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