V4L/DVB (7623): Scripts to maintain the CARDLIST file
[linux-2.6] / 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 "internal.h"
37
38 /*
39  * FIXME: remove all knowledge of the buffer layer from the core VM
40  */
41 #include <linux/buffer_head.h> /* for generic_osync_inode */
42
43 #include <asm/mman.h>
44
45 static ssize_t
46 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
47         loff_t offset, unsigned long nr_segs);
48
49 /*
50  * Shared mappings implemented 30.11.1994. It's not fully working yet,
51  * though.
52  *
53  * Shared mappings now work. 15.8.1995  Bruno.
54  *
55  * finished 'unifying' the page and buffer cache and SMP-threaded the
56  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
57  *
58  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
59  */
60
61 /*
62  * Lock ordering:
63  *
64  *  ->i_mmap_lock               (vmtruncate)
65  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
66  *      ->swap_lock             (exclusive_swap_page, others)
67  *        ->mapping->tree_lock
68  *
69  *  ->i_mutex
70  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
71  *
72  *  ->mmap_sem
73  *    ->i_mmap_lock
74  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
75  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
76  *
77  *  ->mmap_sem
78  *    ->lock_page               (access_process_vm)
79  *
80  *  ->i_mutex                   (generic_file_buffered_write)
81  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
82  *
83  *  ->i_mutex
84  *    ->i_alloc_sem             (various)
85  *
86  *  ->inode_lock
87  *    ->sb_lock                 (fs/fs-writeback.c)
88  *    ->mapping->tree_lock      (__sync_single_inode)
89  *
90  *  ->i_mmap_lock
91  *    ->anon_vma.lock           (vma_adjust)
92  *
93  *  ->anon_vma.lock
94  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
95  *
96  *  ->page_table_lock or pte_lock
97  *    ->swap_lock               (try_to_unmap_one)
98  *    ->private_lock            (try_to_unmap_one)
99  *    ->tree_lock               (try_to_unmap_one)
100  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
101  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
102  *    ->private_lock            (page_remove_rmap->set_page_dirty)
103  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
104  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (zap_pte_range->set_page_dirty)
106  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
107  *
108  *  ->task->proc_lock
109  *    ->dcache_lock             (proc_pid_lookup)
110  */
111
112 /*
113  * Remove a page from the page cache and free it. Caller has to make
114  * sure the page is locked and that nobody else uses it - or that usage
115  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
116  */
117 void __remove_from_page_cache(struct page *page)
118 {
119         struct address_space *mapping = page->mapping;
120
121         mem_cgroup_uncharge_page(page);
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         BUG_ON(page_mapped(page));
127
128         /*
129          * Some filesystems seem to re-dirty the page even after
130          * the VM has canceled the dirty bit (eg ext3 journaling).
131          *
132          * Fix it up by doing a final dirty accounting check after
133          * having removed the page entirely.
134          */
135         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
136                 dec_zone_page_state(page, NR_FILE_DIRTY);
137                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
138         }
139 }
140
141 void remove_from_page_cache(struct page *page)
142 {
143         struct address_space *mapping = page->mapping;
144
145         BUG_ON(!PageLocked(page));
146
147         write_lock_irq(&mapping->tree_lock);
148         __remove_from_page_cache(page);
149         write_unlock_irq(&mapping->tree_lock);
150 }
151
152 static int sync_page(void *word)
153 {
154         struct address_space *mapping;
155         struct page *page;
156
157         page = container_of((unsigned long *)word, struct page, flags);
158
159         /*
160          * page_mapping() is being called without PG_locked held.
161          * Some knowledge of the state and use of the page is used to
162          * reduce the requirements down to a memory barrier.
163          * The danger here is of a stale page_mapping() return value
164          * indicating a struct address_space different from the one it's
165          * associated with when it is associated with one.
166          * After smp_mb(), it's either the correct page_mapping() for
167          * the page, or an old page_mapping() and the page's own
168          * page_mapping() has gone NULL.
169          * The ->sync_page() address_space operation must tolerate
170          * page_mapping() going NULL. By an amazing coincidence,
171          * this comes about because none of the users of the page
172          * in the ->sync_page() methods make essential use of the
173          * page_mapping(), merely passing the page down to the backing
174          * device's unplug functions when it's non-NULL, which in turn
175          * ignore it for all cases but swap, where only page_private(page) is
176          * of interest. When page_mapping() does go NULL, the entire
177          * call stack gracefully ignores the page and returns.
178          * -- wli
179          */
180         smp_mb();
181         mapping = page_mapping(page);
182         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
183                 mapping->a_ops->sync_page(page);
184         io_schedule();
185         return 0;
186 }
187
188 static int sync_page_killable(void *word)
189 {
190         sync_page(word);
191         return fatal_signal_pending(current) ? -EINTR : 0;
192 }
193
194 /**
195  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
196  * @mapping:    address space structure to write
197  * @start:      offset in bytes where the range starts
198  * @end:        offset in bytes where the range ends (inclusive)
199  * @sync_mode:  enable synchronous operation
200  *
201  * Start writeback against all of a mapping's dirty pages that lie
202  * within the byte offsets <start, end> inclusive.
203  *
204  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
205  * opposed to a regular memory cleansing writeback.  The difference between
206  * these two operations is that if a dirty page/buffer is encountered, it must
207  * be waited upon, and not just skipped over.
208  */
209 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
210                                 loff_t end, int sync_mode)
211 {
212         int ret;
213         struct writeback_control wbc = {
214                 .sync_mode = sync_mode,
215                 .nr_to_write = mapping->nrpages * 2,
216                 .range_start = start,
217                 .range_end = end,
218         };
219
220         if (!mapping_cap_writeback_dirty(mapping))
221                 return 0;
222
223         ret = do_writepages(mapping, &wbc);
224         return ret;
225 }
226
227 static inline int __filemap_fdatawrite(struct address_space *mapping,
228         int sync_mode)
229 {
230         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
231 }
232
233 int filemap_fdatawrite(struct address_space *mapping)
234 {
235         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
236 }
237 EXPORT_SYMBOL(filemap_fdatawrite);
238
239 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
240                                 loff_t end)
241 {
242         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
243 }
244
245 /**
246  * filemap_flush - mostly a non-blocking flush
247  * @mapping:    target address_space
248  *
249  * This is a mostly non-blocking flush.  Not suitable for data-integrity
250  * purposes - I/O may not be started against all dirty pages.
251  */
252 int filemap_flush(struct address_space *mapping)
253 {
254         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
255 }
256 EXPORT_SYMBOL(filemap_flush);
257
258 /**
259  * wait_on_page_writeback_range - wait for writeback to complete
260  * @mapping:    target address_space
261  * @start:      beginning page index
262  * @end:        ending page index
263  *
264  * Wait for writeback to complete against pages indexed by start->end
265  * inclusive
266  */
267 int wait_on_page_writeback_range(struct address_space *mapping,
268                                 pgoff_t start, pgoff_t end)
269 {
270         struct pagevec pvec;
271         int nr_pages;
272         int ret = 0;
273         pgoff_t index;
274
275         if (end < start)
276                 return 0;
277
278         pagevec_init(&pvec, 0);
279         index = start;
280         while ((index <= end) &&
281                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
282                         PAGECACHE_TAG_WRITEBACK,
283                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
284                 unsigned i;
285
286                 for (i = 0; i < nr_pages; i++) {
287                         struct page *page = pvec.pages[i];
288
289                         /* until radix tree lookup accepts end_index */
290                         if (page->index > end)
291                                 continue;
292
293                         wait_on_page_writeback(page);
294                         if (PageError(page))
295                                 ret = -EIO;
296                 }
297                 pagevec_release(&pvec);
298                 cond_resched();
299         }
300
301         /* Check for outstanding write errors */
302         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
303                 ret = -ENOSPC;
304         if (test_and_clear_bit(AS_EIO, &mapping->flags))
305                 ret = -EIO;
306
307         return ret;
308 }
309
310 /**
311  * sync_page_range - write and wait on all pages in the passed range
312  * @inode:      target inode
313  * @mapping:    target address_space
314  * @pos:        beginning offset in pages to write
315  * @count:      number of bytes to write
316  *
317  * Write and wait upon all the pages in the passed range.  This is a "data
318  * integrity" operation.  It waits upon in-flight writeout before starting and
319  * waiting upon new writeout.  If there was an IO error, return it.
320  *
321  * We need to re-take i_mutex during the generic_osync_inode list walk because
322  * it is otherwise livelockable.
323  */
324 int sync_page_range(struct inode *inode, struct address_space *mapping,
325                         loff_t pos, loff_t count)
326 {
327         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
328         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
329         int ret;
330
331         if (!mapping_cap_writeback_dirty(mapping) || !count)
332                 return 0;
333         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
334         if (ret == 0) {
335                 mutex_lock(&inode->i_mutex);
336                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
337                 mutex_unlock(&inode->i_mutex);
338         }
339         if (ret == 0)
340                 ret = wait_on_page_writeback_range(mapping, start, end);
341         return ret;
342 }
343 EXPORT_SYMBOL(sync_page_range);
344
345 /**
346  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
347  * @inode:      target inode
348  * @mapping:    target address_space
349  * @pos:        beginning offset in pages to write
350  * @count:      number of bytes to write
351  *
352  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
353  * as it forces O_SYNC writers to different parts of the same file
354  * to be serialised right until io completion.
355  */
356 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
357                            loff_t pos, loff_t count)
358 {
359         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
360         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
361         int ret;
362
363         if (!mapping_cap_writeback_dirty(mapping) || !count)
364                 return 0;
365         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
366         if (ret == 0)
367                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
368         if (ret == 0)
369                 ret = wait_on_page_writeback_range(mapping, start, end);
370         return ret;
371 }
372 EXPORT_SYMBOL(sync_page_range_nolock);
373
374 /**
375  * filemap_fdatawait - wait for all under-writeback pages to complete
376  * @mapping: address space structure to wait for
377  *
378  * Walk the list of under-writeback pages of the given address space
379  * and wait for all of them.
380  */
381 int filemap_fdatawait(struct address_space *mapping)
382 {
383         loff_t i_size = i_size_read(mapping->host);
384
385         if (i_size == 0)
386                 return 0;
387
388         return wait_on_page_writeback_range(mapping, 0,
389                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
390 }
391 EXPORT_SYMBOL(filemap_fdatawait);
392
393 int filemap_write_and_wait(struct address_space *mapping)
394 {
395         int err = 0;
396
397         if (mapping->nrpages) {
398                 err = filemap_fdatawrite(mapping);
399                 /*
400                  * Even if the above returned error, the pages may be
401                  * written partially (e.g. -ENOSPC), so we wait for it.
402                  * But the -EIO is special case, it may indicate the worst
403                  * thing (e.g. bug) happened, so we avoid waiting for it.
404                  */
405                 if (err != -EIO) {
406                         int err2 = filemap_fdatawait(mapping);
407                         if (!err)
408                                 err = err2;
409                 }
410         }
411         return err;
412 }
413 EXPORT_SYMBOL(filemap_write_and_wait);
414
415 /**
416  * filemap_write_and_wait_range - write out & wait on a file range
417  * @mapping:    the address_space for the pages
418  * @lstart:     offset in bytes where the range starts
419  * @lend:       offset in bytes where the range ends (inclusive)
420  *
421  * Write out and wait upon file offsets lstart->lend, inclusive.
422  *
423  * Note that `lend' is inclusive (describes the last byte to be written) so
424  * that this function can be used to write to the very end-of-file (end = -1).
425  */
426 int filemap_write_and_wait_range(struct address_space *mapping,
427                                  loff_t lstart, loff_t lend)
428 {
429         int err = 0;
430
431         if (mapping->nrpages) {
432                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
433                                                  WB_SYNC_ALL);
434                 /* See comment of filemap_write_and_wait() */
435                 if (err != -EIO) {
436                         int err2 = wait_on_page_writeback_range(mapping,
437                                                 lstart >> PAGE_CACHE_SHIFT,
438                                                 lend >> PAGE_CACHE_SHIFT);
439                         if (!err)
440                                 err = err2;
441                 }
442         }
443         return err;
444 }
445
446 /**
447  * add_to_page_cache - add newly allocated pagecache pages
448  * @page:       page to add
449  * @mapping:    the page's address_space
450  * @offset:     page index
451  * @gfp_mask:   page allocation mode
452  *
453  * This function is used to add newly allocated pagecache pages;
454  * the page is new, so we can just run SetPageLocked() against it.
455  * The other page state flags were set by rmqueue().
456  *
457  * This function does not add the page to the LRU.  The caller must do that.
458  */
459 int add_to_page_cache(struct page *page, struct address_space *mapping,
460                 pgoff_t offset, gfp_t gfp_mask)
461 {
462         int error = mem_cgroup_cache_charge(page, current->mm,
463                                         gfp_mask & ~__GFP_HIGHMEM);
464         if (error)
465                 goto out;
466
467         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
468         if (error == 0) {
469                 write_lock_irq(&mapping->tree_lock);
470                 error = radix_tree_insert(&mapping->page_tree, offset, page);
471                 if (!error) {
472                         page_cache_get(page);
473                         SetPageLocked(page);
474                         page->mapping = mapping;
475                         page->index = offset;
476                         mapping->nrpages++;
477                         __inc_zone_page_state(page, NR_FILE_PAGES);
478                 } else
479                         mem_cgroup_uncharge_page(page);
480
481                 write_unlock_irq(&mapping->tree_lock);
482                 radix_tree_preload_end();
483         } else
484                 mem_cgroup_uncharge_page(page);
485 out:
486         return error;
487 }
488 EXPORT_SYMBOL(add_to_page_cache);
489
490 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
491                                 pgoff_t offset, gfp_t gfp_mask)
492 {
493         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
494         if (ret == 0)
495                 lru_cache_add(page);
496         return ret;
497 }
498
499 #ifdef CONFIG_NUMA
500 struct page *__page_cache_alloc(gfp_t gfp)
501 {
502         if (cpuset_do_page_mem_spread()) {
503                 int n = cpuset_mem_spread_node();
504                 return alloc_pages_node(n, gfp, 0);
505         }
506         return alloc_pages(gfp, 0);
507 }
508 EXPORT_SYMBOL(__page_cache_alloc);
509 #endif
510
511 static int __sleep_on_page_lock(void *word)
512 {
513         io_schedule();
514         return 0;
515 }
516
517 /*
518  * In order to wait for pages to become available there must be
519  * waitqueues associated with pages. By using a hash table of
520  * waitqueues where the bucket discipline is to maintain all
521  * waiters on the same queue and wake all when any of the pages
522  * become available, and for the woken contexts to check to be
523  * sure the appropriate page became available, this saves space
524  * at a cost of "thundering herd" phenomena during rare hash
525  * collisions.
526  */
527 static wait_queue_head_t *page_waitqueue(struct page *page)
528 {
529         const struct zone *zone = page_zone(page);
530
531         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
532 }
533
534 static inline void wake_up_page(struct page *page, int bit)
535 {
536         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
537 }
538
539 void wait_on_page_bit(struct page *page, int bit_nr)
540 {
541         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
542
543         if (test_bit(bit_nr, &page->flags))
544                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
545                                                         TASK_UNINTERRUPTIBLE);
546 }
547 EXPORT_SYMBOL(wait_on_page_bit);
548
549 /**
550  * unlock_page - unlock a locked page
551  * @page: the page
552  *
553  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
554  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
555  * mechananism between PageLocked pages and PageWriteback pages is shared.
556  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
557  *
558  * The first mb is necessary to safely close the critical section opened by the
559  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
560  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
561  * parallel wait_on_page_locked()).
562  */
563 void unlock_page(struct page *page)
564 {
565         smp_mb__before_clear_bit();
566         if (!TestClearPageLocked(page))
567                 BUG();
568         smp_mb__after_clear_bit(); 
569         wake_up_page(page, PG_locked);
570 }
571 EXPORT_SYMBOL(unlock_page);
572
573 /**
574  * end_page_writeback - end writeback against a page
575  * @page: the page
576  */
577 void end_page_writeback(struct page *page)
578 {
579         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
580                 if (!test_clear_page_writeback(page))
581                         BUG();
582         }
583         smp_mb__after_clear_bit();
584         wake_up_page(page, PG_writeback);
585 }
586 EXPORT_SYMBOL(end_page_writeback);
587
588 /**
589  * __lock_page - get a lock on the page, assuming we need to sleep to get it
590  * @page: the page to lock
591  *
592  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
593  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
594  * chances are that on the second loop, the block layer's plug list is empty,
595  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
596  */
597 void __lock_page(struct page *page)
598 {
599         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
600
601         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
602                                                         TASK_UNINTERRUPTIBLE);
603 }
604 EXPORT_SYMBOL(__lock_page);
605
606 int __lock_page_killable(struct page *page)
607 {
608         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
609
610         return __wait_on_bit_lock(page_waitqueue(page), &wait,
611                                         sync_page_killable, TASK_KILLABLE);
612 }
613
614 /**
615  * __lock_page_nosync - get a lock on the page, without calling sync_page()
616  * @page: the page to lock
617  *
618  * Variant of lock_page that does not require the caller to hold a reference
619  * on the page's mapping.
620  */
621 void __lock_page_nosync(struct page *page)
622 {
623         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
624         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
625                                                         TASK_UNINTERRUPTIBLE);
626 }
627
628 /**
629  * find_get_page - find and get a page reference
630  * @mapping: the address_space to search
631  * @offset: the page index
632  *
633  * Is there a pagecache struct page at the given (mapping, offset) tuple?
634  * If yes, increment its refcount and return it; if no, return NULL.
635  */
636 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
637 {
638         struct page *page;
639
640         read_lock_irq(&mapping->tree_lock);
641         page = radix_tree_lookup(&mapping->page_tree, offset);
642         if (page)
643                 page_cache_get(page);
644         read_unlock_irq(&mapping->tree_lock);
645         return page;
646 }
647 EXPORT_SYMBOL(find_get_page);
648
649 /**
650  * find_lock_page - locate, pin and lock a pagecache page
651  * @mapping: the address_space to search
652  * @offset: the page index
653  *
654  * Locates the desired pagecache page, locks it, increments its reference
655  * count and returns its address.
656  *
657  * Returns zero if the page was not present. find_lock_page() may sleep.
658  */
659 struct page *find_lock_page(struct address_space *mapping,
660                                 pgoff_t offset)
661 {
662         struct page *page;
663
664 repeat:
665         read_lock_irq(&mapping->tree_lock);
666         page = radix_tree_lookup(&mapping->page_tree, offset);
667         if (page) {
668                 page_cache_get(page);
669                 if (TestSetPageLocked(page)) {
670                         read_unlock_irq(&mapping->tree_lock);
671                         __lock_page(page);
672
673                         /* Has the page been truncated while we slept? */
674                         if (unlikely(page->mapping != mapping)) {
675                                 unlock_page(page);
676                                 page_cache_release(page);
677                                 goto repeat;
678                         }
679                         VM_BUG_ON(page->index != offset);
680                         goto out;
681                 }
682         }
683         read_unlock_irq(&mapping->tree_lock);
684 out:
685         return page;
686 }
687 EXPORT_SYMBOL(find_lock_page);
688
689 /**
690  * find_or_create_page - locate or add a pagecache page
691  * @mapping: the page's address_space
692  * @index: the page's index into the mapping
693  * @gfp_mask: page allocation mode
694  *
695  * Locates a page in the pagecache.  If the page is not present, a new page
696  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
697  * LRU list.  The returned page is locked and has its reference count
698  * incremented.
699  *
700  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
701  * allocation!
702  *
703  * find_or_create_page() returns the desired page's address, or zero on
704  * memory exhaustion.
705  */
706 struct page *find_or_create_page(struct address_space *mapping,
707                 pgoff_t index, gfp_t gfp_mask)
708 {
709         struct page *page;
710         int err;
711 repeat:
712         page = find_lock_page(mapping, index);
713         if (!page) {
714                 page = __page_cache_alloc(gfp_mask);
715                 if (!page)
716                         return NULL;
717                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
718                 if (unlikely(err)) {
719                         page_cache_release(page);
720                         page = NULL;
721                         if (err == -EEXIST)
722                                 goto repeat;
723                 }
724         }
725         return page;
726 }
727 EXPORT_SYMBOL(find_or_create_page);
728
729 /**
730  * find_get_pages - gang pagecache lookup
731  * @mapping:    The address_space to search
732  * @start:      The starting page index
733  * @nr_pages:   The maximum number of pages
734  * @pages:      Where the resulting pages are placed
735  *
736  * find_get_pages() will search for and return a group of up to
737  * @nr_pages pages in the mapping.  The pages are placed at @pages.
738  * find_get_pages() takes a reference against the returned pages.
739  *
740  * The search returns a group of mapping-contiguous pages with ascending
741  * indexes.  There may be holes in the indices due to not-present pages.
742  *
743  * find_get_pages() returns the number of pages which were found.
744  */
745 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
746                             unsigned int nr_pages, struct page **pages)
747 {
748         unsigned int i;
749         unsigned int ret;
750
751         read_lock_irq(&mapping->tree_lock);
752         ret = radix_tree_gang_lookup(&mapping->page_tree,
753                                 (void **)pages, start, nr_pages);
754         for (i = 0; i < ret; i++)
755                 page_cache_get(pages[i]);
756         read_unlock_irq(&mapping->tree_lock);
757         return ret;
758 }
759
760 /**
761  * find_get_pages_contig - gang contiguous pagecache lookup
762  * @mapping:    The address_space to search
763  * @index:      The starting page index
764  * @nr_pages:   The maximum number of pages
765  * @pages:      Where the resulting pages are placed
766  *
767  * find_get_pages_contig() works exactly like find_get_pages(), except
768  * that the returned number of pages are guaranteed to be contiguous.
769  *
770  * find_get_pages_contig() returns the number of pages which were found.
771  */
772 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
773                                unsigned int nr_pages, struct page **pages)
774 {
775         unsigned int i;
776         unsigned int ret;
777
778         read_lock_irq(&mapping->tree_lock);
779         ret = radix_tree_gang_lookup(&mapping->page_tree,
780                                 (void **)pages, index, nr_pages);
781         for (i = 0; i < ret; i++) {
782                 if (pages[i]->mapping == NULL || pages[i]->index != index)
783                         break;
784
785                 page_cache_get(pages[i]);
786                 index++;
787         }
788         read_unlock_irq(&mapping->tree_lock);
789         return i;
790 }
791 EXPORT_SYMBOL(find_get_pages_contig);
792
793 /**
794  * find_get_pages_tag - find and return pages that match @tag
795  * @mapping:    the address_space to search
796  * @index:      the starting page index
797  * @tag:        the tag index
798  * @nr_pages:   the maximum number of pages
799  * @pages:      where the resulting pages are placed
800  *
801  * Like find_get_pages, except we only return pages which are tagged with
802  * @tag.   We update @index to index the next page for the traversal.
803  */
804 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
805                         int tag, unsigned int nr_pages, struct page **pages)
806 {
807         unsigned int i;
808         unsigned int ret;
809
810         read_lock_irq(&mapping->tree_lock);
811         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
812                                 (void **)pages, *index, nr_pages, tag);
813         for (i = 0; i < ret; i++)
814                 page_cache_get(pages[i]);
815         if (ret)
816                 *index = pages[ret - 1]->index + 1;
817         read_unlock_irq(&mapping->tree_lock);
818         return ret;
819 }
820 EXPORT_SYMBOL(find_get_pages_tag);
821
822 /**
823  * grab_cache_page_nowait - returns locked page at given index in given cache
824  * @mapping: target address_space
825  * @index: the page index
826  *
827  * Same as grab_cache_page(), but do not wait if the page is unavailable.
828  * This is intended for speculative data generators, where the data can
829  * be regenerated if the page couldn't be grabbed.  This routine should
830  * be safe to call while holding the lock for another page.
831  *
832  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
833  * and deadlock against the caller's locked page.
834  */
835 struct page *
836 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
837 {
838         struct page *page = find_get_page(mapping, index);
839
840         if (page) {
841                 if (!TestSetPageLocked(page))
842                         return page;
843                 page_cache_release(page);
844                 return NULL;
845         }
846         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
847         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
848                 page_cache_release(page);
849                 page = NULL;
850         }
851         return page;
852 }
853 EXPORT_SYMBOL(grab_cache_page_nowait);
854
855 /*
856  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
857  * a _large_ part of the i/o request. Imagine the worst scenario:
858  *
859  *      ---R__________________________________________B__________
860  *         ^ reading here                             ^ bad block(assume 4k)
861  *
862  * read(R) => miss => readahead(R...B) => media error => frustrating retries
863  * => failing the whole request => read(R) => read(R+1) =>
864  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
865  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
866  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
867  *
868  * It is going insane. Fix it by quickly scaling down the readahead size.
869  */
870 static void shrink_readahead_size_eio(struct file *filp,
871                                         struct file_ra_state *ra)
872 {
873         if (!ra->ra_pages)
874                 return;
875
876         ra->ra_pages /= 4;
877 }
878
879 /**
880  * do_generic_file_read - generic file read routine
881  * @filp:       the file to read
882  * @ppos:       current file position
883  * @desc:       read_descriptor
884  * @actor:      read method
885  *
886  * This is a generic file read routine, and uses the
887  * mapping->a_ops->readpage() function for the actual low-level stuff.
888  *
889  * This is really ugly. But the goto's actually try to clarify some
890  * of the logic when it comes to error handling etc.
891  */
892 static void do_generic_file_read(struct file *filp, loff_t *ppos,
893                 read_descriptor_t *desc, read_actor_t actor)
894 {
895         struct address_space *mapping = filp->f_mapping;
896         struct inode *inode = mapping->host;
897         struct file_ra_state *ra = &filp->f_ra;
898         pgoff_t index;
899         pgoff_t last_index;
900         pgoff_t prev_index;
901         unsigned long offset;      /* offset into pagecache page */
902         unsigned int prev_offset;
903         int error;
904
905         index = *ppos >> PAGE_CACHE_SHIFT;
906         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
907         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
908         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
909         offset = *ppos & ~PAGE_CACHE_MASK;
910
911         for (;;) {
912                 struct page *page;
913                 pgoff_t end_index;
914                 loff_t isize;
915                 unsigned long nr, ret;
916
917                 cond_resched();
918 find_page:
919                 page = find_get_page(mapping, index);
920                 if (!page) {
921                         page_cache_sync_readahead(mapping,
922                                         ra, filp,
923                                         index, last_index - index);
924                         page = find_get_page(mapping, index);
925                         if (unlikely(page == NULL))
926                                 goto no_cached_page;
927                 }
928                 if (PageReadahead(page)) {
929                         page_cache_async_readahead(mapping,
930                                         ra, filp, page,
931                                         index, last_index - index);
932                 }
933                 if (!PageUptodate(page))
934                         goto page_not_up_to_date;
935 page_ok:
936                 /*
937                  * i_size must be checked after we know the page is Uptodate.
938                  *
939                  * Checking i_size after the check allows us to calculate
940                  * the correct value for "nr", which means the zero-filled
941                  * part of the page is not copied back to userspace (unless
942                  * another truncate extends the file - this is desired though).
943                  */
944
945                 isize = i_size_read(inode);
946                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
947                 if (unlikely(!isize || index > end_index)) {
948                         page_cache_release(page);
949                         goto out;
950                 }
951
952                 /* nr is the maximum number of bytes to copy from this page */
953                 nr = PAGE_CACHE_SIZE;
954                 if (index == end_index) {
955                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
956                         if (nr <= offset) {
957                                 page_cache_release(page);
958                                 goto out;
959                         }
960                 }
961                 nr = nr - offset;
962
963                 /* If users can be writing to this page using arbitrary
964                  * virtual addresses, take care about potential aliasing
965                  * before reading the page on the kernel side.
966                  */
967                 if (mapping_writably_mapped(mapping))
968                         flush_dcache_page(page);
969
970                 /*
971                  * When a sequential read accesses a page several times,
972                  * only mark it as accessed the first time.
973                  */
974                 if (prev_index != index || offset != prev_offset)
975                         mark_page_accessed(page);
976                 prev_index = index;
977
978                 /*
979                  * Ok, we have the page, and it's up-to-date, so
980                  * now we can copy it to user space...
981                  *
982                  * The actor routine returns how many bytes were actually used..
983                  * NOTE! This may not be the same as how much of a user buffer
984                  * we filled up (we may be padding etc), so we can only update
985                  * "pos" here (the actor routine has to update the user buffer
986                  * pointers and the remaining count).
987                  */
988                 ret = actor(desc, page, offset, nr);
989                 offset += ret;
990                 index += offset >> PAGE_CACHE_SHIFT;
991                 offset &= ~PAGE_CACHE_MASK;
992                 prev_offset = offset;
993
994                 page_cache_release(page);
995                 if (ret == nr && desc->count)
996                         continue;
997                 goto out;
998
999 page_not_up_to_date:
1000                 /* Get exclusive access to the page ... */
1001                 if (lock_page_killable(page))
1002                         goto readpage_eio;
1003
1004                 /* Did it get truncated before we got the lock? */
1005                 if (!page->mapping) {
1006                         unlock_page(page);
1007                         page_cache_release(page);
1008                         continue;
1009                 }
1010
1011                 /* Did somebody else fill it already? */
1012                 if (PageUptodate(page)) {
1013                         unlock_page(page);
1014                         goto page_ok;
1015                 }
1016
1017 readpage:
1018                 /* Start the actual read. The read will unlock the page. */
1019                 error = mapping->a_ops->readpage(filp, page);
1020
1021                 if (unlikely(error)) {
1022                         if (error == AOP_TRUNCATED_PAGE) {
1023                                 page_cache_release(page);
1024                                 goto find_page;
1025                         }
1026                         goto readpage_error;
1027                 }
1028
1029                 if (!PageUptodate(page)) {
1030                         if (lock_page_killable(page))
1031                                 goto readpage_eio;
1032                         if (!PageUptodate(page)) {
1033                                 if (page->mapping == NULL) {
1034                                         /*
1035                                          * invalidate_inode_pages got it
1036                                          */
1037                                         unlock_page(page);
1038                                         page_cache_release(page);
1039                                         goto find_page;
1040                                 }
1041                                 unlock_page(page);
1042                                 shrink_readahead_size_eio(filp, ra);
1043                                 goto readpage_eio;
1044                         }
1045                         unlock_page(page);
1046                 }
1047
1048                 goto page_ok;
1049
1050 readpage_eio:
1051                 error = -EIO;
1052 readpage_error:
1053                 /* UHHUH! A synchronous read error occurred. Report it */
1054                 desc->error = error;
1055                 page_cache_release(page);
1056                 goto out;
1057
1058 no_cached_page:
1059                 /*
1060                  * Ok, it wasn't cached, so we need to create a new
1061                  * page..
1062                  */
1063                 page = page_cache_alloc_cold(mapping);
1064                 if (!page) {
1065                         desc->error = -ENOMEM;
1066                         goto out;
1067                 }
1068                 error = add_to_page_cache_lru(page, mapping,
1069                                                 index, GFP_KERNEL);
1070                 if (error) {
1071                         page_cache_release(page);
1072                         if (error == -EEXIST)
1073                                 goto find_page;
1074                         desc->error = error;
1075                         goto out;
1076                 }
1077                 goto readpage;
1078         }
1079
1080 out:
1081         ra->prev_pos = prev_index;
1082         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1083         ra->prev_pos |= prev_offset;
1084
1085         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1086         if (filp)
1087                 file_accessed(filp);
1088 }
1089
1090 int file_read_actor(read_descriptor_t *desc, struct page *page,
1091                         unsigned long offset, unsigned long size)
1092 {
1093         char *kaddr;
1094         unsigned long left, count = desc->count;
1095
1096         if (size > count)
1097                 size = count;
1098
1099         /*
1100          * Faults on the destination of a read are common, so do it before
1101          * taking the kmap.
1102          */
1103         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1104                 kaddr = kmap_atomic(page, KM_USER0);
1105                 left = __copy_to_user_inatomic(desc->arg.buf,
1106                                                 kaddr + offset, size);
1107                 kunmap_atomic(kaddr, KM_USER0);
1108                 if (left == 0)
1109                         goto success;
1110         }
1111
1112         /* Do it the slow way */
1113         kaddr = kmap(page);
1114         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1115         kunmap(page);
1116
1117         if (left) {
1118                 size -= left;
1119                 desc->error = -EFAULT;
1120         }
1121 success:
1122         desc->count = count - size;
1123         desc->written += size;
1124         desc->arg.buf += size;
1125         return size;
1126 }
1127
1128 /*
1129  * Performs necessary checks before doing a write
1130  * @iov:        io vector request
1131  * @nr_segs:    number of segments in the iovec
1132  * @count:      number of bytes to write
1133  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1134  *
1135  * Adjust number of segments and amount of bytes to write (nr_segs should be
1136  * properly initialized first). Returns appropriate error code that caller
1137  * should return or zero in case that write should be allowed.
1138  */
1139 int generic_segment_checks(const struct iovec *iov,
1140                         unsigned long *nr_segs, size_t *count, int access_flags)
1141 {
1142         unsigned long   seg;
1143         size_t cnt = 0;
1144         for (seg = 0; seg < *nr_segs; seg++) {
1145                 const struct iovec *iv = &iov[seg];
1146
1147                 /*
1148                  * If any segment has a negative length, or the cumulative
1149                  * length ever wraps negative then return -EINVAL.
1150                  */
1151                 cnt += iv->iov_len;
1152                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1153                         return -EINVAL;
1154                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1155                         continue;
1156                 if (seg == 0)
1157                         return -EFAULT;
1158                 *nr_segs = seg;
1159                 cnt -= iv->iov_len;     /* This segment is no good */
1160                 break;
1161         }
1162         *count = cnt;
1163         return 0;
1164 }
1165 EXPORT_SYMBOL(generic_segment_checks);
1166
1167 /**
1168  * generic_file_aio_read - generic filesystem read routine
1169  * @iocb:       kernel I/O control block
1170  * @iov:        io vector request
1171  * @nr_segs:    number of segments in the iovec
1172  * @pos:        current file position
1173  *
1174  * This is the "read()" routine for all filesystems
1175  * that can use the page cache directly.
1176  */
1177 ssize_t
1178 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1179                 unsigned long nr_segs, loff_t pos)
1180 {
1181         struct file *filp = iocb->ki_filp;
1182         ssize_t retval;
1183         unsigned long seg;
1184         size_t count;
1185         loff_t *ppos = &iocb->ki_pos;
1186
1187         count = 0;
1188         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1189         if (retval)
1190                 return retval;
1191
1192         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1193         if (filp->f_flags & O_DIRECT) {
1194                 loff_t size;
1195                 struct address_space *mapping;
1196                 struct inode *inode;
1197
1198                 mapping = filp->f_mapping;
1199                 inode = mapping->host;
1200                 retval = 0;
1201                 if (!count)
1202                         goto out; /* skip atime */
1203                 size = i_size_read(inode);
1204                 if (pos < size) {
1205                         retval = generic_file_direct_IO(READ, iocb,
1206                                                 iov, pos, nr_segs);
1207                         if (retval > 0)
1208                                 *ppos = pos + retval;
1209                 }
1210                 if (likely(retval != 0)) {
1211                         file_accessed(filp);
1212                         goto out;
1213                 }
1214         }
1215
1216         retval = 0;
1217         if (count) {
1218                 for (seg = 0; seg < nr_segs; seg++) {
1219                         read_descriptor_t desc;
1220
1221                         desc.written = 0;
1222                         desc.arg.buf = iov[seg].iov_base;
1223                         desc.count = iov[seg].iov_len;
1224                         if (desc.count == 0)
1225                                 continue;
1226                         desc.error = 0;
1227                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1228                         retval += desc.written;
1229                         if (desc.error) {
1230                                 retval = retval ?: desc.error;
1231                                 break;
1232                         }
1233                         if (desc.count > 0)
1234                                 break;
1235                 }
1236         }
1237 out:
1238         return retval;
1239 }
1240 EXPORT_SYMBOL(generic_file_aio_read);
1241
1242 static ssize_t
1243 do_readahead(struct address_space *mapping, struct file *filp,
1244              pgoff_t index, unsigned long nr)
1245 {
1246         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1247                 return -EINVAL;
1248
1249         force_page_cache_readahead(mapping, filp, index,
1250                                         max_sane_readahead(nr));
1251         return 0;
1252 }
1253
1254 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1255 {
1256         ssize_t ret;
1257         struct file *file;
1258
1259         ret = -EBADF;
1260         file = fget(fd);
1261         if (file) {
1262                 if (file->f_mode & FMODE_READ) {
1263                         struct address_space *mapping = file->f_mapping;
1264                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1265                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1266                         unsigned long len = end - start + 1;
1267                         ret = do_readahead(mapping, file, start, len);
1268                 }
1269                 fput(file);
1270         }
1271         return ret;
1272 }
1273
1274 #ifdef CONFIG_MMU
1275 /**
1276  * page_cache_read - adds requested page to the page cache if not already there
1277  * @file:       file to read
1278  * @offset:     page index
1279  *
1280  * This adds the requested page to the page cache if it isn't already there,
1281  * and schedules an I/O to read in its contents from disk.
1282  */
1283 static int page_cache_read(struct file *file, pgoff_t offset)
1284 {
1285         struct address_space *mapping = file->f_mapping;
1286         struct page *page; 
1287         int ret;
1288
1289         do {
1290                 page = page_cache_alloc_cold(mapping);
1291                 if (!page)
1292                         return -ENOMEM;
1293
1294                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1295                 if (ret == 0)
1296                         ret = mapping->a_ops->readpage(file, page);
1297                 else if (ret == -EEXIST)
1298                         ret = 0; /* losing race to add is OK */
1299
1300                 page_cache_release(page);
1301
1302         } while (ret == AOP_TRUNCATED_PAGE);
1303                 
1304         return ret;
1305 }
1306
1307 #define MMAP_LOTSAMISS  (100)
1308
1309 /**
1310  * filemap_fault - read in file data for page fault handling
1311  * @vma:        vma in which the fault was taken
1312  * @vmf:        struct vm_fault containing details of the fault
1313  *
1314  * filemap_fault() is invoked via the vma operations vector for a
1315  * mapped memory region to read in file data during a page fault.
1316  *
1317  * The goto's are kind of ugly, but this streamlines the normal case of having
1318  * it in the page cache, and handles the special cases reasonably without
1319  * having a lot of duplicated code.
1320  */
1321 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1322 {
1323         int error;
1324         struct file *file = vma->vm_file;
1325         struct address_space *mapping = file->f_mapping;
1326         struct file_ra_state *ra = &file->f_ra;
1327         struct inode *inode = mapping->host;
1328         struct page *page;
1329         pgoff_t size;
1330         int did_readaround = 0;
1331         int ret = 0;
1332
1333         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1334         if (vmf->pgoff >= size)
1335                 return VM_FAULT_SIGBUS;
1336
1337         /* If we don't want any read-ahead, don't bother */
1338         if (VM_RandomReadHint(vma))
1339                 goto no_cached_page;
1340
1341         /*
1342          * Do we have something in the page cache already?
1343          */
1344 retry_find:
1345         page = find_lock_page(mapping, vmf->pgoff);
1346         /*
1347          * For sequential accesses, we use the generic readahead logic.
1348          */
1349         if (VM_SequentialReadHint(vma)) {
1350                 if (!page) {
1351                         page_cache_sync_readahead(mapping, ra, file,
1352                                                            vmf->pgoff, 1);
1353                         page = find_lock_page(mapping, vmf->pgoff);
1354                         if (!page)
1355                                 goto no_cached_page;
1356                 }
1357                 if (PageReadahead(page)) {
1358                         page_cache_async_readahead(mapping, ra, file, page,
1359                                                            vmf->pgoff, 1);
1360                 }
1361         }
1362
1363         if (!page) {
1364                 unsigned long ra_pages;
1365
1366                 ra->mmap_miss++;
1367
1368                 /*
1369                  * Do we miss much more than hit in this file? If so,
1370                  * stop bothering with read-ahead. It will only hurt.
1371                  */
1372                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1373                         goto no_cached_page;
1374
1375                 /*
1376                  * To keep the pgmajfault counter straight, we need to
1377                  * check did_readaround, as this is an inner loop.
1378                  */
1379                 if (!did_readaround) {
1380                         ret = VM_FAULT_MAJOR;
1381                         count_vm_event(PGMAJFAULT);
1382                 }
1383                 did_readaround = 1;
1384                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1385                 if (ra_pages) {
1386                         pgoff_t start = 0;
1387
1388                         if (vmf->pgoff > ra_pages / 2)
1389                                 start = vmf->pgoff - ra_pages / 2;
1390                         do_page_cache_readahead(mapping, file, start, ra_pages);
1391                 }
1392                 page = find_lock_page(mapping, vmf->pgoff);
1393                 if (!page)
1394                         goto no_cached_page;
1395         }
1396
1397         if (!did_readaround)
1398                 ra->mmap_miss--;
1399
1400         /*
1401          * We have a locked page in the page cache, now we need to check
1402          * that it's up-to-date. If not, it is going to be due to an error.
1403          */
1404         if (unlikely(!PageUptodate(page)))
1405                 goto page_not_uptodate;
1406
1407         /* Must recheck i_size under page lock */
1408         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1409         if (unlikely(vmf->pgoff >= size)) {
1410                 unlock_page(page);
1411                 page_cache_release(page);
1412                 return VM_FAULT_SIGBUS;
1413         }
1414
1415         /*
1416          * Found the page and have a reference on it.
1417          */
1418         mark_page_accessed(page);
1419         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1420         vmf->page = page;
1421         return ret | VM_FAULT_LOCKED;
1422
1423 no_cached_page:
1424         /*
1425          * We're only likely to ever get here if MADV_RANDOM is in
1426          * effect.
1427          */
1428         error = page_cache_read(file, vmf->pgoff);
1429
1430         /*
1431          * The page we want has now been added to the page cache.
1432          * In the unlikely event that someone removed it in the
1433          * meantime, we'll just come back here and read it again.
1434          */
1435         if (error >= 0)
1436                 goto retry_find;
1437
1438         /*
1439          * An error return from page_cache_read can result if the
1440          * system is low on memory, or a problem occurs while trying
1441          * to schedule I/O.
1442          */
1443         if (error == -ENOMEM)
1444                 return VM_FAULT_OOM;
1445         return VM_FAULT_SIGBUS;
1446
1447 page_not_uptodate:
1448         /* IO error path */
1449         if (!did_readaround) {
1450                 ret = VM_FAULT_MAJOR;
1451                 count_vm_event(PGMAJFAULT);
1452         }
1453
1454         /*
1455          * Umm, take care of errors if the page isn't up-to-date.
1456          * Try to re-read it _once_. We do this synchronously,
1457          * because there really aren't any performance issues here
1458          * and we need to check for errors.
1459          */
1460         ClearPageError(page);
1461         error = mapping->a_ops->readpage(file, page);
1462         page_cache_release(page);
1463
1464         if (!error || error == AOP_TRUNCATED_PAGE)
1465                 goto retry_find;
1466
1467         /* Things didn't work out. Return zero to tell the mm layer so. */
1468         shrink_readahead_size_eio(file, ra);
1469         return VM_FAULT_SIGBUS;
1470 }
1471 EXPORT_SYMBOL(filemap_fault);
1472
1473 struct vm_operations_struct generic_file_vm_ops = {
1474         .fault          = filemap_fault,
1475 };
1476
1477 /* This is used for a general mmap of a disk file */
1478
1479 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1480 {
1481         struct address_space *mapping = file->f_mapping;
1482
1483         if (!mapping->a_ops->readpage)
1484                 return -ENOEXEC;
1485         file_accessed(file);
1486         vma->vm_ops = &generic_file_vm_ops;
1487         vma->vm_flags |= VM_CAN_NONLINEAR;
1488         return 0;
1489 }
1490
1491 /*
1492  * This is for filesystems which do not implement ->writepage.
1493  */
1494 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1495 {
1496         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1497                 return -EINVAL;
1498         return generic_file_mmap(file, vma);
1499 }
1500 #else
1501 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1502 {
1503         return -ENOSYS;
1504 }
1505 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1506 {
1507         return -ENOSYS;
1508 }
1509 #endif /* CONFIG_MMU */
1510
1511 EXPORT_SYMBOL(generic_file_mmap);
1512 EXPORT_SYMBOL(generic_file_readonly_mmap);
1513
1514 static struct page *__read_cache_page(struct address_space *mapping,
1515                                 pgoff_t index,
1516                                 int (*filler)(void *,struct page*),
1517                                 void *data)
1518 {
1519         struct page *page;
1520         int err;
1521 repeat:
1522         page = find_get_page(mapping, index);
1523         if (!page) {
1524                 page = page_cache_alloc_cold(mapping);
1525                 if (!page)
1526                         return ERR_PTR(-ENOMEM);
1527                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1528                 if (unlikely(err)) {
1529                         page_cache_release(page);
1530                         if (err == -EEXIST)
1531                                 goto repeat;
1532                         /* Presumably ENOMEM for radix tree node */
1533                         return ERR_PTR(err);
1534                 }
1535                 err = filler(data, page);
1536                 if (err < 0) {
1537                         page_cache_release(page);
1538                         page = ERR_PTR(err);
1539                 }
1540         }
1541         return page;
1542 }
1543
1544 /**
1545  * read_cache_page_async - read into page cache, fill it if needed
1546  * @mapping:    the page's address_space
1547  * @index:      the page index
1548  * @filler:     function to perform the read
1549  * @data:       destination for read data
1550  *
1551  * Same as read_cache_page, but don't wait for page to become unlocked
1552  * after submitting it to the filler.
1553  *
1554  * Read into the page cache. If a page already exists, and PageUptodate() is
1555  * not set, try to fill the page but don't wait for it to become unlocked.
1556  *
1557  * If the page does not get brought uptodate, return -EIO.
1558  */
1559 struct page *read_cache_page_async(struct address_space *mapping,
1560                                 pgoff_t index,
1561                                 int (*filler)(void *,struct page*),
1562                                 void *data)
1563 {
1564         struct page *page;
1565         int err;
1566
1567 retry:
1568         page = __read_cache_page(mapping, index, filler, data);
1569         if (IS_ERR(page))
1570                 return page;
1571         if (PageUptodate(page))
1572                 goto out;
1573
1574         lock_page(page);
1575         if (!page->mapping) {
1576                 unlock_page(page);
1577                 page_cache_release(page);
1578                 goto retry;
1579         }
1580         if (PageUptodate(page)) {
1581                 unlock_page(page);
1582                 goto out;
1583         }
1584         err = filler(data, page);
1585         if (err < 0) {
1586                 page_cache_release(page);
1587                 return ERR_PTR(err);
1588         }
1589 out:
1590         mark_page_accessed(page);
1591         return page;
1592 }
1593 EXPORT_SYMBOL(read_cache_page_async);
1594
1595 /**
1596  * read_cache_page - read into page cache, fill it if needed
1597  * @mapping:    the page's address_space
1598  * @index:      the page index
1599  * @filler:     function to perform the read
1600  * @data:       destination for read data
1601  *
1602  * Read into the page cache. If a page already exists, and PageUptodate() is
1603  * not set, try to fill the page then wait for it to become unlocked.
1604  *
1605  * If the page does not get brought uptodate, return -EIO.
1606  */
1607 struct page *read_cache_page(struct address_space *mapping,
1608                                 pgoff_t index,
1609                                 int (*filler)(void *,struct page*),
1610                                 void *data)
1611 {
1612         struct page *page;
1613
1614         page = read_cache_page_async(mapping, index, filler, data);
1615         if (IS_ERR(page))
1616                 goto out;
1617         wait_on_page_locked(page);
1618         if (!PageUptodate(page)) {
1619                 page_cache_release(page);
1620                 page = ERR_PTR(-EIO);
1621         }
1622  out:
1623         return page;
1624 }
1625 EXPORT_SYMBOL(read_cache_page);
1626
1627 /*
1628  * The logic we want is
1629  *
1630  *      if suid or (sgid and xgrp)
1631  *              remove privs
1632  */
1633 int should_remove_suid(struct dentry *dentry)
1634 {
1635         mode_t mode = dentry->d_inode->i_mode;
1636         int kill = 0;
1637
1638         /* suid always must be killed */
1639         if (unlikely(mode & S_ISUID))
1640                 kill = ATTR_KILL_SUID;
1641
1642         /*
1643          * sgid without any exec bits is just a mandatory locking mark; leave
1644          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1645          */
1646         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1647                 kill |= ATTR_KILL_SGID;
1648
1649         if (unlikely(kill && !capable(CAP_FSETID)))
1650                 return kill;
1651
1652         return 0;
1653 }
1654 EXPORT_SYMBOL(should_remove_suid);
1655
1656 int __remove_suid(struct dentry *dentry, int kill)
1657 {
1658         struct iattr newattrs;
1659
1660         newattrs.ia_valid = ATTR_FORCE | kill;
1661         return notify_change(dentry, &newattrs);
1662 }
1663
1664 int remove_suid(struct dentry *dentry)
1665 {
1666         int killsuid = should_remove_suid(dentry);
1667         int killpriv = security_inode_need_killpriv(dentry);
1668         int error = 0;
1669
1670         if (killpriv < 0)
1671                 return killpriv;
1672         if (killpriv)
1673                 error = security_inode_killpriv(dentry);
1674         if (!error && killsuid)
1675                 error = __remove_suid(dentry, killsuid);
1676
1677         return error;
1678 }
1679 EXPORT_SYMBOL(remove_suid);
1680
1681 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1682                         const struct iovec *iov, size_t base, size_t bytes)
1683 {
1684         size_t copied = 0, left = 0;
1685
1686         while (bytes) {
1687                 char __user *buf = iov->iov_base + base;
1688                 int copy = min(bytes, iov->iov_len - base);
1689
1690                 base = 0;
1691                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1692                 copied += copy;
1693                 bytes -= copy;
1694                 vaddr += copy;
1695                 iov++;
1696
1697                 if (unlikely(left))
1698                         break;
1699         }
1700         return copied - left;
1701 }
1702
1703 /*
1704  * Copy as much as we can into the page and return the number of bytes which
1705  * were sucessfully copied.  If a fault is encountered then return the number of
1706  * bytes which were copied.
1707  */
1708 size_t iov_iter_copy_from_user_atomic(struct page *page,
1709                 struct iov_iter *i, unsigned long offset, size_t bytes)
1710 {
1711         char *kaddr;
1712         size_t copied;
1713
1714         BUG_ON(!in_atomic());
1715         kaddr = kmap_atomic(page, KM_USER0);
1716         if (likely(i->nr_segs == 1)) {
1717                 int left;
1718                 char __user *buf = i->iov->iov_base + i->iov_offset;
1719                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1720                                                         buf, bytes);
1721                 copied = bytes - left;
1722         } else {
1723                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1724                                                 i->iov, i->iov_offset, bytes);
1725         }
1726         kunmap_atomic(kaddr, KM_USER0);
1727
1728         return copied;
1729 }
1730 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1731
1732 /*
1733  * This has the same sideeffects and return value as
1734  * iov_iter_copy_from_user_atomic().
1735  * The difference is that it attempts to resolve faults.
1736  * Page must not be locked.
1737  */
1738 size_t iov_iter_copy_from_user(struct page *page,
1739                 struct iov_iter *i, unsigned long offset, size_t bytes)
1740 {
1741         char *kaddr;
1742         size_t copied;
1743
1744         kaddr = kmap(page);
1745         if (likely(i->nr_segs == 1)) {
1746                 int left;
1747                 char __user *buf = i->iov->iov_base + i->iov_offset;
1748                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1749                 copied = bytes - left;
1750         } else {
1751                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1752                                                 i->iov, i->iov_offset, bytes);
1753         }
1754         kunmap(page);
1755         return copied;
1756 }
1757 EXPORT_SYMBOL(iov_iter_copy_from_user);
1758
1759 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1760 {
1761         BUG_ON(i->count < bytes);
1762
1763         if (likely(i->nr_segs == 1)) {
1764                 i->iov_offset += bytes;
1765                 i->count -= bytes;
1766         } else {
1767                 const struct iovec *iov = i->iov;
1768                 size_t base = i->iov_offset;
1769
1770                 /*
1771                  * The !iov->iov_len check ensures we skip over unlikely
1772                  * zero-length segments (without overruning the iovec).
1773                  */
1774                 while (bytes || unlikely(!iov->iov_len && i->count)) {
1775                         int copy;
1776
1777                         copy = min(bytes, iov->iov_len - base);
1778                         BUG_ON(!i->count || i->count < copy);
1779                         i->count -= copy;
1780                         bytes -= copy;
1781                         base += copy;
1782                         if (iov->iov_len == base) {
1783                                 iov++;
1784                                 base = 0;
1785                         }
1786                 }
1787                 i->iov = iov;
1788                 i->iov_offset = base;
1789         }
1790 }
1791 EXPORT_SYMBOL(iov_iter_advance);
1792
1793 /*
1794  * Fault in the first iovec of the given iov_iter, to a maximum length
1795  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1796  * accessed (ie. because it is an invalid address).
1797  *
1798  * writev-intensive code may want this to prefault several iovecs -- that
1799  * would be possible (callers must not rely on the fact that _only_ the
1800  * first iovec will be faulted with the current implementation).
1801  */
1802 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1803 {
1804         char __user *buf = i->iov->iov_base + i->iov_offset;
1805         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1806         return fault_in_pages_readable(buf, bytes);
1807 }
1808 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1809
1810 /*
1811  * Return the count of just the current iov_iter segment.
1812  */
1813 size_t iov_iter_single_seg_count(struct iov_iter *i)
1814 {
1815         const struct iovec *iov = i->iov;
1816         if (i->nr_segs == 1)
1817                 return i->count;
1818         else
1819                 return min(i->count, iov->iov_len - i->iov_offset);
1820 }
1821 EXPORT_SYMBOL(iov_iter_single_seg_count);
1822
1823 /*
1824  * Performs necessary checks before doing a write
1825  *
1826  * Can adjust writing position or amount of bytes to write.
1827  * Returns appropriate error code that caller should return or
1828  * zero in case that write should be allowed.
1829  */
1830 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1831 {
1832         struct inode *inode = file->f_mapping->host;
1833         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1834
1835         if (unlikely(*pos < 0))
1836                 return -EINVAL;
1837
1838         if (!isblk) {
1839                 /* FIXME: this is for backwards compatibility with 2.4 */
1840                 if (file->f_flags & O_APPEND)
1841                         *pos = i_size_read(inode);
1842
1843                 if (limit != RLIM_INFINITY) {
1844                         if (*pos >= limit) {
1845                                 send_sig(SIGXFSZ, current, 0);
1846                                 return -EFBIG;
1847                         }
1848                         if (*count > limit - (typeof(limit))*pos) {
1849                                 *count = limit - (typeof(limit))*pos;
1850                         }
1851                 }
1852         }
1853
1854         /*
1855          * LFS rule
1856          */
1857         if (unlikely(*pos + *count > MAX_NON_LFS &&
1858                                 !(file->f_flags & O_LARGEFILE))) {
1859                 if (*pos >= MAX_NON_LFS) {
1860                         return -EFBIG;
1861                 }
1862                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1863                         *count = MAX_NON_LFS - (unsigned long)*pos;
1864                 }
1865         }
1866
1867         /*
1868          * Are we about to exceed the fs block limit ?
1869          *
1870          * If we have written data it becomes a short write.  If we have
1871          * exceeded without writing data we send a signal and return EFBIG.
1872          * Linus frestrict idea will clean these up nicely..
1873          */
1874         if (likely(!isblk)) {
1875                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1876                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1877                                 return -EFBIG;
1878                         }
1879                         /* zero-length writes at ->s_maxbytes are OK */
1880                 }
1881
1882                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1883                         *count = inode->i_sb->s_maxbytes - *pos;
1884         } else {
1885 #ifdef CONFIG_BLOCK
1886                 loff_t isize;
1887                 if (bdev_read_only(I_BDEV(inode)))
1888                         return -EPERM;
1889                 isize = i_size_read(inode);
1890                 if (*pos >= isize) {
1891                         if (*count || *pos > isize)
1892                                 return -ENOSPC;
1893                 }
1894
1895                 if (*pos + *count > isize)
1896                         *count = isize - *pos;
1897 #else
1898                 return -EPERM;
1899 #endif
1900         }
1901         return 0;
1902 }
1903 EXPORT_SYMBOL(generic_write_checks);
1904
1905 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1906                                 loff_t pos, unsigned len, unsigned flags,
1907                                 struct page **pagep, void **fsdata)
1908 {
1909         const struct address_space_operations *aops = mapping->a_ops;
1910
1911         if (aops->write_begin) {
1912                 return aops->write_begin(file, mapping, pos, len, flags,
1913                                                         pagep, fsdata);
1914         } else {
1915                 int ret;
1916                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1917                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1918                 struct inode *inode = mapping->host;
1919                 struct page *page;
1920 again:
1921                 page = __grab_cache_page(mapping, index);
1922                 *pagep = page;
1923                 if (!page)
1924                         return -ENOMEM;
1925
1926                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1927                         /*
1928                          * There is no way to resolve a short write situation
1929                          * for a !Uptodate page (except by double copying in
1930                          * the caller done by generic_perform_write_2copy).
1931                          *
1932                          * Instead, we have to bring it uptodate here.
1933                          */
1934                         ret = aops->readpage(file, page);
1935                         page_cache_release(page);
1936                         if (ret) {
1937                                 if (ret == AOP_TRUNCATED_PAGE)
1938                                         goto again;
1939                                 return ret;
1940                         }
1941                         goto again;
1942                 }
1943
1944                 ret = aops->prepare_write(file, page, offset, offset+len);
1945                 if (ret) {
1946                         unlock_page(page);
1947                         page_cache_release(page);
1948                         if (pos + len > inode->i_size)
1949                                 vmtruncate(inode, inode->i_size);
1950                 }
1951                 return ret;
1952         }
1953 }
1954 EXPORT_SYMBOL(pagecache_write_begin);
1955
1956 int pagecache_write_end(struct file *file, struct address_space *mapping,
1957                                 loff_t pos, unsigned len, unsigned copied,
1958                                 struct page *page, void *fsdata)
1959 {
1960         const struct address_space_operations *aops = mapping->a_ops;
1961         int ret;
1962
1963         if (aops->write_end) {
1964                 mark_page_accessed(page);
1965                 ret = aops->write_end(file, mapping, pos, len, copied,
1966                                                         page, fsdata);
1967         } else {
1968                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1969                 struct inode *inode = mapping->host;
1970
1971                 flush_dcache_page(page);
1972                 ret = aops->commit_write(file, page, offset, offset+len);
1973                 unlock_page(page);
1974                 mark_page_accessed(page);
1975                 page_cache_release(page);
1976
1977                 if (ret < 0) {
1978                         if (pos + len > inode->i_size)
1979                                 vmtruncate(inode, inode->i_size);
1980                 } else if (ret > 0)
1981                         ret = min_t(size_t, copied, ret);
1982                 else
1983                         ret = copied;
1984         }
1985
1986         return ret;
1987 }
1988 EXPORT_SYMBOL(pagecache_write_end);
1989
1990 ssize_t
1991 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1992                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1993                 size_t count, size_t ocount)
1994 {
1995         struct file     *file = iocb->ki_filp;
1996         struct address_space *mapping = file->f_mapping;
1997         struct inode    *inode = mapping->host;
1998         ssize_t         written;
1999
2000         if (count != ocount)
2001                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2002
2003         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2004         if (written > 0) {
2005                 loff_t end = pos + written;
2006                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2007                         i_size_write(inode,  end);
2008                         mark_inode_dirty(inode);
2009                 }
2010                 *ppos = end;
2011         }
2012
2013         /*
2014          * Sync the fs metadata but not the minor inode changes and
2015          * of course not the data as we did direct DMA for the IO.
2016          * i_mutex is held, which protects generic_osync_inode() from
2017          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2018          */
2019         if ((written >= 0 || written == -EIOCBQUEUED) &&
2020             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2021                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2022                 if (err < 0)
2023                         written = err;
2024         }
2025         return written;
2026 }
2027 EXPORT_SYMBOL(generic_file_direct_write);
2028
2029 /*
2030  * Find or create a page at the given pagecache position. Return the locked
2031  * page. This function is specifically for buffered writes.
2032  */
2033 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2034 {
2035         int status;
2036         struct page *page;
2037 repeat:
2038         page = find_lock_page(mapping, index);
2039         if (likely(page))
2040                 return page;
2041
2042         page = page_cache_alloc(mapping);
2043         if (!page)
2044                 return NULL;
2045         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2046         if (unlikely(status)) {
2047                 page_cache_release(page);
2048                 if (status == -EEXIST)
2049                         goto repeat;
2050                 return NULL;
2051         }
2052         return page;
2053 }
2054 EXPORT_SYMBOL(__grab_cache_page);
2055
2056 static ssize_t generic_perform_write_2copy(struct file *file,
2057                                 struct iov_iter *i, loff_t pos)
2058 {
2059         struct address_space *mapping = file->f_mapping;
2060         const struct address_space_operations *a_ops = mapping->a_ops;
2061         struct inode *inode = mapping->host;
2062         long status = 0;
2063         ssize_t written = 0;
2064
2065         do {
2066                 struct page *src_page;
2067                 struct page *page;
2068                 pgoff_t index;          /* Pagecache index for current page */
2069                 unsigned long offset;   /* Offset into pagecache page */
2070                 unsigned long bytes;    /* Bytes to write to page */
2071                 size_t copied;          /* Bytes copied from user */
2072
2073                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2074                 index = pos >> PAGE_CACHE_SHIFT;
2075                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2076                                                 iov_iter_count(i));
2077
2078                 /*
2079                  * a non-NULL src_page indicates that we're doing the
2080                  * copy via get_user_pages and kmap.
2081                  */
2082                 src_page = NULL;
2083
2084                 /*
2085                  * Bring in the user page that we will copy from _first_.
2086                  * Otherwise there's a nasty deadlock on copying from the
2087                  * same page as we're writing to, without it being marked
2088                  * up-to-date.
2089                  *
2090                  * Not only is this an optimisation, but it is also required
2091                  * to check that the address is actually valid, when atomic
2092                  * usercopies are used, below.
2093                  */
2094                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2095                         status = -EFAULT;
2096                         break;
2097                 }
2098
2099                 page = __grab_cache_page(mapping, index);
2100                 if (!page) {
2101                         status = -ENOMEM;
2102                         break;
2103                 }
2104
2105                 /*
2106                  * non-uptodate pages cannot cope with short copies, and we
2107                  * cannot take a pagefault with the destination page locked.
2108                  * So pin the source page to copy it.
2109                  */
2110                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2111                         unlock_page(page);
2112
2113                         src_page = alloc_page(GFP_KERNEL);
2114                         if (!src_page) {
2115                                 page_cache_release(page);
2116                                 status = -ENOMEM;
2117                                 break;
2118                         }
2119
2120                         /*
2121                          * Cannot get_user_pages with a page locked for the
2122                          * same reason as we can't take a page fault with a
2123                          * page locked (as explained below).
2124                          */
2125                         copied = iov_iter_copy_from_user(src_page, i,
2126                                                                 offset, bytes);
2127                         if (unlikely(copied == 0)) {
2128                                 status = -EFAULT;
2129                                 page_cache_release(page);
2130                                 page_cache_release(src_page);
2131                                 break;
2132                         }
2133                         bytes = copied;
2134
2135                         lock_page(page);
2136                         /*
2137                          * Can't handle the page going uptodate here, because
2138                          * that means we would use non-atomic usercopies, which
2139                          * zero out the tail of the page, which can cause
2140                          * zeroes to become transiently visible. We could just
2141                          * use a non-zeroing copy, but the APIs aren't too
2142                          * consistent.
2143                          */
2144                         if (unlikely(!page->mapping || PageUptodate(page))) {
2145                                 unlock_page(page);
2146                                 page_cache_release(page);
2147                                 page_cache_release(src_page);
2148                                 continue;
2149                         }
2150                 }
2151
2152                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2153                 if (unlikely(status))
2154                         goto fs_write_aop_error;
2155
2156                 if (!src_page) {
2157                         /*
2158                          * Must not enter the pagefault handler here, because
2159                          * we hold the page lock, so we might recursively
2160                          * deadlock on the same lock, or get an ABBA deadlock
2161                          * against a different lock, or against the mmap_sem
2162                          * (which nests outside the page lock).  So increment
2163                          * preempt count, and use _atomic usercopies.
2164                          *
2165                          * The page is uptodate so we are OK to encounter a
2166                          * short copy: if unmodified parts of the page are
2167                          * marked dirty and written out to disk, it doesn't
2168                          * really matter.
2169                          */
2170                         pagefault_disable();
2171                         copied = iov_iter_copy_from_user_atomic(page, i,
2172                                                                 offset, bytes);
2173                         pagefault_enable();
2174                 } else {
2175                         void *src, *dst;
2176                         src = kmap_atomic(src_page, KM_USER0);
2177                         dst = kmap_atomic(page, KM_USER1);
2178                         memcpy(dst + offset, src + offset, bytes);
2179                         kunmap_atomic(dst, KM_USER1);
2180                         kunmap_atomic(src, KM_USER0);
2181                         copied = bytes;
2182                 }
2183                 flush_dcache_page(page);
2184
2185                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2186                 if (unlikely(status < 0))
2187                         goto fs_write_aop_error;
2188                 if (unlikely(status > 0)) /* filesystem did partial write */
2189                         copied = min_t(size_t, copied, status);
2190
2191                 unlock_page(page);
2192                 mark_page_accessed(page);
2193                 page_cache_release(page);
2194                 if (src_page)
2195                         page_cache_release(src_page);
2196
2197                 iov_iter_advance(i, copied);
2198                 pos += copied;
2199                 written += copied;
2200
2201                 balance_dirty_pages_ratelimited(mapping);
2202                 cond_resched();
2203                 continue;
2204
2205 fs_write_aop_error:
2206                 unlock_page(page);
2207                 page_cache_release(page);
2208                 if (src_page)
2209                         page_cache_release(src_page);
2210
2211                 /*
2212                  * prepare_write() may have instantiated a few blocks
2213                  * outside i_size.  Trim these off again. Don't need
2214                  * i_size_read because we hold i_mutex.
2215                  */
2216                 if (pos + bytes > inode->i_size)
2217                         vmtruncate(inode, inode->i_size);
2218                 break;
2219         } while (iov_iter_count(i));
2220
2221         return written ? written : status;
2222 }
2223
2224 static ssize_t generic_perform_write(struct file *file,
2225                                 struct iov_iter *i, loff_t pos)
2226 {
2227         struct address_space *mapping = file->f_mapping;
2228         const struct address_space_operations *a_ops = mapping->a_ops;
2229         long status = 0;
2230         ssize_t written = 0;
2231         unsigned int flags = 0;
2232
2233         /*
2234          * Copies from kernel address space cannot fail (NFSD is a big user).
2235          */
2236         if (segment_eq(get_fs(), KERNEL_DS))
2237                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2238
2239         do {
2240                 struct page *page;
2241                 pgoff_t index;          /* Pagecache index for current page */
2242                 unsigned long offset;   /* Offset into pagecache page */
2243                 unsigned long bytes;    /* Bytes to write to page */
2244                 size_t copied;          /* Bytes copied from user */
2245                 void *fsdata;
2246
2247                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2248                 index = pos >> PAGE_CACHE_SHIFT;
2249                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2250                                                 iov_iter_count(i));
2251
2252 again:
2253
2254                 /*
2255                  * Bring in the user page that we will copy from _first_.
2256                  * Otherwise there's a nasty deadlock on copying from the
2257                  * same page as we're writing to, without it being marked
2258                  * up-to-date.
2259                  *
2260                  * Not only is this an optimisation, but it is also required
2261                  * to check that the address is actually valid, when atomic
2262                  * usercopies are used, below.
2263                  */
2264                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2265                         status = -EFAULT;
2266                         break;
2267                 }
2268
2269                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2270                                                 &page, &fsdata);
2271                 if (unlikely(status))
2272                         break;
2273
2274                 pagefault_disable();
2275                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2276                 pagefault_enable();
2277                 flush_dcache_page(page);
2278
2279                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2280                                                 page, fsdata);
2281                 if (unlikely(status < 0))
2282                         break;
2283                 copied = status;
2284
2285                 cond_resched();
2286
2287                 iov_iter_advance(i, copied);
2288                 if (unlikely(copied == 0)) {
2289                         /*
2290                          * If we were unable to copy any data at all, we must
2291                          * fall back to a single segment length write.
2292                          *
2293                          * If we didn't fallback here, we could livelock
2294                          * because not all segments in the iov can be copied at
2295                          * once without a pagefault.
2296                          */
2297                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2298                                                 iov_iter_single_seg_count(i));
2299                         goto again;
2300                 }
2301                 pos += copied;
2302                 written += copied;
2303
2304                 balance_dirty_pages_ratelimited(mapping);
2305
2306         } while (iov_iter_count(i));
2307
2308         return written ? written : status;
2309 }
2310
2311 ssize_t
2312 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2313                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2314                 size_t count, ssize_t written)
2315 {
2316         struct file *file = iocb->ki_filp;
2317         struct address_space *mapping = file->f_mapping;
2318         const struct address_space_operations *a_ops = mapping->a_ops;
2319         struct inode *inode = mapping->host;
2320         ssize_t status;
2321         struct iov_iter i;
2322
2323         iov_iter_init(&i, iov, nr_segs, count, written);
2324         if (a_ops->write_begin)
2325                 status = generic_perform_write(file, &i, pos);
2326         else
2327                 status = generic_perform_write_2copy(file, &i, pos);
2328
2329         if (likely(status >= 0)) {
2330                 written += status;
2331                 *ppos = pos + status;
2332
2333                 /*
2334                  * For now, when the user asks for O_SYNC, we'll actually give
2335                  * O_DSYNC
2336                  */
2337                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2338                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2339                                 status = generic_osync_inode(inode, mapping,
2340                                                 OSYNC_METADATA|OSYNC_DATA);
2341                 }
2342         }
2343         
2344         /*
2345          * If we get here for O_DIRECT writes then we must have fallen through
2346          * to buffered writes (block instantiation inside i_size).  So we sync
2347          * the file data here, to try to honour O_DIRECT expectations.
2348          */
2349         if (unlikely(file->f_flags & O_DIRECT) && written)
2350                 status = filemap_write_and_wait(mapping);
2351
2352         return written ? written : status;
2353 }
2354 EXPORT_SYMBOL(generic_file_buffered_write);
2355
2356 static ssize_t
2357 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2358                                 unsigned long nr_segs, loff_t *ppos)
2359 {
2360         struct file *file = iocb->ki_filp;
2361         struct address_space * mapping = file->f_mapping;
2362         size_t ocount;          /* original count */
2363         size_t count;           /* after file limit checks */
2364         struct inode    *inode = mapping->host;
2365         loff_t          pos;
2366         ssize_t         written;
2367         ssize_t         err;
2368
2369         ocount = 0;
2370         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2371         if (err)
2372                 return err;
2373
2374         count = ocount;
2375         pos = *ppos;
2376
2377         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2378
2379         /* We can write back this queue in page reclaim */
2380         current->backing_dev_info = mapping->backing_dev_info;
2381         written = 0;
2382
2383         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2384         if (err)
2385                 goto out;
2386
2387         if (count == 0)
2388                 goto out;
2389
2390         err = remove_suid(file->f_path.dentry);
2391         if (err)
2392                 goto out;
2393
2394         file_update_time(file);
2395
2396         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2397         if (unlikely(file->f_flags & O_DIRECT)) {
2398                 loff_t endbyte;
2399                 ssize_t written_buffered;
2400
2401                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2402                                                         ppos, count, ocount);
2403                 if (written < 0 || written == count)
2404                         goto out;
2405                 /*
2406                  * direct-io write to a hole: fall through to buffered I/O
2407                  * for completing the rest of the request.
2408                  */
2409                 pos += written;
2410                 count -= written;
2411                 written_buffered = generic_file_buffered_write(iocb, iov,
2412                                                 nr_segs, pos, ppos, count,
2413                                                 written);
2414                 /*
2415                  * If generic_file_buffered_write() retuned a synchronous error
2416                  * then we want to return the number of bytes which were
2417                  * direct-written, or the error code if that was zero.  Note
2418                  * that this differs from normal direct-io semantics, which
2419                  * will return -EFOO even if some bytes were written.
2420                  */
2421                 if (written_buffered < 0) {
2422                         err = written_buffered;
2423                         goto out;
2424                 }
2425
2426                 /*
2427                  * We need to ensure that the page cache pages are written to
2428                  * disk and invalidated to preserve the expected O_DIRECT
2429                  * semantics.
2430                  */
2431                 endbyte = pos + written_buffered - written - 1;
2432                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2433                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2434                                             SYNC_FILE_RANGE_WRITE|
2435                                             SYNC_FILE_RANGE_WAIT_AFTER);
2436                 if (err == 0) {
2437                         written = written_buffered;
2438                         invalidate_mapping_pages(mapping,
2439                                                  pos >> PAGE_CACHE_SHIFT,
2440                                                  endbyte >> PAGE_CACHE_SHIFT);
2441                 } else {
2442                         /*
2443                          * We don't know how much we wrote, so just return
2444                          * the number of bytes which were direct-written
2445                          */
2446                 }
2447         } else {
2448                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2449                                 pos, ppos, count, written);
2450         }
2451 out:
2452         current->backing_dev_info = NULL;
2453         return written ? written : err;
2454 }
2455
2456 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2457                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2458 {
2459         struct file *file = iocb->ki_filp;
2460         struct address_space *mapping = file->f_mapping;
2461         struct inode *inode = mapping->host;
2462         ssize_t ret;
2463
2464         BUG_ON(iocb->ki_pos != pos);
2465
2466         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2467                         &iocb->ki_pos);
2468
2469         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2470                 ssize_t err;
2471
2472                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2473                 if (err < 0)
2474                         ret = err;
2475         }
2476         return ret;
2477 }
2478 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2479
2480 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2481                 unsigned long nr_segs, loff_t pos)
2482 {
2483         struct file *file = iocb->ki_filp;
2484         struct address_space *mapping = file->f_mapping;
2485         struct inode *inode = mapping->host;
2486         ssize_t ret;
2487
2488         BUG_ON(iocb->ki_pos != pos);
2489
2490         mutex_lock(&inode->i_mutex);
2491         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2492                         &iocb->ki_pos);
2493         mutex_unlock(&inode->i_mutex);
2494
2495         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2496                 ssize_t err;
2497
2498                 err = sync_page_range(inode, mapping, pos, ret);
2499                 if (err < 0)
2500                         ret = err;
2501         }
2502         return ret;
2503 }
2504 EXPORT_SYMBOL(generic_file_aio_write);
2505
2506 /*
2507  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2508  * went wrong during pagecache shootdown.
2509  */
2510 static ssize_t
2511 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2512         loff_t offset, unsigned long nr_segs)
2513 {
2514         struct file *file = iocb->ki_filp;
2515         struct address_space *mapping = file->f_mapping;
2516         ssize_t retval;
2517         size_t write_len;
2518         pgoff_t end = 0; /* silence gcc */
2519
2520         /*
2521          * If it's a write, unmap all mmappings of the file up-front.  This
2522          * will cause any pte dirty bits to be propagated into the pageframes
2523          * for the subsequent filemap_write_and_wait().
2524          */
2525         if (rw == WRITE) {
2526                 write_len = iov_length(iov, nr_segs);
2527                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2528                 if (mapping_mapped(mapping))
2529                         unmap_mapping_range(mapping, offset, write_len, 0);
2530         }
2531
2532         retval = filemap_write_and_wait(mapping);
2533         if (retval)
2534                 goto out;
2535
2536         /*
2537          * After a write we want buffered reads to be sure to go to disk to get
2538          * the new data.  We invalidate clean cached page from the region we're
2539          * about to write.  We do this *before* the write so that we can return
2540          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2541          */
2542         if (rw == WRITE && mapping->nrpages) {
2543                 retval = invalidate_inode_pages2_range(mapping,
2544                                         offset >> PAGE_CACHE_SHIFT, end);
2545                 if (retval)
2546                         goto out;
2547         }
2548
2549         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2550
2551         /*
2552          * Finally, try again to invalidate clean pages which might have been
2553          * cached by non-direct readahead, or faulted in by get_user_pages()
2554          * if the source of the write was an mmap'ed region of the file
2555          * we're writing.  Either one is a pretty crazy thing to do,
2556          * so we don't support it 100%.  If this invalidation
2557          * fails, tough, the write still worked...
2558          */
2559         if (rw == WRITE && mapping->nrpages) {
2560                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2561         }
2562 out:
2563         return retval;
2564 }
2565
2566 /**
2567  * try_to_release_page() - release old fs-specific metadata on a page
2568  *
2569  * @page: the page which the kernel is trying to free
2570  * @gfp_mask: memory allocation flags (and I/O mode)
2571  *
2572  * The address_space is to try to release any data against the page
2573  * (presumably at page->private).  If the release was successful, return `1'.
2574  * Otherwise return zero.
2575  *
2576  * The @gfp_mask argument specifies whether I/O may be performed to release
2577  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2578  *
2579  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2580  */
2581 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2582 {
2583         struct address_space * const mapping = page->mapping;
2584
2585         BUG_ON(!PageLocked(page));
2586         if (PageWriteback(page))
2587                 return 0;
2588
2589         if (mapping && mapping->a_ops->releasepage)
2590                 return mapping->a_ops->releasepage(page, gfp_mask);
2591         return try_to_free_buffers(page);
2592 }
2593
2594 EXPORT_SYMBOL(try_to_release_page);