Merge branch 'master'
[linux-2.6.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/config.h>
13 #include <linux/module.h>
14 #include <linux/slab.h>
15 #include <linux/compiler.h>
16 #include <linux/fs.h>
17 #include <linux/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  * Like find_get_pages, except we only return pages which are tagged with
702  * `tag'.   We update *index to index the next page for the traversal.
703  */
704 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
705                         int tag, unsigned int nr_pages, struct page **pages)
706 {
707         unsigned int i;
708         unsigned int ret;
709
710         read_lock_irq(&mapping->tree_lock);
711         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
712                                 (void **)pages, *index, nr_pages, tag);
713         for (i = 0; i < ret; i++)
714                 page_cache_get(pages[i]);
715         if (ret)
716                 *index = pages[ret - 1]->index + 1;
717         read_unlock_irq(&mapping->tree_lock);
718         return ret;
719 }
720
721 /*
722  * Same as grab_cache_page, but do not wait if the page is unavailable.
723  * This is intended for speculative data generators, where the data can
724  * be regenerated if the page couldn't be grabbed.  This routine should
725  * be safe to call while holding the lock for another page.
726  *
727  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
728  * and deadlock against the caller's locked page.
729  */
730 struct page *
731 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
732 {
733         struct page *page = find_get_page(mapping, index);
734         gfp_t gfp_mask;
735
736         if (page) {
737                 if (!TestSetPageLocked(page))
738                         return page;
739                 page_cache_release(page);
740                 return NULL;
741         }
742         gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
743         page = alloc_pages(gfp_mask, 0);
744         if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
745                 page_cache_release(page);
746                 page = NULL;
747         }
748         return page;
749 }
750
751 EXPORT_SYMBOL(grab_cache_page_nowait);
752
753 /*
754  * This is a generic file read routine, and uses the
755  * mapping->a_ops->readpage() function for the actual low-level
756  * stuff.
757  *
758  * This is really ugly. But the goto's actually try to clarify some
759  * of the logic when it comes to error handling etc.
760  *
761  * Note the struct file* is only passed for the use of readpage.  It may be
762  * NULL.
763  */
764 void do_generic_mapping_read(struct address_space *mapping,
765                              struct file_ra_state *_ra,
766                              struct file *filp,
767                              loff_t *ppos,
768                              read_descriptor_t *desc,
769                              read_actor_t actor)
770 {
771         struct inode *inode = mapping->host;
772         unsigned long index;
773         unsigned long end_index;
774         unsigned long offset;
775         unsigned long last_index;
776         unsigned long next_index;
777         unsigned long prev_index;
778         loff_t isize;
779         struct page *cached_page;
780         int error;
781         struct file_ra_state ra = *_ra;
782
783         cached_page = NULL;
784         index = *ppos >> PAGE_CACHE_SHIFT;
785         next_index = index;
786         prev_index = ra.prev_page;
787         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
788         offset = *ppos & ~PAGE_CACHE_MASK;
789
790         isize = i_size_read(inode);
791         if (!isize)
792                 goto out;
793
794         end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
795         for (;;) {
796                 struct page *page;
797                 unsigned long nr, ret;
798
799                 /* nr is the maximum number of bytes to copy from this page */
800                 nr = PAGE_CACHE_SIZE;
801                 if (index >= end_index) {
802                         if (index > end_index)
803                                 goto out;
804                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
805                         if (nr <= offset) {
806                                 goto out;
807                         }
808                 }
809                 nr = nr - offset;
810
811                 cond_resched();
812                 if (index == next_index)
813                         next_index = page_cache_readahead(mapping, &ra, filp,
814                                         index, last_index - index);
815
816 find_page:
817                 page = find_get_page(mapping, index);
818                 if (unlikely(page == NULL)) {
819                         handle_ra_miss(mapping, &ra, index);
820                         goto no_cached_page;
821                 }
822                 if (!PageUptodate(page))
823                         goto page_not_up_to_date;
824 page_ok:
825
826                 /* If users can be writing to this page using arbitrary
827                  * virtual addresses, take care about potential aliasing
828                  * before reading the page on the kernel side.
829                  */
830                 if (mapping_writably_mapped(mapping))
831                         flush_dcache_page(page);
832
833                 /*
834                  * When (part of) the same page is read multiple times
835                  * in succession, only mark it as accessed the first time.
836                  */
837                 if (prev_index != index)
838                         mark_page_accessed(page);
839                 prev_index = index;
840
841                 /*
842                  * Ok, we have the page, and it's up-to-date, so
843                  * now we can copy it to user space...
844                  *
845                  * The actor routine returns how many bytes were actually used..
846                  * NOTE! This may not be the same as how much of a user buffer
847                  * we filled up (we may be padding etc), so we can only update
848                  * "pos" here (the actor routine has to update the user buffer
849                  * pointers and the remaining count).
850                  */
851                 ret = actor(desc, page, offset, nr);
852                 offset += ret;
853                 index += offset >> PAGE_CACHE_SHIFT;
854                 offset &= ~PAGE_CACHE_MASK;
855
856                 page_cache_release(page);
857                 if (ret == nr && desc->count)
858                         continue;
859                 goto out;
860
861 page_not_up_to_date:
862                 /* Get exclusive access to the page ... */
863                 lock_page(page);
864
865                 /* Did it get unhashed before we got the lock? */
866                 if (!page->mapping) {
867                         unlock_page(page);
868                         page_cache_release(page);
869                         continue;
870                 }
871
872                 /* Did somebody else fill it already? */
873                 if (PageUptodate(page)) {
874                         unlock_page(page);
875                         goto page_ok;
876                 }
877
878 readpage:
879                 /* Start the actual read. The read will unlock the page. */
880                 error = mapping->a_ops->readpage(filp, page);
881
882                 if (unlikely(error)) {
883                         if (error == AOP_TRUNCATED_PAGE) {
884                                 page_cache_release(page);
885                                 goto find_page;
886                         }
887                         goto readpage_error;
888                 }
889
890                 if (!PageUptodate(page)) {
891                         lock_page(page);
892                         if (!PageUptodate(page)) {
893                                 if (page->mapping == NULL) {
894                                         /*
895                                          * invalidate_inode_pages got it
896                                          */
897                                         unlock_page(page);
898                                         page_cache_release(page);
899                                         goto find_page;
900                                 }
901                                 unlock_page(page);
902                                 error = -EIO;
903                                 goto readpage_error;
904                         }
905                         unlock_page(page);
906                 }
907
908                 /*
909                  * i_size must be checked after we have done ->readpage.
910                  *
911                  * Checking i_size after the readpage allows us to calculate
912                  * the correct value for "nr", which means the zero-filled
913                  * part of the page is not copied back to userspace (unless
914                  * another truncate extends the file - this is desired though).
915                  */
916                 isize = i_size_read(inode);
917                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
918                 if (unlikely(!isize || index > end_index)) {
919                         page_cache_release(page);
920                         goto out;
921                 }
922
923                 /* nr is the maximum number of bytes to copy from this page */
924                 nr = PAGE_CACHE_SIZE;
925                 if (index == end_index) {
926                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
927                         if (nr <= offset) {
928                                 page_cache_release(page);
929                                 goto out;
930                         }
931                 }
932                 nr = nr - offset;
933                 goto page_ok;
934
935 readpage_error:
936                 /* UHHUH! A synchronous read error occurred. Report it */
937                 desc->error = error;
938                 page_cache_release(page);
939                 goto out;
940
941 no_cached_page:
942                 /*
943                  * Ok, it wasn't cached, so we need to create a new
944                  * page..
945                  */
946                 if (!cached_page) {
947                         cached_page = page_cache_alloc_cold(mapping);
948                         if (!cached_page) {
949                                 desc->error = -ENOMEM;
950                                 goto out;
951                         }
952                 }
953                 error = add_to_page_cache_lru(cached_page, mapping,
954                                                 index, GFP_KERNEL);
955                 if (error) {
956                         if (error == -EEXIST)
957                                 goto find_page;
958                         desc->error = error;
959                         goto out;
960                 }
961                 page = cached_page;
962                 cached_page = NULL;
963                 goto readpage;
964         }
965
966 out:
967         *_ra = ra;
968
969         *ppos = ((loff_t) index << PAGE_CACHE_SHIFT) + offset;
970         if (cached_page)
971                 page_cache_release(cached_page);
972         if (filp)
973                 file_accessed(filp);
974 }
975
976 EXPORT_SYMBOL(do_generic_mapping_read);
977
978 int file_read_actor(read_descriptor_t *desc, struct page *page,
979                         unsigned long offset, unsigned long size)
980 {
981         char *kaddr;
982         unsigned long left, count = desc->count;
983
984         if (size > count)
985                 size = count;
986
987         /*
988          * Faults on the destination of a read are common, so do it before
989          * taking the kmap.
990          */
991         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
992                 kaddr = kmap_atomic(page, KM_USER0);
993                 left = __copy_to_user_inatomic(desc->arg.buf,
994                                                 kaddr + offset, size);
995                 kunmap_atomic(kaddr, KM_USER0);
996                 if (left == 0)
997                         goto success;
998         }
999
1000         /* Do it the slow way */
1001         kaddr = kmap(page);
1002         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1003         kunmap(page);
1004
1005         if (left) {
1006                 size -= left;
1007                 desc->error = -EFAULT;
1008         }
1009 success:
1010         desc->count = count - size;
1011         desc->written += size;
1012         desc->arg.buf += size;
1013         return size;
1014 }
1015 EXPORT_SYMBOL(file_read_actor);
1016
1017 /*
1018  * This is the "read()" routine for all filesystems
1019  * that can use the page cache directly.
1020  */
1021 ssize_t
1022 __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1023                 unsigned long nr_segs, loff_t *ppos)
1024 {
1025         struct file *filp = iocb->ki_filp;
1026         ssize_t retval;
1027         unsigned long seg;
1028         size_t count;
1029
1030         count = 0;
1031         for (seg = 0; seg < nr_segs; seg++) {
1032                 const struct iovec *iv = &iov[seg];
1033
1034                 /*
1035                  * If any segment has a negative length, or the cumulative
1036                  * length ever wraps negative then return -EINVAL.
1037                  */
1038                 count += iv->iov_len;
1039                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
1040                         return -EINVAL;
1041                 if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
1042                         continue;
1043                 if (seg == 0)
1044                         return -EFAULT;
1045                 nr_segs = seg;
1046                 count -= iv->iov_len;   /* This segment is no good */
1047                 break;
1048         }
1049
1050         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1051         if (filp->f_flags & O_DIRECT) {
1052                 loff_t pos = *ppos, size;
1053                 struct address_space *mapping;
1054                 struct inode *inode;
1055
1056                 mapping = filp->f_mapping;
1057                 inode = mapping->host;
1058                 retval = 0;
1059                 if (!count)
1060                         goto out; /* skip atime */
1061                 size = i_size_read(inode);
1062                 if (pos < size) {
1063                         retval = generic_file_direct_IO(READ, iocb,
1064                                                 iov, pos, nr_segs);
1065                         if (retval > 0 && !is_sync_kiocb(iocb))
1066                                 retval = -EIOCBQUEUED;
1067                         if (retval > 0)
1068                                 *ppos = pos + retval;
1069                 }
1070                 file_accessed(filp);
1071                 goto out;
1072         }
1073
1074         retval = 0;
1075         if (count) {
1076                 for (seg = 0; seg < nr_segs; seg++) {
1077                         read_descriptor_t desc;
1078
1079                         desc.written = 0;
1080                         desc.arg.buf = iov[seg].iov_base;
1081                         desc.count = iov[seg].iov_len;
1082                         if (desc.count == 0)
1083                                 continue;
1084                         desc.error = 0;
1085                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1086                         retval += desc.written;
1087                         if (desc.error) {
1088                                 retval = retval ?: desc.error;
1089                                 break;
1090                         }
1091                 }
1092         }
1093 out:
1094         return retval;
1095 }
1096
1097 EXPORT_SYMBOL(__generic_file_aio_read);
1098
1099 ssize_t
1100 generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
1101 {
1102         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1103
1104         BUG_ON(iocb->ki_pos != pos);
1105         return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
1106 }
1107
1108 EXPORT_SYMBOL(generic_file_aio_read);
1109
1110 ssize_t
1111 generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
1112 {
1113         struct iovec local_iov = { .iov_base = buf, .iov_len = count };
1114         struct kiocb kiocb;
1115         ssize_t ret;
1116
1117         init_sync_kiocb(&kiocb, filp);
1118         ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
1119         if (-EIOCBQUEUED == ret)
1120                 ret = wait_on_sync_kiocb(&kiocb);
1121         return ret;
1122 }
1123
1124 EXPORT_SYMBOL(generic_file_read);
1125
1126 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
1127 {
1128         ssize_t written;
1129         unsigned long count = desc->count;
1130         struct file *file = desc->arg.data;
1131
1132         if (size > count)
1133                 size = count;
1134
1135         written = file->f_op->sendpage(file, page, offset,
1136                                        size, &file->f_pos, size<count);
1137         if (written < 0) {
1138                 desc->error = written;
1139                 written = 0;
1140         }
1141         desc->count = count - written;
1142         desc->written += written;
1143         return written;
1144 }
1145
1146 ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
1147                          size_t count, read_actor_t actor, void *target)
1148 {
1149         read_descriptor_t desc;
1150
1151         if (!count)
1152                 return 0;
1153
1154         desc.written = 0;
1155         desc.count = count;
1156         desc.arg.data = target;
1157         desc.error = 0;
1158
1159         do_generic_file_read(in_file, ppos, &desc, actor);
1160         if (desc.written)
1161                 return desc.written;
1162         return desc.error;
1163 }
1164
1165 EXPORT_SYMBOL(generic_file_sendfile);
1166
1167 static ssize_t
1168 do_readahead(struct address_space *mapping, struct file *filp,
1169              unsigned long index, unsigned long nr)
1170 {
1171         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1172                 return -EINVAL;
1173
1174         force_page_cache_readahead(mapping, filp, index,
1175                                         max_sane_readahead(nr));
1176         return 0;
1177 }
1178
1179 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1180 {
1181         ssize_t ret;
1182         struct file *file;
1183
1184         ret = -EBADF;
1185         file = fget(fd);
1186         if (file) {
1187                 if (file->f_mode & FMODE_READ) {
1188                         struct address_space *mapping = file->f_mapping;
1189                         unsigned long start = offset >> PAGE_CACHE_SHIFT;
1190                         unsigned long end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1191                         unsigned long len = end - start + 1;
1192                         ret = do_readahead(mapping, file, start, len);
1193                 }
1194                 fput(file);
1195         }
1196         return ret;
1197 }
1198
1199 #ifdef CONFIG_MMU
1200 /*
1201  * This adds the requested page to the page cache if it isn't already there,
1202  * and schedules an I/O to read in its contents from disk.
1203  */
1204 static int FASTCALL(page_cache_read(struct file * file, unsigned long offset));
1205 static int fastcall page_cache_read(struct file * file, unsigned long offset)
1206 {
1207         struct address_space *mapping = file->f_mapping;
1208         struct page *page; 
1209         int ret;
1210
1211         do {
1212                 page = page_cache_alloc_cold(mapping);
1213                 if (!page)
1214                         return -ENOMEM;
1215
1216                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1217                 if (ret == 0)
1218                         ret = mapping->a_ops->readpage(file, page);
1219                 else if (ret == -EEXIST)
1220                         ret = 0; /* losing race to add is OK */
1221
1222                 page_cache_release(page);
1223
1224         } while (ret == AOP_TRUNCATED_PAGE);
1225                 
1226         return ret;
1227 }
1228
1229 #define MMAP_LOTSAMISS  (100)
1230
1231 /*
1232  * filemap_nopage() is invoked via the vma operations vector for a
1233  * mapped memory region to read in file data during a page fault.
1234  *
1235  * The goto's are kind of ugly, but this streamlines the normal case of having
1236  * it in the page cache, and handles the special cases reasonably without
1237  * having a lot of duplicated code.
1238  */
1239 struct page *filemap_nopage(struct vm_area_struct *area,
1240                                 unsigned long address, int *type)
1241 {
1242         int error;
1243         struct file *file = area->vm_file;
1244         struct address_space *mapping = file->f_mapping;
1245         struct file_ra_state *ra = &file->f_ra;
1246         struct inode *inode = mapping->host;
1247         struct page *page;
1248         unsigned long size, pgoff;
1249         int did_readaround = 0, majmin = VM_FAULT_MINOR;
1250
1251         pgoff = ((address-area->vm_start) >> PAGE_CACHE_SHIFT) + area->vm_pgoff;
1252
1253 retry_all:
1254         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1255         if (pgoff >= size)
1256                 goto outside_data_content;
1257
1258         /* If we don't want any read-ahead, don't bother */
1259         if (VM_RandomReadHint(area))
1260                 goto no_cached_page;
1261
1262         /*
1263          * The readahead code wants to be told about each and every page
1264          * so it can build and shrink its windows appropriately
1265          *
1266          * For sequential accesses, we use the generic readahead logic.
1267          */
1268         if (VM_SequentialReadHint(area))
1269                 page_cache_readahead(mapping, ra, file, pgoff, 1);
1270
1271         /*
1272          * Do we have something in the page cache already?
1273          */
1274 retry_find:
1275         page = find_get_page(mapping, pgoff);
1276         if (!page) {
1277                 unsigned long ra_pages;
1278
1279                 if (VM_SequentialReadHint(area)) {
1280                         handle_ra_miss(mapping, ra, pgoff);
1281                         goto no_cached_page;
1282                 }
1283                 ra->mmap_miss++;
1284
1285                 /*
1286                  * Do we miss much more than hit in this file? If so,
1287                  * stop bothering with read-ahead. It will only hurt.
1288                  */
1289                 if (ra->mmap_miss > ra->mmap_hit + MMAP_LOTSAMISS)
1290                         goto no_cached_page;
1291
1292                 /*
1293                  * To keep the pgmajfault counter straight, we need to
1294                  * check did_readaround, as this is an inner loop.
1295                  */
1296                 if (!did_readaround) {
1297                         majmin = VM_FAULT_MAJOR;
1298                         inc_page_state(pgmajfault);
1299                 }
1300                 did_readaround = 1;
1301                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1302                 if (ra_pages) {
1303                         pgoff_t start = 0;
1304
1305                         if (pgoff > ra_pages / 2)
1306                                 start = pgoff - ra_pages / 2;
1307                         do_page_cache_readahead(mapping, file, start, ra_pages);
1308                 }
1309                 page = find_get_page(mapping, pgoff);
1310                 if (!page)
1311                         goto no_cached_page;
1312         }
1313
1314         if (!did_readaround)
1315                 ra->mmap_hit++;
1316
1317         /*
1318          * Ok, found a page in the page cache, now we need to check
1319          * that it's up-to-date.
1320          */
1321         if (!PageUptodate(page))
1322                 goto page_not_uptodate;
1323
1324 success:
1325         /*
1326          * Found the page and have a reference on it.
1327          */
1328         mark_page_accessed(page);
1329         if (type)
1330                 *type = majmin;
1331         return page;
1332
1333 outside_data_content:
1334         /*
1335          * An external ptracer can access pages that normally aren't
1336          * accessible..
1337          */
1338         if (area->vm_mm == current->mm)
1339                 return NULL;
1340         /* Fall through to the non-read-ahead case */
1341 no_cached_page:
1342         /*
1343          * We're only likely to ever get here if MADV_RANDOM is in
1344          * effect.
1345          */
1346         error = page_cache_read(file, pgoff);
1347         grab_swap_token();
1348
1349         /*
1350          * The page we want has now been added to the page cache.
1351          * In the unlikely event that someone removed it in the
1352          * meantime, we'll just come back here and read it again.
1353          */
1354         if (error >= 0)
1355                 goto retry_find;
1356
1357         /*
1358          * An error return from page_cache_read can result if the
1359          * system is low on memory, or a problem occurs while trying
1360          * to schedule I/O.
1361          */
1362         if (error == -ENOMEM)
1363                 return NOPAGE_OOM;
1364         return NULL;
1365
1366 page_not_uptodate:
1367         if (!did_readaround) {
1368                 majmin = VM_FAULT_MAJOR;
1369                 inc_page_state(pgmajfault);
1370         }
1371         lock_page(page);
1372
1373         /* Did it get unhashed while we waited for it? */
1374         if (!page->mapping) {
1375                 unlock_page(page);
1376                 page_cache_release(page);
1377                 goto retry_all;
1378         }
1379
1380         /* Did somebody else get it up-to-date? */
1381         if (PageUptodate(page)) {
1382                 unlock_page(page);
1383                 goto success;
1384         }
1385
1386         error = mapping->a_ops->readpage(file, page);
1387         if (!error) {
1388                 wait_on_page_locked(page);
1389                 if (PageUptodate(page))
1390                         goto success;
1391         } else if (error == AOP_TRUNCATED_PAGE) {
1392                 page_cache_release(page);
1393                 goto retry_find;
1394         }
1395
1396         /*
1397          * Umm, take care of errors if the page isn't up-to-date.
1398          * Try to re-read it _once_. We do this synchronously,
1399          * because there really aren't any performance issues here
1400          * and we need to check for errors.
1401          */
1402         lock_page(page);
1403
1404         /* Somebody truncated the page on us? */
1405         if (!page->mapping) {
1406                 unlock_page(page);
1407                 page_cache_release(page);
1408                 goto retry_all;
1409         }
1410
1411         /* Somebody else successfully read it in? */
1412         if (PageUptodate(page)) {
1413                 unlock_page(page);
1414                 goto success;
1415         }
1416         ClearPageError(page);
1417         error = mapping->a_ops->readpage(file, page);
1418         if (!error) {
1419                 wait_on_page_locked(page);
1420                 if (PageUptodate(page))
1421                         goto success;
1422         } else if (error == AOP_TRUNCATED_PAGE) {
1423                 page_cache_release(page);
1424                 goto retry_find;
1425         }
1426
1427         /*
1428          * Things didn't work out. Return zero to tell the
1429          * mm layer so, possibly freeing the page cache page first.
1430          */
1431         page_cache_release(page);
1432         return NULL;
1433 }
1434
1435 EXPORT_SYMBOL(filemap_nopage);
1436
1437 static struct page * filemap_getpage(struct file *file, unsigned long pgoff,
1438                                         int nonblock)
1439 {
1440         struct address_space *mapping = file->f_mapping;
1441         struct page *page;
1442         int error;
1443
1444         /*
1445          * Do we have something in the page cache already?
1446          */
1447 retry_find:
1448         page = find_get_page(mapping, pgoff);
1449         if (!page) {
1450                 if (nonblock)
1451                         return NULL;
1452                 goto no_cached_page;
1453         }
1454
1455         /*
1456          * Ok, found a page in the page cache, now we need to check
1457          * that it's up-to-date.
1458          */
1459         if (!PageUptodate(page)) {
1460                 if (nonblock) {
1461                         page_cache_release(page);
1462                         return NULL;
1463                 }
1464                 goto page_not_uptodate;
1465         }
1466
1467 success:
1468         /*
1469          * Found the page and have a reference on it.
1470          */
1471         mark_page_accessed(page);
1472         return page;
1473
1474 no_cached_page:
1475         error = page_cache_read(file, pgoff);
1476
1477         /*
1478          * The page we want has now been added to the page cache.
1479          * In the unlikely event that someone removed it in the
1480          * meantime, we'll just come back here and read it again.
1481          */
1482         if (error >= 0)
1483                 goto retry_find;
1484
1485         /*
1486          * An error return from page_cache_read can result if the
1487          * system is low on memory, or a problem occurs while trying
1488          * to schedule I/O.
1489          */
1490         return NULL;
1491
1492 page_not_uptodate:
1493         lock_page(page);
1494
1495         /* Did it get unhashed while we waited for it? */
1496         if (!page->mapping) {
1497                 unlock_page(page);
1498                 goto err;
1499         }
1500
1501         /* Did somebody else get it up-to-date? */
1502         if (PageUptodate(page)) {
1503                 unlock_page(page);
1504                 goto success;
1505         }
1506
1507         error = mapping->a_ops->readpage(file, page);
1508         if (!error) {
1509                 wait_on_page_locked(page);
1510                 if (PageUptodate(page))
1511                         goto success;
1512         } else if (error == AOP_TRUNCATED_PAGE) {
1513                 page_cache_release(page);
1514                 goto retry_find;
1515         }
1516
1517         /*
1518          * Umm, take care of errors if the page isn't up-to-date.
1519          * Try to re-read it _once_. We do this synchronously,
1520          * because there really aren't any performance issues here
1521          * and we need to check for errors.
1522          */
1523         lock_page(page);
1524
1525         /* Somebody truncated the page on us? */
1526         if (!page->mapping) {
1527                 unlock_page(page);
1528                 goto err;
1529         }
1530         /* Somebody else successfully read it in? */
1531         if (PageUptodate(page)) {
1532                 unlock_page(page);
1533                 goto success;
1534         }
1535
1536         ClearPageError(page);
1537         error = mapping->a_ops->readpage(file, page);
1538         if (!error) {
1539                 wait_on_page_locked(page);
1540                 if (PageUptodate(page))
1541                         goto success;
1542         } else if (error == AOP_TRUNCATED_PAGE) {
1543                 page_cache_release(page);
1544                 goto retry_find;
1545         }
1546
1547         /*
1548          * Things didn't work out. Return zero to tell the
1549          * mm layer so, possibly freeing the page cache page first.
1550          */
1551 err:
1552         page_cache_release(page);
1553
1554         return NULL;
1555 }
1556
1557 int filemap_populate(struct vm_area_struct *vma, unsigned long addr,
1558                 unsigned long len, pgprot_t prot, unsigned long pgoff,
1559                 int nonblock)
1560 {
1561         struct file *file = vma->vm_file;
1562         struct address_space *mapping = file->f_mapping;
1563         struct inode *inode = mapping->host;
1564         unsigned long size;
1565         struct mm_struct *mm = vma->vm_mm;
1566         struct page *page;
1567         int err;
1568
1569         if (!nonblock)
1570                 force_page_cache_readahead(mapping, vma->vm_file,
1571                                         pgoff, len >> PAGE_CACHE_SHIFT);
1572
1573 repeat:
1574         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1575         if (pgoff + (len >> PAGE_CACHE_SHIFT) > size)
1576                 return -EINVAL;
1577
1578         page = filemap_getpage(file, pgoff, nonblock);
1579
1580         /* XXX: This is wrong, a filesystem I/O error may have happened. Fix that as
1581          * done in shmem_populate calling shmem_getpage */
1582         if (!page && !nonblock)
1583                 return -ENOMEM;
1584
1585         if (page) {
1586                 err = install_page(mm, vma, addr, page, prot);
1587                 if (err) {
1588                         page_cache_release(page);
1589                         return err;
1590                 }
1591         } else if (vma->vm_flags & VM_NONLINEAR) {
1592                 /* No page was found just because we can't read it in now (being
1593                  * here implies nonblock != 0), but the page may exist, so set
1594                  * the PTE to fault it in later. */
1595                 err = install_file_pte(mm, vma, addr, pgoff, prot);
1596                 if (err)
1597                         return err;
1598         }
1599
1600         len -= PAGE_SIZE;
1601         addr += PAGE_SIZE;
1602         pgoff++;
1603         if (len)
1604                 goto repeat;
1605
1606         return 0;
1607 }
1608 EXPORT_SYMBOL(filemap_populate);
1609
1610 struct vm_operations_struct generic_file_vm_ops = {
1611         .nopage         = filemap_nopage,
1612         .populate       = filemap_populate,
1613 };
1614
1615 /* This is used for a general mmap of a disk file */
1616
1617 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1618 {
1619         struct address_space *mapping = file->f_mapping;
1620
1621         if (!mapping->a_ops->readpage)
1622                 return -ENOEXEC;
1623         file_accessed(file);
1624         vma->vm_ops = &generic_file_vm_ops;
1625         return 0;
1626 }
1627
1628 /*
1629  * This is for filesystems which do not implement ->writepage.
1630  */
1631 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1632 {
1633         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1634                 return -EINVAL;
1635         return generic_file_mmap(file, vma);
1636 }
1637 #else
1638 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1639 {
1640         return -ENOSYS;
1641 }
1642 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1643 {
1644         return -ENOSYS;
1645 }
1646 #endif /* CONFIG_MMU */
1647
1648 EXPORT_SYMBOL(generic_file_mmap);
1649 EXPORT_SYMBOL(generic_file_readonly_mmap);
1650
1651 static inline struct page *__read_cache_page(struct address_space *mapping,
1652                                 unsigned long index,
1653                                 int (*filler)(void *,struct page*),
1654                                 void *data)
1655 {
1656         struct page *page, *cached_page = NULL;
1657         int err;
1658 repeat:
1659         page = find_get_page(mapping, index);
1660         if (!page) {
1661                 if (!cached_page) {
1662                         cached_page = page_cache_alloc_cold(mapping);
1663                         if (!cached_page)
1664                                 return ERR_PTR(-ENOMEM);
1665                 }
1666                 err = add_to_page_cache_lru(cached_page, mapping,
1667                                         index, GFP_KERNEL);
1668                 if (err == -EEXIST)
1669                         goto repeat;
1670                 if (err < 0) {
1671                         /* Presumably ENOMEM for radix tree node */
1672                         page_cache_release(cached_page);
1673                         return ERR_PTR(err);
1674                 }
1675                 page = cached_page;
1676                 cached_page = NULL;
1677                 err = filler(data, page);
1678                 if (err < 0) {
1679                         page_cache_release(page);
1680                         page = ERR_PTR(err);
1681                 }
1682         }
1683         if (cached_page)
1684                 page_cache_release(cached_page);
1685         return page;
1686 }
1687
1688 /*
1689  * Read into the page cache. If a page already exists,
1690  * and PageUptodate() is not set, try to fill the page.
1691  */
1692 struct page *read_cache_page(struct address_space *mapping,
1693                                 unsigned long index,
1694                                 int (*filler)(void *,struct page*),
1695                                 void *data)
1696 {
1697         struct page *page;
1698         int err;
1699
1700 retry:
1701         page = __read_cache_page(mapping, index, filler, data);
1702         if (IS_ERR(page))
1703                 goto out;
1704         mark_page_accessed(page);
1705         if (PageUptodate(page))
1706                 goto out;
1707
1708         lock_page(page);
1709         if (!page->mapping) {
1710                 unlock_page(page);
1711                 page_cache_release(page);
1712                 goto retry;
1713         }
1714         if (PageUptodate(page)) {
1715                 unlock_page(page);
1716                 goto out;
1717         }
1718         err = filler(data, page);
1719         if (err < 0) {
1720                 page_cache_release(page);
1721                 page = ERR_PTR(err);
1722         }
1723  out:
1724         return page;
1725 }
1726
1727 EXPORT_SYMBOL(read_cache_page);
1728
1729 /*
1730  * If the page was newly created, increment its refcount and add it to the
1731  * caller's lru-buffering pagevec.  This function is specifically for
1732  * generic_file_write().
1733  */
1734 static inline struct page *
1735 __grab_cache_page(struct address_space *mapping, unsigned long index,
1736                         struct page **cached_page, struct pagevec *lru_pvec)
1737 {
1738         int err;
1739         struct page *page;
1740 repeat:
1741         page = find_lock_page(mapping, index);
1742         if (!page) {
1743                 if (!*cached_page) {
1744                         *cached_page = page_cache_alloc(mapping);
1745                         if (!*cached_page)
1746                                 return NULL;
1747                 }
1748                 err = add_to_page_cache(*cached_page, mapping,
1749                                         index, GFP_KERNEL);
1750                 if (err == -EEXIST)
1751                         goto repeat;
1752                 if (err == 0) {
1753                         page = *cached_page;
1754                         page_cache_get(page);
1755                         if (!pagevec_add(lru_pvec, page))
1756                                 __pagevec_lru_add(lru_pvec);
1757                         *cached_page = NULL;
1758                 }
1759         }
1760         return page;
1761 }
1762
1763 /*
1764  * The logic we want is
1765  *
1766  *      if suid or (sgid and xgrp)
1767  *              remove privs
1768  */
1769 int remove_suid(struct dentry *dentry)
1770 {
1771         mode_t mode = dentry->d_inode->i_mode;
1772         int kill = 0;
1773         int result = 0;
1774
1775         /* suid always must be killed */
1776         if (unlikely(mode & S_ISUID))
1777                 kill = ATTR_KILL_SUID;
1778
1779         /*
1780          * sgid without any exec bits is just a mandatory locking mark; leave
1781          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1782          */
1783         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1784                 kill |= ATTR_KILL_SGID;
1785
1786         if (unlikely(kill && !capable(CAP_FSETID))) {
1787                 struct iattr newattrs;
1788
1789                 newattrs.ia_valid = ATTR_FORCE | kill;
1790                 result = notify_change(dentry, &newattrs);
1791         }
1792         return result;
1793 }
1794 EXPORT_SYMBOL(remove_suid);
1795
1796 size_t
1797 __filemap_copy_from_user_iovec(char *vaddr, 
1798                         const struct iovec *iov, size_t base, size_t bytes)
1799 {
1800         size_t copied = 0, left = 0;
1801
1802         while (bytes) {
1803                 char __user *buf = iov->iov_base + base;
1804                 int copy = min(bytes, iov->iov_len - base);
1805
1806                 base = 0;
1807                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1808                 copied += copy;
1809                 bytes -= copy;
1810                 vaddr += copy;
1811                 iov++;
1812
1813                 if (unlikely(left)) {
1814                         /* zero the rest of the target like __copy_from_user */
1815                         if (bytes)
1816                                 memset(vaddr, 0, bytes);
1817                         break;
1818                 }
1819         }
1820         return copied - left;
1821 }
1822
1823 /*
1824  * Performs necessary checks before doing a write
1825  *
1826  * Can adjust writing position aor amount of bytes to write.
1827  * Returns appropriate error code that caller should return or
1828  * zero in case that write should be allowed.
1829  */
1830 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1831 {
1832         struct inode *inode = file->f_mapping->host;
1833         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1834
1835         if (unlikely(*pos < 0))
1836                 return -EINVAL;
1837
1838         if (!isblk) {
1839                 /* FIXME: this is for backwards compatibility with 2.4 */
1840                 if (file->f_flags & O_APPEND)
1841                         *pos = i_size_read(inode);
1842
1843                 if (limit != RLIM_INFINITY) {
1844                         if (*pos >= limit) {
1845                                 send_sig(SIGXFSZ, current, 0);
1846                                 return -EFBIG;
1847                         }
1848                         if (*count > limit - (typeof(limit))*pos) {
1849                                 *count = limit - (typeof(limit))*pos;
1850                         }
1851                 }
1852         }
1853
1854         /*
1855          * LFS rule
1856          */
1857         if (unlikely(*pos + *count > MAX_NON_LFS &&
1858                                 !(file->f_flags & O_LARGEFILE))) {
1859                 if (*pos >= MAX_NON_LFS) {
1860                         send_sig(SIGXFSZ, current, 0);
1861                         return -EFBIG;
1862                 }
1863                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1864                         *count = MAX_NON_LFS - (unsigned long)*pos;
1865                 }
1866         }
1867
1868         /*
1869          * Are we about to exceed the fs block limit ?
1870          *
1871          * If we have written data it becomes a short write.  If we have
1872          * exceeded without writing data we send a signal and return EFBIG.
1873          * Linus frestrict idea will clean these up nicely..
1874          */
1875         if (likely(!isblk)) {
1876                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1877                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1878                                 send_sig(SIGXFSZ, current, 0);
1879                                 return -EFBIG;
1880                         }
1881                         /* zero-length writes at ->s_maxbytes are OK */
1882                 }
1883
1884                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1885                         *count = inode->i_sb->s_maxbytes - *pos;
1886         } else {
1887                 loff_t isize;
1888                 if (bdev_read_only(I_BDEV(inode)))
1889                         return -EPERM;
1890                 isize = i_size_read(inode);
1891                 if (*pos >= isize) {
1892                         if (*count || *pos > isize)
1893                                 return -ENOSPC;
1894                 }
1895
1896                 if (*pos + *count > isize)
1897                         *count = isize - *pos;
1898         }
1899         return 0;
1900 }
1901 EXPORT_SYMBOL(generic_write_checks);
1902
1903 ssize_t
1904 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
1905                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
1906                 size_t count, size_t ocount)
1907 {
1908         struct file     *file = iocb->ki_filp;
1909         struct address_space *mapping = file->f_mapping;
1910         struct inode    *inode = mapping->host;
1911         ssize_t         written;
1912
1913         if (count != ocount)
1914                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
1915
1916         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
1917         if (written > 0) {
1918                 loff_t end = pos + written;
1919                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
1920                         i_size_write(inode,  end);
1921                         mark_inode_dirty(inode);
1922                 }
1923                 *ppos = end;
1924         }
1925
1926         /*
1927          * Sync the fs metadata but not the minor inode changes and
1928          * of course not the data as we did direct DMA for the IO.
1929          * i_mutex is held, which protects generic_osync_inode() from
1930          * livelocking.
1931          */
1932         if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
1933                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
1934                 if (err < 0)
1935                         written = err;
1936         }
1937         if (written == count && !is_sync_kiocb(iocb))
1938                 written = -EIOCBQUEUED;
1939         return written;
1940 }
1941 EXPORT_SYMBOL(generic_file_direct_write);
1942
1943 ssize_t
1944 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
1945                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
1946                 size_t count, ssize_t written)
1947 {
1948         struct file *file = iocb->ki_filp;
1949         struct address_space * mapping = file->f_mapping;
1950         struct address_space_operations *a_ops = mapping->a_ops;
1951         struct inode    *inode = mapping->host;
1952         long            status = 0;
1953         struct page     *page;
1954         struct page     *cached_page = NULL;
1955         size_t          bytes;
1956         struct pagevec  lru_pvec;
1957         const struct iovec *cur_iov = iov; /* current iovec */
1958         size_t          iov_base = 0;      /* offset in the current iovec */
1959         char __user     *buf;
1960
1961         pagevec_init(&lru_pvec, 0);
1962
1963         /*
1964          * handle partial DIO write.  Adjust cur_iov if needed.
1965          */
1966         if (likely(nr_segs == 1))
1967                 buf = iov->iov_base + written;
1968         else {
1969                 filemap_set_next_iovec(&cur_iov, &iov_base, written);
1970                 buf = cur_iov->iov_base + iov_base;
1971         }
1972
1973         do {
1974                 unsigned long index;
1975                 unsigned long offset;
1976                 unsigned long maxlen;
1977                 size_t copied;
1978
1979                 offset = (pos & (PAGE_CACHE_SIZE -1)); /* Within page */
1980                 index = pos >> PAGE_CACHE_SHIFT;
1981                 bytes = PAGE_CACHE_SIZE - offset;
1982                 if (bytes > count)
1983                         bytes = count;
1984
1985                 /*
1986                  * Bring in the user page that we will copy from _first_.
1987                  * Otherwise there's a nasty deadlock on copying from the
1988                  * same page as we're writing to, without it being marked
1989                  * up-to-date.
1990                  */
1991                 maxlen = cur_iov->iov_len - iov_base;
1992                 if (maxlen > bytes)
1993                         maxlen = bytes;
1994                 fault_in_pages_readable(buf, maxlen);
1995
1996                 page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
1997                 if (!page) {
1998                         status = -ENOMEM;
1999                         break;
2000                 }
2001
2002                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2003                 if (unlikely(status)) {
2004                         loff_t isize = i_size_read(inode);
2005
2006                         if (status != AOP_TRUNCATED_PAGE)
2007                                 unlock_page(page);
2008                         page_cache_release(page);
2009                         if (status == AOP_TRUNCATED_PAGE)
2010                                 continue;
2011                         /*
2012                          * prepare_write() may have instantiated a few blocks
2013                          * outside i_size.  Trim these off again.
2014                          */
2015                         if (pos + bytes > isize)
2016                                 vmtruncate(inode, isize);
2017                         break;
2018                 }
2019                 if (likely(nr_segs == 1))
2020                         copied = filemap_copy_from_user(page, offset,
2021                                                         buf, bytes);
2022                 else
2023                         copied = filemap_copy_from_user_iovec(page, offset,
2024                                                 cur_iov, iov_base, bytes);
2025                 flush_dcache_page(page);
2026                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2027                 if (status == AOP_TRUNCATED_PAGE) {
2028                         page_cache_release(page);
2029                         continue;
2030                 }
2031                 if (likely(copied > 0)) {
2032                         if (!status)
2033                                 status = copied;
2034
2035                         if (status >= 0) {
2036                                 written += status;
2037                                 count -= status;
2038                                 pos += status;
2039                                 buf += status;
2040                                 if (unlikely(nr_segs > 1)) {
2041                                         filemap_set_next_iovec(&cur_iov,
2042                                                         &iov_base, status);
2043                                         if (count)
2044                                                 buf = cur_iov->iov_base +
2045                                                         iov_base;
2046                                 } else {
2047                                         iov_base += status;
2048                                 }
2049                         }
2050                 }
2051                 if (unlikely(copied != bytes))
2052                         if (status >= 0)
2053                                 status = -EFAULT;
2054                 unlock_page(page);
2055                 mark_page_accessed(page);
2056                 page_cache_release(page);
2057                 if (status < 0)
2058                         break;
2059                 balance_dirty_pages_ratelimited(mapping);
2060                 cond_resched();
2061         } while (count);
2062         *ppos = pos;
2063
2064         if (cached_page)
2065                 page_cache_release(cached_page);
2066
2067         /*
2068          * For now, when the user asks for O_SYNC, we'll actually give O_DSYNC
2069          */
2070         if (likely(status >= 0)) {
2071                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2072                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2073                                 status = generic_osync_inode(inode, mapping,
2074                                                 OSYNC_METADATA|OSYNC_DATA);
2075                 }
2076         }
2077         
2078         /*
2079          * If we get here for O_DIRECT writes then we must have fallen through
2080          * to buffered writes (block instantiation inside i_size).  So we sync
2081          * the file data here, to try to honour O_DIRECT expectations.
2082          */
2083         if (unlikely(file->f_flags & O_DIRECT) && written)
2084                 status = filemap_write_and_wait(mapping);
2085
2086         pagevec_lru_add(&lru_pvec);
2087         return written ? written : status;
2088 }
2089 EXPORT_SYMBOL(generic_file_buffered_write);
2090
2091 static ssize_t
2092 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2093                                 unsigned long nr_segs, loff_t *ppos)
2094 {
2095         struct file *file = iocb->ki_filp;
2096         struct address_space * mapping = file->f_mapping;
2097         size_t ocount;          /* original count */
2098         size_t count;           /* after file limit checks */
2099         struct inode    *inode = mapping->host;
2100         unsigned long   seg;
2101         loff_t          pos;
2102         ssize_t         written;
2103         ssize_t         err;
2104
2105         ocount = 0;
2106         for (seg = 0; seg < nr_segs; seg++) {
2107                 const struct iovec *iv = &iov[seg];
2108
2109                 /*
2110                  * If any segment has a negative length, or the cumulative
2111                  * length ever wraps negative then return -EINVAL.
2112                  */
2113                 ocount += iv->iov_len;
2114                 if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
2115                         return -EINVAL;
2116                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2117                         continue;
2118                 if (seg == 0)
2119                         return -EFAULT;
2120                 nr_segs = seg;
2121                 ocount -= iv->iov_len;  /* This segment is no good */
2122                 break;
2123         }
2124
2125         count = ocount;
2126         pos = *ppos;
2127
2128         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2129
2130         /* We can write back this queue in page reclaim */
2131         current->backing_dev_info = mapping->backing_dev_info;
2132         written = 0;
2133
2134         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2135         if (err)
2136                 goto out;
2137
2138         if (count == 0)
2139                 goto out;
2140
2141         err = remove_suid(file->f_dentry);
2142         if (err)
2143                 goto out;
2144
2145         file_update_time(file);
2146
2147         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2148         if (unlikely(file->f_flags & O_DIRECT)) {
2149                 written = generic_file_direct_write(iocb, iov,
2150                                 &nr_segs, pos, ppos, count, ocount);
2151                 if (written < 0 || written == count)
2152                         goto out;
2153                 /*
2154                  * direct-io write to a hole: fall through to buffered I/O
2155                  * for completing the rest of the request.
2156                  */
2157                 pos += written;
2158                 count -= written;
2159         }
2160
2161         written = generic_file_buffered_write(iocb, iov, nr_segs,
2162                         pos, ppos, count, written);
2163 out:
2164         current->backing_dev_info = NULL;
2165         return written ? written : err;
2166 }
2167 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2168
2169 ssize_t
2170 generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2171                                 unsigned long nr_segs, loff_t *ppos)
2172 {
2173         struct file *file = iocb->ki_filp;
2174         struct address_space *mapping = file->f_mapping;
2175         struct inode *inode = mapping->host;
2176         ssize_t ret;
2177         loff_t pos = *ppos;
2178
2179         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
2180
2181         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2182                 int err;
2183
2184                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2185                 if (err < 0)
2186                         ret = err;
2187         }
2188         return ret;
2189 }
2190
2191 static ssize_t
2192 __generic_file_write_nolock(struct file *file, const struct iovec *iov,
2193                                 unsigned long nr_segs, loff_t *ppos)
2194 {
2195         struct kiocb kiocb;
2196         ssize_t ret;
2197
2198         init_sync_kiocb(&kiocb, file);
2199         ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2200         if (ret == -EIOCBQUEUED)
2201                 ret = wait_on_sync_kiocb(&kiocb);
2202         return ret;
2203 }
2204
2205 ssize_t
2206 generic_file_write_nolock(struct file *file, const struct iovec *iov,
2207                                 unsigned long nr_segs, loff_t *ppos)
2208 {
2209         struct kiocb kiocb;
2210         ssize_t ret;
2211
2212         init_sync_kiocb(&kiocb, file);
2213         ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2214         if (-EIOCBQUEUED == ret)
2215                 ret = wait_on_sync_kiocb(&kiocb);
2216         return ret;
2217 }
2218 EXPORT_SYMBOL(generic_file_write_nolock);
2219
2220 ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
2221                                size_t count, loff_t pos)
2222 {
2223         struct file *file = iocb->ki_filp;
2224         struct address_space *mapping = file->f_mapping;
2225         struct inode *inode = mapping->host;
2226         ssize_t ret;
2227         struct iovec local_iov = { .iov_base = (void __user *)buf,
2228                                         .iov_len = count };
2229
2230         BUG_ON(iocb->ki_pos != pos);
2231
2232         mutex_lock(&inode->i_mutex);
2233         ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
2234                                                 &iocb->ki_pos);
2235         mutex_unlock(&inode->i_mutex);
2236
2237         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2238                 ssize_t err;
2239
2240                 err = sync_page_range(inode, mapping, pos, ret);
2241                 if (err < 0)
2242                         ret = err;
2243         }
2244         return ret;
2245 }
2246 EXPORT_SYMBOL(generic_file_aio_write);
2247
2248 ssize_t generic_file_write(struct file *file, const char __user *buf,
2249                            size_t count, loff_t *ppos)
2250 {
2251         struct address_space *mapping = file->f_mapping;
2252         struct inode *inode = mapping->host;
2253         ssize_t ret;
2254         struct iovec local_iov = { .iov_base = (void __user *)buf,
2255                                         .iov_len = count };
2256
2257         mutex_lock(&inode->i_mutex);
2258         ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
2259         mutex_unlock(&inode->i_mutex);
2260
2261         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2262                 ssize_t err;
2263
2264                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2265                 if (err < 0)
2266                         ret = err;
2267         }
2268         return ret;
2269 }
2270 EXPORT_SYMBOL(generic_file_write);
2271
2272 ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
2273                         unsigned long nr_segs, loff_t *ppos)
2274 {
2275         struct kiocb kiocb;
2276         ssize_t ret;
2277
2278         init_sync_kiocb(&kiocb, filp);
2279         ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
2280         if (-EIOCBQUEUED == ret)
2281                 ret = wait_on_sync_kiocb(&kiocb);
2282         return ret;
2283 }
2284 EXPORT_SYMBOL(generic_file_readv);
2285
2286 ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
2287                         unsigned long nr_segs, loff_t *ppos)
2288 {
2289         struct address_space *mapping = file->f_mapping;
2290         struct inode *inode = mapping->host;
2291         ssize_t ret;
2292
2293         mutex_lock(&inode->i_mutex);
2294         ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
2295         mutex_unlock(&inode->i_mutex);
2296
2297         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2298                 int err;
2299
2300                 err = sync_page_range(inode, mapping, *ppos - ret, ret);
2301                 if (err < 0)
2302                         ret = err;
2303         }
2304         return ret;
2305 }
2306 EXPORT_SYMBOL(generic_file_writev);
2307
2308 /*
2309  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2310  * went wrong during pagecache shootdown.
2311  */
2312 static ssize_t
2313 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2314         loff_t offset, unsigned long nr_segs)
2315 {
2316         struct file *file = iocb->ki_filp;
2317         struct address_space *mapping = file->f_mapping;
2318         ssize_t retval;
2319         size_t write_len = 0;
2320
2321         /*
2322          * If it's a write, unmap all mmappings of the file up-front.  This
2323          * will cause any pte dirty bits to be propagated into the pageframes
2324          * for the subsequent filemap_write_and_wait().
2325          */
2326         if (rw == WRITE) {
2327                 write_len = iov_length(iov, nr_segs);
2328                 if (mapping_mapped(mapping))
2329                         unmap_mapping_range(mapping, offset, write_len, 0);
2330         }
2331
2332         retval = filemap_write_and_wait(mapping);
2333         if (retval == 0) {
2334                 retval = mapping->a_ops->direct_IO(rw, iocb, iov,
2335                                                 offset, nr_segs);
2336                 if (rw == WRITE && mapping->nrpages) {
2337                         pgoff_t end = (offset + write_len - 1)
2338                                                 >> PAGE_CACHE_SHIFT;
2339                         int err = invalidate_inode_pages2_range(mapping,
2340                                         offset >> PAGE_CACHE_SHIFT, end);
2341                         if (err)
2342                                 retval = err;
2343                 }
2344         }
2345         return retval;
2346 }