read_cache_page(struct address_space *mapping,
1741                                 pgoff_t index,
1742                                 int (*filler)(void *,struct page*),
1743                                 void *data)
1744 {
1745         struct page *page;
1746
1747         page = read_cache_page_async(mapping, index, filler, data);
1748         if (IS_ERR(page))
1749                 goto out;
1750         wait_on_page_locked(page);
1751         if (!PageUptodate(page)) {
1752                 page_cache_release(page);
1753                 page = ERR_PTR(-EIO);
1754         }
1755  out:
1756         return page;
1757 }
1758 EXPORT_SYMBOL(read_cache_page);
1759
1760 /*
1761  * The logic we want is
1762  *
1763  *      if suid or (sgid and xgrp)
1764  *              remove privs
1765  */
1766 int should_remove_suid(struct dentry *dentry)
1767 {
1768         mode_t mode = dentry->d_inode->i_mode;
1769         int kill = 0;
1770
1771         /* suid always must be killed */
1772         if (unlikely(mode & S_ISUID))
1773                 kill = ATTR_KILL_SUID;
1774
1775         /*
1776          * sgid without any exec bits is just a mandatory locking mark; leave
1777          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1778          */
1779         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1780                 kill |= ATTR_KILL_SGID;
1781
1782         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1783                 return kill;
1784
1785         return 0;
1786 }
1787 EXPORT_SYMBOL(should_remove_suid);
1788
1789 static int __remove_suid(struct dentry *dentry, int kill)
1790 {
1791         struct iattr newattrs;
1792
1793         newattrs.ia_valid = ATTR_FORCE | kill;
1794         return notify_change(dentry, &newattrs);
1795 }
1796
1797 int file_remove_suid(struct file *file)
1798 {
1799         struct dentry *dentry = file->f_path.dentry;
1800         int killsuid = should_remove_suid(dentry);
1801         int killpriv = security_inode_need_killpriv(dentry);
1802         int error = 0;
1803
1804         if (killpriv < 0)
1805                 return killpriv;
1806         if (killpriv)
1807                 error = security_inode_killpriv(dentry);
1808         if (!error && killsuid)
1809                 error = __remove_suid(dentry, killsuid);
1810
1811         return error;
1812 }
1813 EXPORT_SYMBOL(file_remove_suid);
1814
1815 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1816                         const struct iovec *iov, size_t base, size_t bytes)
1817 {
1818         size_t copied = 0, left = 0;
1819
1820         while (bytes) {
1821                 char __user *buf = iov->iov_base + base;
1822                 int copy = min(bytes, iov->iov_len - base);
1823
1824                 base = 0;
1825                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1826                 copied += copy;
1827                 bytes -= copy;
1828                 vaddr += copy;
1829                 iov++;
1830
1831                 if (unlikely(left))
1832                         break;
1833         }
1834         return copied - left;
1835 }
1836
1837 /*
1838  * Copy as much as we can into the page and return the number of bytes which
1839  * were sucessfully copied.  If a fault is encountered then return the number of
1840  * bytes which were copied.
1841  */
1842 size_t iov_iter_copy_from_user_atomic(struct page *page,
1843                 struct iov_iter *i, unsigned long offset, size_t bytes)
1844 {
1845         char *kaddr;
1846         size_t copied;
1847
1848         BUG_ON(!in_atomic());
1849         kaddr = kmap_atomic(page, KM_USER0);
1850         if (likely(i->nr_segs == 1)) {
1851                 int left;
1852                 char __user *buf = i->iov->iov_base + i->iov_offset;
1853                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1854                 copied = bytes - left;
1855         } else {
1856                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1857                                                 i->iov, i->iov_offset, bytes);
1858         }
1859         kunmap_atomic(kaddr, KM_USER0);
1860
1861         return copied;
1862 }
1863 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1864
1865 /*
1866  * This has the same sideeffects and return value as
1867  * iov_iter_copy_from_user_atomic().
1868  * The difference is that it attempts to resolve faults.
1869  * Page must not be locked.
1870  */
1871 size_t iov_iter_copy_from_user(struct page *page,
1872                 struct iov_iter *i, unsigned long offset, size_t bytes)
1873 {
1874         char *kaddr;
1875         size_t copied;
1876
1877         kaddr = kmap(page);
1878         if (likely(i->nr_segs == 1)) {
1879                 int left;
1880                 char __user *buf = i->iov->iov_base + i->iov_offset;
1881                 left = __copy_from_user(kaddr + offset, buf, bytes);
1882                 copied = bytes - left;
1883         } else {
1884                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1885                                                 i->iov, i->iov_offset, bytes);
1886         }
1887         kunmap(page);
1888         return copied;
1889 }
1890 EXPORT_SYMBOL(iov_iter_copy_from_user);
1891
1892 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1893 {
1894         BUG_ON(i->count < bytes);
1895
1896         if (likely(i->nr_segs == 1)) {
1897                 i->iov_offset += bytes;
1898                 i->count -= bytes;
1899         } else {
1900                 const struct iovec *iov = i->iov;
1901                 size_t base = i->iov_offset;
1902
1903                 /*
1904                  * The !iov->iov_len check ensures we skip over unlikely
1905                  * zero-length segments (without overruning the iovec).
1906                  */
1907                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1908                         int copy;
1909
1910                         copy = min(bytes, iov->iov_len - base);
1911                         BUG_ON(!i->count || i->count < copy);
1912                         i->count -= copy;
1913                         bytes -= copy;
1914                         base += copy;
1915                         if (iov->iov_len == base) {
1916                                 iov++;
1917                                 base = 0;
1918                         }
1919                 }
1920                 i->iov = iov;
1921                 i->iov_offset = base;
1922         }
1923 }
1924 EXPORT_SYMBOL(iov_iter_advance);
1925
1926 /*
1927  * Fault in the first iovec of the given iov_iter, to a maximum length
1928  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1929  * accessed (ie. because it is an invalid address).
1930  *
1931  * writev-intensive code may want this to prefault several iovecs -- that
1932  * would be possible (callers must not rely on the fact that _only_ the
1933  * first iovec will be faulted with the current implementation).
1934  */
1935 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1936 {
1937         char __user *buf = i->iov->iov_base + i->iov_offset;
1938         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1939         return fault_in_pages_readable(buf, bytes);
1940 }
1941 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1942
1943 /*
1944  * Return the count of just the current iov_iter segment.
1945  */
1946 size_t iov_iter_single_seg_count(struct iov_iter *i)
1947 {
1948         const struct iovec *iov = i->iov;
1949         if (i->nr_segs == 1)
1950                 return i->count;
1951         else
1952                 return min(i->count, iov->iov_len - i->iov_offset);
1953 }
1954 EXPORT_SYMBOL(iov_iter_single_seg_count);
1955
1956 /*
1957  * Performs necessary checks before doing a write
1958  *
1959  * Can adjust writing position or amount of bytes to write.
1960  * Returns appropriate error code that caller should return or
1961  * zero in case that write should be allowed.
1962  */
1963 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1964 {
1965         struct inode *inode = file->f_mapping->host;
1966         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1967
1968         if (unlikely(*pos < 0))
1969                 return -EINVAL;
1970
1971         if (!isblk) {
1972                 /* FIXME: this is for backwards compatibility with 2.4 */
1973                 if (file->f_flags & O_APPEND)
1974                         *pos = i_size_read(inode);
1975
1976                 if (limit != RLIM_INFINITY) {
1977                         if (*pos >= limit) {
1978                                 send_sig(SIGXFSZ, current, 0);
1979                                 return -EFBIG;
1980                         }
1981                         if (*count > limit - (typeof(limit))*pos) {
1982                                 *count = limit - (typeof(limit))*pos;
1983                         }
1984                 }
1985         }
1986
1987         /*
1988          * LFS rule
1989          */
1990         if (unlikely(*pos + *count > MAX_NON_LFS &&
1991                                 !(file->f_flags & O_LARGEFILE))) {
1992                 if (*pos >= MAX_NON_LFS) {
1993                         return -EFBIG;
1994                 }
1995                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1996                         *count = MAX_NON_LFS - (unsigned long)*pos;
1997                 }
1998         }
1999
2000         /*
2001          * Are we about to exceed the fs block limit ?
2002          *
2003          * If we have written data it becomes a short write.  If we have
2004          * exceeded without writing data we send a signal and return EFBIG.
2005          * Linus frestrict idea will clean these up nicely..
2006          */
2007         if (likely(!isblk)) {
2008                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2009                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2010                                 return -EFBIG;
2011                         }
2012                         /* zero-length writes at ->s_maxbytes are OK */
2013                 }
2014
2015                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2016                         *count = inode->i_sb->s_maxbytes - *pos;
2017         } else {
2018 #ifdef CONFIG_BLOCK
2019                 loff_t isize;
2020                 if (bdev_read_only(I_BDEV(inode)))
2021                         return -EPERM;
2022                 isize = i_size_read(inode);
2023                 if (*pos >= isize) {
2024                         if (*count || *pos > isize)
2025                                 return -ENOSPC;
2026                 }
2027
2028                 if (*pos + *count > isize)
2029                         *count = isize - *pos;
2030 #else
2031                 return -EPERM;
2032 #endif
2033         }
2034         return 0;
2035 }
2036 EXPORT_SYMBOL(generic_write_checks);
2037
2038 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2039                                 loff_t pos, unsigned len, unsigned flags,
2040                                 struct page **pagep, void **fsdata)
2041 {
2042         const struct address_space_operations *aops = mapping->a_ops;
2043
2044         return aops->write_begin(file, mapping, pos, len, flags,
2045                                                         pagep, fsdata);
2046 }
2047 EXPORT_SYMBOL(pagecache_write_begin);
2048
2049 int pagecache_write_end(struct file *file, struct address_space *mapping,
2050                                 loff_t pos, unsigned len, unsigned copied,
2051                                 struct page *page, void *fsdata)
2052 {
2053         const struct address_space_operations *aops = mapping->a_ops;
2054
2055         mark_page_accessed(page);
2056         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2057 }
2058 EXPORT_SYMBOL(pagecache_write_end);
2059
2060 ssize_t
2061 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2062                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2063                 size_t count, size_t ocount)
2064 {
2065         struct file     *file = iocb->ki_filp;
2066         struct address_space *mapping = file->f_mapping;
2067         struct inode    *inode = mapping->host;
2068         ssize_t         written;
2069         size_t          write_len;
2070         pgoff_t         end;
2071
2072         if (count != ocount)
2073                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2074
2075         write_len = iov_length(iov, *nr_segs);
2076         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2077
2078         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2079         if (written)
2080                 goto out;
2081
2082         /*
2083          * After a write we want buffered reads to be sure to go to disk to get
2084          * the new data.  We invalidate clean cached page from the region we're
2085          * about to write.  We do this *before* the write so that we can return
2086          * without clobbering -EIOCBQUEUED from ->direct_IO().
2087          */
2088         if (mapping->nrpages) {
2089                 written = invalidate_inode_pages2_range(mapping,
2090                                         pos >> PAGE_CACHE_SHIFT, end);
2091                 /*
2092                  * If a page can not be invalidated, return 0 to fall back
2093                  * to buffered write.
2094                  */
2095                 if (written) {
2096                         if (written == -EBUSY)
2097                                 return 0;
2098                         goto out;
2099                 }
2100         }
2101
2102         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2103
2104         /*
2105          * Finally, try again to invalidate clean pages which might have been
2106          * cached by non-direct readahead, or faulted in by get_user_pages()
2107          * if the source of the write was an mmap'ed region of the file
2108          * we're writing.  Either one is a pretty crazy thing to do,
2109          * so we don't support it 100%.  If this invalidation
2110          * fails, tough, the write still worked...
2111          */
2112         if (mapping->nrpages) {
2113                 invalidate_inode_pages2_range(mapping,
2114                                               pos >> PAGE_CACHE_SHIFT, end);
2115         }
2116
2117         if (written > 0) {
2118                 loff_t end = pos + written;
2119                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2120                         i_size_write(inode,  end);
2121                         mark_inode_dirty(inode);
2122                 }
2123                 *ppos = end;
2124         }
2125 out:
2126         return written;
2127 }
2128 EXPORT_SYMBOL(generic_file_direct_write);
2129
2130 /*
2131  * Find or create a page at the given pagecache position. Return the locked
2132  * page. This function is specifically for buffered writes.
2133  */
2134 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2135                                         pgoff_t index, unsigned flags)
2136 {
2137         int status;
2138         struct page *page;
2139         gfp_t gfp_notmask = 0;
2140         if (flags & AOP_FLAG_NOFS)
2141                 gfp_notmask = __GFP_FS;
2142 repeat:
2143         page = find_lock_page(mapping, index);
2144         if (likely(page))
2145                 return page;
2146
2147         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2148         if (!page)
2149                 return NULL;
2150         status = add_to_page_cache_lru(page, mapping, index,
2151                                                 GFP_KERNEL & ~gfp_notmask);
2152         if (unlikely(status)) {
2153                 page_cache_release(page);
2154                 if (status == -EEXIST)
2155                         goto repeat;
2156                 return NULL;
2157         }
2158         return page;
2159 }
2160 EXPORT_SYMBOL(grab_cache_page_write_begin);
2161
2162 static ssize_t generic_perform_write(struct file *file,
2163                                 struct iov_iter *i, loff_t pos)
2164 {
2165         struct address_space *mapping = file->f_mapping;
2166         const struct address_space_operations *a_ops = mapping->a_ops;
2167         long status = 0;
2168         ssize_t written = 0;
2169         unsigned int flags = 0;
2170
2171         /*
2172          * Copies from kernel address space cannot fail (NFSD is a big user).
2173          */
2174         if (segment_eq(get_fs(), KERNEL_DS))
2175                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2176
2177         do {
2178                 struct page *page;
2179                 pgoff_t index;          /* Pagecache index for current page */
2180                 unsigned long offset;   /* Offset into pagecache page */
2181                 unsigned long bytes;    /* Bytes to write to page */
2182                 size_t copied;          /* Bytes copied from user */
2183                 void *fsdata;
2184
2185                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2186                 index = pos >> PAGE_CACHE_SHIFT;
2187                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2188                                                 iov_iter_count(i));
2189
2190 again:
2191
2192                 /*
2193                  * Bring in the user page that we will copy from _first_.
2194                  * Otherwise there's a nasty deadlock on copying from the
2195                  * same page as we're writing to, without it being marked
2196                  * up-to-date.
2197                  *
2198                  * Not only is this an optimisation, but it is also required
2199                  * to check that the address is actually valid, when atomic
2200                  * usercopies are used, below.
2201                  */
2202                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2203                         status = -EFAULT;
2204                         break;
2205                 }
2206
2207                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2208                                                 &page, &fsdata);
2209                 if (unlikely(status))
2210                         break;
2211
2212                 pagefault_disable();
2213                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2214                 pagefault_enable();
2215                 flush_dcache_page(page);
2216
2217                 mark_page_accessed(page);
2218                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2219                                                 page, fsdata);
2220                 if (unlikely(status < 0))
2221                         break;
2222                 copied = status;
2223
2224                 cond_resched();
2225
2226                 iov_iter_advance(i, copied);
2227                 if (unlikely(copied == 0)) {
2228                         /*
2229                          * If we were unable to copy any data at all, we must
2230                          * fall back to a single segment length write.
2231                          *
2232                          * If we didn't fallback here, we could livelock
2233                          * because not all segments in the iov can be copied at
2234                          * once without a pagefault.
2235                          */
2236                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2237                                                 iov_iter_single_seg_count(i));
2238                         goto again;
2239                 }
2240                 pos += copied;
2241                 written += copied;
2242
2243                 balance_dirty_pages_ratelimited(mapping);
2244
2245         } while (iov_iter_count(i));
2246
2247         return written ? written : status;
2248 }
2249
2250 ssize_t
2251 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2252                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2253                 size_t count, ssize_t written)
2254 {
2255         struct file *file = iocb->ki_filp;
2256         struct address_space *mapping = file->f_mapping;
2257         ssize_t status;
2258         struct iov_iter i;
2259
2260         iov_iter_init(&i, iov, nr_segs, count, written);
2261         status = generic_perform_write(file, &i, pos);
2262
2263         if (likely(status >= 0)) {
2264                 written += status;
2265                 *ppos = pos + status;
2266         }
2267         
2268         /*
2269          * If we get here for O_DIRECT writes then we must have fallen through
2270          * to buffered writes (block instantiation inside i_size).  So we sync
2271          * the file data here, to try to honour O_DIRECT expectations.
2272          */
2273         if (unlikely(file->f_flags & O_DIRECT) && written)
2274                 status = filemap_write_and_wait_range(mapping,
2275                                         pos, pos + written - 1);
2276
2277         return written ? written : status;
2278 }
2279 EXPORT_SYMBOL(generic_file_buffered_write);
2280
2281 /**
2282  * __generic_file_aio_write - write data to a file
2283  * @iocb:       IO state structure (file, offset, etc.)
2284  * @iov:        vector with data to write
2285  * @nr_segs:    number of segments in the vector
2286  * @ppos:       position where to write
2287  *
2288  * This function does all the work needed for actually writing data to a
2289  * file. It does all basic checks, removes SUID from the file, updates
2290  * modification times and calls proper subroutines depending on whether we
2291  * do direct IO or a standard buffered write.
2292  *
2293  * It expects i_mutex to be grabbed unless we work on a block device or similar
2294  * object which does not need locking at all.
2295  *
2296  * This function does *not* take care of syncing data in case of O_SYNC write.
2297  * A caller has to handle it. This is mainly due to the fact that we want to
2298  * avoid syncing under i_mutex.
2299  */
2300 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2301                                  unsigned long nr_segs, loff_t *ppos)
2302 {
2303         struct file *file = iocb->ki_filp;
2304         struct address_space * mapping = file->f_mapping;
2305         size_t ocount;          /* original count */
2306         size_t count;           /* after file limit checks */
2307         struct inode    *inode = mapping->host;
2308         loff_t          pos;
2309         ssize_t         written;
2310         ssize_t         err;
2311
2312         ocount = 0;
2313         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2314         if (err)
2315                 return err;
2316
2317         count = ocount;
2318         pos = *ppos;
2319
2320         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2321
2322         /* We can write back this queue in page reclaim */
2323         current->backing_dev_info = mapping->backing_dev_info;
2324         written = 0;
2325
2326         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2327         if (err)
2328                 goto out;
2329
2330         if (count == 0)
2331                 goto out;
2332
2333         err = file_remove_suid(file);
2334         if (err)
2335                 goto out;
2336
2337         file_update_time(file);
2338
2339         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2340         if (unlikely(file->f_flags & O_DIRECT)) {
2341                 loff_t endbyte;
2342                 ssize_t written_buffered;
2343
2344                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2345                                                         ppos, count, ocount);
2346                 if (written < 0 || written == count)
2347                         goto out;
2348                 /*
2349                  * direct-io write to a hole: fall through to buffered I/O
2350                  * for completing the rest of the request.
2351                  */
2352                 pos += written;
2353                 count -= written;
2354                 written_buffered = generic_file_buffered_write(iocb, iov,
2355                                                 nr_segs, pos, ppos, count,
2356                                                 written);
2357                 /*
2358                  * If generic_file_buffered_write() retuned a synchronous error
2359                  * then we want to return the number of bytes which were
2360                  * direct-written, or the error code if that was zero.  Note
2361                  * that this differs from normal direct-io semantics, which
2362                  * will return -EFOO even if some bytes were written.
2363                  */
2364                 if (written_buffered < 0) {
2365                         err = written_buffered;
2366                         goto out;
2367                 }
2368
2369                 /*
2370                  * We need to ensure that the page cache pages are written to
2371                  * disk and invalidated to preserve the expected O_DIRECT
2372                  * semantics.
2373                  */
2374                 endbyte = pos + written_buffered - written - 1;
2375                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2376                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2377                                             SYNC_FILE_RANGE_WRITE|
2378                                             SYNC_FILE_RANGE_WAIT_AFTER);
2379                 if (err == 0) {
2380                         written = written_buffered;
2381                         invalidate_mapping_pages(mapping,
2382                                                  pos >> PAGE_CACHE_SHIFT,
2383                                                  endbyte >> PAGE_CACHE_SHIFT);
2384                 } else {
2385                         /*
2386                          * We don't know how much we wrote, so just return
2387                          * the number of bytes which were direct-written
2388                          */
2389                 }
2390         } else {
2391                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2392                                 pos, ppos, count, written);
2393         }
2394 out:
2395         current->backing_dev_info = NULL;
2396         return written ? written : err;
2397 }
2398 EXPORT_SYMBOL(__generic_file_aio_write);
2399
2400 /**
2401  * generic_file_aio_write - write data to a file
2402  * @iocb:       IO state structure
2403  * @iov:        vector with data to write
2404  * @nr_segs:    number of segments in the vector
2405  * @pos:        position in file where to write
2406  *
2407  * This is a wrapper around __generic_file_aio_write() to be used by most
2408  * filesystems. It takes care of syncing the file in case of O_SYNC file
2409  * and acquires i_mutex as needed.
2410  */
2411 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2412                 unsigned long nr_segs, loff_t pos)
2413 {
2414         struct file *file = iocb->ki_filp;
2415         struct inode *inode = file->f_mapping->host;
2416         ssize_t ret;
2417
2418         BUG_ON(iocb->ki_pos != pos);
2419
2420         mutex_lock(&inode->i_mutex);
2421         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2422         mutex_unlock(&inode->i_mutex);
2423
2424         if (ret > 0 || ret == -EIOCBQUEUED) {
2425                 ssize_t err;
2426
2427                 err = generic_write_sync(file, pos, ret);
2428                 if (err < 0 && ret > 0)
2429                         ret = err;
2430         }
2431         return ret;
2432 }
2433 EXPORT_SYMBOL(generic_file_aio_write);
2434
2435 /**
2436  * try_to_release_page() - release old fs-specific metadata on a page
2437  *
2438  * @page: the page which the kernel is trying to free
2439  * @gfp_mask: memory allocation flags (and I/O mode)
2440  *
2441  * The address_space is to try to release any data against the page
2442  * (presumably at page->private).  If the release was successful, return `1'.
2443  * Otherwise return zero.
2444  *
2445  * This may also be called if PG_fscache is set on a page, indicating that the
2446  * page is known to the local caching routines.
2447  *
2448  * The @gfp_mask argument specifies whether I/O may be performed to release
2449  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2450  *
2451  */
2452 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2453 {
2454         struct address_space * const mapping = page->mapping;
2455
2456         BUG_ON(!PageLocked(page));
2457         if (PageWriteback(page))
2458                 return 0;
2459
2460         if (mapping && mapping->a_ops->releasepage)
2461                 return mapping->a_ops->releasepage(page, gfp_mask);
2462         return try_to_free_buffers(page);
2463 }
2464
2465 EXPORT_SYMBOL(try_to_release_page);