CacheFiles: Permit the page lock state to be monitored
[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         radix_tree_delete(&mapping->page_tree, page->index);
120         page->mapping = NULL;
121         mapping->nrpages--;
122         __dec_zone_page_state(page, NR_FILE_PAGES);
123         BUG_ON(page_mapped(page));
124         mem_cgroup_uncharge_cache_page(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 = LONG_MAX,
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_RECLAIM_MASK);
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  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
569  * @page - Page defining the wait queue of interest
570  * @waiter - Waiter to add to the queue
571  *
572  * Add an arbitrary @waiter to the wait queue for the nominated @page.
573  */
574 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
575 {
576         wait_queue_head_t *q = page_waitqueue(page);
577         unsigned long flags;
578
579         spin_lock_irqsave(&q->lock, flags);
580         __add_wait_queue(q, waiter);
581         spin_unlock_irqrestore(&q->lock, flags);
582 }
583 EXPORT_SYMBOL_GPL(add_page_wait_queue);
584
585 /**
586  * unlock_page - unlock a locked page
587  * @page: the page
588  *
589  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
590  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
591  * mechananism between PageLocked pages and PageWriteback pages is shared.
592  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
593  *
594  * The mb is necessary to enforce ordering between the clear_bit and the read
595  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
596  */
597 void unlock_page(struct page *page)
598 {
599         VM_BUG_ON(!PageLocked(page));
600         clear_bit_unlock(PG_locked, &page->flags);
601         smp_mb__after_clear_bit();
602         wake_up_page(page, PG_locked);
603 }
604 EXPORT_SYMBOL(unlock_page);
605
606 /**
607  * end_page_writeback - end writeback against a page
608  * @page: the page
609  */
610 void end_page_writeback(struct page *page)
611 {
612         if (TestClearPageReclaim(page))
613                 rotate_reclaimable_page(page);
614
615         if (!test_clear_page_writeback(page))
616                 BUG();
617
618         smp_mb__after_clear_bit();
619         wake_up_page(page, PG_writeback);
620 }
621 EXPORT_SYMBOL(end_page_writeback);
622
623 /**
624  * __lock_page - get a lock on the page, assuming we need to sleep to get it
625  * @page: the page to lock
626  *
627  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
628  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
629  * chances are that on the second loop, the block layer's plug list is empty,
630  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
631  */
632 void __lock_page(struct page *page)
633 {
634         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
635
636         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
637                                                         TASK_UNINTERRUPTIBLE);
638 }
639 EXPORT_SYMBOL(__lock_page);
640
641 int __lock_page_killable(struct page *page)
642 {
643         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
644
645         return __wait_on_bit_lock(page_waitqueue(page), &wait,
646                                         sync_page_killable, TASK_KILLABLE);
647 }
648
649 /**
650  * __lock_page_nosync - get a lock on the page, without calling sync_page()
651  * @page: the page to lock
652  *
653  * Variant of lock_page that does not require the caller to hold a reference
654  * on the page's mapping.
655  */
656 void __lock_page_nosync(struct page *page)
657 {
658         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
659         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
660                                                         TASK_UNINTERRUPTIBLE);
661 }
662
663 /**
664  * find_get_page - find and get a page reference
665  * @mapping: the address_space to search
666  * @offset: the page index
667  *
668  * Is there a pagecache struct page at the given (mapping, offset) tuple?
669  * If yes, increment its refcount and return it; if no, return NULL.
670  */
671 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
672 {
673         void **pagep;
674         struct page *page;
675
676         rcu_read_lock();
677 repeat:
678         page = NULL;
679         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
680         if (pagep) {
681                 page = radix_tree_deref_slot(pagep);
682                 if (unlikely(!page || page == RADIX_TREE_RETRY))
683                         goto repeat;
684
685                 if (!page_cache_get_speculative(page))
686                         goto repeat;
687
688                 /*
689                  * Has the page moved?
690                  * This is part of the lockless pagecache protocol. See
691                  * include/linux/pagemap.h for details.
692                  */
693                 if (unlikely(page != *pagep)) {
694                         page_cache_release(page);
695                         goto repeat;
696                 }
697         }
698         rcu_read_unlock();
699
700         return page;
701 }
702 EXPORT_SYMBOL(find_get_page);
703
704 /**
705  * find_lock_page - locate, pin and lock a pagecache page
706  * @mapping: the address_space to search
707  * @offset: the page index
708  *
709  * Locates the desired pagecache page, locks it, increments its reference
710  * count and returns its address.
711  *
712  * Returns zero if the page was not present. find_lock_page() may sleep.
713  */
714 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
715 {
716         struct page *page;
717
718 repeat:
719         page = find_get_page(mapping, offset);
720         if (page) {
721                 lock_page(page);
722                 /* Has the page been truncated? */
723                 if (unlikely(page->mapping != mapping)) {
724                         unlock_page(page);
725                         page_cache_release(page);
726                         goto repeat;
727                 }
728                 VM_BUG_ON(page->index != offset);
729         }
730         return page;
731 }
732 EXPORT_SYMBOL(find_lock_page);
733
734 /**
735  * find_or_create_page - locate or add a pagecache page
736  * @mapping: the page's address_space
737  * @index: the page's index into the mapping
738  * @gfp_mask: page allocation mode
739  *
740  * Locates a page in the pagecache.  If the page is not present, a new page
741  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
742  * LRU list.  The returned page is locked and has its reference count
743  * incremented.
744  *
745  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
746  * allocation!
747  *
748  * find_or_create_page() returns the desired page's address, or zero on
749  * memory exhaustion.
750  */
751 struct page *find_or_create_page(struct address_space *mapping,
752                 pgoff_t index, gfp_t gfp_mask)
753 {
754         struct page *page;
755         int err;
756 repeat:
757         page = find_lock_page(mapping, index);
758         if (!page) {
759                 page = __page_cache_alloc(gfp_mask);
760                 if (!page)
761                         return NULL;
762                 /*
763                  * We want a regular kernel memory (not highmem or DMA etc)
764                  * allocation for the radix tree nodes, but we need to honour
765                  * the context-specific requirements the caller has asked for.
766                  * GFP_RECLAIM_MASK collects those requirements.
767                  */
768                 err = add_to_page_cache_lru(page, mapping, index,
769                         (gfp_mask & GFP_RECLAIM_MASK));
770                 if (unlikely(err)) {
771                         page_cache_release(page);
772                         page = NULL;
773                         if (err == -EEXIST)
774                                 goto repeat;
775                 }
776         }
777         return page;
778 }
779 EXPORT_SYMBOL(find_or_create_page);
780
781 /**
782  * find_get_pages - gang pagecache lookup
783  * @mapping:    The address_space to search
784  * @start:      The starting page index
785  * @nr_pages:   The maximum number of pages
786  * @pages:      Where the resulting pages are placed
787  *
788  * find_get_pages() will search for and return a group of up to
789  * @nr_pages pages in the mapping.  The pages are placed at @pages.
790  * find_get_pages() takes a reference against the returned pages.
791  *
792  * The search returns a group of mapping-contiguous pages with ascending
793  * indexes.  There may be holes in the indices due to not-present pages.
794  *
795  * find_get_pages() returns the number of pages which were found.
796  */
797 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
798                             unsigned int nr_pages, struct page **pages)
799 {
800         unsigned int i;
801         unsigned int ret;
802         unsigned int nr_found;
803
804         rcu_read_lock();
805 restart:
806         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
807                                 (void ***)pages, start, nr_pages);
808         ret = 0;
809         for (i = 0; i < nr_found; i++) {
810                 struct page *page;
811 repeat:
812                 page = radix_tree_deref_slot((void **)pages[i]);
813                 if (unlikely(!page))
814                         continue;
815                 /*
816                  * this can only trigger if nr_found == 1, making livelock
817                  * a non issue.
818                  */
819                 if (unlikely(page == RADIX_TREE_RETRY))
820                         goto restart;
821
822                 if (!page_cache_get_speculative(page))
823                         goto repeat;
824
825                 /* Has the page moved? */
826                 if (unlikely(page != *((void **)pages[i]))) {
827                         page_cache_release(page);
828                         goto repeat;
829                 }
830
831                 pages[ret] = page;
832                 ret++;
833         }
834         rcu_read_unlock();
835         return ret;
836 }
837
838 /**
839  * find_get_pages_contig - gang contiguous pagecache lookup
840  * @mapping:    The address_space to search
841  * @index:      The starting page index
842  * @nr_pages:   The maximum number of pages
843  * @pages:      Where the resulting pages are placed
844  *
845  * find_get_pages_contig() works exactly like find_get_pages(), except
846  * that the returned number of pages are guaranteed to be contiguous.
847  *
848  * find_get_pages_contig() returns the number of pages which were found.
849  */
850 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
851                                unsigned int nr_pages, struct page **pages)
852 {
853         unsigned int i;
854         unsigned int ret;
855         unsigned int nr_found;
856
857         rcu_read_lock();
858 restart:
859         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
860                                 (void ***)pages, index, nr_pages);
861         ret = 0;
862         for (i = 0; i < nr_found; i++) {
863                 struct page *page;
864 repeat:
865                 page = radix_tree_deref_slot((void **)pages[i]);
866                 if (unlikely(!page))
867                         continue;
868                 /*
869                  * this can only trigger if nr_found == 1, making livelock
870                  * a non issue.
871                  */
872                 if (unlikely(page == RADIX_TREE_RETRY))
873                         goto restart;
874
875                 if (page->mapping == NULL || page->index != index)
876                         break;
877
878                 if (!page_cache_get_speculative(page))
879                         goto repeat;
880
881                 /* Has the page moved? */
882                 if (unlikely(page != *((void **)pages[i]))) {
883                         page_cache_release(page);
884                         goto repeat;
885                 }
886
887                 pages[ret] = page;
888                 ret++;
889                 index++;
890         }
891         rcu_read_unlock();
892         return ret;
893 }
894 EXPORT_SYMBOL(find_get_pages_contig);
895
896 /**
897  * find_get_pages_tag - find and return pages that match @tag
898  * @mapping:    the address_space to search
899  * @index:      the starting page index
900  * @tag:        the tag index
901  * @nr_pages:   the maximum number of pages
902  * @pages:      where the resulting pages are placed
903  *
904  * Like find_get_pages, except we only return pages which are tagged with
905  * @tag.   We update @index to index the next page for the traversal.
906  */
907 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
908                         int tag, unsigned int nr_pages, struct page **pages)
909 {
910         unsigned int i;
911         unsigned int ret;
912         unsigned int nr_found;
913
914         rcu_read_lock();
915 restart:
916         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
917                                 (void ***)pages, *index, nr_pages, tag);
918         ret = 0;
919         for (i = 0; i < nr_found; i++) {
920                 struct page *page;
921 repeat:
922                 page = radix_tree_deref_slot((void **)pages[i]);
923                 if (unlikely(!page))
924                         continue;
925                 /*
926                  * this can only trigger if nr_found == 1, making livelock
927                  * a non issue.
928                  */
929                 if (unlikely(page == RADIX_TREE_RETRY))
930                         goto restart;
931
932                 if (!page_cache_get_speculative(page))
933                         goto repeat;
934
935                 /* Has the page moved? */
936                 if (unlikely(page != *((void **)pages[i]))) {
937                         page_cache_release(page);
938                         goto repeat;
939                 }
940
941                 pages[ret] = page;
942                 ret++;
943         }
944         rcu_read_unlock();
945
946         if (ret)
947                 *index = pages[ret - 1]->index + 1;
948
949         return ret;
950 }
951 EXPORT_SYMBOL(find_get_pages_tag);
952
953 /**
954  * grab_cache_page_nowait - returns locked page at given index in given cache
955  * @mapping: target address_space
956  * @index: the page index
957  *
958  * Same as grab_cache_page(), but do not wait if the page is unavailable.
959  * This is intended for speculative data generators, where the data can
960  * be regenerated if the page couldn't be grabbed.  This routine should
961  * be safe to call while holding the lock for another page.
962  *
963  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
964  * and deadlock against the caller's locked page.
965  */
966 struct page *
967 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
968 {
969         struct page *page = find_get_page(mapping, index);
970
971         if (page) {
972                 if (trylock_page(page))
973                         return page;
974                 page_cache_release(page);
975                 return NULL;
976         }
977         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
978         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
979                 page_cache_release(page);
980                 page = NULL;
981         }
982         return page;
983 }
984 EXPORT_SYMBOL(grab_cache_page_nowait);
985
986 /*
987  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
988  * a _large_ part of the i/o request. Imagine the worst scenario:
989  *
990  *      ---R__________________________________________B__________
991  *         ^ reading here                             ^ bad block(assume 4k)
992  *
993  * read(R) => miss => readahead(R...B) => media error => frustrating retries
994  * => failing the whole request => read(R) => read(R+1) =>
995  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
996  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
997  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
998  *
999  * It is going insane. Fix it by quickly scaling down the readahead size.
1000  */
1001 static void shrink_readahead_size_eio(struct file *filp,
1002                                         struct file_ra_state *ra)
1003 {
1004         if (!ra->ra_pages)
1005                 return;
1006
1007         ra->ra_pages /= 4;
1008 }
1009
1010 /**
1011  * do_generic_file_read - generic file read routine
1012  * @filp:       the file to read
1013  * @ppos:       current file position
1014  * @desc:       read_descriptor
1015  * @actor:      read method
1016  *
1017  * This is a generic file read routine, and uses the
1018  * mapping->a_ops->readpage() function for the actual low-level stuff.
1019  *
1020  * This is really ugly. But the goto's actually try to clarify some
1021  * of the logic when it comes to error handling etc.
1022  */
1023 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1024                 read_descriptor_t *desc, read_actor_t actor)
1025 {
1026         struct address_space *mapping = filp->f_mapping;
1027         struct inode *inode = mapping->host;
1028         struct file_ra_state *ra = &filp->f_ra;
1029         pgoff_t index;
1030         pgoff_t last_index;
1031         pgoff_t prev_index;
1032         unsigned long offset;      /* offset into pagecache page */
1033         unsigned int prev_offset;
1034         int error;
1035
1036         index = *ppos >> PAGE_CACHE_SHIFT;
1037         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1038         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1039         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1040         offset = *ppos & ~PAGE_CACHE_MASK;
1041
1042         for (;;) {
1043                 struct page *page;
1044                 pgoff_t end_index;
1045                 loff_t isize;
1046                 unsigned long nr, ret;
1047
1048                 cond_resched();
1049 find_page:
1050                 page = find_get_page(mapping, index);
1051                 if (!page) {
1052                         page_cache_sync_readahead(mapping,
1053                                         ra, filp,
1054                                         index, last_index - index);
1055                         page = find_get_page(mapping, index);
1056                         if (unlikely(page == NULL))
1057                                 goto no_cached_page;
1058                 }
1059                 if (PageReadahead(page)) {
1060                         page_cache_async_readahead(mapping,
1061                                         ra, filp, page,
1062                                         index, last_index - index);
1063                 }
1064                 if (!PageUptodate(page)) {
1065                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1066                                         !mapping->a_ops->is_partially_uptodate)
1067                                 goto page_not_up_to_date;
1068                         if (!trylock_page(page))
1069                                 goto page_not_up_to_date;
1070                         if (!mapping->a_ops->is_partially_uptodate(page,
1071                                                                 desc, offset))
1072                                 goto page_not_up_to_date_locked;
1073                         unlock_page(page);
1074                 }
1075 page_ok:
1076                 /*
1077                  * i_size must be checked after we know the page is Uptodate.
1078                  *
1079                  * Checking i_size after the check allows us to calculate
1080                  * the correct value for "nr", which means the zero-filled
1081                  * part of the page is not copied back to userspace (unless
1082                  * another truncate extends the file - this is desired though).
1083                  */
1084
1085                 isize = i_size_read(inode);
1086                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1087                 if (unlikely(!isize || index > end_index)) {
1088                         page_cache_release(page);
1089                         goto out;
1090                 }
1091
1092                 /* nr is the maximum number of bytes to copy from this page */
1093                 nr = PAGE_CACHE_SIZE;
1094                 if (index == end_index) {
1095                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1096                         if (nr <= offset) {
1097                                 page_cache_release(page);
1098                                 goto out;
1099                         }
1100                 }
1101                 nr = nr - offset;
1102
1103                 /* If users can be writing to this page using arbitrary
1104                  * virtual addresses, take care about potential aliasing
1105                  * before reading the page on the kernel side.
1106                  */
1107                 if (mapping_writably_mapped(mapping))
1108                         flush_dcache_page(page);
1109
1110                 /*
1111                  * When a sequential read accesses a page several times,
1112                  * only mark it as accessed the first time.
1113                  */
1114                 if (prev_index != index || offset != prev_offset)
1115                         mark_page_accessed(page);
1116                 prev_index = index;
1117
1118                 /*
1119                  * Ok, we have the page, and it's up-to-date, so
1120                  * now we can copy it to user space...
1121                  *
1122                  * The actor routine returns how many bytes were actually used..
1123                  * NOTE! This may not be the same as how much of a user buffer
1124                  * we filled up (we may be padding etc), so we can only update
1125                  * "pos" here (the actor routine has to update the user buffer
1126                  * pointers and the remaining count).
1127                  */
1128                 ret = actor(desc, page, offset, nr);
1129                 offset += ret;
1130                 index += offset >> PAGE_CACHE_SHIFT;
1131                 offset &= ~PAGE_CACHE_MASK;
1132                 prev_offset = offset;
1133
1134                 page_cache_release(page);
1135                 if (ret == nr && desc->count)
1136                         continue;
1137                 goto out;
1138
1139 page_not_up_to_date:
1140                 /* Get exclusive access to the page ... */
1141                 error = lock_page_killable(page);
1142                 if (unlikely(error))
1143                         goto readpage_error;
1144
1145 page_not_up_to_date_locked:
1146                 /* Did it get truncated before we got the lock? */
1147                 if (!page->mapping) {
1148                         unlock_page(page);
1149                         page_cache_release(page);
1150                         continue;
1151                 }
1152
1153                 /* Did somebody else fill it already? */
1154                 if (PageUptodate(page)) {
1155                         unlock_page(page);
1156                         goto page_ok;
1157                 }
1158
1159 readpage:
1160                 /* Start the actual read. The read will unlock the page. */
1161                 error = mapping->a_ops->readpage(filp, page);
1162
1163                 if (unlikely(error)) {
1164                         if (error == AOP_TRUNCATED_PAGE) {
1165                                 page_cache_release(page);
1166                                 goto find_page;
1167                         }
1168                         goto readpage_error;
1169                 }
1170
1171                 if (!PageUptodate(page)) {
1172                         error = lock_page_killable(page);
1173                         if (unlikely(error))
1174                                 goto readpage_error;
1175                         if (!PageUptodate(page)) {
1176                                 if (page->mapping == NULL) {
1177                                         /*
1178                                          * invalidate_inode_pages got it
1179                                          */
1180                                         unlock_page(page);
1181                                         page_cache_release(page);
1182                                         goto find_page;
1183                                 }
1184                                 unlock_page(page);
1185                                 shrink_readahead_size_eio(filp, ra);
1186                                 error = -EIO;
1187                                 goto readpage_error;
1188                         }
1189                         unlock_page(page);
1190                 }
1191
1192                 goto page_ok;
1193
1194 readpage_error:
1195                 /* UHHUH! A synchronous read error occurred. Report it */
1196                 desc->error = error;
1197                 page_cache_release(page);
1198                 goto out;
1199
1200 no_cached_page:
1201                 /*
1202                  * Ok, it wasn't cached, so we need to create a new
1203                  * page..
1204                  */
1205                 page = page_cache_alloc_cold(mapping);
1206                 if (!page) {
1207                         desc->error = -ENOMEM;
1208                         goto out;
1209                 }
1210                 error = add_to_page_cache_lru(page, mapping,
1211                                                 index, GFP_KERNEL);
1212                 if (error) {
1213                         page_cache_release(page);
1214                         if (error == -EEXIST)
1215                                 goto find_page;
1216                         desc->error = error;
1217                         goto out;
1218                 }
1219                 goto readpage;
1220         }
1221
1222 out:
1223         ra->prev_pos = prev_index;
1224         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1225         ra->prev_pos |= prev_offset;
1226
1227         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1228         file_accessed(filp);
1229 }
1230
1231 int file_read_actor(read_descriptor_t *desc, struct page *page,
1232                         unsigned long offset, unsigned long size)
1233 {
1234         char *kaddr;
1235         unsigned long left, count = desc->count;
1236
1237         if (size > count)
1238                 size = count;
1239
1240         /*
1241          * Faults on the destination of a read are common, so do it before
1242          * taking the kmap.
1243          */
1244         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1245                 kaddr = kmap_atomic(page, KM_USER0);
1246                 left = __copy_to_user_inatomic(desc->arg.buf,
1247                                                 kaddr + offset, size);
1248                 kunmap_atomic(kaddr, KM_USER0);
1249                 if (left == 0)
1250                         goto success;
1251         }
1252
1253         /* Do it the slow way */
1254         kaddr = kmap(page);
1255         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1256         kunmap(page);
1257
1258         if (left) {
1259                 size -= left;
1260                 desc->error = -EFAULT;
1261         }
1262 success:
1263         desc->count = count - size;
1264         desc->written += size;
1265         desc->arg.buf += size;
1266         return size;
1267 }
1268
1269 /*
1270  * Performs necessary checks before doing a write
1271  * @iov:        io vector request
1272  * @nr_segs:    number of segments in the iovec
1273  * @count:      number of bytes to write
1274  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1275  *
1276  * Adjust number of segments and amount of bytes to write (nr_segs should be
1277  * properly initialized first). Returns appropriate error code that caller
1278  * should return or zero in case that write should be allowed.
1279  */
1280 int generic_segment_checks(const struct iovec *iov,
1281                         unsigned long *nr_segs, size_t *count, int access_flags)
1282 {
1283         unsigned long   seg;
1284         size_t cnt = 0;
1285         for (seg = 0; seg < *nr_segs; seg++) {
1286                 const struct iovec *iv = &iov[seg];
1287
1288                 /*
1289                  * If any segment has a negative length, or the cumulative
1290                  * length ever wraps negative then return -EINVAL.
1291                  */
1292                 cnt += iv->iov_len;
1293                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1294                         return -EINVAL;
1295                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1296                         continue;
1297                 if (seg == 0)
1298                         return -EFAULT;
1299                 *nr_segs = seg;
1300                 cnt -= iv->iov_len;     /* This segment is no good */
1301                 break;
1302         }
1303         *count = cnt;
1304         return 0;
1305 }
1306 EXPORT_SYMBOL(generic_segment_checks);
1307
1308 /**
1309  * generic_file_aio_read - generic filesystem read routine
1310  * @iocb:       kernel I/O control block
1311  * @iov:        io vector request
1312  * @nr_segs:    number of segments in the iovec
1313  * @pos:        current file position
1314  *
1315  * This is the "read()" routine for all filesystems
1316  * that can use the page cache directly.
1317  */
1318 ssize_t
1319 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1320                 unsigned long nr_segs, loff_t pos)
1321 {
1322         struct file *filp = iocb->ki_filp;
1323         ssize_t retval;
1324         unsigned long seg;
1325         size_t count;
1326         loff_t *ppos = &iocb->ki_pos;
1327
1328         count = 0;
1329         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1330         if (retval)
1331                 return retval;
1332
1333         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1334         if (filp->f_flags & O_DIRECT) {
1335                 loff_t size;
1336                 struct address_space *mapping;
1337                 struct inode *inode;
1338
1339                 mapping = filp->f_mapping;
1340                 inode = mapping->host;
1341                 if (!count)
1342                         goto out; /* skip atime */
1343                 size = i_size_read(inode);
1344                 if (pos < size) {
1345                         retval = filemap_write_and_wait_range(mapping, pos,
1346                                         pos + iov_length(iov, nr_segs) - 1);
1347                         if (!retval) {
1348                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1349                                                         iov, pos, nr_segs);
1350                         }
1351                         if (retval > 0)
1352                                 *ppos = pos + retval;
1353                         if (retval) {
1354                                 file_accessed(filp);
1355                                 goto out;
1356                         }
1357                 }
1358         }
1359
1360         for (seg = 0; seg < nr_segs; seg++) {
1361                 read_descriptor_t desc;
1362
1363                 desc.written = 0;
1364                 desc.arg.buf = iov[seg].iov_base;
1365                 desc.count = iov[seg].iov_len;
1366                 if (desc.count == 0)
1367                         continue;
1368                 desc.error = 0;
1369                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1370                 retval += desc.written;
1371                 if (desc.error) {
1372                         retval = retval ?: desc.error;
1373                         break;
1374                 }
1375                 if (desc.count > 0)
1376                         break;
1377         }
1378 out:
1379         return retval;
1380 }
1381 EXPORT_SYMBOL(generic_file_aio_read);
1382
1383 static ssize_t
1384 do_readahead(struct address_space *mapping, struct file *filp,
1385              pgoff_t index, unsigned long nr)
1386 {
1387         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1388                 return -EINVAL;
1389
1390         force_page_cache_readahead(mapping, filp, index,
1391                                         max_sane_readahead(nr));
1392         return 0;
1393 }
1394
1395 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1396 {
1397         ssize_t ret;
1398         struct file *file;
1399
1400         ret = -EBADF;
1401         file = fget(fd);
1402         if (file) {
1403                 if (file->f_mode & FMODE_READ) {
1404                         struct address_space *mapping = file->f_mapping;
1405                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1406                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1407                         unsigned long len = end - start + 1;
1408                         ret = do_readahead(mapping, file, start, len);
1409                 }
1410                 fput(file);
1411         }
1412         return ret;
1413 }
1414 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1415 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1416 {
1417         return SYSC_readahead((int) fd, offset, (size_t) count);
1418 }
1419 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1420 #endif
1421
1422 #ifdef CONFIG_MMU
1423 /**
1424  * page_cache_read - adds requested page to the page cache if not already there
1425  * @file:       file to read
1426  * @offset:     page index
1427  *
1428  * This adds the requested page to the page cache if it isn't already there,
1429  * and schedules an I/O to read in its contents from disk.
1430  */
1431 static int page_cache_read(struct file *file, pgoff_t offset)
1432 {
1433         struct address_space *mapping = file->f_mapping;
1434         struct page *page; 
1435         int ret;
1436
1437         do {
1438                 page = page_cache_alloc_cold(mapping);
1439                 if (!page)
1440                         return -ENOMEM;
1441
1442                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1443                 if (ret == 0)
1444                         ret = mapping->a_ops->readpage(file, page);
1445                 else if (ret == -EEXIST)
1446                         ret = 0; /* losing race to add is OK */
1447
1448                 page_cache_release(page);
1449
1450         } while (ret == AOP_TRUNCATED_PAGE);
1451                 
1452         return ret;
1453 }
1454
1455 #define MMAP_LOTSAMISS  (100)
1456
1457 /**
1458  * filemap_fault - read in file data for page fault handling
1459  * @vma:        vma in which the fault was taken
1460  * @vmf:        struct vm_fault containing details of the fault
1461  *
1462  * filemap_fault() is invoked via the vma operations vector for a
1463  * mapped memory region to read in file data during a page fault.
1464  *
1465  * The goto's are kind of ugly, but this streamlines the normal case of having
1466  * it in the page cache, and handles the special cases reasonably without
1467  * having a lot of duplicated code.
1468  */
1469 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1470 {
1471         int error;
1472         struct file *file = vma->vm_file;
1473         struct address_space *mapping = file->f_mapping;
1474         struct file_ra_state *ra = &file->f_ra;
1475         struct inode *inode = mapping->host;
1476         struct page *page;
1477         pgoff_t size;
1478         int did_readaround = 0;
1479         int ret = 0;
1480
1481         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1482         if (vmf->pgoff >= size)
1483                 return VM_FAULT_SIGBUS;
1484
1485         /* If we don't want any read-ahead, don't bother */
1486         if (VM_RandomReadHint(vma))
1487                 goto no_cached_page;
1488
1489         /*
1490          * Do we have something in the page cache already?
1491          */
1492 retry_find:
1493         page = find_lock_page(mapping, vmf->pgoff);
1494         /*
1495          * For sequential accesses, we use the generic readahead logic.
1496          */
1497         if (VM_SequentialReadHint(vma)) {
1498                 if (!page) {
1499                         page_cache_sync_readahead(mapping, ra, file,
1500                                                            vmf->pgoff, 1);
1501                         page = find_lock_page(mapping, vmf->pgoff);
1502                         if (!page)
1503                                 goto no_cached_page;
1504                 }
1505                 if (PageReadahead(page)) {
1506                         page_cache_async_readahead(mapping, ra, file, page,
1507                                                            vmf->pgoff, 1);
1508                 }
1509         }
1510
1511         if (!page) {
1512                 unsigned long ra_pages;
1513
1514                 ra->mmap_miss++;
1515
1516                 /*
1517                  * Do we miss much more than hit in this file? If so,
1518                  * stop bothering with read-ahead. It will only hurt.
1519                  */
1520                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1521                         goto no_cached_page;
1522
1523                 /*
1524                  * To keep the pgmajfault counter straight, we need to
1525                  * check did_readaround, as this is an inner loop.
1526                  */
1527                 if (!did_readaround) {
1528                         ret = VM_FAULT_MAJOR;
1529                         count_vm_event(PGMAJFAULT);
1530                 }
1531                 did_readaround = 1;
1532                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1533                 if (ra_pages) {
1534                         pgoff_t start = 0;
1535
1536                         if (vmf->pgoff > ra_pages / 2)
1537                                 start = vmf->pgoff - ra_pages / 2;
1538                         do_page_cache_readahead(mapping, file, start, ra_pages);
1539                 }
1540                 page = find_lock_page(mapping, vmf->pgoff);
1541                 if (!page)
1542                         goto no_cached_page;
1543         }
1544
1545         if (!did_readaround)
1546                 ra->mmap_miss--;
1547
1548         /*
1549          * We have a locked page in the page cache, now we need to check
1550          * that it's up-to-date. If not, it is going to be due to an error.
1551          */
1552         if (unlikely(!PageUptodate(page)))
1553                 goto page_not_uptodate;
1554
1555         /* Must recheck i_size under page lock */
1556         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1557         if (unlikely(vmf->pgoff >= size)) {
1558                 unlock_page(page);
1559                 page_cache_release(page);
1560                 return VM_FAULT_SIGBUS;
1561         }
1562
1563         /*
1564          * Found the page and have a reference on it.
1565          */
1566         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1567         vmf->page = page;
1568         return ret | VM_FAULT_LOCKED;
1569
1570 no_cached_page:
1571         /*
1572          * We're only likely to ever get here if MADV_RANDOM is in
1573          * effect.
1574          */
1575         error = page_cache_read(file, vmf->pgoff);
1576
1577         /*
1578          * The page we want has now been added to the page cache.
1579          * In the unlikely event that someone removed it in the
1580          * meantime, we'll just come back here and read it again.
1581          */
1582         if (error >= 0)
1583                 goto retry_find;
1584
1585         /*
1586          * An error return from page_cache_read can result if the
1587          * system is low on memory, or a problem occurs while trying
1588          * to schedule I/O.
1589          */
1590         if (error == -ENOMEM)
1591                 return VM_FAULT_OOM;
1592         return VM_FAULT_SIGBUS;
1593
1594 page_not_uptodate:
1595         /* IO error path */
1596         if (!did_readaround) {
1597                 ret = VM_FAULT_MAJOR;
1598                 count_vm_event(PGMAJFAULT);
1599         }
1600
1601         /*
1602          * Umm, take care of errors if the page isn't up-to-date.
1603          * Try to re-read it _once_. We do this synchronously,
1604          * because there really aren't any performance issues here
1605          * and we need to check for errors.
1606          */
1607         ClearPageError(page);
1608         error = mapping->a_ops->readpage(file, page);
1609         if (!error) {
1610                 wait_on_page_locked(page);
1611                 if (!PageUptodate(page))
1612                         error = -EIO;
1613         }
1614         page_cache_release(page);
1615
1616         if (!error || error == AOP_TRUNCATED_PAGE)
1617                 goto retry_find;
1618
1619         /* Things didn't work out. Return zero to tell the mm layer so. */
1620         shrink_readahead_size_eio(file, ra);
1621         return VM_FAULT_SIGBUS;
1622 }
1623 EXPORT_SYMBOL(filemap_fault);
1624
1625 struct vm_operations_struct generic_file_vm_ops = {
1626         .fault          = filemap_fault,
1627 };
1628
1629 /* This is used for a general mmap of a disk file */
1630
1631 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1632 {
1633         struct address_space *mapping = file->f_mapping;
1634
1635         if (!mapping->a_ops->readpage)
1636                 return -ENOEXEC;
1637         file_accessed(file);
1638         vma->vm_ops = &generic_file_vm_ops;
1639         vma->vm_flags |= VM_CAN_NONLINEAR;
1640         return 0;
1641 }
1642
1643 /*
1644  * This is for filesystems which do not implement ->writepage.
1645  */
1646 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1647 {
1648         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1649                 return -EINVAL;
1650         return generic_file_mmap(file, vma);
1651 }
1652 #else
1653 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1654 {
1655         return -ENOSYS;
1656 }
1657 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1658 {
1659         return -ENOSYS;
1660 }
1661 #endif /* CONFIG_MMU */
1662
1663 EXPORT_SYMBOL(generic_file_mmap);
1664 EXPORT_SYMBOL(generic_file_readonly_mmap);
1665
1666 static struct page *__read_cache_page(struct address_space *mapping,
1667                                 pgoff_t index,
1668                                 int (*filler)(void *,struct page*),
1669                                 void *data)
1670 {
1671         struct page *page;
1672         int err;
1673 repeat:
1674         page = find_get_page(mapping, index);
1675         if (!page) {
1676                 page = page_cache_alloc_cold(mapping);
1677                 if (!page)
1678                         return ERR_PTR(-ENOMEM);
1679                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1680                 if (unlikely(err)) {
1681                         page_cache_release(page);
1682                         if (err == -EEXIST)
1683                                 goto repeat;
1684                         /* Presumably ENOMEM for radix tree node */
1685                         return ERR_PTR(err);
1686                 }
1687                 err = filler(data, page);
1688                 if (err < 0) {
1689                         page_cache_release(page);
1690                         page = ERR_PTR(err);
1691                 }
1692         }
1693         return page;
1694 }
1695
1696 /**
1697  * read_cache_page_async - read into page cache, fill it if needed
1698  * @mapping:    the page's address_space
1699  * @index:      the page index
1700  * @filler:     function to perform the read
1701  * @data:       destination for read data
1702  *
1703  * Same as read_cache_page, but don't wait for page to become unlocked
1704  * after submitting it to the filler.
1705  *
1706  * Read into the page cache. If a page already exists, and PageUptodate() is
1707  * not set, try to fill the page but don't wait for it to become unlocked.
1708  *
1709  * If the page does not get brought uptodate, return -EIO.
1710  */
1711 struct page *read_cache_page_async(struct address_space *mapping,
1712                                 pgoff_t index,
1713                                 int (*filler)(void *,struct page*),
1714                                 void *data)
1715 {
1716         struct page *page;
1717         int err;
1718
1719 retry:
1720         page = __read_cache_page(mapping, index, filler, data);
1721         if (IS_ERR(page))
1722                 return page;
1723         if (PageUptodate(page))
1724                 goto out;
1725
1726         lock_page(page);
1727         if (!page->mapping) {
1728                 unlock_page(page);
1729                 page_cache_release(page);
1730                 goto retry;
1731         }
1732         if (PageUptodate(page)) {
1733                 unlock_page(page);
1734                 goto out;
1735         }
1736         err = filler(data, page);
1737         if (err < 0) {
1738                 page_cache_release(page);
1739                 return ERR_PTR(err);
1740         }
1741 out:
1742         mark_page_accessed(page);
1743         return page;
1744 }
1745 EXPORT_SYMBOL(read_cache_page_async);
1746
1747 /**
1748  * read_cache_page - read into page cache, fill it if needed
1749  * @mapping:    the page's address_space
1750  * @index:      the page index
1751  * @filler:     function to perform the read
1752  * @data:       destination for read data
1753  *
1754  * Read into the page cache. If a page already exists, and PageUptodate() is
1755  * not set, try to fill the page then wait for it to become unlocked.
1756  *
1757  * If the page does not get brought uptodate, return -EIO.
1758  */
1759 struct page *read_cache_page(struct address_space *mapping,
1760                                 pgoff_t index,
1761                                 int (*filler)(void *,struct page*),
1762                                 void *data)
1763 {
1764         struct page *page;
1765
1766         page = read_cache_page_async(mapping, index, filler, data);
1767         if (IS_ERR(page))
1768                 goto out;
1769         wait_on_page_locked(page);
1770         if (!PageUptodate(page)) {
1771                 page_cache_release(page);
1772                 page = ERR_PTR(-EIO);
1773         }
1774  out:
1775         return page;
1776 }
1777 EXPORT_SYMBOL(read_cache_page);
1778
1779 /*
1780  * The logic we want is
1781  *
1782  *      if suid or (sgid and xgrp)
1783  *              remove privs
1784  */
1785 int should_remove_suid(struct dentry *dentry)
1786 {
1787         mode_t mode = dentry->d_inode->i_mode;
1788         int kill = 0;
1789
1790         /* suid always must be killed */
1791         if (unlikely(mode & S_ISUID))
1792                 kill = ATTR_KILL_SUID;
1793
1794         /*
1795          * sgid without any exec bits is just a mandatory locking mark; leave
1796          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1797          */
1798         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1799                 kill |= ATTR_KILL_SGID;
1800
1801         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1802                 return kill;
1803
1804         return 0;
1805 }
1806 EXPORT_SYMBOL(should_remove_suid);
1807
1808 static int __remove_suid(struct dentry *dentry, int kill)
1809 {
1810         struct iattr newattrs;
1811
1812         newattrs.ia_valid = ATTR_FORCE | kill;
1813         return notify_change(dentry, &newattrs);
1814 }
1815
1816 int file_remove_suid(struct file *file)
1817 {
1818         struct dentry *dentry = file->f_path.dentry;
1819         int killsuid = should_remove_suid(dentry);
1820         int killpriv = security_inode_need_killpriv(dentry);
1821         int error = 0;
1822
1823         if (killpriv < 0)
1824                 return killpriv;
1825         if (killpriv)
1826                 error = security_inode_killpriv(dentry);
1827         if (!error && killsuid)
1828                 error = __remove_suid(dentry, killsuid);
1829
1830         return error;
1831 }
1832 EXPORT_SYMBOL(file_remove_suid);
1833
1834 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1835                         const struct iovec *iov, size_t base, size_t bytes)
1836 {
1837         size_t copied = 0, left = 0;
1838
1839         while (bytes) {
1840                 char __user *buf = iov->iov_base + base;
1841                 int copy = min(bytes, iov->iov_len - base);
1842
1843                 base = 0;
1844                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1845                 copied += copy;
1846                 bytes -= copy;
1847                 vaddr += copy;
1848                 iov++;
1849
1850                 if (unlikely(left))
1851                         break;
1852         }
1853         return copied - left;
1854 }
1855
1856 /*
1857  * Copy as much as we can into the page and return the number of bytes which
1858  * were sucessfully copied.  If a fault is encountered then return the number of
1859  * bytes which were copied.
1860  */
1861 size_t iov_iter_copy_from_user_atomic(struct page *page,
1862                 struct iov_iter *i, unsigned long offset, size_t bytes)
1863 {
1864         char *kaddr;
1865         size_t copied;
1866
1867         BUG_ON(!in_atomic());
1868         kaddr = kmap_atomic(page, KM_USER0);
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_inatomic(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_atomic(kaddr, KM_USER0);
1879
1880         return copied;
1881 }
1882 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1883
1884 /*
1885  * This has the same sideeffects and return value as
1886  * iov_iter_copy_from_user_atomic().
1887  * The difference is that it attempts to resolve faults.
1888  * Page must not be locked.
1889  */
1890 size_t iov_iter_copy_from_user(struct page *page,
1891                 struct iov_iter *i, unsigned long offset, size_t bytes)
1892 {
1893         char *kaddr;
1894         size_t copied;
1895
1896         kaddr = kmap(page);
1897         if (likely(i->nr_segs == 1)) {
1898                 int left;
1899                 char __user *buf = i->iov->iov_base + i->iov_offset;
1900                 left = __copy_from_user(kaddr + offset, buf, bytes);
1901                 copied = bytes - left;
1902         } else {
1903                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1904                                                 i->iov, i->iov_offset, bytes);
1905         }
1906         kunmap(page);
1907         return copied;
1908 }
1909 EXPORT_SYMBOL(iov_iter_copy_from_user);
1910
1911 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1912 {
1913         BUG_ON(i->count < bytes);
1914
1915         if (likely(i->nr_segs == 1)) {
1916                 i->iov_offset += bytes;
1917                 i->count -= bytes;
1918         } else {
1919                 const struct iovec *iov = i->iov;
1920                 size_t base = i->iov_offset;
1921
1922                 /*
1923                  * The !iov->iov_len check ensures we skip over unlikely
1924                  * zero-length segments (without overruning the iovec).
1925                  */
1926                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1927                         int copy;
1928
1929                         copy = min(bytes, iov->iov_len - base);
1930                         BUG_ON(!i->count || i->count < copy);
1931                         i->count -= copy;
1932                         bytes -= copy;
1933                         base += copy;
1934                         if (iov->iov_len == base) {
1935                                 iov++;
1936                                 base = 0;
1937                         }
1938                 }
1939                 i->iov = iov;
1940                 i->iov_offset = base;
1941         }
1942 }
1943 EXPORT_SYMBOL(iov_iter_advance);
1944
1945 /*
1946  * Fault in the first iovec of the given iov_iter, to a maximum length
1947  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1948  * accessed (ie. because it is an invalid address).
1949  *
1950  * writev-intensive code may want this to prefault several iovecs -- that
1951  * would be possible (callers must not rely on the fact that _only_ the
1952  * first iovec will be faulted with the current implementation).
1953  */
1954 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1955 {
1956         char __user *buf = i->iov->iov_base + i->iov_offset;
1957         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1958         return fault_in_pages_readable(buf, bytes);
1959 }
1960 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1961
1962 /*
1963  * Return the count of just the current iov_iter segment.
1964  */
1965 size_t iov_iter_single_seg_count(struct iov_iter *i)
1966 {
1967         const struct iovec *iov = i->iov;
1968         if (i->nr_segs == 1)
1969                 return i->count;
1970         else
1971                 return min(i->count, iov->iov_len - i->iov_offset);
1972 }
1973 EXPORT_SYMBOL(iov_iter_single_seg_count);
1974
1975 /*
1976  * Performs necessary checks before doing a write
1977  *
1978  * Can adjust writing position or amount of bytes to write.
1979  * Returns appropriate error code that caller should return or
1980  * zero in case that write should be allowed.
1981  */
1982 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1983 {
1984         struct inode *inode = file->f_mapping->host;
1985         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1986
1987         if (unlikely(*pos < 0))
1988                 return -EINVAL;
1989
1990         if (!isblk) {
1991                 /* FIXME: this is for backwards compatibility with 2.4 */
1992                 if (file->f_flags & O_APPEND)
1993                         *pos = i_size_read(inode);
1994
1995                 if (limit != RLIM_INFINITY) {
1996                         if (*pos >= limit) {
1997                                 send_sig(SIGXFSZ, current, 0);
1998                                 return -EFBIG;
1999                         }
2000                         if (*count > limit - (typeof(limit))*pos) {
2001                                 *count = limit - (typeof(limit))*pos;
2002                         }
2003                 }
2004         }
2005
2006         /*
2007          * LFS rule
2008          */
2009         if (unlikely(*pos + *count > MAX_NON_LFS &&
2010                                 !(file->f_flags & O_LARGEFILE))) {
2011                 if (*pos >= MAX_NON_LFS) {
2012                         return -EFBIG;
2013                 }
2014                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2015                         *count = MAX_NON_LFS - (unsigned long)*pos;
2016                 }
2017         }
2018
2019         /*
2020          * Are we about to exceed the fs block limit ?
2021          *
2022          * If we have written data it becomes a short write.  If we have
2023          * exceeded without writing data we send a signal and return EFBIG.
2024          * Linus frestrict idea will clean these up nicely..
2025          */
2026         if (likely(!isblk)) {
2027                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2028                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2029                                 return -EFBIG;
2030                         }
2031                         /* zero-length writes at ->s_maxbytes are OK */
2032                 }
2033
2034                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2035                         *count = inode->i_sb->s_maxbytes - *pos;
2036         } else {
2037 #ifdef CONFIG_BLOCK
2038                 loff_t isize;
2039                 if (bdev_read_only(I_BDEV(inode)))
2040                         return -EPERM;
2041                 isize = i_size_read(inode);
2042                 if (*pos >= isize) {
2043                         if (*count || *pos > isize)
2044                                 return -ENOSPC;
2045                 }
2046
2047                 if (*pos + *count > isize)
2048                         *count = isize - *pos;
2049 #else
2050                 return -EPERM;
2051 #endif
2052         }
2053         return 0;
2054 }
2055 EXPORT_SYMBOL(generic_write_checks);
2056
2057 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2058                                 loff_t pos, unsigned len, unsigned flags,
2059                                 struct page **pagep, void **fsdata)
2060 {
2061         const struct address_space_operations *aops = mapping->a_ops;
2062
2063         return aops->write_begin(file, mapping, pos, len, flags,
2064                                                         pagep, fsdata);
2065 }
2066 EXPORT_SYMBOL(pagecache_write_begin);
2067
2068 int pagecache_write_end(struct file *file, struct address_space *mapping,
2069                                 loff_t pos, unsigned len, unsigned copied,
2070                                 struct page *page, void *fsdata)
2071 {
2072         const struct address_space_operations *aops = mapping->a_ops;
2073
2074         mark_page_accessed(page);
2075         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2076 }
2077 EXPORT_SYMBOL(pagecache_write_end);
2078
2079 ssize_t
2080 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2081                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2082                 size_t count, size_t ocount)
2083 {
2084         struct file     *file = iocb->ki_filp;
2085         struct address_space *mapping = file->f_mapping;
2086         struct inode    *inode = mapping->host;
2087         ssize_t         written;
2088         size_t          write_len;
2089         pgoff_t         end;
2090
2091         if (count != ocount)
2092                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2093
2094         write_len = iov_length(iov, *nr_segs);
2095         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2096
2097         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2098         if (written)
2099                 goto out;
2100
2101         /*
2102          * After a write we want buffered reads to be sure to go to disk to get
2103          * the new data.  We invalidate clean cached page from the region we're
2104          * about to write.  We do this *before* the write so that we can return
2105          * without clobbering -EIOCBQUEUED from ->direct_IO().
2106          */
2107         if (mapping->nrpages) {
2108                 written = invalidate_inode_pages2_range(mapping,
2109                                         pos >> PAGE_CACHE_SHIFT, end);
2110                 /*
2111                  * If a page can not be invalidated, return 0 to fall back
2112                  * to buffered write.
2113                  */
2114                 if (written) {
2115                         if (written == -EBUSY)
2116                                 return 0;
2117                         goto out;
2118                 }
2119         }
2120
2121         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2122
2123         /*
2124          * Finally, try again to invalidate clean pages which might have been
2125          * cached by non-direct readahead, or faulted in by get_user_pages()
2126          * if the source of the write was an mmap'ed region of the file
2127          * we're writing.  Either one is a pretty crazy thing to do,
2128          * so we don't support it 100%.  If this invalidation
2129          * fails, tough, the write still worked...
2130          */
2131         if (mapping->nrpages) {
2132                 invalidate_inode_pages2_range(mapping,
2133                                               pos >> PAGE_CACHE_SHIFT, end);
2134         }
2135
2136         if (written > 0) {
2137                 loff_t end = pos + written;
2138                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2139                         i_size_write(inode,  end);
2140                         mark_inode_dirty(inode);
2141                 }
2142                 *ppos = end;
2143         }
2144
2145         /*
2146          * Sync the fs metadata but not the minor inode changes and
2147          * of course not the data as we did direct DMA for the IO.
2148          * i_mutex is held, which protects generic_osync_inode() from
2149          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2150          */
2151 out:
2152         if ((written >= 0 || written == -EIOCBQUEUED) &&
2153             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2154                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2155                 if (err < 0)
2156                         written = err;
2157         }
2158         return written;
2159 }
2160 EXPORT_SYMBOL(generic_file_direct_write);
2161
2162 /*
2163  * Find or create a page at the given pagecache position. Return the locked
2164  * page. This function is specifically for buffered writes.
2165  */
2166 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2167                                         pgoff_t index, unsigned flags)
2168 {
2169         int status;
2170         struct page *page;
2171         gfp_t gfp_notmask = 0;
2172         if (flags & AOP_FLAG_NOFS)
2173                 gfp_notmask = __GFP_FS;
2174 repeat:
2175         page = find_lock_page(mapping, index);
2176         if (likely(page))
2177                 return page;
2178
2179         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2180         if (!page)
2181                 return NULL;
2182         status = add_to_page_cache_lru(page, mapping, index,
2183                                                 GFP_KERNEL & ~gfp_notmask);
2184         if (unlikely(status)) {
2185                 page_cache_release(page);
2186                 if (status == -EEXIST)
2187                         goto repeat;
2188                 return NULL;
2189         }
2190         return page;
2191 }
2192 EXPORT_SYMBOL(grab_cache_page_write_begin);
2193
2194 static ssize_t generic_perform_write(struct file *file,
2195                                 struct iov_iter *i, loff_t pos)
2196 {
2197         struct address_space *mapping = file->f_mapping;
2198         const struct address_space_operations *a_ops = mapping->a_ops;
2199         long status = 0;
2200         ssize_t written = 0;
2201         unsigned int flags = 0;
2202
2203         /*
2204          * Copies from kernel address space cannot fail (NFSD is a big user).
2205          */
2206         if (segment_eq(get_fs(), KERNEL_DS))
2207                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2208
2209         do {
2210                 struct page *page;
2211                 pgoff_t index;          /* Pagecache index for current page */
2212                 unsigned long offset;   /* Offset into pagecache page */
2213                 unsigned long bytes;    /* Bytes to write to page */
2214                 size_t copied;          /* Bytes copied from user */
2215                 void *fsdata;
2216
2217                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2218                 index = pos >> PAGE_CACHE_SHIFT;
2219                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2220                                                 iov_iter_count(i));
2221
2222 again:
2223
2224                 /*
2225                  * Bring in the user page that we will copy from _first_.
2226                  * Otherwise there's a nasty deadlock on copying from the
2227                  * same page as we're writing to, without it being marked
2228                  * up-to-date.
2229                  *
2230                  * Not only is this an optimisation, but it is also required
2231                  * to check that the address is actually valid, when atomic
2232                  * usercopies are used, below.
2233                  */
2234                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2235                         status = -EFAULT;
2236                         break;
2237                 }
2238
2239                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2240                                                 &page, &fsdata);
2241                 if (unlikely(status))
2242                         break;
2243
2244                 pagefault_disable();
2245                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2246                 pagefault_enable();
2247                 flush_dcache_page(page);
2248
2249                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2250                                                 page, fsdata);
2251                 if (unlikely(status < 0))
2252                         break;
2253                 copied = status;
2254
2255                 cond_resched();
2256
2257                 iov_iter_advance(i, copied);
2258                 if (unlikely(copied == 0)) {
2259                         /*
2260                          * If we were unable to copy any data at all, we must
2261                          * fall back to a single segment length write.
2262                          *
2263                          * If we didn't fallback here, we could livelock
2264                          * because not all segments in the iov can be copied at
2265                          * once without a pagefault.
2266                          */
2267                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2268                                                 iov_iter_single_seg_count(i));
2269                         goto again;
2270                 }
2271                 pos += copied;
2272                 written += copied;
2273
2274                 balance_dirty_pages_ratelimited(mapping);
2275
2276         } while (iov_iter_count(i));
2277
2278         return written ? written : status;
2279 }
2280
2281 ssize_t
2282 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2283                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2284                 size_t count, ssize_t written)
2285 {
2286         struct file *file = iocb->ki_filp;
2287         struct address_space *mapping = file->f_mapping;
2288         const struct address_space_operations *a_ops = mapping->a_ops;
2289         struct inode *inode = mapping->host;
2290         ssize_t status;
2291         struct iov_iter i;
2292
2293         iov_iter_init(&i, iov, nr_segs, count, written);
2294         status = generic_perform_write(file, &i, pos);
2295
2296         if (likely(status >= 0)) {
2297                 written += status;
2298                 *ppos = pos + status;
2299
2300                 /*
2301                  * For now, when the user asks for O_SYNC, we'll actually give
2302                  * O_DSYNC
2303                  */
2304                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2305                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2306                                 status = generic_osync_inode(inode, mapping,
2307                                                 OSYNC_METADATA|OSYNC_DATA);
2308                 }
2309         }
2310         
2311         /*
2312          * If we get here for O_DIRECT writes then we must have fallen through
2313          * to buffered writes (block instantiation inside i_size).  So we sync
2314          * the file data here, to try to honour O_DIRECT expectations.
2315          */
2316         if (unlikely(file->f_flags & O_DIRECT) && written)
2317                 status = filemap_write_and_wait_range(mapping,
2318                                         pos, pos + written - 1);
2319
2320         return written ? written : status;
2321 }
2322 EXPORT_SYMBOL(generic_file_buffered_write);
2323
2324 static ssize_t
2325 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2326                                 unsigned long nr_segs, loff_t *ppos)
2327 {
2328         struct file *file = iocb->ki_filp;
2329         struct address_space * mapping = file->f_mapping;
2330         size_t ocount;          /* original count */
2331         size_t count;           /* after file limit checks */
2332         struct inode    *inode = mapping->host;
2333         loff_t          pos;
2334         ssize_t         written;
2335         ssize_t         err;
2336
2337         ocount = 0;
2338         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2339         if (err)
2340                 return err;
2341
2342         count = ocount;
2343         pos = *ppos;
2344
2345         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2346
2347         /* We can write back this queue in page reclaim */
2348         current->backing_dev_info = mapping->backing_dev_info;
2349         written = 0;
2350
2351         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2352         if (err)
2353                 goto out;
2354
2355         if (count == 0)
2356                 goto out;
2357
2358         err = file_remove_suid(file);
2359         if (err)
2360                 goto out;
2361
2362         file_update_time(file);
2363
2364         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2365         if (unlikely(file->f_flags & O_DIRECT)) {
2366                 loff_t endbyte;
2367                 ssize_t written_buffered;
2368
2369                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2370                                                         ppos, count, ocount);
2371                 if (written < 0 || written == count)
2372                         goto out;
2373                 /*
2374                  * direct-io write to a hole: fall through to buffered I/O
2375                  * for completing the rest of the request.
2376                  */
2377                 pos += written;
2378                 count -= written;
2379                 written_buffered = generic_file_buffered_write(iocb, iov,
2380                                                 nr_segs, pos, ppos, count,
2381                                                 written);
2382                 /*
2383                  * If generic_file_buffered_write() retuned a synchronous error
2384                  * then we want to return the number of bytes which were
2385                  * direct-written, or the error code if that was zero.  Note
2386                  * that this differs from normal direct-io semantics, which
2387                  * will return -EFOO even if some bytes were written.
2388                  */
2389                 if (written_buffered < 0) {
2390                         err = written_buffered;
2391                         goto out;
2392                 }
2393
2394                 /*
2395                  * We need to ensure that the page cache pages are written to
2396                  * disk and invalidated to preserve the expected O_DIRECT
2397                  * semantics.
2398                  */
2399                 endbyte = pos + written_buffered - written - 1;
2400                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2401                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2402                                             SYNC_FILE_RANGE_WRITE|
2403                                             SYNC_FILE_RANGE_WAIT_AFTER);
2404                 if (err == 0) {
2405                         written = written_buffered;
2406                         invalidate_mapping_pages(mapping,
2407                                                  pos >> PAGE_CACHE_SHIFT,
2408                                                  endbyte >> PAGE_CACHE_SHIFT);
2409                 } else {
2410                         /*
2411                          * We don't know how much we wrote, so just return
2412                          * the number of bytes which were direct-written
2413                          */
2414                 }
2415         } else {
2416                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2417                                 pos, ppos, count, written);
2418         }
2419 out:
2420         current->backing_dev_info = NULL;
2421         return written ? written : err;
2422 }
2423
2424 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2425                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2426 {
2427         struct file *file = iocb->ki_filp;
2428         struct address_space *mapping = file->f_mapping;
2429         struct inode *inode = mapping->host;
2430         ssize_t ret;
2431
2432         BUG_ON(iocb->ki_pos != pos);
2433
2434         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2435                         &iocb->ki_pos);
2436
2437         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2438                 ssize_t err;
2439
2440                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2441                 if (err < 0)
2442                         ret = err;
2443         }
2444         return ret;
2445 }
2446 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2447
2448 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2449                 unsigned long nr_segs, loff_t pos)
2450 {
2451         struct file *file = iocb->ki_filp;
2452         struct address_space *mapping = file->f_mapping;
2453         struct inode *inode = mapping->host;
2454         ssize_t ret;
2455
2456         BUG_ON(iocb->ki_pos != pos);
2457
2458         mutex_lock(&inode->i_mutex);
2459         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2460                         &iocb->ki_pos);
2461         mutex_unlock(&inode->i_mutex);
2462
2463         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2464                 ssize_t err;
2465
2466                 err = sync_page_range(inode, mapping, pos, ret);
2467                 if (err < 0)
2468                         ret = err;
2469         }
2470         return ret;
2471 }
2472 EXPORT_SYMBOL(generic_file_aio_write);
2473
2474 /**
2475  * try_to_release_page() - release old fs-specific metadata on a page
2476  *
2477  * @page: the page which the kernel is trying to free
2478  * @gfp_mask: memory allocation flags (and I/O mode)
2479  *
2480  * The address_space is to try to release any data against the page
2481  * (presumably at page->private).  If the release was successful, return `1'.
2482  * Otherwise return zero.
2483  *
2484  * This may also be called if PG_fscache is set on a page, indicating that the
2485  * page is known to the local caching routines.
2486  *
2487  * The @gfp_mask argument specifies whether I/O may be performed to release
2488  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2489  *
2490  */
2491 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2492 {
2493         struct address_space * const mapping = page->mapping;
2494
2495         BUG_ON(!PageLocked(page));
2496         if (PageWriteback(page))
2497                 return 0;
2498
2499         if (mapping && mapping->a_ops->releasepage)
2500                 return mapping->a_ops->releasepage(page, gfp_mask);
2501         return try_to_free_buffers(page);
2502 }
2503
2504 EXPORT_SYMBOL(try_to_release_page);