Merge branch 'master'
[linux-3.10.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/config.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/compiler.h>
16 #include <linux/fs.h>
17 #include <linux/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/uaccess.h>
42 #include <asm/mman.h>
43
44 static ssize_t
45 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
46         loff_t offset, unsigned long nr_segs);
47
48 /*
49  * Shared mappings implemented 30.11.1994. It's not fully working yet,
50  * though.
51  *
52  * Shared mappings now work. 15.8.1995  Bruno.
53  *
54  * finished 'unifying' the page and buffer cache and SMP-threaded the
55  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
56  *
57  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
58  */
59
60 /*
61  * Lock ordering:
62  *
63  *  ->i_mmap_lock               (vmtruncate)
64  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
65  *      ->swap_lock             (exclusive_swap_page, others)
66  *        ->mapping->tree_lock
67  *
68  *  ->i_mutex
69  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
70  *
71  *  ->mmap_sem
72  *    ->i_mmap_lock
73  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
74  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
75  *
76  *  ->mmap_sem
77  *    ->lock_page               (access_process_vm)
78  *
79  *  ->mmap_sem
80  *    ->i_mutex                 (msync)
81  *
82  *  ->i_mutex
83  *    ->i_alloc_sem             (various)
84  *
85  *  ->inode_lock
86  *    ->sb_lock                 (fs/fs-writeback.c)
87  *    ->mapping->tree_lock      (__sync_single_inode)
88  *
89  *  ->i_mmap_lock
90  *    ->anon_vma.lock           (vma_adjust)
91  *
92  *  ->anon_vma.lock
93  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
94  *
95  *  ->page_table_lock or pte_lock
96  *    ->swap_lock               (try_to_unmap_one)
97  *    ->private_lock            (try_to_unmap_one)
98  *    ->tree_lock               (try_to_unmap_one)
99  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
100  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
101  *    ->private_lock            (page_remove_rmap->set_page_dirty)
102  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
103  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
104  *    ->inode_lock              (zap_pte_range->set_page_dirty)
105  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
106  *
107  *  ->task->proc_lock
108  *    ->dcache_lock             (proc_pid_lookup)
109  */
110
111 /*
112  * Remove a page from the page cache and free it. Caller has to make
113  * sure the page is locked and that nobody else uses it - or that usage
114  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
115  */
116 void __remove_from_page_cache(struct page *page)
117 {
118         struct address_space *mapping = page->mapping;
119
120         radix_tree_delete(&mapping->page_tree, page->index);
121         page->mapping = NULL;
122         mapping->nrpages--;
123         pagecache_acct(-1);
124 }
125
126 void remove_from_page_cache(struct page *page)
127 {
128         struct address_space *mapping = page->mapping;
129
130         BUG_ON(!PageLocked(page));
131
132         write_lock_irq(&mapping->tree_lock);
133         __remove_from_page_cache(page);
134         write_unlock_irq(&mapping->tree_lock);
135 }
136
137 static int sync_page(void *word)
138 {
139         struct address_space *mapping;
140         struct page *page;
141
142         page = container_of((unsigned long *)word, struct page, flags);
143
144         /*
145          * page_mapping() is being called without PG_locked held.
146          * Some knowledge of the state and use of the page is used to
147          * reduce the requirements down to a memory barrier.
148          * The danger here is of a stale page_mapping() return value
149          * indicating a struct address_space different from the one it's
150          * associated with when it is associated with one.
151          * After smp_mb(), it's either the correct page_mapping() for
152          * the page, or an old page_mapping() and the page's own
153          * page_mapping() has gone NULL.
154          * The ->sync_page() address_space operation must tolerate
155          * page_mapping() going NULL. By an amazing coincidence,
156          * this comes about because none of the users of the page
157          * in the ->sync_page() methods make essential use of the
158          * page_mapping(), merely passing the page down to the backing
159          * device's unplug functions when it's non-NULL, which in turn
160          * ignore it for all cases but swap, where only page_private(page) is
161          * of interest. When page_mapping() does go NULL, the entire
162          * call stack gracefully ignores the page and returns.
163          * -- wli
164          */
165         smp_mb();
166         mapping = page_mapping(page);
167         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
168                 mapping->a_ops->sync_page(page);
169         io_schedule();
170         return 0;
171 }
172
173 /**
174  * filemap_fdatawrite_range - start writeback against all of a mapping's
175  * dirty pages that lie within the byte offsets <start, end>
176  * @mapping:    address space structure to write
177  * @start:      offset in bytes where the range starts
178  * @end:        offset in bytes where the range ends (inclusive)
179  * @sync_mode:  enable synchronous operation
180  *
181  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
182  * opposed to a regular memory * cleansing writeback.  The difference between
183  * these two operations is that if a dirty page/buffer is encountered, it must
184  * be waited upon, and not just skipped over.
185  */
186 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
187                                 loff_t end, int sync_mode)
188 {
189         int ret;
190         struct writeback_control wbc = {
191                 .sync_mode = sync_mode,
192                 .nr_to_write = mapping->nrpages * 2,
193                 .start = start,
194                 .end = end,
195         };
196
197         if (!mapping_cap_writeback_dirty(mapping))
198                 return 0;
199
200         ret = do_writepages(mapping, &wbc);
201         return ret;
202 }
203
204 static inline int __filemap_fdatawrite(struct address_space *mapping,
205         int sync_mode)
206 {
207         return __filemap_fdatawrite_range(mapping, 0, 0, sync_mode);
208 }
209
210 int filemap_fdatawrite(struct address_space *mapping)
211 {
212         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
213 }
214 EXPORT_SYMBOL(filemap_fdatawrite);
215
216 static int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
217                                 loff_t end)
218 {
219         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
220 }
221
222 /*
223  * This is a mostly non-blocking flush.  Not suitable for data-integrity
224  * purposes - I/O may not be started against all dirty pages.
225  */
226 int filemap_flush(struct address_space *mapping)
227 {
228         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
229 }
230 EXPORT_SYMBOL(filemap_flush);
231
232 /*
233  * Wait for writeback to complete against pages indexed by start->end
234  * inclusive
235  */
236 int wait_on_page_writeback_range(struct address_space *mapping,
237                                 pgoff_t start, pgoff_t end)
238 {
239         struct pagevec pvec;
240         int nr_pages;
241         int ret = 0;
242         pgoff_t index;
243
244         if (end < start)
245                 return 0;
246
247         pagevec_init(&pvec, 0);
248         index = start;
249         while ((index <= end) &&
250                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
251                         PAGECACHE_TAG_WRITEBACK,
252                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
253                 unsigned i;
254
255                 for (i = 0; i < nr_pages; i++) {
256                         struct page *page = pvec.pages[i];
257
258                         /* until radix tree lookup accepts end_index */
259                         if (page->index > end)
260                                 continue;
261
262                         wait_on_page_writeback(page);
263                         if (PageError(page))
264                                 ret = -EIO;
265                 }
266                 pagevec_release(&pvec);
267                 cond_resched();
268         }
269
270         /* Check for outstanding write errors */
271         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
272                 ret = -ENOSPC;
273         if (test_and_clear_bit(AS_EIO, &mapping->flags))
274                 ret = -EIO;
275
276         return ret;
277 }
278
279 /*
280  * Write and wait upon all the pages in the passed range.  This is a "data
281  * integrity" operation.  It waits upon in-flight writeout before starting and
282  * waiting upon new writeout.  If there was an IO error, return it.
283  *
284  * We need to re-take i_mutex during the generic_osync_inode list walk because
285  * it is otherwise livelockable.
286  */
287 int sync_page_range(struct inode *inode, struct address_space *mapping,
288                         loff_t pos, loff_t count)
289 {
290         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
291         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
292         int ret;
293
294         if (!mapping_cap_writeback_dirty(mapping) || !count)
295                 return 0;
296         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
297         if (ret == 0) {
298                 mutex_lock(&inode->i_mutex);
299                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
300                 mutex_unlock(&inode->i_mutex);
301         }
302         if (ret == 0)
303                 ret = wait_on_page_writeback_range(mapping, start, end);
304         return ret;
305 }
306 EXPORT_SYMBOL(sync_page_range);
307
308 /*
309  * Note: Holding i_mutex across sync_page_range_nolock is not a good idea
310  * as it forces O_SYNC writers to different parts of the same file
311  * to be serialised right until io completion.
312  */
313 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
314                            loff_t pos, loff_t count)
315 {
316         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
317         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
318         int ret;
319
320         if (!mapping_cap_writeback_dirty(mapping) || !count)
321                 return 0;
322         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
323         if (ret == 0)
324                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
325         if (ret == 0)
326                 ret = wait_on_page_writeback_range(mapping, start, end);
327         return ret;
328 }
329 EXPORT_SYMBOL(sync_page_range_nolock);
330
331 /**
332  * filemap_fdatawait - walk the list of under-writeback pages of the given
333  *     address space and wait for all of them.
334  *
335  * @mapping: address space structure to wait for
336  */
337 int filemap_fdatawait(struct address_space *mapping)
338 {
339         loff_t i_size = i_size_read(mapping->host);
340
341         if (i_size == 0)
342                 return 0;
343
344         return wait_on_page_writeback_range(mapping, 0,
345                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
346 }
347 EXPORT_SYMBOL(filemap_fdatawait);
348
349 int filemap_write_and_wait(struct address_space *mapping)
350 {
351         int err = 0;
352
353         if (mapping->nrpages) {
354                 err = filemap_fdatawrite(mapping);
355                 /*
356                  * Even if the above returned error, the pages may be
357                  * written partially (e.g. -ENOSPC), so we wait for it.
358                  * But the -EIO is special case, it may indicate the worst
359                  * thing (e.g. bug) happened, so we avoid waiting for it.
360                  */
361                 if (err != -EIO) {
362                         int err2 = filemap_fdatawait(mapping);
363                         if (!err)
364                                 err = err2;
365                 }
366         }
367         return err;
368 }
369 EXPORT_SYMBOL(filemap_write_and_wait);
370
371 /*
372  * Write out and wait upon file offsets lstart->lend, inclusive.
373  *
374  * Note that `lend' is inclusive (describes the last byte to be written) so
375  * that this function can be used to write to the very end-of-file (end = -1).
376  */
377 int filemap_write_and_wait_range(struct address_space *mapping,
378                                  loff_t lstart, loff_t lend)
379 {
380         int err = 0;
381
382         if (mapping->nrpages) {
383                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
384                                                  WB_SYNC_ALL);
385                 /* See comment of filemap_write_and_wait() */
386                 if (err != -EIO) {
387                         int err2 = wait_on_page_writeback_range(mapping,
388                                                 lstart >> PAGE_CACHE_SHIFT,
389                                                 lend >> PAGE_CACHE_SHIFT);
390                         if (!err)
391                                 err = err2;
392                 }
393         }
394         return err;
395 }
396
397 /*
398  * This function is used to add newly allocated pagecache pages:
399  * the page is new, so we can just run SetPageLocked() against it.
400  * The other page state flags were set by rmqueue().
401  *
402  * This function does not add the page to the LRU.  The caller must do that.
403  */
404 int add_to_page_cache(struct page *page, struct address_space *mapping,
405                 pgoff_t offset, gfp_t gfp_mask)
406 {
407         int error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
408
409         if (error == 0) {
410                 write_lock_irq(&mapping->tree_lock);
411                 error = radix_tree_insert(&mapping->page_tree, offset, page);
412                 if (!error) {
413                         page_cache_get(page);
414                         SetPageLocked(page);
415                         page->mapping = mapping;
416                         page->index = offset;
417                         mapping->nrpages++;
418                         pagecache_acct(1);
419                 }
420                 write_unlock_irq(&mapping->tree_lock);
421                 radix_tree_preload_end();
422         }
423         return error;
424 }
425
426 EXPORT_SYMBOL(add_to_page_cache);
427
428 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
429                                 pgoff_t offset, gfp_t gfp_mask)
430 {
431         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
432         if (ret == 0)
433                 lru_cache_add(page);
434         return ret;
435 }
436
437 #ifdef CONFIG_NUMA
438 struct page *page_cache_alloc(struct address_space *x)
439 {
440         if (cpuset_do_page_mem_spread()) {
441                 int n = cpuset_mem_spread_node();
442                 return alloc_pages_node(n, mapping_gfp_mask(x), 0);
443         }
444         return alloc_pages(mapping_gfp_mask(x), 0);
445 }
446 EXPORT_SYMBOL(page_cache_alloc);
447
448 struct page *page_cache_alloc_cold(struct address_space *x)
449 {
450         if (cpuset_do_page_mem_spread()) {
451                 int n = cpuset_mem_spread_node();
452                 return alloc_pages_node(n, mapping_gfp_mask(x)|__GFP_COLD, 0);
453         }
454         return alloc_pages(mapping_gfp_mask(x)|__GFP_COLD, 0);
455 }
456 EXPORT_SYMBOL(page_cache_alloc_cold);
457 #endif
458
459 /*
460  * In order to wait for pages to become available there must be
461  * waitqueues associated with pages. By using a hash table of
462  * waitqueues where the bucket discipline is to maintain all
463  * waiters on the same queue and wake all when any of the pages
464  * become available, and for the woken contexts to check to be
465  * sure the appropriate page became available, this saves space
466  * at a cost of "thundering herd" phenomena during rare hash
467  * collisions.
468  */
469 static wait_queue_head_t *page_waitqueue(struct page *page)
470 {
471         const struct zone *zone = page_zone(page);
472
473         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
474 }
475
476 static inline void wake_up_page(struct page *page, int bit)
477 {
478         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
479 }
480
481 void fastcall wait_on_page_bit(struct page *page, int bit_nr)
482 {
483         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
484
485         if (test_bit(bit_nr, &page->flags))
486                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
487                                                         TASK_UNINTERRUPTIBLE);
488 }
489 EXPORT_SYMBOL(wait_on_page_bit);
490
491 /**
492  * unlock_page() - unlock a locked page
493  *
494  * @page: the page
495  *
496  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
497  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
498  * mechananism between PageLocked pages and PageWriteback pages is shared.
499  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
500  *
501  * The first mb is necessary to safely close the critical section opened by the
502  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
503  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
504  * parallel wait_on_page_locked()).
505  */
506 void fastcall unlock_page(struct page *page)
507 {
508         smp_mb__before_clear_bit();
509         if (!TestClearPageLocked(page))
510                 BUG();
511         smp_mb__after_clear_bit(); 
512         wake_up_page(page, PG_locked);
513 }
514 EXPORT_SYMBOL(unlock_page);
515
516 /*
517  * End writeback against a page.
518  */
519 void end_page_writeback(struct page *page)
520 {
521         if (!TestClearPageReclaim(page) || rotate_reclaimable_page(page)) {
522                 if (!test_clear_page_writeback(page))
523                         BUG();
524         }
525         smp_mb__after_clear_bit();
526         wake_up_page(page, PG_writeback);
527 }
528 EXPORT_SYMBOL(end_page_writeback);
529
530 /*
531  * Get a lock on the page, assuming we need to sleep to get it.
532  *
533  * Ugly: running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
534  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
535  * chances are that on the second loop, the block layer's plug list is empty,
536  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
537  */
538 void fastcall __lock_page(struct page *page)
539 {
540         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
541
542         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
543                                                         TASK_UNINTERRUPTIBLE);
544 }
545 EXPORT_SYMBOL(__lock_page);
546
547 /*
548  * a rather lightweight function, finding and getting a reference to a
549  * hashed page atomically.
550  */
551 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
552 {
553         struct page *page;
554
555         read_lock_irq(&mapping->tree_lock);
556         page = radix_tree_lookup(&mapping->page_tree, offset);
557         if (page)
558                 page_cache_get(page);
559         read_unlock_irq(&mapping->tree_lock);
560         return page;
561 }
562
563 EXPORT_SYMBOL(find_get_page);
564
565 /*
566  * Same as above, but trylock it instead of incrementing the count.
567  */
568 struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
569 {
570         struct page *page;
571
572         read_lock_irq(&mapping->tree_lock);
573         page = radix_tree_lookup(&mapping->page_tree, offset);
574         if (page && TestSetPageLocked(page))
575                 page = NULL;
576         read_unlock_irq(&mapping->tree_lock);
577         return page;
578 }
579
580 EXPORT_SYMBOL(find_trylock_page);
581
582 /**
583  * find_lock_page - locate, pin and lock a pagecache page
584  *
585  * @mapping: the address_space to search
586  * @offset: the page index
587  *
588  * Locates the desired pagecache page, locks it, increments its reference
589  * count and returns its address.
590  *
591  * Returns zero if the page was not present. find_lock_page() may sleep.
592  */
593 struct page *find_lock_page(struct address_space *mapping,
594                                 unsigned long offset)
595 {
596         struct page *page;
597
598         read_lock_irq(&mapping->tree_lock);
599 repeat:
600         page = radix_tree_lookup(&mapping->page_tree, offset);
601         if (page) {
602                 page_cache_get(page);
603                 if (TestSetPageLocked(page)) {
604                         read_unlock_irq(&mapping->tree_lock);
605                         __lock_page(page);
606                         read_lock_irq(&mapping->tree_lock);
607
608                         /* Has the page been truncated while we slept? */
609                         if (unlikely(page->mapping != mapping ||
610                                      page->index != offset)) {
611                                 unlock_page(page);
612                                 page_cache_release(page);
613                                 goto repeat;
614                         }
615                 }
616         }
617         read_unlock_irq(&mapping->tree_lock);
618         return page;
619 }
620
621 EXPORT_SYMBOL(find_lock_page);
622
623 /**
624  * find_or_create_page - locate or add a pagecache page
625  *
626  * @mapping: the page's address_space
627  * @index: the page's index into the mapping
628  * @gfp_mask: page allocation mode
629  *
630  * Locates a page in the pagecache.  If the page is not present, a new page
631  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
632  * LRU list.  The returned page is locked and has its reference count
633  * incremented.
634  *
635  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
636  * allocation!
637  *
638  * find_or_create_page() returns the desired page's address, or zero on
639  * memory exhaustion.
640  */
641 struct page *find_or_create_page(struct address_space *mapping,
642                 unsigned long index, gfp_t gfp_mask)
643 {
644         struct page *page, *cached_page = NULL;
645         int err;
646 repeat:
647         page = find_lock_page(mapping, index);
648         if (!page) {
649                 if (!cached_page) {
650                         cached_page = alloc_page(gfp_mask);
651                         if (!cached_page)
652                                 return NULL;
653                 }
654                 err = add_to_page_cache_lru(cached_page, mapping,
655                                         index, gfp_mask);
656                 if (!err) {
657                         page = cached_page;
658                         cached_page = NULL;
659                 } else if (err == -EEXIST)
660                         goto repeat;
661         }
662         if (cached_page)
663                 page_cache_release(cached_page);
664         return page;
665 }
666
667 EXPORT_SYMBOL(find_or_create_page);
668
669 /**
670  * find_get_pages - gang pagecache lookup
671  * @mapping:    The address_space to search
672  * @start:      The starting page index
673  * @nr_pages:   The maximum number of pages
674  * @pages:      Where the resulting pages are placed
675  *
676  * find_get_pages() will search for and return a group of up to
677  * @nr_pages pages in the mapping.  The pages are placed at @pages.
678  * find_get_pages() takes a reference against the returned pages.
679  *
680  * The search returns a group of mapping-contiguous pages with ascending
681  * indexes.  There may be holes in the indices due to not-present pages.
682  *
683  * find_get_pages() returns the number of pages which were found.
684  */
685 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
686                             unsigned int nr_pages, struct page **pages)
687 {
688         unsigned int i;
689         unsigned int ret;
690
691         read_lock_irq(&mapping->tree_lock);
692         ret = radix_tree_gang_lookup(&mapping->page_tree,
693                                 (void **)pages, start, nr_pages);
694         for (i = 0; i < ret; i++)
695                 page_cache_get(pages[i]);
696         read_unlock_irq(&mapping->tree_lock);
697         return ret;
698 }
699
700 /**
701  * find_get_pages_contig - gang contiguous pagecache lookup
702  * @mapping:    The address_space to search
703  * @index:      The starting page index
704  * @nr_pages:   The maximum number of pages
705  * @pages:      Where the resulting pages are placed
706  *
707  * find_get_pages_contig() works exactly like find_get_pages(), except
708  * that the returned number of pages are guaranteed to be contiguous.
709  *
710  * find_get_pages_contig() returns the number of pages which were found.
711  */
712 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
713                                unsigned int nr_pages, struct page **pages)
714 {
715         unsigned int i;
716         unsigned int ret;
717
718         read_lock_irq(&mapping->tree_lock);
719         ret = radix_tree_gang_lookup(&mapping->page_tree,
720                                 (void **)pages, index, nr_pages);
721         for (i = 0; i < ret; i++) {
722                 if (pages[i]->mapping == NULL || pages[i]->index != index)
723                         break;
724
725                 page_cache_get(pages[i]);
726                 index++;
727         }
728         read_unlock_irq(&mapping->tree_lock);
729         return i;
730 }
731
732 /*
733  * Like find_get_pages, except we only return pages which are tagged with
734  * `tag'.   We update *index to index the next page for the traversal.
735  */
736 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
737                         int tag, unsigned int nr_pages, struct page **pages)
738 {
739         unsigned int i;
740         unsigned int ret;
741
742         read_lock_irq(&mapping->tree_lock);
743         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
744                                 (void **)pages, *index, nr_pages, tag);
745         for (i = 0; i < ret; i++)
746                 page_cache_get(pages[i]);
747         if (ret)
748                 *index = pages[ret - 1]->index + 1;
749         read_unlock_irq(&mapping->tree_lock);
750         return ret;
751 }
752
753 /*
754  * Same as grab_cache_page, but do not wait if the page is unavailable.
755  * This is intended for speculative data generators, where the data can
756  * be regenerated if the page couldn't be grabbed.  This routine should
757  * be safe to call while holding the lock for another page.
758  *
759  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
760  * and deadlock against the caller's locked page.
761  */
762 struct page *
763 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
764 {
765         struct page *page = find_get_page(mapping, index);
766         gfp_t gfp_mask;
767
768         if (page) {
769                 if (!TestSetPageLocked(page))
770                         return page;
771                 page_cache_release(page);
772                 return NULL;
773         }
774         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
775         page = alloc_pages(gfp_mask, 0);
776         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
777                 page_cache_release(page);
778                 page = NULL;
779         }
780         return page;
781 }
782
783 EXPORT_SYMBOL(grab_cache_page_nowait);
784
785 /*
786  * This is a generic file read routine, and uses the
787  * mapping->a_ops->readpage() function for the actual low-level
788  * stuff.
789  *
790  * This is really ugly. But the goto's actually try to clarify some
791  * of the logic when it comes to error handling etc.
792  *
793  * Note the struct file* is only passed for the use of readpage.  It may be
794  * NULL.
795  */
796 void do_generic_mapping_read(struct address_space *mapping,
797                              struct file_ra_state *_ra,
798                              struct file *filp,
799                              loff_t *ppos,
800                              read_descriptor_t *desc,
801                              read_actor_t actor)
802 {
803         struct inode *inode = mapping->host;
804         unsigned long index;
805         unsigned long end_index;
806         unsigned long offset;
807         unsigned long last_index;
808         unsigned long next_index;
809         unsigned long prev_index;
810         loff_t isize;
811         struct page *cached_page;
812         int error;
813         struct file_ra_state ra = *_ra;
814
815         cached_page = NULL;
816         index = *ppos >> PAGE_CACHE_SHIFT;
817         next_index = index;
818         prev_index = ra.prev_page;
819         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
820         offset = *ppos & ~PAGE_CACHE_MASK;
821
822         isize = i_size_read(inode);
823         if (!isize)
824                 goto out;
825
826         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
827         for (;;) {
828                 struct page *page;
829                 unsigned long nr, ret;
830
831                 /* nr is the maximum number of bytes to copy from this page */
832                 nr = PAGE_CACHE_SIZE;
833                 if (index >= end_index) {
834                         if (index > end_index)
835                                 goto out;
836                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
837                         if (nr <= offset) {
838                                 goto out;
839                         }
840                 }
841                 nr = nr - offset;
842
843                 cond_resched();
844                 if (index == next_index)
845                         next_index = page_cache_readahead(mapping, &ra, filp,
846                                         index, last_index - index);
847
848 find_page:
849                 page = find_get_page(mapping, index);
850                 if (unlikely(page == NULL)) {
851                         handle_ra_miss(mapping, &ra, index);
852                         goto no_cached_page;
853                 }
854                 if (!PageUptodate(page))
855                         goto page_not_up_to_date;
856 page_ok:
857
858                 /* If users can be writing to this page using arbitrary
859                  * virtual addresses, take care about potential aliasing
860                  * before reading the page on the kernel side.
861                  */
862                 if (mapping_writably_mapped(mapping))
863                         flush_dcache_page(page);
864
865                 /*
866                  * When (part of) the same page is read multiple times
867                  * in succession, only mark it as accessed the first time.
868                  */
869                 if (prev_index != index)
870                         mark_page_accessed(page);
871                 prev_index = index;
872
873                 /*
874                  * Ok, we have the page, and it's up-to-date, so
875                  * now we can copy it to user space...
876                  *
877                  * The actor routine returns how many bytes were actually used..
878                  * NOTE! This may not be the same as how much of a user buffer
879                  * we filled up (we may be padding etc), so we can only update
880                  * "pos" here (the actor routine has to update the user buffer
881                  * pointers and the remaining count).
882                  */
883                 ret = actor(desc, page, offset, nr);
884                 offset += ret;
885                 index += offset >> PAGE_CACHE_SHIFT;
886                 offset &= ~PAGE_CACHE_MASK;
887
888                 page_cache_release(page);
889                 if (ret == nr && desc->count)
890                         continue;
891                 goto out;
892
893 page_not_up_to_date:
894                 /* Get exclusive access to the page ... */
895                 lock_page(page);
896
897                 /* Did it get unhashed before we got the lock? */
898                 if (!page->mapping) {
899                         unlock_page(page);
900                         page_cache_release(page);
901                         continue;
902                 }
903
904                 /* Did somebody else fill it already? */
905                 if (PageUptodate(page)) {
906                         unlock_page(page);
907                         goto page_ok;
908                 }
909
910 readpage:
911                 /* Start the actual read. The read will unlock the page. */
912                 error = mapping->a_ops->readpage(filp, page);
913
914                 if (unlikely(error)) {
915                         if (error == AOP_TRUNCATED_PAGE) {
916                                 page_cache_release(page);
917                                 goto find_page;
918                         }
919                         goto readpage_error;
920                 }
921
922                 if (!PageUptodate(page)) {
923                         lock_page(page);
924                         if (!PageUptodate(page)) {
925                                 if (page->mapping == NULL) {
926                                         /*
927                                          * invalidate_inode_pages got it
928                                          */
929                                         unlock_page(page);
930                                         page_cache_release(page);
931                                         goto find_page;
932                                 }
933                                 unlock_page(page);
934                                 error = -EIO;
935                                 goto readpage_error;
936                         }
937                         unlock_page(page);
938                 }
939
940                 /*
941                  * i_size must be checked after we have done ->readpage.
942                  *
943                  * Checking i_size after the readpage allows us to calculate
944                  * the correct value for "nr", which means the zero-filled
945                  * part of the page is not copied back to userspace (unless
946                  * another truncate extends the file - this is desired though).
947                  */
948                 isize = i_size_read(inode);
949                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
950                 if (unlikely(!isize || index > end_index)) {
951                         page_cache_release(page);
952                         goto out;
953                 }
954
955                 /* nr is the maximum number of bytes to copy from this page */
956                 nr = PAGE_CACHE_SIZE;
957                 if (index == end_index) {
958                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
959                         if (nr <= offset) {
960                                 page_cache_release(page);
961                                 goto out;
962                         }
963                 }
964                 nr = nr - offset;
965                 goto page_ok;
966
967 readpage_error:
968                 /* UHHUH! A synchronous read error occurred. Report it */
969                 desc->error = error;
970                 page_cache_release(page);
971                 goto out;
972
973 no_cached_page:
974                 /*
975                  * Ok, it wasn't cached, so we need to create a new
976                  * page..
977                  */
978                 if (!cached_page) {
979                         cached_page = page_cache_alloc_cold(mapping);
980                         if (!cached_page) {
981                                 desc->error = -ENOMEM;
982                                 goto out;
983                         }
984                 }
985                 error = add_to_page_cache_lru(cached_page, mapping,
986                                                 index, GFP_KERNEL);
987                 if (error) {
988                         if (error == -EEXIST)
989                                 goto find_page;
990                         desc->error = error;
991                         goto out;
992                 }
993                 page = cached_page;
994                 cached_page = NULL;
995                 goto readpage;
996         }
997
998 out:
999         *_ra = ra;
1000
1001         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
1002         if (cached_page)
1003                 page_cache_release(cached_page);
1004         if (filp)
1005                 file_accessed(filp);
1006 }
1007
1008 EXPORT_SYMBOL(do_generic_mapping_read);
1009
1010 int file_read_actor(read_descriptor_t *desc, struct page *page,
1011                         unsigned long offset, unsigned long size)
1012 {
1013         char *kaddr;
1014         unsigned long left, count = desc->count;
1015
1016         if (size > count)
1017                 size = count;
1018
1019         /*
1020          * Faults on the destination of a read are common, so do it before
1021          * taking the kmap.
1022          */
1023         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1024                 kaddr = kmap_atomic(page, KM_USER0);
1025                 left = __copy_to_user_inatomic(desc->arg.buf,
1026                                                 kaddr + offset, size);
1027                 kunmap_atomic(kaddr, KM_USER0);
1028                 if (left == 0)
1029                         goto success;
1030         }
1031
1032         /* Do it the slow way */
1033         kaddr = kmap(page);
1034         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1035         kunmap(page);
1036
1037         if (left) {
1038                 size -= left;
1039                 desc->error = -EFAULT;
1040         }
1041 success:
1042         desc->count = count - size;
1043         desc->written += size;
1044         desc->arg.buf += size;
1045         return size;
1046 }
1047 EXPORT_SYMBOL(file_read_actor);
1048
1049 /*
1050  * This is the "read()" routine for all filesystems
1051  * that can use the page cache directly.
1052  */
1053 ssize_t
1054 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1055                 unsigned long nr_segs, loff_t *ppos)
1056 {
1057         struct file *filp = iocb->ki_filp;
1058         ssize_t retval;
1059         unsigned long seg;
1060         size_t count;
1061
1062         count = 0;
1063         for (seg = 0; seg < nr_segs; seg++) {
1064                 const struct iovec *iv = &iov[seg];
1065
1066                 /*
1067                  * If any segment has a negative length, or the cumulative
1068                  * length ever wraps negative then return -EINVAL.
1069                  */
1070                 count += iv->iov_len;
1071                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1072                         return -EINVAL;
1073                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1074                         continue;
1075                 if (seg == 0)
1076                         return -EFAULT;
1077                 nr_segs = seg;
1078                 count -= iv->iov_len;   /* This segment is no good */
1079                 break;
1080         }
1081
1082         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1083         if (filp->f_flags & O_DIRECT) {
1084                 loff_t pos = *ppos, size;
1085                 struct address_space *mapping;
1086                 struct inode *inode;
1087
1088                 mapping = filp->f_mapping;
1089                 inode = mapping->host;
1090                 retval = 0;
1091                 if (!count)
1092                         goto out; /* skip atime */
1093                 size = i_size_read(inode);
1094                 if (pos < size) {
1095                         retval = generic_file_direct_IO(READ, iocb,
1096                                                 iov, pos, nr_segs);
1097                         if (retval > 0 && !is_sync_kiocb(iocb))
1098                                 retval = -EIOCBQUEUED;
1099                         if (retval > 0)
1100                                 *ppos = pos + retval;
1101                 }
1102                 file_accessed(filp);
1103                 goto out;
1104         }
1105
1106         retval = 0;
1107         if (count) {
1108                 for (seg = 0; seg < nr_segs; seg++) {
1109                         read_descriptor_t desc;
1110
1111                         desc.written = 0;
1112                         desc.arg.buf = iov[seg].iov_base;
1113                         desc.count = iov[seg].iov_len;
1114                         if (desc.count == 0)
1115                                 continue;
1116                         desc.error = 0;
1117                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1118                         retval += desc.written;
1119                         if (desc.error) {
1120                                 retval = retval ?: desc.error;
1121                                 break;
1122                         }
1123                 }
1124         }
1125 out:
1126         return retval;
1127 }
1128
1129 EXPORT_SYMBOL(__generic_file_aio_read);
1130
1131 ssize_t
1132 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1133 {
1134         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1135
1136         BUG_ON(iocb->ki_pos != pos);
1137         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1138 }
1139
1140 EXPORT_SYMBOL(generic_file_aio_read);
1141
1142 ssize_t
1143 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1144 {
1145         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1146         struct kiocb kiocb;
1147         ssize_t ret;
1148
1149         init_sync_kiocb(&kiocb, filp);
1150         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1151         if (-EIOCBQUEUED == ret)
1152                 ret = wait_on_sync_kiocb(&kiocb);
1153         return ret;
1154 }
1155
1156 EXPORT_SYMBOL(generic_file_read);
1157
1158 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1159 {
1160         ssize_t written;
1161         unsigned long count = desc->count;
1162         struct file *file = desc->arg.data;
1163
1164         if (size > count)
1165                 size = count;
1166
1167         written = file->f_op->sendpage(file, page, offset,
1168                                        size, &file->f_pos, size<count);
1169         if (written < 0) {
1170                 desc->error = written;
1171                 written = 0;
1172         }
1173         desc->count = count - written;
1174         desc->written += written;
1175         return written;
1176 }
1177
1178 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1179                          size_t count, read_actor_t actor, void *target)
1180 {
1181         read_descriptor_t desc;
1182
1183         if (!count)
1184                 return 0;
1185
1186         desc.written = 0;
1187         desc.count = count;
1188         desc.arg.data = target;
1189         desc.error = 0;
1190
1191         do_generic_file_read(in_file, ppos, &desc, actor);
1192         if (desc.written)
1193                 return desc.written;
1194         return desc.error;
1195 }
1196
1197 EXPORT_SYMBOL(generic_file_sendfile);
1198
1199 static ssize_t
1200 do_readahead(struct address_space *mapping, struct file *filp,
1201              unsigned long index, unsigned long nr)
1202 {
1203         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1204                 return -EINVAL;
1205
1206         force_page_cache_readahead(mapping, filp, index,
1207                                         max_sane_readahead(nr));
1208         return 0;
1209 }
1210
1211 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1212 {
1213         ssize_t ret;
1214         struct file *file;
1215
1216         ret = -EBADF;
1217         file = fget(fd);
1218         if (file) {
1219                 if (file->f_mode & FMODE_READ) {
1220                         struct address_space *mapping = file->f_mapping;
1221                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1222                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1223                         unsigned long len = end - start + 1;
1224                         ret = do_readahead(mapping, file, start, len);
1225                 }
1226                 fput(file);
1227         }
1228         return ret;
1229 }
1230
1231 #ifdef CONFIG_MMU
1232 /*
1233  * This adds the requested page to the page cache if it isn't already there,
1234  * and schedules an I/O to read in its contents from disk.
1235  */
1236 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1237 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1238 {
1239         struct address_space *mapping = file->f_mapping;
1240         struct page *page; 
1241         int ret;
1242
1243         do {
1244                 page = page_cache_alloc_cold(mapping);
1245                 if (!page)
1246                         return -ENOMEM;
1247
1248                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1249                 if (ret == 0)
1250                         ret = mapping->a_ops->readpage(file, page);
1251                 else if (ret == -EEXIST)
1252                         ret = 0; /* losing race to add is OK */
1253
1254                 page_cache_release(page);
1255
1256         } while (ret == AOP_TRUNCATED_PAGE);
1257                 
1258         return ret;
1259 }
1260
1261 #define MMAP_LOTSAMISS  (100)
1262
1263 /*
1264  * filemap_nopage() is invoked via the vma operations vector for a
1265  * mapped memory region to read in file data during a page fault.
1266  *
1267  * The goto's are kind of ugly, but this streamlines the normal case of having
1268  * it in the page cache, and handles the special cases reasonably without
1269  * having a lot of duplicated code.
1270  */
1271 struct page *filemap_nopage(struct vm_area_struct *area,
1272                                 unsigned long address, int *type)
1273 {
1274         int error;
1275         struct file *file = area->vm_file;
1276         struct address_space *mapping = file->f_mapping;
1277         struct file_ra_state *ra = &file->f_ra;
1278         struct inode *inode = mapping->host;
1279         struct page *page;
1280         unsigned long size, pgoff;
1281         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1282
1283         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1284
1285 retry_all:
1286         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1287         if (pgoff >= size)
1288                 goto outside_data_content;
1289
1290         /* If we don't want any read-ahead, don't bother */
1291         if (VM_RandomReadHint(area))
1292                 goto no_cached_page;
1293
1294         /*
1295          * The readahead code wants to be told about each and every page
1296          * so it can build and shrink its windows appropriately
1297          *
1298          * For sequential accesses, we use the generic readahead logic.
1299          */
1300         if (VM_SequentialReadHint(area))
1301                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1302
1303         /*
1304          * Do we have something in the page cache already?
1305          */
1306 retry_find:
1307         page = find_get_page(mapping, pgoff);
1308         if (!page) {
1309                 unsigned long ra_pages;
1310
1311                 if (VM_SequentialReadHint(area)) {
1312                         handle_ra_miss(mapping, ra, pgoff);
1313                         goto no_cached_page;
1314                 }
1315                 ra->mmap_miss++;
1316
1317                 /*
1318                  * Do we miss much more than hit in this file? If so,
1319                  * stop bothering with read-ahead. It will only hurt.
1320                  */
1321                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1322                         goto no_cached_page;
1323
1324                 /*
1325                  * To keep the pgmajfault counter straight, we need to
1326                  * check did_readaround, as this is an inner loop.
1327                  */
1328                 if (!did_readaround) {
1329                         majmin = VM_FAULT_MAJOR;
1330                         inc_page_state(pgmajfault);
1331                 }
1332                 did_readaround = 1;
1333                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1334                 if (ra_pages) {
1335                         pgoff_t start = 0;
1336
1337                         if (pgoff > ra_pages / 2)
1338                                 start = pgoff - ra_pages / 2;
1339                         do_page_cache_readahead(mapping, file, start, ra_pages);
1340                 }
1341                 page = find_get_page(mapping, pgoff);
1342                 if (!page)
1343                         goto no_cached_page;
1344         }
1345
1346         if (!did_readaround)
1347                 ra->mmap_hit++;
1348
1349         /*
1350          * Ok, found a page in the page cache, now we need to check
1351          * that it's up-to-date.
1352          */
1353         if (!PageUptodate(page))
1354                 goto page_not_uptodate;
1355
1356 success:
1357         /*
1358          * Found the page and have a reference on it.
1359          */
1360         mark_page_accessed(page);
1361         if (type)
1362                 *type = majmin;
1363         return page;
1364
1365 outside_data_content:
1366         /*
1367          * An external ptracer can access pages that normally aren't
1368          * accessible..
1369          */
1370         if (area->vm_mm == current->mm)
1371                 return NULL;
1372         /* Fall through to the non-read-ahead case */
1373 no_cached_page:
1374         /*
1375          * We're only likely to ever get here if MADV_RANDOM is in
1376          * effect.
1377          */
1378         error = page_cache_read(file, pgoff);
1379         grab_swap_token();
1380
1381         /*
1382          * The page we want has now been added to the page cache.
1383          * In the unlikely event that someone removed it in the
1384          * meantime, we'll just come back here and read it again.
1385          */
1386         if (error >= 0)
1387                 goto retry_find;
1388
1389         /*
1390          * An error return from page_cache_read can result if the
1391          * system is low on memory, or a problem occurs while trying
1392          * to schedule I/O.
1393          */
1394         if (error == -ENOMEM)
1395                 return NOPAGE_OOM;
1396         return NULL;
1397
1398 page_not_uptodate:
1399         if (!did_readaround) {
1400                 majmin = VM_FAULT_MAJOR;
1401                 inc_page_state(pgmajfault);
1402         }
1403         lock_page(page);
1404
1405         /* Did it get unhashed while we waited for it? */
1406         if (!page->mapping) {
1407                 unlock_page(page);
1408                 page_cache_release(page);
1409                 goto retry_all;
1410         }
1411
1412         /* Did somebody else get it up-to-date? */
1413         if (PageUptodate(page)) {
1414                 unlock_page(page);
1415                 goto success;
1416         }
1417
1418         error = mapping->a_ops->readpage(file, page);
1419         if (!error) {
1420                 wait_on_page_locked(page);
1421                 if (PageUptodate(page))
1422                         goto success;
1423         } else if (error == AOP_TRUNCATED_PAGE) {
1424                 page_cache_release(page);
1425                 goto retry_find;
1426         }
1427
1428         /*
1429          * Umm, take care of errors if the page isn't up-to-date.
1430          * Try to re-read it _once_. We do this synchronously,
1431          * because there really aren't any performance issues here
1432          * and we need to check for errors.
1433          */
1434         lock_page(page);
1435
1436         /* Somebody truncated the page on us? */
1437         if (!page->mapping) {
1438                 unlock_page(page);
1439                 page_cache_release(page);
1440                 goto retry_all;
1441         }
1442
1443         /* Somebody else successfully read it in? */
1444         if (PageUptodate(page)) {
1445                 unlock_page(page);
1446                 goto success;
1447         }
1448         ClearPageError(page);
1449         error = mapping->a_ops->readpage(file, page);
1450         if (!error) {
1451                 wait_on_page_locked(page);
1452                 if (PageUptodate(page))
1453                         goto success;
1454         } else if (error == AOP_TRUNCATED_PAGE) {
1455                 page_cache_release(page);
1456                 goto retry_find;
1457         }
1458
1459         /*
1460          * Things didn't work out. Return zero to tell the
1461          * mm layer so, possibly freeing the page cache page first.
1462          */
1463         page_cache_release(page);
1464         return NULL;
1465 }
1466
1467 EXPORT_SYMBOL(filemap_nopage);
1468
1469 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1470                                         int nonblock)
1471 {
1472         struct address_space *mapping = file->f_mapping;
1473         struct page *page;
1474         int error;
1475
1476         /*
1477          * Do we have something in the page cache already?
1478          */
1479 retry_find:
1480         page = find_get_page(mapping, pgoff);
1481         if (!page) {
1482                 if (nonblock)
1483                         return NULL;
1484                 goto no_cached_page;
1485         }
1486
1487         /*
1488          * Ok, found a page in the page cache, now we need to check
1489          * that it's up-to-date.
1490          */
1491         if (!PageUptodate(page)) {
1492                 if (nonblock) {
1493                         page_cache_release(page);
1494                         return NULL;
1495                 }
1496                 goto page_not_uptodate;
1497         }
1498
1499 success:
1500         /*
1501          * Found the page and have a reference on it.
1502          */
1503         mark_page_accessed(page);
1504         return page;
1505
1506 no_cached_page:
1507         error = page_cache_read(file, pgoff);
1508
1509         /*
1510          * The page we want has now been added to the page cache.
1511          * In the unlikely event that someone removed it in the
1512          * meantime, we'll just come back here and read it again.
1513          */
1514         if (error >= 0)
1515                 goto retry_find;
1516
1517         /*
1518          * An error return from page_cache_read can result if the
1519          * system is low on memory, or a problem occurs while trying
1520          * to schedule I/O.
1521          */
1522         return NULL;
1523
1524 page_not_uptodate:
1525         lock_page(page);
1526
1527         /* Did it get unhashed while we waited for it? */
1528         if (!page->mapping) {
1529                 unlock_page(page);
1530                 goto err;
1531         }
1532
1533         /* Did somebody else get it up-to-date? */
1534         if (PageUptodate(page)) {
1535                 unlock_page(page);
1536                 goto success;
1537         }
1538
1539         error = mapping->a_ops->readpage(file, page);
1540         if (!error) {
1541                 wait_on_page_locked(page);
1542                 if (PageUptodate(page))
1543                         goto success;
1544         } else if (error == AOP_TRUNCATED_PAGE) {
1545                 page_cache_release(page);
1546                 goto retry_find;
1547         }
1548
1549         /*
1550          * Umm, take care of errors if the page isn't up-to-date.
1551          * Try to re-read it _once_. We do this synchronously,
1552          * because there really aren't any performance issues here
1553          * and we need to check for errors.
1554          */
1555         lock_page(page);
1556
1557         /* Somebody truncated the page on us? */
1558         if (!page->mapping) {
1559                 unlock_page(page);
1560                 goto err;
1561         }
1562         /* Somebody else successfully read it in? */
1563         if (PageUptodate(page)) {
1564                 unlock_page(page);
1565                 goto success;
1566         }
1567
1568         ClearPageError(page);
1569         error = mapping->a_ops->readpage(file, page);
1570         if (!error) {
1571                 wait_on_page_locked(page);
1572                 if (PageUptodate(page))
1573                         goto success;
1574         } else if (error == AOP_TRUNCATED_PAGE) {
1575                 page_cache_release(page);
1576                 goto retry_find;
1577         }
1578
1579         /*
1580          * Things didn't work out. Return zero to tell the
1581          * mm layer so, possibly freeing the page cache page first.
1582          */
1583 err:
1584         page_cache_release(page);
1585
1586         return NULL;
1587 }
1588
1589 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1590                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1591                 int nonblock)
1592 {
1593         struct file *file = vma->vm_file;
1594         struct address_space *mapping = file->f_mapping;
1595         struct inode *inode = mapping->host;
1596         unsigned long size;
1597         struct mm_struct *mm = vma->vm_mm;
1598         struct page *page;
1599         int err;
1600
1601         if (!nonblock)
1602                 force_page_cache_readahead(mapping, vma->vm_file,
1603                                         pgoff, len >> PAGE_CACHE_SHIFT);
1604
1605 repeat:
1606         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1607         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1608                 return -EINVAL;
1609
1610         page = filemap_getpage(file, pgoff, nonblock);
1611
1612         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1613          * done in shmem_populate calling shmem_getpage */
1614         if (!page && !nonblock)
1615                 return -ENOMEM;
1616
1617         if (page) {
1618                 err = install_page(mm, vma, addr, page, prot);
1619                 if (err) {
1620                         page_cache_release(page);
1621                         return err;
1622                 }
1623         } else if (vma->vm_flags & VM_NONLINEAR) {
1624                 /* No page was found just because we can't read it in now (being
1625                  * here implies nonblock != 0), but the page may exist, so set
1626                  * the PTE to fault it in later. */
1627                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1628                 if (err)
1629                         return err;
1630         }
1631
1632         len -= PAGE_SIZE;
1633         addr += PAGE_SIZE;
1634         pgoff++;
1635         if (len)
1636                 goto repeat;
1637
1638         return 0;
1639 }
1640 EXPORT_SYMBOL(filemap_populate);
1641
1642 struct vm_operations_struct generic_file_vm_ops = {
1643         .nopage         = filemap_nopage,
1644         .populate       = filemap_populate,
1645 };
1646
1647 /* This is used for a general mmap of a disk file */
1648
1649 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1650 {
1651         struct address_space *mapping = file->f_mapping;
1652
1653         if (!mapping->a_ops->readpage)
1654                 return -ENOEXEC;
1655         file_accessed(file);
1656         vma->vm_ops = &generic_file_vm_ops;
1657         return 0;
1658 }
1659
1660 /*
1661  * This is for filesystems which do not implement ->writepage.
1662  */
1663 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1664 {
1665         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1666                 return -EINVAL;
1667         return generic_file_mmap(file, vma);
1668 }
1669 #else
1670 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1671 {
1672         return -ENOSYS;
1673 }
1674 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1675 {
1676         return -ENOSYS;
1677 }
1678 #endif /* CONFIG_MMU */
1679
1680 EXPORT_SYMBOL(generic_file_mmap);
1681 EXPORT_SYMBOL(generic_file_readonly_mmap);
1682
1683 static inline struct page *__read_cache_page(struct address_space *mapping,
1684                                 unsigned long index,
1685                                 int (*filler)(void *,struct page*),
1686                                 void *data)
1687 {
1688         struct page *page, *cached_page = NULL;
1689         int err;
1690 repeat:
1691         page = find_get_page(mapping, index);
1692         if (!page) {
1693                 if (!cached_page) {
1694                         cached_page = page_cache_alloc_cold(mapping);
1695                         if (!cached_page)
1696                                 return ERR_PTR(-ENOMEM);
1697                 }
1698                 err = add_to_page_cache_lru(cached_page, mapping,
1699                                         index, GFP_KERNEL);
1700                 if (err == -EEXIST)
1701                         goto repeat;
1702                 if (err < 0) {
1703                         /* Presumably ENOMEM for radix tree node */
1704                         page_cache_release(cached_page);
1705                         return ERR_PTR(err);
1706                 }
1707                 page = cached_page;
1708                 cached_page = NULL;
1709                 err = filler(data, page);
1710                 if (err < 0) {
1711                         page_cache_release(page);
1712                         page = ERR_PTR(err);
1713                 }
1714         }
1715         if (cached_page)
1716                 page_cache_release(cached_page);
1717         return page;
1718 }
1719
1720 /*
1721  * Read into the page cache. If a page already exists,
1722  * and PageUptodate() is not set, try to fill the page.
1723  */
1724 struct page *read_cache_page(struct address_space *mapping,
1725                                 unsigned long index,
1726                                 int (*filler)(void *,struct page*),
1727                                 void *data)
1728 {
1729         struct page *page;
1730         int err;
1731
1732 retry:
1733         page = __read_cache_page(mapping, index, filler, data);
1734         if (IS_ERR(page))
1735                 goto out;
1736         mark_page_accessed(page);
1737         if (PageUptodate(page))
1738                 goto out;
1739
1740         lock_page(page);
1741         if (!page->mapping) {
1742                 unlock_page(page);
1743                 page_cache_release(page);
1744                 goto retry;
1745         }
1746         if (PageUptodate(page)) {
1747                 unlock_page(page);
1748                 goto out;
1749         }
1750         err = filler(data, page);
1751         if (err < 0) {
1752                 page_cache_release(page);
1753                 page = ERR_PTR(err);
1754         }
1755  out:
1756         return page;
1757 }
1758
1759 EXPORT_SYMBOL(read_cache_page);
1760
1761 /*
1762  * If the page was newly created, increment its refcount and add it to the
1763  * caller's lru-buffering pagevec.  This function is specifically for
1764  * generic_file_write().
1765  */
1766 static inline struct page *
1767 __grab_cache_page(struct address_space *mapping, unsigned long index,
1768                         struct page **cached_page, struct pagevec *lru_pvec)
1769 {
1770         int err;
1771         struct page *page;
1772 repeat:
1773         page = find_lock_page(mapping, index);
1774         if (!page) {
1775                 if (!*cached_page) {
1776                         *cached_page = page_cache_alloc(mapping);
1777                         if (!*cached_page)
1778                                 return NULL;
1779                 }
1780                 err = add_to_page_cache(*cached_page, mapping,
1781                                         index, GFP_KERNEL);
1782                 if (err == -EEXIST)
1783                         goto repeat;
1784                 if (err == 0) {
1785                         page = *cached_page;
1786                         page_cache_get(page);
1787                         if (!pagevec_add(lru_pvec, page))
1788                                 __pagevec_lru_add(lru_pvec);
1789                         *cached_page = NULL;
1790                 }
1791         }
1792         return page;
1793 }
1794
1795 /*
1796  * The logic we want is
1797  *
1798  *      if suid or (sgid and xgrp)
1799  *              remove privs
1800  */
1801 int remove_suid(struct dentry *dentry)
1802 {
1803         mode_t mode = dentry->d_inode->i_mode;
1804         int kill = 0;
1805         int result = 0;
1806
1807         /* suid always must be killed */
1808         if (unlikely(mode & S_ISUID))
1809                 kill = ATTR_KILL_SUID;
1810
1811         /*
1812          * sgid without any exec bits is just a mandatory locking mark; leave
1813          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1814          */
1815         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1816                 kill |= ATTR_KILL_SGID;
1817
1818         if (unlikely(kill && !capable(CAP_FSETID))) {
1819                 struct iattr newattrs;
1820
1821                 newattrs.ia_valid = ATTR_FORCE | kill;
1822                 result = notify_change(dentry, &newattrs);
1823         }
1824         return result;
1825 }
1826 EXPORT_SYMBOL(remove_suid);
1827
1828 size_t
1829 __filemap_copy_from_user_iovec(char *vaddr, 
1830                         const struct iovec *iov, size_t base, size_t bytes)
1831 {
1832         size_t copied = 0, left = 0;
1833
1834         while (bytes) {
1835                 char __user *buf = iov->iov_base + base;
1836                 int copy = min(bytes, iov->iov_len - base);
1837
1838                 base = 0;
1839                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1840                 copied += copy;
1841                 bytes -= copy;
1842                 vaddr += copy;
1843                 iov++;
1844
1845                 if (unlikely(left)) {
1846                         /* zero the rest of the target like __copy_from_user */
1847                         if (bytes)
1848                                 memset(vaddr, 0, bytes);
1849                         break;
1850                 }
1851         }
1852         return copied - left;
1853 }
1854
1855 /*
1856  * Performs necessary checks before doing a write
1857  *
1858  * Can adjust writing position aor amount of bytes to write.
1859  * Returns appropriate error code that caller should return or
1860  * zero in case that write should be allowed.
1861  */
1862 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1863 {
1864         struct inode *inode = file->f_mapping->host;
1865         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1866
1867         if (unlikely(*pos < 0))
1868                 return -EINVAL;
1869
1870         if (!isblk) {
1871                 /* FIXME: this is for backwards compatibility with 2.4 */
1872                 if (file->f_flags & O_APPEND)
1873                         *pos = i_size_read(inode);
1874
1875                 if (limit != RLIM_INFINITY) {
1876                         if (*pos >= limit) {
1877                                 send_sig(SIGXFSZ, current, 0);
1878                                 return -EFBIG;
1879                         }
1880                         if (*count > limit - (typeof(limit))*pos) {
1881                                 *count = limit - (typeof(limit))*pos;
1882                         }
1883                 }
1884         }
1885
1886         /*
1887          * LFS rule
1888          */
1889         if (unlikely(*pos + *count > MAX_NON_LFS &&
1890                                 !(file->f_flags & O_LARGEFILE))) {
1891                 if (*pos >= MAX_NON_LFS) {
1892                         send_sig(SIGXFSZ, current, 0);
1893                         return -EFBIG;
1894                 }
1895                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1896                         *count = MAX_NON_LFS - (unsigned long)*pos;
1897                 }
1898         }
1899
1900         /*
1901          * Are we about to exceed the fs block limit ?
1902          *
1903          * If we have written data it becomes a short write.  If we have
1904          * exceeded without writing data we send a signal and return EFBIG.
1905          * Linus frestrict idea will clean these up nicely..
1906          */
1907         if (likely(!isblk)) {
1908                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1909                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1910                                 send_sig(SIGXFSZ, current, 0);
1911                                 return -EFBIG;
1912                         }
1913                         /* zero-length writes at ->s_maxbytes are OK */
1914                 }
1915
1916                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1917                         *count = inode->i_sb->s_maxbytes - *pos;
1918         } else {
1919                 loff_t isize;
1920                 if (bdev_read_only(I_BDEV(inode)))
1921                         return -EPERM;
1922                 isize = i_size_read(inode);
1923                 if (*pos >= isize) {
1924                         if (*count || *pos > isize)
1925                                 return -ENOSPC;
1926                 }
1927
1928                 if (*pos + *count > isize)
1929                         *count = isize - *pos;
1930         }
1931         return 0;
1932 }
1933 EXPORT_SYMBOL(generic_write_checks);
1934
1935 ssize_t
1936 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1937                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1938                 size_t count, size_t ocount)
1939 {
1940         struct file     *file = iocb->ki_filp;
1941         struct address_space *mapping = file->f_mapping;
1942         struct inode    *inode = mapping->host;
1943         ssize_t         written;
1944
1945         if (count != ocount)
1946                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1947
1948         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1949         if (written > 0) {
1950                 loff_t end = pos + written;
1951                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1952                         i_size_write(inode,  end);
1953                         mark_inode_dirty(inode);
1954                 }
1955                 *ppos = end;
1956         }
1957
1958         /*
1959          * Sync the fs metadata but not the minor inode changes and
1960          * of course not the data as we did direct DMA for the IO.
1961          * i_mutex is held, which protects generic_osync_inode() from
1962          * livelocking.
1963          */
1964         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1965                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1966                 if (err < 0)
1967                         written = err;
1968         }
1969         if (written == count && !is_sync_kiocb(iocb))
1970                 written = -EIOCBQUEUED;
1971         return written;
1972 }
1973 EXPORT_SYMBOL(generic_file_direct_write);
1974
1975 ssize_t
1976 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1977                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1978                 size_t count, ssize_t written)
1979 {
1980         struct file *file = iocb->ki_filp;
1981         struct address_space * mapping = file->f_mapping;
1982         struct address_space_operations *a_ops = mapping->a_ops;
1983         struct inode    *inode = mapping->host;
1984         long            status = 0;
1985         struct page     *page;
1986         struct page     *cached_page = NULL;
1987         size_t          bytes;
1988         struct pagevec  lru_pvec;
1989         const struct iovec *cur_iov = iov; /* current iovec */
1990         size_t          iov_base = 0;      /* offset in the current iovec */
1991         char __user     *buf;
1992
1993         pagevec_init(&lru_pvec, 0);
1994
1995         /*
1996          * handle partial DIO write.  Adjust cur_iov if needed.
1997          */
1998         if (likely(nr_segs == 1))
1999                 buf = iov->iov_base + written;
2000         else {
2001                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
2002                 buf = cur_iov->iov_base + iov_base;
2003         }
2004
2005         do {
2006                 unsigned long index;
2007                 unsigned long offset;
2008                 unsigned long maxlen;
2009                 size_t copied;
2010
2011                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
2012                 index = pos >> PAGE_CACHE_SHIFT;
2013                 bytes = PAGE_CACHE_SIZE - offset;
2014                 if (bytes > count)
2015                         bytes = count;
2016
2017                 /*
2018                  * Bring in the user page that we will copy from _first_.
2019                  * Otherwise there's a nasty deadlock on copying from the
2020                  * same page as we're writing to, without it being marked
2021                  * up-to-date.
2022                  */
2023                 maxlen = cur_iov->iov_len - iov_base;
2024                 if (maxlen > bytes)
2025                         maxlen = bytes;
2026                 fault_in_pages_readable(buf, maxlen);
2027
2028                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
2029                 if (!page) {
2030                         status = -ENOMEM;
2031                         break;
2032                 }
2033
2034                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2035                 if (unlikely(status)) {
2036                         loff_t isize = i_size_read(inode);
2037
2038                         if (status != AOP_TRUNCATED_PAGE)
2039                                 unlock_page(page);
2040                         page_cache_release(page);
2041                         if (status == AOP_TRUNCATED_PAGE)
2042                                 continue;
2043                         /*
2044                          * prepare_write() may have instantiated a few blocks
2045                          * outside i_size.  Trim these off again.
2046                          */
2047                         if (pos + bytes > isize)
2048                                 vmtruncate(inode, isize);
2049                         break;
2050                 }
2051                 if (likely(nr_segs == 1))
2052                         copied = filemap_copy_from_user(page, offset,
2053                                                         buf, bytes);
2054                 else
2055                         copied = filemap_copy_from_user_iovec(page, offset,
2056                                                 cur_iov, iov_base, bytes);
2057                 flush_dcache_page(page);
2058                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2059                 if (status == AOP_TRUNCATED_PAGE) {
2060                         page_cache_release(page);
2061                         continue;
2062                 }
2063                 if (likely(copied > 0)) {
2064                         if (!status)
2065                                 status = copied;
2066
2067                         if (status >= 0) {
2068                                 written += status;
2069                                 count -= status;
2070                                 pos += status;
2071                                 buf += status;
2072                                 if (unlikely(nr_segs > 1)) {
2073                                         filemap_set_next_iovec(&cur_iov,
2074                                                         &iov_base, status);
2075                                         if (count)
2076                                                 buf = cur_iov->iov_base +
2077                                                         iov_base;
2078                                 } else {
2079                                         iov_base += status;
2080                                 }
2081                         }
2082                 }
2083                 if (unlikely(copied != bytes))
2084                         if (status >= 0)
2085                                 status = -EFAULT;
2086                 unlock_page(page);
2087                 mark_page_accessed(page);
2088                 page_cache_release(page);
2089                 if (status < 0)
2090                         break;
2091                 balance_dirty_pages_ratelimited(mapping);
2092                 cond_resched();
2093         } while (count);
2094         *ppos = pos;
2095
2096         if (cached_page)
2097                 page_cache_release(cached_page);
2098
2099         /*
2100          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2101          */
2102         if (likely(status >= 0)) {
2103                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2104                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2105                                 status = generic_osync_inode(inode, mapping,
2106                                                 OSYNC_METADATA|OSYNC_DATA);
2107                 }
2108         }
2109         
2110         /*
2111          * If we get here for O_DIRECT writes then we must have fallen through
2112          * to buffered writes (block instantiation inside i_size).  So we sync
2113          * the file data here, to try to honour O_DIRECT expectations.
2114          */
2115         if (unlikely(file->f_flags & O_DIRECT) && written)
2116                 status = filemap_write_and_wait(mapping);
2117
2118         pagevec_lru_add(&lru_pvec);
2119         return written ? written : status;
2120 }
2121 EXPORT_SYMBOL(generic_file_buffered_write);
2122
2123 static ssize_t
2124 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2125                                 unsigned long nr_segs, loff_t *ppos)
2126 {
2127         struct file *file = iocb->ki_filp;
2128         struct address_space * mapping = file->f_mapping;
2129         size_t ocount;          /* original count */
2130         size_t count;           /* after file limit checks */
2131         struct inode    *inode = mapping->host;
2132         unsigned long   seg;
2133         loff_t          pos;
2134         ssize_t         written;
2135         ssize_t         err;
2136
2137         ocount = 0;
2138         for (seg = 0; seg < nr_segs; seg++) {
2139                 const struct iovec *iv = &iov[seg];
2140
2141                 /*
2142                  * If any segment has a negative length, or the cumulative
2143                  * length ever wraps negative then return -EINVAL.
2144                  */
2145                 ocount += iv->iov_len;
2146                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2147                         return -EINVAL;
2148                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2149                         continue;
2150                 if (seg == 0)
2151                         return -EFAULT;
2152                 nr_segs = seg;
2153                 ocount -= iv->iov_len;  /* This segment is no good */
2154                 break;
2155         }
2156
2157         count = ocount;
2158         pos = *ppos;
2159
2160         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2161
2162         /* We can write back this queue in page reclaim */
2163         current->backing_dev_info = mapping->backing_dev_info;
2164         written = 0;
2165
2166         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2167         if (err)
2168                 goto out;
2169
2170         if (count == 0)
2171                 goto out;
2172
2173         err = remove_suid(file->f_dentry);
2174         if (err)
2175                 goto out;
2176
2177         file_update_time(file);
2178
2179         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2180         if (unlikely(file->f_flags & O_DIRECT)) {
2181                 written = generic_file_direct_write(iocb, iov,
2182                                 &nr_segs, pos, ppos, count, ocount);
2183                 if (written < 0 || written == count)
2184                         goto out;
2185                 /*
2186                  * direct-io write to a hole: fall through to buffered I/O
2187                  * for completing the rest of the request.
2188                  */
2189                 pos += written;
2190                 count -= written;
2191         }
2192
2193         written = generic_file_buffered_write(iocb, iov, nr_segs,
2194                         pos, ppos, count, written);
2195 out:
2196         current->backing_dev_info = NULL;
2197         return written ? written : err;
2198 }
2199 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2200
2201 ssize_t
2202 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2203                                 unsigned long nr_segs, loff_t *ppos)
2204 {
2205         struct file *file = iocb->ki_filp;
2206         struct address_space *mapping = file->f_mapping;
2207         struct inode *inode = mapping->host;
2208         ssize_t ret;
2209         loff_t pos = *ppos;
2210
2211         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2212
2213         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2214                 int err;
2215
2216                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2217                 if (err < 0)
2218                         ret = err;
2219         }
2220         return ret;
2221 }
2222
2223 static ssize_t
2224 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2225                                 unsigned long nr_segs, loff_t *ppos)
2226 {
2227         struct kiocb kiocb;
2228         ssize_t ret;
2229
2230         init_sync_kiocb(&kiocb, file);
2231         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2232         if (ret == -EIOCBQUEUED)
2233                 ret = wait_on_sync_kiocb(&kiocb);
2234         return ret;
2235 }
2236
2237 ssize_t
2238 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2239                                 unsigned long nr_segs, loff_t *ppos)
2240 {
2241         struct kiocb kiocb;
2242         ssize_t ret;
2243
2244         init_sync_kiocb(&kiocb, file);
2245         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2246         if (-EIOCBQUEUED == ret)
2247                 ret = wait_on_sync_kiocb(&kiocb);
2248         return ret;
2249 }
2250 EXPORT_SYMBOL(generic_file_write_nolock);
2251
2252 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2253                                size_t count, loff_t pos)
2254 {
2255         struct file *file = iocb->ki_filp;
2256         struct address_space *mapping = file->f_mapping;
2257         struct inode *inode = mapping->host;
2258         ssize_t ret;
2259         struct iovec local_iov = { .iov_base = (void __user *)buf,
2260                                         .iov_len = count };
2261
2262         BUG_ON(iocb->ki_pos != pos);
2263
2264         mutex_lock(&inode->i_mutex);
2265         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2266                                                 &iocb->ki_pos);
2267         mutex_unlock(&inode->i_mutex);
2268
2269         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2270                 ssize_t err;
2271
2272                 err = sync_page_range(inode, mapping, pos, ret);
2273                 if (err < 0)
2274                         ret = err;
2275         }
2276         return ret;
2277 }
2278 EXPORT_SYMBOL(generic_file_aio_write);
2279
2280 ssize_t generic_file_write(struct file *file, const char __user *buf,
2281                            size_t count, loff_t *ppos)
2282 {
2283         struct address_space *mapping = file->f_mapping;
2284         struct inode *inode = mapping->host;
2285         ssize_t ret;
2286         struct iovec local_iov = { .iov_base = (void __user *)buf,
2287                                         .iov_len = count };
2288
2289         mutex_lock(&inode->i_mutex);
2290         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2291         mutex_unlock(&inode->i_mutex);
2292
2293         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2294                 ssize_t err;
2295
2296                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2297                 if (err < 0)
2298                         ret = err;
2299         }
2300         return ret;
2301 }
2302 EXPORT_SYMBOL(generic_file_write);
2303
2304 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2305                         unsigned long nr_segs, loff_t *ppos)
2306 {
2307         struct kiocb kiocb;
2308         ssize_t ret;
2309
2310         init_sync_kiocb(&kiocb, filp);
2311         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2312         if (-EIOCBQUEUED == ret)
2313                 ret = wait_on_sync_kiocb(&kiocb);
2314         return ret;
2315 }
2316 EXPORT_SYMBOL(generic_file_readv);
2317
2318 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2319                         unsigned long nr_segs, loff_t *ppos)
2320 {
2321         struct address_space *mapping = file->f_mapping;
2322         struct inode *inode = mapping->host;
2323         ssize_t ret;
2324
2325         mutex_lock(&inode->i_mutex);
2326         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2327         mutex_unlock(&inode->i_mutex);
2328
2329         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2330                 int err;
2331
2332                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2333                 if (err < 0)
2334                         ret = err;
2335         }
2336         return ret;
2337 }
2338 EXPORT_SYMBOL(generic_file_writev);
2339
2340 /*
2341  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2342  * went wrong during pagecache shootdown.
2343  */
2344 static ssize_t
2345 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2346         loff_t offset, unsigned long nr_segs)
2347 {
2348         struct file *file = iocb->ki_filp;
2349         struct address_space *mapping = file->f_mapping;
2350         ssize_t retval;
2351         size_t write_len = 0;
2352
2353         /*
2354          * If it's a write, unmap all mmappings of the file up-front.  This
2355          * will cause any pte dirty bits to be propagated into the pageframes
2356          * for the subsequent filemap_write_and_wait().
2357          */
2358         if (rw == WRITE) {
2359                 write_len = iov_length(iov, nr_segs);
2360                 if (mapping_mapped(mapping))
2361                         unmap_mapping_range(mapping, offset, write_len, 0);
2362         }
2363
2364         retval = filemap_write_and_wait(mapping);
2365         if (retval == 0) {
2366                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2367                                                 offset, nr_segs);
2368                 if (rw == WRITE && mapping->nrpages) {
2369                         pgoff_t end = (offset + write_len - 1)
2370                                                 >> PAGE_CACHE_SHIFT;
2371                         int err = invalidate_inode_pages2_range(mapping,
2372                                         offset >> PAGE_CACHE_SHIFT, end);
2373                         if (err)
2374                                 retval = err;
2375                 }
2376         }
2377         return retval;
2378 }