vfs: Rename generic_file_aio_write_nolock
[linux-3.10.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include <linux/mm_inline.h> /* for page_is_file_cache() */
37 #include "internal.h"
38
39 /*
40  * FIXME: remove all knowledge of the buffer layer from the core VM
41  */
42 #include <linux/buffer_head.h> /* for 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
125         /*
126          * Some filesystems seem to re-dirty the page even after
127          * the VM has canceled the dirty bit (eg ext3 journaling).
128          *
129          * Fix it up by doing a final dirty accounting check after
130          * having removed the page entirely.
131          */
132         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
133                 dec_zone_page_state(page, NR_FILE_DIRTY);
134                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
135         }
136 }
137
138 void remove_from_page_cache(struct page *page)
139 {
140         struct address_space *mapping = page->mapping;
141
142         BUG_ON(!PageLocked(page));
143
144         spin_lock_irq(&mapping->tree_lock);
145         __remove_from_page_cache(page);
146         spin_unlock_irq(&mapping->tree_lock);
147         mem_cgroup_uncharge_cache_page(page);
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  * filemap_fdatawait_range - wait for all under-writeback pages to complete in a given range
311  * @mapping: address space structure to wait for
312  * @start:      offset in bytes where the range starts
313  * @end:        offset in bytes where the range ends (inclusive)
314  *
315  * Walk the list of under-writeback pages of the given address space
316  * in the given range and wait for all of them.
317  *
318  * This is just a simple wrapper so that callers don't have to convert offsets
319  * to page indexes themselves
320  */
321 int filemap_fdatawait_range(struct address_space *mapping, loff_t start,
322                             loff_t end)
323 {
324         return wait_on_page_writeback_range(mapping, start >> PAGE_CACHE_SHIFT,
325                                             end >> PAGE_CACHE_SHIFT);
326 }
327 EXPORT_SYMBOL(filemap_fdatawait_range);
328
329 /**
330  * sync_page_range - write and wait on all pages in the passed range
331  * @inode:      target inode
332  * @mapping:    target address_space
333  * @pos:        beginning offset in pages to write
334  * @count:      number of bytes to write
335  *
336  * Write and wait upon all the pages in the passed range.  This is a "data
337  * integrity" operation.  It waits upon in-flight writeout before starting and
338  * waiting upon new writeout.  If there was an IO error, return it.
339  *
340  * We need to re-take i_mutex during the generic_osync_inode list walk because
341  * it is otherwise livelockable.
342  */
343 int sync_page_range(struct inode *inode, struct address_space *mapping,
344                         loff_t pos, loff_t count)
345 {
346         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
347         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
348         int ret;
349
350         if (!mapping_cap_writeback_dirty(mapping) || !count)
351                 return 0;
352         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
353         if (ret == 0) {
354                 mutex_lock(&inode->i_mutex);
355                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
356                 mutex_unlock(&inode->i_mutex);
357         }
358         if (ret == 0)
359                 ret = wait_on_page_writeback_range(mapping, start, end);
360         return ret;
361 }
362 EXPORT_SYMBOL(sync_page_range);
363
364 /**
365  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
366  * @inode:      target inode
367  * @mapping:    target address_space
368  * @pos:        beginning offset in pages to write
369  * @count:      number of bytes to write
370  *
371  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
372  * as it forces O_SYNC writers to different parts of the same file
373  * to be serialised right until io completion.
374  */
375 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
376                            loff_t pos, loff_t count)
377 {
378         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
379         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
380         int ret;
381
382         if (!mapping_cap_writeback_dirty(mapping) || !count)
383                 return 0;
384         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
385         if (ret == 0)
386                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
387         if (ret == 0)
388                 ret = wait_on_page_writeback_range(mapping, start, end);
389         return ret;
390 }
391 EXPORT_SYMBOL(sync_page_range_nolock);
392
393 /**
394  * filemap_fdatawait - wait for all under-writeback pages to complete
395  * @mapping: address space structure to wait for
396  *
397  * Walk the list of under-writeback pages of the given address space
398  * and wait for all of them.
399  */
400 int filemap_fdatawait(struct address_space *mapping)
401 {
402         loff_t i_size = i_size_read(mapping->host);
403
404         if (i_size == 0)
405                 return 0;
406
407         return wait_on_page_writeback_range(mapping, 0,
408                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
409 }
410 EXPORT_SYMBOL(filemap_fdatawait);
411
412 int filemap_write_and_wait(struct address_space *mapping)
413 {
414         int err = 0;
415
416         if (mapping->nrpages) {
417                 err = filemap_fdatawrite(mapping);
418                 /*
419                  * Even if the above returned error, the pages may be
420                  * written partially (e.g. -ENOSPC), so we wait for it.
421                  * But the -EIO is special case, it may indicate the worst
422                  * thing (e.g. bug) happened, so we avoid waiting for it.
423                  */
424                 if (err != -EIO) {
425                         int err2 = filemap_fdatawait(mapping);
426                         if (!err)
427                                 err = err2;
428                 }
429         }
430         return err;
431 }
432 EXPORT_SYMBOL(filemap_write_and_wait);
433
434 /**
435  * filemap_write_and_wait_range - write out & wait on a file range
436  * @mapping:    the address_space for the pages
437  * @lstart:     offset in bytes where the range starts
438  * @lend:       offset in bytes where the range ends (inclusive)
439  *
440  * Write out and wait upon file offsets lstart->lend, inclusive.
441  *
442  * Note that `lend' is inclusive (describes the last byte to be written) so
443  * that this function can be used to write to the very end-of-file (end = -1).
444  */
445 int filemap_write_and_wait_range(struct address_space *mapping,
446                                  loff_t lstart, loff_t lend)
447 {
448         int err = 0;
449
450         if (mapping->nrpages) {
451                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
452                                                  WB_SYNC_ALL);
453                 /* See comment of filemap_write_and_wait() */
454                 if (err != -EIO) {
455                         int err2 = wait_on_page_writeback_range(mapping,
456                                                 lstart >> PAGE_CACHE_SHIFT,
457                                                 lend >> PAGE_CACHE_SHIFT);
458                         if (!err)
459                                 err = err2;
460                 }
461         }
462         return err;
463 }
464 EXPORT_SYMBOL(filemap_write_and_wait_range);
465
466 /**
467  * add_to_page_cache_locked - add a locked page to the pagecache
468  * @page:       page to add
469  * @mapping:    the page's address_space
470  * @offset:     page index
471  * @gfp_mask:   page allocation mode
472  *
473  * This function is used to add a page to the pagecache. It must be locked.
474  * This function does not add the page to the LRU.  The caller must do that.
475  */
476 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
477                 pgoff_t offset, gfp_t gfp_mask)
478 {
479         int error;
480
481         VM_BUG_ON(!PageLocked(page));
482
483         error = mem_cgroup_cache_charge(page, current->mm,
484                                         gfp_mask & GFP_RECLAIM_MASK);
485         if (error)
486                 goto out;
487
488         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
489         if (error == 0) {
490                 page_cache_get(page);
491                 page->mapping = mapping;
492                 page->index = offset;
493
494                 spin_lock_irq(&mapping->tree_lock);
495                 error = radix_tree_insert(&mapping->page_tree, offset, page);
496                 if (likely(!error)) {
497                         mapping->nrpages++;
498                         __inc_zone_page_state(page, NR_FILE_PAGES);
499                         spin_unlock_irq(&mapping->tree_lock);
500                 } else {
501                         page->mapping = NULL;
502                         spin_unlock_irq(&mapping->tree_lock);
503                         mem_cgroup_uncharge_cache_page(page);
504                         page_cache_release(page);
505                 }
506                 radix_tree_preload_end();
507         } else
508                 mem_cgroup_uncharge_cache_page(page);
509 out:
510         return error;
511 }
512 EXPORT_SYMBOL(add_to_page_cache_locked);
513
514 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
515                                 pgoff_t offset, gfp_t gfp_mask)
516 {
517         int ret;
518
519         /*
520          * Splice_read and readahead add shmem/tmpfs pages into the page cache
521          * before shmem_readpage has a chance to mark them as SwapBacked: they
522          * need to go on the active_anon lru below, and mem_cgroup_cache_charge
523          * (called in add_to_page_cache) needs to know where they're going too.
524          */
525         if (mapping_cap_swap_backed(mapping))
526                 SetPageSwapBacked(page);
527
528         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
529         if (ret == 0) {
530                 if (page_is_file_cache(page))
531                         lru_cache_add_file(page);
532                 else
533                         lru_cache_add_active_anon(page);
534         }
535         return ret;
536 }
537 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
538
539 #ifdef CONFIG_NUMA
540 struct page *__page_cache_alloc(gfp_t gfp)
541 {
542         if (cpuset_do_page_mem_spread()) {
543                 int n = cpuset_mem_spread_node();
544                 return alloc_pages_exact_node(n, gfp, 0);
545         }
546         return alloc_pages(gfp, 0);
547 }
548 EXPORT_SYMBOL(__page_cache_alloc);
549 #endif
550
551 static int __sleep_on_page_lock(void *word)
552 {
553         io_schedule();
554         return 0;
555 }
556
557 /*
558  * In order to wait for pages to become available there must be
559  * waitqueues associated with pages. By using a hash table of
560  * waitqueues where the bucket discipline is to maintain all
561  * waiters on the same queue and wake all when any of the pages
562  * become available, and for the woken contexts to check to be
563  * sure the appropriate page became available, this saves space
564  * at a cost of "thundering herd" phenomena during rare hash
565  * collisions.
566  */
567 static wait_queue_head_t *page_waitqueue(struct page *page)
568 {
569         const struct zone *zone = page_zone(page);
570
571         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
572 }
573
574 static inline void wake_up_page(struct page *page, int bit)
575 {
576         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
577 }
578
579 void wait_on_page_bit(struct page *page, int bit_nr)
580 {
581         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
582
583         if (test_bit(bit_nr, &page->flags))
584                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
585                                                         TASK_UNINTERRUPTIBLE);
586 }
587 EXPORT_SYMBOL(wait_on_page_bit);
588
589 /**
590  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
591  * @page: Page defining the wait queue of interest
592  * @waiter: Waiter to add to the queue
593  *
594  * Add an arbitrary @waiter to the wait queue for the nominated @page.
595  */
596 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
597 {
598         wait_queue_head_t *q = page_waitqueue(page);
599         unsigned long flags;
600
601         spin_lock_irqsave(&q->lock, flags);
602         __add_wait_queue(q, waiter);
603         spin_unlock_irqrestore(&q->lock, flags);
604 }
605 EXPORT_SYMBOL_GPL(add_page_wait_queue);
606
607 /**
608  * unlock_page - unlock a locked page
609  * @page: the page
610  *
611  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
612  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
613  * mechananism between PageLocked pages and PageWriteback pages is shared.
614  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
615  *
616  * The mb is necessary to enforce ordering between the clear_bit and the read
617  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
618  */
619 void unlock_page(struct page *page)
620 {
621         VM_BUG_ON(!PageLocked(page));
622         clear_bit_unlock(PG_locked, &page->flags);
623         smp_mb__after_clear_bit();
624         wake_up_page(page, PG_locked);
625 }
626 EXPORT_SYMBOL(unlock_page);
627
628 /**
629  * end_page_writeback - end writeback against a page
630  * @page: the page
631  */
632 void end_page_writeback(struct page *page)
633 {
634         if (TestClearPageReclaim(page))
635                 rotate_reclaimable_page(page);
636
637         if (!test_clear_page_writeback(page))
638                 BUG();
639
640         smp_mb__after_clear_bit();
641         wake_up_page(page, PG_writeback);
642 }
643 EXPORT_SYMBOL(end_page_writeback);
644
645 /**
646  * __lock_page - get a lock on the page, assuming we need to sleep to get it
647  * @page: the page to lock
648  *
649  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
650  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
651  * chances are that on the second loop, the block layer's plug list is empty,
652  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
653  */
654 void __lock_page(struct page *page)
655 {
656         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
657
658         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
659                                                         TASK_UNINTERRUPTIBLE);
660 }
661 EXPORT_SYMBOL(__lock_page);
662
663 int __lock_page_killable(struct page *page)
664 {
665         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
666
667         return __wait_on_bit_lock(page_waitqueue(page), &wait,
668                                         sync_page_killable, TASK_KILLABLE);
669 }
670 EXPORT_SYMBOL_GPL(__lock_page_killable);
671
672 /**
673  * __lock_page_nosync - get a lock on the page, without calling sync_page()
674  * @page: the page to lock
675  *
676  * Variant of lock_page that does not require the caller to hold a reference
677  * on the page's mapping.
678  */
679 void __lock_page_nosync(struct page *page)
680 {
681         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
682         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
683                                                         TASK_UNINTERRUPTIBLE);
684 }
685
686 /**
687  * find_get_page - find and get a page reference
688  * @mapping: the address_space to search
689  * @offset: the page index
690  *
691  * Is there a pagecache struct page at the given (mapping, offset) tuple?
692  * If yes, increment its refcount and return it; if no, return NULL.
693  */
694 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
695 {
696         void **pagep;
697         struct page *page;
698
699         rcu_read_lock();
700 repeat:
701         page = NULL;
702         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
703         if (pagep) {
704                 page = radix_tree_deref_slot(pagep);
705                 if (unlikely(!page || page == RADIX_TREE_RETRY))
706                         goto repeat;
707
708                 if (!page_cache_get_speculative(page))
709                         goto repeat;
710
711                 /*
712                  * Has the page moved?
713                  * This is part of the lockless pagecache protocol. See
714                  * include/linux/pagemap.h for details.
715                  */
716                 if (unlikely(page != *pagep)) {
717                         page_cache_release(page);
718                         goto repeat;
719                 }
720         }
721         rcu_read_unlock();
722
723         return page;
724 }
725 EXPORT_SYMBOL(find_get_page);
726
727 /**
728  * find_lock_page - locate, pin and lock a pagecache page
729  * @mapping: the address_space to search
730  * @offset: the page index
731  *
732  * Locates the desired pagecache page, locks it, increments its reference
733  * count and returns its address.
734  *
735  * Returns zero if the page was not present. find_lock_page() may sleep.
736  */
737 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
738 {
739         struct page *page;
740
741 repeat:
742         page = find_get_page(mapping, offset);
743         if (page) {
744                 lock_page(page);
745                 /* Has the page been truncated? */
746                 if (unlikely(page->mapping != mapping)) {
747                         unlock_page(page);
748                         page_cache_release(page);
749                         goto repeat;
750                 }
751                 VM_BUG_ON(page->index != offset);
752         }
753         return page;
754 }
755 EXPORT_SYMBOL(find_lock_page);
756
757 /**
758  * find_or_create_page - locate or add a pagecache page
759  * @mapping: the page's address_space
760  * @index: the page's index into the mapping
761  * @gfp_mask: page allocation mode
762  *
763  * Locates a page in the pagecache.  If the page is not present, a new page
764  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
765  * LRU list.  The returned page is locked and has its reference count
766  * incremented.
767  *
768  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
769  * allocation!
770  *
771  * find_or_create_page() returns the desired page's address, or zero on
772  * memory exhaustion.
773  */
774 struct page *find_or_create_page(struct address_space *mapping,
775                 pgoff_t index, gfp_t gfp_mask)
776 {
777         struct page *page;
778         int err;
779 repeat:
780         page = find_lock_page(mapping, index);
781         if (!page) {
782                 page = __page_cache_alloc(gfp_mask);
783                 if (!page)
784                         return NULL;
785                 /*
786                  * We want a regular kernel memory (not highmem or DMA etc)
787                  * allocation for the radix tree nodes, but we need to honour
788                  * the context-specific requirements the caller has asked for.
789                  * GFP_RECLAIM_MASK collects those requirements.
790                  */
791                 err = add_to_page_cache_lru(page, mapping, index,
792                         (gfp_mask & GFP_RECLAIM_MASK));
793                 if (unlikely(err)) {
794                         page_cache_release(page);
795                         page = NULL;
796                         if (err == -EEXIST)
797                                 goto repeat;
798                 }
799         }
800         return page;
801 }
802 EXPORT_SYMBOL(find_or_create_page);
803
804 /**
805  * find_get_pages - gang pagecache lookup
806  * @mapping:    The address_space to search
807  * @start:      The starting page index
808  * @nr_pages:   The maximum number of pages
809  * @pages:      Where the resulting pages are placed
810  *
811  * find_get_pages() will search for and return a group of up to
812  * @nr_pages pages in the mapping.  The pages are placed at @pages.
813  * find_get_pages() takes a reference against the returned pages.
814  *
815  * The search returns a group of mapping-contiguous pages with ascending
816  * indexes.  There may be holes in the indices due to not-present pages.
817  *
818  * find_get_pages() returns the number of pages which were found.
819  */
820 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
821                             unsigned int nr_pages, struct page **pages)
822 {
823         unsigned int i;
824         unsigned int ret;
825         unsigned int nr_found;
826
827         rcu_read_lock();
828 restart:
829         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
830                                 (void ***)pages, start, nr_pages);
831         ret = 0;
832         for (i = 0; i < nr_found; i++) {
833                 struct page *page;
834 repeat:
835                 page = radix_tree_deref_slot((void **)pages[i]);
836                 if (unlikely(!page))
837                         continue;
838                 /*
839                  * this can only trigger if nr_found == 1, making livelock
840                  * a non issue.
841                  */
842                 if (unlikely(page == RADIX_TREE_RETRY))
843                         goto restart;
844
845                 if (!page_cache_get_speculative(page))
846                         goto repeat;
847
848                 /* Has the page moved? */
849                 if (unlikely(page != *((void **)pages[i]))) {
850                         page_cache_release(page);
851                         goto repeat;
852                 }
853
854                 pages[ret] = page;
855                 ret++;
856         }
857         rcu_read_unlock();
858         return ret;
859 }
860
861 /**
862  * find_get_pages_contig - gang contiguous pagecache lookup
863  * @mapping:    The address_space to search
864  * @index:      The starting page index
865  * @nr_pages:   The maximum number of pages
866  * @pages:      Where the resulting pages are placed
867  *
868  * find_get_pages_contig() works exactly like find_get_pages(), except
869  * that the returned number of pages are guaranteed to be contiguous.
870  *
871  * find_get_pages_contig() returns the number of pages which were found.
872  */
873 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
874                                unsigned int nr_pages, struct page **pages)
875 {
876         unsigned int i;
877         unsigned int ret;
878         unsigned int nr_found;
879
880         rcu_read_lock();
881 restart:
882         nr_found = radix_tree_gang_lookup_slot(&mapping->page_tree,
883                                 (void ***)pages, index, nr_pages);
884         ret = 0;
885         for (i = 0; i < nr_found; i++) {
886                 struct page *page;
887 repeat:
888                 page = radix_tree_deref_slot((void **)pages[i]);
889                 if (unlikely(!page))
890                         continue;
891                 /*
892                  * this can only trigger if nr_found == 1, making livelock
893                  * a non issue.
894                  */
895                 if (unlikely(page == RADIX_TREE_RETRY))
896                         goto restart;
897
898                 if (page->mapping == NULL || page->index != index)
899                         break;
900
901                 if (!page_cache_get_speculative(page))
902                         goto repeat;
903
904                 /* Has the page moved? */
905                 if (unlikely(page != *((void **)pages[i]))) {
906                         page_cache_release(page);
907                         goto repeat;
908                 }
909
910                 pages[ret] = page;
911                 ret++;
912                 index++;
913         }
914         rcu_read_unlock();
915         return ret;
916 }
917 EXPORT_SYMBOL(find_get_pages_contig);
918
919 /**
920  * find_get_pages_tag - find and return pages that match @tag
921  * @mapping:    the address_space to search
922  * @index:      the starting page index
923  * @tag:        the tag index
924  * @nr_pages:   the maximum number of pages
925  * @pages:      where the resulting pages are placed
926  *
927  * Like find_get_pages, except we only return pages which are tagged with
928  * @tag.   We update @index to index the next page for the traversal.
929  */
930 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
931                         int tag, unsigned int nr_pages, struct page **pages)
932 {
933         unsigned int i;
934         unsigned int ret;
935         unsigned int nr_found;
936
937         rcu_read_lock();
938 restart:
939         nr_found = radix_tree_gang_lookup_tag_slot(&mapping->page_tree,
940                                 (void ***)pages, *index, nr_pages, tag);
941         ret = 0;
942         for (i = 0; i < nr_found; i++) {
943                 struct page *page;
944 repeat:
945                 page = radix_tree_deref_slot((void **)pages[i]);
946                 if (unlikely(!page))
947                         continue;
948                 /*
949                  * this can only trigger if nr_found == 1, making livelock
950                  * a non issue.
951                  */
952                 if (unlikely(page == RADIX_TREE_RETRY))
953                         goto restart;
954
955                 if (!page_cache_get_speculative(page))
956                         goto repeat;
957
958                 /* Has the page moved? */
959                 if (unlikely(page != *((void **)pages[i]))) {
960                         page_cache_release(page);
961                         goto repeat;
962                 }
963
964                 pages[ret] = page;
965                 ret++;
966         }
967         rcu_read_unlock();
968
969         if (ret)
970                 *index = pages[ret - 1]->index + 1;
971
972         return ret;
973 }
974 EXPORT_SYMBOL(find_get_pages_tag);
975
976 /**
977  * grab_cache_page_nowait - returns locked page at given index in given cache
978  * @mapping: target address_space
979  * @index: the page index
980  *
981  * Same as grab_cache_page(), but do not wait if the page is unavailable.
982  * This is intended for speculative data generators, where the data can
983  * be regenerated if the page couldn't be grabbed.  This routine should
984  * be safe to call while holding the lock for another page.
985  *
986  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
987  * and deadlock against the caller's locked page.
988  */
989 struct page *
990 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
991 {
992         struct page *page = find_get_page(mapping, index);
993
994         if (page) {
995                 if (trylock_page(page))
996                         return page;
997                 page_cache_release(page);
998                 return NULL;
999         }
1000         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1001         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1002                 page_cache_release(page);
1003                 page = NULL;
1004         }
1005         return page;
1006 }
1007 EXPORT_SYMBOL(grab_cache_page_nowait);
1008
1009 /*
1010  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1011  * a _large_ part of the i/o request. Imagine the worst scenario:
1012  *
1013  *      ---R__________________________________________B__________
1014  *         ^ reading here                             ^ bad block(assume 4k)
1015  *
1016  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1017  * => failing the whole request => read(R) => read(R+1) =>
1018  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1019  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1020  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1021  *
1022  * It is going insane. Fix it by quickly scaling down the readahead size.
1023  */
1024 static void shrink_readahead_size_eio(struct file *filp,
1025                                         struct file_ra_state *ra)
1026 {
1027         ra->ra_pages /= 4;
1028 }
1029
1030 /**
1031  * do_generic_file_read - generic file read routine
1032  * @filp:       the file to read
1033  * @ppos:       current file position
1034  * @desc:       read_descriptor
1035  * @actor:      read method
1036  *
1037  * This is a generic file read routine, and uses the
1038  * mapping->a_ops->readpage() function for the actual low-level stuff.
1039  *
1040  * This is really ugly. But the goto's actually try to clarify some
1041  * of the logic when it comes to error handling etc.
1042  */
1043 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1044                 read_descriptor_t *desc, read_actor_t actor)
1045 {
1046         struct address_space *mapping = filp->f_mapping;
1047         struct inode *inode = mapping->host;
1048         struct file_ra_state *ra = &filp->f_ra;
1049         pgoff_t index;
1050         pgoff_t last_index;
1051         pgoff_t prev_index;
1052         unsigned long offset;      /* offset into pagecache page */
1053         unsigned int prev_offset;
1054         int error;
1055
1056         index = *ppos >> PAGE_CACHE_SHIFT;
1057         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1058         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1059         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1060         offset = *ppos & ~PAGE_CACHE_MASK;
1061
1062         for (;;) {
1063                 struct page *page;
1064                 pgoff_t end_index;
1065                 loff_t isize;
1066                 unsigned long nr, ret;
1067
1068                 cond_resched();
1069 find_page:
1070                 page = find_get_page(mapping, index);
1071                 if (!page) {
1072                         page_cache_sync_readahead(mapping,
1073                                         ra, filp,
1074                                         index, last_index - index);
1075                         page = find_get_page(mapping, index);
1076                         if (unlikely(page == NULL))
1077                                 goto no_cached_page;
1078                 }
1079                 if (PageReadahead(page)) {
1080                         page_cache_async_readahead(mapping,
1081                                         ra, filp, page,
1082                                         index, last_index - index);
1083                 }
1084                 if (!PageUptodate(page)) {
1085                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1086                                         !mapping->a_ops->is_partially_uptodate)
1087                                 goto page_not_up_to_date;
1088                         if (!trylock_page(page))
1089                                 goto page_not_up_to_date;
1090                         if (!mapping->a_ops->is_partially_uptodate(page,
1091                                                                 desc, offset))
1092                                 goto page_not_up_to_date_locked;
1093                         unlock_page(page);
1094                 }
1095 page_ok:
1096                 /*
1097                  * i_size must be checked after we know the page is Uptodate.
1098                  *
1099                  * Checking i_size after the check allows us to calculate
1100                  * the correct value for "nr", which means the zero-filled
1101                  * part of the page is not copied back to userspace (unless
1102                  * another truncate extends the file - this is desired though).
1103                  */
1104
1105                 isize = i_size_read(inode);
1106                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1107                 if (unlikely(!isize || index > end_index)) {
1108                         page_cache_release(page);
1109                         goto out;
1110                 }
1111
1112                 /* nr is the maximum number of bytes to copy from this page */
1113                 nr = PAGE_CACHE_SIZE;
1114                 if (index == end_index) {
1115                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1116                         if (nr <= offset) {
1117                                 page_cache_release(page);
1118                                 goto out;
1119                         }
1120                 }
1121                 nr = nr - offset;
1122
1123                 /* If users can be writing to this page using arbitrary
1124                  * virtual addresses, take care about potential aliasing
1125                  * before reading the page on the kernel side.
1126                  */
1127                 if (mapping_writably_mapped(mapping))
1128                         flush_dcache_page(page);
1129
1130                 /*
1131                  * When a sequential read accesses a page several times,
1132                  * only mark it as accessed the first time.
1133                  */
1134                 if (prev_index != index || offset != prev_offset)
1135                         mark_page_accessed(page);
1136                 prev_index = index;
1137
1138                 /*
1139                  * Ok, we have the page, and it's up-to-date, so
1140                  * now we can copy it to user space...
1141                  *
1142                  * The actor routine returns how many bytes were actually used..
1143                  * NOTE! This may not be the same as how much of a user buffer
1144                  * we filled up (we may be padding etc), so we can only update
1145                  * "pos" here (the actor routine has to update the user buffer
1146                  * pointers and the remaining count).
1147                  */
1148                 ret = actor(desc, page, offset, nr);
1149                 offset += ret;
1150                 index += offset >> PAGE_CACHE_SHIFT;
1151                 offset &= ~PAGE_CACHE_MASK;
1152                 prev_offset = offset;
1153
1154                 page_cache_release(page);
1155                 if (ret == nr && desc->count)
1156                         continue;
1157                 goto out;
1158
1159 page_not_up_to_date:
1160                 /* Get exclusive access to the page ... */
1161                 error = lock_page_killable(page);
1162                 if (unlikely(error))
1163                         goto readpage_error;
1164
1165 page_not_up_to_date_locked:
1166                 /* Did it get truncated before we got the lock? */
1167                 if (!page->mapping) {
1168                         unlock_page(page);
1169                         page_cache_release(page);
1170                         continue;
1171                 }
1172
1173                 /* Did somebody else fill it already? */
1174                 if (PageUptodate(page)) {
1175                         unlock_page(page);
1176                         goto page_ok;
1177                 }
1178
1179 readpage:
1180                 /* Start the actual read. The read will unlock the page. */
1181                 error = mapping->a_ops->readpage(filp, page);
1182
1183                 if (unlikely(error)) {
1184                         if (error == AOP_TRUNCATED_PAGE) {
1185                                 page_cache_release(page);
1186                                 goto find_page;
1187                         }
1188                         goto readpage_error;
1189                 }
1190
1191                 if (!PageUptodate(page)) {
1192                         error = lock_page_killable(page);
1193                         if (unlikely(error))
1194                                 goto readpage_error;
1195                         if (!PageUptodate(page)) {
1196                                 if (page->mapping == NULL) {
1197                                         /*
1198                                          * invalidate_inode_pages got it
1199                                          */
1200                                         unlock_page(page);
1201                                         page_cache_release(page);
1202                                         goto find_page;
1203                                 }
1204                                 unlock_page(page);
1205                                 shrink_readahead_size_eio(filp, ra);
1206                                 error = -EIO;
1207                                 goto readpage_error;
1208                         }
1209                         unlock_page(page);
1210                 }
1211
1212                 goto page_ok;
1213
1214 readpage_error:
1215                 /* UHHUH! A synchronous read error occurred. Report it */
1216                 desc->error = error;
1217                 page_cache_release(page);
1218                 goto out;
1219
1220 no_cached_page:
1221                 /*
1222                  * Ok, it wasn't cached, so we need to create a new
1223                  * page..
1224                  */
1225                 page = page_cache_alloc_cold(mapping);
1226                 if (!page) {
1227                         desc->error = -ENOMEM;
1228                         goto out;
1229                 }
1230                 error = add_to_page_cache_lru(page, mapping,
1231                                                 index, GFP_KERNEL);
1232                 if (error) {
1233                         page_cache_release(page);
1234                         if (error == -EEXIST)
1235                                 goto find_page;
1236                         desc->error = error;
1237                         goto out;
1238                 }
1239                 goto readpage;
1240         }
1241
1242 out:
1243         ra->prev_pos = prev_index;
1244         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1245         ra->prev_pos |= prev_offset;
1246
1247         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1248         file_accessed(filp);
1249 }
1250
1251 int file_read_actor(read_descriptor_t *desc, struct page *page,
1252                         unsigned long offset, unsigned long size)
1253 {
1254         char *kaddr;
1255         unsigned long left, count = desc->count;
1256
1257         if (size > count)
1258                 size = count;
1259
1260         /*
1261          * Faults on the destination of a read are common, so do it before
1262          * taking the kmap.
1263          */
1264         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1265                 kaddr = kmap_atomic(page, KM_USER0);
1266                 left = __copy_to_user_inatomic(desc->arg.buf,
1267                                                 kaddr + offset, size);
1268                 kunmap_atomic(kaddr, KM_USER0);
1269                 if (left == 0)
1270                         goto success;
1271         }
1272
1273         /* Do it the slow way */
1274         kaddr = kmap(page);
1275         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1276         kunmap(page);
1277
1278         if (left) {
1279                 size -= left;
1280                 desc->error = -EFAULT;
1281         }
1282 success:
1283         desc->count = count - size;
1284         desc->written += size;
1285         desc->arg.buf += size;
1286         return size;
1287 }
1288
1289 /*
1290  * Performs necessary checks before doing a write
1291  * @iov:        io vector request
1292  * @nr_segs:    number of segments in the iovec
1293  * @count:      number of bytes to write
1294  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1295  *
1296  * Adjust number of segments and amount of bytes to write (nr_segs should be
1297  * properly initialized first). Returns appropriate error code that caller
1298  * should return or zero in case that write should be allowed.
1299  */
1300 int generic_segment_checks(const struct iovec *iov,
1301                         unsigned long *nr_segs, size_t *count, int access_flags)
1302 {
1303         unsigned long   seg;
1304         size_t cnt = 0;
1305         for (seg = 0; seg < *nr_segs; seg++) {
1306                 const struct iovec *iv = &iov[seg];
1307
1308                 /*
1309                  * If any segment has a negative length, or the cumulative
1310                  * length ever wraps negative then return -EINVAL.
1311                  */
1312                 cnt += iv->iov_len;
1313                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1314                         return -EINVAL;
1315                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1316                         continue;
1317                 if (seg == 0)
1318                         return -EFAULT;
1319                 *nr_segs = seg;
1320                 cnt -= iv->iov_len;     /* This segment is no good */
1321                 break;
1322         }
1323         *count = cnt;
1324         return 0;
1325 }
1326 EXPORT_SYMBOL(generic_segment_checks);
1327
1328 /**
1329  * generic_file_aio_read - generic filesystem read routine
1330  * @iocb:       kernel I/O control block
1331  * @iov:        io vector request
1332  * @nr_segs:    number of segments in the iovec
1333  * @pos:        current file position
1334  *
1335  * This is the "read()" routine for all filesystems
1336  * that can use the page cache directly.
1337  */
1338 ssize_t
1339 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1340                 unsigned long nr_segs, loff_t pos)
1341 {
1342         struct file *filp = iocb->ki_filp;
1343         ssize_t retval;
1344         unsigned long seg;
1345         size_t count;
1346         loff_t *ppos = &iocb->ki_pos;
1347
1348         count = 0;
1349         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1350         if (retval)
1351                 return retval;
1352
1353         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1354         if (filp->f_flags & O_DIRECT) {
1355                 loff_t size;
1356                 struct address_space *mapping;
1357                 struct inode *inode;
1358
1359                 mapping = filp->f_mapping;
1360                 inode = mapping->host;
1361                 if (!count)
1362                         goto out; /* skip atime */
1363                 size = i_size_read(inode);
1364                 if (pos < size) {
1365                         retval = filemap_write_and_wait_range(mapping, pos,
1366                                         pos + iov_length(iov, nr_segs) - 1);
1367                         if (!retval) {
1368                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1369                                                         iov, pos, nr_segs);
1370                         }
1371                         if (retval > 0)
1372                                 *ppos = pos + retval;
1373                         if (retval) {
1374                                 file_accessed(filp);
1375                                 goto out;
1376                         }
1377                 }
1378         }
1379
1380         for (seg = 0; seg < nr_segs; seg++) {
1381                 read_descriptor_t desc;
1382
1383                 desc.written = 0;
1384                 desc.arg.buf = iov[seg].iov_base;
1385                 desc.count = iov[seg].iov_len;
1386                 if (desc.count == 0)
1387                         continue;
1388                 desc.error = 0;
1389                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1390                 retval += desc.written;
1391                 if (desc.error) {
1392                         retval = retval ?: desc.error;
1393                         break;
1394                 }
1395                 if (desc.count > 0)
1396                         break;
1397         }
1398 out:
1399         return retval;
1400 }
1401 EXPORT_SYMBOL(generic_file_aio_read);
1402
1403 static ssize_t
1404 do_readahead(struct address_space *mapping, struct file *filp,
1405              pgoff_t index, unsigned long nr)
1406 {
1407         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1408                 return -EINVAL;
1409
1410         force_page_cache_readahead(mapping, filp, index, nr);
1411         return 0;
1412 }
1413
1414 SYSCALL_DEFINE(readahead)(int fd, loff_t offset, size_t count)
1415 {
1416         ssize_t ret;
1417         struct file *file;
1418
1419         ret = -EBADF;
1420         file = fget(fd);
1421         if (file) {
1422                 if (file->f_mode & FMODE_READ) {
1423                         struct address_space *mapping = file->f_mapping;
1424                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1425                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1426                         unsigned long len = end - start + 1;
1427                         ret = do_readahead(mapping, file, start, len);
1428                 }
1429                 fput(file);
1430         }
1431         return ret;
1432 }
1433 #ifdef CONFIG_HAVE_SYSCALL_WRAPPERS
1434 asmlinkage long SyS_readahead(long fd, loff_t offset, long count)
1435 {
1436         return SYSC_readahead((int) fd, offset, (size_t) count);
1437 }
1438 SYSCALL_ALIAS(sys_readahead, SyS_readahead);
1439 #endif
1440
1441 #ifdef CONFIG_MMU
1442 /**
1443  * page_cache_read - adds requested page to the page cache if not already there
1444  * @file:       file to read
1445  * @offset:     page index
1446  *
1447  * This adds the requested page to the page cache if it isn't already there,
1448  * and schedules an I/O to read in its contents from disk.
1449  */
1450 static int page_cache_read(struct file *file, pgoff_t offset)
1451 {
1452         struct address_space *mapping = file->f_mapping;
1453         struct page *page; 
1454         int ret;
1455
1456         do {
1457                 page = page_cache_alloc_cold(mapping);
1458                 if (!page)
1459                         return -ENOMEM;
1460
1461                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1462                 if (ret == 0)
1463                         ret = mapping->a_ops->readpage(file, page);
1464                 else if (ret == -EEXIST)
1465                         ret = 0; /* losing race to add is OK */
1466
1467                 page_cache_release(page);
1468
1469         } while (ret == AOP_TRUNCATED_PAGE);
1470                 
1471         return ret;
1472 }
1473
1474 #define MMAP_LOTSAMISS  (100)
1475
1476 /*
1477  * Synchronous readahead happens when we don't even find
1478  * a page in the page cache at all.
1479  */
1480 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1481                                    struct file_ra_state *ra,
1482                                    struct file *file,
1483                                    pgoff_t offset)
1484 {
1485         unsigned long ra_pages;
1486         struct address_space *mapping = file->f_mapping;
1487
1488         /* If we don't want any read-ahead, don't bother */
1489         if (VM_RandomReadHint(vma))
1490                 return;
1491
1492         if (VM_SequentialReadHint(vma) ||
1493                         offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
1494                 page_cache_sync_readahead(mapping, ra, file, offset,
1495                                           ra->ra_pages);
1496                 return;
1497         }
1498
1499         if (ra->mmap_miss < INT_MAX)
1500                 ra->mmap_miss++;
1501
1502         /*
1503          * Do we miss much more than hit in this file? If so,
1504          * stop bothering with read-ahead. It will only hurt.
1505          */
1506         if (ra->mmap_miss > MMAP_LOTSAMISS)
1507                 return;
1508
1509         /*
1510          * mmap read-around
1511          */
1512         ra_pages = max_sane_readahead(ra->ra_pages);
1513         if (ra_pages) {
1514                 ra->start = max_t(long, 0, offset - ra_pages/2);
1515                 ra->size = ra_pages;
1516                 ra->async_size = 0;
1517                 ra_submit(ra, mapping, file);
1518         }
1519 }
1520
1521 /*
1522  * Asynchronous readahead happens when we find the page and PG_readahead,
1523  * so we want to possibly extend the readahead further..
1524  */
1525 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1526                                     struct file_ra_state *ra,
1527                                     struct file *file,
1528                                     struct page *page,
1529                                     pgoff_t offset)
1530 {
1531         struct address_space *mapping = file->f_mapping;
1532
1533         /* If we don't want any read-ahead, don't bother */
1534         if (VM_RandomReadHint(vma))
1535                 return;
1536         if (ra->mmap_miss > 0)
1537                 ra->mmap_miss--;
1538         if (PageReadahead(page))
1539                 page_cache_async_readahead(mapping, ra, file,
1540                                            page, offset, ra->ra_pages);
1541 }
1542
1543 /**
1544  * filemap_fault - read in file data for page fault handling
1545  * @vma:        vma in which the fault was taken
1546  * @vmf:        struct vm_fault containing details of the fault
1547  *
1548  * filemap_fault() is invoked via the vma operations vector for a
1549  * mapped memory region to read in file data during a page fault.
1550  *
1551  * The goto's are kind of ugly, but this streamlines the normal case of having
1552  * it in the page cache, and handles the special cases reasonably without
1553  * having a lot of duplicated code.
1554  */
1555 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1556 {
1557         int error;
1558         struct file *file = vma->vm_file;
1559         struct address_space *mapping = file->f_mapping;
1560         struct file_ra_state *ra = &file->f_ra;
1561         struct inode *inode = mapping->host;
1562         pgoff_t offset = vmf->pgoff;
1563         struct page *page;
1564         pgoff_t size;
1565         int ret = 0;
1566
1567         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1568         if (offset >= size)
1569                 return VM_FAULT_SIGBUS;
1570
1571         /*
1572          * Do we have something in the page cache already?
1573          */
1574         page = find_get_page(mapping, offset);
1575         if (likely(page)) {
1576                 /*
1577                  * We found the page, so try async readahead before
1578                  * waiting for the lock.
1579                  */
1580                 do_async_mmap_readahead(vma, ra, file, page, offset);
1581                 lock_page(page);
1582
1583                 /* Did it get truncated? */
1584                 if (unlikely(page->mapping != mapping)) {
1585                         unlock_page(page);
1586                         put_page(page);
1587                         goto no_cached_page;
1588                 }
1589         } else {
1590                 /* No page in the page cache at all */
1591                 do_sync_mmap_readahead(vma, ra, file, offset);
1592                 count_vm_event(PGMAJFAULT);
1593                 ret = VM_FAULT_MAJOR;
1594 retry_find:
1595                 page = find_lock_page(mapping, offset);
1596                 if (!page)
1597                         goto no_cached_page;
1598         }
1599
1600         /*
1601          * We have a locked page in the page cache, now we need to check
1602          * that it's up-to-date. If not, it is going to be due to an error.
1603          */
1604         if (unlikely(!PageUptodate(page)))
1605                 goto page_not_uptodate;
1606
1607         /*
1608          * Found the page and have a reference on it.
1609          * We must recheck i_size under page lock.
1610          */
1611         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1612         if (unlikely(offset >= size)) {
1613                 unlock_page(page);
1614                 page_cache_release(page);
1615                 return VM_FAULT_SIGBUS;
1616         }
1617
1618         ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
1619         vmf->page = page;
1620         return ret | VM_FAULT_LOCKED;
1621
1622 no_cached_page:
1623         /*
1624          * We're only likely to ever get here if MADV_RANDOM is in
1625          * effect.
1626          */
1627         error = page_cache_read(file, offset);
1628
1629         /*
1630          * The page we want has now been added to the page cache.
1631          * In the unlikely event that someone removed it in the
1632          * meantime, we'll just come back here and read it again.
1633          */
1634         if (error >= 0)
1635                 goto retry_find;
1636
1637         /*
1638          * An error return from page_cache_read can result if the
1639          * system is low on memory, or a problem occurs while trying
1640          * to schedule I/O.
1641          */
1642         if (error == -ENOMEM)
1643                 return VM_FAULT_OOM;
1644         return VM_FAULT_SIGBUS;
1645
1646 page_not_uptodate:
1647         /*
1648          * Umm, take care of errors if the page isn't up-to-date.
1649          * Try to re-read it _once_. We do this synchronously,
1650          * because there really aren't any performance issues here
1651          * and we need to check for errors.
1652          */
1653         ClearPageError(page);
1654         error = mapping->a_ops->readpage(file, page);
1655         if (!error) {
1656                 wait_on_page_locked(page);
1657                 if (!PageUptodate(page))
1658                         error = -EIO;
1659         }
1660         page_cache_release(page);
1661
1662         if (!error || error == AOP_TRUNCATED_PAGE)
1663                 goto retry_find;
1664
1665         /* Things didn't work out. Return zero to tell the mm layer so. */
1666         shrink_readahead_size_eio(file, ra);
1667         return VM_FAULT_SIGBUS;
1668 }
1669 EXPORT_SYMBOL(filemap_fault);
1670
1671 struct vm_operations_struct generic_file_vm_ops = {
1672         .fault          = filemap_fault,
1673 };
1674
1675 /* This is used for a general mmap of a disk file */
1676
1677 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1678 {
1679         struct address_space *mapping = file->f_mapping;
1680
1681         if (!mapping->a_ops->readpage)
1682                 return -ENOEXEC;
1683         file_accessed(file);
1684         vma->vm_ops = &generic_file_vm_ops;
1685         vma->vm_flags |= VM_CAN_NONLINEAR;
1686         return 0;
1687 }
1688
1689 /*
1690  * This is for filesystems which do not implement ->writepage.
1691  */
1692 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1693 {
1694         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1695                 return -EINVAL;
1696         return generic_file_mmap(file, vma);
1697 }
1698 #else
1699 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1700 {
1701         return -ENOSYS;
1702 }
1703 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1704 {
1705         return -ENOSYS;
1706 }
1707 #endif /* CONFIG_MMU */
1708
1709 EXPORT_SYMBOL(generic_file_mmap);
1710 EXPORT_SYMBOL(generic_file_readonly_mmap);
1711
1712 static struct page *__read_cache_page(struct address_space *mapping,
1713                                 pgoff_t index,
1714                                 int (*filler)(void *,struct page*),
1715                                 void *data)
1716 {
1717         struct page *page;
1718         int err;
1719 repeat:
1720         page = find_get_page(mapping, index);
1721         if (!page) {
1722                 page = page_cache_alloc_cold(mapping);
1723                 if (!page)
1724                         return ERR_PTR(-ENOMEM);
1725                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1726                 if (unlikely(err)) {
1727                         page_cache_release(page);
1728                         if (err == -EEXIST)
1729                                 goto repeat;
1730                         /* Presumably ENOMEM for radix tree node */
1731                         return ERR_PTR(err);
1732                 }
1733                 err = filler(data, page);
1734                 if (err < 0) {
1735                         page_cache_release(page);
1736                         page = ERR_PTR(err);
1737                 }
1738         }
1739         return page;
1740 }
1741
1742 /**
1743  * read_cache_page_async - read into page cache, fill it if needed
1744  * @mapping:    the page's address_space
1745  * @index:      the page index
1746  * @filler:     function to perform the read
1747  * @data:       destination for read data
1748  *
1749  * Same as read_cache_page, but don't wait for page to become unlocked
1750  * after submitting it to the filler.
1751  *
1752  * Read into the page cache. If a page already exists, and PageUptodate() is
1753  * not set, try to fill the page but don't wait for it to become unlocked.
1754  *
1755  * If the page does not get brought uptodate, return -EIO.
1756  */
1757 struct page *read_cache_page_async(struct address_space *mapping,
1758                                 pgoff_t index,
1759                                 int (*filler)(void *,struct page*),
1760                                 void *data)
1761 {
1762         struct page *page;
1763         int err;
1764
1765 retry:
1766         page = __read_cache_page(mapping, index, filler, data);
1767         if (IS_ERR(page))
1768                 return page;
1769         if (PageUptodate(page))
1770                 goto out;
1771
1772         lock_page(page);
1773         if (!page->mapping) {
1774                 unlock_page(page);
1775                 page_cache_release(page);
1776                 goto retry;
1777         }
1778         if (PageUptodate(page)) {
1779                 unlock_page(page);
1780                 goto out;
1781         }
1782         err = filler(data, page);
1783         if (err < 0) {
1784                 page_cache_release(page);
1785                 return ERR_PTR(err);
1786         }
1787 out:
1788         mark_page_accessed(page);
1789         return page;
1790 }
1791 EXPORT_SYMBOL(read_cache_page_async);
1792
1793 /**
1794  * read_cache_page - read into page cache, fill it if needed
1795  * @mapping:    the page's address_space
1796  * @index:      the page index
1797  * @filler:     function to perform the read
1798  * @data:       destination for read data
1799  *
1800  * Read into the page cache. If a page already exists, and PageUptodate() is
1801  * not set, try to fill the page then wait for it to become unlocked.
1802  *
1803  * If the page does not get brought uptodate, return -EIO.
1804  */
1805 struct page *read_cache_page(struct address_space *mapping,
1806                                 pgoff_t index,
1807                                 int (*filler)(void *,struct page*),
1808                                 void *data)
1809 {
1810         struct page *page;
1811
1812         page = read_cache_page_async(mapping, index, filler, data);
1813         if (IS_ERR(page))
1814                 goto out;
1815         wait_on_page_locked(page);
1816         if (!PageUptodate(page)) {
1817                 page_cache_release(page);
1818                 page = ERR_PTR(-EIO);
1819         }
1820  out:
1821         return page;
1822 }
1823 EXPORT_SYMBOL(read_cache_page);
1824
1825 /*
1826  * The logic we want is
1827  *
1828  *      if suid or (sgid and xgrp)
1829  *              remove privs
1830  */
1831 int should_remove_suid(struct dentry *dentry)
1832 {
1833         mode_t mode = dentry->d_inode->i_mode;
1834         int kill = 0;
1835
1836         /* suid always must be killed */
1837         if (unlikely(mode & S_ISUID))
1838                 kill = ATTR_KILL_SUID;
1839
1840         /*
1841          * sgid without any exec bits is just a mandatory locking mark; leave
1842          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1843          */
1844         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1845                 kill |= ATTR_KILL_SGID;
1846
1847         if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
1848                 return kill;
1849
1850         return 0;
1851 }
1852 EXPORT_SYMBOL(should_remove_suid);
1853
1854 static int __remove_suid(struct dentry *dentry, int kill)
1855 {
1856         struct iattr newattrs;
1857
1858         newattrs.ia_valid = ATTR_FORCE | kill;
1859         return notify_change(dentry, &newattrs);
1860 }
1861
1862 int file_remove_suid(struct file *file)
1863 {
1864         struct dentry *dentry = file->f_path.dentry;
1865         int killsuid = should_remove_suid(dentry);
1866         int killpriv = security_inode_need_killpriv(dentry);
1867         int error = 0;
1868
1869         if (killpriv < 0)
1870                 return killpriv;
1871         if (killpriv)
1872                 error = security_inode_killpriv(dentry);
1873         if (!error && killsuid)
1874                 error = __remove_suid(dentry, killsuid);
1875
1876         return error;
1877 }
1878 EXPORT_SYMBOL(file_remove_suid);
1879
1880 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1881                         const struct iovec *iov, size_t base, size_t bytes)
1882 {
1883         size_t copied = 0, left = 0;
1884
1885         while (bytes) {
1886                 char __user *buf = iov->iov_base + base;
1887                 int copy = min(bytes, iov->iov_len - base);
1888
1889                 base = 0;
1890                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1891                 copied += copy;
1892                 bytes -= copy;
1893                 vaddr += copy;
1894                 iov++;
1895
1896                 if (unlikely(left))
1897                         break;
1898         }
1899         return copied - left;
1900 }
1901
1902 /*
1903  * Copy as much as we can into the page and return the number of bytes which
1904  * were sucessfully copied.  If a fault is encountered then return the number of
1905  * bytes which were copied.
1906  */
1907 size_t iov_iter_copy_from_user_atomic(struct page *page,
1908                 struct iov_iter *i, unsigned long offset, size_t bytes)
1909 {
1910         char *kaddr;
1911         size_t copied;
1912
1913         BUG_ON(!in_atomic());
1914         kaddr = kmap_atomic(page, KM_USER0);
1915         if (likely(i->nr_segs == 1)) {
1916                 int left;
1917                 char __user *buf = i->iov->iov_base + i->iov_offset;
1918                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1919                 copied = bytes - left;
1920         } else {
1921                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1922                                                 i->iov, i->iov_offset, bytes);
1923         }
1924         kunmap_atomic(kaddr, KM_USER0);
1925
1926         return copied;
1927 }
1928 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1929
1930 /*
1931  * This has the same sideeffects and return value as
1932  * iov_iter_copy_from_user_atomic().
1933  * The difference is that it attempts to resolve faults.
1934  * Page must not be locked.
1935  */
1936 size_t iov_iter_copy_from_user(struct page *page,
1937                 struct iov_iter *i, unsigned long offset, size_t bytes)
1938 {
1939         char *kaddr;
1940         size_t copied;
1941
1942         kaddr = kmap(page);
1943         if (likely(i->nr_segs == 1)) {
1944                 int left;
1945                 char __user *buf = i->iov->iov_base + i->iov_offset;
1946                 left = __copy_from_user(kaddr + offset, buf, bytes);
1947                 copied = bytes - left;
1948         } else {
1949                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1950                                                 i->iov, i->iov_offset, bytes);
1951         }
1952         kunmap(page);
1953         return copied;
1954 }
1955 EXPORT_SYMBOL(iov_iter_copy_from_user);
1956
1957 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1958 {
1959         BUG_ON(i->count < bytes);
1960
1961         if (likely(i->nr_segs == 1)) {
1962                 i->iov_offset += bytes;
1963                 i->count -= bytes;
1964         } else {
1965                 const struct iovec *iov = i->iov;
1966                 size_t base = i->iov_offset;
1967
1968                 /*
1969                  * The !iov->iov_len check ensures we skip over unlikely
1970                  * zero-length segments (without overruning the iovec).
1971                  */
1972                 while (bytes || unlikely(i->count && !iov->iov_len)) {
1973                         int copy;
1974
1975                         copy = min(bytes, iov->iov_len - base);
1976                         BUG_ON(!i->count || i->count < copy);
1977                         i->count -= copy;
1978                         bytes -= copy;
1979                         base += copy;
1980                         if (iov->iov_len == base) {
1981                                 iov++;
1982                                 base = 0;
1983                         }
1984                 }
1985                 i->iov = iov;
1986                 i->iov_offset = base;
1987         }
1988 }
1989 EXPORT_SYMBOL(iov_iter_advance);
1990
1991 /*
1992  * Fault in the first iovec of the given iov_iter, to a maximum length
1993  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1994  * accessed (ie. because it is an invalid address).
1995  *
1996  * writev-intensive code may want this to prefault several iovecs -- that
1997  * would be possible (callers must not rely on the fact that _only_ the
1998  * first iovec will be faulted with the current implementation).
1999  */
2000 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2001 {
2002         char __user *buf = i->iov->iov_base + i->iov_offset;
2003         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2004         return fault_in_pages_readable(buf, bytes);
2005 }
2006 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2007
2008 /*
2009  * Return the count of just the current iov_iter segment.
2010  */
2011 size_t iov_iter_single_seg_count(struct iov_iter *i)
2012 {
2013         const struct iovec *iov = i->iov;
2014         if (i->nr_segs == 1)
2015                 return i->count;
2016         else
2017                 return min(i->count, iov->iov_len - i->iov_offset);
2018 }
2019 EXPORT_SYMBOL(iov_iter_single_seg_count);
2020
2021 /*
2022  * Performs necessary checks before doing a write
2023  *
2024  * Can adjust writing position or amount of bytes to write.
2025  * Returns appropriate error code that caller should return or
2026  * zero in case that write should be allowed.
2027  */
2028 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2029 {
2030         struct inode *inode = file->f_mapping->host;
2031         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
2032
2033         if (unlikely(*pos < 0))
2034                 return -EINVAL;
2035
2036         if (!isblk) {
2037                 /* FIXME: this is for backwards compatibility with 2.4 */
2038                 if (file->f_flags & O_APPEND)
2039                         *pos = i_size_read(inode);
2040
2041                 if (limit != RLIM_INFINITY) {
2042                         if (*pos >= limit) {
2043                                 send_sig(SIGXFSZ, current, 0);
2044                                 return -EFBIG;
2045                         }
2046                         if (*count > limit - (typeof(limit))*pos) {
2047                                 *count = limit - (typeof(limit))*pos;
2048                         }
2049                 }
2050         }
2051
2052         /*
2053          * LFS rule
2054          */
2055         if (unlikely(*pos + *count > MAX_NON_LFS &&
2056                                 !(file->f_flags & O_LARGEFILE))) {
2057                 if (*pos >= MAX_NON_LFS) {
2058                         return -EFBIG;
2059                 }
2060                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2061                         *count = MAX_NON_LFS - (unsigned long)*pos;
2062                 }
2063         }
2064
2065         /*
2066          * Are we about to exceed the fs block limit ?
2067          *
2068          * If we have written data it becomes a short write.  If we have
2069          * exceeded without writing data we send a signal and return EFBIG.
2070          * Linus frestrict idea will clean these up nicely..
2071          */
2072         if (likely(!isblk)) {
2073                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2074                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2075                                 return -EFBIG;
2076                         }
2077                         /* zero-length writes at ->s_maxbytes are OK */
2078                 }
2079
2080                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2081                         *count = inode->i_sb->s_maxbytes - *pos;
2082         } else {
2083 #ifdef CONFIG_BLOCK
2084                 loff_t isize;
2085                 if (bdev_read_only(I_BDEV(inode)))
2086                         return -EPERM;
2087                 isize = i_size_read(inode);
2088                 if (*pos >= isize) {
2089                         if (*count || *pos > isize)
2090                                 return -ENOSPC;
2091                 }
2092
2093                 if (*pos + *count > isize)
2094                         *count = isize - *pos;
2095 #else
2096                 return -EPERM;
2097 #endif
2098         }
2099         return 0;
2100 }
2101 EXPORT_SYMBOL(generic_write_checks);
2102
2103 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2104                                 loff_t pos, unsigned len, unsigned flags,
2105                                 struct page **pagep, void **fsdata)
2106 {
2107         const struct address_space_operations *aops = mapping->a_ops;
2108
2109         return aops->write_begin(file, mapping, pos, len, flags,
2110                                                         pagep, fsdata);
2111 }
2112 EXPORT_SYMBOL(pagecache_write_begin);
2113
2114 int pagecache_write_end(struct file *file, struct address_space *mapping,
2115                                 loff_t pos, unsigned len, unsigned copied,
2116                                 struct page *page, void *fsdata)
2117 {
2118         const struct address_space_operations *aops = mapping->a_ops;
2119
2120         mark_page_accessed(page);
2121         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2122 }
2123 EXPORT_SYMBOL(pagecache_write_end);
2124
2125 ssize_t
2126 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2127                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2128                 size_t count, size_t ocount)
2129 {
2130         struct file     *file = iocb->ki_filp;
2131         struct address_space *mapping = file->f_mapping;
2132         struct inode    *inode = mapping->host;
2133         ssize_t         written;
2134         size_t          write_len;
2135         pgoff_t         end;
2136
2137         if (count != ocount)
2138                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2139
2140         write_len = iov_length(iov, *nr_segs);
2141         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2142
2143         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2144         if (written)
2145                 goto out;
2146
2147         /*
2148          * After a write we want buffered reads to be sure to go to disk to get
2149          * the new data.  We invalidate clean cached page from the region we're
2150          * about to write.  We do this *before* the write so that we can return
2151          * without clobbering -EIOCBQUEUED from ->direct_IO().
2152          */
2153         if (mapping->nrpages) {
2154                 written = invalidate_inode_pages2_range(mapping,
2155                                         pos >> PAGE_CACHE_SHIFT, end);
2156                 /*
2157                  * If a page can not be invalidated, return 0 to fall back
2158                  * to buffered write.
2159                  */
2160                 if (written) {
2161                         if (written == -EBUSY)
2162                                 return 0;
2163                         goto out;
2164                 }
2165         }
2166
2167         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2168
2169         /*
2170          * Finally, try again to invalidate clean pages which might have been
2171          * cached by non-direct readahead, or faulted in by get_user_pages()
2172          * if the source of the write was an mmap'ed region of the file
2173          * we're writing.  Either one is a pretty crazy thing to do,
2174          * so we don't support it 100%.  If this invalidation
2175          * fails, tough, the write still worked...
2176          */
2177         if (mapping->nrpages) {
2178                 invalidate_inode_pages2_range(mapping,
2179                                               pos >> PAGE_CACHE_SHIFT, end);
2180         }
2181
2182         if (written > 0) {
2183                 loff_t end = pos + written;
2184                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2185                         i_size_write(inode,  end);
2186                         mark_inode_dirty(inode);
2187                 }
2188                 *ppos = end;
2189         }
2190 out:
2191         return written;
2192 }
2193 EXPORT_SYMBOL(generic_file_direct_write);
2194
2195 /*
2196  * Find or create a page at the given pagecache position. Return the locked
2197  * page. This function is specifically for buffered writes.
2198  */
2199 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2200                                         pgoff_t index, unsigned flags)
2201 {
2202         int status;
2203         struct page *page;
2204         gfp_t gfp_notmask = 0;
2205         if (flags & AOP_FLAG_NOFS)
2206                 gfp_notmask = __GFP_FS;
2207 repeat:
2208         page = find_lock_page(mapping, index);
2209         if (likely(page))
2210                 return page;
2211
2212         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
2213         if (!page)
2214                 return NULL;
2215         status = add_to_page_cache_lru(page, mapping, index,
2216                                                 GFP_KERNEL & ~gfp_notmask);
2217         if (unlikely(status)) {
2218                 page_cache_release(page);
2219                 if (status == -EEXIST)
2220                         goto repeat;
2221                 return NULL;
2222         }
2223         return page;
2224 }
2225 EXPORT_SYMBOL(grab_cache_page_write_begin);
2226
2227 static ssize_t generic_perform_write(struct file *file,
2228                                 struct iov_iter *i, loff_t pos)
2229 {
2230         struct address_space *mapping = file->f_mapping;
2231         const struct address_space_operations *a_ops = mapping->a_ops;
2232         long status = 0;
2233         ssize_t written = 0;
2234         unsigned int flags = 0;
2235
2236         /*
2237          * Copies from kernel address space cannot fail (NFSD is a big user).
2238          */
2239         if (segment_eq(get_fs(), KERNEL_DS))
2240                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2241
2242         do {
2243                 struct page *page;
2244                 pgoff_t index;          /* Pagecache index for current page */
2245                 unsigned long offset;   /* Offset into pagecache page */
2246                 unsigned long bytes;    /* Bytes to write to page */
2247                 size_t copied;          /* Bytes copied from user */
2248                 void *fsdata;
2249
2250                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2251                 index = pos >> PAGE_CACHE_SHIFT;
2252                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2253                                                 iov_iter_count(i));
2254
2255 again:
2256
2257                 /*
2258                  * Bring in the user page that we will copy from _first_.
2259                  * Otherwise there's a nasty deadlock on copying from the
2260                  * same page as we're writing to, without it being marked
2261                  * up-to-date.
2262                  *
2263                  * Not only is this an optimisation, but it is also required
2264                  * to check that the address is actually valid, when atomic
2265                  * usercopies are used, below.
2266                  */
2267                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2268                         status = -EFAULT;
2269                         break;
2270                 }
2271
2272                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2273                                                 &page, &fsdata);
2274                 if (unlikely(status))
2275                         break;
2276
2277                 pagefault_disable();
2278                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2279                 pagefault_enable();
2280                 flush_dcache_page(page);
2281
2282                 mark_page_accessed(page);
2283                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2284                                                 page, fsdata);
2285                 if (unlikely(status < 0))
2286                         break;
2287                 copied = status;
2288
2289                 cond_resched();
2290
2291                 iov_iter_advance(i, copied);
2292                 if (unlikely(copied == 0)) {
2293                         /*
2294                          * If we were unable to copy any data at all, we must
2295                          * fall back to a single segment length write.
2296                          *
2297                          * If we didn't fallback here, we could livelock
2298                          * because not all segments in the iov can be copied at
2299                          * once without a pagefault.
2300                          */
2301                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2302                                                 iov_iter_single_seg_count(i));
2303                         goto again;
2304                 }
2305                 pos += copied;
2306                 written += copied;
2307
2308                 balance_dirty_pages_ratelimited(mapping);
2309
2310         } while (iov_iter_count(i));
2311
2312         return written ? written : status;
2313 }
2314
2315 ssize_t
2316 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2317                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2318                 size_t count, ssize_t written)
2319 {
2320         struct file *file = iocb->ki_filp;
2321         struct address_space *mapping = file->f_mapping;
2322         ssize_t status;
2323         struct iov_iter i;
2324
2325         iov_iter_init(&i, iov, nr_segs, count, written);
2326         status = generic_perform_write(file, &i, pos);
2327
2328         if (likely(status >= 0)) {
2329                 written += status;
2330                 *ppos = pos + status;
2331         }
2332         
2333         /*
2334          * If we get here for O_DIRECT writes then we must have fallen through
2335          * to buffered writes (block instantiation inside i_size).  So we sync
2336          * the file data here, to try to honour O_DIRECT expectations.
2337          */
2338         if (unlikely(file->f_flags & O_DIRECT) && written)
2339                 status = filemap_write_and_wait_range(mapping,
2340                                         pos, pos + written - 1);
2341
2342         return written ? written : status;
2343 }
2344 EXPORT_SYMBOL(generic_file_buffered_write);
2345
2346 /**
2347  * __generic_file_aio_write - write data to a file
2348  * @iocb:       IO state structure (file, offset, etc.)
2349  * @iov:        vector with data to write
2350  * @nr_segs:    number of segments in the vector
2351  * @ppos:       position where to write
2352  *
2353  * This function does all the work needed for actually writing data to a
2354  * file. It does all basic checks, removes SUID from the file, updates
2355  * modification times and calls proper subroutines depending on whether we
2356  * do direct IO or a standard buffered write.
2357  *
2358  * It expects i_mutex to be grabbed unless we work on a block device or similar
2359  * object which does not need locking at all.
2360  *
2361  * This function does *not* take care of syncing data in case of O_SYNC write.
2362  * A caller has to handle it. This is mainly due to the fact that we want to
2363  * avoid syncing under i_mutex.
2364  */
2365 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2366                                  unsigned long nr_segs, loff_t *ppos)
2367 {
2368         struct file *file = iocb->ki_filp;
2369         struct address_space * mapping = file->f_mapping;
2370         size_t ocount;          /* original count */
2371         size_t count;           /* after file limit checks */
2372         struct inode    *inode = mapping->host;
2373         loff_t          pos;
2374         ssize_t         written;
2375         ssize_t         err;
2376
2377         ocount = 0;
2378         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2379         if (err)
2380                 return err;
2381
2382         count = ocount;
2383         pos = *ppos;
2384
2385         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2386
2387         /* We can write back this queue in page reclaim */
2388         current->backing_dev_info = mapping->backing_dev_info;
2389         written = 0;
2390
2391         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2392         if (err)
2393                 goto out;
2394
2395         if (count == 0)
2396                 goto out;
2397
2398         err = file_remove_suid(file);
2399         if (err)
2400                 goto out;
2401
2402         file_update_time(file);
2403
2404         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2405         if (unlikely(file->f_flags & O_DIRECT)) {
2406                 loff_t endbyte;
2407                 ssize_t written_buffered;
2408
2409                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2410                                                         ppos, count, ocount);
2411                 if (written < 0 || written == count)
2412                         goto out;
2413                 /*
2414                  * direct-io write to a hole: fall through to buffered I/O
2415                  * for completing the rest of the request.
2416                  */
2417                 pos += written;
2418                 count -= written;
2419                 written_buffered = generic_file_buffered_write(iocb, iov,
2420                                                 nr_segs, pos, ppos, count,
2421                                                 written);
2422                 /*
2423                  * If generic_file_buffered_write() retuned a synchronous error
2424                  * then we want to return the number of bytes which were
2425                  * direct-written, or the error code if that was zero.  Note
2426                  * that this differs from normal direct-io semantics, which
2427                  * will return -EFOO even if some bytes were written.
2428                  */
2429                 if (written_buffered < 0) {
2430                         err = written_buffered;
2431                         goto out;
2432                 }
2433
2434                 /*
2435                  * We need to ensure that the page cache pages are written to
2436                  * disk and invalidated to preserve the expected O_DIRECT
2437                  * semantics.
2438                  */
2439                 endbyte = pos + written_buffered - written - 1;
2440                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2441                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2442                                             SYNC_FILE_RANGE_WRITE|
2443                                             SYNC_FILE_RANGE_WAIT_AFTER);
2444                 if (err == 0) {
2445                         written = written_buffered;
2446                         invalidate_mapping_pages(mapping,
2447                                                  pos >> PAGE_CACHE_SHIFT,
2448                                                  endbyte >> PAGE_CACHE_SHIFT);
2449                 } else {
2450                         /*
2451                          * We don't know how much we wrote, so just return
2452                          * the number of bytes which were direct-written
2453                          */
2454                 }
2455         } else {
2456                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2457                                 pos, ppos, count, written);
2458         }
2459 out:
2460         current->backing_dev_info = NULL;
2461         return written ? written : err;
2462 }
2463 EXPORT_SYMBOL(__generic_file_aio_write);
2464
2465 /**
2466  * generic_file_aio_write - write data to a file
2467  * @iocb:       IO state structure
2468  * @iov:        vector with data to write
2469  * @nr_segs:    number of segments in the vector
2470  * @pos:        position in file where to write
2471  *
2472  * This is a wrapper around __generic_file_aio_write() to be used by most
2473  * filesystems. It takes care of syncing the file in case of O_SYNC file
2474  * and acquires i_mutex as needed.
2475  */
2476 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2477                 unsigned long nr_segs, loff_t pos)
2478 {
2479         struct file *file = iocb->ki_filp;
2480         struct address_space *mapping = file->f_mapping;
2481         struct inode *inode = mapping->host;
2482         ssize_t ret;
2483
2484         BUG_ON(iocb->ki_pos != pos);
2485
2486         mutex_lock(&inode->i_mutex);
2487         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2488         mutex_unlock(&inode->i_mutex);
2489
2490         if ((ret > 0 || ret == -EIOCBQUEUED) &&
2491             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2492                 ssize_t err;
2493
2494                 err = sync_page_range(inode, mapping, pos, ret);
2495                 if (err < 0 && ret > 0)
2496                         ret = err;
2497         }
2498         return ret;
2499 }
2500 EXPORT_SYMBOL(generic_file_aio_write);
2501
2502 /**
2503  * try_to_release_page() - release old fs-specific metadata on a page
2504  *
2505  * @page: the page which the kernel is trying to free
2506  * @gfp_mask: memory allocation flags (and I/O mode)
2507  *
2508  * The address_space is to try to release any data against the page
2509  * (presumably at page->private).  If the release was successful, return `1'.
2510  * Otherwise return zero.
2511  *
2512  * This may also be called if PG_fscache is set on a page, indicating that the
2513  * page is known to the local caching routines.
2514  *
2515  * The @gfp_mask argument specifies whether I/O may be performed to release
2516  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2517  *
2518  */
2519 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2520 {
2521         struct address_space * const mapping = page->mapping;
2522
2523         BUG_ON(!PageLocked(page));
2524         if (PageWriteback(page))
2525                 return 0;
2526
2527         if (mapping && mapping->a_ops->releasepage)
2528                 return mapping->a_ops->releasepage(page, gfp_mask);
2529         return try_to_free_buffers(page);
2530 }
2531
2532 EXPORT_SYMBOL(try_to_release_page);