Linux 3.10.61
[linux-3.10.git] / mm / filemap.c
1 /*
2  *      linux/mm/filemap.c
3  *
4  * Copyright (C) 1994-1999  Linus Torvalds
5  */
6
7 /*
8  * This file handles the generic file mmap semantics used by
9  * most "normal" filesystems (but you don't /have/ to use this:
10  * the NFS filesystem used to do this differently, for example)
11  */
12 #include <linux/export.h>
13 #include <linux/compiler.h>
14 #include <linux/fs.h>
15 #include <linux/uaccess.h>
16 #include <linux/aio.h>
17 #include <linux/capability.h>
18 #include <linux/kernel_stat.h>
19 #include <linux/gfp.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/cpuset.h>
33 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
34 #include <linux/memcontrol.h>
35 #include <linux/cleancache.h>
36 #include "internal.h"
37
38 #define CREATE_TRACE_POINTS
39 #include <trace/events/filemap.h>
40
41 /*
42  * FIXME: remove all knowledge of the buffer layer from the core VM
43  */
44 #include <linux/buffer_head.h> /* for try_to_free_buffers */
45
46 #include <asm/mman.h>
47
48 /*
49  * Shared mappings implemented 30.11.1994. It's not fully working yet,
50  * though.
51  *
52  * Shared mappings now work. 15.8.1995  Bruno.
53  *
54  * finished 'unifying' the page and buffer cache and SMP-threaded the
55  * page-cache, 21.05.1999, Ingo Molnar <mingo@redhat.com>
56  *
57  * SMP-threaded pagemap-LRU 1999, Andrea Arcangeli <andrea@suse.de>
58  */
59
60 /*
61  * Lock ordering:
62  *
63  *  ->i_mmap_mutex              (truncate_pagecache)
64  *    ->private_lock            (__free_pte->__set_page_dirty_buffers)
65  *      ->swap_lock             (exclusive_swap_page, others)
66  *        ->mapping->tree_lock
67  *
68  *  ->i_mutex
69  *    ->i_mmap_mutex            (truncate->unmap_mapping_range)
70  *
71  *  ->mmap_sem
72  *    ->i_mmap_mutex
73  *      ->page_table_lock or pte_lock   (various, mainly in memory.c)
74  *        ->mapping->tree_lock  (arch-dependent flush_dcache_mmap_lock)
75  *
76  *  ->mmap_sem
77  *    ->lock_page               (access_process_vm)
78  *
79  *  ->i_mutex                   (generic_file_buffered_write)
80  *    ->mmap_sem                (fault_in_pages_readable->do_page_fault)
81  *
82  *  bdi->wb.list_lock
83  *    sb_lock                   (fs/fs-writeback.c)
84  *    ->mapping->tree_lock      (__sync_single_inode)
85  *
86  *  ->i_mmap_mutex
87  *    ->anon_vma.lock           (vma_adjust)
88  *
89  *  ->anon_vma.lock
90  *    ->page_table_lock or pte_lock     (anon_vma_prepare and various)
91  *
92  *  ->page_table_lock or pte_lock
93  *    ->swap_lock               (try_to_unmap_one)
94  *    ->private_lock            (try_to_unmap_one)
95  *    ->tree_lock               (try_to_unmap_one)
96  *    ->zone.lru_lock           (follow_page->mark_page_accessed)
97  *    ->zone.lru_lock           (check_pte_range->isolate_lru_page)
98  *    ->private_lock            (page_remove_rmap->set_page_dirty)
99  *    ->tree_lock               (page_remove_rmap->set_page_dirty)
100  *    bdi.wb->list_lock         (page_remove_rmap->set_page_dirty)
101  *    ->inode->i_lock           (page_remove_rmap->set_page_dirty)
102  *    bdi.wb->list_lock         (zap_pte_range->set_page_dirty)
103  *    ->inode->i_lock           (zap_pte_range->set_page_dirty)
104  *    ->private_lock            (zap_pte_range->__set_page_dirty_buffers)
105  *
106  * ->i_mmap_mutex
107  *   ->tasklist_lock            (memory_failure, collect_procs_ao)
108  */
109
110 /*
111  * Delete a page from the page cache and free it. Caller has to make
112  * sure the page is locked and that nobody else uses it - or that usage
113  * is safe.  The caller must hold the mapping's tree_lock.
114  */
115 void __delete_from_page_cache(struct page *page)
116 {
117         struct address_space *mapping = page->mapping;
118
119         trace_mm_filemap_delete_from_page_cache(page);
120         /*
121          * if we're uptodate, flush out into the cleancache, otherwise
122          * invalidate any existing cleancache entries.  We can't leave
123          * stale data around in the cleancache once our page is gone
124          */
125         if (PageUptodate(page) && PageMappedToDisk(page))
126                 cleancache_put_page(page);
127         else
128                 cleancache_invalidate_page(mapping, page);
129
130         radix_tree_delete(&mapping->page_tree, page->index);
131         page->mapping = NULL;
132         /* Leave page->index set: truncation lookup relies upon it */
133         mapping->nrpages--;
134         __dec_zone_page_state(page, NR_FILE_PAGES);
135         if (PageSwapBacked(page))
136                 __dec_zone_page_state(page, NR_SHMEM);
137         BUG_ON(page_mapped(page));
138
139         /*
140          * Some filesystems seem to re-dirty the page even after
141          * the VM has canceled the dirty bit (eg ext3 journaling).
142          *
143          * Fix it up by doing a final dirty accounting check after
144          * having removed the page entirely.
145          */
146         if (PageDirty(page) && mapping_cap_account_dirty(mapping)) {
147                 dec_zone_page_state(page, NR_FILE_DIRTY);
148                 dec_bdi_stat(mapping->backing_dev_info, BDI_RECLAIMABLE);
149         }
150 }
151
152 /**
153  * delete_from_page_cache - delete page from page cache
154  * @page: the page which the kernel is trying to remove from page cache
155  *
156  * This must be called only on pages that have been verified to be in the page
157  * cache and locked.  It will never put the page into the free list, the caller
158  * has a reference on the page.
159  */
160 void delete_from_page_cache(struct page *page)
161 {
162         struct address_space *mapping = page->mapping;
163         void (*freepage)(struct page *);
164
165         BUG_ON(!PageLocked(page));
166
167         freepage = mapping->a_ops->freepage;
168         spin_lock_irq(&mapping->tree_lock);
169         __delete_from_page_cache(page);
170         spin_unlock_irq(&mapping->tree_lock);
171         mem_cgroup_uncharge_cache_page(page);
172
173         if (freepage)
174                 freepage(page);
175         page_cache_release(page);
176 }
177 EXPORT_SYMBOL(delete_from_page_cache);
178
179 static int sleep_on_page(void *word)
180 {
181         io_schedule();
182         return 0;
183 }
184
185 static int sleep_on_page_killable(void *word)
186 {
187         sleep_on_page(word);
188         return fatal_signal_pending(current) ? -EINTR : 0;
189 }
190
191 static int filemap_check_errors(struct address_space *mapping)
192 {
193         int ret = 0;
194         /* Check for outstanding write errors */
195         if (test_and_clear_bit(AS_ENOSPC, &mapping->flags))
196                 ret = -ENOSPC;
197         if (test_and_clear_bit(AS_EIO, &mapping->flags))
198                 ret = -EIO;
199         return ret;
200 }
201
202 /**
203  * __filemap_fdatawrite_range - start writeback on mapping dirty pages in range
204  * @mapping:    address space structure to write
205  * @start:      offset in bytes where the range starts
206  * @end:        offset in bytes where the range ends (inclusive)
207  * @sync_mode:  enable synchronous operation
208  *
209  * Start writeback against all of a mapping's dirty pages that lie
210  * within the byte offsets <start, end> inclusive.
211  *
212  * If sync_mode is WB_SYNC_ALL then this is a "data integrity" operation, as
213  * opposed to a regular memory cleansing writeback.  The difference between
214  * these two operations is that if a dirty page/buffer is encountered, it must
215  * be waited upon, and not just skipped over.
216  */
217 int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
218                                 loff_t end, int sync_mode)
219 {
220         int ret;
221         struct writeback_control wbc = {
222                 .sync_mode = sync_mode,
223                 .nr_to_write = LONG_MAX,
224                 .range_start = start,
225                 .range_end = end,
226         };
227
228         if (!mapping_cap_writeback_dirty(mapping))
229                 return 0;
230
231         ret = do_writepages(mapping, &wbc);
232         return ret;
233 }
234
235 static inline int __filemap_fdatawrite(struct address_space *mapping,
236         int sync_mode)
237 {
238         return __filemap_fdatawrite_range(mapping, 0, LLONG_MAX, sync_mode);
239 }
240
241 int filemap_fdatawrite(struct address_space *mapping)
242 {
243         return __filemap_fdatawrite(mapping, WB_SYNC_ALL);
244 }
245 EXPORT_SYMBOL(filemap_fdatawrite);
246
247 int filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
248                                 loff_t end)
249 {
250         return __filemap_fdatawrite_range(mapping, start, end, WB_SYNC_ALL);
251 }
252 EXPORT_SYMBOL(filemap_fdatawrite_range);
253
254 /**
255  * filemap_flush - mostly a non-blocking flush
256  * @mapping:    target address_space
257  *
258  * This is a mostly non-blocking flush.  Not suitable for data-integrity
259  * purposes - I/O may not be started against all dirty pages.
260  */
261 int filemap_flush(struct address_space *mapping)
262 {
263         return __filemap_fdatawrite(mapping, WB_SYNC_NONE);
264 }
265 EXPORT_SYMBOL(filemap_flush);
266
267 /**
268  * filemap_fdatawait_range - wait for writeback to complete
269  * @mapping:            address space structure to wait for
270  * @start_byte:         offset in bytes where the range starts
271  * @end_byte:           offset in bytes where the range ends (inclusive)
272  *
273  * Walk the list of under-writeback pages of the given address space
274  * in the given range and wait for all of them.
275  */
276 int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
277                             loff_t end_byte)
278 {
279         pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
280         pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
281         struct pagevec pvec;
282         int nr_pages;
283         int ret2, ret = 0;
284
285         if (end_byte < start_byte)
286                 goto out;
287
288         pagevec_init(&pvec, 0);
289         while ((index <= end) &&
290                         (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
291                         PAGECACHE_TAG_WRITEBACK,
292                         min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1)) != 0) {
293                 unsigned i;
294
295                 for (i = 0; i < nr_pages; i++) {
296                         struct page *page = pvec.pages[i];
297
298                         /* until radix tree lookup accepts end_index */
299                         if (page->index > end)
300                                 continue;
301
302                         wait_on_page_writeback(page);
303                         if (TestClearPageError(page))
304                                 ret = -EIO;
305                 }
306                 pagevec_release(&pvec);
307                 cond_resched();
308         }
309 out:
310         ret2 = filemap_check_errors(mapping);
311         if (!ret)
312                 ret = ret2;
313
314         return ret;
315 }
316 EXPORT_SYMBOL(filemap_fdatawait_range);
317
318 /**
319  * filemap_fdatawait - wait for all under-writeback pages to complete
320  * @mapping: address space structure to wait for
321  *
322  * Walk the list of under-writeback pages of the given address space
323  * and wait for all of them.
324  */
325 int filemap_fdatawait(struct address_space *mapping)
326 {
327         loff_t i_size = i_size_read(mapping->host);
328
329         if (i_size == 0)
330                 return 0;
331
332         return filemap_fdatawait_range(mapping, 0, i_size - 1);
333 }
334 EXPORT_SYMBOL(filemap_fdatawait);
335
336 int filemap_write_and_wait(struct address_space *mapping)
337 {
338         int err = 0;
339
340         if (mapping->nrpages) {
341                 err = filemap_fdatawrite(mapping);
342                 /*
343                  * Even if the above returned error, the pages may be
344                  * written partially (e.g. -ENOSPC), so we wait for it.
345                  * But the -EIO is special case, it may indicate the worst
346                  * thing (e.g. bug) happened, so we avoid waiting for it.
347                  */
348                 if (err != -EIO) {
349                         int err2 = filemap_fdatawait(mapping);
350                         if (!err)
351                                 err = err2;
352                 }
353         } else {
354                 err = filemap_check_errors(mapping);
355         }
356         return err;
357 }
358 EXPORT_SYMBOL(filemap_write_and_wait);
359
360 /**
361  * filemap_write_and_wait_range - write out & wait on a file range
362  * @mapping:    the address_space for the pages
363  * @lstart:     offset in bytes where the range starts
364  * @lend:       offset in bytes where the range ends (inclusive)
365  *
366  * Write out and wait upon file offsets lstart->lend, inclusive.
367  *
368  * Note that `lend' is inclusive (describes the last byte to be written) so
369  * that this function can be used to write to the very end-of-file (end = -1).
370  */
371 int filemap_write_and_wait_range(struct address_space *mapping,
372                                  loff_t lstart, loff_t lend)
373 {
374         int err = 0;
375
376         if (mapping->nrpages) {
377                 err = __filemap_fdatawrite_range(mapping, lstart, lend,
378                                                  WB_SYNC_ALL);
379                 /* See comment of filemap_write_and_wait() */
380                 if (err != -EIO) {
381                         int err2 = filemap_fdatawait_range(mapping,
382                                                 lstart, lend);
383                         if (!err)
384                                 err = err2;
385                 }
386         } else {
387                 err = filemap_check_errors(mapping);
388         }
389         return err;
390 }
391 EXPORT_SYMBOL(filemap_write_and_wait_range);
392
393 /**
394  * replace_page_cache_page - replace a pagecache page with a new one
395  * @old:        page to be replaced
396  * @new:        page to replace with
397  * @gfp_mask:   allocation mode
398  *
399  * This function replaces a page in the pagecache with a new one.  On
400  * success it acquires the pagecache reference for the new page and
401  * drops it for the old page.  Both the old and new pages must be
402  * locked.  This function does not add the new page to the LRU, the
403  * caller must do that.
404  *
405  * The remove + add is atomic.  The only way this function can fail is
406  * memory allocation failure.
407  */
408 int replace_page_cache_page(struct page *old, struct page *new, gfp_t gfp_mask)
409 {
410         int error;
411
412         VM_BUG_ON(!PageLocked(old));
413         VM_BUG_ON(!PageLocked(new));
414         VM_BUG_ON(new->mapping);
415
416         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
417         if (!error) {
418                 struct address_space *mapping = old->mapping;
419                 void (*freepage)(struct page *);
420
421                 pgoff_t offset = old->index;
422                 freepage = mapping->a_ops->freepage;
423
424                 page_cache_get(new);
425                 new->mapping = mapping;
426                 new->index = offset;
427
428                 spin_lock_irq(&mapping->tree_lock);
429                 __delete_from_page_cache(old);
430                 error = radix_tree_insert(&mapping->page_tree, offset, new);
431                 BUG_ON(error);
432                 mapping->nrpages++;
433                 __inc_zone_page_state(new, NR_FILE_PAGES);
434                 if (PageSwapBacked(new))
435                         __inc_zone_page_state(new, NR_SHMEM);
436                 spin_unlock_irq(&mapping->tree_lock);
437                 /* mem_cgroup codes must not be called under tree_lock */
438                 mem_cgroup_replace_page_cache(old, new);
439                 radix_tree_preload_end();
440                 if (freepage)
441                         freepage(old);
442                 page_cache_release(old);
443         }
444
445         return error;
446 }
447 EXPORT_SYMBOL_GPL(replace_page_cache_page);
448
449 /**
450  * add_to_page_cache_locked - add a locked page to the pagecache
451  * @page:       page to add
452  * @mapping:    the page's address_space
453  * @offset:     page index
454  * @gfp_mask:   page allocation mode
455  *
456  * This function is used to add a page to the pagecache. It must be locked.
457  * This function does not add the page to the LRU.  The caller must do that.
458  */
459 int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
460                 pgoff_t offset, gfp_t gfp_mask)
461 {
462         int error;
463
464         VM_BUG_ON(!PageLocked(page));
465         VM_BUG_ON(PageSwapBacked(page));
466
467         error = mem_cgroup_cache_charge(page, current->mm,
468                                         gfp_mask & GFP_RECLAIM_MASK);
469         if (error)
470                 goto out;
471
472         error = radix_tree_preload(gfp_mask & ~__GFP_HIGHMEM);
473         if (error == 0) {
474                 page_cache_get(page);
475                 page->mapping = mapping;
476                 page->index = offset;
477
478                 spin_lock_irq(&mapping->tree_lock);
479                 error = radix_tree_insert(&mapping->page_tree, offset, page);
480                 if (likely(!error)) {
481                         mapping->nrpages++;
482                         __inc_zone_page_state(page, NR_FILE_PAGES);
483                         spin_unlock_irq(&mapping->tree_lock);
484                         trace_mm_filemap_add_to_page_cache(page);
485                 } else {
486                         page->mapping = NULL;
487                         /* Leave page->index set: truncation relies upon it */
488                         spin_unlock_irq(&mapping->tree_lock);
489                         mem_cgroup_uncharge_cache_page(page);
490                         page_cache_release(page);
491                 }
492                 radix_tree_preload_end();
493         } else
494                 mem_cgroup_uncharge_cache_page(page);
495 out:
496         return error;
497 }
498 EXPORT_SYMBOL(add_to_page_cache_locked);
499
500 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
501                                 pgoff_t offset, gfp_t gfp_mask)
502 {
503         int ret;
504
505         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
506         if (ret == 0)
507                 lru_cache_add_file(page);
508         return ret;
509 }
510 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
511
512 #ifdef CONFIG_NUMA
513 struct page *__page_cache_alloc(gfp_t gfp)
514 {
515         int n;
516         struct page *page;
517
518         if (cpuset_do_page_mem_spread()) {
519                 unsigned int cpuset_mems_cookie;
520                 do {
521                         cpuset_mems_cookie = get_mems_allowed();
522                         n = cpuset_mem_spread_node();
523                         page = alloc_pages_exact_node(n, gfp, 0);
524                 } while (!put_mems_allowed(cpuset_mems_cookie) && !page);
525
526                 return page;
527         }
528         return alloc_pages(gfp, 0);
529 }
530 EXPORT_SYMBOL(__page_cache_alloc);
531 #endif
532
533 /*
534  * In order to wait for pages to become available there must be
535  * waitqueues associated with pages. By using a hash table of
536  * waitqueues where the bucket discipline is to maintain all
537  * waiters on the same queue and wake all when any of the pages
538  * become available, and for the woken contexts to check to be
539  * sure the appropriate page became available, this saves space
540  * at a cost of "thundering herd" phenomena during rare hash
541  * collisions.
542  */
543 static wait_queue_head_t *page_waitqueue(struct page *page)
544 {
545         const struct zone *zone = page_zone(page);
546
547         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
548 }
549
550 static inline void wake_up_page(struct page *page, int bit)
551 {
552         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
553 }
554
555 void wait_on_page_bit(struct page *page, int bit_nr)
556 {
557         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
558
559         if (test_bit(bit_nr, &page->flags))
560                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
561                                                         TASK_UNINTERRUPTIBLE);
562 }
563 EXPORT_SYMBOL(wait_on_page_bit);
564
565 int wait_on_page_bit_killable(struct page *page, int bit_nr)
566 {
567         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
568
569         if (!test_bit(bit_nr, &page->flags))
570                 return 0;
571
572         return __wait_on_bit(page_waitqueue(page), &wait,
573                              sleep_on_page_killable, TASK_KILLABLE);
574 }
575
576 /**
577  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
578  * @page: Page defining the wait queue of interest
579  * @waiter: Waiter to add to the queue
580  *
581  * Add an arbitrary @waiter to the wait queue for the nominated @page.
582  */
583 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
584 {
585         wait_queue_head_t *q = page_waitqueue(page);
586         unsigned long flags;
587
588         spin_lock_irqsave(&q->lock, flags);
589         __add_wait_queue(q, waiter);
590         spin_unlock_irqrestore(&q->lock, flags);
591 }
592 EXPORT_SYMBOL_GPL(add_page_wait_queue);
593
594 /**
595  * unlock_page - unlock a locked page
596  * @page: the page
597  *
598  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
599  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
600  * mechananism between PageLocked pages and PageWriteback pages is shared.
601  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
602  *
603  * The mb is necessary to enforce ordering between the clear_bit and the read
604  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
605  */
606 void unlock_page(struct page *page)
607 {
608         VM_BUG_ON(!PageLocked(page));
609         clear_bit_unlock(PG_locked, &page->flags);
610         smp_mb__after_clear_bit();
611         wake_up_page(page, PG_locked);
612 }
613 EXPORT_SYMBOL(unlock_page);
614
615 /**
616  * end_page_writeback - end writeback against a page
617  * @page: the page
618  */
619 void end_page_writeback(struct page *page)
620 {
621         if (TestClearPageReclaim(page))
622                 rotate_reclaimable_page(page);
623
624         if (!test_clear_page_writeback(page))
625                 BUG();
626
627         smp_mb__after_clear_bit();
628         wake_up_page(page, PG_writeback);
629 }
630 EXPORT_SYMBOL(end_page_writeback);
631
632 /**
633  * __lock_page - get a lock on the page, assuming we need to sleep to get it
634  * @page: the page to lock
635  */
636 void __lock_page(struct page *page)
637 {
638         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
639
640         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
641                                                         TASK_UNINTERRUPTIBLE);
642 }
643 EXPORT_SYMBOL(__lock_page);
644
645 int __lock_page_killable(struct page *page)
646 {
647         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
648
649         return __wait_on_bit_lock(page_waitqueue(page), &wait,
650                                         sleep_on_page_killable, TASK_KILLABLE);
651 }
652 EXPORT_SYMBOL_GPL(__lock_page_killable);
653
654 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
655                          unsigned int flags)
656 {
657         if (flags & FAULT_FLAG_ALLOW_RETRY) {
658                 /*
659                  * CAUTION! In this case, mmap_sem is not released
660                  * even though return 0.
661                  */
662                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
663                         return 0;
664
665                 up_read(&mm->mmap_sem);
666                 if (flags & FAULT_FLAG_KILLABLE)
667                         wait_on_page_locked_killable(page);
668                 else
669                         wait_on_page_locked(page);
670                 return 0;
671         } else {
672                 if (flags & FAULT_FLAG_KILLABLE) {
673                         int ret;
674
675                         ret = __lock_page_killable(page);
676                         if (ret) {
677                                 up_read(&mm->mmap_sem);
678                                 return 0;
679                         }
680                 } else
681                         __lock_page(page);
682                 return 1;
683         }
684 }
685
686 /**
687  * find_get_page - find and get a page reference
688  * @mapping: the address_space to search
689  * @offset: the page index
690  *
691  * Is there a pagecache struct page at the given (mapping, offset) tuple?
692  * If yes, increment its refcount and return it; if no, return NULL.
693  */
694 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
695 {
696         void **pagep;
697         struct page *page;
698
699         rcu_read_lock();
700 repeat:
701         page = NULL;
702         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
703         if (pagep) {
704                 page = radix_tree_deref_slot(pagep);
705                 if (unlikely(!page))
706                         goto out;
707                 if (radix_tree_exception(page)) {
708                         if (radix_tree_deref_retry(page))
709                                 goto repeat;
710                         /*
711                          * Otherwise, shmem/tmpfs must be storing a swap entry
712                          * here as an exceptional entry: so return it without
713                          * attempting to raise page count.
714                          */
715                         goto out;
716                 }
717                 if (!page_cache_get_speculative(page))
718                         goto repeat;
719
720                 /*
721                  * Has the page moved?
722                  * This is part of the lockless pagecache protocol. See
723                  * include/linux/pagemap.h for details.
724                  */
725                 if (unlikely(page != *pagep)) {
726                         page_cache_release(page);
727                         goto repeat;
728                 }
729         }
730 out:
731         rcu_read_unlock();
732
733         return page;
734 }
735 EXPORT_SYMBOL(find_get_page);
736
737 /**
738  * find_lock_page - locate, pin and lock a pagecache page
739  * @mapping: the address_space to search
740  * @offset: the page index
741  *
742  * Locates the desired pagecache page, locks it, increments its reference
743  * count and returns its address.
744  *
745  * Returns zero if the page was not present. find_lock_page() may sleep.
746  */
747 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
748 {
749         struct page *page;
750
751 repeat:
752         page = find_get_page(mapping, offset);
753         if (page && !radix_tree_exception(page)) {
754                 lock_page(page);
755                 /* Has the page been truncated? */
756                 if (unlikely(page->mapping != mapping)) {
757                         unlock_page(page);
758                         page_cache_release(page);
759                         goto repeat;
760                 }
761                 VM_BUG_ON(page->index != offset);
762         }
763         return page;
764 }
765 EXPORT_SYMBOL(find_lock_page);
766
767 /**
768  * find_or_create_page - locate or add a pagecache page
769  * @mapping: the page's address_space
770  * @index: the page's index into the mapping
771  * @gfp_mask: page allocation mode
772  *
773  * Locates a page in the pagecache.  If the page is not present, a new page
774  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
775  * LRU list.  The returned page is locked and has its reference count
776  * incremented.
777  *
778  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
779  * allocation!
780  *
781  * find_or_create_page() returns the desired page's address, or zero on
782  * memory exhaustion.
783  */
784 struct page *find_or_create_page(struct address_space *mapping,
785                 pgoff_t index, gfp_t gfp_mask)
786 {
787         struct page *page;
788         int err;
789 repeat:
790         page = find_lock_page(mapping, index);
791         if (!page) {
792                 page = __page_cache_alloc(gfp_mask);
793                 if (!page)
794                         return NULL;
795                 /*
796                  * We want a regular kernel memory (not highmem or DMA etc)
797                  * allocation for the radix tree nodes, but we need to honour
798                  * the context-specific requirements the caller has asked for.
799                  * GFP_RECLAIM_MASK collects those requirements.
800                  */
801                 err = add_to_page_cache_lru(page, mapping, index,
802                         (gfp_mask & GFP_RECLAIM_MASK));
803                 if (unlikely(err)) {
804                         page_cache_release(page);
805                         page = NULL;
806                         if (err == -EEXIST)
807                                 goto repeat;
808                 }
809         }
810         return page;
811 }
812 EXPORT_SYMBOL(find_or_create_page);
813
814 /**
815  * find_get_pages - gang pagecache lookup
816  * @mapping:    The address_space to search
817  * @start:      The starting page index
818  * @nr_pages:   The maximum number of pages
819  * @pages:      Where the resulting pages are placed
820  *
821  * find_get_pages() will search for and return a group of up to
822  * @nr_pages pages in the mapping.  The pages are placed at @pages.
823  * find_get_pages() takes a reference against the returned pages.
824  *
825  * The search returns a group of mapping-contiguous pages with ascending
826  * indexes.  There may be holes in the indices due to not-present pages.
827  *
828  * find_get_pages() returns the number of pages which were found.
829  */
830 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
831                             unsigned int nr_pages, struct page **pages)
832 {
833         struct radix_tree_iter iter;
834         void **slot;
835         unsigned ret = 0;
836
837         if (unlikely(!nr_pages))
838                 return 0;
839
840         rcu_read_lock();
841 restart:
842         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
843                 struct page *page;
844 repeat:
845                 page = radix_tree_deref_slot(slot);
846                 if (unlikely(!page))
847                         continue;
848
849                 if (radix_tree_exception(page)) {
850                         if (radix_tree_deref_retry(page)) {
851                                 /*
852                                  * Transient condition which can only trigger
853                                  * when entry at index 0 moves out of or back
854                                  * to root: none yet gotten, safe to restart.
855                                  */
856                                 WARN_ON(iter.index);
857                                 goto restart;
858                         }
859                         /*
860                          * Otherwise, shmem/tmpfs must be storing a swap entry
861                          * here as an exceptional entry: so skip over it -
862                          * we only reach this from invalidate_mapping_pages().
863                          */
864                         continue;
865                 }
866
867                 if (!page_cache_get_speculative(page))
868                         goto repeat;
869
870                 /* Has the page moved? */
871                 if (unlikely(page != *slot)) {
872                         page_cache_release(page);
873                         goto repeat;
874                 }
875
876                 pages[ret] = page;
877                 if (++ret == nr_pages)
878                         break;
879         }
880
881         rcu_read_unlock();
882         return ret;
883 }
884
885 /**
886  * find_get_pages_contig - gang contiguous pagecache lookup
887  * @mapping:    The address_space to search
888  * @index:      The starting page index
889  * @nr_pages:   The maximum number of pages
890  * @pages:      Where the resulting pages are placed
891  *
892  * find_get_pages_contig() works exactly like find_get_pages(), except
893  * that the returned number of pages are guaranteed to be contiguous.
894  *
895  * find_get_pages_contig() returns the number of pages which were found.
896  */
897 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
898                                unsigned int nr_pages, struct page **pages)
899 {
900         struct radix_tree_iter iter;
901         void **slot;
902         unsigned int ret = 0;
903
904         if (unlikely(!nr_pages))
905                 return 0;
906
907         rcu_read_lock();
908 restart:
909         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
910                 struct page *page;
911 repeat:
912                 page = radix_tree_deref_slot(slot);
913                 /* The hole, there no reason to continue */
914                 if (unlikely(!page))
915                         break;
916
917                 if (radix_tree_exception(page)) {
918                         if (radix_tree_deref_retry(page)) {
919                                 /*
920                                  * Transient condition which can only trigger
921                                  * when entry at index 0 moves out of or back
922                                  * to root: none yet gotten, safe to restart.
923                                  */
924                                 goto restart;
925                         }
926                         /*
927                          * Otherwise, shmem/tmpfs must be storing a swap entry
928                          * here as an exceptional entry: so stop looking for
929                          * contiguous pages.
930                          */
931                         break;
932                 }
933
934                 if (!page_cache_get_speculative(page))
935                         goto repeat;
936
937                 /* Has the page moved? */
938                 if (unlikely(page != *slot)) {
939                         page_cache_release(page);
940                         goto repeat;
941                 }
942
943                 /*
944                  * must check mapping and index after taking the ref.
945                  * otherwise we can get both false positives and false
946                  * negatives, which is just confusing to the caller.
947                  */
948                 if (page->mapping == NULL || page->index != iter.index) {
949                         page_cache_release(page);
950                         break;
951                 }
952
953                 pages[ret] = page;
954                 if (++ret == nr_pages)
955                         break;
956         }
957         rcu_read_unlock();
958         return ret;
959 }
960 EXPORT_SYMBOL(find_get_pages_contig);
961
962 /**
963  * find_get_pages_tag - find and return pages that match @tag
964  * @mapping:    the address_space to search
965  * @index:      the starting page index
966  * @tag:        the tag index
967  * @nr_pages:   the maximum number of pages
968  * @pages:      where the resulting pages are placed
969  *
970  * Like find_get_pages, except we only return pages which are tagged with
971  * @tag.   We update @index to index the next page for the traversal.
972  */
973 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
974                         int tag, unsigned int nr_pages, struct page **pages)
975 {
976         struct radix_tree_iter iter;
977         void **slot;
978         unsigned ret = 0;
979
980         if (unlikely(!nr_pages))
981                 return 0;
982
983         rcu_read_lock();
984 restart:
985         radix_tree_for_each_tagged(slot, &mapping->page_tree,
986                                    &iter, *index, tag) {
987                 struct page *page;
988 repeat:
989                 page = radix_tree_deref_slot(slot);
990                 if (unlikely(!page))
991                         continue;
992
993                 if (radix_tree_exception(page)) {
994                         if (radix_tree_deref_retry(page)) {
995                                 /*
996                                  * Transient condition which can only trigger
997                                  * when entry at index 0 moves out of or back
998                                  * to root: none yet gotten, safe to restart.
999                                  */
1000                                 goto restart;
1001                         }
1002                         /*
1003                          * This function is never used on a shmem/tmpfs
1004                          * mapping, so a swap entry won't be found here.
1005                          */
1006                         BUG();
1007                 }
1008
1009                 if (!page_cache_get_speculative(page))
1010                         goto repeat;
1011
1012                 /* Has the page moved? */
1013                 if (unlikely(page != *slot)) {
1014                         page_cache_release(page);
1015                         goto repeat;
1016                 }
1017
1018                 pages[ret] = page;
1019                 if (++ret == nr_pages)
1020                         break;
1021         }
1022
1023         rcu_read_unlock();
1024
1025         if (ret)
1026                 *index = pages[ret - 1]->index + 1;
1027
1028         return ret;
1029 }
1030 EXPORT_SYMBOL(find_get_pages_tag);
1031
1032 /**
1033  * grab_cache_page_nowait - returns locked page at given index in given cache
1034  * @mapping: target address_space
1035  * @index: the page index
1036  *
1037  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1038  * This is intended for speculative data generators, where the data can
1039  * be regenerated if the page couldn't be grabbed.  This routine should
1040  * be safe to call while holding the lock for another page.
1041  *
1042  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1043  * and deadlock against the caller's locked page.
1044  */
1045 struct page *
1046 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1047 {
1048         struct page *page = find_get_page(mapping, index);
1049
1050         if (page) {
1051                 if (trylock_page(page))
1052                         return page;
1053                 page_cache_release(page);
1054                 return NULL;
1055         }
1056         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1057         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1058                 page_cache_release(page);
1059                 page = NULL;
1060         }
1061         return page;
1062 }
1063 EXPORT_SYMBOL(grab_cache_page_nowait);
1064
1065 /*
1066  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1067  * a _large_ part of the i/o request. Imagine the worst scenario:
1068  *
1069  *      ---R__________________________________________B__________
1070  *         ^ reading here                             ^ bad block(assume 4k)
1071  *
1072  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1073  * => failing the whole request => read(R) => read(R+1) =>
1074  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1075  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1076  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1077  *
1078  * It is going insane. Fix it by quickly scaling down the readahead size.
1079  */
1080 static void shrink_readahead_size_eio(struct file *filp,
1081                                         struct file_ra_state *ra)
1082 {
1083         ra->ra_pages /= 4;
1084 }
1085
1086 /**
1087  * do_generic_file_read - generic file read routine
1088  * @filp:       the file to read
1089  * @ppos:       current file position
1090  * @desc:       read_descriptor
1091  * @actor:      read method
1092  *
1093  * This is a generic file read routine, and uses the
1094  * mapping->a_ops->readpage() function for the actual low-level stuff.
1095  *
1096  * This is really ugly. But the goto's actually try to clarify some
1097  * of the logic when it comes to error handling etc.
1098  */
1099 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1100                 read_descriptor_t *desc, read_actor_t actor)
1101 {
1102         struct address_space *mapping = filp->f_mapping;
1103         struct inode *inode = mapping->host;
1104         struct file_ra_state *ra = &filp->f_ra;
1105         pgoff_t index;
1106         pgoff_t last_index;
1107         pgoff_t prev_index;
1108         unsigned long offset;      /* offset into pagecache page */
1109         unsigned int prev_offset;
1110         int error;
1111
1112         index = *ppos >> PAGE_CACHE_SHIFT;
1113         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1114         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1115         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1116         offset = *ppos & ~PAGE_CACHE_MASK;
1117
1118         for (;;) {
1119                 struct page *page;
1120                 pgoff_t end_index;
1121                 loff_t isize;
1122                 unsigned long nr, ret;
1123
1124                 cond_resched();
1125 find_page:
1126                 page = find_get_page(mapping, index);
1127                 if (!page) {
1128                         page_cache_sync_readahead(mapping,
1129                                         ra, filp,
1130                                         index, last_index - index);
1131                         page = find_get_page(mapping, index);
1132                         if (unlikely(page == NULL))
1133                                 goto no_cached_page;
1134                 }
1135                 if (PageReadahead(page)) {
1136                         page_cache_async_readahead(mapping,
1137                                         ra, filp, page,
1138                                         index, last_index - index);
1139                 }
1140                 if (!PageUptodate(page)) {
1141                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1142                                         !mapping->a_ops->is_partially_uptodate)
1143                                 goto page_not_up_to_date;
1144                         if (!trylock_page(page))
1145                                 goto page_not_up_to_date;
1146                         /* Did it get truncated before we got the lock? */
1147                         if (!page->mapping)
1148                                 goto page_not_up_to_date_locked;
1149                         if (!mapping->a_ops->is_partially_uptodate(page,
1150                                                                 desc, offset))
1151                                 goto page_not_up_to_date_locked;
1152                         unlock_page(page);
1153                 }
1154 page_ok:
1155                 /*
1156                  * i_size must be checked after we know the page is Uptodate.
1157                  *
1158                  * Checking i_size after the check allows us to calculate
1159                  * the correct value for "nr", which means the zero-filled
1160                  * part of the page is not copied back to userspace (unless
1161                  * another truncate extends the file - this is desired though).
1162                  */
1163
1164                 isize = i_size_read(inode);
1165                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1166                 if (unlikely(!isize || index > end_index)) {
1167                         page_cache_release(page);
1168                         goto out;
1169                 }
1170
1171                 /* nr is the maximum number of bytes to copy from this page */
1172                 nr = PAGE_CACHE_SIZE;
1173                 if (index == end_index) {
1174                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1175                         if (nr <= offset) {
1176                                 page_cache_release(page);
1177                                 goto out;
1178                         }
1179                 }
1180                 nr = nr - offset;
1181
1182                 /* If users can be writing to this page using arbitrary
1183                  * virtual addresses, take care about potential aliasing
1184                  * before reading the page on the kernel side.
1185                  */
1186                 if (mapping_writably_mapped(mapping))
1187                         flush_dcache_page(page);
1188
1189                 /*
1190                  * When a sequential read accesses a page several times,
1191                  * only mark it as accessed the first time.
1192                  */
1193                 if (prev_index != index || offset != prev_offset)
1194                         mark_page_accessed(page);
1195                 prev_index = index;
1196
1197                 /*
1198                  * Ok, we have the page, and it's up-to-date, so
1199                  * now we can copy it to user space...
1200                  *
1201                  * The actor routine returns how many bytes were actually used..
1202                  * NOTE! This may not be the same as how much of a user buffer
1203                  * we filled up (we may be padding etc), so we can only update
1204                  * "pos" here (the actor routine has to update the user buffer
1205                  * pointers and the remaining count).
1206                  */
1207                 ret = actor(desc, page, offset, nr);
1208                 offset += ret;
1209                 index += offset >> PAGE_CACHE_SHIFT;
1210                 offset &= ~PAGE_CACHE_MASK;
1211                 prev_offset = offset;
1212
1213                 page_cache_release(page);
1214                 if (ret == nr && desc->count)
1215                         continue;
1216                 goto out;
1217
1218 page_not_up_to_date:
1219                 /* Get exclusive access to the page ... */
1220                 error = lock_page_killable(page);
1221                 if (unlikely(error))
1222                         goto readpage_error;
1223
1224 page_not_up_to_date_locked:
1225                 /* Did it get truncated before we got the lock? */
1226                 if (!page->mapping) {
1227                         unlock_page(page);
1228                         page_cache_release(page);
1229                         continue;
1230                 }
1231
1232                 /* Did somebody else fill it already? */
1233                 if (PageUptodate(page)) {
1234                         unlock_page(page);
1235                         goto page_ok;
1236                 }
1237
1238 readpage:
1239                 /*
1240                  * A previous I/O error may have been due to temporary
1241                  * failures, eg. multipath errors.
1242                  * PG_error will be set again if readpage fails.
1243                  */
1244                 ClearPageError(page);
1245                 /* Start the actual read. The read will unlock the page. */
1246                 error = mapping->a_ops->readpage(filp, page);
1247
1248                 if (unlikely(error)) {
1249                         if (error == AOP_TRUNCATED_PAGE) {
1250                                 page_cache_release(page);
1251                                 goto find_page;
1252                         }
1253                         goto readpage_error;
1254                 }
1255
1256                 if (!PageUptodate(page)) {
1257                         error = lock_page_killable(page);
1258                         if (unlikely(error))
1259                                 goto readpage_error;
1260                         if (!PageUptodate(page)) {
1261                                 if (page->mapping == NULL) {
1262                                         /*
1263                                          * invalidate_mapping_pages got it
1264                                          */
1265                                         unlock_page(page);
1266                                         page_cache_release(page);
1267                                         goto find_page;
1268                                 }
1269                                 unlock_page(page);
1270                                 shrink_readahead_size_eio(filp, ra);
1271                                 error = -EIO;
1272                                 goto readpage_error;
1273                         }
1274                         unlock_page(page);
1275                 }
1276
1277                 goto page_ok;
1278
1279 readpage_error:
1280                 /* UHHUH! A synchronous read error occurred. Report it */
1281                 desc->error = error;
1282                 page_cache_release(page);
1283                 goto out;
1284
1285 no_cached_page:
1286                 /*
1287                  * Ok, it wasn't cached, so we need to create a new
1288                  * page..
1289                  */
1290                 page = page_cache_alloc_cold(mapping);
1291                 if (!page) {
1292                         desc->error = -ENOMEM;
1293                         goto out;
1294                 }
1295                 error = add_to_page_cache_lru(page, mapping,
1296                                                 index, GFP_KERNEL);
1297                 if (error) {
1298                         page_cache_release(page);
1299                         if (error == -EEXIST)
1300                                 goto find_page;
1301                         desc->error = error;
1302                         goto out;
1303                 }
1304                 goto readpage;
1305         }
1306
1307 out:
1308         ra->prev_pos = prev_index;
1309         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1310         ra->prev_pos |= prev_offset;
1311
1312         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1313         file_accessed(filp);
1314 }
1315
1316 int file_read_actor(read_descriptor_t *desc, struct page *page,
1317                         unsigned long offset, unsigned long size)
1318 {
1319         char *kaddr;
1320         unsigned long left, count = desc->count;
1321
1322         if (size > count)
1323                 size = count;
1324
1325         /*
1326          * Faults on the destination of a read are common, so do it before
1327          * taking the kmap.
1328          */
1329         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1330                 kaddr = kmap_atomic(page);
1331                 left = __copy_to_user_inatomic(desc->arg.buf,
1332                                                 kaddr + offset, size);
1333                 kunmap_atomic(kaddr);
1334                 if (left == 0)
1335                         goto success;
1336         }
1337
1338         /* Do it the slow way */
1339         kaddr = kmap(page);
1340         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1341         kunmap(page);
1342
1343         if (left) {
1344                 size -= left;
1345                 desc->error = -EFAULT;
1346         }
1347 success:
1348         desc->count = count - size;
1349         desc->written += size;
1350         desc->arg.buf += size;
1351         return size;
1352 }
1353
1354 /*
1355  * Performs necessary checks before doing a write
1356  * @iov:        io vector request
1357  * @nr_segs:    number of segments in the iovec
1358  * @count:      number of bytes to write
1359  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1360  *
1361  * Adjust number of segments and amount of bytes to write (nr_segs should be
1362  * properly initialized first). Returns appropriate error code that caller
1363  * should return or zero in case that write should be allowed.
1364  */
1365 int generic_segment_checks(const struct iovec *iov,
1366                         unsigned long *nr_segs, size_t *count, int access_flags)
1367 {
1368         unsigned long   seg;
1369         size_t cnt = 0;
1370         for (seg = 0; seg < *nr_segs; seg++) {
1371                 const struct iovec *iv = &iov[seg];
1372
1373                 /*
1374                  * If any segment has a negative length, or the cumulative
1375                  * length ever wraps negative then return -EINVAL.
1376                  */
1377                 cnt += iv->iov_len;
1378                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1379                         return -EINVAL;
1380                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1381                         continue;
1382                 if (seg == 0)
1383                         return -EFAULT;
1384                 *nr_segs = seg;
1385                 cnt -= iv->iov_len;     /* This segment is no good */
1386                 break;
1387         }
1388         *count = cnt;
1389         return 0;
1390 }
1391 EXPORT_SYMBOL(generic_segment_checks);
1392
1393 /**
1394  * generic_file_aio_read - generic filesystem read routine
1395  * @iocb:       kernel I/O control block
1396  * @iov:        io vector request
1397  * @nr_segs:    number of segments in the iovec
1398  * @pos:        current file position
1399  *
1400  * This is the "read()" routine for all filesystems
1401  * that can use the page cache directly.
1402  */
1403 ssize_t
1404 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1405                 unsigned long nr_segs, loff_t pos)
1406 {
1407         struct file *filp = iocb->ki_filp;
1408         ssize_t retval;
1409         unsigned long seg = 0;
1410         size_t count;
1411         loff_t *ppos = &iocb->ki_pos;
1412
1413         count = 0;
1414         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1415         if (retval)
1416                 return retval;
1417
1418         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1419         if (filp->f_flags & O_DIRECT) {
1420                 loff_t size;
1421                 struct address_space *mapping;
1422                 struct inode *inode;
1423
1424                 mapping = filp->f_mapping;
1425                 inode = mapping->host;
1426                 if (!count)
1427                         goto out; /* skip atime */
1428                 size = i_size_read(inode);
1429                 if (pos < size) {
1430                         retval = filemap_write_and_wait_range(mapping, pos,
1431                                         pos + iov_length(iov, nr_segs) - 1);
1432                         if (!retval) {
1433                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1434                                                         iov, pos, nr_segs);
1435                         }
1436                         if (retval > 0) {
1437                                 *ppos = pos + retval;
1438                                 count -= retval;
1439                         }
1440
1441                         /*
1442                          * Btrfs can have a short DIO read if we encounter
1443                          * compressed extents, so if there was an error, or if
1444                          * we've already read everything we wanted to, or if
1445                          * there was a short read because we hit EOF, go ahead
1446                          * and return.  Otherwise fallthrough to buffered io for
1447                          * the rest of the read.
1448                          */
1449                         if (retval < 0 || !count || *ppos >= size) {
1450                                 file_accessed(filp);
1451                                 goto out;
1452                         }
1453                 }
1454         }
1455
1456         count = retval;
1457         for (seg = 0; seg < nr_segs; seg++) {
1458                 read_descriptor_t desc;
1459                 loff_t offset = 0;
1460
1461                 /*
1462                  * If we did a short DIO read we need to skip the section of the
1463                  * iov that we've already read data into.
1464                  */
1465                 if (count) {
1466                         if (count > iov[seg].iov_len) {
1467                                 count -= iov[seg].iov_len;
1468                                 continue;
1469                         }
1470                         offset = count;
1471                         count = 0;
1472                 }
1473
1474                 desc.written = 0;
1475                 desc.arg.buf = iov[seg].iov_base + offset;
1476                 desc.count = iov[seg].iov_len - offset;
1477                 if (desc.count == 0)
1478                         continue;
1479                 desc.error = 0;
1480                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1481                 retval += desc.written;
1482                 if (desc.error) {
1483                         retval = retval ?: desc.error;
1484                         break;
1485                 }
1486                 if (desc.count > 0)
1487                         break;
1488         }
1489 out:
1490         return retval;
1491 }
1492 EXPORT_SYMBOL(generic_file_aio_read);
1493
1494 #ifdef CONFIG_MMU
1495 /**
1496  * page_cache_read - adds requested page to the page cache if not already there
1497  * @file:       file to read
1498  * @offset:     page index
1499  *
1500  * This adds the requested page to the page cache if it isn't already there,
1501  * and schedules an I/O to read in its contents from disk.
1502  */
1503 static int page_cache_read(struct file *file, pgoff_t offset)
1504 {
1505         struct address_space *mapping = file->f_mapping;
1506         struct page *page; 
1507         int ret;
1508
1509         do {
1510                 page = page_cache_alloc_cold(mapping);
1511                 if (!page)
1512                         return -ENOMEM;
1513
1514                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1515                 if (ret == 0)
1516                         ret = mapping->a_ops->readpage(file, page);
1517                 else if (ret == -EEXIST)
1518                         ret = 0; /* losing race to add is OK */
1519
1520                 page_cache_release(page);
1521
1522         } while (ret == AOP_TRUNCATED_PAGE);
1523                 
1524         return ret;
1525 }
1526
1527 #define MMAP_LOTSAMISS  (100)
1528
1529 /*
1530  * Synchronous readahead happens when we don't even find
1531  * a page in the page cache at all.
1532  */
1533 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1534                                    struct file_ra_state *ra,
1535                                    struct file *file,
1536                                    pgoff_t offset)
1537 {
1538         unsigned long ra_pages;
1539         struct address_space *mapping = file->f_mapping;
1540
1541         /* If we don't want any read-ahead, don't bother */
1542         if (VM_RandomReadHint(vma))
1543                 return;
1544         if (!ra->ra_pages)
1545                 return;
1546
1547         if (VM_SequentialReadHint(vma)) {
1548                 page_cache_sync_readahead(mapping, ra, file, offset,
1549                                           ra->ra_pages);
1550                 return;
1551         }
1552
1553         /* Avoid banging the cache line if not needed */
1554         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1555                 ra->mmap_miss++;
1556
1557         /*
1558          * Do we miss much more than hit in this file? If so,
1559          * stop bothering with read-ahead. It will only hurt.
1560          */
1561         if (ra->mmap_miss > MMAP_LOTSAMISS)
1562                 return;
1563
1564         /*
1565          * mmap read-around
1566          */
1567         ra_pages = max_sane_readahead(ra->ra_pages);
1568         ra->start = max_t(long, 0, offset - ra_pages / 2);
1569         ra->size = ra_pages;
1570         ra->async_size = ra_pages / 4;
1571         ra_submit(ra, mapping, file);
1572 }
1573
1574 /*
1575  * Asynchronous readahead happens when we find the page and PG_readahead,
1576  * so we want to possibly extend the readahead further..
1577  */
1578 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1579                                     struct file_ra_state *ra,
1580                                     struct file *file,
1581                                     struct page *page,
1582                                     pgoff_t offset)
1583 {
1584         struct address_space *mapping = file->f_mapping;
1585
1586         /* If we don't want any read-ahead, don't bother */
1587         if (VM_RandomReadHint(vma))
1588                 return;
1589         if (ra->mmap_miss > 0)
1590                 ra->mmap_miss--;
1591         if (PageReadahead(page))
1592                 page_cache_async_readahead(mapping, ra, file,
1593                                            page, offset, ra->ra_pages);
1594 }
1595
1596 /**
1597  * filemap_fault - read in file data for page fault handling
1598  * @vma:        vma in which the fault was taken
1599  * @vmf:        struct vm_fault containing details of the fault
1600  *
1601  * filemap_fault() is invoked via the vma operations vector for a
1602  * mapped memory region to read in file data during a page fault.
1603  *
1604  * The goto's are kind of ugly, but this streamlines the normal case of having
1605  * it in the page cache, and handles the special cases reasonably without
1606  * having a lot of duplicated code.
1607  */
1608 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1609 {
1610         int error;
1611         struct file *file = vma->vm_file;
1612         struct address_space *mapping = file->f_mapping;
1613         struct file_ra_state *ra = &file->f_ra;
1614         struct inode *inode = mapping->host;
1615         pgoff_t offset = vmf->pgoff;
1616         struct page *page;
1617         pgoff_t size;
1618         int ret = 0;
1619
1620         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1621         if (offset >= size)
1622                 return VM_FAULT_SIGBUS;
1623
1624         /*
1625          * Do we have something in the page cache already?
1626          */
1627         page = find_get_page(mapping, offset);
1628         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
1629                 /*
1630                  * We found the page, so try async readahead before
1631                  * waiting for the lock.
1632                  */
1633                 do_async_mmap_readahead(vma, ra, file, page, offset);
1634         } else if (!page) {
1635                 /* No page in the page cache at all */
1636                 do_sync_mmap_readahead(vma, ra, file, offset);
1637                 count_vm_event(PGMAJFAULT);
1638                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1639                 ret = VM_FAULT_MAJOR;
1640 retry_find:
1641                 page = find_get_page(mapping, offset);
1642                 if (!page)
1643                         goto no_cached_page;
1644         }
1645
1646         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1647                 page_cache_release(page);
1648                 return ret | VM_FAULT_RETRY;
1649         }
1650
1651         /* Did it get truncated? */
1652         if (unlikely(page->mapping != mapping)) {
1653                 unlock_page(page);
1654                 put_page(page);
1655                 goto retry_find;
1656         }
1657         VM_BUG_ON(page->index != offset);
1658
1659         /*
1660          * We have a locked page in the page cache, now we need to check
1661          * that it's up-to-date. If not, it is going to be due to an error.
1662          */
1663         if (unlikely(!PageUptodate(page)))
1664                 goto page_not_uptodate;
1665
1666         /*
1667          * Found the page and have a reference on it.
1668          * We must recheck i_size under page lock.
1669          */
1670         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1671         if (unlikely(offset >= size)) {
1672                 unlock_page(page);
1673                 page_cache_release(page);
1674                 return VM_FAULT_SIGBUS;
1675         }
1676
1677         vmf->page = page;
1678         return ret | VM_FAULT_LOCKED;
1679
1680 no_cached_page:
1681         /*
1682          * We're only likely to ever get here if MADV_RANDOM is in
1683          * effect.
1684          */
1685         error = page_cache_read(file, offset);
1686
1687         /*
1688          * The page we want has now been added to the page cache.
1689          * In the unlikely event that someone removed it in the
1690          * meantime, we'll just come back here and read it again.
1691          */
1692         if (error >= 0)
1693                 goto retry_find;
1694
1695         /*
1696          * An error return from page_cache_read can result if the
1697          * system is low on memory, or a problem occurs while trying
1698          * to schedule I/O.
1699          */
1700         if (error == -ENOMEM)
1701                 return VM_FAULT_OOM;
1702         return VM_FAULT_SIGBUS;
1703
1704 page_not_uptodate:
1705         /*
1706          * Umm, take care of errors if the page isn't up-to-date.
1707          * Try to re-read it _once_. We do this synchronously,
1708          * because there really aren't any performance issues here
1709          * and we need to check for errors.
1710          */
1711         ClearPageError(page);
1712         error = mapping->a_ops->readpage(file, page);
1713         if (!error) {
1714                 wait_on_page_locked(page);
1715                 if (!PageUptodate(page))
1716                         error = -EIO;
1717         }
1718         page_cache_release(page);
1719
1720         if (!error || error == AOP_TRUNCATED_PAGE)
1721                 goto retry_find;
1722
1723         /* Things didn't work out. Return zero to tell the mm layer so. */
1724         shrink_readahead_size_eio(file, ra);
1725         return VM_FAULT_SIGBUS;
1726 }
1727 EXPORT_SYMBOL(filemap_fault);
1728
1729 int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1730 {
1731         struct page *page = vmf->page;
1732         struct inode *inode = file_inode(vma->vm_file);
1733         int ret = VM_FAULT_LOCKED;
1734
1735         sb_start_pagefault(inode->i_sb);
1736         file_update_time(vma->vm_file);
1737         lock_page(page);
1738         if (page->mapping != inode->i_mapping) {
1739                 unlock_page(page);
1740                 ret = VM_FAULT_NOPAGE;
1741                 goto out;
1742         }
1743         /*
1744          * We mark the page dirty already here so that when freeze is in
1745          * progress, we are guaranteed that writeback during freezing will
1746          * see the dirty page and writeprotect it again.
1747          */
1748         set_page_dirty(page);
1749         wait_for_stable_page(page);
1750 out:
1751         sb_end_pagefault(inode->i_sb);
1752         return ret;
1753 }
1754 EXPORT_SYMBOL(filemap_page_mkwrite);
1755
1756 const struct vm_operations_struct generic_file_vm_ops = {
1757         .fault          = filemap_fault,
1758         .page_mkwrite   = filemap_page_mkwrite,
1759         .remap_pages    = generic_file_remap_pages,
1760 };
1761
1762 /* This is used for a general mmap of a disk file */
1763
1764 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1765 {
1766         struct address_space *mapping = file->f_mapping;
1767
1768         if (!mapping->a_ops->readpage)
1769                 return -ENOEXEC;
1770         file_accessed(file);
1771         vma->vm_ops = &generic_file_vm_ops;
1772         return 0;
1773 }
1774
1775 /*
1776  * This is for filesystems which do not implement ->writepage.
1777  */
1778 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1779 {
1780         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1781                 return -EINVAL;
1782         return generic_file_mmap(file, vma);
1783 }
1784 #else
1785 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1786 {
1787         return -ENOSYS;
1788 }
1789 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1790 {
1791         return -ENOSYS;
1792 }
1793 #endif /* CONFIG_MMU */
1794
1795 EXPORT_SYMBOL(generic_file_mmap);
1796 EXPORT_SYMBOL(generic_file_readonly_mmap);
1797
1798 static struct page *__read_cache_page(struct address_space *mapping,
1799                                 pgoff_t index,
1800                                 int (*filler)(void *, struct page *),
1801                                 void *data,
1802                                 gfp_t gfp)
1803 {
1804         struct page *page;
1805         int err;
1806 repeat:
1807         page = find_get_page(mapping, index);
1808         if (!page) {
1809                 page = __page_cache_alloc(gfp | __GFP_COLD);
1810                 if (!page)
1811                         return ERR_PTR(-ENOMEM);
1812                 err = add_to_page_cache_lru(page, mapping, index, gfp);
1813                 if (unlikely(err)) {
1814                         page_cache_release(page);
1815                         if (err == -EEXIST)
1816                                 goto repeat;
1817                         /* Presumably ENOMEM for radix tree node */
1818                         return ERR_PTR(err);
1819                 }
1820                 err = filler(data, page);
1821                 if (err < 0) {
1822                         page_cache_release(page);
1823                         page = ERR_PTR(err);
1824                 }
1825         }
1826         return page;
1827 }
1828
1829 static struct page *do_read_cache_page(struct address_space *mapping,
1830                                 pgoff_t index,
1831                                 int (*filler)(void *, struct page *),
1832                                 void *data,
1833                                 gfp_t gfp)
1834
1835 {
1836         struct page *page;
1837         int err;
1838
1839 retry:
1840         page = __read_cache_page(mapping, index, filler, data, gfp);
1841         if (IS_ERR(page))
1842                 return page;
1843         if (PageUptodate(page))
1844                 goto out;
1845
1846         lock_page(page);
1847         if (!page->mapping) {
1848                 unlock_page(page);
1849                 page_cache_release(page);
1850                 goto retry;
1851         }
1852         if (PageUptodate(page)) {
1853                 unlock_page(page);
1854                 goto out;
1855         }
1856         err = filler(data, page);
1857         if (err < 0) {
1858                 page_cache_release(page);
1859                 return ERR_PTR(err);
1860         }
1861 out:
1862         mark_page_accessed(page);
1863         return page;
1864 }
1865
1866 /**
1867  * read_cache_page_async - read into page cache, fill it if needed
1868  * @mapping:    the page's address_space
1869  * @index:      the page index
1870  * @filler:     function to perform the read
1871  * @data:       first arg to filler(data, page) function, often left as NULL
1872  *
1873  * Same as read_cache_page, but don't wait for page to become unlocked
1874  * after submitting it to the filler.
1875  *
1876  * Read into the page cache. If a page already exists, and PageUptodate() is
1877  * not set, try to fill the page but don't wait for it to become unlocked.
1878  *
1879  * If the page does not get brought uptodate, return -EIO.
1880  */
1881 struct page *read_cache_page_async(struct address_space *mapping,
1882                                 pgoff_t index,
1883                                 int (*filler)(void *, struct page *),
1884                                 void *data)
1885 {
1886         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1887 }
1888 EXPORT_SYMBOL(read_cache_page_async);
1889
1890 static struct page *wait_on_page_read(struct page *page)
1891 {
1892         if (!IS_ERR(page)) {
1893                 wait_on_page_locked(page);
1894                 if (!PageUptodate(page)) {
1895                         page_cache_release(page);
1896                         page = ERR_PTR(-EIO);
1897                 }
1898         }
1899         return page;
1900 }
1901
1902 /**
1903  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1904  * @mapping:    the page's address_space
1905  * @index:      the page index
1906  * @gfp:        the page allocator flags to use if allocating
1907  *
1908  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1909  * any new page allocations done using the specified allocation flags.
1910  *
1911  * If the page does not get brought uptodate, return -EIO.
1912  */
1913 struct page *read_cache_page_gfp(struct address_space *mapping,
1914                                 pgoff_t index,
1915                                 gfp_t gfp)
1916 {
1917         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1918
1919         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1920 }
1921 EXPORT_SYMBOL(read_cache_page_gfp);
1922
1923 /**
1924  * read_cache_page - read into page cache, fill it if needed
1925  * @mapping:    the page's address_space
1926  * @index:      the page index
1927  * @filler:     function to perform the read
1928  * @data:       first arg to filler(data, page) function, often left as NULL
1929  *
1930  * Read into the page cache. If a page already exists, and PageUptodate() is
1931  * not set, try to fill the page then wait for it to become unlocked.
1932  *
1933  * If the page does not get brought uptodate, return -EIO.
1934  */
1935 struct page *read_cache_page(struct address_space *mapping,
1936                                 pgoff_t index,
1937                                 int (*filler)(void *, struct page *),
1938                                 void *data)
1939 {
1940         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1941 }
1942 EXPORT_SYMBOL(read_cache_page);
1943
1944 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1945                         const struct iovec *iov, size_t base, size_t bytes)
1946 {
1947         size_t copied = 0, left = 0;
1948
1949         while (bytes) {
1950                 char __user *buf = iov->iov_base + base;
1951                 int copy = min(bytes, iov->iov_len - base);
1952
1953                 base = 0;
1954                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1955                 copied += copy;
1956                 bytes -= copy;
1957                 vaddr += copy;
1958                 iov++;
1959
1960                 if (unlikely(left))
1961                         break;
1962         }
1963         return copied - left;
1964 }
1965
1966 /*
1967  * Copy as much as we can into the page and return the number of bytes which
1968  * were successfully copied.  If a fault is encountered then return the number of
1969  * bytes which were copied.
1970  */
1971 size_t iov_iter_copy_from_user_atomic(struct page *page,
1972                 struct iov_iter *i, unsigned long offset, size_t bytes)
1973 {
1974         char *kaddr;
1975         size_t copied;
1976
1977         BUG_ON(!in_atomic());
1978         kaddr = kmap_atomic(page);
1979         if (likely(i->nr_segs == 1)) {
1980                 int left;
1981                 char __user *buf = i->iov->iov_base + i->iov_offset;
1982                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1983                 copied = bytes - left;
1984         } else {
1985                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1986                                                 i->iov, i->iov_offset, bytes);
1987         }
1988         kunmap_atomic(kaddr);
1989
1990         return copied;
1991 }
1992 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
1993
1994 /*
1995  * This has the same sideeffects and return value as
1996  * iov_iter_copy_from_user_atomic().
1997  * The difference is that it attempts to resolve faults.
1998  * Page must not be locked.
1999  */
2000 size_t iov_iter_copy_from_user(struct page *page,
2001                 struct iov_iter *i, unsigned long offset, size_t bytes)
2002 {
2003         char *kaddr;
2004         size_t copied;
2005
2006         kaddr = kmap(page);
2007         if (likely(i->nr_segs == 1)) {
2008                 int left;
2009                 char __user *buf = i->iov->iov_base + i->iov_offset;
2010                 left = __copy_from_user(kaddr + offset, buf, bytes);
2011                 copied = bytes - left;
2012         } else {
2013                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2014                                                 i->iov, i->iov_offset, bytes);
2015         }
2016         kunmap(page);
2017         return copied;
2018 }
2019 EXPORT_SYMBOL(iov_iter_copy_from_user);
2020
2021 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2022 {
2023         BUG_ON(i->count < bytes);
2024
2025         if (likely(i->nr_segs == 1)) {
2026                 i->iov_offset += bytes;
2027                 i->count -= bytes;
2028         } else {
2029                 const struct iovec *iov = i->iov;
2030                 size_t base = i->iov_offset;
2031                 unsigned long nr_segs = i->nr_segs;
2032
2033                 /*
2034                  * The !iov->iov_len check ensures we skip over unlikely
2035                  * zero-length segments (without overruning the iovec).
2036                  */
2037                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2038                         int copy;
2039
2040                         copy = min(bytes, iov->iov_len - base);
2041                         BUG_ON(!i->count || i->count < copy);
2042                         i->count -= copy;
2043                         bytes -= copy;
2044                         base += copy;
2045                         if (iov->iov_len == base) {
2046                                 iov++;
2047                                 nr_segs--;
2048                                 base = 0;
2049                         }
2050                 }
2051                 i->iov = iov;
2052                 i->iov_offset = base;
2053                 i->nr_segs = nr_segs;
2054         }
2055 }
2056 EXPORT_SYMBOL(iov_iter_advance);
2057
2058 /*
2059  * Fault in the first iovec of the given iov_iter, to a maximum length
2060  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2061  * accessed (ie. because it is an invalid address).
2062  *
2063  * writev-intensive code may want this to prefault several iovecs -- that
2064  * would be possible (callers must not rely on the fact that _only_ the
2065  * first iovec will be faulted with the current implementation).
2066  */
2067 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2068 {
2069         char __user *buf = i->iov->iov_base + i->iov_offset;
2070         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2071         return fault_in_pages_readable(buf, bytes);
2072 }
2073 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2074
2075 /*
2076  * Return the count of just the current iov_iter segment.
2077  */
2078 size_t iov_iter_single_seg_count(const struct iov_iter *i)
2079 {
2080         const struct iovec *iov = i->iov;
2081         if (i->nr_segs == 1)
2082                 return i->count;
2083         else
2084                 return min(i->count, iov->iov_len - i->iov_offset);
2085 }
2086 EXPORT_SYMBOL(iov_iter_single_seg_count);
2087
2088 /*
2089  * Performs necessary checks before doing a write
2090  *
2091  * Can adjust writing position or amount of bytes to write.
2092  * Returns appropriate error code that caller should return or
2093  * zero in case that write should be allowed.
2094  */
2095 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2096 {
2097         struct inode *inode = file->f_mapping->host;
2098         unsigned long limit = rlimit(RLIMIT_FSIZE);
2099
2100         if (unlikely(*pos < 0))
2101                 return -EINVAL;
2102
2103         if (!isblk) {
2104                 /* FIXME: this is for backwards compatibility with 2.4 */
2105                 if (file->f_flags & O_APPEND)
2106                         *pos = i_size_read(inode);
2107
2108                 if (limit != RLIM_INFINITY) {
2109                         if (*pos >= limit) {
2110                                 send_sig(SIGXFSZ, current, 0);
2111                                 return -EFBIG;
2112                         }
2113                         if (*count > limit - (typeof(limit))*pos) {
2114                                 *count = limit - (typeof(limit))*pos;
2115                         }
2116                 }
2117         }
2118
2119         /*
2120          * LFS rule
2121          */
2122         if (unlikely(*pos + *count > MAX_NON_LFS &&
2123                                 !(file->f_flags & O_LARGEFILE))) {
2124                 if (*pos >= MAX_NON_LFS) {
2125                         return -EFBIG;
2126                 }
2127                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2128                         *count = MAX_NON_LFS - (unsigned long)*pos;
2129                 }
2130         }
2131
2132         /*
2133          * Are we about to exceed the fs block limit ?
2134          *
2135          * If we have written data it becomes a short write.  If we have
2136          * exceeded without writing data we send a signal and return EFBIG.
2137          * Linus frestrict idea will clean these up nicely..
2138          */
2139         if (likely(!isblk)) {
2140                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2141                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2142                                 return -EFBIG;
2143                         }
2144                         /* zero-length writes at ->s_maxbytes are OK */
2145                 }
2146
2147                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2148                         *count = inode->i_sb->s_maxbytes - *pos;
2149         } else {
2150 #ifdef CONFIG_BLOCK
2151                 loff_t isize;
2152                 if (bdev_read_only(I_BDEV(inode)))
2153                         return -EPERM;
2154                 isize = i_size_read(inode);
2155                 if (*pos >= isize) {
2156                         if (*count || *pos > isize)
2157                                 return -ENOSPC;
2158                 }
2159
2160                 if (*pos + *count > isize)
2161                         *count = isize - *pos;
2162 #else
2163                 return -EPERM;
2164 #endif
2165         }
2166         return 0;
2167 }
2168 EXPORT_SYMBOL(generic_write_checks);
2169
2170 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2171                                 loff_t pos, unsigned len, unsigned flags,
2172                                 struct page **pagep, void **fsdata)
2173 {
2174         const struct address_space_operations *aops = mapping->a_ops;
2175
2176         return aops->write_begin(file, mapping, pos, len, flags,
2177                                                         pagep, fsdata);
2178 }
2179 EXPORT_SYMBOL(pagecache_write_begin);
2180
2181 int pagecache_write_end(struct file *file, struct address_space *mapping,
2182                                 loff_t pos, unsigned len, unsigned copied,
2183                                 struct page *page, void *fsdata)
2184 {
2185         const struct address_space_operations *aops = mapping->a_ops;
2186
2187         mark_page_accessed(page);
2188         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2189 }
2190 EXPORT_SYMBOL(pagecache_write_end);
2191
2192 ssize_t
2193 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2194                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2195                 size_t count, size_t ocount)
2196 {
2197         struct file     *file = iocb->ki_filp;
2198         struct address_space *mapping = file->f_mapping;
2199         struct inode    *inode = mapping->host;
2200         ssize_t         written;
2201         size_t          write_len;
2202         pgoff_t         end;
2203
2204         if (count != ocount)
2205                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2206
2207         write_len = iov_length(iov, *nr_segs);
2208         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2209
2210         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2211         if (written)
2212                 goto out;
2213
2214         /*
2215          * After a write we want buffered reads to be sure to go to disk to get
2216          * the new data.  We invalidate clean cached page from the region we're
2217          * about to write.  We do this *before* the write so that we can return
2218          * without clobbering -EIOCBQUEUED from ->direct_IO().
2219          */
2220         if (mapping->nrpages) {
2221                 written = invalidate_inode_pages2_range(mapping,
2222                                         pos >> PAGE_CACHE_SHIFT, end);
2223                 /*
2224                  * If a page can not be invalidated, return 0 to fall back
2225                  * to buffered write.
2226                  */
2227                 if (written) {
2228                         if (written == -EBUSY)
2229                                 return 0;
2230                         goto out;
2231                 }
2232         }
2233
2234         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2235
2236         /*
2237          * Finally, try again to invalidate clean pages which might have been
2238          * cached by non-direct readahead, or faulted in by get_user_pages()
2239          * if the source of the write was an mmap'ed region of the file
2240          * we're writing.  Either one is a pretty crazy thing to do,
2241          * so we don't support it 100%.  If this invalidation
2242          * fails, tough, the write still worked...
2243          */
2244         if (mapping->nrpages) {
2245                 invalidate_inode_pages2_range(mapping,
2246                                               pos >> PAGE_CACHE_SHIFT, end);
2247         }
2248
2249         if (written > 0) {
2250                 pos += written;
2251                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2252                         i_size_write(inode, pos);
2253                         mark_inode_dirty(inode);
2254                 }
2255                 *ppos = pos;
2256         }
2257 out:
2258         return written;
2259 }
2260 EXPORT_SYMBOL(generic_file_direct_write);
2261
2262 /*
2263  * Find or create a page at the given pagecache position. Return the locked
2264  * page. This function is specifically for buffered writes.
2265  */
2266 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2267                                         pgoff_t index, unsigned flags)
2268 {
2269         int status;
2270         gfp_t gfp_mask;
2271         struct page *page;
2272         gfp_t gfp_notmask = 0;
2273
2274         gfp_mask = mapping_gfp_mask(mapping);
2275         if (mapping_cap_account_dirty(mapping))
2276                 gfp_mask |= __GFP_WRITE;
2277         if (flags & AOP_FLAG_NOFS)
2278                 gfp_notmask = __GFP_FS;
2279 repeat:
2280         page = find_lock_page(mapping, index);
2281         if (page)
2282                 goto found;
2283
2284         page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2285         if (!page)
2286                 return NULL;
2287         status = add_to_page_cache_lru(page, mapping, index,
2288                                                 GFP_KERNEL & ~gfp_notmask);
2289         if (unlikely(status)) {
2290                 page_cache_release(page);
2291                 if (status == -EEXIST)
2292                         goto repeat;
2293                 return NULL;
2294         }
2295 found:
2296         wait_for_stable_page(page);
2297         return page;
2298 }
2299 EXPORT_SYMBOL(grab_cache_page_write_begin);
2300
2301 static ssize_t generic_perform_write(struct file *file,
2302                                 struct iov_iter *i, loff_t pos)
2303 {
2304         struct address_space *mapping = file->f_mapping;
2305         const struct address_space_operations *a_ops = mapping->a_ops;
2306         long status = 0;
2307         ssize_t written = 0;
2308         unsigned int flags = 0;
2309
2310         /*
2311          * Copies from kernel address space cannot fail (NFSD is a big user).
2312          */
2313         if (segment_eq(get_fs(), KERNEL_DS))
2314                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2315
2316         do {
2317                 struct page *page;
2318                 unsigned long offset;   /* Offset into pagecache page */
2319                 unsigned long bytes;    /* Bytes to write to page */
2320                 size_t copied;          /* Bytes copied from user */
2321                 void *fsdata;
2322
2323                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2324                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2325                                                 iov_iter_count(i));
2326
2327 again:
2328                 /*
2329                  * Bring in the user page that we will copy from _first_.
2330                  * Otherwise there's a nasty deadlock on copying from the
2331                  * same page as we're writing to, without it being marked
2332                  * up-to-date.
2333                  *
2334                  * Not only is this an optimisation, but it is also required
2335                  * to check that the address is actually valid, when atomic
2336                  * usercopies are used, below.
2337                  */
2338                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2339                         status = -EFAULT;
2340                         break;
2341                 }
2342
2343                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2344                                                 &page, &fsdata);
2345                 if (unlikely(status))
2346                         break;
2347
2348                 if (mapping_writably_mapped(mapping))
2349                         flush_dcache_page(page);
2350
2351                 pagefault_disable();
2352                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2353                 pagefault_enable();
2354                 flush_dcache_page(page);
2355
2356                 mark_page_accessed(page);
2357                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2358                                                 page, fsdata);
2359                 if (unlikely(status < 0))
2360                         break;
2361                 copied = status;
2362
2363                 cond_resched();
2364
2365                 iov_iter_advance(i, copied);
2366                 if (unlikely(copied == 0)) {
2367                         /*
2368                          * If we were unable to copy any data at all, we must
2369                          * fall back to a single segment length write.
2370                          *
2371                          * If we didn't fallback here, we could livelock
2372                          * because not all segments in the iov can be copied at
2373                          * once without a pagefault.
2374                          */
2375                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2376                                                 iov_iter_single_seg_count(i));
2377                         goto again;
2378                 }
2379                 pos += copied;
2380                 written += copied;
2381
2382                 balance_dirty_pages_ratelimited(mapping);
2383                 if (fatal_signal_pending(current)) {
2384                         status = -EINTR;
2385                         break;
2386                 }
2387         } while (iov_iter_count(i));
2388
2389         return written ? written : status;
2390 }
2391
2392 ssize_t
2393 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2394                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2395                 size_t count, ssize_t written)
2396 {
2397         struct file *file = iocb->ki_filp;
2398         ssize_t status;
2399         struct iov_iter i;
2400
2401         iov_iter_init(&i, iov, nr_segs, count, written);
2402         status = generic_perform_write(file, &i, pos);
2403
2404         if (likely(status >= 0)) {
2405                 written += status;
2406                 *ppos = pos + status;
2407         }
2408         
2409         return written ? written : status;
2410 }
2411 EXPORT_SYMBOL(generic_file_buffered_write);
2412
2413 /**
2414  * __generic_file_aio_write - write data to a file
2415  * @iocb:       IO state structure (file, offset, etc.)
2416  * @iov:        vector with data to write
2417  * @nr_segs:    number of segments in the vector
2418  * @ppos:       position where to write
2419  *
2420  * This function does all the work needed for actually writing data to a
2421  * file. It does all basic checks, removes SUID from the file, updates
2422  * modification times and calls proper subroutines depending on whether we
2423  * do direct IO or a standard buffered write.
2424  *
2425  * It expects i_mutex to be grabbed unless we work on a block device or similar
2426  * object which does not need locking at all.
2427  *
2428  * This function does *not* take care of syncing data in case of O_SYNC write.
2429  * A caller has to handle it. This is mainly due to the fact that we want to
2430  * avoid syncing under i_mutex.
2431  */
2432 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2433                                  unsigned long nr_segs, loff_t *ppos)
2434 {
2435         struct file *file = iocb->ki_filp;
2436         struct address_space * mapping = file->f_mapping;
2437         size_t ocount;          /* original count */
2438         size_t count;           /* after file limit checks */
2439         struct inode    *inode = mapping->host;
2440         loff_t          pos;
2441         ssize_t         written;
2442         ssize_t         err;
2443
2444         ocount = 0;
2445         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2446         if (err)
2447                 return err;
2448
2449         count = ocount;
2450         pos = *ppos;
2451
2452         /* We can write back this queue in page reclaim */
2453         current->backing_dev_info = mapping->backing_dev_info;
2454         written = 0;
2455
2456         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2457         if (err)
2458                 goto out;
2459
2460         if (count == 0)
2461                 goto out;
2462
2463         err = file_remove_suid(file);
2464         if (err)
2465                 goto out;
2466
2467         err = file_update_time(file);
2468         if (err)
2469                 goto out;
2470
2471         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2472         if (unlikely(file->f_flags & O_DIRECT)) {
2473                 loff_t endbyte;
2474                 ssize_t written_buffered;
2475
2476                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2477                                                         ppos, count, ocount);
2478                 if (written < 0 || written == count)
2479                         goto out;
2480                 /*
2481                  * direct-io write to a hole: fall through to buffered I/O
2482                  * for completing the rest of the request.
2483                  */
2484                 pos += written;
2485                 count -= written;
2486                 written_buffered = generic_file_buffered_write(iocb, iov,
2487                                                 nr_segs, pos, ppos, count,
2488                                                 written);
2489                 /*
2490                  * If generic_file_buffered_write() retuned a synchronous error
2491                  * then we want to return the number of bytes which were
2492                  * direct-written, or the error code if that was zero.  Note
2493                  * that this differs from normal direct-io semantics, which
2494                  * will return -EFOO even if some bytes were written.
2495                  */
2496                 if (written_buffered < 0) {
2497                         err = written_buffered;
2498                         goto out;
2499                 }
2500
2501                 /*
2502                  * We need to ensure that the page cache pages are written to
2503                  * disk and invalidated to preserve the expected O_DIRECT
2504                  * semantics.
2505                  */
2506                 endbyte = pos + written_buffered - written - 1;
2507                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2508                 if (err == 0) {
2509                         written = written_buffered;
2510                         invalidate_mapping_pages(mapping,
2511                                                  pos >> PAGE_CACHE_SHIFT,
2512                                                  endbyte >> PAGE_CACHE_SHIFT);
2513                 } else {
2514                         /*
2515                          * We don't know how much we wrote, so just return
2516                          * the number of bytes which were direct-written
2517                          */
2518                 }
2519         } else {
2520                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2521                                 pos, ppos, count, written);
2522         }
2523 out:
2524         current->backing_dev_info = NULL;
2525         return written ? written : err;
2526 }
2527 EXPORT_SYMBOL(__generic_file_aio_write);
2528
2529 /**
2530  * generic_file_aio_write - write data to a file
2531  * @iocb:       IO state structure
2532  * @iov:        vector with data to write
2533  * @nr_segs:    number of segments in the vector
2534  * @pos:        position in file where to write
2535  *
2536  * This is a wrapper around __generic_file_aio_write() to be used by most
2537  * filesystems. It takes care of syncing the file in case of O_SYNC file
2538  * and acquires i_mutex as needed.
2539  */
2540 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2541                 unsigned long nr_segs, loff_t pos)
2542 {
2543         struct file *file = iocb->ki_filp;
2544         struct inode *inode = file->f_mapping->host;
2545         ssize_t ret;
2546
2547         BUG_ON(iocb->ki_pos != pos);
2548
2549         mutex_lock(&inode->i_mutex);
2550         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2551         mutex_unlock(&inode->i_mutex);
2552
2553         if (ret > 0 || ret == -EIOCBQUEUED) {
2554                 ssize_t err;
2555
2556                 err = generic_write_sync(file, pos, ret);
2557                 if (err < 0 && ret > 0)
2558                         ret = err;
2559         }
2560         return ret;
2561 }
2562 EXPORT_SYMBOL(generic_file_aio_write);
2563
2564 /**
2565  * try_to_release_page() - release old fs-specific metadata on a page
2566  *
2567  * @page: the page which the kernel is trying to free
2568  * @gfp_mask: memory allocation flags (and I/O mode)
2569  *
2570  * The address_space is to try to release any data against the page
2571  * (presumably at page->private).  If the release was successful, return `1'.
2572  * Otherwise return zero.
2573  *
2574  * This may also be called if PG_fscache is set on a page, indicating that the
2575  * page is known to the local caching routines.
2576  *
2577  * The @gfp_mask argument specifies whether I/O may be performed to release
2578  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2579  *
2580  */
2581 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2582 {
2583         struct address_space * const mapping = page->mapping;
2584
2585         BUG_ON(!PageLocked(page));
2586         if (PageWriteback(page))
2587                 return 0;
2588
2589         if (mapping && mapping->a_ops->releasepage)
2590                 return mapping->a_ops->releasepage(page, gfp_mask);
2591         return try_to_free_buffers(page);
2592 }
2593
2594 EXPORT_SYMBOL(try_to_release_page);