a1ddd2557af230fb70d784e81406e63596bbe984
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for generic_osync_inode */
43
44 #include <asm/mman.h>
45
46
47 /*
48  * Shared mappings implemented 30.11.1994. It's not fully working yet,
49  * though.
50  *
51  * Shared mappings now work. 15.8.1995  Bruno.
52  *
53  * finished 'unifying' the page and buffer cache and SMP-threaded the
54  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
55  *
56  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
57  */
58
59 /*
60  * Lock ordering:
61  *
62  *  ->i_mmap_lock               (vmtruncate)
63  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
64  *      ->swap_lock             (exclusive_swap_page, others)
65  *        ->mapping->tree_lock
66  *
67  *  ->i_mutex
68  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
69  *
70  *  ->mmap_sem
71  *    ->i_mmap_lock
72  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
73  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
74  *
75  *  ->mmap_sem
76  *    ->lock_page               (access_process_vm)
77  *
78  *  ->i_mutex                   (generic_file_buffered_write)
79  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
80  *
81  *  ->i_mutex
82  *    ->i_alloc_sem             (various)
83  *
84  *  ->inode_lock
85  *    ->sb_lock                 (fs/fs-writeback.c)
86  *    ->mapping->tree_lock      (__sync_single_inode)
87  *
88  *  ->i_mmap_lock
89  *    ->anon_vma.lock           (vma_adjust)
90  *
91  *  ->anon_vma.lock
92  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
93  *
94  *  ->page_table_lock or pte_lock
95  *    ->swap_lock               (try_to_unmap_one)
96  *    ->private_lock            (try_to_unmap_one)
97  *    ->tree_lock               (try_to_unmap_one)
98  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
99  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
100  *    ->private_lock            (page_remove_rmap->set_page_dirty)
101  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
102  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
103  *    ->inode_lock              (zap_pte_range->set_page_dirty)
104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
105  *
106  *  ->task->proc_lock
107  *    ->dcache_lock             (proc_pid_lookup)
108  */
109
110 /*
111  * Remove a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold the mapping's tree_lock.
114  */
115 void __remove_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         mem_cgroup_uncharge_cache_page(page);
120         radix_tree_delete(&mapping->page_tree, page->index);
121         page->mapping = NULL;
122         mapping->nrpages--;
123         __dec_zone_page_state(page, NR_FILE_PAGES);
124         BUG_ON(page_mapped(page));
125
126         /*
127          * Some filesystems seem to re-dirty the page even after
128          * the VM has canceled the dirty bit (eg ext3 journaling).
129          *
130          * Fix it up by doing a final dirty accounting check after
131          * having removed the page entirely.
132          */
133         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
134                 dec_zone_page_state(page, NR_FILE_DIRTY);
135                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
136         }
137 }
138
139 void remove_from_page_cache(struct page *page)
140 {
141         struct address_space *mapping = page->mapping;
142
143         BUG_ON(!PageLocked(page));
144
145         spin_lock_irq(&mapping->tree_lock);
146         __remove_from_page_cache(page);
147         spin_unlock_irq(&mapping->tree_lock);
148 }
149
150 static int sync_page(void *word)
151 {
152         struct address_space *mapping;
153         struct page *page;
154
155         page = container_of((unsigned long *)word, struct page, flags);
156
157         /*
158          * page_mapping() is being called without PG_locked held.
159          * Some knowledge of the state and use of the page is used to
160          * reduce the requirements down to a memory barrier.
161          * The danger here is of a stale page_mapping() return value
162          * indicating a struct address_space different from the one it's
163          * associated with when it is associated with one.
164          * After smp_mb(), it's either the correct page_mapping() for
165          * the page, or an old page_mapping() and the page's own
166          * page_mapping() has gone NULL.
167          * The ->sync_page() address_space operation must tolerate
168          * page_mapping() going NULL. By an amazing coincidence,
169          * this comes about because none of the users of the page
170          * in the ->sync_page() methods make essential use of the
171          * page_mapping(), merely passing the page down to the backing
172          * device's unplug functions when it's non-NULL, which in turn
173          * ignore it for all cases but swap, where only page_private(page) is
174          * of interest. When page_mapping() does go NULL, the entire
175          * call stack gracefully ignores the page and returns.
176          * -- wli
177          */
178         smp_mb();
179         mapping = page_mapping(page);
180         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
181                 mapping->a_ops->sync_page(page);
182         io_schedule();
183         return 0;
184 }
185
186 static int sync_page_killable(void *word)
187 {
188         sync_page(word);
189         return fatal_signal_pending(current) ? -EINTR : 0;
190 }
191
192 /**
193  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
194  * @mapping:    address space structure to write
195  * @start:      offset in bytes where the range starts
196  * @end:        offset in bytes where the range ends (inclusive)
197  * @sync_mode:  enable synchronous operation
198  *
199  * Start writeback against all of a mapping's dirty pages that lie
200  * within the byte offsets <start, end> inclusive.
201  *
202  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
203  * opposed to a regular memory cleansing writeback.  The difference between
204  * these two operations is that if a dirty page/buffer is encountered, it must
205  * be waited upon, and not just skipped over.
206  */
207 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
208                                 loff_t end, int sync_mode)
209 {
210         int ret;
211         struct writeback_control wbc = {
212                 .sync_mode = sync_mode,
213                 .nr_to_write = mapping->nrpages * 2,
214                 .range_start = start,
215                 .range_end = end,
216         };
217
218         if (!mapping_cap_writeback_dirty(mapping))
219                 return 0;
220
221         ret = do_writepages(mapping, &wbc);
222         return ret;
223 }
224
225 static inline int __filemap_fdatawrite(struct address_space *mapping,
226         int sync_mode)
227 {
228         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
229 }
230
231 int filemap_fdatawrite(struct address_space *mapping)
232 {
233         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
234 }
235 EXPORT_SYMBOL(filemap_fdatawrite);
236
237 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
238                                 loff_t end)
239 {
240         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
241 }
242 EXPORT_SYMBOL(filemap_fdatawrite_range);
243
244 /**
245  * filemap_flush - mostly a non-blocking flush
246  * @mapping:    target address_space
247  *
248  * This is a mostly non-blocking flush.  Not suitable for data-integrity
249  * purposes - I/O may not be started against all dirty pages.
250  */
251 int filemap_flush(struct address_space *mapping)
252 {
253         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
254 }
255 EXPORT_SYMBOL(filemap_flush);
256
257 /**
258  * wait_on_page_writeback_range - wait for writeback to complete
259  * @mapping:    target address_space
260  * @start:      beginning page index
261  * @end:        ending page index
262  *
263  * Wait for writeback to complete against pages indexed by start->end
264  * inclusive
265  */
266 int wait_on_page_writeback_range(struct address_space *mapping,
267                                 pgoff_t start, pgoff_t end)
268 {
269         struct pagevec pvec;
270         int nr_pages;
271         int ret = 0;
272         pgoff_t index;
273
274         if (end < start)
275                 return 0;
276
277         pagevec_init(&pvec, 0);
278         index = start;
279         while ((index <= end) &&
280                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
281                         PAGECACHE_TAG_WRITEBACK,
282                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
283                 unsigned i;
284
285                 for (i = 0; i < nr_pages; i++) {
286                         struct page *page = pvec.pages[i];
287
288                         /* until radix tree lookup accepts end_index */
289                         if (page->index > end)
290                                 continue;
291
292                         wait_on_page_writeback(page);
293                         if (PageError(page))
294                                 ret = -EIO;
295                 }
296                 pagevec_release(&pvec);
297                 cond_resched();
298         }
299
300         /* Check for outstanding write errors */
301         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
302                 ret = -ENOSPC;
303         if (test_and_clear_bit(AS_EIO, &mapping->flags))
304                 ret = -EIO;
305
306         return ret;
307 }
308
309 /**
310  * sync_page_range - write and wait on all pages in the passed range
311  * @inode:      target inode
312  * @mapping:    target address_space
313  * @pos:        beginning offset in pages to write
314  * @count:      number of bytes to write
315  *
316  * Write and wait upon all the pages in the passed range.  This is a "data
317  * integrity" operation.  It waits upon in-flight writeout before starting and
318  * waiting upon new writeout.  If there was an IO error, return it.
319  *
320  * We need to re-take i_mutex during the generic_osync_inode list walk because
321  * it is otherwise livelockable.
322  */
323 int sync_page_range(struct inode *inode, struct address_space *mapping,
324                         loff_t pos, loff_t count)
325 {
326         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
327         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
328         int ret;
329
330         if (!mapping_cap_writeback_dirty(mapping) || !count)
331                 return 0;
332         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
333         if (ret == 0) {
334                 mutex_lock(&inode->i_mutex);
335                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
336                 mutex_unlock(&inode->i_mutex);
337         }
338         if (ret == 0)
339                 ret = wait_on_page_writeback_range(mapping, start, end);
340         return ret;
341 }
342 EXPORT_SYMBOL(sync_page_range);
343
344 /**
345  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
346  * @inode:      target inode
347  * @mapping:    target address_space
348  * @pos:        beginning offset in pages to write
349  * @count:      number of bytes to write
350  *
351  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
352  * as it forces O_SYNC writers to different parts of the same file
353  * to be serialised right until io completion.
354  */
355 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
356                            loff_t pos, loff_t count)
357 {
358         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
359         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
360         int ret;
361
362         if (!mapping_cap_writeback_dirty(mapping) || !count)
363                 return 0;
364         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
365         if (ret == 0)
366                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
367         if (ret == 0)
368                 ret = wait_on_page_writeback_range(mapping, start, end);
369         return ret;
370 }
371 EXPORT_SYMBOL(sync_page_range_nolock);
372
373 /**
374  * filemap_fdatawait - wait for all under-writeback pages to complete
375  * @mapping: address space structure to wait for
376  *
377  * Walk the list of under-writeback pages of the given address space
378  * and wait for all of them.
379  */
380 int filemap_fdatawait(struct address_space *mapping)
381 {
382         loff_t i_size = i_size_read(mapping->host);
383
384         if (i_size == 0)
385                 return 0;
386
387         return wait_on_page_writeback_range(mapping, 0,
388                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
389 }
390 EXPORT_SYMBOL(filemap_fdatawait);
391
392 int filemap_write_and_wait(struct address_space *mapping)
393 {
394         int err = 0;
395
396         if (mapping->nrpages) {
397                 err = filemap_fdatawrite(mapping);
398                 /*
399                  * Even if the above returned error, the pages may be
400                  * written partially (e.g. -ENOSPC), so we wait for it.
401                  * But the -EIO is special case, it may indicate the worst
402                  * thing (e.g. bug) happened, so we avoid waiting for it.
403                  */
404                 if (err != -EIO) {
405                         int err2 = filemap_fdatawait(mapping);
406                         if (!err)
407                                 err = err2;
408                 }
409         }
410         return err;
411 }
412 EXPORT_SYMBOL(filemap_write_and_wait);
413
414 /**
415  * filemap_write_and_wait_range - write out & wait on a file range
416  * @mapping:    the address_space for the pages
417  * @lstart:     offset in bytes where the range starts
418  * @lend:       offset in bytes where the range ends (inclusive)
419  *
420  * Write out and wait upon file offsets lstart->lend, inclusive.
421  *
422  * Note that `lend' is inclusive (describes the last byte to be written) so
423  * that this function can be used to write to the very end-of-file (end = -1).
424  */
425 int filemap_write_and_wait_range(struct address_space *mapping,
426                                  loff_t lstart, loff_t lend)
427 {
428         int err = 0;
429
430         if (mapping->nrpages) {
431                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
432                                                  WB_SYNC_ALL);
433                 /* See comment of filemap_write_and_wait() */
434                 if (err != -EIO) {
435                         int err2 = wait_on_page_writeback_range(mapping,
436                                                 lstart >> PAGE_CACHE_SHIFT,
437                                                 lend >> PAGE_CACHE_SHIFT);
438                         if (!err)
439                                 err = err2;
440                 }
441         }
442         return err;
443 }
444
445 /**
446  * add_to_page_cache_locked - add a locked page to the pagecache
447  * @page:       page to add
448  * @mapping:    the page's address_space
449  * @offset:     page index
450  * @gfp_mask:   page allocation mode
451  *
452  * This function is used to add a page to the pagecache. It must be locked.
453  * This function does not add the page to the LRU.  The caller must do that.
454  */
455 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
456                 pgoff_t offset, gfp_t gfp_mask)
457 {
458         int error;
459
460         VM_BUG_ON(!PageLocked(page));
461
462         error = mem_cgroup_cache_charge(page, current->mm,
463                                         gfp_mask & ~__GFP_HIGHMEM);
464         if (error)
465                 goto out;
466
467         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
468         if (error == 0) {
469                 page_cache_get(page);
470                 page->mapping = mapping;
471                 page->index = offset;
472
473                 spin_lock_irq(&mapping->tree_lock);
474                 error = radix_tree_insert(&mapping->page_tree, offset, page);
475                 if (likely(!error)) {
476                         mapping->nrpages++;
477                         __inc_zone_page_state(page, NR_FILE_PAGES);
478                 } else {
479                         page->mapping = NULL;
480                         mem_cgroup_uncharge_cache_page(page);
481                         page_cache_release(page);
482                 }
483
484                 spin_unlock_irq(&mapping->tree_lock);
485                 radix_tree_preload_end();
486         } else
487                 mem_cgroup_uncharge_cache_page(page);
488 out:
489         return error;
490 }
491 EXPORT_SYMBOL(add_to_page_cache_locked);
492
493 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
494                                 pgoff_t offset, gfp_t gfp_mask)
495 {
496         int ret;
497
498         /*
499          * Splice_read and readahead add shmem/tmpfs pages into the page cache
500          * before shmem_readpage has a chance to mark them as SwapBacked: they
501          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
502          * (called in add_to_page_cache) needs to know where they're going too.
503          */
504         if (mapping_cap_swap_backed(mapping))
505                 SetPageSwapBacked(page);
506
507         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
508         if (ret == 0) {
509                 if (page_is_file_cache(page))
510                         lru_cache_add_file(page);
511                 else
512                         lru_cache_add_active_anon(page);
513         }
514         return ret;
515 }
516
517 #ifdef CONFIG_NUMA
518 struct page *__page_cache_alloc(gfp_t gfp)
519 {
520         if (cpuset_do_page_mem_spread()) {
521                 int n = cpuset_mem_spread_node();
522                 return alloc_pages_node(n, gfp, 0);
523         }
524         return alloc_pages(gfp, 0);
525 }
526 EXPORT_SYMBOL(__page_cache_alloc);
527 #endif
528
529 static int __sleep_on_page_lock(void *word)
530 {
531         io_schedule();
532         return 0;
533 }
534
535 /*
536  * In order to wait for pages to become available there must be
537  * waitqueues associated with pages. By using a hash table of
538  * waitqueues where the bucket discipline is to maintain all
539  * waiters on the same queue and wake all when any of the pages
540  * become available, and for the woken contexts to check to be
541  * sure the appropriate page became available, this saves space
542  * at a cost of "thundering herd" phenomena during rare hash
543  * collisions.
544  */
545 static wait_queue_head_t *page_waitqueue(struct page *page)
546 {
547         const struct zone *zone = page_zone(page);
548
549         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
550 }
551
552 static inline void wake_up_page(struct page *page, int bit)
553 {
554         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
555 }
556
557 void wait_on_page_bit(struct page *page, int bit_nr)
558 {
559         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
560
561         if (test_bit(bit_nr, &page->flags))
562                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
563                                                         TASK_UNINTERRUPTIBLE);
564 }
565 EXPORT_SYMBOL(wait_on_page_bit);
566
567 /**
568  * unlock_page - unlock a locked page
569  * @page: the page
570  *
571  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
572  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
573  * mechananism between PageLocked pages and PageWriteback pages is shared.
574  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
575  *
576  * The first mb is necessary to safely close the critical section opened by the
577  * test_and_set_bit() to lock the page; the second mb is necessary to enforce
578  * ordering between the clear_bit and the read of the waitqueue (to avoid SMP
579  * races with a parallel wait_on_page_locked()).
580  */
581 void unlock_page(struct page *page)
582 {
583         smp_mb__before_clear_bit();
584         if (!test_and_clear_bit(PG_locked, &page->flags))
585                 BUG();
586         smp_mb__after_clear_bit(); 
587         wake_up_page(page, PG_locked);
588 }
589 EXPORT_SYMBOL(unlock_page);
590
591 /**
592  * end_page_writeback - end writeback against a page
593  * @page: the page
594  */
595 void end_page_writeback(struct page *page)
596 {
597         if (TestClearPageReclaim(page))
598                 rotate_reclaimable_page(page);
599
600         if (!test_clear_page_writeback(page))
601                 BUG();
602
603         smp_mb__after_clear_bit();
604         wake_up_page(page, PG_writeback);
605 }
606 EXPORT_SYMBOL(end_page_writeback);
607
608 /**
609  * __lock_page - get a lock on the page, assuming we need to sleep to get it
610  * @page: the page to lock
611  *
612  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
613  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
614  * chances are that on the second loop, the block layer's plug list is empty,
615  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
616  */
617 void __lock_page(struct page *page)
618 {
619         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
620
621         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
622                                                         TASK_UNINTERRUPTIBLE);
623 }
624 EXPORT_SYMBOL(__lock_page);
625
626 int __lock_page_killable(struct page *page)
627 {
628         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
629
630         return __wait_on_bit_lock(page_waitqueue(page), &wait,
631                                         sync_page_killable, TASK_KILLABLE);
632 }
633
634 /**
635  * __lock_page_nosync - get a lock on the page, without calling sync_page()
636  * @page: the page to lock
637  *
638  * Variant of lock_page that does not require the caller to hold a reference
639  * on the page's mapping.
640  */
641 void __lock_page_nosync(struct page *page)
642 {
643         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
644         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
645                                                         TASK_UNINTERRUPTIBLE);
646 }
647
648 /**
649  * find_get_page - find and get a page reference
650  * @mapping: the address_space to search
651  * @offset: the page index
652  *
653  * Is there a pagecache struct page at the given (mapping, offset) tuple?
654  * If yes, increment its refcount and return it; if no, return NULL.
655  */
656 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
657 {
658         void **pagep;
659         struct page *page;
660
661         rcu_read_lock();
662 repeat:
663         page = NULL;
664         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
665         if (pagep) {
666                 page = radix_tree_deref_slot(pagep);
667                 if (unlikely(!page || page == RADIX_TREE_RETRY))
668                         goto repeat;
669
670                 if (!page_cache_get_speculative(page))
671                         goto repeat;
672
673                 /*
674                  * Has the page moved?
675                  * This is part of the lockless pagecache protocol. See
676                  * include/linux/pagemap.h for details.
677                  */
678                 if (unlikely(page != *pagep)) {
679                         page_cache_release(page);
680                         goto repeat;
681                 }
682         }
683         rcu_read_unlock();
684
685         return page;
686 }
687 EXPORT_SYMBOL(find_get_page);
688
689 /**
690  * find_lock_page - locate, pin and lock a pagecache page
691  * @mapping: the address_space to search
692  * @offset: the page index
693  *
694  * Locates the desired pagecache page, locks it, increments its reference
695  * count and returns its address.
696  *
697  * Returns zero if the page was not present. find_lock_page() may sleep.
698  */
699 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
700 {
701         struct page *page;
702
703 repeat:
704         page = find_get_page(mapping, offset);
705         if (page) {
706                 lock_page(page);
707                 /* Has the page been truncated? */
708                 if (unlikely(page->mapping != mapping)) {
709                         unlock_page(page);
710                         page_cache_release(page);
711                         goto repeat;
712                 }
713                 VM_BUG_ON(page->index != offset);
714         }
715         return page;
716 }
717 EXPORT_SYMBOL(find_lock_page);
718
719 /**
720  * find_or_create_page - locate or add a pagecache page
721  * @mapping: the page's address_space
722  * @index: the page's index into the mapping
723  * @gfp_mask: page allocation mode
724  *
725  * Locates a page in the pagecache.  If the page is not present, a new page
726  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
727  * LRU list.  The returned page is locked and has its reference count
728  * incremented.
729  *
730  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
731  * allocation!
732  *
733  * find_or_create_page() returns the desired page's address, or zero on
734  * memory exhaustion.
735  */
736 struct page *find_or_create_page(struct address_space *mapping,
737                 pgoff_t index, gfp_t gfp_mask)
738 {
739         struct page *page;
740         int err;
741 repeat:
742         page = find_lock_page(mapping, index);
743         if (!page) {
744                 page = __page_cache_alloc(gfp_mask);
745                 if (!page)
746                         return NULL;
747                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
748                 if (unlikely(err)) {
749                         page_cache_release(page);
750                         page = NULL;
751                         if (err == -EEXIST)
752                                 goto repeat;
753                 }
754         }
755         return page;
756 }
757 EXPORT_SYMBOL(find_or_create_page);
758
759 /**
760  * find_get_pages - gang pagecache lookup
761  * @mapping:    The address_space to search
762  * @start:      The starting page index
763  * @nr_pages:   The maximum number of pages
764  * @pages:      Where the resulting pages are placed
765  *
766  * find_get_pages() will search for and return a group of up to
767  * @nr_pages pages in the mapping.  The pages are placed at @pages.
768  * find_get_pages() takes a reference against the returned pages.
769  *
770  * The search returns a group of mapping-contiguous pages with ascending
771  * indexes.  There may be holes in the indices due to not-present pages.
772  *
773  * find_get_pages() returns the number of pages which were found.
774  */
775 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
776                             unsigned int nr_pages, struct page **pages)
777 {
778         unsigned int i;
779         unsigned int ret;
780         unsigned int nr_found;
781
782         rcu_read_lock();
783 restart:
784         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
785                                 (void ***)pages, start, nr_pages);
786         ret = 0;
787         for (i = 0; i < nr_found; i++) {
788                 struct page *page;
789 repeat:
790                 page = radix_tree_deref_slot((void **)pages[i]);
791                 if (unlikely(!page))
792                         continue;
793                 /*
794                  * this can only trigger if nr_found == 1, making livelock
795                  * a non issue.
796                  */
797                 if (unlikely(page == RADIX_TREE_RETRY))
798                         goto restart;
799
800                 if (!page_cache_get_speculative(page))
801                         goto repeat;
802
803                 /* Has the page moved? */
804                 if (unlikely(page != *((void **)pages[i]))) {
805                         page_cache_release(page);
806                         goto repeat;
807                 }
808
809                 pages[ret] = page;
810                 ret++;
811         }
812         rcu_read_unlock();
813         return ret;
814 }
815
816 /**
817  * find_get_pages_contig - gang contiguous pagecache lookup
818  * @mapping:    The address_space to search
819  * @index:      The starting page index
820  * @nr_pages:   The maximum number of pages
821  * @pages:      Where the resulting pages are placed
822  *
823  * find_get_pages_contig() works exactly like find_get_pages(), except
824  * that the returned number of pages are guaranteed to be contiguous.
825  *
826  * find_get_pages_contig() returns the number of pages which were found.
827  */
828 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
829                                unsigned int nr_pages, struct page **pages)
830 {
831         unsigned int i;
832         unsigned int ret;
833         unsigned int nr_found;
834
835         rcu_read_lock();
836 restart:
837         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
838                                 (void ***)pages, index, nr_pages);
839         ret = 0;
840         for (i = 0; i < nr_found; i++) {
841                 struct page *page;
842 repeat:
843                 page = radix_tree_deref_slot((void **)pages[i]);
844                 if (unlikely(!page))
845                         continue;
846                 /*
847                  * this can only trigger if nr_found == 1, making livelock
848                  * a non issue.
849                  */
850                 if (unlikely(page == RADIX_TREE_RETRY))
851                         goto restart;
852
853                 if (page->mapping == NULL || page->index != index)
854                         break;
855
856                 if (!page_cache_get_speculative(page))
857                         goto repeat;
858
859                 /* Has the page moved? */
860                 if (unlikely(page != *((void **)pages[i]))) {
861                         page_cache_release(page);
862                         goto repeat;
863                 }
864
865                 pages[ret] = page;
866                 ret++;
867                 index++;
868         }
869         rcu_read_unlock();
870         return ret;
871 }
872 EXPORT_SYMBOL(find_get_pages_contig);
873
874 /**
875  * find_get_pages_tag - find and return pages that match @tag
876  * @mapping:    the address_space to search
877  * @index:      the starting page index
878  * @tag:        the tag index
879  * @nr_pages:   the maximum number of pages
880  * @pages:      where the resulting pages are placed
881  *
882  * Like find_get_pages, except we only return pages which are tagged with
883  * @tag.   We update @index to index the next page for the traversal.
884  */
885 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
886                         int tag, unsigned int nr_pages, struct page **pages)
887 {
888         unsigned int i;
889         unsigned int ret;
890         unsigned int nr_found;
891
892         rcu_read_lock();
893 restart:
894         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
895                                 (void ***)pages, *index, nr_pages, tag);
896         ret = 0;
897         for (i = 0; i < nr_found; i++) {
898                 struct page *page;
899 repeat:
900                 page = radix_tree_deref_slot((void **)pages[i]);
901                 if (unlikely(!page))
902                         continue;
903                 /*
904                  * this can only trigger if nr_found == 1, making livelock
905                  * a non issue.
906                  */
907                 if (unlikely(page == RADIX_TREE_RETRY))
908                         goto restart;
909
910                 if (!page_cache_get_speculative(page))
911                         goto repeat;
912
913                 /* Has the page moved? */
914                 if (unlikely(page != *((void **)pages[i]))) {
915                         page_cache_release(page);
916                         goto repeat;
917                 }
918
919                 pages[ret] = page;
920                 ret++;
921         }
922         rcu_read_unlock();
923
924         if (ret)
925                 *index = pages[ret - 1]->index + 1;
926
927         return ret;
928 }
929 EXPORT_SYMBOL(find_get_pages_tag);
930
931 /**
932  * grab_cache_page_nowait - returns locked page at given index in given cache
933  * @mapping: target address_space
934  * @index: the page index
935  *
936  * Same as grab_cache_page(), but do not wait if the page is unavailable.
937  * This is intended for speculative data generators, where the data can
938  * be regenerated if the page couldn't be grabbed.  This routine should
939  * be safe to call while holding the lock for another page.
940  *
941  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
942  * and deadlock against the caller's locked page.
943  */
944 struct page *
945 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
946 {
947         struct page *page = find_get_page(mapping, index);
948
949         if (page) {
950                 if (trylock_page(page))
951                         return page;
952                 page_cache_release(page);
953                 return NULL;
954         }
955         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
956         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
957                 page_cache_release(page);
958                 page = NULL;
959         }
960         return page;
961 }
962 EXPORT_SYMBOL(grab_cache_page_nowait);
963
964 /*
965  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
966  * a _large_ part of the i/o request. Imagine the worst scenario:
967  *
968  *      ---R__________________________________________B__________
969  *         ^ reading here                             ^ bad block(assume 4k)
970  *
971  * read(R) => miss => readahead(R...B) => media error => frustrating retries
972  * => failing the whole request => read(R) => read(R+1) =>
973  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
974  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
975  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
976  *
977  * It is going insane. Fix it by quickly scaling down the readahead size.
978  */
979 static void shrink_readahead_size_eio(struct file *filp,
980                                         struct file_ra_state *ra)
981 {
982         if (!ra->ra_pages)
983                 return;
984
985         ra->ra_pages /= 4;
986 }
987
988 /**
989  * do_generic_file_read - generic file read routine
990  * @filp:       the file to read
991  * @ppos:       current file position
992  * @desc:       read_descriptor
993  * @actor:      read method
994  *
995  * This is a generic file read routine, and uses the
996  * mapping->a_ops->readpage() function for the actual low-level stuff.
997  *
998  * This is really ugly. But the goto's actually try to clarify some
999  * of the logic when it comes to error handling etc.
1000  */
1001 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1002                 read_descriptor_t *desc, read_actor_t actor)
1003 {
1004         struct address_space *mapping = filp->f_mapping;
1005         struct inode *inode = mapping->host;
1006         struct file_ra_state *ra = &filp->f_ra;
1007         pgoff_t index;
1008         pgoff_t last_index;
1009         pgoff_t prev_index;
1010         unsigned long offset;      /* offset into pagecache page */
1011         unsigned int prev_offset;
1012         int error;
1013
1014         index = *ppos >> PAGE_CACHE_SHIFT;
1015         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1016         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1017         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1018         offset = *ppos & ~PAGE_CACHE_MASK;
1019
1020         for (;;) {
1021                 struct page *page;
1022                 pgoff_t end_index;
1023                 loff_t isize;
1024                 unsigned long nr, ret;
1025
1026                 cond_resched();
1027 find_page:
1028                 page = find_get_page(mapping, index);
1029                 if (!page) {
1030                         page_cache_sync_readahead(mapping,
1031                                         ra, filp,
1032                                         index, last_index - index);
1033                         page = find_get_page(mapping, index);
1034                         if (unlikely(page == NULL))
1035                                 goto no_cached_page;
1036                 }
1037                 if (PageReadahead(page)) {
1038                         page_cache_async_readahead(mapping,
1039                                         ra, filp, page,
1040                                         index, last_index - index);
1041                 }
1042                 if (!PageUptodate(page)) {
1043                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1044                                         !mapping->a_ops->is_partially_uptodate)
1045                                 goto page_not_up_to_date;
1046                         if (!trylock_page(page))
1047                                 goto page_not_up_to_date;
1048                         if (!mapping->a_ops->is_partially_uptodate(page,
1049                                                                 desc, offset))
1050                                 goto page_not_up_to_date_locked;
1051                         unlock_page(page);
1052                 }
1053 page_ok:
1054                 /*
1055                  * i_size must be checked after we know the page is Uptodate.
1056                  *
1057                  * Checking i_size after the check allows us to calculate
1058                  * the correct value for "nr", which means the zero-filled
1059                  * part of the page is not copied back to userspace (unless
1060                  * another truncate extends the file - this is desired though).
1061                  */
1062
1063                 isize = i_size_read(inode);
1064                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1065                 if (unlikely(!isize || index > end_index)) {
1066                         page_cache_release(page);
1067                         goto out;
1068                 }
1069
1070                 /* nr is the maximum number of bytes to copy from this page */
1071                 nr = PAGE_CACHE_SIZE;
1072                 if (index == end_index) {
1073                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1074                         if (nr <= offset) {
1075                                 page_cache_release(page);
1076                                 goto out;
1077                         }
1078                 }
1079                 nr = nr - offset;
1080
1081                 /* If users can be writing to this page using arbitrary
1082                  * virtual addresses, take care about potential aliasing
1083                  * before reading the page on the kernel side.
1084                  */
1085                 if (mapping_writably_mapped(mapping))
1086                         flush_dcache_page(page);
1087
1088                 /*
1089                  * When a sequential read accesses a page several times,
1090                  * only mark it as accessed the first time.
1091                  */
1092                 if (prev_index != index || offset != prev_offset)
1093                         mark_page_accessed(page);
1094                 prev_index = index;
1095
1096                 /*
1097                  * Ok, we have the page, and it's up-to-date, so
1098                  * now we can copy it to user space...
1099                  *
1100                  * The actor routine returns how many bytes were actually used..
1101                  * NOTE! This may not be the same as how much of a user buffer
1102                  * we filled up (we may be padding etc), so we can only update
1103                  * "pos" here (the actor routine has to update the user buffer
1104                  * pointers and the remaining count).
1105                  */
1106                 ret = actor(desc, page, offset, nr);
1107                 offset += ret;
1108                 index += offset >> PAGE_CACHE_SHIFT;
1109                 offset &= ~PAGE_CACHE_MASK;
1110                 prev_offset = offset;
1111
1112                 page_cache_release(page);
1113                 if (ret == nr && desc->count)
1114                         continue;
1115                 goto out;
1116
1117 page_not_up_to_date:
1118                 /* Get exclusive access to the page ... */
1119                 error = lock_page_killable(page);
1120                 if (unlikely(error))
1121                         goto readpage_error;
1122
1123 page_not_up_to_date_locked:
1124                 /* Did it get truncated before we got the lock? */
1125                 if (!page->mapping) {
1126                         unlock_page(page);
1127                         page_cache_release(page);
1128                         continue;
1129                 }
1130
1131                 /* Did somebody else fill it already? */
1132                 if (PageUptodate(page)) {
1133                         unlock_page(page);
1134                         goto page_ok;
1135                 }
1136
1137 readpage:
1138                 /* Start the actual read. The read will unlock the page. */
1139                 error = mapping->a_ops->readpage(filp, page);
1140
1141                 if (unlikely(error)) {
1142                         if (error == AOP_TRUNCATED_PAGE) {
1143                                 page_cache_release(page);
1144                                 goto find_page;
1145                         }
1146                         goto readpage_error;
1147                 }
1148
1149                 if (!PageUptodate(page)) {
1150                         error = lock_page_killable(page);
1151                         if (unlikely(error))
1152                                 goto readpage_error;
1153                         if (!PageUptodate(page)) {
1154                                 if (page->mapping == NULL) {
1155                                         /*
1156                                          * invalidate_inode_pages got it
1157                                          */
1158                                         unlock_page(page);
1159                                         page_cache_release(page);
1160                                         goto find_page;
1161                                 }
1162                                 unlock_page(page);
1163                                 shrink_readahead_size_eio(filp, ra);
1164                                 error = -EIO;
1165                                 goto readpage_error;
1166                         }
1167                         unlock_page(page);
1168                 }
1169
1170                 goto page_ok;
1171
1172 readpage_error:
1173                 /* UHHUH! A synchronous read error occurred. Report it */
1174                 desc->error = error;
1175                 page_cache_release(page);
1176                 goto out;
1177
1178 no_cached_page:
1179                 /*
1180                  * Ok, it wasn't cached, so we need to create a new
1181                  * page..
1182                  */
1183                 page = page_cache_alloc_cold(mapping);
1184                 if (!page) {
1185                         desc->error = -ENOMEM;
1186                         goto out;
1187                 }
1188                 error = add_to_page_cache_lru(page, mapping,
1189                                                 index, GFP_KERNEL);
1190                 if (error) {
1191                         page_cache_release(page);
1192                         if (error == -EEXIST)
1193                                 goto find_page;
1194                         desc->error = error;
1195                         goto out;
1196                 }
1197                 goto readpage;
1198         }
1199
1200 out:
1201         ra->prev_pos = prev_index;
1202         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1203         ra->prev_pos |= prev_offset;
1204
1205         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1206         file_accessed(filp);
1207 }
1208
1209 int file_read_actor(read_descriptor_t *desc, struct page *page,
1210                         unsigned long offset, unsigned long size)
1211 {
1212         char *kaddr;
1213         unsigned long left, count = desc->count;
1214
1215         if (size > count)
1216                 size = count;
1217
1218         /*
1219          * Faults on the destination of a read are common, so do it before
1220          * taking the kmap.
1221          */
1222         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1223                 kaddr = kmap_atomic(page, KM_USER0);
1224                 left = __copy_to_user_inatomic(desc->arg.buf,
1225                                                 kaddr + offset, size);
1226                 kunmap_atomic(kaddr, KM_USER0);
1227                 if (left == 0)
1228                         goto success;
1229         }
1230
1231         /* Do it the slow way */
1232         kaddr = kmap(page);
1233         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1234         kunmap(page);
1235
1236         if (left) {
1237                 size -= left;
1238                 desc->error = -EFAULT;
1239         }
1240 success:
1241         desc->count = count - size;
1242         desc->written += size;
1243         desc->arg.buf += size;
1244         return size;
1245 }
1246
1247 /*
1248  * Performs necessary checks before doing a write
1249  * @iov:        io vector request
1250  * @nr_segs:    number of segments in the iovec
1251  * @count:      number of bytes to write
1252  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1253  *
1254  * Adjust number of segments and amount of bytes to write (nr_segs should be
1255  * properly initialized first). Returns appropriate error code that caller
1256  * should return or zero in case that write should be allowed.
1257  */
1258 int generic_segment_checks(const struct iovec *iov,
1259                         unsigned long *nr_segs, size_t *count, int access_flags)
1260 {
1261         unsigned long   seg;
1262         size_t cnt = 0;
1263         for (seg = 0; seg < *nr_segs; seg++) {
1264                 const struct iovec *iv = &iov[seg];
1265
1266                 /*
1267                  * If any segment has a negative length, or the cumulative
1268                  * length ever wraps negative then return -EINVAL.
1269                  */
1270                 cnt += iv->iov_len;
1271                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1272                         return -EINVAL;
1273                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1274                         continue;
1275                 if (seg == 0)
1276                         return -EFAULT;
1277                 *nr_segs = seg;
1278                 cnt -= iv->iov_len;     /* This segment is no good */
1279                 break;
1280         }
1281         *count = cnt;
1282         return 0;
1283 }
1284 EXPORT_SYMBOL(generic_segment_checks);
1285
1286 /**
1287  * generic_file_aio_read - generic filesystem read routine
1288  * @iocb:       kernel I/O control block
1289  * @iov:        io vector request
1290  * @nr_segs:    number of segments in the iovec
1291  * @pos:        current file position
1292  *
1293  * This is the "read()" routine for all filesystems
1294  * that can use the page cache directly.
1295  */
1296 ssize_t
1297 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1298                 unsigned long nr_segs, loff_t pos)
1299 {
1300         struct file *filp = iocb->ki_filp;
1301         ssize_t retval;
1302         unsigned long seg;
1303         size_t count;
1304         loff_t *ppos = &iocb->ki_pos;
1305
1306         count = 0;
1307         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1308         if (retval)
1309                 return retval;
1310
1311         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1312         if (filp->f_flags & O_DIRECT) {
1313                 loff_t size;
1314                 struct address_space *mapping;
1315                 struct inode *inode;
1316
1317                 mapping = filp->f_mapping;
1318                 inode = mapping->host;
1319                 if (!count)
1320                         goto out; /* skip atime */
1321                 size = i_size_read(inode);
1322                 if (pos < size) {
1323                         retval = filemap_write_and_wait(mapping);
1324                         if (!retval) {
1325                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1326                                                         iov, pos, nr_segs);
1327                         }
1328                         if (retval > 0)
1329                                 *ppos = pos + retval;
1330                         if (retval) {
1331                                 file_accessed(filp);
1332                                 goto out;
1333                         }
1334                 }
1335         }
1336
1337         for (seg = 0; seg < nr_segs; seg++) {
1338                 read_descriptor_t desc;
1339
1340                 desc.written = 0;
1341                 desc.arg.buf = iov[seg].iov_base;
1342                 desc.count = iov[seg].iov_len;
1343                 if (desc.count == 0)
1344                         continue;
1345                 desc.error = 0;
1346                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1347                 retval += desc.written;
1348                 if (desc.error) {
1349                         retval = retval ?: desc.error;
1350                         break;
1351                 }
1352                 if (desc.count > 0)
1353                         break;
1354         }
1355 out:
1356         return retval;
1357 }
1358 EXPORT_SYMBOL(generic_file_aio_read);
1359
1360 static ssize_t
1361 do_readahead(struct address_space *mapping, struct file *filp,
1362              pgoff_t index, unsigned long nr)
1363 {
1364         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1365                 return -EINVAL;
1366
1367         force_page_cache_readahead(mapping, filp, index,
1368                                         max_sane_readahead(nr));
1369         return 0;
1370 }
1371
1372 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1373 {
1374         ssize_t ret;
1375         struct file *file;
1376
1377         ret = -EBADF;
1378         file = fget(fd);
1379         if (file) {
1380                 if (file->f_mode & FMODE_READ) {
1381                         struct address_space *mapping = file->f_mapping;
1382                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1383                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1384                         unsigned long len = end - start + 1;
1385                         ret = do_readahead(mapping, file, start, len);
1386                 }
1387                 fput(file);
1388         }
1389         return ret;
1390 }
1391
1392 #ifdef CONFIG_MMU
1393 /**
1394  * page_cache_read - adds requested page to the page cache if not already there
1395  * @file:       file to read
1396  * @offset:     page index
1397  *
1398  * This adds the requested page to the page cache if it isn't already there,
1399  * and schedules an I/O to read in its contents from disk.
1400  */
1401 static int page_cache_read(struct file *file, pgoff_t offset)
1402 {
1403         struct address_space *mapping = file->f_mapping;
1404         struct page *page; 
1405         int ret;
1406
1407         do {
1408                 page = page_cache_alloc_cold(mapping);
1409                 if (!page)
1410                         return -ENOMEM;
1411
1412                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1413                 if (ret == 0)
1414                         ret = mapping->a_ops->readpage(file, page);
1415                 else if (ret == -EEXIST)
1416                         ret = 0; /* losing race to add is OK */
1417
1418                 page_cache_release(page);
1419
1420         } while (ret == AOP_TRUNCATED_PAGE);
1421                 
1422         return ret;
1423 }
1424
1425 #define MMAP_LOTSAMISS  (100)
1426
1427 /**
1428  * filemap_fault - read in file data for page fault handling
1429  * @vma:        vma in which the fault was taken
1430  * @vmf:        struct vm_fault containing details of the fault
1431  *
1432  * filemap_fault() is invoked via the vma operations vector for a
1433  * mapped memory region to read in file data during a page fault.
1434  *
1435  * The goto's are kind of ugly, but this streamlines the normal case of having
1436  * it in the page cache, and handles the special cases reasonably without
1437  * having a lot of duplicated code.
1438  */
1439 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1440 {
1441         int error;
1442         struct file *file = vma->vm_file;
1443         struct address_space *mapping = file->f_mapping;
1444         struct file_ra_state *ra = &file->f_ra;
1445         struct inode *inode = mapping->host;
1446         struct page *page;
1447         pgoff_t size;
1448         int did_readaround = 0;
1449         int ret = 0;
1450
1451         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1452         if (vmf->pgoff >= size)
1453                 return VM_FAULT_SIGBUS;
1454
1455         /* If we don't want any read-ahead, don't bother */
1456         if (VM_RandomReadHint(vma))
1457                 goto no_cached_page;
1458
1459         /*
1460          * Do we have something in the page cache already?
1461          */
1462 retry_find:
1463         page = find_lock_page(mapping, vmf->pgoff);
1464         /*
1465          * For sequential accesses, we use the generic readahead logic.
1466          */
1467         if (VM_SequentialReadHint(vma)) {
1468                 if (!page) {
1469                         page_cache_sync_readahead(mapping, ra, file,
1470                                                            vmf->pgoff, 1);
1471                         page = find_lock_page(mapping, vmf->pgoff);
1472                         if (!page)
1473                                 goto no_cached_page;
1474                 }
1475                 if (PageReadahead(page)) {
1476                         page_cache_async_readahead(mapping, ra, file, page,
1477                                                            vmf->pgoff, 1);
1478                 }
1479         }
1480
1481         if (!page) {
1482                 unsigned long ra_pages;
1483
1484                 ra->mmap_miss++;
1485
1486                 /*
1487                  * Do we miss much more than hit in this file? If so,
1488                  * stop bothering with read-ahead. It will only hurt.
1489                  */
1490                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1491                         goto no_cached_page;
1492
1493                 /*
1494                  * To keep the pgmajfault counter straight, we need to
1495                  * check did_readaround, as this is an inner loop.
1496                  */
1497                 if (!did_readaround) {
1498                         ret = VM_FAULT_MAJOR;
1499                         count_vm_event(PGMAJFAULT);
1500                 }
1501                 did_readaround = 1;
1502                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1503                 if (ra_pages) {
1504                         pgoff_t start = 0;
1505
1506                         if (vmf->pgoff > ra_pages / 2)
1507                                 start = vmf->pgoff - ra_pages / 2;
1508                         do_page_cache_readahead(mapping, file, start, ra_pages);
1509                 }
1510                 page = find_lock_page(mapping, vmf->pgoff);
1511                 if (!page)
1512                         goto no_cached_page;
1513         }
1514
1515         if (!did_readaround)
1516                 ra->mmap_miss--;
1517
1518         /*
1519          * We have a locked page in the page cache, now we need to check
1520          * that it's up-to-date. If not, it is going to be due to an error.
1521          */
1522         if (unlikely(!PageUptodate(page)))
1523                 goto page_not_uptodate;
1524
1525         /* Must recheck i_size under page lock */
1526         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1527         if (unlikely(vmf->pgoff >= size)) {
1528                 unlock_page(page);
1529                 page_cache_release(page);
1530                 return VM_FAULT_SIGBUS;
1531         }
1532
1533         /*
1534          * Found the page and have a reference on it.
1535          */
1536         mark_page_accessed(page);
1537         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1538         vmf->page = page;
1539         return ret | VM_FAULT_LOCKED;
1540
1541 no_cached_page:
1542         /*
1543          * We're only likely to ever get here if MADV_RANDOM is in
1544          * effect.
1545          */
1546         error = page_cache_read(file, vmf->pgoff);
1547
1548         /*
1549          * The page we want has now been added to the page cache.
1550          * In the unlikely event that someone removed it in the
1551          * meantime, we'll just come back here and read it again.
1552          */
1553         if (error >= 0)
1554                 goto retry_find;
1555
1556         /*
1557          * An error return from page_cache_read can result if the
1558          * system is low on memory, or a problem occurs while trying
1559          * to schedule I/O.
1560          */
1561         if (error == -ENOMEM)
1562                 return VM_FAULT_OOM;
1563         return VM_FAULT_SIGBUS;
1564
1565 page_not_uptodate:
1566         /* IO error path */
1567         if (!did_readaround) {
1568                 ret = VM_FAULT_MAJOR;
1569                 count_vm_event(PGMAJFAULT);
1570         }
1571
1572         /*
1573          * Umm, take care of errors if the page isn't up-to-date.
1574          * Try to re-read it _once_. We do this synchronously,
1575          * because there really aren't any performance issues here
1576          * and we need to check for errors.
1577          */
1578         ClearPageError(page);
1579         error = mapping->a_ops->readpage(file, page);
1580         if (!error) {
1581                 wait_on_page_locked(page);
1582                 if (!PageUptodate(page))
1583                         error = -EIO;
1584         }
1585         page_cache_release(page);
1586
1587         if (!error || error == AOP_TRUNCATED_PAGE)
1588                 goto retry_find;
1589
1590         /* Things didn't work out. Return zero to tell the mm layer so. */
1591         shrink_readahead_size_eio(file, ra);
1592         return VM_FAULT_SIGBUS;
1593 }
1594 EXPORT_SYMBOL(filemap_fault);
1595
1596 struct vm_operations_struct generic_file_vm_ops = {
1597         .fault          = filemap_fault,
1598 };
1599
1600 /* This is used for a general mmap of a disk file */
1601
1602 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1603 {
1604         struct address_space *mapping = file->f_mapping;
1605
1606         if (!mapping->a_ops->readpage)
1607                 return -ENOEXEC;
1608         file_accessed(file);
1609         vma->vm_ops = &generic_file_vm_ops;
1610         vma->vm_flags |= VM_CAN_NONLINEAR;
1611         return 0;
1612 }
1613
1614 /*
1615  * This is for filesystems which do not implement ->writepage.
1616  */
1617 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1618 {
1619         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1620                 return -EINVAL;
1621         return generic_file_mmap(file, vma);
1622 }
1623 #else
1624 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1625 {
1626         return -ENOSYS;
1627 }
1628 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1629 {
1630         return -ENOSYS;
1631 }
1632 #endif /* CONFIG_MMU */
1633
1634 EXPORT_SYMBOL(generic_file_mmap);
1635 EXPORT_SYMBOL(generic_file_readonly_mmap);
1636
1637 static struct page *__read_cache_page(struct address_space *mapping,
1638                                 pgoff_t index,
1639                                 int (*filler)(void *,struct page*),
1640                                 void *data)
1641 {
1642         struct page *page;
1643         int err;
1644 repeat:
1645         page = find_get_page(mapping, index);
1646         if (!page) {
1647                 page = page_cache_alloc_cold(mapping);
1648                 if (!page)
1649                         return ERR_PTR(-ENOMEM);
1650                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1651                 if (unlikely(err)) {
1652                         page_cache_release(page);
1653                         if (err == -EEXIST)
1654                                 goto repeat;
1655                         /* Presumably ENOMEM for radix tree node */
1656                         return ERR_PTR(err);
1657                 }
1658                 err = filler(data, page);
1659                 if (err < 0) {
1660                         page_cache_release(page);
1661                         page = ERR_PTR(err);
1662                 }
1663         }
1664         return page;
1665 }
1666
1667 /**
1668  * read_cache_page_async - read into page cache, fill it if needed
1669  * @mapping:    the page's address_space
1670  * @index:      the page index
1671  * @filler:     function to perform the read
1672  * @data:       destination for read data
1673  *
1674  * Same as read_cache_page, but don't wait for page to become unlocked
1675  * after submitting it to the filler.
1676  *
1677  * Read into the page cache. If a page already exists, and PageUptodate() is
1678  * not set, try to fill the page but don't wait for it to become unlocked.
1679  *
1680  * If the page does not get brought uptodate, return -EIO.
1681  */
1682 struct page *read_cache_page_async(struct address_space *mapping,
1683                                 pgoff_t index,
1684                                 int (*filler)(void *,struct page*),
1685                                 void *data)
1686 {
1687         struct page *page;
1688         int err;
1689
1690 retry:
1691         page = __read_cache_page(mapping, index, filler, data);
1692         if (IS_ERR(page))
1693                 return page;
1694         if (PageUptodate(page))
1695                 goto out;
1696
1697         lock_page(page);
1698         if (!page->mapping) {
1699                 unlock_page(page);
1700                 page_cache_release(page);
1701                 goto retry;
1702         }
1703         if (PageUptodate(page)) {
1704                 unlock_page(page);
1705                 goto out;
1706         }
1707         err = filler(data, page);
1708         if (err < 0) {
1709                 page_cache_release(page);
1710                 return ERR_PTR(err);
1711         }
1712 out:
1713         mark_page_accessed(page);
1714         return page;
1715 }
1716 EXPORT_SYMBOL(read_cache_page_async);
1717
1718 /**
1719  * read_cache_page - read into page cache, fill it if needed
1720  * @mapping:    the page's address_space
1721  * @index:      the page index
1722  * @filler:     function to perform the read
1723  * @data:       destination for read data
1724  *
1725  * Read into the page cache. If a page already exists, and PageUptodate() is
1726  * not set, try to fill the page then wait for it to become unlocked.
1727  *
1728  * If the page does not get brought uptodate, return -EIO.
1729  */
1730 struct page *read_cache_page(struct address_space *mapping,
1731                                 pgoff_t index,
1732                                 int (*filler)(void *,struct page*),
1733                                 void *data)
1734 {
1735         struct page *page;
1736
1737         page = read_cache_page_async(mapping, index, filler, data);
1738         if (IS_ERR(page))
1739                 goto out;
1740         wait_on_page_locked(page);
1741         if (!PageUptodate(page)) {
1742                 page_cache_release(page);
1743                 page = ERR_PTR(-EIO);
1744         }
1745  out:
1746         return page;
1747 }
1748 EXPORT_SYMBOL(read_cache_page);
1749
1750 /*
1751  * The logic we want is
1752  *
1753  *      if suid or (sgid and xgrp)
1754  *              remove privs
1755  */
1756 int should_remove_suid(struct dentry *dentry)
1757 {
1758         mode_t mode = dentry->d_inode->i_mode;
1759         int kill = 0;
1760
1761         /* suid always must be killed */
1762         if (unlikely(mode & S_ISUID))
1763                 kill = ATTR_KILL_SUID;
1764
1765         /*
1766          * sgid without any exec bits is just a mandatory locking mark; leave
1767          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1768          */
1769         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1770                 kill |= ATTR_KILL_SGID;
1771
1772         if (unlikely(kill && !capable(CAP_FSETID)))
1773                 return kill;
1774
1775         return 0;
1776 }
1777 EXPORT_SYMBOL(should_remove_suid);
1778
1779 static int __remove_suid(struct dentry *dentry, int kill)
1780 {
1781         struct iattr newattrs;
1782
1783         newattrs.ia_valid = ATTR_FORCE | kill;
1784         return notify_change(dentry, &newattrs);
1785 }
1786
1787 int file_remove_suid(struct file *file)
1788 {
1789         struct dentry *dentry = file->f_path.dentry;
1790         int killsuid = should_remove_suid(dentry);
1791         int killpriv = security_inode_need_killpriv(dentry);
1792         int error = 0;
1793
1794         if (killpriv < 0)
1795                 return killpriv;
1796         if (killpriv)
1797                 error = security_inode_killpriv(dentry);
1798         if (!error && killsuid)
1799                 error = __remove_suid(dentry, killsuid);
1800
1801         return error;
1802 }
1803 EXPORT_SYMBOL(file_remove_suid);
1804
1805 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1806                         const struct iovec *iov, size_t base, size_t bytes)
1807 {
1808         size_t copied = 0, left = 0;
1809
1810         while (bytes) {
1811                 char __user *buf = iov->iov_base + base;
1812                 int copy = min(bytes, iov->iov_len - base);
1813
1814                 base = 0;
1815                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1816                 copied += copy;
1817                 bytes -= copy;
1818                 vaddr += copy;
1819                 iov++;
1820
1821                 if (unlikely(left))
1822                         break;
1823         }
1824         return copied - left;
1825 }
1826
1827 /*
1828  * Copy as much as we can into the page and return the number of bytes which
1829  * were sucessfully copied.  If a fault is encountered then return the number of
1830  * bytes which were copied.
1831  */
1832 size_t iov_iter_copy_from_user_atomic(struct page *page,
1833                 struct iov_iter *i, unsigned long offset, size_t bytes)
1834 {
1835         char *kaddr;
1836         size_t copied;
1837
1838         BUG_ON(!in_atomic());
1839         kaddr = kmap_atomic(page, KM_USER0);
1840         if (likely(i->nr_segs == 1)) {
1841                 int left;
1842                 char __user *buf = i->iov->iov_base + i->iov_offset;
1843                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1844                                                         buf, bytes);
1845                 copied = bytes - left;
1846         } else {
1847                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1848                                                 i->iov, i->iov_offset, bytes);
1849         }
1850         kunmap_atomic(kaddr, KM_USER0);
1851
1852         return copied;
1853 }
1854 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1855
1856 /*
1857  * This has the same sideeffects and return value as
1858  * iov_iter_copy_from_user_atomic().
1859  * The difference is that it attempts to resolve faults.
1860  * Page must not be locked.
1861  */
1862 size_t iov_iter_copy_from_user(struct page *page,
1863                 struct iov_iter *i, unsigned long offset, size_t bytes)
1864 {
1865         char *kaddr;
1866         size_t copied;
1867
1868         kaddr = kmap(page);
1869         if (likely(i->nr_segs == 1)) {
1870                 int left;
1871                 char __user *buf = i->iov->iov_base + i->iov_offset;
1872                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1873                 copied = bytes - left;
1874         } else {
1875                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1876                                                 i->iov, i->iov_offset, bytes);
1877         }
1878         kunmap(page);
1879         return copied;
1880 }
1881 EXPORT_SYMBOL(iov_iter_copy_from_user);
1882
1883 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1884 {
1885         BUG_ON(i->count < bytes);
1886
1887         if (likely(i->nr_segs == 1)) {
1888                 i->iov_offset += bytes;
1889                 i->count -= bytes;
1890         } else {
1891                 const struct iovec *iov = i->iov;
1892                 size_t base = i->iov_offset;
1893
1894                 /*
1895                  * The !iov->iov_len check ensures we skip over unlikely
1896                  * zero-length segments (without overruning the iovec).
1897                  */
1898                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1899                         int copy;
1900
1901                         copy = min(bytes, iov->iov_len - base);
1902                         BUG_ON(!i->count || i->count < copy);
1903                         i->count -= copy;
1904                         bytes -= copy;
1905                         base += copy;
1906                         if (iov->iov_len == base) {
1907                                 iov++;
1908                                 base = 0;
1909                         }
1910                 }
1911                 i->iov = iov;
1912                 i->iov_offset = base;
1913         }
1914 }
1915 EXPORT_SYMBOL(iov_iter_advance);
1916
1917 /*
1918  * Fault in the first iovec of the given iov_iter, to a maximum length
1919  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1920  * accessed (ie. because it is an invalid address).
1921  *
1922  * writev-intensive code may want this to prefault several iovecs -- that
1923  * would be possible (callers must not rely on the fact that _only_ the
1924  * first iovec will be faulted with the current implementation).
1925  */
1926 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1927 {
1928         char __user *buf = i->iov->iov_base + i->iov_offset;
1929         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1930         return fault_in_pages_readable(buf, bytes);
1931 }
1932 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1933
1934 /*
1935  * Return the count of just the current iov_iter segment.
1936  */
1937 size_t iov_iter_single_seg_count(struct iov_iter *i)
1938 {
1939         const struct iovec *iov = i->iov;
1940         if (i->nr_segs == 1)
1941                 return i->count;
1942         else
1943                 return min(i->count, iov->iov_len - i->iov_offset);
1944 }
1945 EXPORT_SYMBOL(iov_iter_single_seg_count);
1946
1947 /*
1948  * Performs necessary checks before doing a write
1949  *
1950  * Can adjust writing position or amount of bytes to write.
1951  * Returns appropriate error code that caller should return or
1952  * zero in case that write should be allowed.
1953  */
1954 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1955 {
1956         struct inode *inode = file->f_mapping->host;
1957         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1958
1959         if (unlikely(*pos < 0))
1960                 return -EINVAL;
1961
1962         if (!isblk) {
1963                 /* FIXME: this is for backwards compatibility with 2.4 */
1964                 if (file->f_flags & O_APPEND)
1965                         *pos = i_size_read(inode);
1966
1967                 if (limit != RLIM_INFINITY) {
1968                         if (*pos >= limit) {
1969                                 send_sig(SIGXFSZ, current, 0);
1970                                 return -EFBIG;
1971                         }
1972                         if (*count > limit - (typeof(limit))*pos) {
1973                                 *count = limit - (typeof(limit))*pos;
1974                         }
1975                 }
1976         }
1977
1978         /*
1979          * LFS rule
1980          */
1981         if (unlikely(*pos + *count > MAX_NON_LFS &&
1982                                 !(file->f_flags & O_LARGEFILE))) {
1983                 if (*pos >= MAX_NON_LFS) {
1984                         return -EFBIG;
1985                 }
1986                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1987                         *count = MAX_NON_LFS - (unsigned long)*pos;
1988                 }
1989         }
1990
1991         /*
1992          * Are we about to exceed the fs block limit ?
1993          *
1994          * If we have written data it becomes a short write.  If we have
1995          * exceeded without writing data we send a signal and return EFBIG.
1996          * Linus frestrict idea will clean these up nicely..
1997          */
1998         if (likely(!isblk)) {
1999                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2000                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2001                                 return -EFBIG;
2002                         }
2003                         /* zero-length writes at ->s_maxbytes are OK */
2004                 }
2005
2006                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2007                         *count = inode->i_sb->s_maxbytes - *pos;
2008         } else {
2009 #ifdef CONFIG_BLOCK
2010                 loff_t isize;
2011                 if (bdev_read_only(I_BDEV(inode)))
2012                         return -EPERM;
2013                 isize = i_size_read(inode);
2014                 if (*pos >= isize) {
2015                         if (*count || *pos > isize)
2016                                 return -ENOSPC;
2017                 }
2018
2019                 if (*pos + *count > isize)
2020                         *count = isize - *pos;
2021 #else
2022                 return -EPERM;
2023 #endif
2024         }
2025         return 0;
2026 }
2027 EXPORT_SYMBOL(generic_write_checks);
2028
2029 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2030                                 loff_t pos, unsigned len, unsigned flags,
2031                                 struct page **pagep, void **fsdata)
2032 {
2033         const struct address_space_operations *aops = mapping->a_ops;
2034
2035         if (aops->write_begin) {
2036                 return aops->write_begin(file, mapping, pos, len, flags,
2037                                                         pagep, fsdata);
2038         } else {
2039                 int ret;
2040                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
2041                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2042                 struct inode *inode = mapping->host;
2043                 struct page *page;
2044 again:
2045                 page = __grab_cache_page(mapping, index);
2046                 *pagep = page;
2047                 if (!page)
2048                         return -ENOMEM;
2049
2050                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
2051                         /*
2052                          * There is no way to resolve a short write situation
2053                          * for a !Uptodate page (except by double copying in
2054                          * the caller done by generic_perform_write_2copy).
2055                          *
2056                          * Instead, we have to bring it uptodate here.
2057                          */
2058                         ret = aops->readpage(file, page);
2059                         page_cache_release(page);
2060                         if (ret) {
2061                                 if (ret == AOP_TRUNCATED_PAGE)
2062                                         goto again;
2063                                 return ret;
2064                         }
2065                         goto again;
2066                 }
2067
2068                 ret = aops->prepare_write(file, page, offset, offset+len);
2069                 if (ret) {
2070                         unlock_page(page);
2071                         page_cache_release(page);
2072                         if (pos + len > inode->i_size)
2073                                 vmtruncate(inode, inode->i_size);
2074                 }
2075                 return ret;
2076         }
2077 }
2078 EXPORT_SYMBOL(pagecache_write_begin);
2079
2080 int pagecache_write_end(struct file *file, struct address_space *mapping,
2081                                 loff_t pos, unsigned len, unsigned copied,
2082                                 struct page *page, void *fsdata)
2083 {
2084         const struct address_space_operations *aops = mapping->a_ops;
2085         int ret;
2086
2087         if (aops->write_end) {
2088                 mark_page_accessed(page);
2089                 ret = aops->write_end(file, mapping, pos, len, copied,
2090                                                         page, fsdata);
2091         } else {
2092                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
2093                 struct inode *inode = mapping->host;
2094
2095                 flush_dcache_page(page);
2096                 ret = aops->commit_write(file, page, offset, offset+len);
2097                 unlock_page(page);
2098                 mark_page_accessed(page);
2099                 page_cache_release(page);
2100
2101                 if (ret < 0) {
2102                         if (pos + len > inode->i_size)
2103                                 vmtruncate(inode, inode->i_size);
2104                 } else if (ret > 0)
2105                         ret = min_t(size_t, copied, ret);
2106                 else
2107                         ret = copied;
2108         }
2109
2110         return ret;
2111 }
2112 EXPORT_SYMBOL(pagecache_write_end);
2113
2114 ssize_t
2115 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2116                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2117                 size_t count, size_t ocount)
2118 {
2119         struct file     *file = iocb->ki_filp;
2120         struct address_space *mapping = file->f_mapping;
2121         struct inode    *inode = mapping->host;
2122         ssize_t         written;
2123         size_t          write_len;
2124         pgoff_t         end;
2125
2126         if (count != ocount)
2127                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2128
2129         /*
2130          * Unmap all mmappings of the file up-front.
2131          *
2132          * This will cause any pte dirty bits to be propagated into the
2133          * pageframes for the subsequent filemap_write_and_wait().
2134          */
2135         write_len = iov_length(iov, *nr_segs);
2136         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2137         if (mapping_mapped(mapping))
2138                 unmap_mapping_range(mapping, pos, write_len, 0);
2139
2140         written = filemap_write_and_wait(mapping);
2141         if (written)
2142                 goto out;
2143
2144         /*
2145          * After a write we want buffered reads to be sure to go to disk to get
2146          * the new data.  We invalidate clean cached page from the region we're
2147          * about to write.  We do this *before* the write so that we can return
2148          * without clobbering -EIOCBQUEUED from ->direct_IO().
2149          */
2150         if (mapping->nrpages) {
2151                 written = invalidate_inode_pages2_range(mapping,
2152                                         pos >> PAGE_CACHE_SHIFT, end);
2153                 /*
2154                  * If a page can not be invalidated, return 0 to fall back
2155                  * to buffered write.
2156                  */
2157                 if (written) {
2158                         if (written == -EBUSY)
2159                                 return 0;
2160                         goto out;
2161                 }
2162         }
2163
2164         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2165
2166         /*
2167          * Finally, try again to invalidate clean pages which might have been
2168          * cached by non-direct readahead, or faulted in by get_user_pages()
2169          * if the source of the write was an mmap'ed region of the file
2170          * we're writing.  Either one is a pretty crazy thing to do,
2171          * so we don't support it 100%.  If this invalidation
2172          * fails, tough, the write still worked...
2173          */
2174         if (mapping->nrpages) {
2175                 invalidate_inode_pages2_range(mapping,
2176                                               pos >> PAGE_CACHE_SHIFT, end);
2177         }
2178
2179         if (written > 0) {
2180                 loff_t end = pos + written;
2181                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2182                         i_size_write(inode,  end);
2183                         mark_inode_dirty(inode);
2184                 }
2185                 *ppos = end;
2186         }
2187
2188         /*
2189          * Sync the fs metadata but not the minor inode changes and
2190          * of course not the data as we did direct DMA for the IO.
2191          * i_mutex is held, which protects generic_osync_inode() from
2192          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2193          */
2194 out:
2195         if ((written >= 0 || written == -EIOCBQUEUED) &&
2196             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2197                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2198                 if (err < 0)
2199                         written = err;
2200         }
2201         return written;
2202 }
2203 EXPORT_SYMBOL(generic_file_direct_write);
2204
2205 /*
2206  * Find or create a page at the given pagecache position. Return the locked
2207  * page. This function is specifically for buffered writes.
2208  */
2209 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2210 {
2211         int status;
2212         struct page *page;
2213 repeat:
2214         page = find_lock_page(mapping, index);
2215         if (likely(page))
2216                 return page;
2217
2218         page = page_cache_alloc(mapping);
2219         if (!page)
2220                 return NULL;
2221         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2222         if (unlikely(status)) {
2223                 page_cache_release(page);
2224                 if (status == -EEXIST)
2225                         goto repeat;
2226                 return NULL;
2227         }
2228         return page;
2229 }
2230 EXPORT_SYMBOL(__grab_cache_page);
2231
2232 static ssize_t generic_perform_write_2copy(struct file *file,
2233                                 struct iov_iter *i, loff_t pos)
2234 {
2235         struct address_space *mapping = file->f_mapping;
2236         const struct address_space_operations *a_ops = mapping->a_ops;
2237         struct inode *inode = mapping->host;
2238         long status = 0;
2239         ssize_t written = 0;
2240
2241         do {
2242                 struct page *src_page;
2243                 struct page *page;
2244                 pgoff_t index;          /* Pagecache index for current page */
2245                 unsigned long offset;   /* Offset into pagecache page */
2246                 unsigned long bytes;    /* Bytes to write to page */
2247                 size_t copied;          /* Bytes copied from user */
2248
2249                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2250                 index = pos >> PAGE_CACHE_SHIFT;
2251                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2252                                                 iov_iter_count(i));
2253
2254                 /*
2255                  * a non-NULL src_page indicates that we're doing the
2256                  * copy via get_user_pages and kmap.
2257                  */
2258                 src_page = NULL;
2259
2260                 /*
2261                  * Bring in the user page that we will copy from _first_.
2262                  * Otherwise there's a nasty deadlock on copying from the
2263                  * same page as we're writing to, without it being marked
2264                  * up-to-date.
2265                  *
2266                  * Not only is this an optimisation, but it is also required
2267                  * to check that the address is actually valid, when atomic
2268                  * usercopies are used, below.
2269                  */
2270                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2271                         status = -EFAULT;
2272                         break;
2273                 }
2274
2275                 page = __grab_cache_page(mapping, index);
2276                 if (!page) {
2277                         status = -ENOMEM;
2278                         break;
2279                 }
2280
2281                 /*
2282                  * non-uptodate pages cannot cope with short copies, and we
2283                  * cannot take a pagefault with the destination page locked.
2284                  * So pin the source page to copy it.
2285                  */
2286                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2287                         unlock_page(page);
2288
2289                         src_page = alloc_page(GFP_KERNEL);
2290                         if (!src_page) {
2291                                 page_cache_release(page);
2292                                 status = -ENOMEM;
2293                                 break;
2294                         }
2295
2296                         /*
2297                          * Cannot get_user_pages with a page locked for the
2298                          * same reason as we can't take a page fault with a
2299                          * page locked (as explained below).
2300                          */
2301                         copied = iov_iter_copy_from_user(src_page, i,
2302                                                                 offset, bytes);
2303                         if (unlikely(copied == 0)) {
2304                                 status = -EFAULT;
2305                                 page_cache_release(page);
2306                                 page_cache_release(src_page);
2307                                 break;
2308                         }
2309                         bytes = copied;
2310
2311                         lock_page(page);
2312                         /*
2313                          * Can't handle the page going uptodate here, because
2314                          * that means we would use non-atomic usercopies, which
2315                          * zero out the tail of the page, which can cause
2316                          * zeroes to become transiently visible. We could just
2317                          * use a non-zeroing copy, but the APIs aren't too
2318                          * consistent.
2319                          */
2320                         if (unlikely(!page->mapping || PageUptodate(page))) {
2321                                 unlock_page(page);
2322                                 page_cache_release(page);
2323                                 page_cache_release(src_page);
2324                                 continue;
2325                         }
2326                 }
2327
2328                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2329                 if (unlikely(status))
2330                         goto fs_write_aop_error;
2331
2332                 if (!src_page) {
2333                         /*
2334                          * Must not enter the pagefault handler here, because
2335                          * we hold the page lock, so we might recursively
2336                          * deadlock on the same lock, or get an ABBA deadlock
2337                          * against a different lock, or against the mmap_sem
2338                          * (which nests outside the page lock).  So increment
2339                          * preempt count, and use _atomic usercopies.
2340                          *
2341                          * The page is uptodate so we are OK to encounter a
2342                          * short copy: if unmodified parts of the page are
2343                          * marked dirty and written out to disk, it doesn't
2344                          * really matter.
2345                          */
2346                         pagefault_disable();
2347                         copied = iov_iter_copy_from_user_atomic(page, i,
2348                                                                 offset, bytes);
2349                         pagefault_enable();
2350                 } else {
2351                         void *src, *dst;
2352                         src = kmap_atomic(src_page, KM_USER0);
2353                         dst = kmap_atomic(page, KM_USER1);
2354                         memcpy(dst + offset, src + offset, bytes);
2355                         kunmap_atomic(dst, KM_USER1);
2356                         kunmap_atomic(src, KM_USER0);
2357                         copied = bytes;
2358                 }
2359                 flush_dcache_page(page);
2360
2361                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2362                 if (unlikely(status < 0))
2363                         goto fs_write_aop_error;
2364                 if (unlikely(status > 0)) /* filesystem did partial write */
2365                         copied = min_t(size_t, copied, status);
2366
2367                 unlock_page(page);
2368                 mark_page_accessed(page);
2369                 page_cache_release(page);
2370                 if (src_page)
2371                         page_cache_release(src_page);
2372
2373                 iov_iter_advance(i, copied);
2374                 pos += copied;
2375                 written += copied;
2376
2377                 balance_dirty_pages_ratelimited(mapping);
2378                 cond_resched();
2379                 continue;
2380
2381 fs_write_aop_error:
2382                 unlock_page(page);
2383                 page_cache_release(page);
2384                 if (src_page)
2385                         page_cache_release(src_page);
2386
2387                 /*
2388                  * prepare_write() may have instantiated a few blocks
2389                  * outside i_size.  Trim these off again. Don't need
2390                  * i_size_read because we hold i_mutex.
2391                  */
2392                 if (pos + bytes > inode->i_size)
2393                         vmtruncate(inode, inode->i_size);
2394                 break;
2395         } while (iov_iter_count(i));
2396
2397         return written ? written : status;
2398 }
2399
2400 static ssize_t generic_perform_write(struct file *file,
2401                                 struct iov_iter *i, loff_t pos)
2402 {
2403         struct address_space *mapping = file->f_mapping;
2404         const struct address_space_operations *a_ops = mapping->a_ops;
2405         long status = 0;
2406         ssize_t written = 0;
2407         unsigned int flags = 0;
2408
2409         /*
2410          * Copies from kernel address space cannot fail (NFSD is a big user).
2411          */
2412         if (segment_eq(get_fs(), KERNEL_DS))
2413                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2414
2415         do {
2416                 struct page *page;
2417                 pgoff_t index;          /* Pagecache index for current page */
2418                 unsigned long offset;   /* Offset into pagecache page */
2419                 unsigned long bytes;    /* Bytes to write to page */
2420                 size_t copied;          /* Bytes copied from user */
2421                 void *fsdata;
2422
2423                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2424                 index = pos >> PAGE_CACHE_SHIFT;
2425                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2426                                                 iov_iter_count(i));
2427
2428 again:
2429
2430                 /*
2431                  * Bring in the user page that we will copy from _first_.
2432                  * Otherwise there's a nasty deadlock on copying from the
2433                  * same page as we're writing to, without it being marked
2434                  * up-to-date.
2435                  *
2436                  * Not only is this an optimisation, but it is also required
2437                  * to check that the address is actually valid, when atomic
2438                  * usercopies are used, below.
2439                  */
2440                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2441                         status = -EFAULT;
2442                         break;
2443                 }
2444
2445                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2446                                                 &page, &fsdata);
2447                 if (unlikely(status))
2448                         break;
2449
2450                 pagefault_disable();
2451                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2452                 pagefault_enable();
2453                 flush_dcache_page(page);
2454
2455                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2456                                                 page, fsdata);
2457                 if (unlikely(status < 0))
2458                         break;
2459                 copied = status;
2460
2461                 cond_resched();
2462
2463                 iov_iter_advance(i, copied);
2464                 if (unlikely(copied == 0)) {
2465                         /*
2466                          * If we were unable to copy any data at all, we must
2467                          * fall back to a single segment length write.
2468                          *
2469                          * If we didn't fallback here, we could livelock
2470                          * because not all segments in the iov can be copied at
2471                          * once without a pagefault.
2472                          */
2473                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2474                                                 iov_iter_single_seg_count(i));
2475                         goto again;
2476                 }
2477                 pos += copied;
2478                 written += copied;
2479
2480                 balance_dirty_pages_ratelimited(mapping);
2481
2482         } while (iov_iter_count(i));
2483
2484         return written ? written : status;
2485 }
2486
2487 ssize_t
2488 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2489                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2490                 size_t count, ssize_t written)
2491 {
2492         struct file *file = iocb->ki_filp;
2493         struct address_space *mapping = file->f_mapping;
2494         const struct address_space_operations *a_ops = mapping->a_ops;
2495         struct inode *inode = mapping->host;
2496         ssize_t status;
2497         struct iov_iter i;
2498
2499         iov_iter_init(&i, iov, nr_segs, count, written);
2500         if (a_ops->write_begin)
2501                 status = generic_perform_write(file, &i, pos);
2502         else
2503                 status = generic_perform_write_2copy(file, &i, pos);
2504
2505         if (likely(status >= 0)) {
2506                 written += status;
2507                 *ppos = pos + status;
2508
2509                 /*
2510                  * For now, when the user asks for O_SYNC, we'll actually give
2511                  * O_DSYNC
2512                  */
2513                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2514                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2515                                 status = generic_osync_inode(inode, mapping,
2516                                                 OSYNC_METADATA|OSYNC_DATA);
2517                 }
2518         }
2519         
2520         /*
2521          * If we get here for O_DIRECT writes then we must have fallen through
2522          * to buffered writes (block instantiation inside i_size).  So we sync
2523          * the file data here, to try to honour O_DIRECT expectations.
2524          */
2525         if (unlikely(file->f_flags & O_DIRECT) && written)
2526                 status = filemap_write_and_wait(mapping);
2527
2528         return written ? written : status;
2529 }
2530 EXPORT_SYMBOL(generic_file_buffered_write);
2531
2532 static ssize_t
2533 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2534                                 unsigned long nr_segs, loff_t *ppos)
2535 {
2536         struct file *file = iocb->ki_filp;
2537         struct address_space * mapping = file->f_mapping;
2538         size_t ocount;          /* original count */
2539         size_t count;           /* after file limit checks */
2540         struct inode    *inode = mapping->host;
2541         loff_t          pos;
2542         ssize_t         written;
2543         ssize_t         err;
2544
2545         ocount = 0;
2546         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2547         if (err)
2548                 return err;
2549
2550         count = ocount;
2551         pos = *ppos;
2552
2553         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2554
2555         /* We can write back this queue in page reclaim */
2556         current->backing_dev_info = mapping->backing_dev_info;
2557         written = 0;
2558
2559         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2560         if (err)
2561                 goto out;
2562
2563         if (count == 0)
2564                 goto out;
2565
2566         err = file_remove_suid(file);
2567         if (err)
2568                 goto out;
2569
2570         file_update_time(file);
2571
2572         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2573         if (unlikely(file->f_flags & O_DIRECT)) {
2574                 loff_t endbyte;
2575                 ssize_t written_buffered;
2576
2577                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2578                                                         ppos, count, ocount);
2579                 if (written < 0 || written == count)
2580                         goto out;
2581                 /*
2582                  * direct-io write to a hole: fall through to buffered I/O
2583                  * for completing the rest of the request.
2584                  */
2585                 pos += written;
2586                 count -= written;
2587                 written_buffered = generic_file_buffered_write(iocb, iov,
2588                                                 nr_segs, pos, ppos, count,
2589                                                 written);
2590                 /*
2591                  * If generic_file_buffered_write() retuned a synchronous error
2592                  * then we want to return the number of bytes which were
2593                  * direct-written, or the error code if that was zero.  Note
2594                  * that this differs from normal direct-io semantics, which
2595                  * will return -EFOO even if some bytes were written.
2596                  */
2597                 if (written_buffered < 0) {
2598                         err = written_buffered;
2599                         goto out;
2600                 }
2601
2602                 /*
2603                  * We need to ensure that the page cache pages are written to
2604                  * disk and invalidated to preserve the expected O_DIRECT
2605                  * semantics.
2606                  */
2607                 endbyte = pos + written_buffered - written - 1;
2608                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2609                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2610                                             SYNC_FILE_RANGE_WRITE|
2611                                             SYNC_FILE_RANGE_WAIT_AFTER);
2612                 if (err == 0) {
2613                         written = written_buffered;
2614                         invalidate_mapping_pages(mapping,
2615                                                  pos >> PAGE_CACHE_SHIFT,
2616                                                  endbyte >> PAGE_CACHE_SHIFT);
2617                 } else {
2618                         /*
2619                          * We don't know how much we wrote, so just return
2620                          * the number of bytes which were direct-written
2621                          */
2622                 }
2623         } else {
2624                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2625                                 pos, ppos, count, written);
2626         }
2627 out:
2628         current->backing_dev_info = NULL;
2629         return written ? written : err;
2630 }
2631
2632 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2633                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2634 {
2635         struct file *file = iocb->ki_filp;
2636         struct address_space *mapping = file->f_mapping;
2637         struct inode *inode = mapping->host;
2638         ssize_t ret;
2639
2640         BUG_ON(iocb->ki_pos != pos);
2641
2642         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2643                         &iocb->ki_pos);
2644
2645         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2646                 ssize_t err;
2647
2648                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2649                 if (err < 0)
2650                         ret = err;
2651         }
2652         return ret;
2653 }
2654 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2655
2656 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2657                 unsigned long nr_segs, loff_t pos)
2658 {
2659         struct file *file = iocb->ki_filp;
2660         struct address_space *mapping = file->f_mapping;
2661         struct inode *inode = mapping->host;
2662         ssize_t ret;
2663
2664         BUG_ON(iocb->ki_pos != pos);
2665
2666         mutex_lock(&inode->i_mutex);
2667         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2668                         &iocb->ki_pos);
2669         mutex_unlock(&inode->i_mutex);
2670
2671         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2672                 ssize_t err;
2673
2674                 err = sync_page_range(inode, mapping, pos, ret);
2675                 if (err < 0)
2676                         ret = err;
2677         }
2678         return ret;
2679 }
2680 EXPORT_SYMBOL(generic_file_aio_write);
2681
2682 /**
2683  * try_to_release_page() - release old fs-specific metadata on a page
2684  *
2685  * @page: the page which the kernel is trying to free
2686  * @gfp_mask: memory allocation flags (and I/O mode)
2687  *
2688  * The address_space is to try to release any data against the page
2689  * (presumably at page->private).  If the release was successful, return `1'.
2690  * Otherwise return zero.
2691  *
2692  * The @gfp_mask argument specifies whether I/O may be performed to release
2693  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2694  *
2695  */
2696 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2697 {
2698         struct address_space * const mapping = page->mapping;
2699
2700         BUG_ON(!PageLocked(page));
2701         if (PageWriteback(page))
2702                 return 0;
2703
2704         if (mapping && mapping->a_ops->releasepage)
2705                 return mapping->a_ops->releasepage(page, gfp_mask);
2706         return try_to_free_buffers(page);
2707 }
2708
2709 EXPORT_SYMBOL(try_to_release_page);