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