]> git.karo-electronics.de Git - karo-tx-linux.git/blob - mm/filemap.c
ARM: imx_v6_v7_defconfig: Select CONFIG_PROVE_LOCKING
[karo-tx-linux.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                 return error;
471
472         error = radix_tree_maybe_preload(gfp_mask & ~__GFP_HIGHMEM);
473         if (error) {
474                 mem_cgroup_uncharge_cache_page(page);
475                 return error;
476         }
477
478         page_cache_get(page);
479         page->mapping = mapping;
480         page->index = offset;
481
482         spin_lock_irq(&mapping->tree_lock);
483         error = radix_tree_insert(&mapping->page_tree, offset, page);
484         radix_tree_preload_end();
485         if (unlikely(error))
486                 goto err_insert;
487         mapping->nrpages++;
488         __inc_zone_page_state(page, NR_FILE_PAGES);
489         spin_unlock_irq(&mapping->tree_lock);
490         trace_mm_filemap_add_to_page_cache(page);
491         return 0;
492 err_insert:
493         page->mapping = NULL;
494         /* Leave page->index set: truncation relies upon it */
495         spin_unlock_irq(&mapping->tree_lock);
496         mem_cgroup_uncharge_cache_page(page);
497         page_cache_release(page);
498         return error;
499 }
500 EXPORT_SYMBOL(add_to_page_cache_locked);
501
502 int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
503                                 pgoff_t offset, gfp_t gfp_mask)
504 {
505         int ret;
506
507         ret = add_to_page_cache(page, mapping, offset, gfp_mask);
508         if (ret == 0)
509                 lru_cache_add_file(page);
510         return ret;
511 }
512 EXPORT_SYMBOL_GPL(add_to_page_cache_lru);
513
514 #ifdef CONFIG_NUMA
515 struct page *__page_cache_alloc(gfp_t gfp)
516 {
517         int n;
518         struct page *page;
519
520         if (cpuset_do_page_mem_spread()) {
521                 unsigned int cpuset_mems_cookie;
522                 do {
523                         cpuset_mems_cookie = get_mems_allowed();
524                         n = cpuset_mem_spread_node();
525                         page = alloc_pages_exact_node(n, gfp, 0);
526                 } while (!put_mems_allowed(cpuset_mems_cookie) && !page);
527
528                 return page;
529         }
530         return alloc_pages(gfp, 0);
531 }
532 EXPORT_SYMBOL(__page_cache_alloc);
533 #endif
534
535 /*
536  * In order to wait for pages to become available there must be
537  * waitqueues associated with pages. By using a hash table of
538  * waitqueues where the bucket discipline is to maintain all
539  * waiters on the same queue and wake all when any of the pages
540  * become available, and for the woken contexts to check to be
541  * sure the appropriate page became available, this saves space
542  * at a cost of "thundering herd" phenomena during rare hash
543  * collisions.
544  */
545 static wait_queue_head_t *page_waitqueue(struct page *page)
546 {
547         const struct zone *zone = page_zone(page);
548
549         return &zone->wait_table[hash_ptr(page, zone->wait_table_bits)];
550 }
551
552 static inline void wake_up_page(struct page *page, int bit)
553 {
554         __wake_up_bit(page_waitqueue(page), &page->flags, bit);
555 }
556
557 void wait_on_page_bit(struct page *page, int bit_nr)
558 {
559         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
560
561         if (test_bit(bit_nr, &page->flags))
562                 __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
563                                                         TASK_UNINTERRUPTIBLE);
564 }
565 EXPORT_SYMBOL(wait_on_page_bit);
566
567 int wait_on_page_bit_killable(struct page *page, int bit_nr)
568 {
569         DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
570
571         if (!test_bit(bit_nr, &page->flags))
572                 return 0;
573
574         return __wait_on_bit(page_waitqueue(page), &wait,
575                              sleep_on_page_killable, TASK_KILLABLE);
576 }
577
578 /**
579  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
580  * @page: Page defining the wait queue of interest
581  * @waiter: Waiter to add to the queue
582  *
583  * Add an arbitrary @waiter to the wait queue for the nominated @page.
584  */
585 void add_page_wait_queue(struct page *page, wait_queue_t *waiter)
586 {
587         wait_queue_head_t *q = page_waitqueue(page);
588         unsigned long flags;
589
590         spin_lock_irqsave(&q->lock, flags);
591         __add_wait_queue(q, waiter);
592         spin_unlock_irqrestore(&q->lock, flags);
593 }
594 EXPORT_SYMBOL_GPL(add_page_wait_queue);
595
596 /**
597  * unlock_page - unlock a locked page
598  * @page: the page
599  *
600  * Unlocks the page and wakes up sleepers in ___wait_on_page_locked().
601  * Also wakes sleepers in wait_on_page_writeback() because the wakeup
602  * mechananism between PageLocked pages and PageWriteback pages is shared.
603  * But that's OK - sleepers in wait_on_page_writeback() just go back to sleep.
604  *
605  * The mb is necessary to enforce ordering between the clear_bit and the read
606  * of the waitqueue (to avoid SMP races with a parallel wait_on_page_locked()).
607  */
608 void unlock_page(struct page *page)
609 {
610         VM_BUG_ON(!PageLocked(page));
611         clear_bit_unlock(PG_locked, &page->flags);
612         smp_mb__after_clear_bit();
613         wake_up_page(page, PG_locked);
614 }
615 EXPORT_SYMBOL(unlock_page);
616
617 /**
618  * end_page_writeback - end writeback against a page
619  * @page: the page
620  */
621 void end_page_writeback(struct page *page)
622 {
623         if (TestClearPageReclaim(page))
624                 rotate_reclaimable_page(page);
625
626         if (!test_clear_page_writeback(page))
627                 BUG();
628
629         smp_mb__after_clear_bit();
630         wake_up_page(page, PG_writeback);
631 }
632 EXPORT_SYMBOL(end_page_writeback);
633
634 /**
635  * __lock_page - get a lock on the page, assuming we need to sleep to get it
636  * @page: the page to lock
637  */
638 void __lock_page(struct page *page)
639 {
640         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
641
642         __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
643                                                         TASK_UNINTERRUPTIBLE);
644 }
645 EXPORT_SYMBOL(__lock_page);
646
647 int __lock_page_killable(struct page *page)
648 {
649         DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
650
651         return __wait_on_bit_lock(page_waitqueue(page), &wait,
652                                         sleep_on_page_killable, TASK_KILLABLE);
653 }
654 EXPORT_SYMBOL_GPL(__lock_page_killable);
655
656 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
657                          unsigned int flags)
658 {
659         if (flags & FAULT_FLAG_ALLOW_RETRY) {
660                 /*
661                  * CAUTION! In this case, mmap_sem is not released
662                  * even though return 0.
663                  */
664                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
665                         return 0;
666
667                 up_read(&mm->mmap_sem);
668                 if (flags & FAULT_FLAG_KILLABLE)
669                         wait_on_page_locked_killable(page);
670                 else
671                         wait_on_page_locked(page);
672                 return 0;
673         } else {
674                 if (flags & FAULT_FLAG_KILLABLE) {
675                         int ret;
676
677                         ret = __lock_page_killable(page);
678                         if (ret) {
679                                 up_read(&mm->mmap_sem);
680                                 return 0;
681                         }
682                 } else
683                         __lock_page(page);
684                 return 1;
685         }
686 }
687
688 /**
689  * find_get_page - find and get a page reference
690  * @mapping: the address_space to search
691  * @offset: the page index
692  *
693  * Is there a pagecache struct page at the given (mapping, offset) tuple?
694  * If yes, increment its refcount and return it; if no, return NULL.
695  */
696 struct page *find_get_page(struct address_space *mapping, pgoff_t offset)
697 {
698         void **pagep;
699         struct page *page;
700
701         rcu_read_lock();
702 repeat:
703         page = NULL;
704         pagep = radix_tree_lookup_slot(&mapping->page_tree, offset);
705         if (pagep) {
706                 page = radix_tree_deref_slot(pagep);
707                 if (unlikely(!page))
708                         goto out;
709                 if (radix_tree_exception(page)) {
710                         if (radix_tree_deref_retry(page))
711                                 goto repeat;
712                         /*
713                          * Otherwise, shmem/tmpfs must be storing a swap entry
714                          * here as an exceptional entry: so return it without
715                          * attempting to raise page count.
716                          */
717                         goto out;
718                 }
719                 if (!page_cache_get_speculative(page))
720                         goto repeat;
721
722                 /*
723                  * Has the page moved?
724                  * This is part of the lockless pagecache protocol. See
725                  * include/linux/pagemap.h for details.
726                  */
727                 if (unlikely(page != *pagep)) {
728                         page_cache_release(page);
729                         goto repeat;
730                 }
731         }
732 out:
733         rcu_read_unlock();
734
735         return page;
736 }
737 EXPORT_SYMBOL(find_get_page);
738
739 /**
740  * find_lock_page - locate, pin and lock a pagecache page
741  * @mapping: the address_space to search
742  * @offset: the page index
743  *
744  * Locates the desired pagecache page, locks it, increments its reference
745  * count and returns its address.
746  *
747  * Returns zero if the page was not present. find_lock_page() may sleep.
748  */
749 struct page *find_lock_page(struct address_space *mapping, pgoff_t offset)
750 {
751         struct page *page;
752
753 repeat:
754         page = find_get_page(mapping, offset);
755         if (page && !radix_tree_exception(page)) {
756                 lock_page(page);
757                 /* Has the page been truncated? */
758                 if (unlikely(page->mapping != mapping)) {
759                         unlock_page(page);
760                         page_cache_release(page);
761                         goto repeat;
762                 }
763                 VM_BUG_ON(page->index != offset);
764         }
765         return page;
766 }
767 EXPORT_SYMBOL(find_lock_page);
768
769 /**
770  * find_or_create_page - locate or add a pagecache page
771  * @mapping: the page's address_space
772  * @index: the page's index into the mapping
773  * @gfp_mask: page allocation mode
774  *
775  * Locates a page in the pagecache.  If the page is not present, a new page
776  * is allocated using @gfp_mask and is added to the pagecache and to the VM's
777  * LRU list.  The returned page is locked and has its reference count
778  * incremented.
779  *
780  * find_or_create_page() may sleep, even if @gfp_flags specifies an atomic
781  * allocation!
782  *
783  * find_or_create_page() returns the desired page's address, or zero on
784  * memory exhaustion.
785  */
786 struct page *find_or_create_page(struct address_space *mapping,
787                 pgoff_t index, gfp_t gfp_mask)
788 {
789         struct page *page;
790         int err;
791 repeat:
792         page = find_lock_page(mapping, index);
793         if (!page) {
794                 page = __page_cache_alloc(gfp_mask);
795                 if (!page)
796                         return NULL;
797                 /*
798                  * We want a regular kernel memory (not highmem or DMA etc)
799                  * allocation for the radix tree nodes, but we need to honour
800                  * the context-specific requirements the caller has asked for.
801                  * GFP_RECLAIM_MASK collects those requirements.
802                  */
803                 err = add_to_page_cache_lru(page, mapping, index,
804                         (gfp_mask & GFP_RECLAIM_MASK));
805                 if (unlikely(err)) {
806                         page_cache_release(page);
807                         page = NULL;
808                         if (err == -EEXIST)
809                                 goto repeat;
810                 }
811         }
812         return page;
813 }
814 EXPORT_SYMBOL(find_or_create_page);
815
816 /**
817  * find_get_pages - gang pagecache lookup
818  * @mapping:    The address_space to search
819  * @start:      The starting page index
820  * @nr_pages:   The maximum number of pages
821  * @pages:      Where the resulting pages are placed
822  *
823  * find_get_pages() will search for and return a group of up to
824  * @nr_pages pages in the mapping.  The pages are placed at @pages.
825  * find_get_pages() takes a reference against the returned pages.
826  *
827  * The search returns a group of mapping-contiguous pages with ascending
828  * indexes.  There may be holes in the indices due to not-present pages.
829  *
830  * find_get_pages() returns the number of pages which were found.
831  */
832 unsigned find_get_pages(struct address_space *mapping, pgoff_t start,
833                             unsigned int nr_pages, struct page **pages)
834 {
835         struct radix_tree_iter iter;
836         void **slot;
837         unsigned ret = 0;
838
839         if (unlikely(!nr_pages))
840                 return 0;
841
842         rcu_read_lock();
843 restart:
844         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
845                 struct page *page;
846 repeat:
847                 page = radix_tree_deref_slot(slot);
848                 if (unlikely(!page))
849                         continue;
850
851                 if (radix_tree_exception(page)) {
852                         if (radix_tree_deref_retry(page)) {
853                                 /*
854                                  * Transient condition which can only trigger
855                                  * when entry at index 0 moves out of or back
856                                  * to root: none yet gotten, safe to restart.
857                                  */
858                                 WARN_ON(iter.index);
859                                 goto restart;
860                         }
861                         /*
862                          * Otherwise, shmem/tmpfs must be storing a swap entry
863                          * here as an exceptional entry: so skip over it -
864                          * we only reach this from invalidate_mapping_pages().
865                          */
866                         continue;
867                 }
868
869                 if (!page_cache_get_speculative(page))
870                         goto repeat;
871
872                 /* Has the page moved? */
873                 if (unlikely(page != *slot)) {
874                         page_cache_release(page);
875                         goto repeat;
876                 }
877
878                 pages[ret] = page;
879                 if (++ret == nr_pages)
880                         break;
881         }
882
883         rcu_read_unlock();
884         return ret;
885 }
886
887 /**
888  * find_get_pages_contig - gang contiguous pagecache lookup
889  * @mapping:    The address_space to search
890  * @index:      The starting page index
891  * @nr_pages:   The maximum number of pages
892  * @pages:      Where the resulting pages are placed
893  *
894  * find_get_pages_contig() works exactly like find_get_pages(), except
895  * that the returned number of pages are guaranteed to be contiguous.
896  *
897  * find_get_pages_contig() returns the number of pages which were found.
898  */
899 unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
900                                unsigned int nr_pages, struct page **pages)
901 {
902         struct radix_tree_iter iter;
903         void **slot;
904         unsigned int ret = 0;
905
906         if (unlikely(!nr_pages))
907                 return 0;
908
909         rcu_read_lock();
910 restart:
911         radix_tree_for_each_contig(slot, &mapping->page_tree, &iter, index) {
912                 struct page *page;
913 repeat:
914                 page = radix_tree_deref_slot(slot);
915                 /* The hole, there no reason to continue */
916                 if (unlikely(!page))
917                         break;
918
919                 if (radix_tree_exception(page)) {
920                         if (radix_tree_deref_retry(page)) {
921                                 /*
922                                  * Transient condition which can only trigger
923                                  * when entry at index 0 moves out of or back
924                                  * to root: none yet gotten, safe to restart.
925                                  */
926                                 goto restart;
927                         }
928                         /*
929                          * Otherwise, shmem/tmpfs must be storing a swap entry
930                          * here as an exceptional entry: so stop looking for
931                          * contiguous pages.
932                          */
933                         break;
934                 }
935
936                 if (!page_cache_get_speculative(page))
937                         goto repeat;
938
939                 /* Has the page moved? */
940                 if (unlikely(page != *slot)) {
941                         page_cache_release(page);
942                         goto repeat;
943                 }
944
945                 /*
946                  * must check mapping and index after taking the ref.
947                  * otherwise we can get both false positives and false
948                  * negatives, which is just confusing to the caller.
949                  */
950                 if (page->mapping == NULL || page->index != iter.index) {
951                         page_cache_release(page);
952                         break;
953                 }
954
955                 pages[ret] = page;
956                 if (++ret == nr_pages)
957                         break;
958         }
959         rcu_read_unlock();
960         return ret;
961 }
962 EXPORT_SYMBOL(find_get_pages_contig);
963
964 /**
965  * find_get_pages_tag - find and return pages that match @tag
966  * @mapping:    the address_space to search
967  * @index:      the starting page index
968  * @tag:        the tag index
969  * @nr_pages:   the maximum number of pages
970  * @pages:      where the resulting pages are placed
971  *
972  * Like find_get_pages, except we only return pages which are tagged with
973  * @tag.   We update @index to index the next page for the traversal.
974  */
975 unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
976                         int tag, unsigned int nr_pages, struct page **pages)
977 {
978         struct radix_tree_iter iter;
979         void **slot;
980         unsigned ret = 0;
981
982         if (unlikely(!nr_pages))
983                 return 0;
984
985         rcu_read_lock();
986 restart:
987         radix_tree_for_each_tagged(slot, &mapping->page_tree,
988                                    &iter, *index, tag) {
989                 struct page *page;
990 repeat:
991                 page = radix_tree_deref_slot(slot);
992                 if (unlikely(!page))
993                         continue;
994
995                 if (radix_tree_exception(page)) {
996                         if (radix_tree_deref_retry(page)) {
997                                 /*
998                                  * Transient condition which can only trigger
999                                  * when entry at index 0 moves out of or back
1000                                  * to root: none yet gotten, safe to restart.
1001                                  */
1002                                 goto restart;
1003                         }
1004                         /*
1005                          * This function is never used on a shmem/tmpfs
1006                          * mapping, so a swap entry won't be found here.
1007                          */
1008                         BUG();
1009                 }
1010
1011                 if (!page_cache_get_speculative(page))
1012                         goto repeat;
1013
1014                 /* Has the page moved? */
1015                 if (unlikely(page != *slot)) {
1016                         page_cache_release(page);
1017                         goto repeat;
1018                 }
1019
1020                 pages[ret] = page;
1021                 if (++ret == nr_pages)
1022                         break;
1023         }
1024
1025         rcu_read_unlock();
1026
1027         if (ret)
1028                 *index = pages[ret - 1]->index + 1;
1029
1030         return ret;
1031 }
1032 EXPORT_SYMBOL(find_get_pages_tag);
1033
1034 /**
1035  * grab_cache_page_nowait - returns locked page at given index in given cache
1036  * @mapping: target address_space
1037  * @index: the page index
1038  *
1039  * Same as grab_cache_page(), but do not wait if the page is unavailable.
1040  * This is intended for speculative data generators, where the data can
1041  * be regenerated if the page couldn't be grabbed.  This routine should
1042  * be safe to call while holding the lock for another page.
1043  *
1044  * Clear __GFP_FS when allocating the page to avoid recursion into the fs
1045  * and deadlock against the caller's locked page.
1046  */
1047 struct page *
1048 grab_cache_page_nowait(struct address_space *mapping, pgoff_t index)
1049 {
1050         struct page *page = find_get_page(mapping, index);
1051
1052         if (page) {
1053                 if (trylock_page(page))
1054                         return page;
1055                 page_cache_release(page);
1056                 return NULL;
1057         }
1058         page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
1059         if (page && add_to_page_cache_lru(page, mapping, index, GFP_NOFS)) {
1060                 page_cache_release(page);
1061                 page = NULL;
1062         }
1063         return page;
1064 }
1065 EXPORT_SYMBOL(grab_cache_page_nowait);
1066
1067 /*
1068  * CD/DVDs are error prone. When a medium error occurs, the driver may fail
1069  * a _large_ part of the i/o request. Imagine the worst scenario:
1070  *
1071  *      ---R__________________________________________B__________
1072  *         ^ reading here                             ^ bad block(assume 4k)
1073  *
1074  * read(R) => miss => readahead(R...B) => media error => frustrating retries
1075  * => failing the whole request => read(R) => read(R+1) =>
1076  * readahead(R+1...B+1) => bang => read(R+2) => read(R+3) =>
1077  * readahead(R+3...B+2) => bang => read(R+3) => read(R+4) =>
1078  * readahead(R+4...B+3) => bang => read(R+4) => read(R+5) => ......
1079  *
1080  * It is going insane. Fix it by quickly scaling down the readahead size.
1081  */
1082 static void shrink_readahead_size_eio(struct file *filp,
1083                                         struct file_ra_state *ra)
1084 {
1085         ra->ra_pages /= 4;
1086 }
1087
1088 /**
1089  * do_generic_file_read - generic file read routine
1090  * @filp:       the file to read
1091  * @ppos:       current file position
1092  * @desc:       read_descriptor
1093  * @actor:      read method
1094  *
1095  * This is a generic file read routine, and uses the
1096  * mapping->a_ops->readpage() function for the actual low-level stuff.
1097  *
1098  * This is really ugly. But the goto's actually try to clarify some
1099  * of the logic when it comes to error handling etc.
1100  */
1101 static void do_generic_file_read(struct file *filp, loff_t *ppos,
1102                 read_descriptor_t *desc, read_actor_t actor)
1103 {
1104         struct address_space *mapping = filp->f_mapping;
1105         struct inode *inode = mapping->host;
1106         struct file_ra_state *ra = &filp->f_ra;
1107         pgoff_t index;
1108         pgoff_t last_index;
1109         pgoff_t prev_index;
1110         unsigned long offset;      /* offset into pagecache page */
1111         unsigned int prev_offset;
1112         int error;
1113
1114         index = *ppos >> PAGE_CACHE_SHIFT;
1115         prev_index = ra->prev_pos >> PAGE_CACHE_SHIFT;
1116         prev_offset = ra->prev_pos & (PAGE_CACHE_SIZE-1);
1117         last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
1118         offset = *ppos & ~PAGE_CACHE_MASK;
1119
1120         for (;;) {
1121                 struct page *page;
1122                 pgoff_t end_index;
1123                 loff_t isize;
1124                 unsigned long nr, ret;
1125
1126                 cond_resched();
1127 find_page:
1128                 page = find_get_page(mapping, index);
1129                 if (!page) {
1130                         page_cache_sync_readahead(mapping,
1131                                         ra, filp,
1132                                         index, last_index - index);
1133                         page = find_get_page(mapping, index);
1134                         if (unlikely(page == NULL))
1135                                 goto no_cached_page;
1136                 }
1137                 if (PageReadahead(page)) {
1138                         page_cache_async_readahead(mapping,
1139                                         ra, filp, page,
1140                                         index, last_index - index);
1141                 }
1142                 if (!PageUptodate(page)) {
1143                         if (inode->i_blkbits == PAGE_CACHE_SHIFT ||
1144                                         !mapping->a_ops->is_partially_uptodate)
1145                                 goto page_not_up_to_date;
1146                         if (!trylock_page(page))
1147                                 goto page_not_up_to_date;
1148                         /* Did it get truncated before we got the lock? */
1149                         if (!page->mapping)
1150                                 goto page_not_up_to_date_locked;
1151                         if (!mapping->a_ops->is_partially_uptodate(page,
1152                                                                 desc, offset))
1153                                 goto page_not_up_to_date_locked;
1154                         unlock_page(page);
1155                 }
1156 page_ok:
1157                 /*
1158                  * i_size must be checked after we know the page is Uptodate.
1159                  *
1160                  * Checking i_size after the check allows us to calculate
1161                  * the correct value for "nr", which means the zero-filled
1162                  * part of the page is not copied back to userspace (unless
1163                  * another truncate extends the file - this is desired though).
1164                  */
1165
1166                 isize = i_size_read(inode);
1167                 end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
1168                 if (unlikely(!isize || index > end_index)) {
1169                         page_cache_release(page);
1170                         goto out;
1171                 }
1172
1173                 /* nr is the maximum number of bytes to copy from this page */
1174                 nr = PAGE_CACHE_SIZE;
1175                 if (index == end_index) {
1176                         nr = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
1177                         if (nr <= offset) {
1178                                 page_cache_release(page);
1179                                 goto out;
1180                         }
1181                 }
1182                 nr = nr - offset;
1183
1184                 /* If users can be writing to this page using arbitrary
1185                  * virtual addresses, take care about potential aliasing
1186                  * before reading the page on the kernel side.
1187                  */
1188                 if (mapping_writably_mapped(mapping))
1189                         flush_dcache_page(page);
1190
1191                 /*
1192                  * When a sequential read accesses a page several times,
1193                  * only mark it as accessed the first time.
1194                  */
1195                 if (prev_index != index || offset != prev_offset)
1196                         mark_page_accessed(page);
1197                 prev_index = index;
1198
1199                 /*
1200                  * Ok, we have the page, and it's up-to-date, so
1201                  * now we can copy it to user space...
1202                  *
1203                  * The actor routine returns how many bytes were actually used..
1204                  * NOTE! This may not be the same as how much of a user buffer
1205                  * we filled up (we may be padding etc), so we can only update
1206                  * "pos" here (the actor routine has to update the user buffer
1207                  * pointers and the remaining count).
1208                  */
1209                 ret = actor(desc, page, offset, nr);
1210                 offset += ret;
1211                 index += offset >> PAGE_CACHE_SHIFT;
1212                 offset &= ~PAGE_CACHE_MASK;
1213                 prev_offset = offset;
1214
1215                 page_cache_release(page);
1216                 if (ret == nr && desc->count)
1217                         continue;
1218                 goto out;
1219
1220 page_not_up_to_date:
1221                 /* Get exclusive access to the page ... */
1222                 error = lock_page_killable(page);
1223                 if (unlikely(error))
1224                         goto readpage_error;
1225
1226 page_not_up_to_date_locked:
1227                 /* Did it get truncated before we got the lock? */
1228                 if (!page->mapping) {
1229                         unlock_page(page);
1230                         page_cache_release(page);
1231                         continue;
1232                 }
1233
1234                 /* Did somebody else fill it already? */
1235                 if (PageUptodate(page)) {
1236                         unlock_page(page);
1237                         goto page_ok;
1238                 }
1239
1240 readpage:
1241                 /*
1242                  * A previous I/O error may have been due to temporary
1243                  * failures, eg. multipath errors.
1244                  * PG_error will be set again if readpage fails.
1245                  */
1246                 ClearPageError(page);
1247                 /* Start the actual read. The read will unlock the page. */
1248                 error = mapping->a_ops->readpage(filp, page);
1249
1250                 if (unlikely(error)) {
1251                         if (error == AOP_TRUNCATED_PAGE) {
1252                                 page_cache_release(page);
1253                                 goto find_page;
1254                         }
1255                         goto readpage_error;
1256                 }
1257
1258                 if (!PageUptodate(page)) {
1259                         error = lock_page_killable(page);
1260                         if (unlikely(error))
1261                                 goto readpage_error;
1262                         if (!PageUptodate(page)) {
1263                                 if (page->mapping == NULL) {
1264                                         /*
1265                                          * invalidate_mapping_pages got it
1266                                          */
1267                                         unlock_page(page);
1268                                         page_cache_release(page);
1269                                         goto find_page;
1270                                 }
1271                                 unlock_page(page);
1272                                 shrink_readahead_size_eio(filp, ra);
1273                                 error = -EIO;
1274                                 goto readpage_error;
1275                         }
1276                         unlock_page(page);
1277                 }
1278
1279                 goto page_ok;
1280
1281 readpage_error:
1282                 /* UHHUH! A synchronous read error occurred. Report it */
1283                 desc->error = error;
1284                 page_cache_release(page);
1285                 goto out;
1286
1287 no_cached_page:
1288                 /*
1289                  * Ok, it wasn't cached, so we need to create a new
1290                  * page..
1291                  */
1292                 page = page_cache_alloc_cold(mapping);
1293                 if (!page) {
1294                         desc->error = -ENOMEM;
1295                         goto out;
1296                 }
1297                 error = add_to_page_cache_lru(page, mapping,
1298                                                 index, GFP_KERNEL);
1299                 if (error) {
1300                         page_cache_release(page);
1301                         if (error == -EEXIST)
1302                                 goto find_page;
1303                         desc->error = error;
1304                         goto out;
1305                 }
1306                 goto readpage;
1307         }
1308
1309 out:
1310         ra->prev_pos = prev_index;
1311         ra->prev_pos <<= PAGE_CACHE_SHIFT;
1312         ra->prev_pos |= prev_offset;
1313
1314         *ppos = ((loff_t)index << PAGE_CACHE_SHIFT) + offset;
1315         file_accessed(filp);
1316 }
1317
1318 int file_read_actor(read_descriptor_t *desc, struct page *page,
1319                         unsigned long offset, unsigned long size)
1320 {
1321         char *kaddr;
1322         unsigned long left, count = desc->count;
1323
1324         if (size > count)
1325                 size = count;
1326
1327         /*
1328          * Faults on the destination of a read are common, so do it before
1329          * taking the kmap.
1330          */
1331         if (!fault_in_pages_writeable(desc->arg.buf, size)) {
1332                 kaddr = kmap_atomic(page);
1333                 left = __copy_to_user_inatomic(desc->arg.buf,
1334                                                 kaddr + offset, size);
1335                 kunmap_atomic(kaddr);
1336                 if (left == 0)
1337                         goto success;
1338         }
1339
1340         /* Do it the slow way */
1341         kaddr = kmap(page);
1342         left = __copy_to_user(desc->arg.buf, kaddr + offset, size);
1343         kunmap(page);
1344
1345         if (left) {
1346                 size -= left;
1347                 desc->error = -EFAULT;
1348         }
1349 success:
1350         desc->count = count - size;
1351         desc->written += size;
1352         desc->arg.buf += size;
1353         return size;
1354 }
1355
1356 /*
1357  * Performs necessary checks before doing a write
1358  * @iov:        io vector request
1359  * @nr_segs:    number of segments in the iovec
1360  * @count:      number of bytes to write
1361  * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
1362  *
1363  * Adjust number of segments and amount of bytes to write (nr_segs should be
1364  * properly initialized first). Returns appropriate error code that caller
1365  * should return or zero in case that write should be allowed.
1366  */
1367 int generic_segment_checks(const struct iovec *iov,
1368                         unsigned long *nr_segs, size_t *count, int access_flags)
1369 {
1370         unsigned long   seg;
1371         size_t cnt = 0;
1372         for (seg = 0; seg < *nr_segs; seg++) {
1373                 const struct iovec *iv = &iov[seg];
1374
1375                 /*
1376                  * If any segment has a negative length, or the cumulative
1377                  * length ever wraps negative then return -EINVAL.
1378                  */
1379                 cnt += iv->iov_len;
1380                 if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
1381                         return -EINVAL;
1382                 if (access_ok(access_flags, iv->iov_base, iv->iov_len))
1383                         continue;
1384                 if (seg == 0)
1385                         return -EFAULT;
1386                 *nr_segs = seg;
1387                 cnt -= iv->iov_len;     /* This segment is no good */
1388                 break;
1389         }
1390         *count = cnt;
1391         return 0;
1392 }
1393 EXPORT_SYMBOL(generic_segment_checks);
1394
1395 /**
1396  * generic_file_aio_read - generic filesystem read routine
1397  * @iocb:       kernel I/O control block
1398  * @iov:        io vector request
1399  * @nr_segs:    number of segments in the iovec
1400  * @pos:        current file position
1401  *
1402  * This is the "read()" routine for all filesystems
1403  * that can use the page cache directly.
1404  */
1405 ssize_t
1406 generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
1407                 unsigned long nr_segs, loff_t pos)
1408 {
1409         struct file *filp = iocb->ki_filp;
1410         ssize_t retval;
1411         unsigned long seg = 0;
1412         size_t count;
1413         loff_t *ppos = &iocb->ki_pos;
1414
1415         count = 0;
1416         retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
1417         if (retval)
1418                 return retval;
1419
1420         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
1421         if (filp->f_flags & O_DIRECT) {
1422                 loff_t size;
1423                 struct address_space *mapping;
1424                 struct inode *inode;
1425
1426                 mapping = filp->f_mapping;
1427                 inode = mapping->host;
1428                 if (!count)
1429                         goto out; /* skip atime */
1430                 size = i_size_read(inode);
1431                 if (pos < size) {
1432                         retval = filemap_write_and_wait_range(mapping, pos,
1433                                         pos + iov_length(iov, nr_segs) - 1);
1434                         if (!retval) {
1435                                 retval = mapping->a_ops->direct_IO(READ, iocb,
1436                                                         iov, pos, nr_segs);
1437                         }
1438                         if (retval > 0) {
1439                                 *ppos = pos + retval;
1440                                 count -= retval;
1441                         }
1442
1443                         /*
1444                          * Btrfs can have a short DIO read if we encounter
1445                          * compressed extents, so if there was an error, or if
1446                          * we've already read everything we wanted to, or if
1447                          * there was a short read because we hit EOF, go ahead
1448                          * and return.  Otherwise fallthrough to buffered io for
1449                          * the rest of the read.
1450                          */
1451                         if (retval < 0 || !count || *ppos >= size) {
1452                                 file_accessed(filp);
1453                                 goto out;
1454                         }
1455                 }
1456         }
1457
1458         count = retval;
1459         for (seg = 0; seg < nr_segs; seg++) {
1460                 read_descriptor_t desc;
1461                 loff_t offset = 0;
1462
1463                 /*
1464                  * If we did a short DIO read we need to skip the section of the
1465                  * iov that we've already read data into.
1466                  */
1467                 if (count) {
1468                         if (count > iov[seg].iov_len) {
1469                                 count -= iov[seg].iov_len;
1470                                 continue;
1471                         }
1472                         offset = count;
1473                         count = 0;
1474                 }
1475
1476                 desc.written = 0;
1477                 desc.arg.buf = iov[seg].iov_base + offset;
1478                 desc.count = iov[seg].iov_len - offset;
1479                 if (desc.count == 0)
1480                         continue;
1481                 desc.error = 0;
1482                 do_generic_file_read(filp, ppos, &desc, file_read_actor);
1483                 retval += desc.written;
1484                 if (desc.error) {
1485                         retval = retval ?: desc.error;
1486                         break;
1487                 }
1488                 if (desc.count > 0)
1489                         break;
1490         }
1491 out:
1492         return retval;
1493 }
1494 EXPORT_SYMBOL(generic_file_aio_read);
1495
1496 #ifdef CONFIG_MMU
1497 /**
1498  * page_cache_read - adds requested page to the page cache if not already there
1499  * @file:       file to read
1500  * @offset:     page index
1501  *
1502  * This adds the requested page to the page cache if it isn't already there,
1503  * and schedules an I/O to read in its contents from disk.
1504  */
1505 static int page_cache_read(struct file *file, pgoff_t offset)
1506 {
1507         struct address_space *mapping = file->f_mapping;
1508         struct page *page; 
1509         int ret;
1510
1511         do {
1512                 page = page_cache_alloc_cold(mapping);
1513                 if (!page)
1514                         return -ENOMEM;
1515
1516                 ret = add_to_page_cache_lru(page, mapping, offset, GFP_KERNEL);
1517                 if (ret == 0)
1518                         ret = mapping->a_ops->readpage(file, page);
1519                 else if (ret == -EEXIST)
1520                         ret = 0; /* losing race to add is OK */
1521
1522                 page_cache_release(page);
1523
1524         } while (ret == AOP_TRUNCATED_PAGE);
1525                 
1526         return ret;
1527 }
1528
1529 #define MMAP_LOTSAMISS  (100)
1530
1531 /*
1532  * Synchronous readahead happens when we don't even find
1533  * a page in the page cache at all.
1534  */
1535 static void do_sync_mmap_readahead(struct vm_area_struct *vma,
1536                                    struct file_ra_state *ra,
1537                                    struct file *file,
1538                                    pgoff_t offset)
1539 {
1540         unsigned long ra_pages;
1541         struct address_space *mapping = file->f_mapping;
1542
1543         /* If we don't want any read-ahead, don't bother */
1544         if (vma->vm_flags & VM_RAND_READ)
1545                 return;
1546         if (!ra->ra_pages)
1547                 return;
1548
1549         if (vma->vm_flags & VM_SEQ_READ) {
1550                 page_cache_sync_readahead(mapping, ra, file, offset,
1551                                           ra->ra_pages);
1552                 return;
1553         }
1554
1555         /* Avoid banging the cache line if not needed */
1556         if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
1557                 ra->mmap_miss++;
1558
1559         /*
1560          * Do we miss much more than hit in this file? If so,
1561          * stop bothering with read-ahead. It will only hurt.
1562          */
1563         if (ra->mmap_miss > MMAP_LOTSAMISS)
1564                 return;
1565
1566         /*
1567          * mmap read-around
1568          */
1569         ra_pages = max_sane_readahead(ra->ra_pages);
1570         ra->start = max_t(long, 0, offset - ra_pages / 2);
1571         ra->size = ra_pages;
1572         ra->async_size = ra_pages / 4;
1573         ra_submit(ra, mapping, file);
1574 }
1575
1576 /*
1577  * Asynchronous readahead happens when we find the page and PG_readahead,
1578  * so we want to possibly extend the readahead further..
1579  */
1580 static void do_async_mmap_readahead(struct vm_area_struct *vma,
1581                                     struct file_ra_state *ra,
1582                                     struct file *file,
1583                                     struct page *page,
1584                                     pgoff_t offset)
1585 {
1586         struct address_space *mapping = file->f_mapping;
1587
1588         /* If we don't want any read-ahead, don't bother */
1589         if (vma->vm_flags & VM_RAND_READ)
1590                 return;
1591         if (ra->mmap_miss > 0)
1592                 ra->mmap_miss--;
1593         if (PageReadahead(page))
1594                 page_cache_async_readahead(mapping, ra, file,
1595                                            page, offset, ra->ra_pages);
1596 }
1597
1598 /**
1599  * filemap_fault - read in file data for page fault handling
1600  * @vma:        vma in which the fault was taken
1601  * @vmf:        struct vm_fault containing details of the fault
1602  *
1603  * filemap_fault() is invoked via the vma operations vector for a
1604  * mapped memory region to read in file data during a page fault.
1605  *
1606  * The goto's are kind of ugly, but this streamlines the normal case of having
1607  * it in the page cache, and handles the special cases reasonably without
1608  * having a lot of duplicated code.
1609  */
1610 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1611 {
1612         int error;
1613         struct file *file = vma->vm_file;
1614         struct address_space *mapping = file->f_mapping;
1615         struct file_ra_state *ra = &file->f_ra;
1616         struct inode *inode = mapping->host;
1617         pgoff_t offset = vmf->pgoff;
1618         struct page *page;
1619         bool memcg_oom;
1620         pgoff_t size;
1621         int ret = 0;
1622
1623         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1624         if (offset >= size)
1625                 return VM_FAULT_SIGBUS;
1626
1627         /*
1628          * Do we have something in the page cache already?  Either
1629          * way, try readahead, but disable the memcg OOM killer for it
1630          * as readahead is optional and no errors are propagated up
1631          * the fault stack.  The OOM killer is enabled while trying to
1632          * instantiate the faulting page individually below.
1633          */
1634         page = find_get_page(mapping, offset);
1635         if (likely(page) && !(vmf->flags & FAULT_FLAG_TRIED)) {
1636                 /*
1637                  * We found the page, so try async readahead before
1638                  * waiting for the lock.
1639                  */
1640                 memcg_oom = mem_cgroup_toggle_oom(false);
1641                 do_async_mmap_readahead(vma, ra, file, page, offset);
1642                 mem_cgroup_toggle_oom(memcg_oom);
1643         } else if (!page) {
1644                 /* No page in the page cache at all */
1645                 memcg_oom = mem_cgroup_toggle_oom(false);
1646                 do_sync_mmap_readahead(vma, ra, file, offset);
1647                 mem_cgroup_toggle_oom(memcg_oom);
1648                 count_vm_event(PGMAJFAULT);
1649                 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
1650                 ret = VM_FAULT_MAJOR;
1651 retry_find:
1652                 page = find_get_page(mapping, offset);
1653                 if (!page)
1654                         goto no_cached_page;
1655         }
1656
1657         if (!lock_page_or_retry(page, vma->vm_mm, vmf->flags)) {
1658                 page_cache_release(page);
1659                 return ret | VM_FAULT_RETRY;
1660         }
1661
1662         /* Did it get truncated? */
1663         if (unlikely(page->mapping != mapping)) {
1664                 unlock_page(page);
1665                 put_page(page);
1666                 goto retry_find;
1667         }
1668         VM_BUG_ON(page->index != offset);
1669
1670         /*
1671          * We have a locked page in the page cache, now we need to check
1672          * that it's up-to-date. If not, it is going to be due to an error.
1673          */
1674         if (unlikely(!PageUptodate(page)))
1675                 goto page_not_uptodate;
1676
1677         /*
1678          * Found the page and have a reference on it.
1679          * We must recheck i_size under page lock.
1680          */
1681         size = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1682         if (unlikely(offset >= size)) {
1683                 unlock_page(page);
1684                 page_cache_release(page);
1685                 return VM_FAULT_SIGBUS;
1686         }
1687
1688         vmf->page = page;
1689         return ret | VM_FAULT_LOCKED;
1690
1691 no_cached_page:
1692         /*
1693          * We're only likely to ever get here if MADV_RANDOM is in
1694          * effect.
1695          */
1696         error = page_cache_read(file, offset);
1697
1698         /*
1699          * The page we want has now been added to the page cache.
1700          * In the unlikely event that someone removed it in the
1701          * meantime, we'll just come back here and read it again.
1702          */
1703         if (error >= 0)
1704                 goto retry_find;
1705
1706         /*
1707          * An error return from page_cache_read can result if the
1708          * system is low on memory, or a problem occurs while trying
1709          * to schedule I/O.
1710          */
1711         if (error == -ENOMEM)
1712                 return VM_FAULT_OOM;
1713         return VM_FAULT_SIGBUS;
1714
1715 page_not_uptodate:
1716         /*
1717          * Umm, take care of errors if the page isn't up-to-date.
1718          * Try to re-read it _once_. We do this synchronously,
1719          * because there really aren't any performance issues here
1720          * and we need to check for errors.
1721          */
1722         ClearPageError(page);
1723         error = mapping->a_ops->readpage(file, page);
1724         if (!error) {
1725                 wait_on_page_locked(page);
1726                 if (!PageUptodate(page))
1727                         error = -EIO;
1728         }
1729         page_cache_release(page);
1730
1731         if (!error || error == AOP_TRUNCATED_PAGE)
1732                 goto retry_find;
1733
1734         /* Things didn't work out. Return zero to tell the mm layer so. */
1735         shrink_readahead_size_eio(file, ra);
1736         return VM_FAULT_SIGBUS;
1737 }
1738 EXPORT_SYMBOL(filemap_fault);
1739
1740 int filemap_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1741 {
1742         struct page *page = vmf->page;
1743         struct inode *inode = file_inode(vma->vm_file);
1744         int ret = VM_FAULT_LOCKED;
1745
1746         sb_start_pagefault(inode->i_sb);
1747         file_update_time(vma->vm_file);
1748         lock_page(page);
1749         if (page->mapping != inode->i_mapping) {
1750                 unlock_page(page);
1751                 ret = VM_FAULT_NOPAGE;
1752                 goto out;
1753         }
1754         /*
1755          * We mark the page dirty already here so that when freeze is in
1756          * progress, we are guaranteed that writeback during freezing will
1757          * see the dirty page and writeprotect it again.
1758          */
1759         set_page_dirty(page);
1760         wait_for_stable_page(page);
1761 out:
1762         sb_end_pagefault(inode->i_sb);
1763         return ret;
1764 }
1765 EXPORT_SYMBOL(filemap_page_mkwrite);
1766
1767 const struct vm_operations_struct generic_file_vm_ops = {
1768         .fault          = filemap_fault,
1769         .page_mkwrite   = filemap_page_mkwrite,
1770         .remap_pages    = generic_file_remap_pages,
1771 };
1772
1773 /* This is used for a general mmap of a disk file */
1774
1775 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1776 {
1777         struct address_space *mapping = file->f_mapping;
1778
1779         if (!mapping->a_ops->readpage)
1780                 return -ENOEXEC;
1781         file_accessed(file);
1782         vma->vm_ops = &generic_file_vm_ops;
1783         return 0;
1784 }
1785
1786 /*
1787  * This is for filesystems which do not implement ->writepage.
1788  */
1789 int generic_file_readonly_mmap(struct file *file, struct vm_area_struct *vma)
1790 {
1791         if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_MAYWRITE))
1792                 return -EINVAL;
1793         return generic_file_mmap(file, vma);
1794 }
1795 #else
1796 int generic_file_mmap(struct file * file, struct vm_area_struct * vma)
1797 {
1798         return -ENOSYS;
1799 }
1800 int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
1801 {
1802         return -ENOSYS;
1803 }
1804 #endif /* CONFIG_MMU */
1805
1806 EXPORT_SYMBOL(generic_file_mmap);
1807 EXPORT_SYMBOL(generic_file_readonly_mmap);
1808
1809 static struct page *__read_cache_page(struct address_space *mapping,
1810                                 pgoff_t index,
1811                                 int (*filler)(void *, struct page *),
1812                                 void *data,
1813                                 gfp_t gfp)
1814 {
1815         struct page *page;
1816         int err;
1817 repeat:
1818         page = find_get_page(mapping, index);
1819         if (!page) {
1820                 page = __page_cache_alloc(gfp | __GFP_COLD);
1821                 if (!page)
1822                         return ERR_PTR(-ENOMEM);
1823                 err = add_to_page_cache_lru(page, mapping, index, gfp);
1824                 if (unlikely(err)) {
1825                         page_cache_release(page);
1826                         if (err == -EEXIST)
1827                                 goto repeat;
1828                         /* Presumably ENOMEM for radix tree node */
1829                         return ERR_PTR(err);
1830                 }
1831                 err = filler(data, page);
1832                 if (err < 0) {
1833                         page_cache_release(page);
1834                         page = ERR_PTR(err);
1835                 }
1836         }
1837         return page;
1838 }
1839
1840 static struct page *do_read_cache_page(struct address_space *mapping,
1841                                 pgoff_t index,
1842                                 int (*filler)(void *, struct page *),
1843                                 void *data,
1844                                 gfp_t gfp)
1845
1846 {
1847         struct page *page;
1848         int err;
1849
1850 retry:
1851         page = __read_cache_page(mapping, index, filler, data, gfp);
1852         if (IS_ERR(page))
1853                 return page;
1854         if (PageUptodate(page))
1855                 goto out;
1856
1857         lock_page(page);
1858         if (!page->mapping) {
1859                 unlock_page(page);
1860                 page_cache_release(page);
1861                 goto retry;
1862         }
1863         if (PageUptodate(page)) {
1864                 unlock_page(page);
1865                 goto out;
1866         }
1867         err = filler(data, page);
1868         if (err < 0) {
1869                 page_cache_release(page);
1870                 return ERR_PTR(err);
1871         }
1872 out:
1873         mark_page_accessed(page);
1874         return page;
1875 }
1876
1877 /**
1878  * read_cache_page_async - read into page cache, fill it if needed
1879  * @mapping:    the page's address_space
1880  * @index:      the page index
1881  * @filler:     function to perform the read
1882  * @data:       first arg to filler(data, page) function, often left as NULL
1883  *
1884  * Same as read_cache_page, but don't wait for page to become unlocked
1885  * after submitting it to the filler.
1886  *
1887  * Read into the page cache. If a page already exists, and PageUptodate() is
1888  * not set, try to fill the page but don't wait for it to become unlocked.
1889  *
1890  * If the page does not get brought uptodate, return -EIO.
1891  */
1892 struct page *read_cache_page_async(struct address_space *mapping,
1893                                 pgoff_t index,
1894                                 int (*filler)(void *, struct page *),
1895                                 void *data)
1896 {
1897         return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
1898 }
1899 EXPORT_SYMBOL(read_cache_page_async);
1900
1901 static struct page *wait_on_page_read(struct page *page)
1902 {
1903         if (!IS_ERR(page)) {
1904                 wait_on_page_locked(page);
1905                 if (!PageUptodate(page)) {
1906                         page_cache_release(page);
1907                         page = ERR_PTR(-EIO);
1908                 }
1909         }
1910         return page;
1911 }
1912
1913 /**
1914  * read_cache_page_gfp - read into page cache, using specified page allocation flags.
1915  * @mapping:    the page's address_space
1916  * @index:      the page index
1917  * @gfp:        the page allocator flags to use if allocating
1918  *
1919  * This is the same as "read_mapping_page(mapping, index, NULL)", but with
1920  * any new page allocations done using the specified allocation flags.
1921  *
1922  * If the page does not get brought uptodate, return -EIO.
1923  */
1924 struct page *read_cache_page_gfp(struct address_space *mapping,
1925                                 pgoff_t index,
1926                                 gfp_t gfp)
1927 {
1928         filler_t *filler = (filler_t *)mapping->a_ops->readpage;
1929
1930         return wait_on_page_read(do_read_cache_page(mapping, index, filler, NULL, gfp));
1931 }
1932 EXPORT_SYMBOL(read_cache_page_gfp);
1933
1934 /**
1935  * read_cache_page - read into page cache, fill it if needed
1936  * @mapping:    the page's address_space
1937  * @index:      the page index
1938  * @filler:     function to perform the read
1939  * @data:       first arg to filler(data, page) function, often left as NULL
1940  *
1941  * Read into the page cache. If a page already exists, and PageUptodate() is
1942  * not set, try to fill the page then wait for it to become unlocked.
1943  *
1944  * If the page does not get brought uptodate, return -EIO.
1945  */
1946 struct page *read_cache_page(struct address_space *mapping,
1947                                 pgoff_t index,
1948                                 int (*filler)(void *, struct page *),
1949                                 void *data)
1950 {
1951         return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
1952 }
1953 EXPORT_SYMBOL(read_cache_page);
1954
1955 static size_t __iovec_copy_from_user_inatomic(char *vaddr,
1956                         const struct iovec *iov, size_t base, size_t bytes)
1957 {
1958         size_t copied = 0, left = 0;
1959
1960         while (bytes) {
1961                 char __user *buf = iov->iov_base + base;
1962                 int copy = min(bytes, iov->iov_len - base);
1963
1964                 base = 0;
1965                 left = __copy_from_user_inatomic(vaddr, buf, copy);
1966                 copied += copy;
1967                 bytes -= copy;
1968                 vaddr += copy;
1969                 iov++;
1970
1971                 if (unlikely(left))
1972                         break;
1973         }
1974         return copied - left;
1975 }
1976
1977 /*
1978  * Copy as much as we can into the page and return the number of bytes which
1979  * were successfully copied.  If a fault is encountered then return the number of
1980  * bytes which were copied.
1981  */
1982 size_t iov_iter_copy_from_user_atomic(struct page *page,
1983                 struct iov_iter *i, unsigned long offset, size_t bytes)
1984 {
1985         char *kaddr;
1986         size_t copied;
1987
1988         BUG_ON(!in_atomic());
1989         kaddr = kmap_atomic(page);
1990         if (likely(i->nr_segs == 1)) {
1991                 int left;
1992                 char __user *buf = i->iov->iov_base + i->iov_offset;
1993                 left = __copy_from_user_inatomic(kaddr + offset, buf, bytes);
1994                 copied = bytes - left;
1995         } else {
1996                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
1997                                                 i->iov, i->iov_offset, bytes);
1998         }
1999         kunmap_atomic(kaddr);
2000
2001         return copied;
2002 }
2003 EXPORT_SYMBOL(iov_iter_copy_from_user_atomic);
2004
2005 /*
2006  * This has the same sideeffects and return value as
2007  * iov_iter_copy_from_user_atomic().
2008  * The difference is that it attempts to resolve faults.
2009  * Page must not be locked.
2010  */
2011 size_t iov_iter_copy_from_user(struct page *page,
2012                 struct iov_iter *i, unsigned long offset, size_t bytes)
2013 {
2014         char *kaddr;
2015         size_t copied;
2016
2017         kaddr = kmap(page);
2018         if (likely(i->nr_segs == 1)) {
2019                 int left;
2020                 char __user *buf = i->iov->iov_base + i->iov_offset;
2021                 left = __copy_from_user(kaddr + offset, buf, bytes);
2022                 copied = bytes - left;
2023         } else {
2024                 copied = __iovec_copy_from_user_inatomic(kaddr + offset,
2025                                                 i->iov, i->iov_offset, bytes);
2026         }
2027         kunmap(page);
2028         return copied;
2029 }
2030 EXPORT_SYMBOL(iov_iter_copy_from_user);
2031
2032 void iov_iter_advance(struct iov_iter *i, size_t bytes)
2033 {
2034         BUG_ON(i->count < bytes);
2035
2036         if (likely(i->nr_segs == 1)) {
2037                 i->iov_offset += bytes;
2038                 i->count -= bytes;
2039         } else {
2040                 const struct iovec *iov = i->iov;
2041                 size_t base = i->iov_offset;
2042                 unsigned long nr_segs = i->nr_segs;
2043
2044                 /*
2045                  * The !iov->iov_len check ensures we skip over unlikely
2046                  * zero-length segments (without overruning the iovec).
2047                  */
2048                 while (bytes || unlikely(i->count && !iov->iov_len)) {
2049                         int copy;
2050
2051                         copy = min(bytes, iov->iov_len - base);
2052                         BUG_ON(!i->count || i->count < copy);
2053                         i->count -= copy;
2054                         bytes -= copy;
2055                         base += copy;
2056                         if (iov->iov_len == base) {
2057                                 iov++;
2058                                 nr_segs--;
2059                                 base = 0;
2060                         }
2061                 }
2062                 i->iov = iov;
2063                 i->iov_offset = base;
2064                 i->nr_segs = nr_segs;
2065         }
2066 }
2067 EXPORT_SYMBOL(iov_iter_advance);
2068
2069 /*
2070  * Fault in the first iovec of the given iov_iter, to a maximum length
2071  * of bytes. Returns 0 on success, or non-zero if the memory could not be
2072  * accessed (ie. because it is an invalid address).
2073  *
2074  * writev-intensive code may want this to prefault several iovecs -- that
2075  * would be possible (callers must not rely on the fact that _only_ the
2076  * first iovec will be faulted with the current implementation).
2077  */
2078 int iov_iter_fault_in_readable(struct iov_iter *i, size_t bytes)
2079 {
2080         char __user *buf = i->iov->iov_base + i->iov_offset;
2081         bytes = min(bytes, i->iov->iov_len - i->iov_offset);
2082         return fault_in_pages_readable(buf, bytes);
2083 }
2084 EXPORT_SYMBOL(iov_iter_fault_in_readable);
2085
2086 /*
2087  * Return the count of just the current iov_iter segment.
2088  */
2089 size_t iov_iter_single_seg_count(const struct iov_iter *i)
2090 {
2091         const struct iovec *iov = i->iov;
2092         if (i->nr_segs == 1)
2093                 return i->count;
2094         else
2095                 return min(i->count, iov->iov_len - i->iov_offset);
2096 }
2097 EXPORT_SYMBOL(iov_iter_single_seg_count);
2098
2099 /*
2100  * Performs necessary checks before doing a write
2101  *
2102  * Can adjust writing position or amount of bytes to write.
2103  * Returns appropriate error code that caller should return or
2104  * zero in case that write should be allowed.
2105  */
2106 inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, int isblk)
2107 {
2108         struct inode *inode = file->f_mapping->host;
2109         unsigned long limit = rlimit(RLIMIT_FSIZE);
2110
2111         if (unlikely(*pos < 0))
2112                 return -EINVAL;
2113
2114         if (!isblk) {
2115                 /* FIXME: this is for backwards compatibility with 2.4 */
2116                 if (file->f_flags & O_APPEND)
2117                         *pos = i_size_read(inode);
2118
2119                 if (limit != RLIM_INFINITY) {
2120                         if (*pos >= limit) {
2121                                 send_sig(SIGXFSZ, current, 0);
2122                                 return -EFBIG;
2123                         }
2124                         if (*count > limit - (typeof(limit))*pos) {
2125                                 *count = limit - (typeof(limit))*pos;
2126                         }
2127                 }
2128         }
2129
2130         /*
2131          * LFS rule
2132          */
2133         if (unlikely(*pos + *count > MAX_NON_LFS &&
2134                                 !(file->f_flags & O_LARGEFILE))) {
2135                 if (*pos >= MAX_NON_LFS) {
2136                         return -EFBIG;
2137                 }
2138                 if (*count > MAX_NON_LFS - (unsigned long)*pos) {
2139                         *count = MAX_NON_LFS - (unsigned long)*pos;
2140                 }
2141         }
2142
2143         /*
2144          * Are we about to exceed the fs block limit ?
2145          *
2146          * If we have written data it becomes a short write.  If we have
2147          * exceeded without writing data we send a signal and return EFBIG.
2148          * Linus frestrict idea will clean these up nicely..
2149          */
2150         if (likely(!isblk)) {
2151                 if (unlikely(*pos >= inode->i_sb->s_maxbytes)) {
2152                         if (*count || *pos > inode->i_sb->s_maxbytes) {
2153                                 return -EFBIG;
2154                         }
2155                         /* zero-length writes at ->s_maxbytes are OK */
2156                 }
2157
2158                 if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
2159                         *count = inode->i_sb->s_maxbytes - *pos;
2160         } else {
2161 #ifdef CONFIG_BLOCK
2162                 loff_t isize;
2163                 if (bdev_read_only(I_BDEV(inode)))
2164                         return -EPERM;
2165                 isize = i_size_read(inode);
2166                 if (*pos >= isize) {
2167                         if (*count || *pos > isize)
2168                                 return -ENOSPC;
2169                 }
2170
2171                 if (*pos + *count > isize)
2172                         *count = isize - *pos;
2173 #else
2174                 return -EPERM;
2175 #endif
2176         }
2177         return 0;
2178 }
2179 EXPORT_SYMBOL(generic_write_checks);
2180
2181 int pagecache_write_begin(struct file *file, struct address_space *mapping,
2182                                 loff_t pos, unsigned len, unsigned flags,
2183                                 struct page **pagep, void **fsdata)
2184 {
2185         const struct address_space_operations *aops = mapping->a_ops;
2186
2187         return aops->write_begin(file, mapping, pos, len, flags,
2188                                                         pagep, fsdata);
2189 }
2190 EXPORT_SYMBOL(pagecache_write_begin);
2191
2192 int pagecache_write_end(struct file *file, struct address_space *mapping,
2193                                 loff_t pos, unsigned len, unsigned copied,
2194                                 struct page *page, void *fsdata)
2195 {
2196         const struct address_space_operations *aops = mapping->a_ops;
2197
2198         mark_page_accessed(page);
2199         return aops->write_end(file, mapping, pos, len, copied, page, fsdata);
2200 }
2201 EXPORT_SYMBOL(pagecache_write_end);
2202
2203 ssize_t
2204 generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
2205                 unsigned long *nr_segs, loff_t pos, loff_t *ppos,
2206                 size_t count, size_t ocount)
2207 {
2208         struct file     *file = iocb->ki_filp;
2209         struct address_space *mapping = file->f_mapping;
2210         struct inode    *inode = mapping->host;
2211         ssize_t         written;
2212         size_t          write_len;
2213         pgoff_t         end;
2214
2215         if (count != ocount)
2216                 *nr_segs = iov_shorten((struct iovec *)iov, *nr_segs, count);
2217
2218         write_len = iov_length(iov, *nr_segs);
2219         end = (pos + write_len - 1) >> PAGE_CACHE_SHIFT;
2220
2221         written = filemap_write_and_wait_range(mapping, pos, pos + write_len - 1);
2222         if (written)
2223                 goto out;
2224
2225         /*
2226          * After a write we want buffered reads to be sure to go to disk to get
2227          * the new data.  We invalidate clean cached page from the region we're
2228          * about to write.  We do this *before* the write so that we can return
2229          * without clobbering -EIOCBQUEUED from ->direct_IO().
2230          */
2231         if (mapping->nrpages) {
2232                 written = invalidate_inode_pages2_range(mapping,
2233                                         pos >> PAGE_CACHE_SHIFT, end);
2234                 /*
2235                  * If a page can not be invalidated, return 0 to fall back
2236                  * to buffered write.
2237                  */
2238                 if (written) {
2239                         if (written == -EBUSY)
2240                                 return 0;
2241                         goto out;
2242                 }
2243         }
2244
2245         written = mapping->a_ops->direct_IO(WRITE, iocb, iov, pos, *nr_segs);
2246
2247         /*
2248          * Finally, try again to invalidate clean pages which might have been
2249          * cached by non-direct readahead, or faulted in by get_user_pages()
2250          * if the source of the write was an mmap'ed region of the file
2251          * we're writing.  Either one is a pretty crazy thing to do,
2252          * so we don't support it 100%.  If this invalidation
2253          * fails, tough, the write still worked...
2254          */
2255         if (mapping->nrpages) {
2256                 invalidate_inode_pages2_range(mapping,
2257                                               pos >> PAGE_CACHE_SHIFT, end);
2258         }
2259
2260         if (written > 0) {
2261                 pos += written;
2262                 if (pos > i_size_read(inode) && !S_ISBLK(inode->i_mode)) {
2263                         i_size_write(inode, pos);
2264                         mark_inode_dirty(inode);
2265                 }
2266                 *ppos = pos;
2267         }
2268 out:
2269         return written;
2270 }
2271 EXPORT_SYMBOL(generic_file_direct_write);
2272
2273 /*
2274  * Find or create a page at the given pagecache position. Return the locked
2275  * page. This function is specifically for buffered writes.
2276  */
2277 struct page *grab_cache_page_write_begin(struct address_space *mapping,
2278                                         pgoff_t index, unsigned flags)
2279 {
2280         int status;
2281         gfp_t gfp_mask;
2282         struct page *page;
2283         gfp_t gfp_notmask = 0;
2284
2285         gfp_mask = mapping_gfp_mask(mapping);
2286         if (mapping_cap_account_dirty(mapping))
2287                 gfp_mask |= __GFP_WRITE;
2288         if (flags & AOP_FLAG_NOFS)
2289                 gfp_notmask = __GFP_FS;
2290 repeat:
2291         page = find_lock_page(mapping, index);
2292         if (page)
2293                 goto found;
2294
2295         page = __page_cache_alloc(gfp_mask & ~gfp_notmask);
2296         if (!page)
2297                 return NULL;
2298         status = add_to_page_cache_lru(page, mapping, index,
2299                                                 GFP_KERNEL & ~gfp_notmask);
2300         if (unlikely(status)) {
2301                 page_cache_release(page);
2302                 if (status == -EEXIST)
2303                         goto repeat;
2304                 return NULL;
2305         }
2306 found:
2307         wait_for_stable_page(page);
2308         return page;
2309 }
2310 EXPORT_SYMBOL(grab_cache_page_write_begin);
2311
2312 static ssize_t generic_perform_write(struct file *file,
2313                                 struct iov_iter *i, loff_t pos)
2314 {
2315         struct address_space *mapping = file->f_mapping;
2316         const struct address_space_operations *a_ops = mapping->a_ops;
2317         long status = 0;
2318         ssize_t written = 0;
2319         unsigned int flags = 0;
2320
2321         /*
2322          * Copies from kernel address space cannot fail (NFSD is a big user).
2323          */
2324         if (segment_eq(get_fs(), KERNEL_DS))
2325                 flags |= AOP_FLAG_UNINTERRUPTIBLE;
2326
2327         do {
2328                 struct page *page;
2329                 unsigned long offset;   /* Offset into pagecache page */
2330                 unsigned long bytes;    /* Bytes to write to page */
2331                 size_t copied;          /* Bytes copied from user */
2332                 void *fsdata;
2333
2334                 offset = (pos & (PAGE_CACHE_SIZE - 1));
2335                 bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2336                                                 iov_iter_count(i));
2337
2338 again:
2339                 /*
2340                  * Bring in the user page that we will copy from _first_.
2341                  * Otherwise there's a nasty deadlock on copying from the
2342                  * same page as we're writing to, without it being marked
2343                  * up-to-date.
2344                  *
2345                  * Not only is this an optimisation, but it is also required
2346                  * to check that the address is actually valid, when atomic
2347                  * usercopies are used, below.
2348                  */
2349                 if (unlikely(iov_iter_fault_in_readable(i, bytes))) {
2350                         status = -EFAULT;
2351                         break;
2352                 }
2353
2354                 status = a_ops->write_begin(file, mapping, pos, bytes, flags,
2355                                                 &page, &fsdata);
2356                 if (unlikely(status))
2357                         break;
2358
2359                 if (mapping_writably_mapped(mapping))
2360                         flush_dcache_page(page);
2361
2362                 pagefault_disable();
2363                 copied = iov_iter_copy_from_user_atomic(page, i, offset, bytes);
2364                 pagefault_enable();
2365                 flush_dcache_page(page);
2366
2367                 mark_page_accessed(page);
2368                 status = a_ops->write_end(file, mapping, pos, bytes, copied,
2369                                                 page, fsdata);
2370                 if (unlikely(status < 0))
2371                         break;
2372                 copied = status;
2373
2374                 cond_resched();
2375
2376                 iov_iter_advance(i, copied);
2377                 if (unlikely(copied == 0)) {
2378                         /*
2379                          * If we were unable to copy any data at all, we must
2380                          * fall back to a single segment length write.
2381                          *
2382                          * If we didn't fallback here, we could livelock
2383                          * because not all segments in the iov can be copied at
2384                          * once without a pagefault.
2385                          */
2386                         bytes = min_t(unsigned long, PAGE_CACHE_SIZE - offset,
2387                                                 iov_iter_single_seg_count(i));
2388                         goto again;
2389                 }
2390                 pos += copied;
2391                 written += copied;
2392
2393                 balance_dirty_pages_ratelimited(mapping);
2394                 if (fatal_signal_pending(current)) {
2395                         status = -EINTR;
2396                         break;
2397                 }
2398         } while (iov_iter_count(i));
2399
2400         return written ? written : status;
2401 }
2402
2403 ssize_t
2404 generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
2405                 unsigned long nr_segs, loff_t pos, loff_t *ppos,
2406                 size_t count, ssize_t written)
2407 {
2408         struct file *file = iocb->ki_filp;
2409         ssize_t status;
2410         struct iov_iter i;
2411
2412         iov_iter_init(&i, iov, nr_segs, count, written);
2413         status = generic_perform_write(file, &i, pos);
2414
2415         if (likely(status >= 0)) {
2416                 written += status;
2417                 *ppos = pos + status;
2418         }
2419         
2420         return written ? written : status;
2421 }
2422 EXPORT_SYMBOL(generic_file_buffered_write);
2423
2424 /**
2425  * __generic_file_aio_write - write data to a file
2426  * @iocb:       IO state structure (file, offset, etc.)
2427  * @iov:        vector with data to write
2428  * @nr_segs:    number of segments in the vector
2429  * @ppos:       position where to write
2430  *
2431  * This function does all the work needed for actually writing data to a
2432  * file. It does all basic checks, removes SUID from the file, updates
2433  * modification times and calls proper subroutines depending on whether we
2434  * do direct IO or a standard buffered write.
2435  *
2436  * It expects i_mutex to be grabbed unless we work on a block device or similar
2437  * object which does not need locking at all.
2438  *
2439  * This function does *not* take care of syncing data in case of O_SYNC write.
2440  * A caller has to handle it. This is mainly due to the fact that we want to
2441  * avoid syncing under i_mutex.
2442  */
2443 ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2444                                  unsigned long nr_segs, loff_t *ppos)
2445 {
2446         struct file *file = iocb->ki_filp;
2447         struct address_space * mapping = file->f_mapping;
2448         size_t ocount;          /* original count */
2449         size_t count;           /* after file limit checks */
2450         struct inode    *inode = mapping->host;
2451         loff_t          pos;
2452         ssize_t         written;
2453         ssize_t         err;
2454
2455         ocount = 0;
2456         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
2457         if (err)
2458                 return err;
2459
2460         count = ocount;
2461         pos = *ppos;
2462
2463         /* We can write back this queue in page reclaim */
2464         current->backing_dev_info = mapping->backing_dev_info;
2465         written = 0;
2466
2467         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2468         if (err)
2469                 goto out;
2470
2471         if (count == 0)
2472                 goto out;
2473
2474         err = file_remove_suid(file);
2475         if (err)
2476                 goto out;
2477
2478         err = file_update_time(file);
2479         if (err)
2480                 goto out;
2481
2482         /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
2483         if (unlikely(file->f_flags & O_DIRECT)) {
2484                 loff_t endbyte;
2485                 ssize_t written_buffered;
2486
2487                 written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
2488                                                         ppos, count, ocount);
2489                 if (written < 0 || written == count)
2490                         goto out;
2491                 /*
2492                  * direct-io write to a hole: fall through to buffered I/O
2493                  * for completing the rest of the request.
2494                  */
2495                 pos += written;
2496                 count -= written;
2497                 written_buffered = generic_file_buffered_write(iocb, iov,
2498                                                 nr_segs, pos, ppos, count,
2499                                                 written);
2500                 /*
2501                  * If generic_file_buffered_write() retuned a synchronous error
2502                  * then we want to return the number of bytes which were
2503                  * direct-written, or the error code if that was zero.  Note
2504                  * that this differs from normal direct-io semantics, which
2505                  * will return -EFOO even if some bytes were written.
2506                  */
2507                 if (written_buffered < 0) {
2508                         err = written_buffered;
2509                         goto out;
2510                 }
2511
2512                 /*
2513                  * We need to ensure that the page cache pages are written to
2514                  * disk and invalidated to preserve the expected O_DIRECT
2515                  * semantics.
2516                  */
2517                 endbyte = pos + written_buffered - written - 1;
2518                 err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
2519                 if (err == 0) {
2520                         written = written_buffered;
2521                         invalidate_mapping_pages(mapping,
2522                                                  pos >> PAGE_CACHE_SHIFT,
2523                                                  endbyte >> PAGE_CACHE_SHIFT);
2524                 } else {
2525                         /*
2526                          * We don't know how much we wrote, so just return
2527                          * the number of bytes which were direct-written
2528                          */
2529                 }
2530         } else {
2531                 written = generic_file_buffered_write(iocb, iov, nr_segs,
2532                                 pos, ppos, count, written);
2533         }
2534 out:
2535         current->backing_dev_info = NULL;
2536         return written ? written : err;
2537 }
2538 EXPORT_SYMBOL(__generic_file_aio_write);
2539
2540 /**
2541  * generic_file_aio_write - write data to a file
2542  * @iocb:       IO state structure
2543  * @iov:        vector with data to write
2544  * @nr_segs:    number of segments in the vector
2545  * @pos:        position in file where to write
2546  *
2547  * This is a wrapper around __generic_file_aio_write() to be used by most
2548  * filesystems. It takes care of syncing the file in case of O_SYNC file
2549  * and acquires i_mutex as needed.
2550  */
2551 ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2552                 unsigned long nr_segs, loff_t pos)
2553 {
2554         struct file *file = iocb->ki_filp;
2555         struct inode *inode = file->f_mapping->host;
2556         ssize_t ret;
2557
2558         BUG_ON(iocb->ki_pos != pos);
2559
2560         mutex_lock(&inode->i_mutex);
2561         ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
2562         mutex_unlock(&inode->i_mutex);
2563
2564         if (ret > 0) {
2565                 ssize_t err;
2566
2567                 err = generic_write_sync(file, pos, ret);
2568                 if (err < 0 && ret > 0)
2569                         ret = err;
2570         }
2571         return ret;
2572 }
2573 EXPORT_SYMBOL(generic_file_aio_write);
2574
2575 /**
2576  * try_to_release_page() - release old fs-specific metadata on a page
2577  *
2578  * @page: the page which the kernel is trying to free
2579  * @gfp_mask: memory allocation flags (and I/O mode)
2580  *
2581  * The address_space is to try to release any data against the page
2582  * (presumably at page->private).  If the release was successful, return `1'.
2583  * Otherwise return zero.
2584  *
2585  * This may also be called if PG_fscache is set on a page, indicating that the
2586  * page is known to the local caching routines.
2587  *
2588  * The @gfp_mask argument specifies whether I/O may be performed to release
2589  * this page (__GFP_IO), and whether the call may block (__GFP_WAIT & __GFP_FS).
2590  *
2591  */
2592 int try_to_release_page(struct page *page, gfp_t gfp_mask)
2593 {
2594         struct address_space * const mapping = page->mapping;
2595
2596         BUG_ON(!PageLocked(page));
2597         if (PageWriteback(page))
2598                 return 0;
2599
2600         if (mapping && mapping->a_ops->releasepage)
2601                 return mapping->a_ops->releasepage(page, gfp_mask);
2602         return try_to_free_buffers(page);
2603 }
2604
2605 EXPORT_SYMBOL(try_to_release_page);