ACPICA: Cleanup debug operand dump mechanism
[linux-2.6] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/module.h>
13 #include <linux/slab.h>
14 #include <linux/compiler.h>
15 #include <linux/fs.h>
16 #include <linux/uaccess.h>
17 #include <linux/aio.h>
18 #include <linux/capability.h>
19 #include <linux/kernel_stat.h>
20 #include <linux/mm.h>
21 #include <linux/swap.h>
22 #include <linux/mman.h>
23 #include <linux/pagemap.h>
24 #include <linux/file.h>
25 #include <linux/uio.h>
26 #include <linux/hash.h>
27 #include <linux/writeback.h>
28 #include <linux/backing-dev.h>
29 #include <linux/pagevec.h>
30 #include <linux/blkdev.h>
31 #include <linux/security.h>
32 #include <linux/syscalls.h>
33 #include <linux/cpuset.h>
34 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
35 #include <linux/memcontrol.h>
36 #include "internal.h"
37
38 /*
39  * FIXME: remove all knowledge of the buffer layer from the core VM
40  */
41 #include <linux/buffer_head.h> /* for generic_osync_inode */
42
43 #include <asm/mman.h>
44
45 static ssize_t
46 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
47         loff_t offset, unsigned long nr_segs);
48
49 /*
50  * Shared mappings implemented 30.11.1994. It's not fully working yet,
51  * though.
52  *
53  * Shared mappings now work. 15.8.1995  Bruno.
54  *
55  * finished 'unifying' the page and buffer cache and SMP-threaded the
56  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
57  *
58  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
59  */
60
61 /*
62  * Lock ordering:
63  *
64  *  ->i_mmap_lock               (vmtruncate)
65  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
66  *      ->swap_lock             (exclusive_swap_page, others)
67  *        ->mapping->tree_lock
68  *
69  *  ->i_mutex
70  *    ->i_mmap_lock             (truncate->unmap_mapping_range)
71  *
72  *  ->mmap_sem
73  *    ->i_mmap_lock
74  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
75  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
76  *
77  *  ->mmap_sem
78  *    ->lock_page               (access_process_vm)
79  *
80  *  ->i_mutex                   (generic_file_buffered_write)
81  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
82  *
83  *  ->i_mutex
84  *    ->i_alloc_sem             (various)
85  *
86  *  ->inode_lock
87  *    ->sb_lock                 (fs/fs-writeback.c)
88  *    ->mapping->tree_lock      (__sync_single_inode)
89  *
90  *  ->i_mmap_lock
91  *    ->anon_vma.lock           (vma_adjust)
92  *
93  *  ->anon_vma.lock
94  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
95  *
96  *  ->page_table_lock or pte_lock
97  *    ->swap_lock               (try_to_unmap_one)
98  *    ->private_lock            (try_to_unmap_one)
99  *    ->tree_lock               (try_to_unmap_one)
100  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
101  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
102  *    ->private_lock            (page_remove_rmap->set_page_dirty)
103  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
104  *    ->inode_lock              (page_remove_rmap->set_page_dirty)
105  *    ->inode_lock              (zap_pte_range->set_page_dirty)
106  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
107  *
108  *  ->task->proc_lock
109  *    ->dcache_lock             (proc_pid_lookup)
110  */
111
112 /*
113  * Remove a page from the page cache and free it. Caller has to make
114  * sure the page is locked and that nobody else uses it - or that usage
115  * is safe.  The caller must hold a write_lock on the mapping's tree_lock.
116  */
117 void __remove_from_page_cache(struct page *page)
118 {
119         struct address_space *mapping = page->mapping;
120
121         mem_cgroup_uncharge_page(page);
122         radix_tree_delete(&mapping->page_tree, page->index);
123         page->mapping = NULL;
124         mapping->nrpages--;
125         __dec_zone_page_state(page, NR_FILE_PAGES);
126         BUG_ON(page_mapped(page));
127
128         /*
129          * Some filesystems seem to re-dirty the page even after
130          * the VM has canceled the dirty bit (eg ext3 journaling).
131          *
132          * Fix it up by doing a final dirty accounting check after
133          * having removed the page entirely.
134          */
135         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
136                 dec_zone_page_state(page, NR_FILE_DIRTY);
137                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
138         }
139 }
140
141 void remove_from_page_cache(struct page *page)
142 {
143         struct address_space *mapping = page->mapping;
144
145         BUG_ON(!PageLocked(page));
146
147         write_lock_irq(&mapping->tree_lock);
148         __remove_from_page_cache(page);
149         write_unlock_irq(&mapping->tree_lock);
150 }
151
152 static int sync_page(void *word)
153 {
154         struct address_space *mapping;
155         struct page *page;
156
157         page = container_of((unsigned long *)word, struct page, flags);
158
159         /*
160          * page_mapping() is being called without PG_locked held.
161          * Some knowledge of the state and use of the page is used to
162          * reduce the requirements down to a memory barrier.
163          * The danger here is of a stale page_mapping() return value
164          * indicating a struct address_space different from the one it's
165          * associated with when it is associated with one.
166          * After smp_mb(), it's either the correct page_mapping() for
167          * the page, or an old page_mapping() and the page's own
168          * page_mapping() has gone NULL.
169          * The ->sync_page() address_space operation must tolerate
170          * page_mapping() going NULL. By an amazing coincidence,
171          * this comes about because none of the users of the page
172          * in the ->sync_page() methods make essential use of the
173          * page_mapping(), merely passing the page down to the backing
174          * device's unplug functions when it's non-NULL, which in turn
175          * ignore it for all cases but swap, where only page_private(page) is
176          * of interest. When page_mapping() does go NULL, the entire
177          * call stack gracefully ignores the page and returns.
178          * -- wli
179          */
180         smp_mb();
181         mapping = page_mapping(page);
182         if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
183                 mapping->a_ops->sync_page(page);
184         io_schedule();
185         return 0;
186 }
187
188 static int sync_page_killable(void *word)
189 {
190         sync_page(word);
191         return fatal_signal_pending(current) ? -EINTR : 0;
192 }
193
194 /**
195  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
196  * @mapping:    address space structure to write
197  * @start:      offset in bytes where the range starts
198  * @end:        offset in bytes where the range ends (inclusive)
199  * @sync_mode:  enable synchronous operation
200  *
201  * Start writeback against all of a mapping's dirty pages that lie
202  * within the byte offsets <start, end> inclusive.
203  *
204  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
205  * opposed to a regular memory cleansing writeback.  The difference between
206  * these two operations is that if a dirty page/buffer is encountered, it must
207  * be waited upon, and not just skipped over.
208  */
209 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
210                                 loff_t end, int sync_mode)
211 {
212         int ret;
213         struct writeback_control wbc = {
214                 .sync_mode = sync_mode,
215                 .nr_to_write = mapping->nrpages * 2,
216                 .range_start = start,
217                 .range_end = end,
218         };
219
220         if (!mapping_cap_writeback_dirty(mapping))
221                 return 0;
222
223         ret = do_writepages(mapping, &wbc);
224         return ret;
225 }
226
227 static inline int __filemap_fdatawrite(struct address_space *mapping,
228         int sync_mode)
229 {
230         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
231 }
232
233 int filemap_fdatawrite(struct address_space *mapping)
234 {
235         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
236 }
237 EXPORT_SYMBOL(filemap_fdatawrite);
238
239 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
240                                 loff_t end)
241 {
242         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
243 }
244 EXPORT_SYMBOL(filemap_fdatawrite_range);
245
246 /**
247  * filemap_flush - mostly a non-blocking flush
248  * @mapping:    target address_space
249  *
250  * This is a mostly non-blocking flush.  Not suitable for data-integrity
251  * purposes - I/O may not be started against all dirty pages.
252  */
253 int filemap_flush(struct address_space *mapping)
254 {
255         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
256 }
257 EXPORT_SYMBOL(filemap_flush);
258
259 /**
260  * wait_on_page_writeback_range - wait for writeback to complete
261  * @mapping:    target address_space
262  * @start:      beginning page index
263  * @end:        ending page index
264  *
265  * Wait for writeback to complete against pages indexed by start->end
266  * inclusive
267  */
268 int wait_on_page_writeback_range(struct address_space *mapping,
269                                 pgoff_t start, pgoff_t end)
270 {
271         struct pagevec pvec;
272         int nr_pages;
273         int ret = 0;
274         pgoff_t index;
275
276         if (end < start)
277                 return 0;
278
279         pagevec_init(&pvec, 0);
280         index = start;
281         while ((index <= end) &&
282                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
283                         PAGECACHE_TAG_WRITEBACK,
284                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
285                 unsigned i;
286
287                 for (i = 0; i < nr_pages; i++) {
288                         struct page *page = pvec.pages[i];
289
290                         /* until radix tree lookup accepts end_index */
291                         if (page->index > end)
292                                 continue;
293
294                         wait_on_page_writeback(page);
295                         if (PageError(page))
296                                 ret = -EIO;
297                 }
298                 pagevec_release(&pvec);
299                 cond_resched();
300         }
301
302         /* Check for outstanding write errors */
303         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
304                 ret = -ENOSPC;
305         if (test_and_clear_bit(AS_EIO, &mapping->flags))
306                 ret = -EIO;
307
308         return ret;
309 }
310
311 /**
312  * sync_page_range - write and wait on all pages in the passed range
313  * @inode:      target inode
314  * @mapping:    target address_space
315  * @pos:        beginning offset in pages to write
316  * @count:      number of bytes to write
317  *
318  * Write and wait upon all the pages in the passed range.  This is a "data
319  * integrity" operation.  It waits upon in-flight writeout before starting and
320  * waiting upon new writeout.  If there was an IO error, return it.
321  *
322  * We need to re-take i_mutex during the generic_osync_inode list walk because
323  * it is otherwise livelockable.
324  */
325 int sync_page_range(struct inode *inode, struct address_space *mapping,
326                         loff_t pos, loff_t count)
327 {
328         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
329         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
330         int ret;
331
332         if (!mapping_cap_writeback_dirty(mapping) || !count)
333                 return 0;
334         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
335         if (ret == 0) {
336                 mutex_lock(&inode->i_mutex);
337                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
338                 mutex_unlock(&inode->i_mutex);
339         }
340         if (ret == 0)
341                 ret = wait_on_page_writeback_range(mapping, start, end);
342         return ret;
343 }
344 EXPORT_SYMBOL(sync_page_range);
345
346 /**
347  * sync_page_range_nolock - write & wait on all pages in the passed range without locking
348  * @inode:      target inode
349  * @mapping:    target address_space
350  * @pos:        beginning offset in pages to write
351  * @count:      number of bytes to write
352  *
353  * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
354  * as it forces O_SYNC writers to different parts of the same file
355  * to be serialised right until io completion.
356  */
357 int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
358                            loff_t pos, loff_t count)
359 {
360         pgoff_t start = pos >> PAGE_CACHE_SHIFT;
361         pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
362         int ret;
363
364         if (!mapping_cap_writeback_dirty(mapping) || !count)
365                 return 0;
366         ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
367         if (ret == 0)
368                 ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
369         if (ret == 0)
370                 ret = wait_on_page_writeback_range(mapping, start, end);
371         return ret;
372 }
373 EXPORT_SYMBOL(sync_page_range_nolock);
374
375 /**
376  * filemap_fdatawait - wait for all under-writeback pages to complete
377  * @mapping: address space structure to wait for
378  *
379  * Walk the list of under-writeback pages of the given address space
380  * and wait for all of them.
381  */
382 int filemap_fdatawait(struct address_space *mapping)
383 {
384         loff_t i_size = i_size_read(mapping->host);
385
386         if (i_size == 0)
387                 return 0;
388
389         return wait_on_page_writeback_range(mapping, 0,
390                                 (i_size - 1) >> PAGE_CACHE_SHIFT);
391 }
392 EXPORT_SYMBOL(filemap_fdatawait);
393
394 int filemap_write_and_wait(struct address_space *mapping)
395 {
396         int err = 0;
397
398         if (mapping->nrpages) {
399                 err = filemap_fdatawrite(mapping);
400                 /*
401                  * Even if the above returned error, the pages may be
402                  * written partially (e.g. -ENOSPC), so we wait for it.
403                  * But the -EIO is special case, it may indicate the worst
404                  * thing (e.g. bug) happened, so we avoid waiting for it.
405                  */
406                 if (err != -EIO) {
407                         int err2 = filemap_fdatawait(mapping);
408                         if (!err)
409                                 err = err2;
410                 }
411         }
412         return err;
413 }
414 EXPORT_SYMBOL(filemap_write_and_wait);
415
416 /**
417  * filemap_write_and_wait_range - write out & wait on a file range
418  * @mapping:    the address_space for the pages
419  * @lstart:     offset in bytes where the range starts
420  * @lend:       offset in bytes where the range ends (inclusive)
421  *
422  * Write out and wait upon file offsets lstart->lend, inclusive.
423  *
424  * Note that `lend' is inclusive (describes the last byte to be written) so
425  * that this function can be used to write to the very end-of-file (end = -1).
426  */
427 int filemap_write_and_wait_range(struct address_space *mapping,
428                                  loff_t lstart, loff_t lend)
429 {
430         int err = 0;
431
432         if (mapping->nrpages) {
433                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
434                                                  WB_SYNC_ALL);
435                 /* See comment of filemap_write_and_wait() */
436                 if (err != -EIO) {
437                         int err2 = wait_on_page_writeback_range(mapping,
438                                                 lstart >> PAGE_CACHE_SHIFT,
439                                                 lend >> PAGE_CACHE_SHIFT);
440                         if (!err)
441                                 err = err2;
442                 }
443         }
444         return err;
445 }
446
447 /**
448  * add_to_page_cache - add newly allocated pagecache pages
449  * @page:       page to add
450  * @mapping:    the page's address_space
451  * @offset:     page index
452  * @gfp_mask:   page allocation mode
453  *
454  * This function is used to add newly allocated pagecache pages;
455  * the page is new, so we can just run SetPageLocked() against it.
456  * The other page state flags were set by rmqueue().
457  *
458  * This function does not add the page to the LRU.  The caller must do that.
459  */
460 int add_to_page_cache(struct page *page, struct address_space *mapping,
461                 pgoff_t offset, gfp_t gfp_mask)
462 {
463         int error = mem_cgroup_cache_charge(page, current->mm,
464                                         gfp_mask & ~__GFP_HIGHMEM);
465         if (error)
466                 goto out;
467
468         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
469         if (error == 0) {
470                 write_lock_irq(&mapping->tree_lock);
471                 error = radix_tree_insert(&mapping->page_tree, offset, page);
472                 if (!error) {
473                         page_cache_get(page);
474                         SetPageLocked(page);
475                         page->mapping = mapping;
476                         page->index = offset;
477                         mapping->nrpages++;
478                         __inc_zone_page_state(page, NR_FILE_PAGES);
479                 } else
480                         mem_cgroup_uncharge_page(page);
481
482                 write_unlock_irq(&mapping->tree_lock);
483                 radix_tree_preload_end();
484         } else
485                 mem_cgroup_uncharge_page(page);
486 out:
487         return error;
488 }
489 EXPORT_SYMBOL(add_to_page_cache);
490
491 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
492                                 pgoff_t offset, gfp_t gfp_mask)
493 {
494         int ret = add_to_page_cache(page, mapping, offset, gfp_mask);
495         if (ret == 0)
496                 lru_cache_add(page);
497         return ret;
498 }
499
500 #ifdef CONFIG_NUMA
501 struct page *__page_cache_alloc(gfp_t gfp)
502 {
503         if (cpuset_do_page_mem_spread()) {
504                 int n = cpuset_mem_spread_node();
505                 return alloc_pages_node(n, gfp, 0);
506         }
507         return alloc_pages(gfp, 0);
508 }
509 EXPORT_SYMBOL(__page_cache_alloc);
510 #endif
511
512 static int __sleep_on_page_lock(void *word)
513 {
514         io_schedule();
515         return 0;
516 }
517
518 /*
519  * In order to wait for pages to become available there must be
520  * waitqueues associated with pages. By using a hash table of
521  * waitqueues where the bucket discipline is to maintain all
522  * waiters on the same queue and wake all when any of the pages
523  * become available, and for the woken contexts to check to be
524  * sure the appropriate page became available, this saves space
525  * at a cost of "thundering herd" phenomena during rare hash
526  * collisions.
527  */
528 static wait_queue_head_t *page_waitqueue(struct page *page)
529 {
530         const struct zone *zone = page_zone(page);
531
532         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
533 }
534
535 static inline void wake_up_page(struct page *page, int bit)
536 {
537         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
538 }
539
540 void wait_on_page_bit(struct page *page, int bit_nr)
541 {
542         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
543
544         if (test_bit(bit_nr, &page->flags))
545                 __wait_on_bit(page_waitqueue(page), &wait, sync_page,
546                                                         TASK_UNINTERRUPTIBLE);
547 }
548 EXPORT_SYMBOL(wait_on_page_bit);
549
550 /**
551  * unlock_page - unlock a locked page
552  * @page: the page
553  *
554  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
555  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
556  * mechananism between PageLocked pages and PageWriteback pages is shared.
557  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
558  *
559  * The first mb is necessary to safely close the critical section opened by the
560  * TestSetPageLocked(), the second mb is necessary to enforce ordering between
561  * the clear_bit and the read of the waitqueue (to avoid SMP races with a
562  * parallel wait_on_page_locked()).
563  */
564 void unlock_page(struct page *page)
565 {
566         smp_mb__before_clear_bit();
567         if (!TestClearPageLocked(page))
568                 BUG();
569         smp_mb__after_clear_bit(); 
570         wake_up_page(page, PG_locked);
571 }
572 EXPORT_SYMBOL(unlock_page);
573
574 /**
575  * end_page_writeback - end writeback against a page
576  * @page: the page
577  */
578 void end_page_writeback(struct page *page)
579 {
580         if (TestClearPageReclaim(page))
581                 rotate_reclaimable_page(page);
582
583         if (!test_clear_page_writeback(page))
584                 BUG();
585
586         smp_mb__after_clear_bit();
587         wake_up_page(page, PG_writeback);
588 }
589 EXPORT_SYMBOL(end_page_writeback);
590
591 /**
592  * __lock_page - get a lock on the page, assuming we need to sleep to get it
593  * @page: the page to lock
594  *
595  * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
596  * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
597  * chances are that on the second loop, the block layer's plug list is empty,
598  * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
599  */
600 void __lock_page(struct page *page)
601 {
602         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
603
604         __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
605                                                         TASK_UNINTERRUPTIBLE);
606 }
607 EXPORT_SYMBOL(__lock_page);
608
609 int __lock_page_killable(struct page *page)
610 {
611         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
612
613         return __wait_on_bit_lock(page_waitqueue(page), &wait,
614                                         sync_page_killable, TASK_KILLABLE);
615 }
616
617 /**
618  * __lock_page_nosync - get a lock on the page, without calling sync_page()
619  * @page: the page to lock
620  *
621  * Variant of lock_page that does not require the caller to hold a reference
622  * on the page's mapping.
623  */
624 void __lock_page_nosync(struct page *page)
625 {
626         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
627         __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
628                                                         TASK_UNINTERRUPTIBLE);
629 }
630
631 /**
632  * find_get_page - find and get a page reference
633  * @mapping: the address_space to search
634  * @offset: the page index
635  *
636  * Is there a pagecache struct page at the given (mapping, offset) tuple?
637  * If yes, increment its refcount and return it; if no, return NULL.
638  */
639 struct page * find_get_page(struct address_space *mapping, pgoff_t offset)
640 {
641         struct page *page;
642
643         read_lock_irq(&mapping->tree_lock);
644         page = radix_tree_lookup(&mapping->page_tree, offset);
645         if (page)
646                 page_cache_get(page);
647         read_unlock_irq(&mapping->tree_lock);
648         return page;
649 }
650 EXPORT_SYMBOL(find_get_page);
651
652 /**
653  * find_lock_page - locate, pin and lock a pagecache page
654  * @mapping: the address_space to search
655  * @offset: the page index
656  *
657  * Locates the desired pagecache page, locks it, increments its reference
658  * count and returns its address.
659  *
660  * Returns zero if the page was not present. find_lock_page() may sleep.
661  */
662 struct page *find_lock_page(struct address_space *mapping,
663                                 pgoff_t offset)
664 {
665         struct page *page;
666
667 repeat:
668         read_lock_irq(&mapping->tree_lock);
669         page = radix_tree_lookup(&mapping->page_tree, offset);
670         if (page) {
671                 page_cache_get(page);
672                 if (TestSetPageLocked(page)) {
673                         read_unlock_irq(&mapping->tree_lock);
674                         __lock_page(page);
675
676                         /* Has the page been truncated while we slept? */
677                         if (unlikely(page->mapping != mapping)) {
678                                 unlock_page(page);
679                                 page_cache_release(page);
680                                 goto repeat;
681                         }
682                         VM_BUG_ON(page->index != offset);
683                         goto out;
684                 }
685         }
686         read_unlock_irq(&mapping->tree_lock);
687 out:
688         return page;
689 }
690 EXPORT_SYMBOL(find_lock_page);
691
692 /**
693  * find_or_create_page - locate or add a pagecache page
694  * @mapping: the page's address_space
695  * @index: the page's index into the mapping
696  * @gfp_mask: page allocation mode
697  *
698  * Locates a page in the pagecache.  If the page is not present, a new page
699  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
700  * LRU list.  The returned page is locked and has its reference count
701  * incremented.
702  *
703  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
704  * allocation!
705  *
706  * find_or_create_page() returns the desired page's address, or zero on
707  * memory exhaustion.
708  */
709 struct page *find_or_create_page(struct address_space *mapping,
710                 pgoff_t index, gfp_t gfp_mask)
711 {
712         struct page *page;
713         int err;
714 repeat:
715         page = find_lock_page(mapping, index);
716         if (!page) {
717                 page = __page_cache_alloc(gfp_mask);
718                 if (!page)
719                         return NULL;
720                 err = add_to_page_cache_lru(page, mapping, index, gfp_mask);
721                 if (unlikely(err)) {
722                         page_cache_release(page);
723                         page = NULL;
724                         if (err == -EEXIST)
725                                 goto repeat;
726                 }
727         }
728         return page;
729 }
730 EXPORT_SYMBOL(find_or_create_page);
731
732 /**
733  * find_get_pages - gang pagecache lookup
734  * @mapping:    The address_space to search
735  * @start:      The starting page index
736  * @nr_pages:   The maximum number of pages
737  * @pages:      Where the resulting pages are placed
738  *
739  * find_get_pages() will search for and return a group of up to
740  * @nr_pages pages in the mapping.  The pages are placed at @pages.
741  * find_get_pages() takes a reference against the returned pages.
742  *
743  * The search returns a group of mapping-contiguous pages with ascending
744  * indexes.  There may be holes in the indices due to not-present pages.
745  *
746  * find_get_pages() returns the number of pages which were found.
747  */
748 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
749                             unsigned int nr_pages, struct page **pages)
750 {
751         unsigned int i;
752         unsigned int ret;
753
754         read_lock_irq(&mapping->tree_lock);
755         ret = radix_tree_gang_lookup(&mapping->page_tree,
756                                 (void **)pages, start, nr_pages);
757         for (i = 0; i < ret; i++)
758                 page_cache_get(pages[i]);
759         read_unlock_irq(&mapping->tree_lock);
760         return ret;
761 }
762
763 /**
764  * find_get_pages_contig - gang contiguous pagecache lookup
765  * @mapping:    The address_space to search
766  * @index:      The starting page index
767  * @nr_pages:   The maximum number of pages
768  * @pages:      Where the resulting pages are placed
769  *
770  * find_get_pages_contig() works exactly like find_get_pages(), except
771  * that the returned number of pages are guaranteed to be contiguous.
772  *
773  * find_get_pages_contig() returns the number of pages which were found.
774  */
775 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
776                                unsigned int nr_pages, struct page **pages)
777 {
778         unsigned int i;
779         unsigned int ret;
780
781         read_lock_irq(&mapping->tree_lock);
782         ret = radix_tree_gang_lookup(&mapping->page_tree,
783                                 (void **)pages, index, nr_pages);
784         for (i = 0; i < ret; i++) {
785                 if (pages[i]->mapping == NULL || pages[i]->index != index)
786                         break;
787
788                 page_cache_get(pages[i]);
789                 index++;
790         }
791         read_unlock_irq(&mapping->tree_lock);
792         return i;
793 }
794 EXPORT_SYMBOL(find_get_pages_contig);
795
796 /**
797  * find_get_pages_tag - find and return pages that match @tag
798  * @mapping:    the address_space to search
799  * @index:      the starting page index
800  * @tag:        the tag index
801  * @nr_pages:   the maximum number of pages
802  * @pages:      where the resulting pages are placed
803  *
804  * Like find_get_pages, except we only return pages which are tagged with
805  * @tag.   We update @index to index the next page for the traversal.
806  */
807 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
808                         int tag, unsigned int nr_pages, struct page **pages)
809 {
810         unsigned int i;
811         unsigned int ret;
812
813         read_lock_irq(&mapping->tree_lock);
814         ret = radix_tree_gang_lookup_tag(&mapping->page_tree,
815                                 (void **)pages, *index, nr_pages, tag);
816         for (i = 0; i < ret; i++)
817                 page_cache_get(pages[i]);
818         if (ret)
819                 *index = pages[ret - 1]->index + 1;
820         read_unlock_irq(&mapping->tree_lock);
821         return ret;
822 }
823 EXPORT_SYMBOL(find_get_pages_tag);
824
825 /**
826  * grab_cache_page_nowait - returns locked page at given index in given cache
827  * @mapping: target address_space
828  * @index: the page index
829  *
830  * Same as grab_cache_page(), but do not wait if the page is unavailable.
831  * This is intended for speculative data generators, where the data can
832  * be regenerated if the page couldn't be grabbed.  This routine should
833  * be safe to call while holding the lock for another page.
834  *
835  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
836  * and deadlock against the caller's locked page.
837  */
838 struct page *
839 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
840 {
841         struct page *page = find_get_page(mapping, index);
842
843         if (page) {
844                 if (!TestSetPageLocked(page))
845                         return page;
846                 page_cache_release(page);
847                 return NULL;
848         }
849         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
850         if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
851                 page_cache_release(page);
852                 page = NULL;
853         }
854         return page;
855 }
856 EXPORT_SYMBOL(grab_cache_page_nowait);
857
858 /*
859  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
860  * a _large_ part of the i/o request. Imagine the worst scenario:
861  *
862  *      ---R__________________________________________B__________
863  *         ^ reading here                             ^ bad block(assume 4k)
864  *
865  * read(R) => miss => readahead(R...B) => media error => frustrating retries
866  * => failing the whole request => read(R) => read(R+1) =>
867  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
868  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
869  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
870  *
871  * It is going insane. Fix it by quickly scaling down the readahead size.
872  */
873 static void shrink_readahead_size_eio(struct file *filp,
874                                         struct file_ra_state *ra)
875 {
876         if (!ra->ra_pages)
877                 return;
878
879         ra->ra_pages /= 4;
880 }
881
882 /**
883  * do_generic_file_read - generic file read routine
884  * @filp:       the file to read
885  * @ppos:       current file position
886  * @desc:       read_descriptor
887  * @actor:      read method
888  *
889  * This is a generic file read routine, and uses the
890  * mapping->a_ops->readpage() function for the actual low-level stuff.
891  *
892  * This is really ugly. But the goto's actually try to clarify some
893  * of the logic when it comes to error handling etc.
894  */
895 static void do_generic_file_read(struct file *filp, loff_t *ppos,
896                 read_descriptor_t *desc, read_actor_t actor)
897 {
898         struct address_space *mapping = filp->f_mapping;
899         struct inode *inode = mapping->host;
900         struct file_ra_state *ra = &filp->f_ra;
901         pgoff_t index;
902         pgoff_t last_index;
903         pgoff_t prev_index;
904         unsigned long offset;      /* offset into pagecache page */
905         unsigned int prev_offset;
906         int error;
907
908         index = *ppos >> PAGE_CACHE_SHIFT;
909         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
910         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
911         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
912         offset = *ppos & ~PAGE_CACHE_MASK;
913
914         for (;;) {
915                 struct page *page;
916                 pgoff_t end_index;
917                 loff_t isize;
918                 unsigned long nr, ret;
919
920                 cond_resched();
921 find_page:
922                 page = find_get_page(mapping, index);
923                 if (!page) {
924                         page_cache_sync_readahead(mapping,
925                                         ra, filp,
926                                         index, last_index - index);
927                         page = find_get_page(mapping, index);
928                         if (unlikely(page == NULL))
929                                 goto no_cached_page;
930                 }
931                 if (PageReadahead(page)) {
932                         page_cache_async_readahead(mapping,
933                                         ra, filp, page,
934                                         index, last_index - index);
935                 }
936                 if (!PageUptodate(page))
937                         goto page_not_up_to_date;
938 page_ok:
939                 /*
940                  * i_size must be checked after we know the page is Uptodate.
941                  *
942                  * Checking i_size after the check allows us to calculate
943                  * the correct value for "nr", which means the zero-filled
944                  * part of the page is not copied back to userspace (unless
945                  * another truncate extends the file - this is desired though).
946                  */
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
966                 /* If users can be writing to this page using arbitrary
967                  * virtual addresses, take care about potential aliasing
968                  * before reading the page on the kernel side.
969                  */
970                 if (mapping_writably_mapped(mapping))
971                         flush_dcache_page(page);
972
973                 /*
974                  * When a sequential read accesses a page several times,
975                  * only mark it as accessed the first time.
976                  */
977                 if (prev_index != index || offset != prev_offset)
978                         mark_page_accessed(page);
979                 prev_index = index;
980
981                 /*
982                  * Ok, we have the page, and it's up-to-date, so
983                  * now we can copy it to user space...
984                  *
985                  * The actor routine returns how many bytes were actually used..
986                  * NOTE! This may not be the same as how much of a user buffer
987                  * we filled up (we may be padding etc), so we can only update
988                  * "pos" here (the actor routine has to update the user buffer
989                  * pointers and the remaining count).
990                  */
991                 ret = actor(desc, page, offset, nr);
992                 offset += ret;
993                 index += offset >> PAGE_CACHE_SHIFT;
994                 offset &= ~PAGE_CACHE_MASK;
995                 prev_offset = offset;
996
997                 page_cache_release(page);
998                 if (ret == nr && desc->count)
999                         continue;
1000                 goto out;
1001
1002 page_not_up_to_date:
1003                 /* Get exclusive access to the page ... */
1004                 if (lock_page_killable(page))
1005                         goto readpage_eio;
1006
1007                 /* Did it get truncated before we got the lock? */
1008                 if (!page->mapping) {
1009                         unlock_page(page);
1010                         page_cache_release(page);
1011                         continue;
1012                 }
1013
1014                 /* Did somebody else fill it already? */
1015                 if (PageUptodate(page)) {
1016                         unlock_page(page);
1017                         goto page_ok;
1018                 }
1019
1020 readpage:
1021                 /* Start the actual read. The read will unlock the page. */
1022                 error = mapping->a_ops->readpage(filp, page);
1023
1024                 if (unlikely(error)) {
1025                         if (error == AOP_TRUNCATED_PAGE) {
1026                                 page_cache_release(page);
1027                                 goto find_page;
1028                         }
1029                         goto readpage_error;
1030                 }
1031
1032                 if (!PageUptodate(page)) {
1033                         if (lock_page_killable(page))
1034                                 goto readpage_eio;
1035                         if (!PageUptodate(page)) {
1036                                 if (page->mapping == NULL) {
1037                                         /*
1038                                          * invalidate_inode_pages got it
1039                                          */
1040                                         unlock_page(page);
1041                                         page_cache_release(page);
1042                                         goto find_page;
1043                                 }
1044                                 unlock_page(page);
1045                                 shrink_readahead_size_eio(filp, ra);
1046                                 goto readpage_eio;
1047                         }
1048                         unlock_page(page);
1049                 }
1050
1051                 goto page_ok;
1052
1053 readpage_eio:
1054                 error = -EIO;
1055 readpage_error:
1056                 /* UHHUH! A synchronous read error occurred. Report it */
1057                 desc->error = error;
1058                 page_cache_release(page);
1059                 goto out;
1060
1061 no_cached_page:
1062                 /*
1063                  * Ok, it wasn't cached, so we need to create a new
1064                  * page..
1065                  */
1066                 page = page_cache_alloc_cold(mapping);
1067                 if (!page) {
1068                         desc->error = -ENOMEM;
1069                         goto out;
1070                 }
1071                 error = add_to_page_cache_lru(page, mapping,
1072                                                 index, GFP_KERNEL);
1073                 if (error) {
1074                         page_cache_release(page);
1075                         if (error == -EEXIST)
1076                                 goto find_page;
1077                         desc->error = error;
1078                         goto out;
1079                 }
1080                 goto readpage;
1081         }
1082
1083 out:
1084         ra->prev_pos = prev_index;
1085         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1086         ra->prev_pos |= prev_offset;
1087
1088         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1089         if (filp)
1090                 file_accessed(filp);
1091 }
1092
1093 int file_read_actor(read_descriptor_t *desc, struct page *page,
1094                         unsigned long offset, unsigned long size)
1095 {
1096         char *kaddr;
1097         unsigned long left, count = desc->count;
1098
1099         if (size > count)
1100                 size = count;
1101
1102         /*
1103          * Faults on the destination of a read are common, so do it before
1104          * taking the kmap.
1105          */
1106         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1107                 kaddr = kmap_atomic(page, KM_USER0);
1108                 left = __copy_to_user_inatomic(desc->arg.buf,
1109                                                 kaddr + offset, size);
1110                 kunmap_atomic(kaddr, KM_USER0);
1111                 if (left == 0)
1112                         goto success;
1113         }
1114
1115         /* Do it the slow way */
1116         kaddr = kmap(page);
1117         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1118         kunmap(page);
1119
1120         if (left) {
1121                 size -= left;
1122                 desc->error = -EFAULT;
1123         }
1124 success:
1125         desc->count = count - size;
1126         desc->written += size;
1127         desc->arg.buf += size;
1128         return size;
1129 }
1130
1131 /*
1132  * Performs necessary checks before doing a write
1133  * @iov:        io vector request
1134  * @nr_segs:    number of segments in the iovec
1135  * @count:      number of bytes to write
1136  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1137  *
1138  * Adjust number of segments and amount of bytes to write (nr_segs should be
1139  * properly initialized first). Returns appropriate error code that caller
1140  * should return or zero in case that write should be allowed.
1141  */
1142 int generic_segment_checks(const struct iovec *iov,
1143                         unsigned long *nr_segs, size_t *count, int access_flags)
1144 {
1145         unsigned long   seg;
1146         size_t cnt = 0;
1147         for (seg = 0; seg < *nr_segs; seg++) {
1148                 const struct iovec *iv = &iov[seg];
1149
1150                 /*
1151                  * If any segment has a negative length, or the cumulative
1152                  * length ever wraps negative then return -EINVAL.
1153                  */
1154                 cnt += iv->iov_len;
1155                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1156                         return -EINVAL;
1157                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1158                         continue;
1159                 if (seg == 0)
1160                         return -EFAULT;
1161                 *nr_segs = seg;
1162                 cnt -= iv->iov_len;     /* This segment is no good */
1163                 break;
1164         }
1165         *count = cnt;
1166         return 0;
1167 }
1168 EXPORT_SYMBOL(generic_segment_checks);
1169
1170 /**
1171  * generic_file_aio_read - generic filesystem read routine
1172  * @iocb:       kernel I/O control block
1173  * @iov:        io vector request
1174  * @nr_segs:    number of segments in the iovec
1175  * @pos:        current file position
1176  *
1177  * This is the "read()" routine for all filesystems
1178  * that can use the page cache directly.
1179  */
1180 ssize_t
1181 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1182                 unsigned long nr_segs, loff_t pos)
1183 {
1184         struct file *filp = iocb->ki_filp;
1185         ssize_t retval;
1186         unsigned long seg;
1187         size_t count;
1188         loff_t *ppos = &iocb->ki_pos;
1189
1190         count = 0;
1191         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1192         if (retval)
1193                 return retval;
1194
1195         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1196         if (filp->f_flags & O_DIRECT) {
1197                 loff_t size;
1198                 struct address_space *mapping;
1199                 struct inode *inode;
1200
1201                 mapping = filp->f_mapping;
1202                 inode = mapping->host;
1203                 retval = 0;
1204                 if (!count)
1205                         goto out; /* skip atime */
1206                 size = i_size_read(inode);
1207                 if (pos < size) {
1208                         retval = generic_file_direct_IO(READ, iocb,
1209                                                 iov, pos, nr_segs);
1210                         if (retval > 0)
1211                                 *ppos = pos + retval;
1212                 }
1213                 if (likely(retval != 0)) {
1214                         file_accessed(filp);
1215                         goto out;
1216                 }
1217         }
1218
1219         retval = 0;
1220         if (count) {
1221                 for (seg = 0; seg < nr_segs; seg++) {
1222                         read_descriptor_t desc;
1223
1224                         desc.written = 0;
1225                         desc.arg.buf = iov[seg].iov_base;
1226                         desc.count = iov[seg].iov_len;
1227                         if (desc.count == 0)
1228                                 continue;
1229                         desc.error = 0;
1230                         do_generic_file_read(filp,ppos,&desc,file_read_actor);
1231                         retval += desc.written;
1232                         if (desc.error) {
1233                                 retval = retval ?: desc.error;
1234                                 break;
1235                         }
1236                         if (desc.count > 0)
1237                                 break;
1238                 }
1239         }
1240 out:
1241         return retval;
1242 }
1243 EXPORT_SYMBOL(generic_file_aio_read);
1244
1245 static ssize_t
1246 do_readahead(struct address_space *mapping, struct file *filp,
1247              pgoff_t index, unsigned long nr)
1248 {
1249         if (!mapping || !mapping->a_ops || !mapping->a_ops->readpage)
1250                 return -EINVAL;
1251
1252         force_page_cache_readahead(mapping, filp, index,
1253                                         max_sane_readahead(nr));
1254         return 0;
1255 }
1256
1257 asmlinkage ssize_t sys_readahead(int fd, loff_t offset, size_t count)
1258 {
1259         ssize_t ret;
1260         struct file *file;
1261
1262         ret = -EBADF;
1263         file = fget(fd);
1264         if (file) {
1265                 if (file->f_mode & FMODE_READ) {
1266                         struct address_space *mapping = file->f_mapping;
1267                         pgoff_t start = offset >> PAGE_CACHE_SHIFT;
1268                         pgoff_t end = (offset + count - 1) >> PAGE_CACHE_SHIFT;
1269                         unsigned long len = end - start + 1;
1270                         ret = do_readahead(mapping, file, start, len);
1271                 }
1272                 fput(file);
1273         }
1274         return ret;
1275 }
1276
1277 #ifdef CONFIG_MMU
1278 /**
1279  * page_cache_read - adds requested page to the page cache if not already there
1280  * @file:       file to read
1281  * @offset:     page index
1282  *
1283  * This adds the requested page to the page cache if it isn't already there,
1284  * and schedules an I/O to read in its contents from disk.
1285  */
1286 static int page_cache_read(struct file *file, pgoff_t offset)
1287 {
1288         struct address_space *mapping = file->f_mapping;
1289         struct page *page; 
1290         int ret;
1291
1292         do {
1293                 page = page_cache_alloc_cold(mapping);
1294                 if (!page)
1295                         return -ENOMEM;
1296
1297                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1298                 if (ret == 0)
1299                         ret = mapping->a_ops->readpage(file, page);
1300                 else if (ret == -EEXIST)
1301                         ret = 0; /* losing race to add is OK */
1302
1303                 page_cache_release(page);
1304
1305         } while (ret == AOP_TRUNCATED_PAGE);
1306                 
1307         return ret;
1308 }
1309
1310 #define MMAP_LOTSAMISS  (100)
1311
1312 /**
1313  * filemap_fault - read in file data for page fault handling
1314  * @vma:        vma in which the fault was taken
1315  * @vmf:        struct vm_fault containing details of the fault
1316  *
1317  * filemap_fault() is invoked via the vma operations vector for a
1318  * mapped memory region to read in file data during a page fault.
1319  *
1320  * The goto's are kind of ugly, but this streamlines the normal case of having
1321  * it in the page cache, and handles the special cases reasonably without
1322  * having a lot of duplicated code.
1323  */
1324 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1325 {
1326         int error;
1327         struct file *file = vma->vm_file;
1328         struct address_space *mapping = file->f_mapping;
1329         struct file_ra_state *ra = &file->f_ra;
1330         struct inode *inode = mapping->host;
1331         struct page *page;
1332         pgoff_t size;
1333         int did_readaround = 0;
1334         int ret = 0;
1335
1336         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1337         if (vmf->pgoff >= size)
1338                 return VM_FAULT_SIGBUS;
1339
1340         /* If we don't want any read-ahead, don't bother */
1341         if (VM_RandomReadHint(vma))
1342                 goto no_cached_page;
1343
1344         /*
1345          * Do we have something in the page cache already?
1346          */
1347 retry_find:
1348         page = find_lock_page(mapping, vmf->pgoff);
1349         /*
1350          * For sequential accesses, we use the generic readahead logic.
1351          */
1352         if (VM_SequentialReadHint(vma)) {
1353                 if (!page) {
1354                         page_cache_sync_readahead(mapping, ra, file,
1355                                                            vmf->pgoff, 1);
1356                         page = find_lock_page(mapping, vmf->pgoff);
1357                         if (!page)
1358                                 goto no_cached_page;
1359                 }
1360                 if (PageReadahead(page)) {
1361                         page_cache_async_readahead(mapping, ra, file, page,
1362                                                            vmf->pgoff, 1);
1363                 }
1364         }
1365
1366         if (!page) {
1367                 unsigned long ra_pages;
1368
1369                 ra->mmap_miss++;
1370
1371                 /*
1372                  * Do we miss much more than hit in this file? If so,
1373                  * stop bothering with read-ahead. It will only hurt.
1374                  */
1375                 if (ra->mmap_miss > MMAP_LOTSAMISS)
1376                         goto no_cached_page;
1377
1378                 /*
1379                  * To keep the pgmajfault counter straight, we need to
1380                  * check did_readaround, as this is an inner loop.
1381                  */
1382                 if (!did_readaround) {
1383                         ret = VM_FAULT_MAJOR;
1384                         count_vm_event(PGMAJFAULT);
1385                 }
1386                 did_readaround = 1;
1387                 ra_pages = max_sane_readahead(file->f_ra.ra_pages);
1388                 if (ra_pages) {
1389                         pgoff_t start = 0;
1390
1391                         if (vmf->pgoff > ra_pages / 2)
1392                                 start = vmf->pgoff - ra_pages / 2;
1393                         do_page_cache_readahead(mapping, file, start, ra_pages);
1394                 }
1395                 page = find_lock_page(mapping, vmf->pgoff);
1396                 if (!page)
1397                         goto no_cached_page;
1398         }
1399
1400         if (!did_readaround)
1401                 ra->mmap_miss--;
1402
1403         /*
1404          * We have a locked page in the page cache, now we need to check
1405          * that it's up-to-date. If not, it is going to be due to an error.
1406          */
1407         if (unlikely(!PageUptodate(page)))
1408                 goto page_not_uptodate;
1409
1410         /* Must recheck i_size under page lock */
1411         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1412         if (unlikely(vmf->pgoff >= size)) {
1413                 unlock_page(page);
1414                 page_cache_release(page);
1415                 return VM_FAULT_SIGBUS;
1416         }
1417
1418         /*
1419          * Found the page and have a reference on it.
1420          */
1421         mark_page_accessed(page);
1422         ra->prev_pos = (loff_t)page->index << PAGE_CACHE_SHIFT;
1423         vmf->page = page;
1424         return ret | VM_FAULT_LOCKED;
1425
1426 no_cached_page:
1427         /*
1428          * We're only likely to ever get here if MADV_RANDOM is in
1429          * effect.
1430          */
1431         error = page_cache_read(file, vmf->pgoff);
1432
1433         /*
1434          * The page we want has now been added to the page cache.
1435          * In the unlikely event that someone removed it in the
1436          * meantime, we'll just come back here and read it again.
1437          */
1438         if (error >= 0)
1439                 goto retry_find;
1440
1441         /*
1442          * An error return from page_cache_read can result if the
1443          * system is low on memory, or a problem occurs while trying
1444          * to schedule I/O.
1445          */
1446         if (error == -ENOMEM)
1447                 return VM_FAULT_OOM;
1448         return VM_FAULT_SIGBUS;
1449
1450 page_not_uptodate:
1451         /* IO error path */
1452         if (!did_readaround) {
1453                 ret = VM_FAULT_MAJOR;
1454                 count_vm_event(PGMAJFAULT);
1455         }
1456
1457         /*
1458          * Umm, take care of errors if the page isn't up-to-date.
1459          * Try to re-read it _once_. We do this synchronously,
1460          * because there really aren't any performance issues here
1461          * and we need to check for errors.
1462          */
1463         ClearPageError(page);
1464         error = mapping->a_ops->readpage(file, page);
1465         if (!error) {
1466                 wait_on_page_locked(page);
1467                 if (!PageUptodate(page))
1468                         error = -EIO;
1469         }
1470         page_cache_release(page);
1471
1472         if (!error || error == AOP_TRUNCATED_PAGE)
1473                 goto retry_find;
1474
1475         /* Things didn't work out. Return zero to tell the mm layer so. */
1476         shrink_readahead_size_eio(file, ra);
1477         return VM_FAULT_SIGBUS;
1478 }
1479 EXPORT_SYMBOL(filemap_fault);
1480
1481 struct vm_operations_struct generic_file_vm_ops = {
1482         .fault          = filemap_fault,
1483 };
1484
1485 /* This is used for a general mmap of a disk file */
1486
1487 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1488 {
1489         struct address_space *mapping = file->f_mapping;
1490
1491         if (!mapping->a_ops->readpage)
1492                 return -ENOEXEC;
1493         file_accessed(file);
1494         vma->vm_ops = &generic_file_vm_ops;
1495         vma->vm_flags |= VM_CAN_NONLINEAR;
1496         return 0;
1497 }
1498
1499 /*
1500  * This is for filesystems which do not implement ->writepage.
1501  */
1502 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1503 {
1504         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1505                 return -EINVAL;
1506         return generic_file_mmap(file, vma);
1507 }
1508 #else
1509 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1510 {
1511         return -ENOSYS;
1512 }
1513 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1514 {
1515         return -ENOSYS;
1516 }
1517 #endif /* CONFIG_MMU */
1518
1519 EXPORT_SYMBOL(generic_file_mmap);
1520 EXPORT_SYMBOL(generic_file_readonly_mmap);
1521
1522 static struct page *__read_cache_page(struct address_space *mapping,
1523                                 pgoff_t index,
1524                                 int (*filler)(void *,struct page*),
1525                                 void *data)
1526 {
1527         struct page *page;
1528         int err;
1529 repeat:
1530         page = find_get_page(mapping, index);
1531         if (!page) {
1532                 page = page_cache_alloc_cold(mapping);
1533                 if (!page)
1534                         return ERR_PTR(-ENOMEM);
1535                 err = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
1536                 if (unlikely(err)) {
1537                         page_cache_release(page);
1538                         if (err == -EEXIST)
1539                                 goto repeat;
1540                         /* Presumably ENOMEM for radix tree node */
1541                         return ERR_PTR(err);
1542                 }
1543                 err = filler(data, page);
1544                 if (err < 0) {
1545                         page_cache_release(page);
1546                         page = ERR_PTR(err);
1547                 }
1548         }
1549         return page;
1550 }
1551
1552 /**
1553  * read_cache_page_async - read into page cache, fill it if needed
1554  * @mapping:    the page's address_space
1555  * @index:      the page index
1556  * @filler:     function to perform the read
1557  * @data:       destination for read data
1558  *
1559  * Same as read_cache_page, but don't wait for page to become unlocked
1560  * after submitting it to the filler.
1561  *
1562  * Read into the page cache. If a page already exists, and PageUptodate() is
1563  * not set, try to fill the page but don't wait for it to become unlocked.
1564  *
1565  * If the page does not get brought uptodate, return -EIO.
1566  */
1567 struct page *read_cache_page_async(struct address_space *mapping,
1568                                 pgoff_t index,
1569                                 int (*filler)(void *,struct page*),
1570                                 void *data)
1571 {
1572         struct page *page;
1573         int err;
1574
1575 retry:
1576         page = __read_cache_page(mapping, index, filler, data);
1577         if (IS_ERR(page))
1578                 return page;
1579         if (PageUptodate(page))
1580                 goto out;
1581
1582         lock_page(page);
1583         if (!page->mapping) {
1584                 unlock_page(page);
1585                 page_cache_release(page);
1586                 goto retry;
1587         }
1588         if (PageUptodate(page)) {
1589                 unlock_page(page);
1590                 goto out;
1591         }
1592         err = filler(data, page);
1593         if (err < 0) {
1594                 page_cache_release(page);
1595                 return ERR_PTR(err);
1596         }
1597 out:
1598         mark_page_accessed(page);
1599         return page;
1600 }
1601 EXPORT_SYMBOL(read_cache_page_async);
1602
1603 /**
1604  * read_cache_page - read into page cache, fill it if needed
1605  * @mapping:    the page's address_space
1606  * @index:      the page index
1607  * @filler:     function to perform the read
1608  * @data:       destination for read data
1609  *
1610  * Read into the page cache. If a page already exists, and PageUptodate() is
1611  * not set, try to fill the page then wait for it to become unlocked.
1612  *
1613  * If the page does not get brought uptodate, return -EIO.
1614  */
1615 struct page *read_cache_page(struct address_space *mapping,
1616                                 pgoff_t index,
1617                                 int (*filler)(void *,struct page*),
1618                                 void *data)
1619 {
1620         struct page *page;
1621
1622         page = read_cache_page_async(mapping, index, filler, data);
1623         if (IS_ERR(page))
1624                 goto out;
1625         wait_on_page_locked(page);
1626         if (!PageUptodate(page)) {
1627                 page_cache_release(page);
1628                 page = ERR_PTR(-EIO);
1629         }
1630  out:
1631         return page;
1632 }
1633 EXPORT_SYMBOL(read_cache_page);
1634
1635 /*
1636  * The logic we want is
1637  *
1638  *      if suid or (sgid and xgrp)
1639  *              remove privs
1640  */
1641 int should_remove_suid(struct dentry *dentry)
1642 {
1643         mode_t mode = dentry->d_inode->i_mode;
1644         int kill = 0;
1645
1646         /* suid always must be killed */
1647         if (unlikely(mode & S_ISUID))
1648                 kill = ATTR_KILL_SUID;
1649
1650         /*
1651          * sgid without any exec bits is just a mandatory locking mark; leave
1652          * it alone.  If some exec bits are set, it's a real sgid; kill it.
1653          */
1654         if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
1655                 kill |= ATTR_KILL_SGID;
1656
1657         if (unlikely(kill && !capable(CAP_FSETID)))
1658                 return kill;
1659
1660         return 0;
1661 }
1662 EXPORT_SYMBOL(should_remove_suid);
1663
1664 static int __remove_suid(struct dentry *dentry, int kill)
1665 {
1666         struct iattr newattrs;
1667
1668         newattrs.ia_valid = ATTR_FORCE | kill;
1669         return notify_change(dentry, &newattrs);
1670 }
1671
1672 int remove_suid(struct dentry *dentry)
1673 {
1674         int killsuid = should_remove_suid(dentry);
1675         int killpriv = security_inode_need_killpriv(dentry);
1676         int error = 0;
1677
1678         if (killpriv < 0)
1679                 return killpriv;
1680         if (killpriv)
1681                 error = security_inode_killpriv(dentry);
1682         if (!error && killsuid)
1683                 error = __remove_suid(dentry, killsuid);
1684
1685         return error;
1686 }
1687 EXPORT_SYMBOL(remove_suid);
1688
1689 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1690                         const struct iovec *iov, size_t base, size_t bytes)
1691 {
1692         size_t copied = 0, left = 0;
1693
1694         while (bytes) {
1695                 char __user *buf = iov->iov_base + base;
1696                 int copy = min(bytes, iov->iov_len - base);
1697
1698                 base = 0;
1699                 left = __copy_from_user_inatomic_nocache(vaddr, buf, copy);
1700                 copied += copy;
1701                 bytes -= copy;
1702                 vaddr += copy;
1703                 iov++;
1704
1705                 if (unlikely(left))
1706                         break;
1707         }
1708         return copied - left;
1709 }
1710
1711 /*
1712  * Copy as much as we can into the page and return the number of bytes which
1713  * were sucessfully copied.  If a fault is encountered then return the number of
1714  * bytes which were copied.
1715  */
1716 size_t iov_iter_copy_from_user_atomic(struct page *page,
1717                 struct iov_iter *i, unsigned long offset, size_t bytes)
1718 {
1719         char *kaddr;
1720         size_t copied;
1721
1722         BUG_ON(!in_atomic());
1723         kaddr = kmap_atomic(page, KM_USER0);
1724         if (likely(i->nr_segs == 1)) {
1725                 int left;
1726                 char __user *buf = i->iov->iov_base + i->iov_offset;
1727                 left = __copy_from_user_inatomic_nocache(kaddr + offset,
1728                                                         buf, bytes);
1729                 copied = bytes - left;
1730         } else {
1731                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1732                                                 i->iov, i->iov_offset, bytes);
1733         }
1734         kunmap_atomic(kaddr, KM_USER0);
1735
1736         return copied;
1737 }
1738 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1739
1740 /*
1741  * This has the same sideeffects and return value as
1742  * iov_iter_copy_from_user_atomic().
1743  * The difference is that it attempts to resolve faults.
1744  * Page must not be locked.
1745  */
1746 size_t iov_iter_copy_from_user(struct page *page,
1747                 struct iov_iter *i, unsigned long offset, size_t bytes)
1748 {
1749         char *kaddr;
1750         size_t copied;
1751
1752         kaddr = kmap(page);
1753         if (likely(i->nr_segs == 1)) {
1754                 int left;
1755                 char __user *buf = i->iov->iov_base + i->iov_offset;
1756                 left = __copy_from_user_nocache(kaddr + offset, buf, bytes);
1757                 copied = bytes - left;
1758         } else {
1759                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1760                                                 i->iov, i->iov_offset, bytes);
1761         }
1762         kunmap(page);
1763         return copied;
1764 }
1765 EXPORT_SYMBOL(iov_iter_copy_from_user);
1766
1767 void iov_iter_advance(struct iov_iter *i, size_t bytes)
1768 {
1769         BUG_ON(i->count < bytes);
1770
1771         if (likely(i->nr_segs == 1)) {
1772                 i->iov_offset += bytes;
1773                 i->count -= bytes;
1774         } else {
1775                 const struct iovec *iov = i->iov;
1776                 size_t base = i->iov_offset;
1777
1778                 /*
1779                  * The !iov->iov_len check ensures we skip over unlikely
1780                  * zero-length segments (without overruning the iovec).
1781                  */
1782                 while (bytes || unlikely(!iov->iov_len && i->count)) {
1783                         int copy;
1784
1785                         copy = min(bytes, iov->iov_len - base);
1786                         BUG_ON(!i->count || i->count < copy);
1787                         i->count -= copy;
1788                         bytes -= copy;
1789                         base += copy;
1790                         if (iov->iov_len == base) {
1791                                 iov++;
1792                                 base = 0;
1793                         }
1794                 }
1795                 i->iov = iov;
1796                 i->iov_offset = base;
1797         }
1798 }
1799 EXPORT_SYMBOL(iov_iter_advance);
1800
1801 /*
1802  * Fault in the first iovec of the given iov_iter, to a maximum length
1803  * of bytes. Returns 0 on success, or non-zero if the memory could not be
1804  * accessed (ie. because it is an invalid address).
1805  *
1806  * writev-intensive code may want this to prefault several iovecs -- that
1807  * would be possible (callers must not rely on the fact that _only_ the
1808  * first iovec will be faulted with the current implementation).
1809  */
1810 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
1811 {
1812         char __user *buf = i->iov->iov_base + i->iov_offset;
1813         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
1814         return fault_in_pages_readable(buf, bytes);
1815 }
1816 EXPORT_SYMBOL(iov_iter_fault_in_readable);
1817
1818 /*
1819  * Return the count of just the current iov_iter segment.
1820  */
1821 size_t iov_iter_single_seg_count(struct iov_iter *i)
1822 {
1823         const struct iovec *iov = i->iov;
1824         if (i->nr_segs == 1)
1825                 return i->count;
1826         else
1827                 return min(i->count, iov->iov_len - i->iov_offset);
1828 }
1829 EXPORT_SYMBOL(iov_iter_single_seg_count);
1830
1831 /*
1832  * Performs necessary checks before doing a write
1833  *
1834  * Can adjust writing position or amount of bytes to write.
1835  * Returns appropriate error code that caller should return or
1836  * zero in case that write should be allowed.
1837  */
1838 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
1839 {
1840         struct inode *inode = file->f_mapping->host;
1841         unsigned long limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1842
1843         if (unlikely(*pos < 0))
1844                 return -EINVAL;
1845
1846         if (!isblk) {
1847                 /* FIXME: this is for backwards compatibility with 2.4 */
1848                 if (file->f_flags & O_APPEND)
1849                         *pos = i_size_read(inode);
1850
1851                 if (limit != RLIM_INFINITY) {
1852                         if (*pos >= limit) {
1853                                 send_sig(SIGXFSZ, current, 0);
1854                                 return -EFBIG;
1855                         }
1856                         if (*count > limit - (typeof(limit))*pos) {
1857                                 *count = limit - (typeof(limit))*pos;
1858                         }
1859                 }
1860         }
1861
1862         /*
1863          * LFS rule
1864          */
1865         if (unlikely(*pos + *count > MAX_NON_LFS &&
1866                                 !(file->f_flags & O_LARGEFILE))) {
1867                 if (*pos >= MAX_NON_LFS) {
1868                         return -EFBIG;
1869                 }
1870                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
1871                         *count = MAX_NON_LFS - (unsigned long)*pos;
1872                 }
1873         }
1874
1875         /*
1876          * Are we about to exceed the fs block limit ?
1877          *
1878          * If we have written data it becomes a short write.  If we have
1879          * exceeded without writing data we send a signal and return EFBIG.
1880          * Linus frestrict idea will clean these up nicely..
1881          */
1882         if (likely(!isblk)) {
1883                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
1884                         if (*count || *pos > inode->i_sb->s_maxbytes) {
1885                                 return -EFBIG;
1886                         }
1887                         /* zero-length writes at ->s_maxbytes are OK */
1888                 }
1889
1890                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
1891                         *count = inode->i_sb->s_maxbytes - *pos;
1892         } else {
1893 #ifdef CONFIG_BLOCK
1894                 loff_t isize;
1895                 if (bdev_read_only(I_BDEV(inode)))
1896                         return -EPERM;
1897                 isize = i_size_read(inode);
1898                 if (*pos >= isize) {
1899                         if (*count || *pos > isize)
1900                                 return -ENOSPC;
1901                 }
1902
1903                 if (*pos + *count > isize)
1904                         *count = isize - *pos;
1905 #else
1906                 return -EPERM;
1907 #endif
1908         }
1909         return 0;
1910 }
1911 EXPORT_SYMBOL(generic_write_checks);
1912
1913 int pagecache_write_begin(struct file *file, struct address_space *mapping,
1914                                 loff_t pos, unsigned len, unsigned flags,
1915                                 struct page **pagep, void **fsdata)
1916 {
1917         const struct address_space_operations *aops = mapping->a_ops;
1918
1919         if (aops->write_begin) {
1920                 return aops->write_begin(file, mapping, pos, len, flags,
1921                                                         pagep, fsdata);
1922         } else {
1923                 int ret;
1924                 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1925                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1926                 struct inode *inode = mapping->host;
1927                 struct page *page;
1928 again:
1929                 page = __grab_cache_page(mapping, index);
1930                 *pagep = page;
1931                 if (!page)
1932                         return -ENOMEM;
1933
1934                 if (flags & AOP_FLAG_UNINTERRUPTIBLE && !PageUptodate(page)) {
1935                         /*
1936                          * There is no way to resolve a short write situation
1937                          * for a !Uptodate page (except by double copying in
1938                          * the caller done by generic_perform_write_2copy).
1939                          *
1940                          * Instead, we have to bring it uptodate here.
1941                          */
1942                         ret = aops->readpage(file, page);
1943                         page_cache_release(page);
1944                         if (ret) {
1945                                 if (ret == AOP_TRUNCATED_PAGE)
1946                                         goto again;
1947                                 return ret;
1948                         }
1949                         goto again;
1950                 }
1951
1952                 ret = aops->prepare_write(file, page, offset, offset+len);
1953                 if (ret) {
1954                         unlock_page(page);
1955                         page_cache_release(page);
1956                         if (pos + len > inode->i_size)
1957                                 vmtruncate(inode, inode->i_size);
1958                 }
1959                 return ret;
1960         }
1961 }
1962 EXPORT_SYMBOL(pagecache_write_begin);
1963
1964 int pagecache_write_end(struct file *file, struct address_space *mapping,
1965                                 loff_t pos, unsigned len, unsigned copied,
1966                                 struct page *page, void *fsdata)
1967 {
1968         const struct address_space_operations *aops = mapping->a_ops;
1969         int ret;
1970
1971         if (aops->write_end) {
1972                 mark_page_accessed(page);
1973                 ret = aops->write_end(file, mapping, pos, len, copied,
1974                                                         page, fsdata);
1975         } else {
1976                 unsigned offset = pos & (PAGE_CACHE_SIZE - 1);
1977                 struct inode *inode = mapping->host;
1978
1979                 flush_dcache_page(page);
1980                 ret = aops->commit_write(file, page, offset, offset+len);
1981                 unlock_page(page);
1982                 mark_page_accessed(page);
1983                 page_cache_release(page);
1984
1985                 if (ret < 0) {
1986                         if (pos + len > inode->i_size)
1987                                 vmtruncate(inode, inode->i_size);
1988                 } else if (ret > 0)
1989                         ret = min_t(size_t, copied, ret);
1990                 else
1991                         ret = copied;
1992         }
1993
1994         return ret;
1995 }
1996 EXPORT_SYMBOL(pagecache_write_end);
1997
1998 ssize_t
1999 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2000                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2001                 size_t count, size_t ocount)
2002 {
2003         struct file     *file = iocb->ki_filp;
2004         struct address_space *mapping = file->f_mapping;
2005         struct inode    *inode = mapping->host;
2006         ssize_t         written;
2007
2008         if (count != ocount)
2009                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2010
2011         written = generic_file_direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2012         if (written > 0) {
2013                 loff_t end = pos + written;
2014                 if (end > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2015                         i_size_write(inode,  end);
2016                         mark_inode_dirty(inode);
2017                 }
2018                 *ppos = end;
2019         }
2020
2021         /*
2022          * Sync the fs metadata but not the minor inode changes and
2023          * of course not the data as we did direct DMA for the IO.
2024          * i_mutex is held, which protects generic_osync_inode() from
2025          * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
2026          */
2027         if ((written >= 0 || written == -EIOCBQUEUED) &&
2028             ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2029                 int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
2030                 if (err < 0)
2031                         written = err;
2032         }
2033         return written;
2034 }
2035 EXPORT_SYMBOL(generic_file_direct_write);
2036
2037 /*
2038  * Find or create a page at the given pagecache position. Return the locked
2039  * page. This function is specifically for buffered writes.
2040  */
2041 struct page *__grab_cache_page(struct address_space *mapping, pgoff_t index)
2042 {
2043         int status;
2044         struct page *page;
2045 repeat:
2046         page = find_lock_page(mapping, index);
2047         if (likely(page))
2048                 return page;
2049
2050         page = page_cache_alloc(mapping);
2051         if (!page)
2052                 return NULL;
2053         status = add_to_page_cache_lru(page, mapping, index, GFP_KERNEL);
2054         if (unlikely(status)) {
2055                 page_cache_release(page);
2056                 if (status == -EEXIST)
2057                         goto repeat;
2058                 return NULL;
2059         }
2060         return page;
2061 }
2062 EXPORT_SYMBOL(__grab_cache_page);
2063
2064 static ssize_t generic_perform_write_2copy(struct file *file,
2065                                 struct iov_iter *i, loff_t pos)
2066 {
2067         struct address_space *mapping = file->f_mapping;
2068         const struct address_space_operations *a_ops = mapping->a_ops;
2069         struct inode *inode = mapping->host;
2070         long status = 0;
2071         ssize_t written = 0;
2072
2073         do {
2074                 struct page *src_page;
2075                 struct page *page;
2076                 pgoff_t index;          /* Pagecache index for current page */
2077                 unsigned long offset;   /* Offset into pagecache page */
2078                 unsigned long bytes;    /* Bytes to write to page */
2079                 size_t copied;          /* Bytes copied from user */
2080
2081                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2082                 index = pos >> PAGE_CACHE_SHIFT;
2083                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2084                                                 iov_iter_count(i));
2085
2086                 /*
2087                  * a non-NULL src_page indicates that we're doing the
2088                  * copy via get_user_pages and kmap.
2089                  */
2090                 src_page = NULL;
2091
2092                 /*
2093                  * Bring in the user page that we will copy from _first_.
2094                  * Otherwise there's a nasty deadlock on copying from the
2095                  * same page as we're writing to, without it being marked
2096                  * up-to-date.
2097                  *
2098                  * Not only is this an optimisation, but it is also required
2099                  * to check that the address is actually valid, when atomic
2100                  * usercopies are used, below.
2101                  */
2102                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2103                         status = -EFAULT;
2104                         break;
2105                 }
2106
2107                 page = __grab_cache_page(mapping, index);
2108                 if (!page) {
2109                         status = -ENOMEM;
2110                         break;
2111                 }
2112
2113                 /*
2114                  * non-uptodate pages cannot cope with short copies, and we
2115                  * cannot take a pagefault with the destination page locked.
2116                  * So pin the source page to copy it.
2117                  */
2118                 if (!PageUptodate(page) && !segment_eq(get_fs(), KERNEL_DS)) {
2119                         unlock_page(page);
2120
2121                         src_page = alloc_page(GFP_KERNEL);
2122                         if (!src_page) {
2123                                 page_cache_release(page);
2124                                 status = -ENOMEM;
2125                                 break;
2126                         }
2127
2128                         /*
2129                          * Cannot get_user_pages with a page locked for the
2130                          * same reason as we can't take a page fault with a
2131                          * page locked (as explained below).
2132                          */
2133                         copied = iov_iter_copy_from_user(src_page, i,
2134                                                                 offset, bytes);
2135                         if (unlikely(copied == 0)) {
2136                                 status = -EFAULT;
2137                                 page_cache_release(page);
2138                                 page_cache_release(src_page);
2139                                 break;
2140                         }
2141                         bytes = copied;
2142
2143                         lock_page(page);
2144                         /*
2145                          * Can't handle the page going uptodate here, because
2146                          * that means we would use non-atomic usercopies, which
2147                          * zero out the tail of the page, which can cause
2148                          * zeroes to become transiently visible. We could just
2149                          * use a non-zeroing copy, but the APIs aren't too
2150                          * consistent.
2151                          */
2152                         if (unlikely(!page->mapping || PageUptodate(page))) {
2153                                 unlock_page(page);
2154                                 page_cache_release(page);
2155                                 page_cache_release(src_page);
2156                                 continue;
2157                         }
2158                 }
2159
2160                 status = a_ops->prepare_write(file, page, offset, offset+bytes);
2161                 if (unlikely(status))
2162                         goto fs_write_aop_error;
2163
2164                 if (!src_page) {
2165                         /*
2166                          * Must not enter the pagefault handler here, because
2167                          * we hold the page lock, so we might recursively
2168                          * deadlock on the same lock, or get an ABBA deadlock
2169                          * against a different lock, or against the mmap_sem
2170                          * (which nests outside the page lock).  So increment
2171                          * preempt count, and use _atomic usercopies.
2172                          *
2173                          * The page is uptodate so we are OK to encounter a
2174                          * short copy: if unmodified parts of the page are
2175                          * marked dirty and written out to disk, it doesn't
2176                          * really matter.
2177                          */
2178                         pagefault_disable();
2179                         copied = iov_iter_copy_from_user_atomic(page, i,
2180                                                                 offset, bytes);
2181                         pagefault_enable();
2182                 } else {
2183                         void *src, *dst;
2184                         src = kmap_atomic(src_page, KM_USER0);
2185                         dst = kmap_atomic(page, KM_USER1);
2186                         memcpy(dst + offset, src + offset, bytes);
2187                         kunmap_atomic(dst, KM_USER1);
2188                         kunmap_atomic(src, KM_USER0);
2189                         copied = bytes;
2190                 }
2191                 flush_dcache_page(page);
2192
2193                 status = a_ops->commit_write(file, page, offset, offset+bytes);
2194                 if (unlikely(status < 0))
2195                         goto fs_write_aop_error;
2196                 if (unlikely(status > 0)) /* filesystem did partial write */
2197                         copied = min_t(size_t, copied, status);
2198
2199                 unlock_page(page);
2200                 mark_page_accessed(page);
2201                 page_cache_release(page);
2202                 if (src_page)
2203                         page_cache_release(src_page);
2204
2205                 iov_iter_advance(i, copied);
2206                 pos += copied;
2207                 written += copied;
2208
2209                 balance_dirty_pages_ratelimited(mapping);
2210                 cond_resched();
2211                 continue;
2212
2213 fs_write_aop_error:
2214                 unlock_page(page);
2215                 page_cache_release(page);
2216                 if (src_page)
2217                         page_cache_release(src_page);
2218
2219                 /*
2220                  * prepare_write() may have instantiated a few blocks
2221                  * outside i_size.  Trim these off again. Don't need
2222                  * i_size_read because we hold i_mutex.
2223                  */
2224                 if (pos + bytes > inode->i_size)
2225                         vmtruncate(inode, inode->i_size);
2226                 break;
2227         } while (iov_iter_count(i));
2228
2229         return written ? written : status;
2230 }
2231
2232 static ssize_t generic_perform_write(struct file *file,
2233                                 struct iov_iter *i, loff_t pos)
2234 {
2235         struct address_space *mapping = file->f_mapping;
2236         const struct address_space_operations *a_ops = mapping->a_ops;
2237         long status = 0;
2238         ssize_t written = 0;
2239         unsigned int flags = 0;
2240
2241         /*
2242          * Copies from kernel address space cannot fail (NFSD is a big user).
2243          */
2244         if (segment_eq(get_fs(), KERNEL_DS))
2245                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2246
2247         do {
2248                 struct page *page;
2249                 pgoff_t index;          /* Pagecache index for current page */
2250                 unsigned long offset;   /* Offset into pagecache page */
2251                 unsigned long bytes;    /* Bytes to write to page */
2252                 size_t copied;          /* Bytes copied from user */
2253                 void *fsdata;
2254
2255                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2256                 index = pos >> PAGE_CACHE_SHIFT;
2257                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2258                                                 iov_iter_count(i));
2259
2260 again:
2261
2262                 /*
2263                  * Bring in the user page that we will copy from _first_.
2264                  * Otherwise there's a nasty deadlock on copying from the
2265                  * same page as we're writing to, without it being marked
2266                  * up-to-date.
2267                  *
2268                  * Not only is this an optimisation, but it is also required
2269                  * to check that the address is actually valid, when atomic
2270                  * usercopies are used, below.
2271                  */
2272                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2273                         status = -EFAULT;
2274                         break;
2275                 }
2276
2277                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2278                                                 &page, &fsdata);
2279                 if (unlikely(status))
2280                         break;
2281
2282                 pagefault_disable();
2283                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2284                 pagefault_enable();
2285                 flush_dcache_page(page);
2286
2287                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2288                                                 page, fsdata);
2289                 if (unlikely(status < 0))
2290                         break;
2291                 copied = status;
2292
2293                 cond_resched();
2294
2295                 iov_iter_advance(i, copied);
2296                 if (unlikely(copied == 0)) {
2297                         /*
2298                          * If we were unable to copy any data at all, we must
2299                          * fall back to a single segment length write.
2300                          *
2301                          * If we didn't fallback here, we could livelock
2302                          * because not all segments in the iov can be copied at
2303                          * once without a pagefault.
2304                          */
2305                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2306                                                 iov_iter_single_seg_count(i));
2307                         goto again;
2308                 }
2309                 pos += copied;
2310                 written += copied;
2311
2312                 balance_dirty_pages_ratelimited(mapping);
2313
2314         } while (iov_iter_count(i));
2315
2316         return written ? written : status;
2317 }
2318
2319 ssize_t
2320 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2321                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2322                 size_t count, ssize_t written)
2323 {
2324         struct file *file = iocb->ki_filp;
2325         struct address_space *mapping = file->f_mapping;
2326         const struct address_space_operations *a_ops = mapping->a_ops;
2327         struct inode *inode = mapping->host;
2328         ssize_t status;
2329         struct iov_iter i;
2330
2331         iov_iter_init(&i, iov, nr_segs, count, written);
2332         if (a_ops->write_begin)
2333                 status = generic_perform_write(file, &i, pos);
2334         else
2335                 status = generic_perform_write_2copy(file, &i, pos);
2336
2337         if (likely(status >= 0)) {
2338                 written += status;
2339                 *ppos = pos + status;
2340
2341                 /*
2342                  * For now, when the user asks for O_SYNC, we'll actually give
2343                  * O_DSYNC
2344                  */
2345                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2346                         if (!a_ops->writepage || !is_sync_kiocb(iocb))
2347                                 status = generic_osync_inode(inode, mapping,
2348                                                 OSYNC_METADATA|OSYNC_DATA);
2349                 }
2350         }
2351         
2352         /*
2353          * If we get here for O_DIRECT writes then we must have fallen through
2354          * to buffered writes (block instantiation inside i_size).  So we sync
2355          * the file data here, to try to honour O_DIRECT expectations.
2356          */
2357         if (unlikely(file->f_flags & O_DIRECT) && written)
2358                 status = filemap_write_and_wait(mapping);
2359
2360         return written ? written : status;
2361 }
2362 EXPORT_SYMBOL(generic_file_buffered_write);
2363
2364 static ssize_t
2365 __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
2366                                 unsigned long nr_segs, loff_t *ppos)
2367 {
2368         struct file *file = iocb->ki_filp;
2369         struct address_space * mapping = file->f_mapping;
2370         size_t ocount;          /* original count */
2371         size_t count;           /* after file limit checks */
2372         struct inode    *inode = mapping->host;
2373         loff_t          pos;
2374         ssize_t         written;
2375         ssize_t         err;
2376
2377         ocount = 0;
2378         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2379         if (err)
2380                 return err;
2381
2382         count = ocount;
2383         pos = *ppos;
2384
2385         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2386
2387         /* We can write back this queue in page reclaim */
2388         current->backing_dev_info = mapping->backing_dev_info;
2389         written = 0;
2390
2391         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2392         if (err)
2393                 goto out;
2394
2395         if (count == 0)
2396                 goto out;
2397
2398         err = remove_suid(file->f_path.dentry);
2399         if (err)
2400                 goto out;
2401
2402         file_update_time(file);
2403
2404         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2405         if (unlikely(file->f_flags & O_DIRECT)) {
2406                 loff_t endbyte;
2407                 ssize_t written_buffered;
2408
2409                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2410                                                         ppos, count, ocount);
2411                 if (written < 0 || written == count)
2412                         goto out;
2413                 /*
2414                  * direct-io write to a hole: fall through to buffered I/O
2415                  * for completing the rest of the request.
2416                  */
2417                 pos += written;
2418                 count -= written;
2419                 written_buffered = generic_file_buffered_write(iocb, iov,
2420                                                 nr_segs, pos, ppos, count,
2421                                                 written);
2422                 /*
2423                  * If generic_file_buffered_write() retuned a synchronous error
2424                  * then we want to return the number of bytes which were
2425                  * direct-written, or the error code if that was zero.  Note
2426                  * that this differs from normal direct-io semantics, which
2427                  * will return -EFOO even if some bytes were written.
2428                  */
2429                 if (written_buffered < 0) {
2430                         err = written_buffered;
2431                         goto out;
2432                 }
2433
2434                 /*
2435                  * We need to ensure that the page cache pages are written to
2436                  * disk and invalidated to preserve the expected O_DIRECT
2437                  * semantics.
2438                  */
2439                 endbyte = pos + written_buffered - written - 1;
2440                 err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
2441                                             SYNC_FILE_RANGE_WAIT_BEFORE|
2442                                             SYNC_FILE_RANGE_WRITE|
2443                                             SYNC_FILE_RANGE_WAIT_AFTER);
2444                 if (err == 0) {
2445                         written = written_buffered;
2446                         invalidate_mapping_pages(mapping,
2447                                                  pos >> PAGE_CACHE_SHIFT,
2448                                                  endbyte >> PAGE_CACHE_SHIFT);
2449                 } else {
2450                         /*
2451                          * We don't know how much we wrote, so just return
2452                          * the number of bytes which were direct-written
2453                          */
2454                 }
2455         } else {
2456                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2457                                 pos, ppos, count, written);
2458         }
2459 out:
2460         current->backing_dev_info = NULL;
2461         return written ? written : err;
2462 }
2463
2464 ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
2465                 const struct iovec *iov, unsigned long nr_segs, loff_t pos)
2466 {
2467         struct file *file = iocb->ki_filp;
2468         struct address_space *mapping = file->f_mapping;
2469         struct inode *inode = mapping->host;
2470         ssize_t ret;
2471
2472         BUG_ON(iocb->ki_pos != pos);
2473
2474         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2475                         &iocb->ki_pos);
2476
2477         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2478                 ssize_t err;
2479
2480                 err = sync_page_range_nolock(inode, mapping, pos, ret);
2481                 if (err < 0)
2482                         ret = err;
2483         }
2484         return ret;
2485 }
2486 EXPORT_SYMBOL(generic_file_aio_write_nolock);
2487
2488 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2489                 unsigned long nr_segs, loff_t pos)
2490 {
2491         struct file *file = iocb->ki_filp;
2492         struct address_space *mapping = file->f_mapping;
2493         struct inode *inode = mapping->host;
2494         ssize_t ret;
2495
2496         BUG_ON(iocb->ki_pos != pos);
2497
2498         mutex_lock(&inode->i_mutex);
2499         ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
2500                         &iocb->ki_pos);
2501         mutex_unlock(&inode->i_mutex);
2502
2503         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2504                 ssize_t err;
2505
2506                 err = sync_page_range(inode, mapping, pos, ret);
2507                 if (err < 0)
2508                         ret = err;
2509         }
2510         return ret;
2511 }
2512 EXPORT_SYMBOL(generic_file_aio_write);
2513
2514 /*
2515  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
2516  * went wrong during pagecache shootdown.
2517  */
2518 static ssize_t
2519 generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
2520         loff_t offset, unsigned long nr_segs)
2521 {
2522         struct file *file = iocb->ki_filp;
2523         struct address_space *mapping = file->f_mapping;
2524         ssize_t retval;
2525         size_t write_len;
2526         pgoff_t end = 0; /* silence gcc */
2527
2528         /*
2529          * If it's a write, unmap all mmappings of the file up-front.  This
2530          * will cause any pte dirty bits to be propagated into the pageframes
2531          * for the subsequent filemap_write_and_wait().
2532          */
2533         if (rw == WRITE) {
2534                 write_len = iov_length(iov, nr_segs);
2535                 end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
2536                 if (mapping_mapped(mapping))
2537                         unmap_mapping_range(mapping, offset, write_len, 0);
2538         }
2539
2540         retval = filemap_write_and_wait(mapping);
2541         if (retval)
2542                 goto out;
2543
2544         /*
2545          * After a write we want buffered reads to be sure to go to disk to get
2546          * the new data.  We invalidate clean cached page from the region we're
2547          * about to write.  We do this *before* the write so that we can return
2548          * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
2549          */
2550         if (rw == WRITE && mapping->nrpages) {
2551                 retval = invalidate_inode_pages2_range(mapping,
2552                                         offset >> PAGE_CACHE_SHIFT, end);
2553                 if (retval)
2554                         goto out;
2555         }
2556
2557         retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
2558
2559         /*
2560          * Finally, try again to invalidate clean pages which might have been
2561          * cached by non-direct readahead, or faulted in by get_user_pages()
2562          * if the source of the write was an mmap'ed region of the file
2563          * we're writing.  Either one is a pretty crazy thing to do,
2564          * so we don't support it 100%.  If this invalidation
2565          * fails, tough, the write still worked...
2566          */
2567         if (rw == WRITE && mapping->nrpages) {
2568                 invalidate_inode_pages2_range(mapping, offset >> PAGE_CACHE_SHIFT, end);
2569         }
2570 out:
2571         return retval;
2572 }
2573
2574 /**
2575  * try_to_release_page() - release old fs-specific metadata on a page
2576  *
2577  * @page: the page which the kernel is trying to free
2578  * @gfp_mask: memory allocation flags (and I/O mode)
2579  *
2580  * The address_space is to try to release any data against the page
2581  * (presumably at page->private).  If the release was successful, return `1'.
2582  * Otherwise return zero.
2583  *
2584  * The @gfp_mask argument specifies whether I/O may be performed to release
2585  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
2586  *
2587  * NOTE: @gfp_mask may go away, and this function may become non-blocking.
2588  */
2589 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2590 {
2591         struct address_space * const mapping = page->mapping;
2592
2593         BUG_ON(!PageLocked(page));
2594         if (PageWriteback(page))
2595                 return 0;
2596
2597         if (mapping && mapping->a_ops->releasepage)
2598                 return mapping->a_ops->releasepage(page, gfp_mask);
2599         return try_to_free_buffers(page);
2600 }
2601
2602 EXPORT_SYMBOL(try_to_release_page);